proc A: proc B:
- writeback_sb_inodes
- __writeback_single_inode
- do_writepages
- f2fs_write_node_pages
- f2fs_balance_fs_bg - write_checkpoint
- build_free_nids - flush_nat_entries
- __build_free_nids - __flush_nat_entry_set
- ra_meta_pages - get_next_nat_page
- current_nat_addr - set_to_next_nat
[do nat_bitmap checking] - f2fs_change_bit
For proc A, nat_bitmap and nat_bitmap_mir would be compared without lock_op and
nm_i->nat_tree_lock, while proc B is changing nat_bitmap/nat_bitmap_ver in cp.
So it is normal for nat_bitmap/nat_bitmap diffrence under such scenario.
This patch fix this by removing the monitoring point.
[Fix: 599a09b f2fs: check in-memory nat version bitmap]
Signed-off-by: Hou Pengyang <houpengyang@huawei.com>
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
Currently, if we call fsync after updating the xattr date belongs to the
file, f2fs needs to trigger checkpoint to keep xattr data consistent. But,
this policy cause low performance as checkpoint will block most foreground
operations and cause unneeded and unrelated IOs around checkpoint.
This patch will reuse regular file recovery policy for xattr node block,
so, we change to write xattr node block tagged with fsync flag to warm
area instead of cold area, and during recovery, we search warm node chain
for fsynced xattr block, and do the recovery.
So, for below application IO pattern, performance can be improved
obviously:
- touch file
- create/update/delete xattr entry in file
- fsync file
Signed-off-by: Chao Yu <yuchao0@huawei.com>
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
This patch adds a mirror for nat version bitmap, and use it to detect
in-memory bitmap corruption which may be caused by bit-transition of
cache or memory overflow.
Signed-off-by: Chao Yu <yuchao0@huawei.com>
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
During free nid allocation, in order to do preallocation, we will tag free
nid entry as allocated one and still leave it in free nid list, for other
allocators who want to grab free nids, it needs to traverse the free nid
list for lookup. It becomes overhead in scenario of allocating free nid
intensively by multithreads.
This patch splits free nid list to two list: {free,alloc}_nid_list, to
keep free nids and preallocated free nids separately, after that, traverse
latency will be gone, besides split nid_cnt for separate statistic.
Additionally, introduce __insert_nid_to_list and __remove_nid_from_list for
cleanup.
Signed-off-by: Chao Yu <yuchao0@huawei.com>
[Jaegeuk Kim: modify f2fs_bug_on to avoid needless branches]
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
f2fs contained a number of endianness conversion bugs.
Also, one function should have been 'static'.
Found with sparse by running 'make C=2 CF=-D__CHECK_ENDIAN__ fs/f2fs/'
Signed-off-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
This patch introduces spinlock to protect updating process of ckpt_flags
field in struct f2fs_checkpoint, it avoids incorrectly updating in race
condition.
Signed-off-by: Chao Yu <yuchao0@huawei.com>
[Jaegeuk Kim: add __is_set_ckpt_flags likewise __set_ckpt_flags]
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
Previously, we used cp_version only to detect recoverable dnodes.
In order to avoid same garbage cp_version, we needed to truncate the next
dnode during checkpoint, resulting in additional discard or data write.
If we can distinguish this by using crc in addition to cp_version, we can
remove this overhead.
There is backward compatibility concern where it changes node_footer layout.
So, this patch introduces a new checkpoint flag, CP_CRC_RECOVERY_FLAG, to
detect new layout. New layout will be activated only when this flag is set.
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
The readahead nat pages are more likely to be reclaimed quickly, so it'd better
to gather more free nids in advance.
And, let's keep some free nids as much as possible.
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
In write_begin, if storage supports stable_page, we don't need to wait for
writeback to update its contents.
This patch introduces to use wait_for_stable_page instead of
wait_on_page_writeback.
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
The sceanrio is:
1. create fully node blocks
2. flush node blocks
3. write inline_data for all the node blocks again
4. flush node blocks redundantly
So, this patch tries to flush inline_data when flushing node blocks.
Reviewed-by: Chao Yu <chao2.yu@samsung.com>
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
This patch exports a new sysfs entry 'dirty_nat_ratio' to control threshold
of dirty nat entries, if current ratio exceeds configured threshold,
checkpoint will be triggered in f2fs_balance_fs_bg for flushing dirty nats.
Signed-off-by: Chao Yu <chao2.yu@samsung.com>
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
When testing f2fs with xfstest, generic/251 is stuck for long time,
the case uses below serials to obtain fresh released space in device,
in order to prepare for following fstrim test.
1. rm -rf /mnt/dir
2. mkdir /mnt/dir/
3. cp -axT `pwd`/ /mnt/dir/
4. goto 1
During preparing step, all nat entries will be cached in nat cache,
most of them are dirty entries with invalid blkaddr, which means
nodes related to these entries have been truncated, and they could
be reused after the dirty entries been checkpointed.
However, there was no checkpoint been triggered, so nid allocators
(e.g. mkdir, creat) will run into long journey of iterating all NAT
pages, looking for free nids in alloc_nid->build_free_nids.
Here, in f2fs_balance_fs_bg we give another chance to do checkpoint
to flush nat entries for reusing them in free nid cache when dirty
entry count exceeds 10% of max count.
Signed-off-by: Chao Yu <chao2.yu@samsung.com>
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
Use sbi->blocks_per_seg directly to avoid unnecessary calculation when using
1 << sbi->log_blocks_per_seg.
Signed-off-by: Chao Yu <chao2.yu@samsung.com>
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
After finishing building free nid cache, we will try to readahead
asynchronously 4 more pages for the next reloading, the count of
readahead nid pages is fixed.
In some case, like SMR drive, read less sectors with fixed count
each time we trigger RA may be low efficient, since we will face
high seeking overhead, so we'd better let user to configure this
parameter from sysfs in specific workload.
Signed-off-by: Chao Yu <chao2.yu@samsung.com>
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
The periodic checkpoint can resolve the previous issue.
So, now we can use this again to improve the reported performance regression:
https://lkml.org/lkml/2015/10/8/20
This reverts commit 15bec0ff5a9ba6d203178fa8772259df6207942a.
Previously, we skip dentry block writes when wbc is SYNC_NONE with no memory
pressure and the number of dirty pages is pretty small.
But, we didn't skip for normal data writes, which gives us not much big impact
on overall performance.
Moreover, by skipping some data writes, kworker falls into infinite loop to try
to write blocks, when many dir inodes have only one dentry block.
So, this patch removes skipping data writes.
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
Introduce infra macro and data structure for rb-tree based extent cache:
Macros:
* EXT_TREE_VEC_SIZE: indicate vector size for gang lookup in extent tree.
* F2FS_MIN_EXTENT_LEN: indicate minimum length of extent managed in cache.
* EXTENT_CACHE_SHRINK_NUMBER: indicate number of extent in cache will be shrunk.
Basic data structures for extent cache:
* struct extent_tree: extent tree entry per inode.
* struct extent_node: extent info node linked in extent tree.
Besides, adding new extent cache related fields in f2fs_sb_info.
Signed-off-by: Chao Yu <chao2.yu@samsung.com>
Signed-off-by: Changman Lee <cm224.lee@samsung.com>
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
In the normal case, the radix_tree_nodes are freed successfully.
But, when cp_error was detected, we should destroy them forcefully.
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
In do_recover_data, we find and update previous node pages after updating
its new block addresses.
After then, we call fill_node_footer without reset field, we erase its
cold bit so that this new cold node block is written to wrong log area.
This patch fixes not to miss its old flag.
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
This patch moves one member of struct nat_entry: _flag_ to struct node_info,
so _version_ in struct node_info and _flag_ which are unsigned char type will
merge to one 32-bit space in register/memory. So the size of nat_entry will be
reduced from 28 bytes to 24 bytes (for 64-bit machine, reduce its size from 40
bytes to 32 bytes) and then slab memory using by f2fs will be reduced.
changes from v2:
o update description of memory usage gain for 64-bit machine suggested by
Changman Lee.
changes from v1:
o introduce inline copy_node_info() to copy valid data from node info suggested
by Jaegeuk Kim, it can avoid bug.
Reviewed-by: Changman Lee <cm224.lee@samsung.com>
Signed-off-by: Chao Yu <chao2.yu@samsung.com>
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
This patch adds two new ioctls to release inmemory pages grabbed by atomic
writes.
o f2fs_ioc_abort_volatile_write
- If transaction was failed, all the grabbed pages and data should be written.
o f2fs_ioc_release_volatile_write
- This is to enhance the performance of PERSIST mode in sqlite.
In order to avoid huge memory consumption which causes OOM, this patch changes
volatile writes to use normal dirty pages, instead blocked flushing to the disk
as long as system does not suffer from memory pressure.
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
This patch adds to control the memory footprint used by ino entries.
This will conduct best effort, not strictly.
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
Introduce f2fs_change_bit to simplify the change bit logic in
function set_to_next_nat{sit}.
Signed-off-by: Gu Zheng <guz.fnst@cn.fujitsu.com>
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
Previously, f2fs tries to reorganize the dirty nat entries into multiple sets
according to its nid ranges. This can improve the flushing nat pages, however,
if there are a lot of cached nat entries, it becomes a bottleneck.
This patch introduces a new set management flow by removing dirty nat list and
adding a series of set operations when the nat entry becomes dirty.
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
This patch revisited whole the recovery information during the f2fs_sync_file.
In this patch, there are three information to make a decision.
a) IS_CHECKPOINTED, /* is it checkpointed before? */
b) HAS_FSYNCED_INODE, /* is the inode fsynced before? */
c) HAS_LAST_FSYNC, /* has the latest node fsync mark? */
And, the scenarios for our rule are based on:
[Term] F: fsync_mark, D: dentry_mark
1. inode(x) | CP | inode(x) | dnode(F)
2. inode(x) | CP | inode(F) | dnode(F)
3. inode(x) | CP | dnode(F) | inode(x) | inode(F)
4. inode(x) | CP | dnode(F) | inode(F)
5. CP | inode(x) | dnode(F) | inode(DF)
6. CP | inode(DF) | dnode(F)
7. CP | dnode(F) | inode(DF)
8. CP | dnode(F) | inode(x) | inode(DF)
For example, #3, the three conditions should be changed as follows.
inode(x) | CP | dnode(F) | inode(x) | inode(F)
a) x o o o o
b) x x x x o
c) x o o x o
If f2fs_sync_file stops ------^,
it should write inode(F) --------------^
So, the need_inode_block_update should return true, since
c) get_nat_flag(e, HAS_LAST_FSYNC), is false.
For example, #8,
CP | alloc | dnode(F) | inode(x) | inode(DF)
a) o x x x x
b) x x x o
c) o o x o
If f2fs_sync_file stops -------^,
it should write inode(DF) --------------^
Note that, the roll-forward policy should follow this rule, which means,
if there are any missing blocks, we doesn't need to recover that inode.
Signed-off-by: Huang Ying <ying.huang@intel.com>
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
This patch introduces a flag in the nat entry structure to merge various
information such as checkpointed and fsync_done marks.
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
The nm_i->fcnt checking is executed before spin_lock, so if another
thread delete the last free_nid from the list, the wrong nid may be
gotten. So fix the race condition by moving the nm_i->fnct checking
into spin_lock.
Signed-off-by: Huang, Ying <ying.huang@intel.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>
fix the following checkpatch warning:
WARNING: do {} while (0) macros should not be semicolon terminated
Signed-off-by: Zhang Zhen <zhenzhang.zhang@huawei.com>
Signed-off-by: Jaegeuk Kim <jaegeuk.kim@samsung.com>
This patch splits grab_cache_page_write_begin into grab_cache_page and
wait_on_page_writeback for node pages.
This patch intends to enhance the latency to get node pages by alleviating
unnecessary wait_on_page_writeback.
Signed-off-by: Chao Yu <chao2.yu@samsung.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 introduce raw_nat_from_node_info() to simplfy some codes, and also
use exist function node_info_from_raw_nat() to do the same job.
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>
The NM_WOUT_THRESHOLD is now obsolete since f2fs starts to control on a basis
of the memory footprint.
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>
This patch introduces a help function f2fs_has_xattr_block for better
readability.
Signed-off-by: Chao Yu <chao2.yu@samsung.com>
Signed-off-by: Jaegeuk Kim <jaegeuk.kim@samsung.com>
The nat cache entry maintains a status whether it is checkpointed or not.
So, if a new cache entry is loaded from the last checkpoint,
nat_entry->checkpointed should be true.
If the cache entry is modified as being dirty, nat_entry->checkpoint should
be false.
Signed-off-by: Jaegeuk Kim <jaegeuk.kim@samsung.com>
Update several comments:
1. use f2fs_{un}lock_op install of mutex_{un}lock_op.
2. update comment of get_data_block().
3. update description of node offset.
Signed-off-by: Chao Yu <chao2.yu@samsung.com>
Signed-off-by: Jaegeuk Kim <jaegeuk.kim@samsung.com>
This patch fixes the use of XATTR_NODE_OFFSET.
o The offset should not use several MSB bits which are used by marking node
blocks.
o IS_DNODE should handle XATTR_NODE_OFFSET to avoid potential abnormality
during the fsync call.
Signed-off-by: Jaegeuk Kim <jaegeuk.kim@samsung.com>
Introduce help function F2FS_NODE() to simplify the conversion of node_page to
f2fs_node.
Signed-off-by: Gu Zheng <guz.fnst@cn.fujitsu.com>
Signed-off-by: Jaegeuk Kim <jaegeuk.kim@samsung.com>
If a file is linked, f2fs loose its parent inode number so that fsync calls
for the linked file should do checkpoint all the time.
But, if we can recover its parent inode number after the checkpoint, we can
adjust roll-forward mechanism for the further fsync calls, which is able to
improve the fsync performance significatly.
Signed-off-by: Jaegeuk Kim <jaegeuk.kim@samsung.com>
There are various functions with common code which could be separated
out to make common routines. So, made new routines and in order to
retain the same call path and no major changes, written some macros
to access those routines.
Signed-off-by: Namjae Jeon <namjae.jeon@samsung.com>
Signed-off-by: Amit Sahrawat <a.sahrawat@samsung.com>
Signed-off-by: Jaegeuk Kim <jaegeuk.kim@samsung.com>
If (ofs % (NIDS_PER_BLOCK + 1) == 0), the node is an indirect node block.
Signed-off-by: Zhihui Zhang <zzhsuny@gmail.com>
Signed-off-by: Jaegeuk Kim <jaegeuk.kim@samsung.com>
Hou Pengyang