We use flush cmd control to collect many flush cmds, and flush them
together. In this case, we use two list to manage the flush cmds
(collect and dispatch), and one spin lock is used to protect this.
In fact, the lock-less list(llist) is very suitable to this case,
and we use simplify this routine.
-
v2:
-use llist_for_each_entry_safe to fix possible use-after-free issue.
-remove the unused field from struct flush_cmd.
Thanks for Yu's suggestion.
-
Signed-off-by: Gu Zheng <guz.fnst@cn.fujitsu.com>
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
In commit aec71382c6 ("f2fs: refactor flush_nat_entries codes for reducing NAT
writes"), we descripte the issue as below:
"Although building NAT journal in cursum reduce the read/write work for NAT
block, but previous design leave us lower performance when write checkpoint
frequently for these cases:
1. if journal in cursum has already full, it's a bit of waste that we flush all
nat entries to page for persistence, but not to cache any entries.
2. if journal in cursum is not full, we fill nat entries to journal util
journal is full, then flush the left dirty entries to disk without merge
journaled entries, so these journaled entries may be flushed to disk at next
checkpoint but lost chance to flushed last time."
Actually, we have the same problem in using SIT journal area.
In this patch, firstly we will update sit journal with dirty entries as many as
possible. Secondly if there is no space in sit journal, we will remove all
entries in journal and walk through the whole dirty entry bitmap of sit,
accounting dirty sit entries located in same SIT block to sit entry set. All
entry sets are linked to list sit_entry_set in sm_info, sorted ascending order
by count of entries in set. Later we flush entries in set which have fewest
entries into journal as many as we can, and then flush dense set with merged
entries to disk.
In this way we can use sit journal area more effectively, also we will reduce
SIT update, result in gaining in performance and saving lifetime of flash
device.
In my testing environment, it shows this patch can help to reduce SIT block
update obviously.
virtual machine + hard disk:
fsstress -p 20 -n 400 -l 5
sit page num cp count sit pages/cp
based 2006.50 1349.75 1.486
patched 1566.25 1463.25 1.070
Our latency of merging op is small when handling a great number of dirty SIT
entries in flush_sit_entries:
latency(ns) dirty sit count
36038 2151
49168 2123
37174 2232
Signed-off-by: Chao Yu <chao2.yu@samsung.com>
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
I think we need to let the dirty node pages remain in the page cache instead
of rewriting them in their places.
So, after done with successful recovery, write_checkpoint will flush all of them
through the normal write path.
Through this, we can avoid potential error cases in terms of block allocation.
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
The init_inode_metadata calls truncate_blocks when error is occurred.
The callers holds f2fs_lock_op, so we should not call it again in
truncate_blocks.
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
The generic_shutdown_super calls sync_filesystem, evict_inode, and then
f2fs_put_super. In f2fs_evict_inode, we remain some dirty inode information
so we should release them at f2fs_put_super.
Reviewed-by: Chao Yu <chao2.yu@samsung.com>
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
This patch fixes not to skip xattr recovery and inline xattr/data recovery
order.
Reviewed-by: Chao Yu <chao2.yu@samsung.com>
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
This patch adds a parentheses to make clear for condition check.
And also it changes the return type for better meanings.
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
If mkwrite is called to an inode having inline_data, it can overwrite the data
index space as NEW_ADDR. (e.g., the first 4 bytes are coincidently zero)
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
Fix typo and some grammatical errors.
The words "filesystem" and "readahead" are being used without the space treewide.
Signed-off-by: Park Ju Hyung <qkrwngud825@gmail.com>
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
When we recover data of inode in roll-forward procedure, and the inode has both
inline data and inline xattr. We may skip recovering inline xattr if we recover
inline data form node page first.
This patch will fix the problem that we lost inline xattr data in above
scenario.
Signed-off-by: Chao Yu <chao2.yu@samsung.com>
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
We do not need to block on ->node_write among different node page writers e.g.
fsync/flush, unless we have a node page writer from write_checkpoint.
So it's better use rw_semaphore instead of mutex type for ->node_write to
promote performance.
Signed-off-by: Chao Yu <chao2.yu@samsung.com>
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
If the bit is already set, we don't need to reset it, and vice versa.
Because we don't need to make the caches dirty for that.
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
This patch enforces in-place-updates only when fdatasync is requested.
If we adopt this in-place-updates for the fdatasync, we can skip to write the
recovery information.
Reviewed-by: Chao Yu <chao2.yu@samsung.com>
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
This patch introduces a inode number list in which represents inodes having
appended data writes or updated data writes after last checkpoint.
This will be used at fsync to determine whether the recovery information
should be written or not.
Reviewed-by: Chao Yu <chao2.yu@samsung.com>
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
For better ino management, this patch replaces the data structure from list
to radix tree.
Reviewed-by: Chao Yu <chao2.yu@samsung.com>
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
This patch changes the naming of orphan-related data structures to use as
inode numbers managed globally.
Later, we can use this facility for managing any inode number lists.
Reviewed-by: Chao Yu <chao2.yu@samsung.com>
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
This patch adds a mount option, nobarrier, in f2fs.
The assumption in here is that file system keeps the IO ordering, but
doesn't care about cache flushes inside the storages.
Reviewed-by: Chao Yu <chao2.yu@samsung.com>
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
Although building NAT journal in cursum reduce the read/write work for NAT
block, but previous design leave us lower performance when write checkpoint
frequently for these cases:
1. if journal in cursum has already full, it's a bit of waste that we flush all
nat entries to page for persistence, but not to cache any entries.
2. if journal in cursum is not full, we fill nat entries to journal util
journal is full, then flush the left dirty entries to disk without merge
journaled entries, so these journaled entries may be flushed to disk at next
checkpoint but lost chance to flushed last time.
In this patch we merge dirty entries located in same NAT block to nat entry set,
and linked all set to list, sorted ascending order by entries' count of set.
Later we flush entries in sparse set into journal as many as we can, and then
flush merged entries to disk. In this way we can not only gain in performance,
but also save lifetime of flash device.
In my testing environment, it shows this patch can help to reduce NAT block
writes obviously. In hard disk test case: cost time of fsstress is stablely
reduced by about 5%.
1. virtual machine + hard disk:
fsstress -p 20 -n 200 -l 5
node num cp count nodes/cp
based 4599.6 1803.0 2.551
patched 2714.6 1829.6 1.483
2. virtual machine + 32g micro SD card:
fsstress -p 20 -n 200 -l 1 -w -f chown=0 -f creat=4 -f dwrite=0
-f fdatasync=4 -f fsync=4 -f link=0 -f mkdir=4 -f mknod=4 -f rename=5
-f rmdir=5 -f symlink=0 -f truncate=4 -f unlink=5 -f write=0 -S
node num cp count nodes/cp
based 84.5 43.7 1.933
patched 49.2 40.0 1.23
Our latency of merging op shows not bad when handling extreme case like:
merging a great number of dirty nats:
latency(ns) dirty nat count
3089219 24922
5129423 27422
4000250 24523
change log from v1:
o fix wrong logic in add_nat_entry when grab a new nat entry set.
o swith to create slab cache in create_node_manager_caches.
o use GFP_ATOMIC instead of GFP_NOFS to avoid potential long latency.
change log from v2:
o make comment position more appropriate suggested by Jaegeuk Kim.
Signed-off-by: Chao Yu <chao2.yu@samsung.com>
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
This patch adds f2fs_do_tmpfile to eliminate the redundant init_inode_metadata
flow.
Throught this, we can provide the consistent lock usage, e.g., fi->i_sem, and
this will enable better debugging stuffs.
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
This patch add lower bound verification for nid in check_nid_range, so nids
reserved like 0, node, meta passed by caller could be checked there.
And then check_nid_range could be used in f2fs_nfs_get_inode for simplifying
code.
Signed-off-by: Chao Yu <chao2.yu@samsung.com>
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
If data are overwritten through dio, previous f2fs doesn't remain the fsync mark
due to no additional node writes.
Note that this patch should resolve the xfstests:311.
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
f2fs's cp has one page which consists of struct f2fs_checkpoint and
version bitmap of sit and nat. To support lots of segments, we need more
blocks for sit bitmap. So let's arrange sit bitmap as following:
+-----------------+------------+
| f2fs_checkpoint | sit bitmap |
| + nat bitmap | |
+-----------------+------------+
0 4k N blocks
Signed-off-by: Changman Lee <cm224.lee@samsung.com>
[Jaegeuk Kim: simple code change for readability]
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
Previous we do not truncate inline data in inode page when setattr, so following
case could still read the inline data which has already truncated:
1.write inline data
2.ftruncate size to 0
3.ftruncate size to max inline data size
4.read from offset 0
This patch introduces truncate_inline_data() to fix this problem.
change log from v1:
o fix a bug and do not truncate first page data after truncate inline data.
Signed-off-by: Chao Yu <chao2.yu@samsung.com>
Signed-off-by: Jaegeuk Kim <jaegeuk.kim@samsung.com>
We have no so such readahead mechanism in ->iterate() path as the one in
->read() path, it cause low performance when we read large directory.
This patch add readahead in f2fs_readdir() for better performance.
Signed-off-by: Chao Yu <chao2.yu@samsung.com>
Signed-off-by: Jaegeuk Kim <jaegeuk.kim@samsung.com>
In This patch we introduce f2fs_seek_block to support SEEK_{DATA,HOLE} of
lseek(2).
change log from v1:
o fix bug when lseek from middle of page and fix wrong calculation of
PGOFS_OF_NEXT_DNODE macro.
Signed-off-by: Chao Yu <chao2.yu@samsung.com>
Signed-off-by: Jaegeuk Kim <jaegeuk.kim@samsung.com>
Introduce help function {create,destroy}_flush_cmd_control to clean up
the create/destory flush merge operation.
Signed-off-by: Gu Zheng <guz.fnst@cn.fujitsu.com>
Signed-off-by: Jaegeuk Kim <jaegeuk.kim@samsung.com>
Split the flush_merge fields from sm_i, and use the new struct flush_cmd_control
to wrap it, so that we can igonre these fileds if flush_merge is disable, and
it alse can the structs more neat.
Signed-off-by: Gu Zheng <guz.fnst@cn.fujitsu.com>
Signed-off-by: Jaegeuk Kim <jaegeuk.kim@samsung.com>
If so many dirty dentry blocks are cached, not reached to the flush condition,
we should fall into livelock in balance_dirty_pages.
So, let's consider the mem size for the condition.
Signed-off-by: Jaegeuk Kim <jaegeuk.kim@samsung.com>
This patch introduces available_nids for alloc_nids() and fixes max_nid for
build_free_nids() and scan_nat_pages().
Signed-off-by: Chao Yu <chao2.yu@samsung.com>
Reviewed-by: Chao Yu <chao2.yu@samsung.com>
Signed-off-by: Jaegeuk Kim <jaegeuk.kim@samsung.com>
Add the *remount* handle of flush_merge option, so that the users
can enable flush_merge in the runtime, such as the underlying device
handles the cache_flush command relatively slowly.
Signed-off-by: Gu Zheng <guz.fnst@cn.fujitsu.com>
Signed-off-by: Jaegeuk Kim <jaegeuk.kim@samsung.com>
This patch removes list opeations in handling dirty dir inodes.
Previously, F2FS traverses whole the list of dirty dir inodes to check whether
there is an existing inode or not, resulting in heavy CPU overheads.
So this patch removes such the traverse operations by adding FI_DIRTY_DIR to
indicate the inode lies on the list or not.
Through this simple flag, we can remove redundant operations gracefully.
Signed-off-by: Jaegeuk Kim <jaegeuk.kim@samsung.com>
The f2fs always scans the next chain of direct node blocks.
But some garbage blocks are able to be remained due to no discard support or
SSR triggers.
This occasionally wreaks recovering wrong inodes that were used or BUG_ONs
due to reallocating node ids as follows.
When mount this f2fs image:
http://linuxtesting.org/downloads/f2fs_fault_image.zip
BUG_ON is triggered in f2fs driver (messages below are generated on
kernel 3.13.2; for other kernels output is similar):
kernel BUG at fs/f2fs/node.c:215!
Call Trace:
[<ffffffffa032ebad>] recover_inode_page+0x1fd/0x3e0 [f2fs]
[<ffffffff811446e7>] ? __lock_page+0x67/0x70
[<ffffffff81089990>] ? autoremove_wake_function+0x50/0x50
[<ffffffffa0337788>] recover_fsync_data+0x1398/0x15d0 [f2fs]
[<ffffffff812b9e5c>] ? selinux_d_instantiate+0x1c/0x20
[<ffffffff811cb20b>] ? d_instantiate+0x5b/0x80
[<ffffffffa0321044>] f2fs_fill_super+0xb04/0xbf0 [f2fs]
[<ffffffff811b861e>] ? mount_bdev+0x7e/0x210
[<ffffffff811b8769>] mount_bdev+0x1c9/0x210
[<ffffffffa0320540>] ? validate_superblock+0x210/0x210 [f2fs]
[<ffffffffa031cf8d>] f2fs_mount+0x1d/0x30 [f2fs]
[<ffffffff811b9497>] mount_fs+0x47/0x1c0
[<ffffffff81166e00>] ? __alloc_percpu+0x10/0x20
[<ffffffff811d4032>] vfs_kern_mount+0x72/0x110
[<ffffffff811d6763>] do_mount+0x493/0x910
[<ffffffff811615cb>] ? strndup_user+0x5b/0x80
[<ffffffff811d6c70>] SyS_mount+0x90/0xe0
[<ffffffff8166f8d9>] system_call_fastpath+0x16/0x1b
Found by Linux File System Verification project (linuxtesting.org).
Reported-by: Andrey Tsyvarev <tsyvarev@ispras.ru>
Signed-off-by: Jaegeuk Kim <jaegeuk.kim@samsung.com>
Some storage devices show relatively high latencies to complete cache_flush
commands, even though their normal IO speed is prettry much high. In such
the case, it needs to merge cache_flush commands as much as possible to avoid
issuing them redundantly.
So, this patch introduces a mount option, "-o flush_merge", to mitigate such
the overhead.
If this option is enabled by user, F2FS merges the cache_flush commands and then
issues just one cache_flush on behalf of them. Once the single command is
finished, F2FS sends a completion signal to all the pending threads.
Note that, this option can be used under a workload consisting of very intensive
concurrent fsync calls, while the storage handles cache_flush commands slowly.
Signed-off-by: Jaegeuk Kim <jaegeuk.kim@samsung.com>
This patch introduce is_merged_page() to check whether current page is merged
in f2fs bio cache. When page is not in cache, we can avoid submitting bio cache,
resulting in having more chance to merge pages.
Signed-off-by: Chao Yu <chao2.yu@samsung.com>
Signed-off-by: Jaegeuk Kim <jaegeuk.kim@samsung.com>
If multiple redundant fsync calls are triggered, we don't need to write its
node pages with fsync mark continuously.
So, this patch adds FI_NEED_FSYNC to track whether the latest node block is
written with the fsync mark or not.
If the mark was set, a new fsync doesn't need to write a node block.
Otherwise, we should do a new node block with the mark for roll-forward
recovery.
Signed-off-by: Jaegeuk Kim <jaegeuk.kim@samsung.com>
This patch introduces fi->i_sem to protect fi's info that includes xattr_ver,
pino, i_nlink.
This enables to remove i_mutex during f2fs_sync_file, resulting in performance
improvement when a number of fsync calls are triggered from many concurrent
threads.
Signed-off-by: Jaegeuk Kim <jaegeuk.kim@samsung.com>
This patch introduces ram_thresh, a sysfs entry, which controls the memory
footprint used by the free nid list and the nat cache.
Previously, the free nid list was controlled by MAX_FREE_NIDS, while the nat
cache was managed by NM_WOUT_THRESHOLD.
However, this approach cannot be applied dynamically according to the system.
So, this patch adds ram_thresh that users can specify the threshold, which is
in order of 1 / 1024.
For example, if the total ram size is 4GB and the value is set to 10 by default,
f2fs tries to control the number of free nids and nat caches not to consume over
10 * (4GB / 1024) = 10MB.
Signed-off-by: Jaegeuk Kim <jaegeuk.kim@samsung.com>
Gu Zheng