When logging a directory we start by flushing all its delayed items.
That results in adding dir index items to the subvolume btree, for new
dentries, and removing dir index items from the subvolume btree for any
dentries that were deleted.
This makes it straightforward to log a directory simply by iterating over
all the modified subvolume btree leaves, especially when we used to log
both dir index keys and dir item keys (before commit 339d035424
("btrfs: only copy dir index keys when logging a directory") and when we
used to copy old dir index entries for leaves modified in the current
transaction (before commit 732d591a5d ("btrfs: stop copying old dir
items when logging a directory")).
From an efficiency point of view this has a couple of drawbacks:
1) Adds extra latency, due to copying delayed items to the subvolume btree
and deleting dir index items from the btree.
Further if there are other tasks accessing the btree, which is common
(syscalls like creat, mkdir, rename, link, unlink, truncate, reflinks,
etc, finishing an ordered extent, etc), lock contention can cause
further delays, both to the task logging a directory and to the other
tasks accessing the btree;
2) More time spent overall flushing delayed items, if after logging the
directory further changes are done to the directory in the same
transaction.
For example, if we add 10 dentries to a directory, fsync it, add more
10 dentries, fsync it again, then add more 10 dentries and fsync it
again, then we end up inserting 3 batches of 10 items to the subvolume
btree. With the changes from this patch, we flush all the delayed items
to the btree only once - a single batch of 30 items, and outside the
logging code (transaction commit or when delayed items are flushed
asynchronously).
This change simply skips the flushing of delayed items every time we log a
directory. Instead we copy the delayed insertion items directly to the log
tree and delete delayed deletion items directly from the log tree.
Therefore avoiding changing first the subvolume btree and then scanning it
for new items to copy from it to the log tree and detecting deletions
by observing gaps in consecutive dir index keys in subvolume btree leaves.
Running the following tests on a non-debug kernel (Debian's default kernel
config), on a box with a NVMe device, a 12 cores Intel CPU and 64G of ram,
produced the results below.
The results compare a branch without this patch and all the other patches
it depends on versus the same branch with the patchset applied.
The patchset is comprised of the following patches:
btrfs: don't drop dir index range items when logging a directory
btrfs: remove the root argument from log_new_dir_dentries()
btrfs: update stale comment for log_new_dir_dentries()
btrfs: free list element sooner at log_new_dir_dentries()
btrfs: avoid memory allocation at log_new_dir_dentries() for common case
btrfs: remove root argument from btrfs_delayed_item_reserve_metadata()
btrfs: store index number instead of key in struct btrfs_delayed_item
btrfs: remove unused logic when looking up delayed items
btrfs: shrink the size of struct btrfs_delayed_item
btrfs: search for last logged dir index if it's not cached in the inode
btrfs: move need_log_inode() to above log_conflicting_inodes()
btrfs: move log_new_dir_dentries() above btrfs_log_inode()
btrfs: log conflicting inodes without holding log mutex of the initial inode
btrfs: skip logging parent dir when conflicting inode is not a dir
btrfs: use delayed items when logging a directory
Custom test script for testing time spent at btrfs_log_inode():
#!/bin/bash
DEV=/dev/nvme0n1
MNT=/mnt/nvme0n1
# Total number of files to create in the test directory.
NUM_FILES=10000
# Fsync after creating or renaming N files.
FSYNC_AFTER=100
umount $DEV &> /dev/null
mkfs.btrfs -f $DEV
mount -o ssd $DEV $MNT
TEST_DIR=$MNT/testdir
mkdir $TEST_DIR
echo "Creating files..."
for ((i = 1; i <= $NUM_FILES; i++)); do
echo -n > $TEST_DIR/file_$i
if (( ($i % $FSYNC_AFTER) == 0 )); then
xfs_io -c "fsync" $TEST_DIR
fi
done
sync
echo "Renaming files..."
for ((i = 1; i <= $NUM_FILES; i++)); do
mv $TEST_DIR/file_$i $TEST_DIR/file_$i.renamed
if (( ($i % $FSYNC_AFTER) == 0 )); then
xfs_io -c "fsync" $TEST_DIR
fi
done
umount $MNT
And using the following bpftrace script to capture the total time that is
spent at btrfs_log_inode():
#!/usr/bin/bpftrace
k:btrfs_log_inode
{
@start_log_inode[tid] = nsecs;
}
kr:btrfs_log_inode
/@start_log_inode[tid]/
{
$dur = (nsecs - @start_log_inode[tid]) / 1000;
@btrfs_log_inode_total_time = sum($dur);
delete(@start_log_inode[tid]);
}
END
{
clear(@start_log_inode);
}
Result before applying patchset:
@btrfs_log_inode_total_time: 622642
Result after applying patchset:
@btrfs_log_inode_total_time: 354134 (-43.1% time spent)
The following dbench script was also used for testing:
#!/bin/bash
NUM_JOBS=$(nproc --all)
DEV=/dev/nvme0n1
MNT=/mnt/nvme0n1
MOUNT_OPTIONS="-o ssd"
MKFS_OPTIONS="-O no-holes -R free-space-tree"
echo "performance" | \
tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
umount $DEV &> /dev/null
mkfs.btrfs -f $MKFS_OPTIONS $DEV
mount $MOUNT_OPTIONS $DEV $MNT
dbench -D $MNT --skip-cleanup -t 120 -S $NUM_JOBS
umount $MNT
Before patchset:
Operation Count AvgLat MaxLat
----------------------------------------
NTCreateX 3322265 0.034 21.032
Close 2440562 0.002 0.994
Rename 140664 1.150 269.633
Unlink 670796 1.093 269.678
Deltree 96 5.481 15.510
Mkdir 48 0.004 0.052
Qpathinfo 3010924 0.014 8.127
Qfileinfo 528055 0.001 0.518
Qfsinfo 552113 0.003 0.372
Sfileinfo 270575 0.005 0.688
Find 1164176 0.052 13.931
WriteX 1658537 0.019 5.918
ReadX 5207412 0.003 1.034
LockX 10818 0.003 0.079
UnlockX 10818 0.002 0.313
Flush 232811 1.027 269.735
Throughput 869.867 MB/sec (sync dirs) 12 clients 12 procs max_latency=269.741 ms
After patchset:
Operation Count AvgLat MaxLat
----------------------------------------
NTCreateX 4152738 0.029 20.863
Close 3050770 0.002 1.119
Rename 175829 0.871 211.741
Unlink 838447 0.845 211.724
Deltree 120 4.798 14.162
Mkdir 60 0.003 0.005
Qpathinfo 3763807 0.011 4.673
Qfileinfo 660111 0.001 0.400
Qfsinfo 690141 0.003 0.429
Sfileinfo 338260 0.005 0.725
Find 1455273 0.046 6.787
WriteX 2073307 0.017 5.690
ReadX 6509193 0.003 1.171
LockX 13522 0.003 0.077
UnlockX 13522 0.002 0.125
Flush 291044 0.811 211.631
Throughput 1089.27 MB/sec (sync dirs) 12 clients 12 procs max_latency=211.750 ms
(+25.2% throughput, -21.5% max latency)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
When we find a conflicting inode (an inode that had the same name and
parent directory as the inode we are logging now) that was deleted in the
current transaction, we always end up logging its parent directory.
This is to deal with the case where the conflicting inode corresponds to
a deleted subvolume/snapshot or a directory that had subvolumes/snapshots
(or some subdirectory inside it had subvolumes/snapshots, etc), because
we can't deal with dropping subvolumes/snapshots during log replay. So
if we log the parent directory, and if we are dealing with these special
cases, then we fallback to a transaction commit when logging the parent,
because its last_unlink_trans will match the current transaction (which
gets set and propagated when a subvolume/snapshot is deleted).
This change skips the logging of the parent directory when the conflicting
inode is not a directory (or a subvolume/snapshot). This is ok because in
this case logging the current inode is enough to trigger an unlink of the
conflicting inode during log replay.
So for a case like this:
$ mkdir /mnt/dir
$ echo -n "first foo data" > /mnt/dir/foo
$ sync
$ rm -f /mnt/dir/foo
$ echo -n "second foo data" > /mnt/dir/foo
$ xfs_io -c "fsync" /mnt/dir/foo
We avoid logging parent directory "dir" when logging the new file "foo".
In other cases it avoids falling back to a transaction commit, when the
parent directory has a last_unlink_trans value that matches the current
transaction, due to moving a file from it to some other directory.
This is a case that happens frequently with dbench for example, where a
new file that has the name/parent of another file that was deleted in the
current transaction, is fsynced.
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
When logging an inode, if we detect the inode has a reference that
conflicts with some other inode that got renamed, we log that other inode
while holding the log mutex of the current inode. We then find out if
there are other inodes that conflict with the first conflicting inode,
and log them while under the log mutex of the original inode. This is
fine because the recursion can only happen once.
For the upcoming work where we directly log delayed items without flushing
them first to the subvolume tree, this recursion adds a lot of complexity
and it's hard to keep lockdep happy about it.
So collect a list of conflicting inodes and then log the inodes after
unlocking the log mutex of the inode we started with.
Also limit the maximum number of conflict inodes we log to 10, to avoid
spending too much time logging (and maybe allocating too many list
elements too), as typically we don't have more than 1 or 2 conflicting
inodes - if we go over the limit, simply fallback to a transaction commit.
It is possible to have a very long list of conflicting inodes to be
intentionally created by a user if he/she creates a very long succession
of renames like this:
(...)
rename E to F
rename D to E
rename C to D
rename B to C
rename A to B
touch A (create a new file named A)
fsync A
If that happened for a sequence of hundreds or thousands of renames, it
could massively slow down the logging and cause other secondary effects
like for example blocking other fsync operations and transaction commits
for a very long time (assuming it wouldn't run into -ENOSPC or -ENOMEM
first). However such cases are very uncommon to happen in practice,
nevertheless it's better to be prepared for them and avoid chaos.
Such long sequence of conflicting inodes could be created before this
change.
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
The static function log_new_dir_dentries() is currently defined below
btrfs_log_inode(), but in an upcoming patch a new function is introduced
that is called by btrfs_log_inode() and this new function needs to call
log_new_dir_dentries(). So move log_new_dir_dentries() to a location
between btrfs_log_inode() and need_log_inode() (the later is called by
log_new_dir_dentries()).
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
The static function need_log_inode() is defined below btrfs_log_inode()
and log_conflicting_inodes(), but in the next patches in the series we
will need to call need_log_inode() in a couple new functions that will be
used by btrfs_log_inode(). So move its definition to a location above
log_conflicting_inodes().
Also make its arguments 'const', since they are not supposed to be
modified.
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
The key offset of the last dir index item that was logged is stored in
the inode's last_dir_index_offset field. However that field is not
persisted in the inode item or elsewhere, so if the inode gets evicted
and reloaded, it gets a value of (u64)-1, so that when we are logging
dir index items we check if they were logged before, to avoid attempts
to insert duplicated keys and fallback to a transaction commit.
Improve on this by searching for the last dir index that was logged when
we start logging a directory if the inode's last_dir_index_offset is not
set (has a value of (u64)-1) and it was logged before. This avoids
checking if each dir index item we find was already logged before, and
simplifies the logging of dir index items (process_dir_items_leaf()).
This will also be needed for an incoming change where we start logging
delayed items directly, without flushing them first.
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Currently struct btrfs_delayed_item has a base size of 96 bytes, but its
size can be decreased by doing the following 2 tweaks:
1) Change data_len from u32 to u16. Our maximum possible leaf size is 64K,
so the data_len can never be larger than that, and in fact it is always
much smaller than that. The max length for a dentry's name is ensured
at the VFS level (PATH_MAX, 4096 bytes) and in struct btrfs_inode_ref
and btrfs_dir_item we use a u16 to store the name's length;
2) Change 'ins_or_del' to a 1 bit enum, which is all we need since it
can only have 2 values. After this there's also no longer the need to
BUG_ON() before using 'ins_or_del' in several places. Also rename the
field from 'ins_or_del' to 'type', which is more clear.
These two tweaks decrease the size of struct btrfs_delayed_item from 96
bytes down to 88 bytes. A previous patch already reduced the size of this
structure by 16 bytes, but an upcoming change will increase its size by
16 bytes (adding a struct list_head element).
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
All callers pass NULL to the 'prev' and 'next' arguments of the function
__btrfs_lookup_delayed_item(), so remove these arguments. Also, remove
the unnecessary wrapper __btrfs_lookup_delayed_insertion_item(), making
btrfs_delete_delayed_insertion_item() directly call
__btrfs_lookup_delayed_item().
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
All delayed items are for dir index keys, so there's really no point of
having an embedded struct btrfs_key in struct btrfs_delayed_item, which
makes the structure use more space than necessary (and adds a hole of 7
bytes).
So replace the key field with an index number (u64), which reduces the
size of struct btrfs_delayed_item from 112 bytes down to 96 bytes.
Some upcoming work will increase the structure size by 16 bytes, so this
change compensates for that future size increase.
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
The root argument of btrfs_delayed_item_reserve_metadata() is used only
to get the fs_info object, but we already have a transaction handle, which
we can use to get the fs_info. So remove the root argument.
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
At log_new_dir_dentries() we always start by allocating a list element
for the starting inode and then do a while loop with the condition being
a list emptiness check.
This however is not needed, we can avoid allocating this initial list
element and then just check for the list emptiness at the end of the
loop's body. So just do that to save one memory allocation from the
kmalloc-32 slab.
This allows for not doing any memory allocation when we don't have any
subdirectory to log, which is a very common case.
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
At log_new_dir_dentries(), there's no need to keep the current list
element allocated while processing the leaves with directory items for
the current directory, and while logging other inodes. Plus in case we
find a subdirectory, we also end up allocating a new list element while
the current one is still allocated, temporarily using more memory than
necessary.
So free the current list element early on, before processing leaves.
Also make the removal and release of all list elements in case of an
error more simple by eliminating the label and goto, adding an explicit
loop to release all list elements in case an error happens.
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
The comment refers to the function log_dir_items() in order to check why
the inodes of new directory entries need to be logged, but the relevant
comments are no longer at log_dir_items(), they were moved to the function
process_dir_items_leaf() in commit eb10d85ee7 ("btrfs: factor out the
copying loop of dir items from log_dir_items()"). So update it with the
current function name.
Also remove references with i_mutex to "VFS lock", since the inode lock
is no longer a mutex since 2016 (it's now a rw semaphore).
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
There's no point in passing a root argument to log_new_dir_dentries()
because it always corresponds to the root of the given inode. So remove
it and extract the root from the given inode.
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
When logging a directory that was previously logged in the current
transaction, we drop all the range items (BTRFS_DIR_LOG_INDEX_KEY key
type). This is because we will process all leaves in the subvolume's tree
that were changed in the current transaction and then add range items for
covering new dir index items and deleted dir index items, which could
cover now a larger range than before.
We used to fail if we tried to insert a range item key that already
exists, so we dropped all range items to avoid failing. However nowadays,
since commit 750ee45490 ("btrfs: fix assertion failure when logging
directory key range item"), we simply update any range item that already
exists, increasing its range's last dir index if needed. Since the range
covered by a range item can never decrease, due to the fact that dir index
values come from a monotonically increasing counter and are never reused,
we can stop dropping all range items before we start logging a directory.
By not dropping the items we can avoid having occasional tree rebalance
operations.
This will also be needed for an incoming change where we start logging
delayed items directly, without flushing them first.
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
[PROBLEM]
The existing scrub code for data extents always limit the block size to
sectorsize.
This causes quite some extra scrub_block being allocated:
(there is a data extent at logical bytenr 298844160, length 64KiB)
alloc_scrub_block: new block: logical=298844160 physical=298844160 mirror=1
alloc_scrub_block: new block: logical=298848256 physical=298848256 mirror=1
alloc_scrub_block: new block: logical=298852352 physical=298852352 mirror=1
alloc_scrub_block: new block: logical=298856448 physical=298856448 mirror=1
alloc_scrub_block: new block: logical=298860544 physical=298860544 mirror=1
alloc_scrub_block: new block: logical=298864640 physical=298864640 mirror=1
alloc_scrub_block: new block: logical=298868736 physical=298868736 mirror=1
alloc_scrub_block: new block: logical=298872832 physical=298872832 mirror=1
alloc_scrub_block: new block: logical=298876928 physical=298876928 mirror=1
alloc_scrub_block: new block: logical=298881024 physical=298881024 mirror=1
alloc_scrub_block: new block: logical=298885120 physical=298885120 mirror=1
alloc_scrub_block: new block: logical=298889216 physical=298889216 mirror=1
alloc_scrub_block: new block: logical=298893312 physical=298893312 mirror=1
alloc_scrub_block: new block: logical=298897408 physical=298897408 mirror=1
alloc_scrub_block: new block: logical=298901504 physical=298901504 mirror=1
alloc_scrub_block: new block: logical=298905600 physical=298905600 mirror=1
...
scrub_block_put: free block: logical=298844160 physical=298844160 len=4096 mirror=1
scrub_block_put: free block: logical=298848256 physical=298848256 len=4096 mirror=1
scrub_block_put: free block: logical=298852352 physical=298852352 len=4096 mirror=1
scrub_block_put: free block: logical=298856448 physical=298856448 len=4096 mirror=1
scrub_block_put: free block: logical=298860544 physical=298860544 len=4096 mirror=1
scrub_block_put: free block: logical=298864640 physical=298864640 len=4096 mirror=1
scrub_block_put: free block: logical=298868736 physical=298868736 len=4096 mirror=1
scrub_block_put: free block: logical=298872832 physical=298872832 len=4096 mirror=1
scrub_block_put: free block: logical=298876928 physical=298876928 len=4096 mirror=1
scrub_block_put: free block: logical=298881024 physical=298881024 len=4096 mirror=1
scrub_block_put: free block: logical=298885120 physical=298885120 len=4096 mirror=1
scrub_block_put: free block: logical=298889216 physical=298889216 len=4096 mirror=1
scrub_block_put: free block: logical=298893312 physical=298893312 len=4096 mirror=1
scrub_block_put: free block: logical=298897408 physical=298897408 len=4096 mirror=1
scrub_block_put: free block: logical=298901504 physical=298901504 len=4096 mirror=1
scrub_block_put: free block: logical=298905600 physical=298905600 len=4096 mirror=1
This behavior will waste a lot of memory, especially after we have moved
quite some members from scrub_sector to scrub_block.
[FIX]
To reduce the allocation of scrub_block, and to reduce memory usage, use
BTRFS_STRIPE_LEN instead of sectorsize as the block size to scrub data
extents.
This results only one scrub_block to be allocated for above data extent:
alloc_scrub_block: new block: logical=298844160 physical=298844160 mirror=1
scrub_block_put: free block: logical=298844160 physical=298844160 len=65536 mirror=1
This would greatly reduce the memory usage (even it's just transient)
for larger data extents scrub.
For above example, the memory usage would be:
Old: num_sectors * (sizeof(scrub_block) + sizeof(scrub_sector))
16 * (408 + 96) = 8065
New: sizeof(scrub_block) + num_sectors * sizeof(scrub_sector)
408 + 16 * 96 = 1944
A good reduction of 75.9%.
Signed-off-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Currently we store the following members in scrub_sector:
- logical
- physical
- physical_for_dev_replace
- dev
- mirror_num
However the current scrub code has ensured that scrub_blocks never cross
stripe boundary.
This is caused by the entry functions (scrub_simple_mirror,
scrub_simple_stripe), thus every scrub_block will not cross stripe
boundary.
Thus this makes it possible to move those members into scrub_block other
than putting them into scrub_sector.
This should save quite some memory, as a scrub_block can be as large as 64
sectors, even for metadata it's 16 sectors byte default.
Signed-off-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Although scrub currently works for subpage (PAGE_SIZE > sectorsize) cases,
it will allocate one page for each scrub_sector, which can cause extra
unnecessary memory usage.
Utilize scrub_block::pages[] instead of allocating page for each
scrub_sector, this allows us to integrate larger extents while using
less memory.
For example, if our page size is 64K, sectorsize is 4K, and we got an
32K sized extent.
We will only allocate one page for scrub_block, and all 8 scrub sectors
will point to that page.
To do that properly, here we introduce several small helpers:
- scrub_page_get_logical()
Get the logical bytenr of a page.
We store the logical bytenr of the page range into page::private.
But for 32bit systems, their (void *) is not large enough to contain
a u64, so in that case we will need to allocate extra memory for it.
For 64bit systems, we can use page::private directly.
- scrub_block_get_logical()
Just get the logical bytenr of the first page.
- scrub_sector_get_page()
Return the page which the scrub_sector points to.
- scrub_sector_get_page_offset()
Return the offset inside the page which the scrub_sector points to.
- scrub_sector_get_kaddr()
Return the address which the scrub_sector points to.
Just a wrapper using scrub_sector_get_page() and
scrub_sector_get_page_offset()
- bio_add_scrub_sector()
Please note that, even with this patch, we're still allocating one page
for one sector for data extents.
This is because in scrub_extent() we split the data extent using
sectorsize.
The memory usage reduction will need extra work to make scrub to work
like data read to only use the correct sector(s).
Signed-off-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
[BACKGROUND]
Currently for scrub, we allocate one page for one sector, this is fine
for PAGE_SIZE == sectorsize support, but can waste extra memory for
subpage support.
[CODE CHANGE]
Make scrub_block contain all the pages, so if we're scrubbing an extent
sized 64K, and our page size is also 64K, we only need to allocate one
page.
[LIFESPAN CHANGE]
Since now scrub_sector no longer holds a page, but is using
scrub_block::pages[] instead, we have to ensure scrub_block has a longer
lifespan for write bio. The lifespan for read bio is already large
enough.
Now scrub_block will only be released after the write bio finished.
[COMING NEXT]
Currently we only added scrub_block::pages[] for this purpose, but
scrub_sector is still utilizing the old scrub_sector::page.
The switch will happen in the next patch.
Signed-off-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
The allocation and initialization is shared by 3 call sites, and we're
going to change the initialization of some members in the upcoming
patches.
So factor out the allocation and initialization of scrub_sector into a
helper, alloc_scrub_sector(), which will do the following work:
- Allocate the memory for scrub_sector
- Allocate a page for scrub_sector::page
- Initialize scrub_sector::refs to 1
- Attach the allocated scrub_sector to scrub_block
The attachment is bidirectional, which means scrub_block::sectorv[]
will be updated and scrub_sector::sblock will also be updated.
- Update scrub_block::sector_count and do extra sanity check on it
Signed-off-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Although there are only two callers, we are going to add some members
for scrub_block in the incoming patches. Factoring out the
initialization code will make later expansion easier.
One thing to note is, even scrub_handle_errored_block() doesn't utilize
scrub_block::refs, we still use alloc_scrub_block() to initialize
sblock::ref, allowing us to use scrub_block_put() to do cleanup.
Signed-off-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
In function scrub_handle_errored_block(), we use @sblocks_for_recheck
pointer to hold one scrub_block for each mirror, and uses kcalloc() to
allocate an array.
But this one pointer for an array is not readable due to the member
offsets done by addition and not [].
Change this pointer to struct scrub_block *[BTRFS_MAX_MIRRORS], this
will slightly increase the stack memory usage.
Since function scrub_handle_errored_block() won't get iterative calls,
this extra cost would completely be acceptable.
And since we're here, also set sblock->refs and use scrub_block_put() to
clean them up, as later we will add extra members in scrub_block, which
needs scrub_block_put() to clean them up.
Signed-off-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Preserve the fs-verity status of a btrfs file across send/recv.
There is no facility for installing the Merkle tree contents directly on
the receiving filesystem, so we package up the parameters used to enable
verity found in the verity descriptor. This gives the receive side
enough information to properly enable verity again. Note that this means
that receive will have to re-compute the whole Merkle tree, similar to
how compression worked before encoded_write.
Since the file becomes read-only after verity is enabled, it is
important that verity is added to the send stream after any file writes.
Therefore, when we process a verity item, merely note that it happened,
then actually create the command in the send stream during
'finish_inode_if_needed'.
This also creates V3 of the send stream format, without any format
changes besides adding the new commands and attributes.
Signed-off-by: Boris Burkov <boris@bur.io>
Signed-off-by: David Sterba <dsterba@suse.com>
Use `atomic_try_cmpxchg(ptr, &old, new)` instead of
`atomic_cmpxchg(ptr, old, new) == old` in free_extent_buffer. This
has two benefits:
- The x86 cmpxchg instruction returns success in the ZF flag, so this
change saves a compare after cmpxchg, as well as a related move
instruction in the front of cmpxchg.
- atomic_try_cmpxchg implicitly assigns the *ptr value to &old when
cmpxchg fails, enabling further code simplifications.
This patch has no functional change.
Reviewed-by: Boris Burkov <boris@bur.io>
Signed-off-by: Uros Bizjak <ubizjak@gmail.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
There are several sanity checks which are no longer possible to trigger
inside btrfs_scrub_dev().
Since we have mount time check against super block nodesize/sectorsize,
and our fixed macro is hardcoded to handle even the worst combination.
Thus those sanity checks are no longer needed, can be easily removed.
But this patch still uses some ASSERT()s as a safe net just in case we
change some features in the future to trigger those impossible
combinations.
Signed-off-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
We used to use this in a few spots, but now we only use it directly
inside of block-group.c, so remove the helper and just open code where
we were using it.
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Before when this was modifying the bit field we had to protect it with
the bg->lock, however now we're using bit helpers so we can stop
using the bg->lock.
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
This is used mostly to determine if we need to look at the caching ctl
list and clean up any references to this block group. However we never
clear this flag, specifically because we need to know if we have to
remove a caching ctl we have for this block group still. This is in the
remove block group path which isn't a fast path, so the optimization
doesn't really matter, simplify this logic and remove the flag.
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
We're breaking out and re-searching for the next block group while
evicting any of the block group cache inodes. This is not needed, the
block groups aren't disappearing here, we can simply loop through the
block groups like normal and iput any inode that we find.
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
We use this during device replace for zoned devices, we were simply
taking the lock because it was in a bit field and we needed the lock to
be safe with other modifications in the bitfield. With the bit helpers
we no longer require that locking.
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
We use a bit field in the btrfs_block_group for different flags, however
this is awkward because we have to hold the block_group->lock for any
modification of any of these fields, and makes the code clunky for a few
of these flags. Convert these to a properly flags setup so we can
utilize the bit helpers.
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
We previously had the pattern of
btrfs_update_space_info(all, the, bg, fields, &space_info);
link_block_group(bg);
bg->space_info = space_info;
Now that we're passing the bg into btrfs_add_bg_to_space_info we can do
the linking in that function, transforming this to simply
btrfs_add_bg_to_space_info(fs_info, bg);
and put the link_block_group() and bg->space_info assignment directly in
btrfs_add_bg_to_space_info.
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
This function has grown a bunch of new arguments, and it just boils down
to passing in all the block group fields as arguments. Simplify this by
passing in the block group itself and updating the space_info fields
based on the block group fields directly.
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
For both unused bg deletion and async balance work we'll happily run if
the fs is closing. However I want to move these to their own worker
thread, and they can be long running jobs, so add a check to see if
we're closing and simply bail.
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
btrfs_insert_file_extent() is only ever used to insert holes, so rename
it and remove the redundant parameters.
Reviewed-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: Omar Sandoval <osandov@osandov.com>
Signed-off-by: Sweet Tea Dorminy <sweettea-kernel@dorminy.me>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
We have own string matching helper that duplicates what sysfs_streq
does, with a slight difference that it skips initial whitespace. So far
this is used for the drive allocation policy. The initial whitespace
of written sysfs values should be rather discouraged and we should use a
standard helper.
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Reviewed-by: Anand Jain <anand.jain@oracle.com>
Signed-off-by: David Sterba <dsterba@suse.com>
[BUG]
The following script shows that, although scrub can detect super block
errors, it never tries to fix it:
mkfs.btrfs -f -d raid1 -m raid1 $dev1 $dev2
xfs_io -c "pwrite 67108864 4k" $dev2
mount $dev1 $mnt
btrfs scrub start -B $dev2
btrfs scrub start -Br $dev2
umount $mnt
The first scrub reports the super error correctly:
scrub done for f3289218-abd3-41ac-a630-202f766c0859
Scrub started: Tue Aug 2 14:44:11 2022
Status: finished
Duration: 0:00:00
Total to scrub: 1.26GiB
Rate: 0.00B/s
Error summary: super=1
Corrected: 0
Uncorrectable: 0
Unverified: 0
But the second read-only scrub still reports the same super error:
Scrub started: Tue Aug 2 14:44:11 2022
Status: finished
Duration: 0:00:00
Total to scrub: 1.26GiB
Rate: 0.00B/s
Error summary: super=1
Corrected: 0
Uncorrectable: 0
Unverified: 0
[CAUSE]
The comments already shows that super block can be easily fixed by
committing a transaction:
/*
* If we find an error in a super block, we just report it.
* They will get written with the next transaction commit
* anyway
*/
But the truth is, such assumption is not always true, and since scrub
should try to repair every error it found (except for read-only scrub),
we should really actively commit a transaction to fix this.
[FIX]
Just commit a transaction if we found any super block errors, after
everything else is done.
We cannot do this just after scrub_supers(), as
btrfs_commit_transaction() will try to pause and wait for the running
scrub, thus we can not call it with scrub_lock hold.
Signed-off-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
[PROBLEM]
Unlike data/metadata corruption, if scrub detected some error in the
super block, the only error message is from the updated device status:
BTRFS info (device dm-1): scrub: started on devid 2
BTRFS error (device dm-1): bdev /dev/mapper/test-scratch2 errs: wr 0, rd 0, flush 0, corrupt 1, gen 0
BTRFS info (device dm-1): scrub: finished on devid 2 with status: 0
This is not helpful at all.
[CAUSE]
Unlike data/metadata error reporting, there is no visible report in
kernel dmesg to report supper block errors.
In fact, return value of scrub_checksum_super() is intentionally
skipped, thus scrub_handle_errored_block() will never be called for
super blocks.
[FIX]
Make super block errors to output an error message, now the full
dmesg would looks like this:
BTRFS info (device dm-1): scrub: started on devid 2
BTRFS warning (device dm-1): super block error on device /dev/mapper/test-scratch2, physical 67108864
BTRFS error (device dm-1): bdev /dev/mapper/test-scratch2 errs: wr 0, rd 0, flush 0, corrupt 1, gen 0
BTRFS info (device dm-1): scrub: finished on devid 2 with status: 0
BTRFS info (device dm-1): scrub: started on devid 2
This fix involves:
- Move the super_errors reporting to scrub_handle_errored_block()
This allows the device status message to show after the super block
error message.
But now we no longer distinguish super block corruption and generation
mismatch, now all counted as corruption.
- Properly check the return value from scrub_checksum_super()
- Add extra super block error reporting for scrub_print_warning().
Signed-off-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
With CONFIG_READ_ONLY_THP_FOR_FS, the Linux kernel supports using THPs for
read-only mmapped files, such as shared libraries. However, the kernel
makes no attempt to actually align those mappings on 2MB boundaries,
which makes it impossible to use those THPs most of the time. This issue
applies to general file mapping THP as well as existing setups using
CONFIG_READ_ONLY_THP_FOR_FS. This is easily fixed by using
thp_get_unmapped_area for the unmapped_area function in btrfs, which
is what ext2, ext4, fuse, and xfs all use.
Initially btrfs had been left out in commit 8c07fc452ac0 ("btrfs: fix
alignment of VMA for memory mapped files on THP") as btrfs does not support
DAX. However, commit 1854bc6e24 ("mm/readahead: Align file mappings
for non-DAX") removed the DAX requirement. We should now be able to call
thp_get_unmapped_area() for btrfs.
The problem can be seen in /proc/PID/smaps where THPeligible is set to 0
on mappings to eligible shared object files as shown below.
Before this patch:
7fc6a7e18000-7fc6a80cc000 r-xp 00000000 00:1e 199856
/usr/lib64/libcrypto.so.1.1.1k
Size: 2768 kB
THPeligible: 0
VmFlags: rd ex mr mw me
With this patch the library is mapped at a 2MB aligned address:
fbdfe200000-7fbdfe4b4000 r-xp 00000000 00:1e 199856
/usr/lib64/libcrypto.so.1.1.1k
Size: 2768 kB
THPeligible: 1
VmFlags: rd ex mr mw me
This fixes the alignment of VMAs for any mmap of a file that has the
rd and ex permissions and size >= 2MB. The VMA alignment and
THPeligible field for anonymous memory is handled separately and
is thus not effected by this change.
CC: stable@vger.kernel.org # 5.18+
Signed-off-by: Alexander Zhu <alexlzhu@fb.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
This wait event is very similar to the pending ordered wait event in the
sense that it occurs in a different context than the condition signaling
for the event. The signaling occurs in btrfs_remove_ordered_extent()
while the wait event is implemented in btrfs_start_ordered_extent() in
fs/btrfs/ordered-data.c
However, in this case a thread must not acquire the lockdep map for the
ordered extents wait event when the ordered extent is related to a free
space inode. That is because lockdep creates dependencies between locks
acquired both in execution paths related to normal inodes and paths
related to free space inodes, thus leading to false positives.
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Ioannis Angelakopoulos <iangelak@fb.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Reinitialize the class of the lockdep map for struct inode's
mapping->invalidate_lock in load_free_space_cache() function in
fs/btrfs/free-space-cache.c. This will prevent lockdep from producing
false positives related to execution paths that make use of free space
inodes and paths that make use of normal inodes.
Specifically, with this change lockdep will create separate lock
dependencies that include the invalidate_lock, in the case that free
space inodes are used and in the case that normal inodes are used.
The lockdep class for this lock was first initialized in
inode_init_always() in fs/inode.c.
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Ioannis Angelakopoulos <iangelak@fb.com>
Signed-off-by: David Sterba <dsterba@suse.com>
In contrast to the num_writers and num_extwriters wait events, the
condition for the pending ordered wait event is signaled in a different
context from the wait event itself. The condition signaling occurs in
btrfs_remove_ordered_extent() in fs/btrfs/ordered-data.c while the wait
event is implemented in btrfs_commit_transaction() in
fs/btrfs/transaction.c
Thus the thread signaling the condition has to acquire the lockdep map
as a reader at the start of btrfs_remove_ordered_extent() and release it
after it has signaled the condition. In this case some dependencies
might be left out due to the placement of the annotation, but it is
better than no annotation at all.
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Ioannis Angelakopoulos <iangelak@fb.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Add lockdep annotations for the transaction states that have wait
events;
1) TRANS_STATE_COMMIT_START
2) TRANS_STATE_UNBLOCKED
3) TRANS_STATE_SUPER_COMMITTED
4) TRANS_STATE_COMPLETED
The new macros introduced here to annotate the transaction states wait
events have the same effect as the generic lockdep annotation macros.
With the exception of the lockdep annotation for TRANS_STATE_COMMIT_START
the transaction thread has to acquire the lockdep maps for the
transaction states as reader after the lockdep map for num_writers is
released so that lockdep does not complain.
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Ioannis Angelakopoulos <iangelak@fb.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Similarly to the num_writers wait event in fs/btrfs/transaction.c add a
lockdep annotation for the num_extwriters wait event.
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Ioannis Angelakopoulos <iangelak@fb.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Annotate the num_writers wait event in fs/btrfs/transaction.c with
lockdep in order to catch deadlocks involving this wait event.
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Ioannis Angelakopoulos <iangelak@fb.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Introduce four macros that are used to annotate wait events in btrfs code
with lockdep;
1) the btrfs_lockdep_init_map
2) the btrfs_lockdep_acquire,
3) the btrfs_lockdep_release
4) the btrfs_might_wait_for_event macros.
The btrfs_lockdep_init_map macro is used to initialize a lockdep map.
The btrfs_lockdep_<acquire,release> macros are used by threads to take
the lockdep map as readers (shared lock) and release it, respectively.
The btrfs_might_wait_for_event macro is used by threads to take the
lockdep map as writers (exclusive lock) and release it.
In general, the lockdep annotation for wait events work as follows:
The condition for a wait event can be modified and signaled at the same
time by multiple threads. These threads hold the lockdep map as readers
when they enter a context in which blocking would prevent signaling the
condition. Frequently, this occurs when a thread violates a condition
(lockdep map acquire), before restoring it and signaling it at a later
point (lockdep map release).
The threads that block on the wait event take the lockdep map as writers
(exclusive lock). These threads have to block until all the threads that
hold the lockdep map as readers signal the condition for the wait event
and release the lockdep map.
The lockdep annotation is used to warn about potential deadlock scenarios
that involve the threads that modify and signal the wait event condition
and threads that block on the wait event. A simple example is illustrated
below:
Without lockdep:
TA TB
cond = false
lock(A)
wait_event(w, cond)
unlock(A)
lock(A)
cond = true
signal(w)
unlock(A)
With lockdep:
TA TB
rwsem_acquire_read(lockdep_map)
cond = false
lock(A)
rwsem_acquire(lockdep_map)
rwsem_release(lockdep_map)
wait_event(w, cond)
unlock(A)
lock(A)
cond = true
signal(w)
unlock(A)
rwsem_release(lockdep_map)
In the second case, with the lockdep annotation, lockdep would warn about
an ABBA deadlock, while the first case would just deadlock at some point.
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Ioannis Angelakopoulos <iangelak@fb.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
There is an internal report on hitting the following ASSERT() in
recalculate_thresholds():
ASSERT(ctl->total_bitmaps <= max_bitmaps);
Above @max_bitmaps is calculated using the following variables:
- bytes_per_bg
8 * 4096 * 4096 (128M) for x86_64/x86.
- block_group->length
The length of the block group.
@max_bitmaps is the rounded up value of block_group->length / 128M.
Normally one free space cache should not have more bitmaps than above
value, but when it happens the ASSERT() can be triggered if
CONFIG_BTRFS_ASSERT is also enabled.
But the ASSERT() itself won't provide enough info to know which is going
wrong.
Is the bg too small thus it only allows one bitmap?
Or is there something else wrong?
So although I haven't found extra reports or crash dump to do further
investigation, add the extra info to make it more helpful to debug.
Reviewed-by: Anand Jain <anand.jain@oracle.com>
Signed-off-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
-----BEGIN PGP SIGNATURE-----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=4Hu8
-----END PGP SIGNATURE-----
Merge tag 'for-6.0-rc6-tag' of git://git.kernel.org/pub/scm/linux/kernel/git/kdave/linux
Pull btrfs fixes from David Sterba:
- two fixes for hangs in the umount sequence where threads depend on
each other and the work must be finished in the right order
- in zoned mode, wait for flushing all block group metadata IO before
finishing the zone
* tag 'for-6.0-rc6-tag' of git://git.kernel.org/pub/scm/linux/kernel/git/kdave/linux:
btrfs: zoned: wait for extent buffer IOs before finishing a zone
btrfs: fix hang during unmount when stopping a space reclaim worker
btrfs: fix hang during unmount when stopping block group reclaim worker
Before sending REQ_OP_ZONE_FINISH to a zone, we need to ensure that
ongoing IOs already finished. Or, we will see a "Zone Is Full" error for
the IOs, as the ZONE_FINISH command makes the zone full.
We ensure that with btrfs_wait_block_group_reservations() and
btrfs_wait_ordered_roots() for a data block group. And, for a metadata
block group, the comparison of alloc_offset vs meta_write_pointer mostly
ensures IOs for the allocated region already sent. However, there still
can be a little time frame where the IOs are sent but not yet completed.
Introduce wait_eb_writebacks() to ensure such IOs are completed for a
metadata block group. It walks the buffer_radix to find extent buffers in
the block group and calls wait_on_extent_buffer_writeback() on them.
Fixes: afba2bc036 ("btrfs: zoned: implement active zone tracking")
CC: stable@vger.kernel.org # 5.19+
Signed-off-by: Naohiro Aota <naohiro.aota@wdc.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Often when running generic/562 from fstests we can hang during unmount,
resulting in a trace like this:
Sep 07 11:52:00 debian9 unknown: run fstests generic/562 at 2022-09-07 11:52:00
Sep 07 11:55:32 debian9 kernel: INFO: task umount:49438 blocked for more than 120 seconds.
Sep 07 11:55:32 debian9 kernel: Not tainted 6.0.0-rc2-btrfs-next-122 #1
Sep 07 11:55:32 debian9 kernel: "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
Sep 07 11:55:32 debian9 kernel: task:umount state:D stack: 0 pid:49438 ppid: 25683 flags:0x00004000
Sep 07 11:55:32 debian9 kernel: Call Trace:
Sep 07 11:55:32 debian9 kernel: <TASK>
Sep 07 11:55:32 debian9 kernel: __schedule+0x3c8/0xec0
Sep 07 11:55:32 debian9 kernel: ? rcu_read_lock_sched_held+0x12/0x70
Sep 07 11:55:32 debian9 kernel: schedule+0x5d/0xf0
Sep 07 11:55:32 debian9 kernel: schedule_timeout+0xf1/0x130
Sep 07 11:55:32 debian9 kernel: ? lock_release+0x224/0x4a0
Sep 07 11:55:32 debian9 kernel: ? lock_acquired+0x1a0/0x420
Sep 07 11:55:32 debian9 kernel: ? trace_hardirqs_on+0x2c/0xd0
Sep 07 11:55:32 debian9 kernel: __wait_for_common+0xac/0x200
Sep 07 11:55:32 debian9 kernel: ? usleep_range_state+0xb0/0xb0
Sep 07 11:55:32 debian9 kernel: __flush_work+0x26d/0x530
Sep 07 11:55:32 debian9 kernel: ? flush_workqueue_prep_pwqs+0x140/0x140
Sep 07 11:55:32 debian9 kernel: ? trace_clock_local+0xc/0x30
Sep 07 11:55:32 debian9 kernel: __cancel_work_timer+0x11f/0x1b0
Sep 07 11:55:32 debian9 kernel: ? close_ctree+0x12b/0x5b3 [btrfs]
Sep 07 11:55:32 debian9 kernel: ? __trace_bputs+0x10b/0x170
Sep 07 11:55:32 debian9 kernel: close_ctree+0x152/0x5b3 [btrfs]
Sep 07 11:55:32 debian9 kernel: ? evict_inodes+0x166/0x1c0
Sep 07 11:55:32 debian9 kernel: generic_shutdown_super+0x71/0x120
Sep 07 11:55:32 debian9 kernel: kill_anon_super+0x14/0x30
Sep 07 11:55:32 debian9 kernel: btrfs_kill_super+0x12/0x20 [btrfs]
Sep 07 11:55:32 debian9 kernel: deactivate_locked_super+0x2e/0xa0
Sep 07 11:55:32 debian9 kernel: cleanup_mnt+0x100/0x160
Sep 07 11:55:32 debian9 kernel: task_work_run+0x59/0xa0
Sep 07 11:55:32 debian9 kernel: exit_to_user_mode_prepare+0x1a6/0x1b0
Sep 07 11:55:32 debian9 kernel: syscall_exit_to_user_mode+0x16/0x40
Sep 07 11:55:32 debian9 kernel: do_syscall_64+0x48/0x90
Sep 07 11:55:32 debian9 kernel: entry_SYSCALL_64_after_hwframe+0x63/0xcd
Sep 07 11:55:32 debian9 kernel: RIP: 0033:0x7fcde59a57a7
Sep 07 11:55:32 debian9 kernel: RSP: 002b:00007ffe914217c8 EFLAGS: 00000246 ORIG_RAX: 00000000000000a6
Sep 07 11:55:32 debian9 kernel: RAX: 0000000000000000 RBX: 00007fcde5ae8264 RCX: 00007fcde59a57a7
Sep 07 11:55:32 debian9 kernel: RDX: 0000000000000000 RSI: 0000000000000000 RDI: 000055b57556cdd0
Sep 07 11:55:32 debian9 kernel: RBP: 000055b57556cba0 R08: 0000000000000000 R09: 00007ffe91420570
Sep 07 11:55:32 debian9 kernel: R10: 0000000000000000 R11: 0000000000000246 R12: 0000000000000000
Sep 07 11:55:32 debian9 kernel: R13: 000055b57556cdd0 R14: 000055b57556ccb8 R15: 0000000000000000
Sep 07 11:55:32 debian9 kernel: </TASK>
What happens is the following:
1) The cleaner kthread tries to start a transaction to delete an unused
block group, but the metadata reservation can not be satisfied right
away, so a reservation ticket is created and it starts the async
metadata reclaim task (fs_info->async_reclaim_work);
2) Writeback for all the filler inodes with an i_size of 2K starts
(generic/562 creates a lot of 2K files with the goal of filling
metadata space). We try to create an inline extent for them, but we
fail when trying to insert the inline extent with -ENOSPC (at
cow_file_range_inline()) - since this is not critical, we fallback
to non-inline mode (back to cow_file_range()), reserve extents, create
extent maps and create the ordered extents;
3) An unmount starts, enters close_ctree();
4) The async reclaim task is flushing stuff, entering the flush states one
by one, until it reaches RUN_DELAYED_IPUTS. There it runs all current
delayed iputs.
After running the delayed iputs and before calling
btrfs_wait_on_delayed_iputs(), one or more ordered extents complete,
and btrfs_add_delayed_iput() is called for each one through
btrfs_finish_ordered_io() -> btrfs_put_ordered_extent(). This results
in bumping fs_info->nr_delayed_iputs from 0 to some positive value.
So the async reclaim task blocks at btrfs_wait_on_delayed_iputs() waiting
for fs_info->nr_delayed_iputs to become 0;
5) The current transaction is committed by the transaction kthread, we then
start unpinning extents and end up calling btrfs_try_granting_tickets()
through unpin_extent_range(), since we released some space.
This results in satisfying the ticket created by the cleaner kthread at
step 1, waking up the cleaner kthread;
6) At close_ctree() we ask the cleaner kthread to park;
7) The cleaner kthread starts the transaction, deletes the unused block
group, and then calls kthread_should_park(), which returns true, so it
parks. And at this point we have the delayed iputs added by the
completion of the ordered extents still pending;
8) Then later at close_ctree(), when we call:
cancel_work_sync(&fs_info->async_reclaim_work);
We hang forever, since the cleaner was parked and no one else can run
delayed iputs after that, while the reclaim task is waiting for the
remaining delayed iputs to be completed.
Fix this by waiting for all ordered extents to complete and running the
delayed iputs before attempting to stop the async reclaim tasks. Note that
we can not wait for ordered extents with btrfs_wait_ordered_roots() (or
other similar functions) because that waits for the BTRFS_ORDERED_COMPLETE
flag to be set on an ordered extent, but the delayed iput is added after
that, when doing the final btrfs_put_ordered_extent(). So instead wait for
the work queues used for executing ordered extent completion to be empty,
which works because we do the final put on an ordered extent at
btrfs_finish_ordered_io() (while we are in the unmount context).
Fixes: d6fd0ae25c ("Btrfs: fix missing delayed iputs on unmount")
CC: stable@vger.kernel.org # 5.15+
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
