JBD2 layer support triggers which are called when journaling layer moves
buffer to a certain state. We can use the frozen trigger, which gets
called when buffer data is frozen and about to be written out to the
journal, to compute block checksums for some buffer types (similarly as
does ocfs2). This avoids unnecessary repeated recomputation of the
checksum (at the cost of larger window where memory corruption won't be
caught by checksumming) and is even necessary when there are
unsynchronized updaters of the checksummed data.
So add superblock and journal trigger type arguments to
ext4_journal_get_write_access() and ext4_journal_get_create_access() so
that frozen triggers can be set accordingly. Also add inode argument to
ext4_walk_page_buffers() and all the callbacks used with that function
for the same purpose. This patch is mostly only a change of prototype of
the above mentioned functions and a few small helpers. Real checksumming
will come later.
Reviewed-by: Theodore Ts'o <tytso@mit.edu>
Signed-off-by: Jan Kara <jack@suse.cz>
Link: https://lore.kernel.org/r/20210816095713.16537-1-jack@suse.cz
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Add sparse annotations to suppress false positive context imbalance
warnings, and use NULL instead of 0 in EXT_MAX_{EXTENT,INDEX}.
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
The background discard kwork tries to mark blocks used and issue
discard. This can make filesystem suffer from NOSPC error, xfstest
generic/371 can fail due to it. Fix it by flushing discard kwork
in ext4_should_retry_alloc. At the same time, give up discard at
the moment.
Signed-off-by: Wang Jianchao <wangjianchao@kuaishou.com>
Link: https://lore.kernel.org/r/20210830075246.12516-6-jianchao.wan9@gmail.com
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Right now, discard is issued and waited to be completed in jbd2
commit kthread context after the logs are committed. When large
amount of files are deleted and discard is flooding, jbd2 commit
kthread can be blocked for long time. Then all of the metadata
operations can be blocked to wait the log space.
One case is the page fault path with read mm->mmap_sem held, which
wants to update the file time but has to wait for the log space.
When other threads in the task wants to do mmap, then write mmap_sem
is blocked. Finally all of the following read mmap_sem requirements
are blocked, even the ps command which need to read the /proc/pid/
-cmdline. Our monitor service which needs to read /proc/pid/cmdline
used to be blocked for 5 mins.
This patch frees the blocks back to buddy after commit and then do
discard in a async kworker context in fstrim fashion, namely,
- mark blocks to be discarded as used if they have not been allocated
- do discard
- mark them free
After this, jbd2 commit kthread won't be blocked any more by discard
and we won't get NOSPC even if the discard is slow or throttled.
Link: https://marc.info/?l=linux-kernel&m=162143690731901&w=2
Suggested-by: Theodore Ts'o <tytso@mit.edu>
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Wang Jianchao <wangjianchao@kuaishou.com>
Link: https://lore.kernel.org/r/20210830075246.12516-5-jianchao.wan9@gmail.com
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
There is no functional change in this patch but just split the
codes, which serachs free block and does trim, into a new function
ext4_try_to_trim_range. This is preparing for the following async
backgroup discard.
Reviewed-by: Andreas Dilger <adilger@dilger.ca>
Signed-off-by: Wang Jianchao <wangjianchao@kuaishou.com>
Reviewed-by: Jan Kara <jack@suse.cz>
Link: https://lore.kernel.org/r/20210724074124.25731-3-jianchao.wan9@gmail.com
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Get rid of the 'group' parameter of ext4_trim_extent as we can get
it from the 'e4b'.
Reviewed-by: Andreas Dilger <adilger@dilger.ca>
Signed-off-by: Wang Jianchao <wangjianchao@kuaishou.com>
Reviewed-by: Jan Kara <jack@suse.cz>
Link: https://lore.kernel.org/r/20210724074124.25731-2-jianchao.wan9@gmail.com
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Commit 5d1b1b3f49 ("ext4: fix BUG when calling ext4_error with locked
block group") introduces ext4_grp_locked_error to handle unlocking a
group in error cases. Otherwise, there is a possibility of a sleep while
atomic. However, since 43c73221b3 ("ext4: replace BUG_ON with WARN_ON
in mb_find_extent()"), mb_find_extent() has contained a ext4_error()
call while a group spinlock is held. Replace this with
ext4_grp_locked_error.
Fixes: 43c73221b3 ("ext4: replace BUG_ON with WARN_ON in mb_find_extent()")
Cc: <stable@vger.kernel.org> # 4.14+
Signed-off-by: Stephen Brennan <stephen.s.brennan@oracle.com>
Reviewed-by: Lukas Czerner <lczerner@redhat.com>
Reviewed-by: Junxiao Bi <junxiao.bi@oracle.com>
Link: https://lore.kernel.org/r/20210623232114.34457-1-stephen.s.brennan@oracle.com
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Made suggested modifications from checkpatch in reference to ERROR:
trailing whitespace
Signed-off-by: Jack Qiu <jack.qiu@huawei.com>
Reviewed-by: Ritesh Harjani <riteshh@linux.ibm.com>
Link: https://lore.kernel.org/r/20210409042035.15516-1-jack.qiu@huawei.com
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Instead of traversing through groups linearly, scan groups in specific
orders at cr 0 and cr 1. At cr 0, we want to find groups that have the
largest free order >= the order of the request. So, with this patch,
we maintain lists for each possible order and insert each group into a
list based on the largest free order in its buddy bitmap. During cr 0
allocation, we traverse these lists in the increasing order of largest
free orders. This allows us to find a group with the best available cr
0 match in constant time. If nothing can be found, we fallback to cr 1
immediately.
At CR1, the story is slightly different. We want to traverse in the
order of increasing average fragment size. For CR1, we maintain a rb
tree of groupinfos which is sorted by average fragment size. Instead
of traversing linearly, at CR1, we traverse in the order of increasing
average fragment size, starting at the most optimal group. This brings
down cr 1 search complexity to log(num groups).
For cr >= 2, we just perform the linear search as before. Also, in
case of lock contention, we intermittently fallback to linear search
even in CR 0 and CR 1 cases. This allows us to proceed during the
allocation path even in case of high contention.
There is an opportunity to do optimization at CR2 too. That's because
at CR2 we only consider groups where bb_free counter (number of free
blocks) is greater than the request extent size. That's left as future
work.
All the changes introduced in this patch are protected under a new
mount option "mb_optimize_scan".
With this patchset, following experiment was performed:
Created a highly fragmented disk of size 65TB. The disk had no
contiguous 2M regions. Following command was run consecutively for 3
times:
time dd if=/dev/urandom of=file bs=2M count=10
Here are the results with and without cr 0/1 optimizations introduced
in this patch:
|---------+------------------------------+---------------------------|
| | Without CR 0/1 Optimizations | With CR 0/1 Optimizations |
|---------+------------------------------+---------------------------|
| 1st run | 5m1.871s | 2m47.642s |
| 2nd run | 2m28.390s | 0m0.611s |
| 3rd run | 2m26.530s | 0m1.255s |
|---------+------------------------------+---------------------------|
Signed-off-by: Harshad Shirwadkar <harshadshirwadkar@gmail.com>
Reported-by: kernel test robot <lkp@intel.com>
Reported-by: Dan Carpenter <dan.carpenter@oracle.com>
Reviewed-by: Andreas Dilger <adilger@dilger.ca>
Link: https://lore.kernel.org/r/20210401172129.189766-6-harshadshirwadkar@gmail.com
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
A few arrays in mballoc.c use the total number of valid orders as
their size. Currently, this value is set as "sb->s_blocksize_bits +
2". This makes code harder to read. So, instead add a new macro
MB_NUM_ORDERS(sb) to make the code more readable.
Signed-off-by: Harshad Shirwadkar <harshadshirwadkar@gmail.com>
Reviewed-by: Andreas Dilger <adilger@dilger.ca>
Reviewed-by: Ritesh Harjani <ritesh.list@gmail.com>
Link: https://lore.kernel.org/r/20210401172129.189766-5-harshadshirwadkar@gmail.com
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
s_mb_buddies_generated gets used later in this patch series to
determine if the cr 0 and cr 1 optimziations should be performed or
not. Currently, s_mb_buddies_generated is protected under a
spin_lock. In the allocation path, it is better if we don't depend on
the lock and instead read the value atomically. In order to do that,
we drop s_bal_lock altogether and we convert the only two protected
fields by it s_mb_buddies_generated and s_mb_generation_time to atomic
type.
Signed-off-by: Harshad Shirwadkar <harshadshirwadkar@gmail.com>
Reviewed-by: Andreas Dilger <adilger@dilger.ca>
Reviewed-by: Ritesh Harjani <ritesh.list@gmail.com>
Link: https://lore.kernel.org/r/20210401172129.189766-2-harshadshirwadkar@gmail.com
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
syzbot found UBSAN: shift-out-of-bounds in ext4_mb_init [1], when
1 << sbi->s_es->s_log_groups_per_flex is bigger than UINT_MAX,
where sbi->s_mb_prefetch is unsigned integer type.
32 is the maximum allowed power of s_log_groups_per_flex. Following if
check will also trigger UBSAN shift-out-of-bound:
if (1 << sbi->s_es->s_log_groups_per_flex >= UINT_MAX) {
So I'm checking it against the raw number, perhaps there is another way
to calculate UINT_MAX max power. Also use min_t as to make sure it's
uint type.
[1] UBSAN: shift-out-of-bounds in fs/ext4/mballoc.c:2713:24
shift exponent 60 is too large for 32-bit type 'int'
Call Trace:
__dump_stack lib/dump_stack.c:79 [inline]
dump_stack+0x137/0x1be lib/dump_stack.c:120
ubsan_epilogue lib/ubsan.c:148 [inline]
__ubsan_handle_shift_out_of_bounds+0x432/0x4d0 lib/ubsan.c:395
ext4_mb_init_backend fs/ext4/mballoc.c:2713 [inline]
ext4_mb_init+0x19bc/0x19f0 fs/ext4/mballoc.c:2898
ext4_fill_super+0xc2ec/0xfbe0 fs/ext4/super.c:4983
Reported-by: syzbot+a8b4b0c60155e87e9484@syzkaller.appspotmail.com
Signed-off-by: Sabyrzhan Tasbolatov <snovitoll@gmail.com>
Reviewed-by: Jan Kara <jack@suse.cz>
Link: https://lore.kernel.org/r/20210224095800.3350002-1-snovitoll@gmail.com
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Commit cfd7323772 ("ext4: add prefetching for block allocation
bitmaps") introduced block bitmap prefetch, and expects to read block
bitmaps of flex_bg through an IO. However, it seems to ignore the
value range of s_log_groups_per_flex. In the scenario where the value
of s_log_groups_per_flex is greater than 27, s_mb_prefetch or
s_mb_prefetch_limit will overflow, cause a divide zero exception.
In addition, the logic of calculating nr is also flawed, because the
size of flexbg is fixed during a single mount, but s_mb_prefetch can
be modified, which causes nr to fail to meet the value condition of
[1, flexbg_size].
To solve this problem, we need to set the upper limit of
s_mb_prefetch. Since we expect to load block bitmaps of a flex_bg
through an IO, we can consider determining a reasonable upper limit
among the IO limit parameters. After consideration, we chose
BLK_MAX_SEGMENT_SIZE. This is a good choice to solve divide zero
problem and avoiding performance degradation.
[ Some minor code simplifications to make the changes easy to follow -- TYT ]
Reported-by: Tosk Robot <tencent_os_robot@tencent.com>
Signed-off-by: Chunguang Xu <brookxu@tencent.com>
Reviewed-by: Samuel Liao <samuelliao@tencent.com>
Reviewed-by: Andreas Dilger <adilger@dilger.ca>
Link: https://lore.kernel.org/r/1607051143-24508-1-git-send-email-brookxu@tencent.com
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
When freeing metadata, we will create an ext4_free_data and
insert it into the pending free list. After the current
transaction is committed, the object will be freed.
ext4_mb_free_metadata() will check whether the area to be freed
overlaps with the pending free list. If true, return directly. At this
time, ext4_free_data is leaked. Fortunately, the probability of this
problem is small, since it only occurs if the file system is corrupted
such that a block is claimed by more one inode and those inodes are
deleted within a single jbd2 transaction.
Signed-off-by: Chunguang Xu <brookxu@tencent.com>
Link: https://lore.kernel.org/r/1604764698-4269-8-git-send-email-brookxu@tencent.com
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Cc: stable@kernel.org
The code of mb_find_order_for_block is a bit obscure, but we can
simplify it with mb_find_buddy(), make the code more concise.
Signed-off-by: Chunguang Xu <brookxu@tencent.com>
Reviewed-by: Andreas Dilger <adilger@dilger.ca>
Link: https://lore.kernel.org/r/1604764698-4269-3-git-send-email-brookxu@tencent.com
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
After this patch (163a203), if an abnormal bitmap is detected, we
will mark the group as corrupt, and we will not use this group in
the future. Therefore, it should be meaningless to regenerate the
buddy bitmap of this group, It might be better to delete it.
Signed-off-by: Chunguang Xu <brookxu@tencent.com>
Reviewed-by: Andreas Dilger <adilger@dilger.ca>
Link: https://lore.kernel.org/r/1604764698-4269-2-git-send-email-brookxu@tencent.com
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Fast commit file system states are recorded in
sbi->s_mount_flags. Fast commit expects these bit manipulations to be
atomic. This patch adds helpers to make those modifications atomic.
Suggested-by: Jan Kara <jack@suse.cz>
Signed-off-by: Harshad Shirwadkar <harshadshirwadkar@gmail.com>
Link: https://lore.kernel.org/r/20201106035911.1942128-21-harshadshirwadkar@gmail.com
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Smatch complains that "i" can be uninitialized if we don't enter the
loop. I don't know if it's possible but we may as well silence this
warning.
[ Initialize i to sb->s_blocksize instead of 0. The only way the for
loop could be skipped entirely is the in-memory data structures, in
particular the bh->b_data for the on-disk superblock has gotten
corrupted enough that calculated value of group is >= to
ext4_get_groups_count(sb). In that case, we want to exit
immediately without allocating a block. -- TYT ]
Fixes: 8016e29f43 ("ext4: fast commit recovery path")
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Link: https://lore.kernel.org/r/20201030114620.GB3251003@mwanda
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Cc: stable@kernel.org
This patch adds fast commit recovery path support for Ext4 file
system. We add several helper functions that are similar in spirit to
e2fsprogs journal recovery path handlers. Example of such functions
include - a simple block allocator, idempotent block bitmap update
function etc. Using these routines and the fast commit log in the fast
commit area, the recovery path (ext4_fc_replay()) performs fast commit
log recovery.
Reported-by: kernel test robot <lkp@intel.com>
Signed-off-by: Harshad Shirwadkar <harshadshirwadkar@gmail.com>
Link: https://lore.kernel.org/r/20201015203802.3597742-8-harshadshirwadkar@gmail.com
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Delete repeated words in fs/ext4/.
{the, this, of, we, after}
Also change spelling of "xttr" in inline.c to "xattr" in 2 places.
Signed-off-by: Randy Dunlap <rdunlap@infradead.org>
Reviewed-by: Jan Kara <jack@suse.cz>
Link: https://lore.kernel.org/r/20200805024850.12129-1-rdunlap@infradead.org
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
ext4_mb_discard_group_preallocations() can be releasing group lock with
preallocations accumulated on its local list. Thus although
discard_pa_seq was incremented and concurrent allocating processes will
be retrying allocations, it can happen that premature ENOSPC error is
returned because blocks used for preallocations are not available for
reuse yet. Make sure we always free locally accumulated preallocations
before releasing group lock.
Fixes: 07b5b8e1ac ("ext4: mballoc: introduce pcpu seqcnt for freeing PA to improve ENOSPC handling")
Signed-off-by: Jan Kara <jack@suse.cz>
Link: https://lore.kernel.org/r/20200924150959.4335-1-jack@suse.cz
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
As we test disk offline/online with running fsstress, we find fsstress
process is keeping running state.
kworker/u32:3-262 [004] ...1 140.787471: ext4_mb_discard_preallocations: dev 8,32 needed 114
....
kworker/u32:3-262 [004] ...1 140.787471: ext4_mb_discard_preallocations: dev 8,32 needed 114
ext4_mb_new_blocks
repeat:
ext4_mb_discard_preallocations_should_retry(sb, ac, &seq)
freed = ext4_mb_discard_preallocations
ext4_mb_discard_group_preallocations
this_cpu_inc(discard_pa_seq);
---> freed == 0
seq_retry = ext4_get_discard_pa_seq_sum
for_each_possible_cpu(__cpu)
__seq += per_cpu(discard_pa_seq, __cpu);
if (seq_retry != *seq) {
*seq = seq_retry;
ret = true;
}
As we see seq_retry is sum of discard_pa_seq every cpu, if
ext4_mb_discard_group_preallocations return zero discard_pa_seq in this
cpu maybe increase one, so condition "seq_retry != *seq" have always
been met.
Ritesh Harjani suggest to in ext4_mb_discard_group_preallocations function we
only increase discard_pa_seq when there is some PA to free.
Fixes: 07b5b8e1ac ("ext4: mballoc: introduce pcpu seqcnt for freeing PA to improve ENOSPC handling")
Signed-off-by: Ye Bin <yebin10@huawei.com>
Reviewed-by: Jan Kara <jack@suse.cz>
Reviewed-by: Ritesh Harjani <riteshh@linux.ibm.com>
Link: https://lore.kernel.org/r/20200916113859.1556397-3-yebin10@huawei.com
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
In the scenario of writing sparse files, the per-inode prealloc list may
be very long, resulting in high overhead for ext4_mb_use_preallocated().
To circumvent this problem, we limit the maximum length of per-inode
prealloc list to 512 and allow users to modify it.
After patching, we observed that the sys ratio of cpu has dropped, and
the system throughput has increased significantly. We created a process
to write the sparse file, and the running time of the process on the
fixed kernel was significantly reduced, as follows:
Running time on unfixed kernel：
[root@TENCENT64 ~]# time taskset 0x01 ./sparse /data1/sparce.dat
real 0m2.051s
user 0m0.008s
sys 0m2.026s
Running time on fixed kernel：
[root@TENCENT64 ~]# time taskset 0x01 ./sparse /data1/sparce.dat
real 0m0.471s
user 0m0.004s
sys 0m0.395s
Signed-off-by: Chunguang Xu <brookxu@tencent.com>
Link: https://lore.kernel.org/r/d7a98178-056b-6db5-6bce-4ead23f4a257@gmail.com
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Reorganize the if statement of ext4_mb_release_context(), make it
easier to read.
Signed-off-by: Chunguang Xu <brookxu@tencent.com>
Link: https://lore.kernel.org/r/5439ac6f-db79-ad68-76c1-a4dda9aa0cc3@gmail.com
Reviewed-by: Ritesh Harjani <riteshh@linux.ibm.com>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Lost chunks are when some other process raced with the current thread
to grab a particular block allocation. Add mb_debug log for
developers who wants to see how often this is happening for a
particular workload.
Signed-off-by: Chunguang Xu <brookxu@tencent.com>
Link: https://lore.kernel.org/r/0a165ac0-1912-aebd-8a0d-b42e7cd1aea1@gmail.com
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
seq_puts is a lot cheaper than seq_printf, so use that to print
literal strings.
Signed-off-by: Xu Wang <vulab@iscas.ac.cn>
Reviewed-by: Ritesh Harjani <riteshh@linux.ibm.com>
Link: https://lore.kernel.org/r/20200810022158.9167-1-vulab@iscas.ac.cn
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
It might be better to adjust the code in two places:
1. Determine whether grp is currupt or not should be placed first.
2. (cr<=2 && free <ac->ac_g_ex.fe_len)should may belong to the crx
strategy, and it may be more appropriate to put it in the
subsequent switch statement block. For cr1, cr2, the conditions
in switch potentially realize the above judgment. For cr0, we
should add (free <ac->ac_g_ex.fe_len) judgment, and then delete
(free / fragments) >= ac->ac_g_ex.fe_len), because cr0 returns
true by default.
Signed-off-by: Chunguang Xu <brookxu@tencent.com>
Reviewed-by: Andreas Dilger <adilger@dilger.ca>
Reviewed-by: Ritesh Harjani <riteshh@linux.ibm.com>
Link: https://lore.kernel.org/r/e20b2d8f-1154-adb7-3831-a9e11ba842e9@gmail.com
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
These comments do not seem to be related to ext4_mb_check_limits(),
it may be invalid.
Signed-off-by: Chunguang Xu <brookxu@tencent.com>
Reviewed-by: Ritesh Harjani <riteshh@linux.ibm.com>
Link: https://lore.kernel.org/r/c49faf0c-d5d5-9c51-6911-9e0ff57c6bfa@gmail.com
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Currently, system zones just track ranges of block, that are "important"
fs metadata (bitmaps, group descriptors, journal blocks, etc.). This
however complicates how extent tree (or indirect blocks) can be checked
for inodes that actually track such metadata - currently the journal
inode but arguably we should be treating quota files or resize inode
similarly. We cannot run __ext4_ext_check() on such metadata inodes when
loading their extents as that would immediately trigger the validity
checks and so we just hack around that and special-case the journal
inode. This however leads to a situation that a journal inode which has
extent tree of depth at least one can have invalid extent tree that gets
unnoticed until ext4_cache_extents() crashes.
To overcome this limitation, track inode number each system zone belongs
to (0 is used for zones not belonging to any inode). We can then verify
inode number matches the expected one when verifying extent tree and
thus avoid the false errors. With this there's no need to to
special-case journal inode during extent tree checking anymore so remove
it.
Fixes: 0a944e8a6c ("ext4: don't perform block validity checks on the journal inode")
Reported-by: Wolfgang Frisch <wolfgang.frisch@suse.com>
Reviewed-by: Lukas Czerner <lczerner@redhat.com>
Signed-off-by: Jan Kara <jack@suse.cz>
Link: https://lore.kernel.org/r/20200728130437.7804-4-jack@suse.cz
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
For file systems where we can afford to keep the buddy bitmaps cached,
we can speed up initial writes to large file systems by starting to
load the block allocation bitmaps as soon as the file system is
mounted. This won't work well for _super_ large file systems, or
memory constrained systems, so we only enable this when it is
requested via a mount option.
Addresses-Google-Bug: 159488342
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Reviewed-by: Andreas Dilger <adilger@dilger.ca>
cr=0 is supposed to be an optimization to save CPU cycles, but if
buddy data (in memory) is not initialized then all this makes no sense
as we have to do sync IO taking a lot of cycles. Also, at cr=0
mballoc doesn't choose any available chunk. cr=1 also skips groups
using heuristic based on avg. fragment size. It's more useful to skip
such groups and switch to cr=2 where groups will be scanned for
available chunks. However, we always read the first block group in a
flex_bg so metadata blocks will get read into the first flex_bg if
possible.
Using sparse image and dm-slow virtual device of 120TB was
simulated, then the image was formatted and filled using debugfs to
mark ~85% of available space as busy. mount process w/o the patch
couldn't complete in half an hour (according to vmstat it would take
~10-11 hours). With the patch applied mount took ~20 seconds.
Lustre-bug-id: https://jira.whamcloud.com/browse/LU-12988
Signed-off-by: Alex Zhuravlev <azhuravlev@whamcloud.com>
Reviewed-by: Andreas Dilger <adilger@whamcloud.com>
Reviewed-by: Artem Blagodarenko <artem.blagodarenko@gmail.com>
This should significantly improve bitmap loading, especially for flex
groups as it tries to load all bitmaps within a flex.group instead of
one by one synchronously.
Prefetching is done in 8 * flex_bg groups, so it should be 8
read-ahead reads for a single allocating thread. At the end of
allocation the thread waits for read-ahead completion and initializes
buddy information so that read-aheads are not lost in case of memory
pressure.
At cr=0 the number of prefetching IOs is limited per allocation
context to prevent a situation when mballoc loads thousands of bitmaps
looking for a perfect group and ignoring groups with good chunks.
Together with the patch "ext4: limit scanning of uninitialized groups"
the mount time (which includes few tiny allocations) of a 1PB
filesystem is reduced significantly:
0% full 50%-full unpatched patched
mount time 33s 9279s 563s
[ Restructured by tytso; removed the state flags in the allocation
context, so it can be used to lazily prefetch the allocation bitmaps
immediately after the file system is mounted. Skip prefetching
block groups which are uninitialized. Finally pass in the
REQ_RAHEAD flag to the block layer while prefetching. ]
Signed-off-by: Alex Zhuravlev <bzzz@whamcloud.com>
Reviewed-by: Andreas Dilger <adilger@whamcloud.com>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Currently while doing block allocation grp->bb_free may be getting
modified if discard is happening in parallel.
For e.g. consider a case where there are lot of threads who have
preallocated lot of blocks and there is a thread which is trying
to discard all of this group's PA. Now it could happen that
we see all of those group's bb_free is zero and fail the allocation
while there is sufficient space if we free up all the PA.
So this patch adds another flag "EXT4_MB_STRICT_CHECK" which will be set
if we are unable to allocate any blocks in the first try (since we may
not have considered blocks about to be discarded from PA lists).
So during retry attempt to allocate blocks we will use ext4_lock_group()
for checking if the group is good or not.
Signed-off-by: Ritesh Harjani <riteshh@linux.ibm.com>
Link: https://lore.kernel.org/r/9cb740a117c958c36596f167b12af1beae9a68b7.1589955723.git.riteshh@linux.ibm.com
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
ext4_mb_good_group() definition was changed some time back
and now it even initializes the buddy cache (via ext4_mb_init_group()),
if in case the EXT4_MB_GRP_NEED_INIT() is true for a group.
Note that ext4_mb_init_group() could sleep and so should not be called
under a spinlock held.
This is fine as of now because ext4_mb_good_group() is called before
loading the buddy bitmap without ext4_lock_group() held
and again called after loading the bitmap, only this time with
ext4_lock_group() held.
But still this whole thing is confusing.
So this patch refactors out ext4_mb_good_group_nolock() which should be
called when without holding ext4_lock_group().
Also in further patches we hold the spinlock (ext4_lock_group()) while
doing any calculations which involves grp->bb_free or grp->bb_fragments.
Signed-off-by: Ritesh Harjani <riteshh@linux.ibm.com>
Link: https://lore.kernel.org/r/d9f7d031a5fbe1c943fae6bf1ff5cdf0604ae722.1589955723.git.riteshh@linux.ibm.com
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
There could be a race in function ext4_mb_discard_group_preallocations()
where the 1st thread may iterate through group's bb_prealloc_list and
remove all the PAs and add to function's local list head.
Now if the 2nd thread comes in to discard the group preallocations,
it will see that the group->bb_prealloc_list is empty and will return 0.
Consider for a case where we have less number of groups
(for e.g. just group 0),
this may even return an -ENOSPC error from ext4_mb_new_blocks()
(where we call for ext4_mb_discard_group_preallocations()).
But that is wrong, since 2nd thread should have waited for 1st thread
to release all the PAs and should have retried for allocation.
Since 1st thread was anyway going to discard the PAs.
The algorithm using this percpu seq counter goes below:
1. We sample the percpu discard_pa_seq counter before trying for block
allocation in ext4_mb_new_blocks().
2. We increment this percpu discard_pa_seq counter when we either allocate
or free these blocks i.e. while marking those blocks as used/free in
mb_mark_used()/mb_free_blocks().
3. We also increment this percpu seq counter when we successfully identify
that the bb_prealloc_list is not empty and hence proceed for discarding
of those PAs inside ext4_mb_discard_group_preallocations().
Now to make sure that the regular fast path of block allocation is not
affected, as a small optimization we only sample the percpu seq counter
on that cpu. Only when the block allocation fails and when freed blocks
found were 0, that is when we sample percpu seq counter for all cpus using
below function ext4_get_discard_pa_seq_sum(). This happens after making
sure that all the PAs on grp->bb_prealloc_list got freed or if it's empty.
It can be well argued that why don't just check for grp->bb_free to
see if there are any free blocks to be allocated. So here are the two
concerns which were discussed:-
1. If for some reason the blocks available in the group are not
appropriate for allocation logic (say for e.g.
EXT4_MB_HINT_GOAL_ONLY, although this is not yet implemented), then
the retry logic may result into infinte looping since grp->bb_free is
non-zero.
2. Also before preallocation was clubbed with block allocation with the
same ext4_lock_group() held, there were lot of races where grp->bb_free
could not be reliably relied upon.
Due to above, this patch considers discard_pa_seq logic to determine if
we should retry for block allocation. Say if there are are n threads
trying for block allocation and none of those could allocate or discard
any of the blocks, then all of those n threads will fail the block
allocation and return -ENOSPC error. (Since the seq counter for all of
those will match as no block allocation/discard was done during that
duration).
Signed-off-by: Ritesh Harjani <riteshh@linux.ibm.com>
Link: https://lore.kernel.org/r/7f254686903b87c419d798742fd9a1be34f0657b.1589955723.git.riteshh@linux.ibm.com
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Implement ext4_mb_discard_preallocations_should_retry()
which we will need in later patches to add more logic
like check for sequence number match to see if we should
retry for block allocation or not.
There should be no functionality change in this patch.
Signed-off-by: Ritesh Harjani <riteshh@linux.ibm.com>
Link: https://lore.kernel.org/r/1cfae0098d2aa9afbeb59331401258182868c8f2.1589955723.git.riteshh@linux.ibm.com
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
ext4_mb_discard_preallocations() only checks for grp->bb_prealloc_list
of every group to discard the group's PA to free up the space if
allocation request fails. Consider below race:-
Process A Process B
1. allocate blocks
1. Fails block allocation from
ext4_mb_regular_allocator()
ext4_lock_group()
allocated blocks
more than ac_o_ex.fe_len
ext4_unlock_group()
2. Scans the
grp->bb_prealloc_list (under
ext4_lock_group()) and
find nothing and thus return
-ENOSPC.
2. Add the additional blocks to PA list
ext4_lock_group()
add blocks to grp->bb_prealloc_list
ext4_unlock_group()
Above race could be avoided if we add those additional blocks to
grp->bb_prealloc_list at the same time with block allocation when
ext4_lock_group() was still held.
With this discard-PA will know if there are actually any blocks which
could be freed from the PA
Signed-off-by: Ritesh Harjani <riteshh@linux.ibm.com>
Link: https://lore.kernel.org/r/a2217dd782585b42328981832e6d396abaaccb80.1589955723.git.riteshh@linux.ibm.com
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
mb_debug() msg had only 1 control level for all type of msgs.
And if we enable mballoc_debug then all of those msgs would be enabled.
Instead of adding multiple debug levels for mb_debug() msgs, use
pr_debug() with which we could have finer control to print msgs at all
of different levels (i.e. at file, func, line no.).
Also add process name/pid, superblk id, and other info in mb_debug()
msg. This also kills the mballoc_debug module parameter, since it is
not needed any more.
Signed-off-by: Ritesh Harjani <riteshh@linux.ibm.com>
Link: https://lore.kernel.org/r/f0c660cbde9e2edbe95c67942ca9ad80dd2231eb.1589086800.git.riteshh@linux.ibm.com
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Make sure to check for e4b->bd_info->bb_bitmap == NULL, in
mb_cmp_bitmaps() and return if NULL, to avoid possible NULL ptr
dereference. Similar to how we do this in other ifdef DOUBLE_CHECK
functions.
Also remove the BUG_ON() logic if kmalloc() or ext4_read_block_bitmap()
fails. We should simply mark grp->bb_bitmap as NULL if above happens.
In fact ext4_read_block_bitmap() may even return an error in case of resize
ioctl. Hence remove this BUG_ON logic (fstests ext4/032 may trigger
this).
Link: https://lore.kernel.org/r/9a54f8a696ff17c057cd571be3d15ac3ec1407f1.1589086800.git.riteshh@linux.ibm.com
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
This patch implemets mb_group_bb_bitmap_alloc() and
mb_group_bb_bitmap_free() function to remove #ifdef DOUBLE_CHECK macro
and it's related code from inside
ext4_mb_add_groupinfo()/ext4_mb_release().
There should be no functionality change in this patch.
Signed-off-by: Ritesh Harjani <riteshh@linux.ibm.com>
Link: https://lore.kernel.org/r/8c2095d74b779f0254a19b24982490dc6f07c4f9.1589086800.git.riteshh@linux.ibm.com
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
