13 Oct, 2015
9 commits
-
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
Signed-off-by: Jaegeuk Kim -
When there is no free nid in nid cache, all new node allocaters stop their
job to wait for reloading of free nids, however reloading is synchronous as
we will read 4 NAT pages for building nid cache, it cause the long latency.This patch tries to readahead more NAT pages with READA request flag after
reloading of free nids. It helps to improve performance when users allocate
node id intensively.Env: Sandisk 32G sd card
time for i in `seq 1 60000`; { echo -n > /mnt/f2fs/$i; echo XXXXXX > /mnt/f2fs/$i;}Before:
real 0m2.814s
user 0m1.220s
sys 0m1.536sAfter:
real 0m2.711s
user 0m1.136s
sys 0m1.568sSigned-off-by: Chao Yu
Signed-off-by: Jaegeuk Kim -
Now, we use ra_meta_pages to reads continuous physical blocks as much as
possible to improve performance of following reads. However, ra_meta_pages
uses a synchronous readahead approach by submitting bio with READ, as READ
is with high priority, it can not be used in the case of preloading blocks,
and it's not sure when these RAed pages will be used.This patch supports asynchronous readahead in ra_meta_pages by tagging bio
with READA flag in order to allow preloading.Signed-off-by: Chao Yu
Signed-off-by: Jaegeuk Kim -
In recovery or checkpoint flow, we grab pages temperarily in meta inode's
mapping for caching temperary data, actually, datas in these pages were
not meta data of f2fs, but still we tag them with REQ_META flag. However,
lower device like eMMC may do some optimization for data of such type.
So in order to avoid wrong optimization, we'd better remove such flag
for temperary non-meta pages.Signed-off-by: Chao Yu
Signed-off-by: Jaegeuk Kim -
This patch adds a tracepoint for f2fs_read_data_pages to trace when pages
are readahead by VFS.Signed-off-by: Chao Yu
Signed-off-by: Jaegeuk Kim -
For normal inodes, their pages are allocated with __GFP_FS, which can cause
filesystem calls when reclaiming memory.
This can incur a dead lock condition accordingly.So, this patch addresses this problem by introducing
f2fs_grab_cache_page(.., bool for_write), which calls
grab_cache_page_write_begin() with AOP_FLAG_NOFS.Signed-off-by: Jaegeuk Kim
-
The f2fs_collapse_range and f2fs_insert_range changes the block addresses
directly. But that can cause uncovered SSA updates.
In that case, we need to give up to change the block addresses and do buffered
writes to keep filesystem consistency.Signed-off-by: Jaegeuk Kim
-
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.
-
This patch introduces F2FS_GOING_DOWN_METAFLUSH which flushes meta pages like
SSA blocks and then blocks all the writes.
This can be used by power-failure tests.Reviewed-by: Chao Yu
Signed-off-by: Jaegeuk Kim
10 Oct, 2015
31 commits
-
This patch tries to merge IOs as many as possible when background flusher
conducts flushing the dirty meta pages.[Before]
...
2fs_submit_write_bio: dev = (8,18), WRITE_SYNC(MP), META, sector = 124320, size = 4096
f2fs_submit_write_bio: dev = (8,18), WRITE_SYNC(MP), META, sector = 124560, size = 32768
f2fs_submit_write_bio: dev = (8,18), WRITE_SYNC(MP), META, sector = 95720, size = 987136
f2fs_submit_write_bio: dev = (8,18), WRITE_SYNC(MP), META, sector = 123928, size = 4096
f2fs_submit_write_bio: dev = (8,18), WRITE_SYNC(MP), META, sector = 123944, size = 8192
f2fs_submit_write_bio: dev = (8,18), WRITE_SYNC(MP), META, sector = 123968, size = 45056
f2fs_submit_write_bio: dev = (8,18), WRITE_SYNC(MP), META, sector = 124064, size = 4096
f2fs_submit_write_bio: dev = (8,18), WRITE_SYNC(MP), META, sector = 97648, size = 1007616
f2fs_submit_write_bio: dev = (8,18), WRITE_SYNC(MP), META, sector = 123776, size = 8192
f2fs_submit_write_bio: dev = (8,18), WRITE_SYNC(MP), META, sector = 123800, size = 32768
f2fs_submit_write_bio: dev = (8,18), WRITE_SYNC(MP), META, sector = 124624, size = 4096
f2fs_submit_write_bio: dev = (8,18), WRITE_SYNC(MP), META, sector = 99616, size = 921600
f2fs_submit_write_bio: dev = (8,18), WRITE_SYNC(MP), META, sector = 123608, size = 4096
f2fs_submit_write_bio: dev = (8,18), WRITE_SYNC(MP), META, sector = 123624, size = 77824
f2fs_submit_write_bio: dev = (8,18), WRITE_SYNC(MP), META, sector = 123792, size = 4096
f2fs_submit_write_bio: dev = (8,18), WRITE_SYNC(MP), META, sector = 123864, size = 32768
...[After]
...
f2fs_submit_write_bio: dev = (8,18), WRITE_SYNC(MP), META, sector = 92168, size = 892928
f2fs_submit_write_bio: dev = (8,18), WRITE_SYNC(MP), META, sector = 93912, size = 753664
f2fs_submit_write_bio: dev = (8,18), WRITE_SYNC(MP), META, sector = 95384, size = 716800
f2fs_submit_write_bio: dev = (8,18), WRITE_SYNC(MP), META, sector = 96784, size = 712704
f2fs_submit_write_bio: dev = (8,18), WRITE_SYNC(MP), META, sector = 104160, size = 364544
f2fs_submit_write_bio: dev = (8,18), WRITE_SYNC(MP), META, sector = 104872, size = 356352
f2fs_submit_write_bio: dev = (8,18), WRITE_SYNC(MP), META, sector = 105568, size = 278528
f2fs_submit_write_bio: dev = (8,18), WRITE_SYNC(MP), META, sector = 106112, size = 319488
f2fs_submit_write_bio: dev = (8,18), WRITE_SYNC(MP), META, sector = 106736, size = 258048
f2fs_submit_write_bio: dev = (8,18), WRITE_SYNC(MP), META, sector = 107240, size = 270336
f2fs_submit_write_bio: dev = (8,18), WRITE_SYNC(MP), META, sector = 107768, size = 180224
...Signed-off-by: Jaegeuk Kim
-
This patch introduces a periodic checkpoint feature.
Note that, this is not enforcing to conduct checkpoints very strictly in terms
of trigger timing, instead just hope to help user experiences.
The default value is 60 seconds.Reviewed-by: Chao Yu
Signed-off-by: Jaegeuk Kim -
This patch introduces a tracepoint to monitor background gc behaviors.
Reviewed-by: Chao Yu
Signed-off-by: Jaegeuk Kim -
This patch introduce background_gc=sync enabling synchronous cleaning in
background.Reviewed-by: Chao Yu
Signed-off-by: Jaegeuk Kim -
This patch introduce a new ioctl for those users who want to trigger
checkpoint from userspace through ioctl.Signed-off-by: Chao Yu
Signed-off-by: Jaegeuk Kim -
This patch drops in batches gc triggered through ioctl, since user
can easily control the gc by designing the loop around the ->ioctl.We support synchronous gc by forcing using FG_GC in f2fs_gc, so with
it, user can make sure that in this round all blocks gced were
persistent in the device until ioctl returned.Signed-off-by: Chao Yu
Signed-off-by: Jaegeuk Kim -
When searching victim during gc, if there are no dirty segments in
filesystem, we will still take the time to search the whole dirty segment
map, it's not needed, it's better to skip in this condition.Signed-off-by: Chao Yu
Signed-off-by: Jaegeuk Kim -
When doing gc, we search a victim in dirty map, starting from position of
last victim, we will reset the current searching position until we touch
the end of dirty map, and then search the whole diryt map. So sometimes we
will search the range [victim, last] twice, it's redundant, this patch
avoids this issue.Signed-off-by: Chao Yu
Signed-off-by: Jaegeuk Kim -
Our hit stat of extent cache will increase all the time until remount,
and we use atomic_t type for the stat variable, so it may easily incur
overflow when we query extent cache frequently in a long time running
fs.So to avoid that, this patch uses atomic64_t for hit stat variables.
Signed-off-by: Chao Yu
Signed-off-by: Jaegeuk Kim -
This patch introduces f2fs_kvmalloc to avoid -ENOMEM during mount.
Signed-off-by: Jaegeuk Kim
-
If we do not call get_victim first, we cannot get a new victim for retrial
path.Reviewed-by: Chao Yu
Signed-off-by: Jaegeuk Kim -
This patch fixes to maintain the right section count freed in garbage
collecting when triggering a foreground gc.Besides, when a foreground gc is running on current selected section, once
we fail to gc one segment, it's better to abandon gcing the left segments
in current section, because anyway we will select next victim for
foreground gc, so gc on the left segments in previous section will become
overhead and also cause the long latency for caller.Signed-off-by: Chao Yu
Signed-off-by: Jaegeuk Kim -
This patch fixes to update ctime and atime correctly when truncating
larger in ->setattr.The bug is reported by xfstest generic/313 as below:
generic/313 2s ... - output mismatch (see ./results/generic/313.out.bad)
--- tests/generic/313.out 2015-08-04 15:28:53.430798882 +0800
+++ results/generic/313.out.bad 2015-09-28 17:04:27.294278016 +0800
@@ -1,2 +1,4 @@
QA output created by 313
Silence is golden
+ctime not updated after truncate up
+mtime not updated after truncate up
...
(Run 'diff -u tests/generic/313.out tests/generic/313.out.bad' to see the entire diff)
Ran: generic/313
Failures: generic/313
Failed 1 of 1 testsSigned-off-by: Chao Yu
Signed-off-by: Jaegeuk Kim -
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
-
Protecting recovery flow by using cp_rwsem is not needed, since we have
prevent triggering any checkpoint by locking cp_mutex previously.Signed-off-by: Chao Yu
Signed-off-by: Jaegeuk Kim -
In update_sit_info, we use div_u64 to handle 'u64 divide u64' case, but
div_u64 can only handle 32-bits divisor, so our divisor with u64 type
passed to div_u64 will overflow, result in the wrong calculation when
show debug info of f2fs as below:BDF: 464, avg. vblocks: 23509
(BDF should never exceed 100)So change to use div64_u64 to handle this case correctly.
Signed-off-by: Chao Yu
Signed-off-by: Jaegeuk Kim -
This patch adds a new helper __try_update_largest_extent for cleanup.
Signed-off-by: Chao Yu
Signed-off-by: Jaegeuk Kim -
This fixes error handling for calls to various functions in the
function recover_inline_data to check if these particular functions
either return a error code or the boolean value false to signal their
caller they have failed internally and if this arises return false
to signal failure immediately to the caller of recover_inline_data
as we cannot continue after failures to calling either the function
truncate_inline_inode or truncate_blocks.Signed-off-by: Nicholas Krause
Signed-off-by: Jaegeuk Kim -
Swith extent_cache option dynamically when remount may casue consistency
issue between extent cache and dnode page. Fix in this patch to avoid
that condition.Signed-off-by: Chao Yu
Signed-off-by: Jaegeuk Kim -
We introduce F2FS_GET_BLOCK_READ in commit e2b4e2bc8865 ("f2fs: fix
incorrect mapping for bmap"), but forget to use this flag in the right
place, fix it.Signed-off-by: Chao Yu
Signed-off-by: Jaegeuk Kim -
Here is a oops reported as following message when testing generic/019 of
xfstest:------------[ cut here ]------------
kernel BUG at /home/yuchao/git/f2fs-dev/segment.c:882!
invalid opcode: 0000 [#1] SMP
Modules linked in: zram lz4_compress lz4_decompress f2fs(O) ip6table_filter ip6_tables ebtable_nat ebtables nf_conntrack_ipv4
nf_def
CPU: 2 PID: 25441 Comm: fio Tainted: G O 4.3.0-rc1+ #6
Hardware name: Hewlett-Packard HP Z220 CMT Workstation/1790, BIOS K51 v01.61 05/16/2013
task: ffff8803f4e85580 ti: ffff8803fd61c000 task.ti: ffff8803fd61c000
RIP: 0010:[] [] new_curseg+0x321/0x330 [f2fs]
RSP: 0018:ffff8803fd61f918 EFLAGS: 00010246
RAX: 00000000000007ed RBX: 0000000000000224 RCX: 000000000000001f
RDX: 0000000000000800 RSI: ffffffffffffffff RDI: ffff8803f56f4300
RBP: ffff8803fd61f978 R08: 0000000000000000 R09: 0000000000000000
R10: 0000000000000024 R11: ffff8800d23bbd78 R12: ffff8800d0ef0000
R13: 0000000000000224 R14: 0000000000000000 R15: 0000000000000001
FS: 00007f827ff85700(0000) GS:ffff88041ea80000(0000) knlGS:0000000000000000
CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: ffffffffff600000 CR3: 00000003fef17000 CR4: 00000000001406e0
Stack:
000007ea00000002 0000000100000001 ffff8803f6456248 000007ed0000002b
0000000000000224 ffff880404d1aa20 ffff8803fd61f9c8 ffff8800d0ef0000
ffff8803f6456248 0000000000000001 00000000ffffffff ffffffffa078f358
Call Trace:
[] allocate_segment_by_default+0x1a7/0x1f0 [f2fs]
[] allocate_data_block+0x17c/0x360 [f2fs]
[] __allocate_data_block+0x131/0x1d0 [f2fs]
[] f2fs_direct_IO+0x4b5/0x580 [f2fs]
[] generic_file_direct_write+0xae/0x160
[] __generic_file_write_iter+0xd5/0x1f0
[] generic_file_write_iter+0xf7/0x200
[] ? apparmor_file_permission+0x18/0x20
[] ? f2fs_fallocate+0x1190/0x1190 [f2fs]
[] f2fs_file_write_iter+0x46/0x90 [f2fs]
[] aio_run_iocb+0x1ee/0x290
[] ? mutex_lock+0x1e/0x50
[] ? aio_read_events+0x207/0x2b0
[] do_io_submit+0x373/0x630
[] ? SyS_io_getevents+0x56/0xb0
[] SyS_io_submit+0x10/0x20
[] entry_SYSCALL_64_fastpath+0x12/0x6a
Code: 45 c8 48 8b 78 10 e8 9f 23 bf e0 41 8b 8c 24 cc 03 00 00 89 c7 31 d2 89 c6 89 d8 29 df f7 f1 29 d1 39 cf 0f 83 be fd ff ff eb
RIP [] new_curseg+0x321/0x330 [f2fs]
RSP
---[ end trace 2e577d7f711ddb86 ]---The reason is that: in the test of generic/019, we will trigger a manmade
IO error in block layer through debugfs, after that, prefree segment will
no longer be freed, because we always skip doing gc or checkpoint when
there occurs an IO error.Meanwhile fio with aio engine generated a large number of direct IOs,
which continue allocating spaces in free segment until we run out of them,
eventually, results in panic in new_curseg as no more free segment was
found.So, this patch changes to return EIO in direct_IO for this condition.
Signed-off-by: Chao Yu
Signed-off-by: Jaegeuk Kim -
truncate_data_blocks_range can do in batches truncation which makes all
changes in dnode page content, dnode page status, extent cache, block
count updating together.But previously, truncate_hole() always truncates one block in dnode page
at a time by invoking truncate_data_blocks_range(,1), which make thing
slow.This patch changes truncate_hole() to do in batches truncation for all
target blocks in one direct node inside truncate_data_blocks_range, which
can make our punch hole operation in ->fallocate more efficent.Signed-off-by: Chao Yu
Signed-off-by: Jaegeuk Kim -
Fix 2 potential problems:
1. when largest extent needs to be invalidated, it will be reset in
__drop_largest_extent, which makes __is_extent_same after always
return false, and largest extent unchanged. Now we update it properly.2. when extent is split and the latter part remains in tree, next_en
should be the latter part instead of next extent of original extent.
It will cause merge failure if there is in-place update, although
there is not, I think this fix will still makes codes less ambiguous.This patch also simplifies codes of invalidating extents, and optimizes the
procedues that split extent into two.
There are a few modifications after last patch:
1. prev_en now is updated properly.
2. more codes and branches are simplified.Signed-off-by: Fan li
Signed-off-by: Jaegeuk Kim -
now we update extent by range, fofs may not be on the largest
extent if the new extent overlaps with it. so add a new function
to drop largest extent properly.Signed-off-by: Fan li
Signed-off-by: Jaegeuk Kim -
This patch avoids to produce new checkpoint blocks before the previous meta
pages were written completely.Signed-off-by: Jaegeuk Kim
-
We got dentry pages from high_mem, and its address space directly goes into the
decryption path via f2fs_fname_disk_to_usr.
But, sg_init_one assumes the address is not from high_mem, so we can get this
panic since it doesn't call kmap_high but kunmap_high is triggered at the end.kernel BUG at ../../../../../../kernel/mm/highmem.c:290!
Internal error: Oops - BUG: 0 [#1] PREEMPT SMP ARM
...
(kunmap_high+0xb0/0xb8) from [] (__kunmap_atomic+0xa0/0xa4)
(__kunmap_atomic+0xa0/0xa4) from [] (blkcipher_walk_done+0x128/0x1ec)
(blkcipher_walk_done+0x128/0x1ec) from [] (crypto_cbc_decrypt+0xc0/0x170)
(crypto_cbc_decrypt+0xc0/0x170) from [] (crypto_cts_decrypt+0xc0/0x114)
(crypto_cts_decrypt+0xc0/0x114) from [] (async_decrypt+0x40/0x48)
(async_decrypt+0x40/0x48) from [] (f2fs_fname_disk_to_usr+0x124/0x304)
(f2fs_fname_disk_to_usr+0x124/0x304) from [] (f2fs_fill_dentries+0xac/0x188)
(f2fs_fill_dentries+0xac/0x188) from [] (f2fs_readdir+0x1f0/0x300)
(f2fs_readdir+0x1f0/0x300) from [] (vfs_readdir+0x90/0xb4)
(vfs_readdir+0x90/0xb4) from [] (SyS_getdents64+0x64/0xcc)
(SyS_getdents64+0x64/0xcc) from [] (ret_fast_syscall+0x0/0x30)Cc:
Reviewed-by: Chao Yu
Signed-off-by: Jaegeuk Kim -
In this patch, we try to reorganize f2fs_map_blocks to make block mapping
flow more clear by using following structure:/* check status of mapping */
if (unmapped) {
/* blkaddr == NULL_ADDR || blkaddr == NEW_ADDR */if (create) {
/* write path, handle dio write case here */
alloc_and_map;
} else {
/*
* handle read cases from all call paths:
* 1. generic read;
* 2. dio read;
* 3. fiemap;
* 4. bmap
*/
}
}/* map buffer_header */
Besides, this patch handles the missing case correctly for dio write:
When we fail in __allocate_data_blocks, then in f2fs_map_blocks, we will
not allocate blocks correctly for preallocated blocks, but returning with
an unmapped buffer head, which will result in failure of dio write.Signed-off-by: Chao Yu
Signed-off-by: Jaegeuk Kim -
This function should be static.
Signed-off-by: Jaegeuk Kim
-
We have potential overflow issue when calculating size of object, when
we left shift index with PAGE_CACHE_SHIFT bits, if type of index has only
32-bits space in 32-bit architecture, left shifting will incur overflow,
i.e:pgoff_t index = 0xFFFFFFFF;
loff_t size = index << PAGE_CACHE_SHIFT;
size: 0xFFFFF000So we should cast index with 64-bits type to avoid this issue.
Signed-off-by: Chao Yu
Signed-off-by: Jaegeuk Kim -
When shrinking extent cache, we have two steps in the flow:
1) shrink objects which are unreferenced by inodes;
2) shrink objects from LRU list of extent cache.In step 1, if we haven't shrunk enough number of objects, we will try
step 2, but before that we didn't update the searching position which
may point to last inode index in global extent tree, result in failing
to shrink objects by traversing the all inodes' extent tree.In this patch, we reset searching position to beginning of global extent
tree for fixing.Signed-off-by: Chao Yu
Signed-off-by: Jaegeuk Kim -
This patch changes to verify file type early in f2fs_fallocate for
cleanup, meanwhile this also fixes to add missing verification for
expand_inode_data.Signed-off-by: Chao Yu
Signed-off-by: Jaegeuk Kim
