The buffer cache hash is showing typical hash scalability problems.
In large scale testing the number of cached items growing far larger
than the hash can efficiently handle. Hence we need to move to a
self-scaling cache indexing mechanism.
I have selected rbtrees for indexing becuse they can have O(log n)
search scalability, and insert and remove cost is not excessive,
even on large trees. Hence we should be able to cache large numbers
of buffers without incurring the excessive cache miss search
penalties that the hash is imposing on us.
To ensure we still have parallel access to the cache, we need
multiple trees. Rather than hashing the buffers by disk address to
select a tree, it seems more sensible to separate trees by typical
access patterns. Most operations use buffers from within a single AG
at a time, so rather than searching lots of different lists,
separate the buffer indexes out into per-AG rbtrees. This means that
searches during metadata operation have a much higher chance of
hitting cache resident nodes, and that updates of the tree are less
likely to disturb trees being accessed on other CPUs doing
independent operations.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Alex Elder <aelder@sgi.com>
Memory reclaim via shrinkers has a terrible habit of having N+M
concurrent shrinker executions (N = num CPUs, M = num kswapds) all
trying to shrink the same cache. When the cache they are all working
on is protected by a single spinlock, massive contention an
slowdowns occur.
Wrap the per-ag inode caches with a reclaim mutex to serialise
reclaim access to the AG. This will block concurrent reclaim in each
AG but still allow reclaim to scan multiple AGs concurrently. Allow
shrinkers to move on to the next AG if it can't get the lock, and if
we can't get any AG, then start blocking on locks.
To prevent reclaimers from continually scanning the same inodes in
each AG, add a cursor that tracks where the last reclaim got up to
and start from that point on the next reclaim. This should avoid
only ever scanning a small number of inodes at the satart of each AG
and not making progress. If we have a non-shrinker based reclaim
pass, ignore the cursor and reset it to zero once we are done.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Alex Elder <aelder@sgi.com>
When we start taking a reference to the per-ag for every cached
buffer in the system, kernel lockstat profiling on an 8-way create
workload shows the mp->m_perag_lock has higher acquisition rates
than the inode lock and has significantly more contention. That is,
it becomes the highest contended lock in the system.
The perag lookup is trivial to convert to lock-less RCU lookups
because perag structures never go away. Hence the only thing we need
to protect against is tree structure changes during a grow. This can
be done simply by replacing the locking in xfs_perag_get() with RCU
read locking. This removes the mp->m_perag_lock completely from this
path.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Alex Elder <aelder@sgi.com>
If a filesystem is mounted without the inode64 mount option we
should still be able to access inodes not fitting into 32 bits, just
not created new ones. For this to work we need to make sure the
inode cache radix tree is initialized for all allocation groups, not
just those we plan to allocate inodes from. This patch makes sure
we initialize the inode cache radix tree for all allocation groups,
and also cleans xfs_initialize_perag up a bit to separate the
inode32 logical from the general perag structure setup.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
When we free a metadata extent, we record it in the per-AG busy
extent array so that it is not re-used before the freeing
transaction hits the disk. This array is fixed size, so when it
overflows we make further allocation transactions synchronous
because we cannot track more freed extents until those transactions
hit the disk and are completed. Under heavy mixed allocation and
freeing workloads with large log buffers, we can overflow this array
quite easily.
Further, the array is sparsely populated, which means that inserts
need to search for a free slot, and array searches often have to
search many more slots that are actually used to check all the
busy extents. Quite inefficient, really.
To enable this aspect of extent freeing to scale better, we need
a structure that can grow dynamically. While in other areas of
XFS we have used radix trees, the extents being freed are at random
locations on disk so are better suited to being indexed by an rbtree.
So, use a per-AG rbtree indexed by block number to track busy
extents. This incures a memory allocation when marking an extent
busy, but should not occur too often in low memory situations. This
should scale to an arbitrary number of extents so should not be a
limitation for features such as in-memory aggregation of
transactions.
However, there are still situations where we can't avoid allocating
busy extents (such as allocation from the AGFL). To minimise the
overhead of such occurences, we need to avoid doing a synchronous
log force while holding the AGF locked to ensure that the previous
transactions are safely on disk before we use the extent. We can do
this by marking the transaction doing the allocation as synchronous
rather issuing a log force.
Because of the locking involved and the ordering of transactions,
the synchronous transaction provides the same guarantees as a
synchronous log force because it ensures that all the prior
transactions are already on disk when the synchronous transaction
hits the disk. i.e. it preserves the free->allocate order of the
extent correctly in recovery.
By doing this, we avoid holding the AGF locked while log writes are
in progress, hence reducing the length of time the lock is held and
therefore we increase the rate at which we can allocate and free
from the allocation group, thereby increasing overall throughput.
The only problem with this approach is that when a metadata buffer is
marked stale (e.g. a directory block is removed), then buffer remains
pinned and locked until the log goes to disk. The issue here is that
if that stale buffer is reallocated in a subsequent transaction, the
attempt to lock that buffer in the transaction will hang waiting
the log to go to disk to unlock and unpin the buffer. Hence if
someone tries to lock a pinned, stale, locked buffer we need to
push on the log to get it unlocked ASAP. Effectively we are trading
off a guaranteed log force for a much less common trigger for log
force to occur.
Ideally we should not reallocate busy extents. That is a much more
complex fix to the problem as it involves direct intervention in the
allocation btree searches in many places. This is left to a future
set of modifications.
Finally, now that we track busy extents in allocated memory, we
don't need the descriptors in the transaction structure to point to
them. We can replace the complex busy chunk infrastructure with a
simple linked list of busy extents. This allows us to remove a large
chunk of code, making the overall change a net reduction in code
size.
Signed-off-by: Dave Chinner <david@fromorbit.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
On low memory boxes or those with highmem, kernel can OOM before the
background reclaims inodes via xfssyncd. Add a shrinker to run inode
reclaim so that it inode reclaim is expedited when memory is low.
This is more complex than it needs to be because the VM folk don't
want a context added to the shrinker infrastructure. Hence we need
to add a global list of XFS mount structures so the shrinker can
traverse them.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Now that the perag structure is allocated memory rather than held in
an array, we don't need to have the busy extent array external to
the structure. Embed it into the perag structure to avoid needing an
extra allocation when setting up.
Signed-off-by: Dave Chinner <david@fromorbit.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
Uninline xfs_perag_{get,put} so that tracepoints can be inserted
into them to speed debugging of reference count problems.
Signed-off-by: Dave Chinner <david@fromorbit.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
Reference count the per-ag structures to ensure that we keep get/put
pairs balanced. Assert that the reference counts are zero at unmount
time to catch leaks. In future, reference counts will enable us to
safely remove perag structures by allowing us to detect when they
are no longer in use.
Signed-off-by: Dave Chinner <david@fromorbit.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
Convert the old xfs tracing support that could only be used with the
out of tree kdb and xfsidbg patches to use the generic event tracer.
To use it make sure CONFIG_EVENT_TRACING is enabled and then enable
all xfs trace channels by:
echo 1 > /sys/kernel/debug/tracing/events/xfs/enable
or alternatively enable single events by just doing the same in one
event subdirectory, e.g.
echo 1 > /sys/kernel/debug/tracing/events/xfs/xfs_ihold/enable
or set more complex filters, etc. In Documentation/trace/events.txt
all this is desctribed in more detail. To reads the events do a
cat /sys/kernel/debug/tracing/trace
Compared to the last posting this patch converts the tracing mostly to
the one tracepoint per callsite model that other users of the new
tracing facility also employ. This allows a very fine-grained control
of the tracing, a cleaner output of the traces and also enables the
perf tool to use each tracepoint as a virtual performance counter,
allowing us to e.g. count how often certain workloads git various
spots in XFS. Take a look at
http://lwn.net/Articles/346470/
for some examples.
Also the btree tracing isn't included at all yet, as it will require
additional core tracing features not in mainline yet, I plan to
deliver it later.
And the really nice thing about this patch is that it actually removes
many lines of code while adding this nice functionality:
fs/xfs/Makefile | 8
fs/xfs/linux-2.6/xfs_acl.c | 1
fs/xfs/linux-2.6/xfs_aops.c | 52 -
fs/xfs/linux-2.6/xfs_aops.h | 2
fs/xfs/linux-2.6/xfs_buf.c | 117 +--
fs/xfs/linux-2.6/xfs_buf.h | 33
fs/xfs/linux-2.6/xfs_fs_subr.c | 3
fs/xfs/linux-2.6/xfs_ioctl.c | 1
fs/xfs/linux-2.6/xfs_ioctl32.c | 1
fs/xfs/linux-2.6/xfs_iops.c | 1
fs/xfs/linux-2.6/xfs_linux.h | 1
fs/xfs/linux-2.6/xfs_lrw.c | 87 --
fs/xfs/linux-2.6/xfs_lrw.h | 45 -
fs/xfs/linux-2.6/xfs_super.c | 104 ---
fs/xfs/linux-2.6/xfs_super.h | 7
fs/xfs/linux-2.6/xfs_sync.c | 1
fs/xfs/linux-2.6/xfs_trace.c | 75 ++
fs/xfs/linux-2.6/xfs_trace.h | 1369 +++++++++++++++++++++++++++++++++++++++++
fs/xfs/linux-2.6/xfs_vnode.h | 4
fs/xfs/quota/xfs_dquot.c | 110 ---
fs/xfs/quota/xfs_dquot.h | 21
fs/xfs/quota/xfs_qm.c | 40 -
fs/xfs/quota/xfs_qm_syscalls.c | 4
fs/xfs/support/ktrace.c | 323 ---------
fs/xfs/support/ktrace.h | 85 --
fs/xfs/xfs.h | 16
fs/xfs/xfs_ag.h | 14
fs/xfs/xfs_alloc.c | 230 +-----
fs/xfs/xfs_alloc.h | 27
fs/xfs/xfs_alloc_btree.c | 1
fs/xfs/xfs_attr.c | 107 ---
fs/xfs/xfs_attr.h | 10
fs/xfs/xfs_attr_leaf.c | 14
fs/xfs/xfs_attr_sf.h | 40 -
fs/xfs/xfs_bmap.c | 507 +++------------
fs/xfs/xfs_bmap.h | 49 -
fs/xfs/xfs_bmap_btree.c | 6
fs/xfs/xfs_btree.c | 5
fs/xfs/xfs_btree_trace.h | 17
fs/xfs/xfs_buf_item.c | 87 --
fs/xfs/xfs_buf_item.h | 20
fs/xfs/xfs_da_btree.c | 3
fs/xfs/xfs_da_btree.h | 7
fs/xfs/xfs_dfrag.c | 2
fs/xfs/xfs_dir2.c | 8
fs/xfs/xfs_dir2_block.c | 20
fs/xfs/xfs_dir2_leaf.c | 21
fs/xfs/xfs_dir2_node.c | 27
fs/xfs/xfs_dir2_sf.c | 26
fs/xfs/xfs_dir2_trace.c | 216 ------
fs/xfs/xfs_dir2_trace.h | 72 --
fs/xfs/xfs_filestream.c | 8
fs/xfs/xfs_fsops.c | 2
fs/xfs/xfs_iget.c | 111 ---
fs/xfs/xfs_inode.c | 67 --
fs/xfs/xfs_inode.h | 76 --
fs/xfs/xfs_inode_item.c | 5
fs/xfs/xfs_iomap.c | 85 --
fs/xfs/xfs_iomap.h | 8
fs/xfs/xfs_log.c | 181 +----
fs/xfs/xfs_log_priv.h | 20
fs/xfs/xfs_log_recover.c | 1
fs/xfs/xfs_mount.c | 2
fs/xfs/xfs_quota.h | 8
fs/xfs/xfs_rename.c | 1
fs/xfs/xfs_rtalloc.c | 1
fs/xfs/xfs_rw.c | 3
fs/xfs/xfs_trans.h | 47 +
fs/xfs/xfs_trans_buf.c | 62 -
fs/xfs/xfs_vnodeops.c | 8
70 files changed, 2151 insertions(+), 2592 deletions(-)
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
Don't search too far - abort if it is outside a certain radius and simply do
a linear search for the first free inode. In AGs with a million inodes this
can speed up allocation speed by 3-4x.
[hch: ported to the new xfs_ialloc.c world order]
Signed-off-by: Dave Chinner <dgc@sgi.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Alex Elder <aelder@sgi.com>
Signed-off-by: Felix Blyakher <felixb@sgi.com>
CONFIG_XFS_DEBUG builds still need xfs_read_agf to be
non-static, oops.
Signed-off-by: Eric Sandeen <sandeen@sandeen.net>
Reviewed-by: Felix Blyakher <felixb@sgi.com>
Signed-off-by: Felix Blyakher <felixb@sgi.com>
A lot more functions could be made static, but they need
forward declarations; this does some easy ones, and also
found a few unused functions in the process.
Signed-off-by: Eric Sandeen <sandeen@sandeen.net>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Felix Blyakher <felixb@sgi.com>
Given that we walk across the per-ag inode lists so often, it makes sense to
introduce an iterator for this.
Convert the sync and reclaim code to use this new iterator, quota code will
follow in the next patch.
Also change xfs_reclaim_inode to return -EGAIN instead of 1 for an inode
already under reclaim. This simplifies the AG iterator and doesn't
matter for the only other caller.
[hch: merged the lookup and execute callbacks back into one to get the
pag_ici_lock locking correct and simplify the code flow]
Signed-off-by: Dave Chinner <david@fromorbit.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Eric Sandeen <sandeen@sandeen.net>
Just another set of types obsfucating the code, remove them.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <david@fromorbit.com>
Remove the last of the macros-defined-to-static-functions.
Signed-off-by: Eric Sandeen <sandeen@sandeen.net>
Reviewed-by: Christoph Hellwig <hch@infradead.org>
Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
Add a helper to read the AGF header and perform basic verification.
Based on hunks from a larger patch from Dave Chinner.
(First sent on Juli 23rd)
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <david@fromorbit.com>
Signed-off-by: Niv Sardi <xaiki@sgi.com>
Add a helper to read the AGI header and perform basic verification.
Based on hunks from a larger patch from Dave Chinner.
(First sent on Juli 23rd)
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <david@fromorbit.com>
Signed-off-by: Niv Sardi <xaiki@sgi.com>
Prepare for removing the deleted inode list by marking inodes for reclaim
in the inode radix trees so that we can use the radix trees to find
reclaimable inodes.
SGI-PV: 988142
SGI-Modid: xfs-linux-melb:xfs-kern:32331a
Signed-off-by: David Chinner <david@fromorbit.com>
Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
Signed-off-by: Christoph Hellwig <hch@infradead.org>
Un-obfuscate pagb_lock, remove mutex_lock->spin_lock macros, call
spin_lock directly, remove extraneous cookie holdover from old xfs code,
and change lock type to spinlock_t.
SGI-PV: 970382
SGI-Modid: xfs-linux-melb:xfs-kern:29743a
Signed-off-by: Eric Sandeen <sandeen@sandeen.net>
Signed-off-by: Donald Douwsma <donaldd@sgi.com>
Signed-off-by: Tim Shimmin <tes@sgi.com>
One of the perpetual scaling problems XFS has is indexing it's incore
inodes. We currently uses hashes and the default hash sizes chosen can
only ever be a tradeoff between memory consumption and the maximum
realistic size of the cache.
As a result, anyone who has millions of inodes cached on a filesystem
needs to tunes the size of the cache via the ihashsize mount option to
allow decent scalability with inode cache operations.
A further problem is the separate inode cluster hash, whose size is based
on the ihashsize but is smaller, and so under certain conditions (sparse
cluster cache population) this can become a limitation long before the
inode hash is causing issues.
The following patchset removes the inode hash and cluster hash and
replaces them with radix trees to avoid the scalability limitations of the
hashes. It also reduces the size of the inodes by 3 pointers....
SGI-PV: 969561
SGI-Modid: xfs-linux-melb:xfs-kern:29481a
Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Christoph Hellwig <hch@infradead.org>
Signed-off-by: Tim Shimmin <tes@sgi.com>
In media spaces, video is often stored in a frame-per-file format. When
dealing with uncompressed realtime HD video streams in this format, it is
crucial that files do not get fragmented and that multiple files a placed
contiguously on disk.
When multiple streams are being ingested and played out at the same time,
it is critical that the filesystem does not cross the streams and
interleave them together as this creates seek and readahead cache miss
latency and prevents both ingest and playout from meeting frame rate
targets.
This patch set creates a "stream of files" concept into the allocator to
place all the data from a single stream contiguously on disk so that RAID
array readahead can be used effectively. Each additional stream gets
placed in different allocation groups within the filesystem, thereby
ensuring that we don't cross any streams. When an AG fills up, we select a
new AG for the stream that is not in use.
The core of the functionality is the stream tracking - each inode that we
create in a directory needs to be associated with the directories' stream.
Hence every time we create a file, we look up the directories' stream
object and associate the new file with that object.
Once we have a stream object for a file, we use the AG that the stream
object point to for allocations. If we can't allocate in that AG (e.g. it
is full) we move the entire stream to another AG. Other inodes in the same
stream are moved to the new AG on their next allocation (i.e. lazy
update).
Stream objects are kept in a cache and hold a reference on the inode.
Hence the inode cannot be reclaimed while there is an outstanding stream
reference. This means that on unlink we need to remove the stream
association and we also need to flush all the associations on certain
events that want to reclaim all unreferenced inodes (e.g. filesystem
freeze).
SGI-PV: 964469
SGI-Modid: xfs-linux-melb:xfs-kern:29096a
Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Barry Naujok <bnaujok@sgi.com>
Signed-off-by: Donald Douwsma <donaldd@sgi.com>
Signed-off-by: Christoph Hellwig <hch@infradead.org>
Signed-off-by: Tim Shimmin <tes@sgi.com>
Signed-off-by: Vlad Apostolov <vapo@sgi.com>
When we have a couple of hundred transactions on the fly at once, they all
typically modify the on disk superblock in some way.
create/unclink/mkdir/rmdir modify inode counts, allocation/freeing modify
free block counts.
When these counts are modified in a transaction, they must eventually lock
the superblock buffer and apply the mods. The buffer then remains locked
until the transaction is committed into the incore log buffer. The result
of this is that with enough transactions on the fly the incore superblock
buffer becomes a bottleneck.
The result of contention on the incore superblock buffer is that
transaction rates fall - the more pressure that is put on the superblock
buffer, the slower things go.
The key to removing the contention is to not require the superblock fields
in question to be locked. We do that by not marking the superblock dirty
in the transaction. IOWs, we modify the incore superblock but do not
modify the cached superblock buffer. In short, we do not log superblock
modifications to critical fields in the superblock on every transaction.
In fact we only do it just before we write the superblock to disk every
sync period or just before unmount.
This creates an interesting problem - if we don't log or write out the
fields in every transaction, then how do the values get recovered after a
crash? the answer is simple - we keep enough duplicate, logged information
in other structures that we can reconstruct the correct count after log
recovery has been performed.
It is the AGF and AGI structures that contain the duplicate information;
after recovery, we walk every AGI and AGF and sum their individual
counters to get the correct value, and we do a transaction into the log to
correct them. An optimisation of this is that if we have a clean unmount
record, we know the value in the superblock is correct, so we can avoid
the summation walk under normal conditions and so mount/recovery times do
not change under normal operation.
One wrinkle that was discovered during development was that the blocks
used in the freespace btrees are never accounted for in the AGF counters.
This was once a valid optimisation to make; when the filesystem is full,
the free space btrees are empty and consume no space. Hence when it
matters, the "accounting" is correct. But that means the when we do the
AGF summations, we would not have a correct count and xfs_check would
complain. Hence a new counter was added to track the number of blocks used
by the free space btrees. This is an *on-disk format change*.
As a result of this, lazy superblock counters are a mkfs option and at the
moment on linux there is no way to convert an old filesystem. This is
possible - xfs_db can be used to twiddle the right bits and then
xfs_repair will do the format conversion for you. Similarly, you can
convert backwards as well. At some point we'll add functionality to
xfs_admin to do the bit twiddling easily....
SGI-PV: 964999
SGI-Modid: xfs-linux-melb:xfs-kern:28652a
Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Christoph Hellwig <hch@infradead.org>
Signed-off-by: Tim Shimmin <tes@sgi.com>
used for ondisk values.
SGI-PV: 954580
SGI-Modid: xfs-linux-melb:xfs-kern:26553a
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Nathan Scott <nathans@sgi.com>
Signed-off-by: Tim Shimmin <tes@sgi.com>
Initial git repository build. I'm not bothering with the full history,
even though we have it. We can create a separate "historical" git
archive of that later if we want to, and in the meantime it's about
3.2GB when imported into git - space that would just make the early
git days unnecessarily complicated, when we don't have a lot of good
infrastructure for it.
Let it rip!
Dave Chinner