Filipe Manana [Tue, 1 Nov 2022 16:15:46 +0000 (16:15 +0000)]
btrfs: use a single argument for extent offset in backref walking functions
The interface for find_parent_nodes() has two extent offset related
arguments:
1) One u64 pointer argument for the extent offset;
2) One boolean argument to tell if the extent offset should be ignored or
not.
These are confusing, becase the extent offset pointer can be NULL and in
some cases callers pass a NULL value as a way to tell the backref walking
code to ignore offsets in file extent items (and simply consider all file
extent items that point to the target data extent).
The boolean argument was added in commit c995ab3cda3f ("btrfs: add a flag
to iterate_inodes_from_logical to find all extent refs for uncompressed
extents"), but it was never really necessary, it was enough if it could
find a way to get a NULL value passed to the "extent_item_pos" argument of
find_parent_nodes(). The arguments are also passed to functions called
by find_parent_nodes() and respective helper functions, which further
makes everything more complicated than needed.
Then we have several backref walking related functions that end up calling
find_parent_nodes(), either directly or through some other function that
they call, and for many we have to use an "extent_item_pos" (u64) argument
and a boolean "ignore_offset" argument too.
This is confusing and not really necessary. So use a single argument to
specify the extent offset, as a simple u64 and not as a pointer, but
using a special value of (u64)-1, defined as a documented constant, to
indicate when the extent offset should be ignored.
This is also preparation work for the upcoming patches in the series that
add other arguments to find_parent_nodes() and other related functions
that use it.
Filipe Manana [Tue, 1 Nov 2022 16:15:45 +0000 (16:15 +0000)]
btrfs: send: optimize clone detection to increase extent sharing
Currently send does not do the best decisions when it comes to decide
between multiple clone sources, which results in clone operations for
partial extent ranges, which has the following disadvantages:
1) We get less shared extents at the destination;
2) We have to read more data during the send operation and emit more
write commands.
Besides not being optimal behaviour, it also breaks user expectations and
is often reported by users, with a recent example in the Link tag at the
bottom of this change log.
Part of the reason for this non-optimal behaviour is that the backref
walking code does not provide information about the length of the file
extent items that were found for each backref, so send is blind about
which backref is the best to chose as a cloning source.
The other existing reasons are just silliness, namely always prefering
the inode with the lowest number when multiple are found for the same
root and when we can clone from multiple roots, always prefer the send
root over any of the other clone roots. This does not make any sense
since any inode or root is fine and as good as any other inode/root.
Fix this by making backref walking pass information about the number of
bytes referenced by each file extent item and then have send's backref
callback pick the inode with the highest number of bytes for each root.
Finally select the root from which we can clone more bytes from.
Total Exclusive Set shared Filename
2.00MiB 1.00MiB - /mnt/sdi/foo
2.00MiB 0.00B - /mnt/sdi/bar
2.00MiB 0.00B - /mnt/sdi/baz
6.00MiB 1.00MiB 2.00MiB /mnt/sdi
Create a readonly snapshot of '/mnt/sdi' in '/mnt/sdi/snap'
At subvol /mnt/sdi/snap
At subvol snap
Now there's a much better sharing, files bar and baz share 1M of the
extent of file foo and the second extent of files bar and baz is shared
between themselves.
This will later be turned into a test case for fstests.
Filipe Manana [Tue, 1 Nov 2022 16:15:44 +0000 (16:15 +0000)]
btrfs: send: avoid unnecessary backref lookups when finding clone source
At find_extent_clone(), unless we are given an inline extent, a file
extent item that represents hole or an extent that starts beyond the
i_size, we always do backref walking to look for clone sources, unless
if we have more than SEND_MAX_EXTENT_REFS (64) known references on the
extent.
However if we know we only have one reference in the extent item and only
one clone source (the send root), then it's pointless to do the backref
walking to search for clone sources, as we can't clone from any other
root. So skip the backref walking in that case.
The following test was run on a non-debug kernel (Debian's default kernel
config):
$ cat test.sh
#!/bin/bash
DEV=/dev/sdi
MNT=/mnt/sdi
mkfs.btrfs -f $DEV
mount $DEV $MNT
# Create an extent tree that's not too small and none of the
# extents is shared.
for ((i = 1; i <= 50000; i++)); do
xfs_io -f -c "pwrite 0 4K" $MNT/file_$i > /dev/null
echo -ne "\r$i files created..."
done
echo
Filipe Manana [Tue, 1 Nov 2022 16:15:43 +0000 (16:15 +0000)]
btrfs: send: drop unnecessary backref context field initializations
At find_extent_clone() we are initializing to zero the 'found_itself' and
'found' fields of the backref context before we use it but we have already
initialized the structure to zeroes when we declared it on stack, so it's
pointless to initialize those fields and they are unnecessarily increasing
the object text size with two "mov" instructions (x86_64).
Similarly make the 'extent_len' initialization more clear by using an if-
-then-else instead of a double assignment to it in case the extent's end
crosses the i_size boundary.
Before this change:
$ size fs/btrfs/send.o
text data bss dec hex filename
68694 4252 16 72962 11d02 fs/btrfs/send.o
After this change:
$ size fs/btrfs/send.o
text data bss dec hex filename
68678 4252 16 72946 11cf2 fs/btrfs/send.o
Filipe Manana [Tue, 1 Nov 2022 16:15:42 +0000 (16:15 +0000)]
btrfs: send: update comment at find_extent_clone()
We have this unclear comment at find_extent_clone() about extents starting
at a file offset greater than or equals to the i_size of the inode. It's
not really informative and it's misleading, since it mentions the author
found such extents with snapshots and large files.
Such extents are a result of fallocate with FALLOC_FL_KEEP_SIZE and there
is no relation to snapshots or large files (all write paths update the
i_size before inserting a new file extent item). So update the comment to
be precise about it and why we don't bother looking for clone sources in
that case.
Filipe Manana [Tue, 1 Nov 2022 16:15:41 +0000 (16:15 +0000)]
btrfs: send: avoid unnecessary path allocations when finding extent clone
When looking for an extent clone, at find_extent_clone(), we start by
allocating a path and then check for cases where we can't have clones
and exit immediately in those cases. It's a waste of time to allocate
the path before those cases, so reorder the logic so that we check for
those cases before allocating the path.
Qu Wenruo [Tue, 1 Nov 2022 11:16:12 +0000 (19:16 +0800)]
btrfs: remove the unused endio_raid56_workers and btrfs_raid_bio::end_io_work
Since we have switched all raid56 workload to submit-and-wait method,
there is no use for btrfs_fs_info::endio_raid56_workers workqueue and
btrfs_raid_bio::end_io_work.
Qu Wenruo [Tue, 1 Nov 2022 11:16:10 +0000 (19:16 +0800)]
btrfs: raid56: extract scrub read bio list assembly code into a helper
Just like what we did for write/recovery, also extract the read bio
assembly code into a helper for scrub.
The difference between the three are:
- rmw_assemble_read_bios() only submit reads for missing sectors
Thus it will skip cached sectors, but will also read sectors which
is not covered by any full stripe. (For cache usage)
- recover_assemble_read_bios() reads every sector which has not failed
- scrub_assemble_read_bios() has extra check for vertical stripes
It's mostly the same as rmw_assemble_read_bios(), but will skip
sectors which is not covered by a vertical stripe.
Qu Wenruo [Tue, 1 Nov 2022 11:16:09 +0000 (19:16 +0800)]
btrfs: raid56: switch write path to rmw_rbio()
This includes the following changes:
- Implement new raid_unplug() functions
Now we don't need a workqueue to run the plug, as all our
work is just queue rmw_rbio_work() call, which can be executed
without sleep.
- Implement a rmw_rbio_work_locked() helper
This is for unlock_stripe(), which is already holding the full stripe
lock.
- Remove all the old functions
This should already shows how complex the old functions are, as we
ended up removing the following functions:
Qu Wenruo [Tue, 1 Nov 2022 11:16:08 +0000 (19:16 +0800)]
btrfs: raid56: introduce the main entrance for RMW path
The new entrance will be called rmw_rbio(), it will have a streamlined
workflow by using submit-and-wait method.
Thus there will be no weird jumps between tons of functions, thus way
more reader friendly, and will make later expansion easier, as it's now
a straight workflow, the timing is way more clear.
Unfortunately we can not yet migrate the RMW path to use this new
entrance as we still need extra work to address the plug and
unlock_stripe() function.
Qu Wenruo [Tue, 1 Nov 2022 11:16:05 +0000 (19:16 +0800)]
btrfs: raid56: switch recovery path to a single function
Currently btrfs uses end_io functions to jump between different stages
of recovery.
For example, we go the following different functions:
- raid56_bio_end_io()
This handles the read for all the sectors (except the missing device).
- __raid_recover_end_io()
This does the real work, it's called inside the delayed work function
raid_recover_end_io_work().
This one recovery path involves at least 3 different functions, which is
a big burden for readers.
This patch will change the behavior by:
- Introduce a unified recovery entrance, recover_rbio()
- Use submit-and-wait method
So the workflow is not interrupted by the endio function jump.
This doesn't bring performance change, but reduce the burden for
reviewers.
- Run the main function in the rmw_workers workqueue
Now raid56_parity_recover() only needs to setup the work, and
queue the work using start_async_work().
Now readers only need to do one function jump (start_async_work()) to
find out the main entrance of recovery path.
Furthermore, recover_rbio() function can easily be reused by other paths.
The old recovery path is still utilized by degraded write path.
It will be cleaned up when we have migrated the write path.
Qu Wenruo [Tue, 1 Nov 2022 11:16:01 +0000 (19:16 +0800)]
btrfs: raid56: extract the vertical stripe recovery code into recover_vertical()
This refactor includes the following behavior change first:
- Don't error out if only P/Q is corrupted
The old code will directly error out if only P/Q is corrupted.
Although it is an logical error if we go into rebuild path with
only P/Q corrupted, there is no need to error out.
Just skip the rebuild and return the already good data.
Then comes the following refactor which shouldn't cause behavior
changes:
- Introduce a helper to do vertical stripe recovery
This not only reduce one indent level, but also paves the road for
later data checksum verification in RMW cycles.
- Sort rbio->faila/b before recovery
So we don't need to do the same swap every vertical stripe
- Replace a BUG_ON() with ASSERT()
Or checkpatch won't let me pass.
- Mark recovered sectors uptodate after the recover loop
- Do the cleanup for pointers unconditionally
We only need to initialize @pointers and @unmap_array to NULL, so
we can safely free them unconditionally.
- Mark the repaired sector uptodate in recover_vertical()
David Sterba [Thu, 27 Oct 2022 11:07:05 +0000 (13:07 +0200)]
btrfs: merge struct extent_page_data to btrfs_bio_ctrl
The two structures appear on the same call paths, btrfs_bio_ctrl is
embedded in extent_page_data and we pass bio_ctrl to some functions.
After merging there are fewer indirections and we have only one control
structure. The packing remains same.
The btrfs_bio_ctrl was selected as the target structure as the operation
is closer to bio processing.
David Sterba [Thu, 27 Oct 2022 11:07:03 +0000 (13:07 +0200)]
btrfs: switch extent_page_data bit fields to bools
The semantics of the two members is a boolean, so change the type
accordingly. We have space in extent_page_data due to alignment there's
no change in size.
The div_factor* helpers calculate fraction or percentage fraction. The
name is a bit confusing, we use it only for percentage calculations and
there are two helpers.
There's a helper mult_frac that's for general fractions, that tries to
be accurate but we multiply and divide by small numbers so we can use
the div_u64 helper.
Rename the div_factor* helpers and use 1..100 percentage range, also drop
the case checking for percentage == 100, it's never hit.
The conversions:
* div_factor calculates tenths and the numbers need to be adjusted
* div_factor_fine is direct replacement
Filipe Manana [Mon, 31 Oct 2022 11:43:56 +0000 (11:43 +0000)]
btrfs: update stale comment for nowait direct IO writes
If when doing a direct IO write we need to fallback to buffered IO, we
this comment at btrfs_direct_write() that says we can't directly fallback
to buffered IO if we have a NOWAIT iocb, because we have no support for
NOWAIT buffered writes. That is not true anymore, as support for NOWAIT
buffered writes was added recently in commit 926078b21db9 ("btrfs: enable
nowait async buffered writes").
However we still can't fallback to a buffered write in case we have a
NOWAIT iocb, because we'll need to flush delalloc and wait for it to
complete after doing the buffered write, and that can block for several
reasons, the main reason being waiting for IO to complete.
btrfs: extract the inline extent read code into its own function
Currently we have inline extent read code behind two levels of
indentation, factor them them out into a new function,
read_inline_extent(), to make it a little easier to read.
Since we're here, also remove @extent_offset and @pg_offset arguments
from uncompress_inline() function, as it's not possible to have inline
extents at non-inline file offset.
btrfs: make inline extent read calculation much simpler
Currently we calculate inline extent read in a way that inline extent
can start at non-zero offset.
This is consistent with the inode selftests, which puts an inline extent
at file offset 5.
Meanwhile the inline extent creation code will only create inline extent
at file offset 0.
Furthermore with the introduction of tree-checker on file extents, we are
actively rejecting inline extent which starts at non-zero file offset.
And so far we haven't yet seen any report of rejected inline extents at
non-zero file offset.
This all means, the extra calculation to support inline extents at
non-zero file offset is mostly paper weight, and damaging the
readability of the code.
Thus this patch will:
- Add extra ASSERT()s to make sure involved file offset are all 0
- Remove @extent_offset calculation
- Simplify the involved code
As several variables are now single-use, no need to declare them as
a variable anymore.
Josef Bacik [Wed, 26 Oct 2022 19:08:22 +0000 (15:08 -0400)]
btrfs: rename tree-defrag.c to defrag.c
This currently has only one helper in it, and it's for tree based
defrag. We have the various defrag code in 3 different places, so
rename this to defrag.c. Followup patches will move the code into this
new file.
Josef Bacik [Wed, 26 Oct 2022 19:08:21 +0000 (15:08 -0400)]
btrfs: move inode prototypes to btrfs_inode.h
I initially wanted to make a new header file for this, but these
prototypes do naturally fit into btrfs_inode.h. If we want to extract
vfs from pure btrfs code in the future we may need to split this up, but
btrfs_inode embeds the vfs_inode, so it makes sense to put the
prototypes in this header for now.
Josef Bacik [Wed, 26 Oct 2022 19:08:20 +0000 (15:08 -0400)]
btrfs: move the printk and assert helpers to messages.c
These helpers are core to btrfs, and in order to more easily sync
various parts of the btrfs kernel code into btrfs-progs we need to be
able to carry these helpers with us. However we want to have our own
implementation for the helpers themselves, currently they're implemented
in different files that we want to sync inside of btrfs-progs itself.
Move these into their own C file, this will allow us to contain our
overrides in btrfs-progs in it's own file without messing with the rest
of the codebase.
In copying things over I fixed up a few whitespace errors that already
existed.
Josef Bacik [Wed, 26 Oct 2022 19:08:19 +0000 (15:08 -0400)]
btrfs: add blk_types.h include to compression.h
When moving the printk messages into their own file I got a compiler
error because the includes grabbed compression.h, but nothing pulled in
the blk_types.h dependency that compression.h has because it uses
blkstatus_t. Add blk_types.h to compression.h so that this sort of
thing doesn't happen in the future.
Josef Bacik [Wed, 26 Oct 2022 19:08:18 +0000 (15:08 -0400)]
btrfs: add dependencies to fs.h and block-rsv.h
There's several structures that are embedded inside of fs_info.h, so if
we don't have all the proper includes when we include fs.h we'll get a
variety of compile errors. I fixed this by adding a temporary c file
that just had #include "fs.h" and then added include files until the
compiler stopped complaining.
Josef Bacik [Tue, 25 Oct 2022 15:13:06 +0000 (11:13 -0400)]
btrfs: remove unused btrfs_cond_migrate_bytes
The last user of this was removed in 7f9fe6144076 ("btrfs: improve
global reserve stealing logic"), drop this code as it's no longer called
by anybody.
Josef Bacik [Mon, 24 Oct 2022 18:46:53 +0000 (14:46 -0400)]
btrfs: move the lockdep helpers into locking.h
These more naturally fit in with the locking related code, and they're
all defines so they can easily go anywhere, move them out of ctree.h
into locking.h
Omar Sandoval [Thu, 20 Oct 2022 16:58:28 +0000 (12:58 -0400)]
btrfs: extend btrfs_dir_item type to store encryption status
For directories with encrypted files/filenames, we need to store a flag
indicating this fact. There's no room in other fields, so we'll need to
borrow a bit from dir_type. Since it's now a combination of type and
flags, we rename it to dir_flags to reflect its new usage.
The new flag, FT_ENCRYPTED, indicates a directory containing encrypted
data, which is orthogonal to file type; therefore, add the new
flag, and make conversion from directory type to file type strip the
flag.
As the file types almost never change we can afford to use the bits.
Actual usage will be guarded behind an incompat bit, this patch only
adds the support for later use by fscrypt.
btrfs: use struct fscrypt_str instead of struct qstr
While struct qstr is more natural without fscrypt, since it's provided
by dentries, struct fscrypt_str is provided by the fscrypt handlers
processing dentries, and is thus more natural in the fscrypt world.
Replace all of the struct qstr uses with struct fscrypt_str.
btrfs: setup qstr from dentrys using fscrypt helper
Most places where we get a struct qstr, we are doing so from a dentry.
With fscrypt, the dentry's name may be encrypted on-disk, so fscrypt
provides a helper to convert a dentry name to the appropriate disk name
if necessary. Convert each of the dentry name accesses to use
fscrypt_setup_filename(), then convert the resulting fscrypt_name back
to an unencrypted qstr. This does not work for nokey names, but the
specific locations that could spawn nokey names are noted.
At present, since there are no encrypted directories, nothing goes down
the filename encryption paths.
btrfs: use struct qstr instead of name and namelen pairs
Many functions throughout btrfs take name buffer and name length
arguments. Most of these functions at the highest level are usually
called with these arguments extracted from a supplied dentry's name.
But the entire name can be passed instead, making each function a little
more elegant.
Each function whose arguments are currently the name and length
extracted from a dentry is herein converted to instead take a pointer to
the name in the dentry. The couple of calls to these calls without a
struct dentry are converted to create an appropriate qstr to pass in.
Additionally, every function which is only called with a name/len
extracted directly from a qstr is also converted.
This change has positive effect on stack consumption, frame of many
functions is reduced but this will be used in the future for fscrypt
related structures.
Anand Jain [Wed, 19 Oct 2022 16:21:42 +0000 (21:51 +0530)]
btrfs: merge module cleanup sequence to one helper
The module exit function exit_btrfs_fs() is duplicating a section of code
in init_btrfs_fs(). Add a helper to remove the duplicated code. Due
to the init/exit section requirements the function must be inline and
not a plain static as it could cause section mismatch.
David Sterba [Fri, 14 Oct 2022 13:59:35 +0000 (15:59 +0200)]
btrfs: switch GFP_NOFS to GFP_KERNEL in scrub_setup_recheck_block
There's only one caller that calls scrub_setup_recheck_block in the
memalloc_nofs_save/_restore protection so it's effectively already
GFP_NOFS and it's safe to use GFP_KERNEL.
Josef Bacik [Wed, 19 Oct 2022 14:51:00 +0000 (10:51 -0400)]
btrfs: move accessor helpers into accessors.h
This is a large patch, but because they're all macros it's impossible to
split up. Simply copy all of the item accessors in ctree.h and paste
them in accessors.h, and then update any files to include the header so
everything compiles.
Josef Bacik [Wed, 19 Oct 2022 14:50:59 +0000 (10:50 -0400)]
btrfs: move btrfs_map_token to accessors
This is specific to the item-accessor code, move it out of ctree.h into
accessor.h/.c and then update the users to include the new header file.
This un-inlines btrfs_init_map_token, however this is only called once
per function so it's not critical to be inlined. This also saves 904
bytes of code on a release build.
Josef Bacik [Wed, 19 Oct 2022 14:50:58 +0000 (10:50 -0400)]
btrfs: rename struct-funcs.c to accessors.c
Rename struct-funcs.c to accessors.c so we can move the item accessors
out of ctree.h. accessors.c is a better description of the code that is
contained in these files.
Josef Bacik [Wed, 19 Oct 2022 14:50:56 +0000 (10:50 -0400)]
btrfs: remove fs_info::pending_changes and related code
Now that we're not using this code anywhere we can remove it as well as
the member from fs_info.
We don't have any mount options or on/off features that would utilize
the pending infrastructure, the last one was inode_cache.
There was a patchset [1] to enable some features from sysfs that would
break things if it would be set immediately. In case we'll need that
kind of logic again the patch can be reverted, but for the current use
it can be replaced by the single state bit to do the commit.
Josef Bacik [Wed, 19 Oct 2022 14:50:55 +0000 (10:50 -0400)]
btrfs: add a BTRFS_FS_NEED_TRANS_COMMIT flag
Currently we are only using fs_info->pending_changes to indicate that we
need a transaction commit. The original users for this were removed
years ago and we don't have more usage in sight, so this is the only
remaining reason to have this field. Add a flag so we can remove this
code.
Josef Bacik [Wed, 19 Oct 2022 14:50:52 +0000 (10:50 -0400)]
btrfs: convert incompat and compat flag test helpers to macros
These helpers use functions not defined in fs.h, they're simply
accessors of the super block in fs_info, convert them to macros so
that we don't have a weird dependency between fs.h and accessors.h.
Josef Bacik [Wed, 19 Oct 2022 14:50:51 +0000 (10:50 -0400)]
btrfs: move BTRFS_FS_STATE* definitions and helpers to fs.h
We're going to use fs.h to hold fs wide related helpers and definitions,
move the FS_STATE enum and related helpers to fs.h, and then update all
files that need these definitions to include fs.h.
Josef Bacik [Wed, 19 Oct 2022 14:50:50 +0000 (10:50 -0400)]
btrfs: push printk index code into their respective helpers
The printk index work can be pushed into the printk helpers themselves,
this allows us to further sanitize messages.h, removing the last
include in the header itself.
Josef Bacik [Wed, 19 Oct 2022 14:50:49 +0000 (10:50 -0400)]
btrfs: move the printk helpers out of ctree.h
We have a bunch of printk helpers that are in ctree.h. These have
nothing to do with ctree.c, so move them into their own header.
Subsequent patches will cleanup the printk helpers.
Josef Bacik [Wed, 19 Oct 2022 14:50:48 +0000 (10:50 -0400)]
btrfs: move assert helpers out of ctree.h
These call functions that aren't defined in, or will be moved out of,
ctree.h Move them to super.c where the other assert/error message code
is defined. Drop the __noreturn attribute for btrfs_assertfail as
objtool does not like it and fails with warnings like
fs/btrfs/dir-item.o: warning: objtool: .text.unlikely: unexpected end of section
fs/btrfs/xattr.o: warning: objtool: btrfs_setxattr() falls through to next function btrfs_setxattr_trans.cold()
fs/btrfs/xattr.o: warning: objtool: .text.unlikely: unexpected end of section
Josef Bacik [Wed, 19 Oct 2022 14:50:47 +0000 (10:50 -0400)]
btrfs: move fs wide helpers out of ctree.h
We have several fs wide related helpers in ctree.h. The bulk of these
are the incompat flag test helpers, but there are things such as
btrfs_fs_closing() and the read only helpers that also aren't directly
related to the ctree code. Move these into a fs.h header, which will
serve as the location for file system wide related helpers.
David Sterba [Tue, 18 Oct 2022 14:06:38 +0000 (16:06 +0200)]
btrfs: simplify generation check in btrfs_get_dentry
Callers that pass non-zero generation always want to perform the
generation check, we can simply encode that in one parameter and drop
check_generation. Add function documentation.
David Sterba [Tue, 26 Jul 2022 18:54:10 +0000 (20:54 +0200)]
btrfs: auto enable discard=async when possible
There's a request to automatically enable async discard for capable
devices. We can do that, the async mode is designed to wait for larger
freed extents and is not intrusive, with limits to iops, kbps or latency.
The status and tunables will be exported in /sys/fs/btrfs/FSID/discard .
The automatic selection is done if there's at least one discard capable
device in the filesystem (not capable devices are skipped). Mounting
with any other discard option will honor that option, notably mounting
with nodiscard will keep it disabled.
Josef Bacik [Fri, 14 Oct 2022 14:00:41 +0000 (10:00 -0400)]
btrfs: do not panic if we can't allocate a prealloc extent state
We sometimes have to allocate new extent states when clearing or setting
new bits in an extent io tree. Generally we preallocate this before
taking the tree spin lock, but we can use this preallocated extent state
sometimes and then need to try to do a GFP_ATOMIC allocation under the
lock.
Unfortunately sometimes this fails, and then we hit the BUG_ON() and
bring the box down. This happens roughly 20 times a week in our fleet.
However the vast majority of callers use GFP_NOFS, which means that if
this GFP_ATOMIC allocation fails, we could simply drop the spin lock, go
back and allocate a new extent state with our given gfp mask, and begin
again from where we left off.
For the remaining callers that do not use GFP_NOFS, they are generally
using GFP_NOWAIT, which still allows for some reclaim. So allow these
allocations to attempt to happen outside of the spin lock so we don't
need to rely on GFP_ATOMIC allocations.
This in essence creates an infinite loop for anything that isn't
GFP_NOFS. To address this we may want to migrate to using mempools for
extent states so that we will always have emergency reserves in order to
make our allocations.
Josef Bacik [Fri, 14 Oct 2022 14:00:39 +0000 (10:00 -0400)]
btrfs: do not use GFP_ATOMIC in the read endio
We have done read endio in an async thread for a very, very long time,
which makes the use of GFP_ATOMIC and unlock_extent_atomic() unneeded in
our read endio path. We've noticed under heavy memory pressure in our
fleet that we can fail these allocations, and then often trip a
BUG_ON(!allocation), which isn't an ideal outcome. Begin to address
this by simply not using GFP_ATOMIC, which will allow us to do things
like actually allocate a extent state when doing
set_extent_bits(UPTODATE) in the endio handler.
End io handlers are not called in atomic context, besides we have been
allocating failrec with GFP_NOFS so we'd notice there's a problem.
btrfs: skip update of block group item if used bytes are the same
[BACKGROUND]
When committing a transaction, we will update block group items for all
dirty block groups.
But in fact, dirty block groups don't always need to update their block
group items.
It's pretty common to have a metadata block group which experienced
several COW operations, but still have the same amount of used bytes.
In that case, we may unnecessarily COW a tree block doing nothing.
[ENHANCEMENT]
This patch will introduce btrfs_block_group::commit_used member to
remember the last used bytes, and use that new member to skip
unnecessary block group item update.
This would be more common for large filesystems, where metadata block
group can be as large as 1GiB, containing at most 64K metadata items.
In that case, if COW added and then deleted one metadata item near the
end of the block group, then it's completely possible we don't need to
touch the block group item at all.
[BENCHMARK]
The change itself can have quite a high chance (20~80%) to skip block
group item updates in lot of workloads.
As a result, it would result shorter time spent on
btrfs_write_dirty_block_groups(), and overall reduce the execution time
of the critical section of btrfs_commit_transaction().
Here comes a fio command, which will do random writes in 4K block size,
causing a very heavy metadata updates.
The file size (512M) and number of threads (4) means 2GiB file size in
total, but during the full 300s run time, my dedicated SATA SSD is able
to write around 20~25GiB, which is over 10 times the file size.
Thus after we fill the initial 2G, we should not cause much block group
item updates.
Please note, the fio numbers by themselves don't have much change, but
if we look deeper, there is some reduced execution time, especially for
the critical section of btrfs_commit_transaction().
I added extra trace_printk() to measure the following per-transaction
execution time:
- Critical section of btrfs_commit_transaction()
By re-using the existing update_commit_stats() function, which
has already calculated the interval correctly.
- The while() loop for btrfs_write_dirty_block_groups()
Although this includes the execution time of btrfs_run_delayed_refs(),
it should still be representative overall.
Both result involves transid 7~30, the same amount of transaction
committed.
The result looks like this:
| Before | After | Diff
----------------------+-------------------+----------------+--------
Transaction interval | 229247198.5 | 215016933.6 | -6.2%
Block group interval | 23133.33333 | 18970.83333 | -18.0%
The change in block group item updates is more obvious, as skipped block
group item updates also mean less delayed refs.
And the overall execution time for that block group update loop is
pretty small, thus we can assume the extent tree is already mostly
cached. If we can skip an uncached tree block, it would cause more
obvious change.
Unfortunately the overall reduction in commit transaction critical
section is much smaller, as the block group item updates loop is not
really the major part, at least not for the above fio script.
But still we have a observable reduction in the critical section.
David Sterba [Fri, 9 Sep 2022 15:43:06 +0000 (17:43 +0200)]
btrfs: convert __TRANS_* defines to enum bits
The base transaction bits can be defined as bits in a contiguous
sequence, although right now there's a hole from bit 1 to 8.
The bits are used for btrfs_trans_handle::type, and there's another set
of TRANS_STATE_* defines that are for btrfs_transaction::state. They are
mutually exclusive though the hole in the sequence looks like was made
for the states.
projects / linux.git / log