linux/fs/xfs/libxfs/xfs_attr_remote.c

653 lines
16 KiB
C
Raw Normal View History

// SPDX-License-Identifier: GPL-2.0
/*
* Copyright (c) 2000-2005 Silicon Graphics, Inc.
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
* Copyright (c) 2013 Red Hat, Inc.
* All Rights Reserved.
*/
#include "xfs.h"
#include "xfs_fs.h"
#include "xfs_shared.h"
#include "xfs_format.h"
#include "xfs_log_format.h"
#include "xfs_trans_resv.h"
#include "xfs_bit.h"
#include "xfs_mount.h"
#include "xfs_defer.h"
#include "xfs_da_format.h"
#include "xfs_da_btree.h"
#include "xfs_inode.h"
#include "xfs_alloc.h"
#include "xfs_trans.h"
#include "xfs_inode_item.h"
#include "xfs_bmap.h"
#include "xfs_bmap_util.h"
#include "xfs_attr.h"
#include "xfs_attr_leaf.h"
#include "xfs_attr_remote.h"
#include "xfs_trans_space.h"
#include "xfs_trace.h"
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
#include "xfs_cksum.h"
#include "xfs_buf_item.h"
#include "xfs_error.h"
#define ATTR_RMTVALUE_MAPSIZE 1 /* # of map entries at once */
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
/*
* Each contiguous block has a header, so it is not just a simple attribute
* length to FSB conversion.
*/
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
int
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
xfs_attr3_rmt_blocks(
struct xfs_mount *mp,
int attrlen)
{
if (xfs_sb_version_hascrc(&mp->m_sb)) {
int buflen = XFS_ATTR3_RMT_BUF_SPACE(mp, mp->m_sb.sb_blocksize);
return (attrlen + buflen - 1) / buflen;
}
return XFS_B_TO_FSB(mp, attrlen);
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
}
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
/*
* Checking of the remote attribute header is split into two parts. The verifier
* does CRC, location and bounds checking, the unpacking function checks the
* attribute parameters and owner.
*/
static xfs_failaddr_t
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
xfs_attr3_rmt_hdr_ok(
void *ptr,
xfs_ino_t ino,
uint32_t offset,
uint32_t size,
xfs_daddr_t bno)
{
struct xfs_attr3_rmt_hdr *rmt = ptr;
if (bno != be64_to_cpu(rmt->rm_blkno))
return __this_address;
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
if (offset != be32_to_cpu(rmt->rm_offset))
return __this_address;
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
if (size != be32_to_cpu(rmt->rm_bytes))
return __this_address;
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
if (ino != be64_to_cpu(rmt->rm_owner))
return __this_address;
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
/* ok */
return NULL;
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
}
static xfs_failaddr_t
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
xfs_attr3_rmt_verify(
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
struct xfs_mount *mp,
void *ptr,
int fsbsize,
xfs_daddr_t bno)
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
{
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
struct xfs_attr3_rmt_hdr *rmt = ptr;
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
if (!xfs_sb_version_hascrc(&mp->m_sb))
return __this_address;
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
if (rmt->rm_magic != cpu_to_be32(XFS_ATTR3_RMT_MAGIC))
return __this_address;
if (!uuid_equal(&rmt->rm_uuid, &mp->m_sb.sb_meta_uuid))
return __this_address;
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
if (be64_to_cpu(rmt->rm_blkno) != bno)
return __this_address;
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
if (be32_to_cpu(rmt->rm_bytes) > fsbsize - sizeof(*rmt))
return __this_address;
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
if (be32_to_cpu(rmt->rm_offset) +
be32_to_cpu(rmt->rm_bytes) > XFS_XATTR_SIZE_MAX)
return __this_address;
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
if (rmt->rm_owner == 0)
return __this_address;
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
return NULL;
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
}
static int
__xfs_attr3_rmt_read_verify(
struct xfs_buf *bp,
bool check_crc,
xfs_failaddr_t *failaddr)
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
{
struct xfs_mount *mp = bp->b_target->bt_mount;
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
char *ptr;
int len;
xfs_daddr_t bno;
int blksize = mp->m_attr_geo->blksize;
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
/* no verification of non-crc buffers */
if (!xfs_sb_version_hascrc(&mp->m_sb))
return 0;
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
ptr = bp->b_addr;
bno = bp->b_bn;
len = BBTOB(bp->b_length);
ASSERT(len >= blksize);
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
while (len > 0) {
if (check_crc &&
!xfs_verify_cksum(ptr, blksize, XFS_ATTR3_RMT_CRC_OFF)) {
*failaddr = __this_address;
return -EFSBADCRC;
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
}
*failaddr = xfs_attr3_rmt_verify(mp, ptr, blksize, bno);
if (*failaddr)
return -EFSCORRUPTED;
len -= blksize;
ptr += blksize;
bno += BTOBB(blksize);
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
}
if (len != 0) {
*failaddr = __this_address;
return -EFSCORRUPTED;
}
return 0;
}
static void
xfs_attr3_rmt_read_verify(
struct xfs_buf *bp)
{
xfs_failaddr_t fa;
int error;
error = __xfs_attr3_rmt_read_verify(bp, true, &fa);
if (error)
xfs_verifier_error(bp, error, fa);
}
static xfs_failaddr_t
xfs_attr3_rmt_verify_struct(
struct xfs_buf *bp)
{
xfs_failaddr_t fa;
int error;
error = __xfs_attr3_rmt_read_verify(bp, false, &fa);
return error ? fa : NULL;
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
}
static void
xfs_attr3_rmt_write_verify(
struct xfs_buf *bp)
{
struct xfs_mount *mp = bp->b_target->bt_mount;
xfs_failaddr_t fa;
xfs: remote attribute headers contain an invalid LSN In recent testing, a system that crashed failed log recovery on restart with a bad symlink buffer magic number: XFS (vda): Starting recovery (logdev: internal) XFS (vda): Bad symlink block magic! XFS: Assertion failed: 0, file: fs/xfs/xfs_log_recover.c, line: 2060 On examination of the log via xfs_logprint, none of the symlink buffers in the log had a bad magic number, nor were any other types of buffer log format headers mis-identified as symlink buffers. Tracing was used to find the buffer the kernel was tripping over, and xfs_db identified it's contents as: 000: 5841524d 00000000 00000346 64d82b48 8983e692 d71e4680 a5f49e2c b317576e 020: 00000000 00602038 00000000 006034ce d0020000 00000000 4d4d4d4d 4d4d4d4d 040: 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d 060: 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d ..... This is a remote attribute buffer, which are notable in that they are not logged but are instead written synchronously by the remote attribute code so that they exist on disk before the attribute transactions are committed to the journal. The above remote attribute block has an invalid LSN in it - cycle 0xd002000, block 0 - which means when log recovery comes along to determine if the transaction that writes to the underlying block should be replayed, it sees a block that has a future LSN and so does not replay the buffer data in the transaction. Instead, it validates the buffer magic number and attaches the buffer verifier to it. It is this buffer magic number check that is failing in the above assert, indicating that we skipped replay due to the LSN of the underlying buffer. The problem here is that the remote attribute buffers cannot have a valid LSN placed into them, because the transaction that contains the attribute tree pointer changes and the block allocation that the attribute data is being written to hasn't yet been committed. Hence the LSN field in the attribute block is completely unwritten, thereby leaving the underlying contents of the block in the LSN field. It could have any value, and hence a future overwrite of the block by log recovery may or may not work correctly. Fix this by always writing an invalid LSN to the remote attribute block, as any buffer in log recovery that needs to write over the remote attribute should occur. We are protected from having old data written over the attribute by the fact that freeing the block before the remote attribute is written will result in the buffer being marked stale in the log and so all changes prior to the buffer stale transaction will be cancelled by log recovery. Hence it is safe to ignore the LSN in the case or synchronously written, unlogged metadata such as remote attribute blocks, and to ensure we do that correctly, we need to write an invalid LSN to all remote attribute blocks to trigger immediate recovery of metadata that is written over the top. As a further protection for filesystems that may already have remote attribute blocks with bad LSNs on disk, change the log recovery code to always trigger immediate recovery of metadata over remote attribute blocks. cc: <stable@vger.kernel.org> Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Brian Foster <bfoster@redhat.com> Signed-off-by: Dave Chinner <david@fromorbit.com>
2015-07-29 09:48:01 +08:00
int blksize = mp->m_attr_geo->blksize;
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
char *ptr;
int len;
xfs_daddr_t bno;
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
/* no verification of non-crc buffers */
if (!xfs_sb_version_hascrc(&mp->m_sb))
return;
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
ptr = bp->b_addr;
bno = bp->b_bn;
len = BBTOB(bp->b_length);
ASSERT(len >= blksize);
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
while (len > 0) {
xfs: remote attribute headers contain an invalid LSN In recent testing, a system that crashed failed log recovery on restart with a bad symlink buffer magic number: XFS (vda): Starting recovery (logdev: internal) XFS (vda): Bad symlink block magic! XFS: Assertion failed: 0, file: fs/xfs/xfs_log_recover.c, line: 2060 On examination of the log via xfs_logprint, none of the symlink buffers in the log had a bad magic number, nor were any other types of buffer log format headers mis-identified as symlink buffers. Tracing was used to find the buffer the kernel was tripping over, and xfs_db identified it's contents as: 000: 5841524d 00000000 00000346 64d82b48 8983e692 d71e4680 a5f49e2c b317576e 020: 00000000 00602038 00000000 006034ce d0020000 00000000 4d4d4d4d 4d4d4d4d 040: 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d 060: 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d ..... This is a remote attribute buffer, which are notable in that they are not logged but are instead written synchronously by the remote attribute code so that they exist on disk before the attribute transactions are committed to the journal. The above remote attribute block has an invalid LSN in it - cycle 0xd002000, block 0 - which means when log recovery comes along to determine if the transaction that writes to the underlying block should be replayed, it sees a block that has a future LSN and so does not replay the buffer data in the transaction. Instead, it validates the buffer magic number and attaches the buffer verifier to it. It is this buffer magic number check that is failing in the above assert, indicating that we skipped replay due to the LSN of the underlying buffer. The problem here is that the remote attribute buffers cannot have a valid LSN placed into them, because the transaction that contains the attribute tree pointer changes and the block allocation that the attribute data is being written to hasn't yet been committed. Hence the LSN field in the attribute block is completely unwritten, thereby leaving the underlying contents of the block in the LSN field. It could have any value, and hence a future overwrite of the block by log recovery may or may not work correctly. Fix this by always writing an invalid LSN to the remote attribute block, as any buffer in log recovery that needs to write over the remote attribute should occur. We are protected from having old data written over the attribute by the fact that freeing the block before the remote attribute is written will result in the buffer being marked stale in the log and so all changes prior to the buffer stale transaction will be cancelled by log recovery. Hence it is safe to ignore the LSN in the case or synchronously written, unlogged metadata such as remote attribute blocks, and to ensure we do that correctly, we need to write an invalid LSN to all remote attribute blocks to trigger immediate recovery of metadata that is written over the top. As a further protection for filesystems that may already have remote attribute blocks with bad LSNs on disk, change the log recovery code to always trigger immediate recovery of metadata over remote attribute blocks. cc: <stable@vger.kernel.org> Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Brian Foster <bfoster@redhat.com> Signed-off-by: Dave Chinner <david@fromorbit.com>
2015-07-29 09:48:01 +08:00
struct xfs_attr3_rmt_hdr *rmt = (struct xfs_attr3_rmt_hdr *)ptr;
fa = xfs_attr3_rmt_verify(mp, ptr, blksize, bno);
if (fa) {
xfs_verifier_error(bp, -EFSCORRUPTED, fa);
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
return;
}
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
xfs: remote attribute headers contain an invalid LSN In recent testing, a system that crashed failed log recovery on restart with a bad symlink buffer magic number: XFS (vda): Starting recovery (logdev: internal) XFS (vda): Bad symlink block magic! XFS: Assertion failed: 0, file: fs/xfs/xfs_log_recover.c, line: 2060 On examination of the log via xfs_logprint, none of the symlink buffers in the log had a bad magic number, nor were any other types of buffer log format headers mis-identified as symlink buffers. Tracing was used to find the buffer the kernel was tripping over, and xfs_db identified it's contents as: 000: 5841524d 00000000 00000346 64d82b48 8983e692 d71e4680 a5f49e2c b317576e 020: 00000000 00602038 00000000 006034ce d0020000 00000000 4d4d4d4d 4d4d4d4d 040: 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d 060: 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d ..... This is a remote attribute buffer, which are notable in that they are not logged but are instead written synchronously by the remote attribute code so that they exist on disk before the attribute transactions are committed to the journal. The above remote attribute block has an invalid LSN in it - cycle 0xd002000, block 0 - which means when log recovery comes along to determine if the transaction that writes to the underlying block should be replayed, it sees a block that has a future LSN and so does not replay the buffer data in the transaction. Instead, it validates the buffer magic number and attaches the buffer verifier to it. It is this buffer magic number check that is failing in the above assert, indicating that we skipped replay due to the LSN of the underlying buffer. The problem here is that the remote attribute buffers cannot have a valid LSN placed into them, because the transaction that contains the attribute tree pointer changes and the block allocation that the attribute data is being written to hasn't yet been committed. Hence the LSN field in the attribute block is completely unwritten, thereby leaving the underlying contents of the block in the LSN field. It could have any value, and hence a future overwrite of the block by log recovery may or may not work correctly. Fix this by always writing an invalid LSN to the remote attribute block, as any buffer in log recovery that needs to write over the remote attribute should occur. We are protected from having old data written over the attribute by the fact that freeing the block before the remote attribute is written will result in the buffer being marked stale in the log and so all changes prior to the buffer stale transaction will be cancelled by log recovery. Hence it is safe to ignore the LSN in the case or synchronously written, unlogged metadata such as remote attribute blocks, and to ensure we do that correctly, we need to write an invalid LSN to all remote attribute blocks to trigger immediate recovery of metadata that is written over the top. As a further protection for filesystems that may already have remote attribute blocks with bad LSNs on disk, change the log recovery code to always trigger immediate recovery of metadata over remote attribute blocks. cc: <stable@vger.kernel.org> Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Brian Foster <bfoster@redhat.com> Signed-off-by: Dave Chinner <david@fromorbit.com>
2015-07-29 09:48:01 +08:00
/*
* Ensure we aren't writing bogus LSNs to disk. See
* xfs_attr3_rmt_hdr_set() for the explanation.
*/
if (rmt->rm_lsn != cpu_to_be64(NULLCOMMITLSN)) {
xfs_verifier_error(bp, -EFSCORRUPTED, __this_address);
xfs: remote attribute headers contain an invalid LSN In recent testing, a system that crashed failed log recovery on restart with a bad symlink buffer magic number: XFS (vda): Starting recovery (logdev: internal) XFS (vda): Bad symlink block magic! XFS: Assertion failed: 0, file: fs/xfs/xfs_log_recover.c, line: 2060 On examination of the log via xfs_logprint, none of the symlink buffers in the log had a bad magic number, nor were any other types of buffer log format headers mis-identified as symlink buffers. Tracing was used to find the buffer the kernel was tripping over, and xfs_db identified it's contents as: 000: 5841524d 00000000 00000346 64d82b48 8983e692 d71e4680 a5f49e2c b317576e 020: 00000000 00602038 00000000 006034ce d0020000 00000000 4d4d4d4d 4d4d4d4d 040: 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d 060: 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d ..... This is a remote attribute buffer, which are notable in that they are not logged but are instead written synchronously by the remote attribute code so that they exist on disk before the attribute transactions are committed to the journal. The above remote attribute block has an invalid LSN in it - cycle 0xd002000, block 0 - which means when log recovery comes along to determine if the transaction that writes to the underlying block should be replayed, it sees a block that has a future LSN and so does not replay the buffer data in the transaction. Instead, it validates the buffer magic number and attaches the buffer verifier to it. It is this buffer magic number check that is failing in the above assert, indicating that we skipped replay due to the LSN of the underlying buffer. The problem here is that the remote attribute buffers cannot have a valid LSN placed into them, because the transaction that contains the attribute tree pointer changes and the block allocation that the attribute data is being written to hasn't yet been committed. Hence the LSN field in the attribute block is completely unwritten, thereby leaving the underlying contents of the block in the LSN field. It could have any value, and hence a future overwrite of the block by log recovery may or may not work correctly. Fix this by always writing an invalid LSN to the remote attribute block, as any buffer in log recovery that needs to write over the remote attribute should occur. We are protected from having old data written over the attribute by the fact that freeing the block before the remote attribute is written will result in the buffer being marked stale in the log and so all changes prior to the buffer stale transaction will be cancelled by log recovery. Hence it is safe to ignore the LSN in the case or synchronously written, unlogged metadata such as remote attribute blocks, and to ensure we do that correctly, we need to write an invalid LSN to all remote attribute blocks to trigger immediate recovery of metadata that is written over the top. As a further protection for filesystems that may already have remote attribute blocks with bad LSNs on disk, change the log recovery code to always trigger immediate recovery of metadata over remote attribute blocks. cc: <stable@vger.kernel.org> Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Brian Foster <bfoster@redhat.com> Signed-off-by: Dave Chinner <david@fromorbit.com>
2015-07-29 09:48:01 +08:00
return;
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
}
xfs_update_cksum(ptr, blksize, XFS_ATTR3_RMT_CRC_OFF);
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
len -= blksize;
ptr += blksize;
bno += BTOBB(blksize);
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
}
if (len != 0)
xfs_verifier_error(bp, -EFSCORRUPTED, __this_address);
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
}
const struct xfs_buf_ops xfs_attr3_rmt_buf_ops = {
.name = "xfs_attr3_rmt",
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
.verify_read = xfs_attr3_rmt_read_verify,
.verify_write = xfs_attr3_rmt_write_verify,
.verify_struct = xfs_attr3_rmt_verify_struct,
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
};
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
STATIC int
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
xfs_attr3_rmt_hdr_set(
struct xfs_mount *mp,
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
void *ptr,
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
xfs_ino_t ino,
uint32_t offset,
uint32_t size,
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
xfs_daddr_t bno)
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
{
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
struct xfs_attr3_rmt_hdr *rmt = ptr;
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
if (!xfs_sb_version_hascrc(&mp->m_sb))
return 0;
rmt->rm_magic = cpu_to_be32(XFS_ATTR3_RMT_MAGIC);
rmt->rm_offset = cpu_to_be32(offset);
rmt->rm_bytes = cpu_to_be32(size);
uuid_copy(&rmt->rm_uuid, &mp->m_sb.sb_meta_uuid);
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
rmt->rm_owner = cpu_to_be64(ino);
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
rmt->rm_blkno = cpu_to_be64(bno);
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
xfs: remote attribute headers contain an invalid LSN In recent testing, a system that crashed failed log recovery on restart with a bad symlink buffer magic number: XFS (vda): Starting recovery (logdev: internal) XFS (vda): Bad symlink block magic! XFS: Assertion failed: 0, file: fs/xfs/xfs_log_recover.c, line: 2060 On examination of the log via xfs_logprint, none of the symlink buffers in the log had a bad magic number, nor were any other types of buffer log format headers mis-identified as symlink buffers. Tracing was used to find the buffer the kernel was tripping over, and xfs_db identified it's contents as: 000: 5841524d 00000000 00000346 64d82b48 8983e692 d71e4680 a5f49e2c b317576e 020: 00000000 00602038 00000000 006034ce d0020000 00000000 4d4d4d4d 4d4d4d4d 040: 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d 060: 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d 4d4d4d4d ..... This is a remote attribute buffer, which are notable in that they are not logged but are instead written synchronously by the remote attribute code so that they exist on disk before the attribute transactions are committed to the journal. The above remote attribute block has an invalid LSN in it - cycle 0xd002000, block 0 - which means when log recovery comes along to determine if the transaction that writes to the underlying block should be replayed, it sees a block that has a future LSN and so does not replay the buffer data in the transaction. Instead, it validates the buffer magic number and attaches the buffer verifier to it. It is this buffer magic number check that is failing in the above assert, indicating that we skipped replay due to the LSN of the underlying buffer. The problem here is that the remote attribute buffers cannot have a valid LSN placed into them, because the transaction that contains the attribute tree pointer changes and the block allocation that the attribute data is being written to hasn't yet been committed. Hence the LSN field in the attribute block is completely unwritten, thereby leaving the underlying contents of the block in the LSN field. It could have any value, and hence a future overwrite of the block by log recovery may or may not work correctly. Fix this by always writing an invalid LSN to the remote attribute block, as any buffer in log recovery that needs to write over the remote attribute should occur. We are protected from having old data written over the attribute by the fact that freeing the block before the remote attribute is written will result in the buffer being marked stale in the log and so all changes prior to the buffer stale transaction will be cancelled by log recovery. Hence it is safe to ignore the LSN in the case or synchronously written, unlogged metadata such as remote attribute blocks, and to ensure we do that correctly, we need to write an invalid LSN to all remote attribute blocks to trigger immediate recovery of metadata that is written over the top. As a further protection for filesystems that may already have remote attribute blocks with bad LSNs on disk, change the log recovery code to always trigger immediate recovery of metadata over remote attribute blocks. cc: <stable@vger.kernel.org> Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Brian Foster <bfoster@redhat.com> Signed-off-by: Dave Chinner <david@fromorbit.com>
2015-07-29 09:48:01 +08:00
/*
* Remote attribute blocks are written synchronously, so we don't
* have an LSN that we can stamp in them that makes any sense to log
* recovery. To ensure that log recovery handles overwrites of these
* blocks sanely (i.e. once they've been freed and reallocated as some
* other type of metadata) we need to ensure that the LSN has a value
* that tells log recovery to ignore the LSN and overwrite the buffer
* with whatever is in it's log. To do this, we use the magic
* NULLCOMMITLSN to indicate that the LSN is invalid.
*/
rmt->rm_lsn = cpu_to_be64(NULLCOMMITLSN);
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
return sizeof(struct xfs_attr3_rmt_hdr);
}
/*
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
* Helper functions to copy attribute data in and out of the one disk extents
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
*/
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
STATIC int
xfs_attr_rmtval_copyout(
struct xfs_mount *mp,
struct xfs_buf *bp,
xfs_ino_t ino,
int *offset,
int *valuelen,
uint8_t **dst)
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
{
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
char *src = bp->b_addr;
xfs_daddr_t bno = bp->b_bn;
int len = BBTOB(bp->b_length);
int blksize = mp->m_attr_geo->blksize;
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
ASSERT(len >= blksize);
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
while (len > 0 && *valuelen > 0) {
int hdr_size = 0;
int byte_cnt = XFS_ATTR3_RMT_BUF_SPACE(mp, blksize);
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
byte_cnt = min(*valuelen, byte_cnt);
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
if (xfs_sb_version_hascrc(&mp->m_sb)) {
if (xfs_attr3_rmt_hdr_ok(src, ino, *offset,
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
byte_cnt, bno)) {
xfs_alert(mp,
"remote attribute header mismatch bno/off/len/owner (0x%llx/0x%x/Ox%x/0x%llx)",
bno, *offset, byte_cnt, ino);
return -EFSCORRUPTED;
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
}
hdr_size = sizeof(struct xfs_attr3_rmt_hdr);
}
memcpy(*dst, src + hdr_size, byte_cnt);
/* roll buffer forwards */
len -= blksize;
src += blksize;
bno += BTOBB(blksize);
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
/* roll attribute data forwards */
*valuelen -= byte_cnt;
*dst += byte_cnt;
*offset += byte_cnt;
}
return 0;
}
STATIC void
xfs_attr_rmtval_copyin(
struct xfs_mount *mp,
struct xfs_buf *bp,
xfs_ino_t ino,
int *offset,
int *valuelen,
uint8_t **src)
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
{
char *dst = bp->b_addr;
xfs_daddr_t bno = bp->b_bn;
int len = BBTOB(bp->b_length);
int blksize = mp->m_attr_geo->blksize;
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
ASSERT(len >= blksize);
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
while (len > 0 && *valuelen > 0) {
int hdr_size;
int byte_cnt = XFS_ATTR3_RMT_BUF_SPACE(mp, blksize);
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
byte_cnt = min(*valuelen, byte_cnt);
hdr_size = xfs_attr3_rmt_hdr_set(mp, dst, ino, *offset,
byte_cnt, bno);
memcpy(dst + hdr_size, *src, byte_cnt);
/*
* If this is the last block, zero the remainder of it.
* Check that we are actually the last block, too.
*/
if (byte_cnt + hdr_size < blksize) {
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
ASSERT(*valuelen - byte_cnt == 0);
ASSERT(len == blksize);
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
memset(dst + hdr_size + byte_cnt, 0,
blksize - hdr_size - byte_cnt);
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
}
/* roll buffer forwards */
len -= blksize;
dst += blksize;
bno += BTOBB(blksize);
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
/* roll attribute data forwards */
*valuelen -= byte_cnt;
*src += byte_cnt;
*offset += byte_cnt;
}
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
}
/*
* Read the value associated with an attribute from the out-of-line buffer
* that we stored it in.
*/
int
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
xfs_attr_rmtval_get(
struct xfs_da_args *args)
{
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
struct xfs_bmbt_irec map[ATTR_RMTVALUE_MAPSIZE];
struct xfs_mount *mp = args->dp->i_mount;
struct xfs_buf *bp;
xfs_dablk_t lblkno = args->rmtblkno;
uint8_t *dst = args->value;
xfs: remote attribute overwrite causes transaction overrun Commit e461fcb ("xfs: remote attribute lookups require the value length") passes the remote attribute length in the xfs_da_args structure on lookup so that CRC calculations and validity checking can be performed correctly by related code. This, unfortunately has the side effect of changing the args->valuelen parameter in cases where it shouldn't. That is, when we replace a remote attribute, the incoming replacement stores the value and length in args->value and args->valuelen, but then the lookup which finds the existing remote attribute overwrites args->valuelen with the length of the remote attribute being replaced. Hence when we go to create the new attribute, we create it of the size of the existing remote attribute, not the size it is supposed to be. When the new attribute is much smaller than the old attribute, this results in a transaction overrun and an ASSERT() failure on a debug kernel: XFS: Assertion failed: tp->t_blk_res_used <= tp->t_blk_res, file: fs/xfs/xfs_trans.c, line: 331 Fix this by keeping the remote attribute value length separate to the attribute value length in the xfs_da_args structure. The enables us to pass the length of the remote attribute to be removed without overwriting the new attribute's length. Also, ensure that when we save remote block contexts for a later rename we zero the original state variables so that we don't confuse the state of the attribute to be removes with the state of the new attribute that we just added. [Spotted by Brain Foster.] Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Brian Foster <bfoster@redhat.com> Signed-off-by: Dave Chinner <david@fromorbit.com>
2014-05-06 05:37:31 +08:00
int valuelen;
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
int nmap;
int error;
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
int blkcnt = args->rmtblkcnt;
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
int i;
int offset = 0;
trace_xfs_attr_rmtval_get(args);
ASSERT(!(args->flags & ATTR_KERNOVAL));
xfs: remote attribute overwrite causes transaction overrun Commit e461fcb ("xfs: remote attribute lookups require the value length") passes the remote attribute length in the xfs_da_args structure on lookup so that CRC calculations and validity checking can be performed correctly by related code. This, unfortunately has the side effect of changing the args->valuelen parameter in cases where it shouldn't. That is, when we replace a remote attribute, the incoming replacement stores the value and length in args->value and args->valuelen, but then the lookup which finds the existing remote attribute overwrites args->valuelen with the length of the remote attribute being replaced. Hence when we go to create the new attribute, we create it of the size of the existing remote attribute, not the size it is supposed to be. When the new attribute is much smaller than the old attribute, this results in a transaction overrun and an ASSERT() failure on a debug kernel: XFS: Assertion failed: tp->t_blk_res_used <= tp->t_blk_res, file: fs/xfs/xfs_trans.c, line: 331 Fix this by keeping the remote attribute value length separate to the attribute value length in the xfs_da_args structure. The enables us to pass the length of the remote attribute to be removed without overwriting the new attribute's length. Also, ensure that when we save remote block contexts for a later rename we zero the original state variables so that we don't confuse the state of the attribute to be removes with the state of the new attribute that we just added. [Spotted by Brain Foster.] Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Brian Foster <bfoster@redhat.com> Signed-off-by: Dave Chinner <david@fromorbit.com>
2014-05-06 05:37:31 +08:00
ASSERT(args->rmtvaluelen == args->valuelen);
xfs: remote attribute overwrite causes transaction overrun Commit e461fcb ("xfs: remote attribute lookups require the value length") passes the remote attribute length in the xfs_da_args structure on lookup so that CRC calculations and validity checking can be performed correctly by related code. This, unfortunately has the side effect of changing the args->valuelen parameter in cases where it shouldn't. That is, when we replace a remote attribute, the incoming replacement stores the value and length in args->value and args->valuelen, but then the lookup which finds the existing remote attribute overwrites args->valuelen with the length of the remote attribute being replaced. Hence when we go to create the new attribute, we create it of the size of the existing remote attribute, not the size it is supposed to be. When the new attribute is much smaller than the old attribute, this results in a transaction overrun and an ASSERT() failure on a debug kernel: XFS: Assertion failed: tp->t_blk_res_used <= tp->t_blk_res, file: fs/xfs/xfs_trans.c, line: 331 Fix this by keeping the remote attribute value length separate to the attribute value length in the xfs_da_args structure. The enables us to pass the length of the remote attribute to be removed without overwriting the new attribute's length. Also, ensure that when we save remote block contexts for a later rename we zero the original state variables so that we don't confuse the state of the attribute to be removes with the state of the new attribute that we just added. [Spotted by Brain Foster.] Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Brian Foster <bfoster@redhat.com> Signed-off-by: Dave Chinner <david@fromorbit.com>
2014-05-06 05:37:31 +08:00
valuelen = args->rmtvaluelen;
while (valuelen > 0) {
nmap = ATTR_RMTVALUE_MAPSIZE;
error = xfs_bmapi_read(args->dp, (xfs_fileoff_t)lblkno,
blkcnt, map, &nmap,
XFS_BMAPI_ATTRFORK);
if (error)
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
return error;
ASSERT(nmap >= 1);
for (i = 0; (i < nmap) && (valuelen > 0); i++) {
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
xfs_daddr_t dblkno;
int dblkcnt;
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
ASSERT((map[i].br_startblock != DELAYSTARTBLOCK) &&
(map[i].br_startblock != HOLESTARTBLOCK));
dblkno = XFS_FSB_TO_DADDR(mp, map[i].br_startblock);
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
dblkcnt = XFS_FSB_TO_BB(mp, map[i].br_blockcount);
error = xfs_trans_read_buf(mp, args->trans,
mp->m_ddev_targp,
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
dblkno, dblkcnt, 0, &bp,
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
&xfs_attr3_rmt_buf_ops);
if (error)
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
return error;
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
error = xfs_attr_rmtval_copyout(mp, bp, args->dp->i_ino,
&offset, &valuelen,
&dst);
xfs_trans_brelse(args->trans, bp);
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
if (error)
return error;
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
/* roll attribute extent map forwards */
lblkno += map[i].br_blockcount;
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
blkcnt -= map[i].br_blockcount;
}
}
ASSERT(valuelen == 0);
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
return 0;
}
/*
* Write the value associated with an attribute into the out-of-line buffer
* that we have defined for it.
*/
int
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
xfs_attr_rmtval_set(
struct xfs_da_args *args)
{
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
struct xfs_inode *dp = args->dp;
struct xfs_mount *mp = dp->i_mount;
struct xfs_bmbt_irec map;
xfs_dablk_t lblkno;
xfs_fileoff_t lfileoff = 0;
uint8_t *src = args->value;
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
int blkcnt;
int valuelen;
int nmap;
int error;
int offset = 0;
trace_xfs_attr_rmtval_set(args);
/*
* Find a "hole" in the attribute address space large enough for
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
* us to drop the new attribute's value into. Because CRC enable
* attributes have headers, we can't just do a straight byte to FSB
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
* conversion and have to take the header space into account.
*/
xfs: remote attribute overwrite causes transaction overrun Commit e461fcb ("xfs: remote attribute lookups require the value length") passes the remote attribute length in the xfs_da_args structure on lookup so that CRC calculations and validity checking can be performed correctly by related code. This, unfortunately has the side effect of changing the args->valuelen parameter in cases where it shouldn't. That is, when we replace a remote attribute, the incoming replacement stores the value and length in args->value and args->valuelen, but then the lookup which finds the existing remote attribute overwrites args->valuelen with the length of the remote attribute being replaced. Hence when we go to create the new attribute, we create it of the size of the existing remote attribute, not the size it is supposed to be. When the new attribute is much smaller than the old attribute, this results in a transaction overrun and an ASSERT() failure on a debug kernel: XFS: Assertion failed: tp->t_blk_res_used <= tp->t_blk_res, file: fs/xfs/xfs_trans.c, line: 331 Fix this by keeping the remote attribute value length separate to the attribute value length in the xfs_da_args structure. The enables us to pass the length of the remote attribute to be removed without overwriting the new attribute's length. Also, ensure that when we save remote block contexts for a later rename we zero the original state variables so that we don't confuse the state of the attribute to be removes with the state of the new attribute that we just added. [Spotted by Brain Foster.] Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Brian Foster <bfoster@redhat.com> Signed-off-by: Dave Chinner <david@fromorbit.com>
2014-05-06 05:37:31 +08:00
blkcnt = xfs_attr3_rmt_blocks(mp, args->rmtvaluelen);
error = xfs_bmap_first_unused(args->trans, args->dp, blkcnt, &lfileoff,
XFS_ATTR_FORK);
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
if (error)
return error;
args->rmtblkno = lblkno = (xfs_dablk_t)lfileoff;
args->rmtblkcnt = blkcnt;
/*
* Roll through the "value", allocating blocks on disk as required.
*/
while (blkcnt > 0) {
/*
* Allocate a single extent, up to the size of the value.
*
* Note that we have to consider this a data allocation as we
* write the remote attribute without logging the contents.
* Hence we must ensure that we aren't using blocks that are on
* the busy list so that we don't overwrite blocks which have
* recently been freed but their transactions are not yet
* committed to disk. If we overwrite the contents of a busy
* extent and then crash then the block may not contain the
* correct metadata after log recovery occurs.
*/
xfs_defer_init(args->trans->t_dfops, args->firstblock);
nmap = 1;
error = xfs_bmapi_write(args->trans, dp, (xfs_fileoff_t)lblkno,
blkcnt, XFS_BMAPI_ATTRFORK, args->firstblock,
args->total, &map, &nmap);
if (error)
goto out_defer_cancel;
xfs_defer_ijoin(args->trans->t_dfops, dp);
error = xfs_defer_finish(&args->trans, args->trans->t_dfops);
if (error)
goto out_defer_cancel;
ASSERT(nmap == 1);
ASSERT((map.br_startblock != DELAYSTARTBLOCK) &&
(map.br_startblock != HOLESTARTBLOCK));
lblkno += map.br_blockcount;
blkcnt -= map.br_blockcount;
/*
* Start the next trans in the chain.
*/
error = xfs_trans_roll_inode(&args->trans, dp);
if (error)
return error;
}
/*
* Roll through the "value", copying the attribute value to the
* already-allocated blocks. Blocks are written synchronously
* so that we can know they are all on disk before we turn off
* the INCOMPLETE flag.
*/
lblkno = args->rmtblkno;
blkcnt = args->rmtblkcnt;
xfs: remote attribute overwrite causes transaction overrun Commit e461fcb ("xfs: remote attribute lookups require the value length") passes the remote attribute length in the xfs_da_args structure on lookup so that CRC calculations and validity checking can be performed correctly by related code. This, unfortunately has the side effect of changing the args->valuelen parameter in cases where it shouldn't. That is, when we replace a remote attribute, the incoming replacement stores the value and length in args->value and args->valuelen, but then the lookup which finds the existing remote attribute overwrites args->valuelen with the length of the remote attribute being replaced. Hence when we go to create the new attribute, we create it of the size of the existing remote attribute, not the size it is supposed to be. When the new attribute is much smaller than the old attribute, this results in a transaction overrun and an ASSERT() failure on a debug kernel: XFS: Assertion failed: tp->t_blk_res_used <= tp->t_blk_res, file: fs/xfs/xfs_trans.c, line: 331 Fix this by keeping the remote attribute value length separate to the attribute value length in the xfs_da_args structure. The enables us to pass the length of the remote attribute to be removed without overwriting the new attribute's length. Also, ensure that when we save remote block contexts for a later rename we zero the original state variables so that we don't confuse the state of the attribute to be removes with the state of the new attribute that we just added. [Spotted by Brain Foster.] Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Brian Foster <bfoster@redhat.com> Signed-off-by: Dave Chinner <david@fromorbit.com>
2014-05-06 05:37:31 +08:00
valuelen = args->rmtvaluelen;
while (valuelen > 0) {
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
struct xfs_buf *bp;
xfs_daddr_t dblkno;
int dblkcnt;
ASSERT(blkcnt > 0);
xfs_defer_init(args->trans->t_dfops, args->firstblock);
nmap = 1;
error = xfs_bmapi_read(dp, (xfs_fileoff_t)lblkno,
blkcnt, &map, &nmap,
XFS_BMAPI_ATTRFORK);
if (error)
return error;
ASSERT(nmap == 1);
ASSERT((map.br_startblock != DELAYSTARTBLOCK) &&
(map.br_startblock != HOLESTARTBLOCK));
dblkno = XFS_FSB_TO_DADDR(mp, map.br_startblock),
dblkcnt = XFS_FSB_TO_BB(mp, map.br_blockcount);
bp = xfs_buf_get(mp->m_ddev_targp, dblkno, dblkcnt, 0);
if (!bp)
return -ENOMEM;
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
bp->b_ops = &xfs_attr3_rmt_buf_ops;
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
xfs_attr_rmtval_copyin(mp, bp, args->dp->i_ino, &offset,
&valuelen, &src);
error = xfs_bwrite(bp); /* GROT: NOTE: synchronous write */
xfs_buf_relse(bp);
if (error)
return error;
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
/* roll attribute extent map forwards */
lblkno += map.br_blockcount;
blkcnt -= map.br_blockcount;
}
ASSERT(valuelen == 0);
xfs: add CRC protection to remote attributes There are two ways of doing this - the first is to add a CRC to the remote attribute entry in the attribute block. The second is to treat them similar to the remote symlink, where each fragment has it's own header and identifies fragment location in the attribute. The problem with the CRC in the remote attr entry is that we cannot identify the owner of the metadata from the metadata blocks themselves, or where the blocks fit into the remote attribute. The down side to this approach is that we never know when the attribute has been read from disk or not and so we have to verify it every time it is read, and we must calculate it during the create transaction and log it. We do not log CRCs for any other metadata, and so this creates a unique set of coherency problems that, in general, are best avoided. Adding an identifying header to each allocated block allows us to identify each fragment and where in the attribute it is located. It enables us to rebuild the remote attribute from just the raw blocks containing the attribute. It also provides us to do per-block CRCs verification at IO time rather than during the transaction context that creates it or every time it is read into a user buffer. Hence it avoids all the problems that an external, logged CRC has, and provides all the benefits of self identifying metadata. The only complexity is that we have to add a header per fragment, and we don't know how many fragments will be needed prior to allocations. If we take the symlink example, the header is 56 bytes and hence for a 4k block size filesystem, in the worst case 16 headers requires 1 extra block for the 64k attribute data. For 512 byte filesystems the worst case is an extra block for every 9 fragments (i.e. 16 extra blocks in the worse case). This will be very rare and so it's not really a major concern. Because allocation is done in two steps - the first finds a hole large enough in the attribute file, the second does the allocation - we only need to find a hole big enough for a worst case allocation. We only need to allocate enough extra blocks for number of headers required by the fragments, and we can calculate that as we go.... Hence it really only makes sense to use the same model as for symlinks - it doesn't add that much complexity, does not require an attribute tree format change, and does not require logging calculated CRC values. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 13:11:28 +08:00
return 0;
out_defer_cancel:
xfs_defer_cancel(args->trans->t_dfops);
args->trans = NULL;
return error;
}
/*
* Remove the value associated with an attribute by deleting the
* out-of-line buffer that it is stored on.
*/
int
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
xfs_attr_rmtval_remove(
struct xfs_da_args *args)
{
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
struct xfs_mount *mp = args->dp->i_mount;
xfs_dablk_t lblkno;
int blkcnt;
int error;
int done;
trace_xfs_attr_rmtval_remove(args);
/*
* Roll through the "value", invalidating the attribute value's blocks.
*/
lblkno = args->rmtblkno;
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
blkcnt = args->rmtblkcnt;
while (blkcnt > 0) {
struct xfs_bmbt_irec map;
struct xfs_buf *bp;
xfs_daddr_t dblkno;
int dblkcnt;
int nmap;
/*
* Try to remember where we decided to put the value.
*/
nmap = 1;
error = xfs_bmapi_read(args->dp, (xfs_fileoff_t)lblkno,
blkcnt, &map, &nmap, XFS_BMAPI_ATTRFORK);
if (error)
return error;
ASSERT(nmap == 1);
ASSERT((map.br_startblock != DELAYSTARTBLOCK) &&
(map.br_startblock != HOLESTARTBLOCK));
dblkno = XFS_FSB_TO_DADDR(mp, map.br_startblock),
dblkcnt = XFS_FSB_TO_BB(mp, map.br_blockcount);
/*
* If the "remote" value is in the cache, remove it.
*/
bp = xfs_buf_incore(mp->m_ddev_targp, dblkno, dblkcnt, XBF_TRYLOCK);
if (bp) {
xfs_buf_stale(bp);
xfs_buf_relse(bp);
bp = NULL;
}
lblkno += map.br_blockcount;
blkcnt -= map.br_blockcount;
}
/*
* Keep de-allocating extents until the remote-value region is gone.
*/
lblkno = args->rmtblkno;
xfs: rework remote attr CRCs Note: this changes the on-disk remote attribute format. I assert that this is OK to do as CRCs are marked experimental and the first kernel it is included in has not yet reached release yet. Further, the userspace utilities are still evolving and so anyone using this stuff right now is a developer or tester using volatile filesystems for testing this feature. Hence changing the format right now to save longer term pain is the right thing to do. The fundamental change is to move from a header per extent in the attribute to a header per filesytem block in the attribute. This means there are more header blocks and the parsing of the attribute data is slightly more complex, but it has the advantage that we always know the size of the attribute on disk based on the length of the data it contains. This is where the header-per-extent method has problems. We don't know the size of the attribute on disk without first knowing how many extents are used to hold it. And we can't tell from a mapping lookup, either, because remote attributes can be allocated contiguously with other attribute blocks and so there is no obvious way of determining the actual size of the atribute on disk short of walking and mapping buffers. The problem with this approach is that if we map a buffer incorrectly (e.g. we make the last buffer for the attribute data too long), we then get buffer cache lookup failure when we map it correctly. i.e. we get a size mismatch on lookup. This is not necessarily fatal, but it's a cache coherency problem that can lead to returning the wrong data to userspace or writing the wrong data to disk. And debug kernels will assert fail if this occurs. I found lots of niggly little problems trying to fix this issue on a 4k block size filesystem, finally getting it to pass with lots of fixes. The thing is, 1024 byte filesystems still failed, and it was getting really complex handling all the corner cases that were showing up. And there were clearly more that I hadn't found yet. It is complex, fragile code, and if we don't fix it now, it will be complex, fragile code forever more. Hence the simple fix is to add a header to each filesystem block. This gives us the same relationship between the attribute data length and the number of blocks on disk as we have without CRCs - it's a linear mapping and doesn't require us to guess anything. It is simple to implement, too - the remote block count calculated at lookup time can be used by the remote attribute set/get/remove code without modification for both CRC and non-CRC filesystems. The world becomes sane again. Because the copy-in and copy-out now need to iterate over each filesystem block, I moved them into helper functions so we separate the block mapping and buffer manupulations from the attribute data and CRC header manipulations. The code becomes much clearer as a result, and it is a lot easier to understand and debug. It also appears to be much more robust - once it worked on 4k block size filesystems, it has worked without failure on 1k block size filesystems, too. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit ad1858d77771172e08016890f0eb2faedec3ecee)
2013-05-21 16:02:08 +08:00
blkcnt = args->rmtblkcnt;
done = 0;
while (!done) {
xfs_defer_init(args->trans->t_dfops, args->firstblock);
error = xfs_bunmapi(args->trans, args->dp, lblkno, blkcnt,
XFS_BMAPI_ATTRFORK, 1, args->firstblock,
&done);
if (error)
goto out_defer_cancel;
xfs_defer_ijoin(args->trans->t_dfops, args->dp);
error = xfs_defer_finish(&args->trans, args->trans->t_dfops);
if (error)
goto out_defer_cancel;
/*
* Close out trans and start the next one in the chain.
*/
error = xfs_trans_roll_inode(&args->trans, args->dp);
if (error)
return error;
}
return 0;
out_defer_cancel:
xfs_defer_cancel(args->trans->t_dfops);
args->trans = NULL;
return error;
}