Commits · 3e49363be6330f49e670240e8f46e6fe0bf5628a · Kirill Smelkov / linux

15 Feb, 2023 3 commits

btrfs: send: cache utimes operations for directories if possible · 3e49363b

Filipe Manana authored Jan 11, 2023

Whenever we add or remove an entry to a directory, we issue an utimes
command for the directory. If we add 1000 entries to a directory (create
1000 files under it or move 1000 files to it), then we issue the same
utimes command 1000 times, which increases the send stream size, results
in more pipe IO, one search in the send b+tree, allocating one path for
the search, etc, as well as making the receiver do a system call for each
duplicated utimes command.

We also issue an utimes command when we create a new directory, but later
we might add entries to it corresponding to inodes with an higher inode
number, so it's pointless to issue the utimes command before we create
the last inode under the directory.

So use a lru cache to track directories for which we must send a utimes
command. When we need to remove an entry from the cache, we issue the
utimes command for the respective directory. When finishing the send
operation, we go over each cache element and issue the respective utimes
command. Finally the caching is entirely optional, just a performance
optimization, meaning that if we fail to cache (due to memory allocation
failure), we issue the utimes command right away, that is, we fallback
to the previous, unoptimized, behaviour.

This patch belongs to a patchset comprised of the following patches:

  btrfs: send: directly return from did_overwrite_ref() and simplify it
  btrfs: send: avoid unnecessary generation search at did_overwrite_ref()
  btrfs: send: directly return from will_overwrite_ref() and simplify it
  btrfs: send: avoid extra b+tree searches when checking reference overrides
  btrfs: send: remove send_progress argument from can_rmdir()
  btrfs: send: avoid duplicated orphan dir allocation and initialization
  btrfs: send: avoid unnecessary orphan dir rbtree search at can_rmdir()
  btrfs: send: reduce searches on parent root when checking if dir can be removed
  btrfs: send: iterate waiting dir move rbtree only once when processing refs
  btrfs: send: initialize all the red black trees earlier
  btrfs: send: genericize the backref cache to allow it to be reused
  btrfs: adapt lru cache to allow for 64 bits keys on 32 bits systems
  btrfs: send: cache information about created directories
  btrfs: allow a generation number to be associated with lru cache entries
  btrfs: add an api to delete a specific entry from the lru cache
  btrfs: send: use the lru cache to implement the name cache
  btrfs: send: update size of roots array for backref cache entries
  btrfs: send: cache utimes operations for directories if possible

The following test was run before and after applying the whole patchset,
and on a non-debug kernel (Debian's default kernel config):

   #!/bin/bash

   MNT=/mnt/sdi
   DEV=/dev/sdi

   mkfs.btrfs -f $DEV > /dev/null
   mount $DEV $MNT

   mkdir $MNT/A
   for ((i = 1; i <= 20000; i++)); do
       echo -n > $MNT/A/file_$i
   done

   btrfs subvolume snapshot -r $MNT $MNT/snap1

   mkdir $MNT/B
   for ((i = 20000; i <= 40000; i++)); do
       echo -n > $MNT/B/file_$i
   done

   mv $MNT/A/file_* $MNT/B/

   btrfs subvolume snapshot -r $MNT $MNT/snap2

   start=$(date +%s%N)
   btrfs send -p $MNT/snap1 $MNT/snap2 > /dev/null
   end=$(date +%s%N)

   dur=$(( (end - start) / 1000000 ))
   echo "Incremental send took $dur milliseconds"

   umount $MNT

Before the whole patchset: 18408 milliseconds
After the whole patchset:   1942 milliseconds  (9.5x speedup)

Using 60000 files instead of 40000:

Before the whole patchset: 39764 milliseconds
After the whole patchset:   3076 milliseconds  (12.9x speedup)

Using 20000 files instead of 40000:

Before the whole patchset:  5072 milliseconds
After the whole patchset:    916 milliseconds  (5.5x speedup)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>

3e49363b

btrfs: send: update size of roots array for backref cache entries · ace79df8

Filipe Manana authored Jan 11, 2023

Currently we limit the size of the roots array, for backref cache entries,
to 12 elements. This is because that number is enough for most cases and
to make the backref cache entry size to be exactly 128 bytes, so that
memory is allocated from the kmalloc-128 slab and no space is wasted.

However recent changes in the series refactored the backref cache to be
more generic and allow it to be reused for other purposes, which resulted
in increasing the size of the embedded structure btrfs_lru_cache_entry in
order to allow for supporting inode numbers as keys on 32 bits system and
allow multiple generations per key. This resulted in increasing the size
of struct backref_cache_entry from 128 bytes to 152 bytes. Since the cache
entries are allocated with kmalloc(), it means we end up using the slab
kmalloc-192, so we end up wasting 40 bytes of memory. So bump the size of
the roots array from 12 elements to 17 elements, so we end up using 192
bytes for each backref cache entry.

This patch is part of a larger patchset and the changelog of the last
patch in the series contains a sample performance test and results.
The patches that comprise the patchset are the following:

btrfs: send: directly return from did_overwrite_ref() and simplify it
btrfs: send: avoid unnecessary generation search at did_overwrite_ref()
btrfs: send: directly return from will_overwrite_ref() and simplify it
btrfs: send: avoid extra b+tree searches when checking reference overrides
btrfs: send: remove send_progress argument from can_rmdir()
btrfs: send: avoid duplicated orphan dir allocation and initialization
btrfs: send: avoid unnecessary orphan dir rbtree search at can_rmdir()
btrfs: send: reduce searches on parent root when checking if dir can be removed
btrfs: send: iterate waiting dir move rbtree only once when processing refs
btrfs: send: initialize all the red black trees earlier
btrfs: send: genericize the backref cache to allow it to be reused
btrfs: adapt lru cache to allow for 64 bits keys on 32 bits systems
btrfs: send: cache information about created directories
btrfs: allow a generation number to be associated with lru cache entries
btrfs: add an api to delete a specific entry from the lru cache
btrfs: send: use the lru cache to implement the name cache
btrfs: send: update size of roots array for backref cache entries
btrfs: send: cache utimes operations for directories if possible
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>

ace79df8

btrfs: send: use the lru cache to implement the name cache · c48545de

Filipe Manana authored Jan 11, 2023

The name cache in send is basically a lru cache implemented with a radix
tree and linked lists, very similar to the lru cache module which is used
for the send backref cache and the cache of previously created directories
during a send operation. So remove all the custom caching code for the
name cache and make it use the lru cache instead.

One particular detail to note is that the current cache behaves a bit
differently when it comes to eviction of entries. Namely when after
inserting a new name in the cache, if the cache now has 256 entries, we
evict the last 128 LRU entries. The lru_cache.{c,h} module behaves a bit
differently in that once we reach the cache limit, we evict a single LRU
entry. In practice this doesn't make much difference, but it's actually
better to evict just one entry instead of half of the entries, as there's
always a chance we will need a name stored in one of that last 128 removed
entries.

This patch is part of a larger patchset and the changelog of the last
patch in the series contains a sample performance test and results.
The patches that comprise the patchset are the following:

c48545de

13 Feb, 2023 37 commits

btrfs: add an api to delete a specific entry from the lru cache · d588adae

Filipe Manana authored Jan 11, 2023

In order to replace the open coded name cache in send with the lru cache,
we need an API for the lru cache to delete a specific entry for which we
did a previous lookup. This adds the API for it, and a next patch in the
series will use it.

This patch is part of a larger patchset and the changelog of the last
patch in the series contains a sample performance test and results.
The patches that comprise the patchset are the following:

d588adae

btrfs: allow a generation number to be associated with lru cache entries · 0da0c560

Filipe Manana authored Jan 11, 2023

This allows an optional generation number to be associated to each entry
of the lru cache. Entries with the same key but different generations, are
stored in the linked list to which the maple tree points to. This is meant
to be used when there's a small number of different generations, so the
impact of searching a linked list is negligible. The goal is to get rid of
the open coded name cache in the send code (which uses a radix tree and
a similar linked list of values/entries) and use instead the lru cache
module. For that particular use case we have at most 2 generations that
are associated to each key (inode number): one generation for the send
root and another generation for the parent root. The actual migration of
the send name cache is done in the next patch in the series.

This patch is part of a larger patchset and the changelog of the last
patch in the series contains a sample performance test and results.
The patches that comprise the patchset are the following:

0da0c560

btrfs: send: cache information about created directories · e8a7f49d

Filipe Manana authored Jan 11, 2023

During an incremental send, when processing the reference for an inode
we need to check if the directory where the new reference is located was
already created before creating the new reference. This check, which is
done by the helper did_create_dir(), can be expensive if the directory
has many entries, since it consists in searching the send root's b+tree
and visiting every single dir index key until we either find one which
points to an inode with a number smaller than the current inode's number
or until we visited all index keys. So it doesn't scale well for very
large directories.

So improve on this by caching created directories using a lru cache, and
limiting its size to 64 entries, which results in using at most 4096
bytes of memory. The caching is optional, if we fail to allocate memory,
we just proceed as before and use the existing slower path.

This patch is part of a larger patchset and the changelog of the last
patch in the series contains a sample performance test and results.
The patches that comprise the patchset are the following:

e8a7f49d

btrfs: adapt lru cache to allow for 64 bits keys on 32 bits systems · 6273ee62

Filipe Manana authored Jan 11, 2023

The lru cache is backed by a maple tree, which uses the unsigned long
type for keys, and that type has a width of 32 bits on 32 bits systems
and a width of 64 bits on 64 bits systems.

Currently there is only one user of the lru cache, the send backref cache,
which uses a sector number as a key, a logical address right shifted by
fs_info->sectorsize_bits, so a 32 bits width is not yet a problem (the
same happens with the radix tree we use to track extent buffers,
fs_info->buffer_radix).

However the next patches in the series will start using the lru cache for
cases where inode numbers are the keys, and the inode numbers are always
64 bits, even if we are running on a 32 bits system.

So adapt the lru cache to allow multiple values under the same key, by
having the maple tree store a head entry that points to a list of entries
instead of pointing to a single entry. This is a similar approach to what
we currently do for the name cache in send (which uses a radix tree that
has indexes with an unsigned long type as well), and will allow later to
use the lru cache for the send name cache as well.

This patch is part of a larger patchset and the changelog of the last
patch in the series contains a sample performance test and results.
The patches that comprise the patchset are the following:

6273ee62

btrfs: send: genericize the backref cache to allow it to be reused · 90b90d4a

Filipe Manana authored Jan 11, 2023

The backref cache is a cache backed by a maple tree and a linked list to
keep track of temporal access to cached entries (the LRU entry always at
the head of the list). This type of caching method is going to be useful
in other scenarios, so make the cache implementation more generic and
move it into its own header and source files.

This patch is part of a larger patchset and the changelog of the last
patch in the series contains a sample performance test and results.
The patches that comprise the patchset are the following:

90b90d4a

btrfs: send: initialize all the red black trees earlier · d307d2f3

Filipe Manana authored Jan 11, 2023

After we allocate the send context object and before we initialize all
the red black trees, we can jump to the 'out' label if some errors happen,
and then under the 'out' label we use RB_EMPTY_ROOT() against some of the
those trees, which we have not yet initialized. This happens to work out
ok because the send context object was initialized to zeroes with kzalloc
and the RB_ROOT initializer just happens to have the following definition:

#define RB_ROOT (struct rb_root) { NULL, }

But it's really neither clean nor a good practice as RB_ROOT is supposed
to be opaque and in case it changes or we change those red black trees to
some other data structure, it leaves us in a precarious situation.

So initialize all the red black trees immediately after allocating the
send context and before any jump into the 'out' label.

This patch is part of a larger patchset and the changelog of the last
patch in the series contains a sample performance test and results.
The patches that comprise the patchset are the following:

d307d2f3

btrfs: send: iterate waiting dir move rbtree only once when processing refs · 8c139e1d

Filipe Manana authored Jan 11, 2023

When processing the new references for an inode, we unnecessarily iterate
twice the waiting dir moves rbtree, once with is_waiting_for_move() and
if we found an entry in the rbtree, we iterate it again with a call to
get_waiting_dir_move(). This is pointless, we can make this simpler and
more efficient by calling only get_waiting_dir_move(), so just do that.

This patch is part of a larger patchset and the changelog of the last
patch in the series contains a sample performance test and results.
The patches that comprise the patchset are the following:

8c139e1d

btrfs: send: reduce searches on parent root when checking if dir can be removed · 474e4761

Filipe Manana authored Jan 11, 2023

During an incremental send, every time we remove a reference (dentry) for
an inode and the parent directory does not exists anymore in the send
root, we go check if we can remove the directory by making a call to
can_rmdir(). This helper can only return true (value 1) if all dentries
were already removed, and for that it always does a search on the parent
root for dir index keys - if it finds any dentry referring to an inode
with a number higher then the inode currently being processed, then the
directory can not be removed and it must return false (value 0).

However that means if a directory that was deleted had 1000 dentries, and
each one pointed to an inode with a number higher then the number of the
directory's inode, we end up doing 1000 searches on the parent root.
Typically files are created in a directory after the directory was created
and therefore they get an higher inode number than the directory. It's
also common to have the each dentry pointing to an inode with a higher
number then the inodes the previous dentries point to, for example when
creating a series of files inside a directory, a very common pattern.

So improve on that by having the first call to can_rmdir() for a directory
to check the number of the inode that the last dentry points to and cache
that inode number in the orphan dir structure. Then every subsequent call
to can_rmdir() can avoid doing a search on the parent root if the number
of the inode currently being processed is smaller than cached inode number
at the directory's orphan dir structure.

This patch is part of a larger patchset and the changelog of the last
patch in the series contains a sample performance test and results.
The patches that comprise the patchset are the following:

474e4761

btrfs: send: avoid unnecessary orphan dir rbtree search at can_rmdir() · 78cf1a95

Filipe Manana authored Jan 11, 2023

At can_rmdir() we start by searching the orphan dirs rbtree for an orphan
dir object for the target directory. Later when iterating over the dir
index keys, if we find that any dir entry points to inode for which there
is a pending dir move or the inode was not yet processed, we exit because
we can't remove the directory yet. However we end up always calling
add_orphan_dir_info(), which will iterate again the rbtree and if there is
already an orphan dir object (created by the first call to can_rmdir()),
it returns the existing object. This is unnecessary work because in case
there is already an existing orphan dir object, we got a reference to it
at the start of can_rmdir(). So skip the call to add_orphan_dir_info()
if we already have a reference for an orphan dir object.

This patch is part of a larger patchset and the changelog of the last
patch in the series contains a sample performance test and results.
The patches that comprise the patchset are the following:

78cf1a95

btrfs: send: avoid duplicated orphan dir allocation and initialization · d921b9cf

Filipe Manana authored Jan 11, 2023

At can_rmdir() we are allocating and initializing an orphan dir object
twice. This can be deduplicated outside of the loop that iterates over
the dir index keys. So deduplicate that code, even because other patch
in the series will need to add more initialization code and another one
will add one more condition.

This patch is part of a larger patchset and the changelog of the last
patch in the series contains a sample performance test and results.
The patches that comprise the patchset are the following:

d921b9cf

btrfs: send: remove send_progress argument from can_rmdir() · 24970ccb

Filipe Manana authored Jan 11, 2023

All callers of can_rmdir() pass sctx->cur_ino as the value for the
send_progress argument, so remove the argument and directly use
sctx->cur_ino.

This patch is part of a larger patchset and the changelog of the last
patch in the series contains a sample performance test and results.
The patches that comprise the patchset are the following:

24970ccb

btrfs: send: avoid extra b+tree searches when checking reference overrides · 498581f3

Filipe Manana authored Jan 11, 2023

During an incremental send, when processing the new references of an inode
(either it's a new inode or an existing one renamed/moved), he will search
the b+tree of the send or parent roots in order to find out the inode item
of the parent directory and extract its generation. However we are doing
that search twice, once with is_inode_existent() -> get_cur_inode_state()
and then again at did_overwrite_ref() or will_overwrite_ref().

So avoid that and get the generation at get_cur_inode_state() and then
propagate it up to did_overwrite_ref() and will_overwrite_ref().

This patch is part of a larger patchset and the changelog of the last
patch in the series contains a sample performance test and results.
The patches that comprise the patchset are the following:

498581f3

btrfs: send: directly return from will_overwrite_ref() and simplify it · b3047a42

Filipe Manana authored Jan 11, 2023

There are no resources to release before will_overwrite_ref() returns, so
we don't really need the 'out' label and jumping to it when conditions are
met - we can directly return and get rid of the label and jumps. Also we
can deal with -ENOENT and other errors in a single if-else logic, as it's
more straightforward.

This helps the next patch in the series to be more simple as well.

This patch is part of a larger patchset and the changelog of the last
patch in the series contains a sample performance test and results.
The patches that comprise the patchset are the following:

b3047a42

btrfs: send: avoid unnecessary generation search at did_overwrite_ref() · cb689481

Filipe Manana authored Jan 11, 2023

At did_overwrite_ref() we always call get_inode_gen() to find out the
generation of the inode 'ow_inode'. However we don't always need to use
that generation, and in fact it's very common to not use it, so we end
up doing a b+tree search on the send root, allocating a path, etc, for
nothing. So improve on this by getting the generation only if we need
to use it.

This patch is part of a larger patchset and the changelog of the last
patch in the series contains a sample performance test and results.
The patches that comprise the patchset are the following:

cb689481

btrfs: send: directly return from did_overwrite_ref() and simplify it · e739ba30

Filipe Manana authored Jan 11, 2023

There are no resources to release before did_overwrite_ref() returns, so
we don't really need the 'out' label and jumping to it when conditions are
met - we can directly return and get rid of the label and jumps. Also we
can deal with -ENOENT and other errors in a single if-else logic, as it's
more straightforward.

This helps the next patch in the series to be more simple as well.

This patch is part of a larger patchset and the changelog of the last
patch in the series contains a sample performance test and results.
The patches that comprise the patchset are the following:

e739ba30

btrfs: sysfs: update fs features directory asynchronously · b7625f46

Qu Wenruo authored Jan 13, 2023

[BUG]
Since the introduction of per-fs feature sysfs interface
(/sys/fs/btrfs/<UUID>/features/), the content of that directory is never
updated.

Thus for the following case, that directory will not show the new
features like RAID56:

  # mkfs.btrfs -f $dev1 $dev2 $dev3
  # mount $dev1 $mnt
  # btrfs balance start -f -mconvert=raid5 $mnt
  # ls /sys/fs/btrfs/$uuid/features/
  extended_iref  free_space_tree  no_holes  skinny_metadata

While after unmount and mount, we got the correct features:

  # umount $mnt
  # mount $dev1 $mnt
  # ls /sys/fs/btrfs/$uuid/features/
  extended_iref  free_space_tree  no_holes  raid56 skinny_metadata

[CAUSE]
Because we never really try to update the content of per-fs features/
directory.

We had an attempt to update the features directory dynamically in commit
14e46e04 ("btrfs: synchronize incompat feature bits with sysfs
files"), but unfortunately it get reverted in commit e410e34f
("Revert "btrfs: synchronize incompat feature bits with sysfs files"").
The problem in the original patch is, in the context of
btrfs_create_chunk(), we can not afford to update the sysfs group.

The exported but never utilized function, btrfs_sysfs_feature_update()
is the leftover of such attempt.  As even if we go sysfs_update_group(),
new files will need extra memory allocation, and we have no way to
specify the sysfs update to go GFP_NOFS.

[FIX]
This patch will address the old problem by doing asynchronous sysfs
update in the cleaner thread.

This involves the following changes:

- Make __btrfs_(set|clear)_fs_(incompat|compat_ro) helpers to set
  BTRFS_FS_FEATURE_CHANGED flag when needed

- Update btrfs_sysfs_feature_update() to use sysfs_update_group()
  And drop unnecessary arguments.

- Call btrfs_sysfs_feature_update() in cleaner_kthread
  If we have the BTRFS_FS_FEATURE_CHANGED flag set.

- Wake up cleaner_kthread in btrfs_commit_transaction if we have
  BTRFS_FS_FEATURE_CHANGED flag

By this, all the previously dangerous call sites like
btrfs_create_chunk() need no new changes, as above helpers would
have already set the BTRFS_FS_FEATURE_CHANGED flag.

The real work happens at cleaner_kthread, thus we pay the cost of
delaying the update to sysfs directory, but the delayed time should be
small enough that end user can not distinguish though it might get
delayed if the cleaner thread is busy with removing subvolumes or
defrag.

CC: stable@vger.kernel.org # 4.14+
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>

b7625f46

btrfs: remove duplicate include header in extent-tree.c · 58e36c2a

ye xingchen authored Jan 16, 2023

extent-tree.h is included more than once, added in a0231804 ("btrfs:
move extent-tree helpers into their own header file").
Signed-off-by: ye xingchen <ye.xingchen@zte.com.cn>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>

58e36c2a

btrfs: scrub: improve tree block error reporting · 28232909

Qu Wenruo authored Dec 17, 2022

[BUG]
When debugging a scrub related metadata error, it turns out that our
metadata error reporting is not ideal.

The only 3 error messages are:

- BTRFS error (device dm-2): bdev /dev/mapper/test-scratch1 errs: wr 0, rd 0, flush 0, corrupt 0, gen 1
Showing we have metadata generation mismatch errors.

- BTRFS error (device dm-2): unable to fixup (regular) error at logical 7110656 on dev /dev/mapper/test-scratch1
Showing which tree blocks are corrupted.

- BTRFS warning (device dm-2): checksum/header error at logical 24772608 on dev /dev/mapper/test-scratch2, physical 3801088: metadata node (level 1) in tree 5
Showing which physical range the corrupted metadata is at.

We have to combine the above 3 to know we have a corrupted metadata with
generation mismatch.

And this is already the better case, if we have other problems, like
fsid mismatch, we can not even know the cause.

[CAUSE]
The problem is caused by the fact that, scrub_checksum_tree_block()
never outputs any error message.

It just return two bits for scrub: sblock->header_error, and
sblock->generation_error.

And later we report error in scrub_print_warning(), but unfortunately we
only have two bits, there is not really much thing we can done to print
any detailed errors.

[FIX]
This patch will do the following to enhance the error reporting of
metadata scrub:

- Add extra warning (ratelimited) for every error we hit
This can help us to distinguish the different types of errors.
Some errors can help us to know what's going wrong immediately,
like bytenr mismatch.

- Re-order the checks
Currently we check bytenr first, then immediately generation.
This can lead to false generation mismatch reports, while the fsid
mismatches.

Here is the new output for the bug I'm debugging (we forgot to
writeback tree blocks for commit roots):

BTRFS warning (device dm-2): tree block 24117248 mirror 1 has bad fsid, has b77cd862-f150-4c71-90ec-7baf0544d83f want 17df6abf-23cd-445f-b350-5b3e40bfd2fc
BTRFS warning (device dm-2): tree block 24117248 mirror 0 has bad fsid, has b77cd862-f150-4c71-90ec-7baf0544d83f want 17df6abf-23cd-445f-b350-5b3e40bfd2fc

Now we can immediately know it's some tree blocks didn't even get written
back, other than the original confusing generation mismatch.
Signed-off-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>

28232909

btrfs: don't use size classes for zoned file systems · cb0922f2

Boris Burkov authored Dec 15, 2022

When a file system has ZNS devices which are constrained by a maximum
number of active block groups, then not being able to use all the block
groups for every allocation is not ideal, and could cause us to loop a
ton with mixed size allocations.

In general, since zoned doesn't write into gaps behind where block
groups are writing, it is not susceptible to the same sort of
fragmentation that size classes are designed to solve, so we can skip
size classes for zoned file systems in general, even though there would
probably be no harm for SMR devices.
Signed-off-by: Boris Burkov <boris@bur.io>
Signed-off-by: David Sterba <dsterba@suse.com>

cb0922f2

btrfs: load block group size class when caching · c7eec3d9

Boris Burkov authored Dec 15, 2022

Since the size class is an artifact of an arbitrary anti fragmentation
strategy, it doesn't really make sense to persist it. Furthermore, most
of the size class logic assumes fresh block groups. That is of course
not a reasonable assumption -- we will be upgrading kernels with
existing filesystems whose block groups are not classified.

To work around those issues, implement logic to compute the size class
of the block groups as we cache them in. To perfectly assess the state
of a block group, we would have to read the entire extent tree (since
the free space cache mashes together contiguous extent items) which
would be prohibitively expensive for larger file systems with more
extents.

We can do it relatively cheaply by implementing a simple heuristic of
sampling a handful of extents and picking the smallest one we see. In
the happy case where the block group was classified, we will only see
extents of the correct size. In the unhappy case, we will hopefully find
one of the smaller extents, but there is no perfect answer anyway.
Autorelocation will eventually churn up the block group if there is
significant freeing anyway.

There was no regression in mount performance at end state of the fsperf
test suite, and the delay until the block group is marked cached is
minimized by the constant number of extent samples.
Signed-off-by: Boris Burkov <boris@bur.io>
Signed-off-by: David Sterba <dsterba@suse.com>

c7eec3d9

btrfs: introduce size class to block group allocator · 52bb7a21

Boris Burkov authored Dec 15, 2022

The aim of this patch is to reduce the fragmentation of block groups
under certain unhappy workloads. It is particularly effective when the
size of extents correlates with their lifetime, which is something we
have observed causing fragmentation in the fleet at Meta.

This patch categorizes extents into size classes:

- x < 128KiB: "small"
- 128KiB < x < 8MiB: "medium"
- x > 8MiB: "large"

and as much as possible reduces allocations of extents into block groups
that don't match the size class. This takes advantage of any (possible)
correlation between size and lifetime and also leaves behind predictable
re-usable gaps when extents are freed; small writes don't gum up bigger
holes.

Size classes are implemented in the following way:

- Mark each new block group with a size class of the first allocation
  that goes into it.

- Add two new passes to ffe: "unset size class" and "wrong size class".
  First, try only matching block groups, then try unset ones, then allow
  allocation of new ones, and finally allow mismatched block groups.

- Filtering is done just by skipping inappropriate ones, there is no
  special size class indexing.

Other solutions I considered were:

- A best fit allocator with an rb-tree. This worked well, as small
  writes didn't leak big holes from large freed extents, but led to
  regressions in ffe and write performance due to lock contention on
  the rb-tree with every allocation possibly updating it in parallel.
  Perhaps something clever could be done to do the updates in the
  background while being "right enough".

- A fixed size "working set". This prevents freeing an extent
  drastically changing where writes currently land, and seems like a
  good option too. Doesn't take advantage of size in any way.

- The same size class idea, but implemented with xarray marks. This
  turned out to be slower than looping the linked list and skipping
  wrong block groups, and is also less flexible since we must have only
  3 size classes (max #marks). With the current approach we can have as
  many as we like.

Performance testing was done via: https://github.com/josefbacik/fsperf
Of particular relevance are the new fragmentation specific tests.

A brief summary of the testing results:

- Neutral results on existing tests. There are some minor regressions
  and improvements here and there, but nothing that truly stands out as
  notable.
- Improvement on new tests where size class and extent lifetime are
  correlated. Fragmentation in these cases is completely eliminated
  and write performance is generally a little better. There is also
  significant improvement where extent sizes are just a bit larger than
  the size class boundaries.
- Regression on one new tests: where the allocations are sized
  intentionally a hair under the borders of the size classes. Results
  are neutral on the test that intentionally attacks this new scheme by
  mixing extent size and lifetime.

The full dump of the performance results can be found here:
https://bur.io/fsperf/size-class-2022-11-15.txt
(there are ANSI escape codes, so best to curl and view in terminal)

Here is a snippet from the full results for a new test which mixes
buffered writes appending to a long lived set of files and large short
lived fallocates:

bufferedappendvsfallocate results
         metric             baseline       current        stdev            diff
======================================================================================
avg_commit_ms                    31.13         29.20          2.67     -6.22%
bg_count                            14         15.60             0     11.43%
commits                          11.10         12.20          0.32      9.91%
elapsed                          27.30         26.40          2.98     -3.30%
end_state_mount_ns         11122551.90   10635118.90     851143.04     -4.38%
end_state_umount_ns           1.36e+09      1.35e+09   12248056.65     -1.07%
find_free_extent_calls       116244.30     114354.30        964.56     -1.63%
find_free_extent_ns_max      599507.20    1047168.20     103337.08     74.67%
find_free_extent_ns_mean       3607.19       3672.11        101.20      1.80%
find_free_extent_ns_min            500           512          6.67      2.40%
find_free_extent_ns_p50           2848          2876         37.65      0.98%
find_free_extent_ns_p95           4916          5000         75.45      1.71%
find_free_extent_ns_p99       20734.49      20920.48       1670.93      0.90%
frag_pct_max                     61.67             0          8.05   -100.00%
frag_pct_mean                    43.59             0          6.10   -100.00%
frag_pct_min                     25.91             0         16.60   -100.00%
frag_pct_p50                     42.53             0          7.25   -100.00%
frag_pct_p95                     61.67             0          8.05   -100.00%
frag_pct_p99                     61.67             0          8.05   -100.00%
fragmented_bg_count               6.10             0          1.45   -100.00%
max_commit_ms                    49.80            46          5.37     -7.63%
sys_cpu                           2.59          2.62          0.29      1.39%
write_bw_bytes                1.62e+08      1.68e+08   17975843.50      3.23%
write_clat_ns_mean            57426.39      54475.95       2292.72     -5.14%
write_clat_ns_p50             46950.40      42905.60       2101.35     -8.62%
write_clat_ns_p99            148070.40     143769.60       2115.17     -2.90%
write_io_kbytes                4194304       4194304             0      0.00%
write_iops                     2476.15       2556.10        274.29      3.23%
write_lat_ns_max            2101667.60    2251129.50     370556.59      7.11%
write_lat_ns_mean             59374.91      55682.00       2523.09     -6.22%
write_lat_ns_min              17353.10         16250       1646.08     -6.36%

There are some mixed improvements/regressions in most metrics along with
an elimination of fragmentation in this workload.

On the balance, the drastic 1->0 improvement in the happy cases seems
worth the mix of regressions and improvements we do observe.

Some considerations for future work:

- Experimenting with more size classes
- More hinting/search ordering work to approximate a best-fit allocator
Signed-off-by: Boris Burkov <boris@bur.io>
Signed-off-by: David Sterba <dsterba@suse.com>

52bb7a21

btrfs: add more find_free_extent tracepoints · 854c2f36

Boris Burkov authored Dec 15, 2022

find_free_extent is a complicated function. It consists (at least) of:

- a hint that jumps into the middle of a for loop macro
- a middle loop trying every raid level
- an outer loop ascending through ffe loop levels
- complicated logic for skipping some of those ffe loop levels
- multiple underlying in-bg allocators (zoned, cluster, no cluster)

Which is all to say that more tracing is helpful for debugging its
behavior. Add two new tracepoints: at the entrance to the block_groups
loop (hit for every raid level and every ffe_ctl loop) and at the point
we seriously consider a block_group for allocation. This way we can see
the whole path through the algorithm, including hints, multiple loops,
etc.
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Boris Burkov <boris@bur.io>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>

854c2f36

btrfs: pass find_free_extent_ctl to allocator tracepoints · cfc2de0f

Boris Burkov authored Dec 15, 2022

The allocator tracepoints currently have a pile of values from ffe_ctl.
In modifying the allocator and adding more tracepoints, I found myself
adding to the already long argument list of the tracepoints. It makes it
a lot simpler to just send in the ffe_ctl itself.
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Reviewed-by: Anand Jain <anand.jain@oracle.com>
Signed-off-by: Boris Burkov <boris@bur.io>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>

cfc2de0f

btrfs: remove the wait argument to btrfs_start_ordered_extent · 36d45567

Christoph Hellwig authored Dec 12, 2022

Given that wait is always set to 1, so remove the argument.
Last use of wait with 0 was in 0c304304 ("Btrfs: remove
csum_bytes_left").
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>

36d45567

btrfs: use a single variable to track return value for log_dir_items() · 235e1c7b

Filipe Manana authored Jan 10, 2023

We currently use 'ret' and 'err' to track the return value for
log_dir_items(), which is confusing and likely the cause for previous
bugs where log_dir_items() did not return an error when it should, fixed
in previous patches.

So change this and use only a single variable, 'ret', to track the return
value. This is simpler and makes it similar to most of the existing code.
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>

235e1c7b

btrfs: use a negative value for BTRFS_LOG_FORCE_COMMIT · 5cce1780

Filipe Manana authored Jan 10, 2023

Currently we use the value 1 for BTRFS_LOG_FORCE_COMMIT, but that value
has a few inconveniences:

1) If it's ever used by btrfs_log_inode(), or any function down the call
   chain, we have to remember to btrfs_set_log_full_commit(), which is
   repetitive and has a chance to be forgotten in future use cases.
   btrfs_log_inode_parent() only calls btrfs_set_log_full_commit() when
   it gets a negative value from btrfs_log_inode();

2) Down the call chain of btrfs_log_inode(), we may have functions that
   need to force a log commit, but can return either an error (negative
   value), false (0) or true (1). So they are forced to return some
   random negative to force a log commit - using BTRFS_LOG_FORCE_COMMIT
   would make the intention more clear. Currently the only example is
   flush_dir_items_batch().

So turn BTRFS_LOG_FORCE_COMMIT into a negative value. The chosen value
is -(MAX_ERRNO + 1), so that it does not overlap any errno value and makes
it easier to debug.
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>

5cce1780

btrfs: use PAGE_{ALIGN, ALIGNED, ALIGN_DOWN} macro · ce394a7f

Yushan Zhou authored Jan 03, 2023

The header file linux/mm.h provides PAGE_ALIGN, PAGE_ALIGNED,
PAGE_ALIGN_DOWN macros. Use these macros to make code more
concise.
Signed-off-by: Yushan Zhou <katrinzhou@tencent.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>

ce394a7f

btrfs: go to matching label when cleaning em in btrfs_submit_direct · d31de378

Peng Hao authored Jan 09, 2023

When btrfs_get_chunk_map fails to allocate a new em the cleanup does not
need to be done so the goto target is out_err, which is consistent with
current coding style.
Signed-off-by: Peng Hao <flyingpeng@tencent.com>
Reviewed-by: David Sterba <dsterba@suse.com>
[ update changelog ]
Signed-off-by: David Sterba <dsterba@suse.com>

d31de378

btrfs: turn on -Wmaybe-uninitialized · 1ec49744

Josef Bacik authored Dec 16, 2022

We had a recent bug that would have been caught by a newer compiler with
-Wmaybe-uninitialized and would have saved us a month of failing tests
that I didn't have time to investigate.
Reviewed-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>

1ec49744

btrfs: fix uninitialized variable warning in run_one_async_start · a6ca692e

Josef Bacik authored Dec 16, 2022

With -Wmaybe-uninitialized compiler complains about ret being possibly
uninitialized, which isn't possible as the WQ_ constants are set only
from our code, however we can handle the default case and get rid of the
warning.

The value is set to BLK_STS_IOERR so it does not issue any IO and could
be potentially detected, but this is basically a "cannot happen" error.
To catch any problems during development use the assert.
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
[ set the error in default: ]
Signed-off-by: David Sterba <dsterba@suse.com>

a6ca692e

btrfs: zoned: fix uninitialized variable warning in btrfs_get_dev_zones · cd30d3bc

Naohiro Aota authored Dec 21, 2022

Fix an uninitialized warning we get with -Wmaybe-uninitialized where it
thought zno may have been uninitialized, in both cases it depends on
zinfo->zone_cache but we know the value won't change between checks.
Reported-by: Josef Bacik <josef@toxicpanda.com>
Link: https://lore.kernel.org/linux-btrfs/af6c527cbd8bdc782e50bd33996ee83acc3a16fb.1671221596.git.josef@toxicpanda.com/Signed-off-by: Naohiro Aota <naohiro.aota@wdc.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>

cd30d3bc

btrfs: fix uninitialized variable warning in btrfs_sb_log_location · 12adffe6

Josef Bacik authored Dec 16, 2022

We only have 3 possible mirrors, and we have ASSERT()'s to make sure
we're not passing in an invalid super mirror into this function, so
technically this value isn't uninitialized.  However
-Wmaybe-uninitialized will complain, so set it to U64_MAX so if we don't
have ASSERT()'s turned on it'll error out later on when it see's the
zone is beyond our maximum zones.
Reviewed-by: Naohiro Aota <naohiro.aota@wdc.com>
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>

12adffe6

btrfs: fix uninitialized variable warnings in __set_extent_bit and convert_extent_bit · 59864325

Josef Bacik authored Dec 16, 2022

We will pass in the parent and p pointer into our tree_search function
to avoid doing a second search when inserting a new extent state into
the tree.  However because this is conditional upon passing in these
pointers the compiler seems to think these values can be uninitialized
if we're using -Wmaybe-uninitialized.  Fix this by initializing these
values.
Reviewed-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>

59864325

btrfs: fix uninitialized variable warning in btrfs_update_block_group · efbf35a1

Josef Bacik authored Dec 16, 2022

reclaim isn't set in the alloc case, however we only care about
reclaim in the !alloc case.  This isn't an actual problem, however
-Wmaybe-uninitialized will complain, so initialize reclaim to quiet the
compiler.
Reviewed-by: Qu Wenruo <wqu@suse.com>
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>

efbf35a1

btrfs: fix uninitialized variable warning in get_inode_gen · ab199013

Josef Bacik authored Dec 16, 2022

Anybody that calls get_inode_gen() can have an uninitialized gen if
there's an error.  This isn't a big deal because all the users just exit
if they get an error, however it makes -Wmaybe-uninitialized complain,
so fix this up to always initialize the passed in gen, this quiets all
of the uninitialized warnings in send.c.
Reviewed-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>

ab199013

btrfs: fix uninitialized variable warning in btrfs_cleanup_ordered_extents · 0e47b25c

Josef Bacik authored Dec 16, 2022

We can conditionally pass in a locked page, and then we'll use that page
range to skip marking errors as that will happen in another layer.
However this causes the compiler to complain because it doesn't
understand we only use these values when we have the page.  Make the
compiler stop complaining by setting these values to 0.
Reviewed-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Reviewed-by: Qu Wenruo <wqu@suse.com>
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>

0e47b25c

btrfs: move btrfs_abort_transaction to transaction.c · fccf0c84

Josef Bacik authored Dec 07, 2022

While trying to sync messages.[ch] I ended up with this dependency on
messages.h in the rest of btrfs-progs code base because it's where
btrfs_abort_transaction() was now held.  We want to keep messages.[ch]
limited to the kernel code, and the btrfs_abort_transaction() code
better fits in the transaction code and not in messages.
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
[ move the __cold attributes ]
Signed-off-by: David Sterba <dsterba@suse.com>

fccf0c84