rcu_pending, btree key cache rework: this solves lock contenting in the
key cache, eliminating the biggest source of the srcu lock hold time
warnings, and drastically improving performance on some metadata heavy
workloads - on multithreaded creates we're now 3-4x faster than xfs.
We're now using an rhashtable instead of the system inode hash table;
this is another significant performance improvement on multithreaded
metadata workloads, eliminating more lock contention.
for_each_btree_key_in_subvolume_upto(): new helper for iterating over
keys within a specific subvolume, eliminating a lot of open coded
"subvolume_get_snapshot()" and also fixing another source of srcu lock
time warnings, by running each loop iteration in its own transaction (as
the existing for_each_btree_key() does).
More work on btree_trans locking asserts; we now assert that we don't
hold btree node locks when trans->locked is false, which is important
because we don't use lockdep for tracking individual btree node locks.
Some cleanups and improvements in the bset.c btree node lookup code,
from Alan.
Rework of btree node pinning, which we use in backpointers fsck. The old
hacky implementation, where the shrinker just skipped over nodes in the
pinned range, was causing OOMs; instead we now use another shrinker with
a much higher seeks number for pinned nodes.
Rebalance now uses BCH_WRITE_ONLY_SPECIFIED_DEVS; this fixes an issue
where rebalance would sometimes fall back to allocating from the full
filesystem, which is not what we want when it's trying to move data to a
specific target.
Use __GFP_ACCOUNT, GFP_RECLAIMABLE for btree node, key cache
allocations.
Idmap mounts are now supported - Hongbo.
Rename whiteouts are now supported - Hongbo.
Erasure coding can now handle devices being marked as failed, or
forcibly removed. We still need the evacuate path for erasure coding,
but it's getting very close to ready for people to start using.
Status, and when will we be taking off experimental:
----------------------------------------------------
Going by critical, user facing bugs getting found and fixed, we're
nearly there. There are a couple key items that need to be finished
before we can take off the experimental label:
- The end-user experience is still pretty painful when the root
filesystem needs a fsck; we need some form of limited self healing so
that necessary repair gets run automatically. Errors (by type) are
recorded in the superblock, so what we need to do next is convert
remaining inconsistent() errors to fsck() errors (so that all runtime
inconsistencies are logged in the superblock), and we need to go
through the list of fsck errors and classify them by which fsck passes
are needed to repair them.
- We need comprehensive torture testing for all our repair paths, to
shake out remaining bugs there. Thomas has been working on the tooling
for this, so this is coming soonish.
Slightly less critical items:
- We need to improve the end-user experience for degraded mounts: right
now, a degraded root filesystem means dropping to an initramfs shell
or somehow inputting mount options manually (we don't want to allow
degraded mounts without some form of user input, except on unattended
servers) - we need the mount helper to prompt the user to allow
mounting degraded, and make sure this works with systemd.
- Scalabiity: we have users running 100TB+ filesystems, and that's
effectively the limit right now due to fsck times. We have some
reworks in the pipeline to address this, we're aiming to make petabyte
sized filesystems practical.
-----BEGIN PGP SIGNATURE-----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=oVCy
-----END PGP SIGNATURE-----
Merge tag 'bcachefs-2024-09-21' of git://evilpiepirate.org/bcachefs
Pull bcachefs updates from Kent Overstreet:
- rcu_pending, btree key cache rework: this solves lock contenting in
the key cache, eliminating the biggest source of the srcu lock hold
time warnings, and drastically improving performance on some metadata
heavy workloads - on multithreaded creates we're now 3-4x faster than
xfs.
- We're now using an rhashtable instead of the system inode hash table;
this is another significant performance improvement on multithreaded
metadata workloads, eliminating more lock contention.
- for_each_btree_key_in_subvolume_upto(): new helper for iterating over
keys within a specific subvolume, eliminating a lot of open coded
"subvolume_get_snapshot()" and also fixing another source of srcu
lock time warnings, by running each loop iteration in its own
transaction (as the existing for_each_btree_key() does).
- More work on btree_trans locking asserts; we now assert that we don't
hold btree node locks when trans->locked is false, which is important
because we don't use lockdep for tracking individual btree node
locks.
- Some cleanups and improvements in the bset.c btree node lookup code,
from Alan.
- Rework of btree node pinning, which we use in backpointers fsck. The
old hacky implementation, where the shrinker just skipped over nodes
in the pinned range, was causing OOMs; instead we now use another
shrinker with a much higher seeks number for pinned nodes.
- Rebalance now uses BCH_WRITE_ONLY_SPECIFIED_DEVS; this fixes an issue
where rebalance would sometimes fall back to allocating from the full
filesystem, which is not what we want when it's trying to move data
to a specific target.
- Use __GFP_ACCOUNT, GFP_RECLAIMABLE for btree node, key cache
allocations.
- Idmap mounts are now supported (Hongbo Li)
- Rename whiteouts are now supported (Hongbo Li)
- Erasure coding can now handle devices being marked as failed, or
forcibly removed. We still need the evacuate path for erasure coding,
but it's getting very close to ready for people to start using.
* tag 'bcachefs-2024-09-21' of git://evilpiepirate.org/bcachefs: (99 commits)
bcachefs: return err ptr instead of null in read sb clean
bcachefs: Remove duplicated include in backpointers.c
bcachefs: Don't drop devices with stripe pointers
bcachefs: bch2_ec_stripe_head_get() now checks for change in rw devices
bcachefs: bch_fs.rw_devs_change_count
bcachefs: bch2_dev_remove_stripes()
bcachefs: bch2_trigger_ptr() calculates sectors even when no device
bcachefs: improve error messages in bch2_ec_read_extent()
bcachefs: improve error message on too few devices for ec
bcachefs: improve bch2_new_stripe_to_text()
bcachefs: ec_stripe_head.nr_created
bcachefs: bch_stripe.disk_label
bcachefs: stripe_to_mem()
bcachefs: EIO errcode cleanup
bcachefs: Rework btree node pinning
bcachefs: split up btree cache counters for live, freeable
bcachefs: btree cache counters should be size_t
bcachefs: Don't count "skipped access bit" as touched in btree cache scan
bcachefs: Failed devices no longer require mounting in degraded mode
bcachefs: bch2_dev_rcu_noerror()
...
-----BEGIN PGP SIGNATURE-----
iHUEABYKAB0WIQRAhzRXHqcMeLMyaSiRxhvAZXjcogUCZuQEvgAKCRCRxhvAZXjc
ou77AQD3U1KjbdgzbUi6kaUmiiWOPhfYTlm8mho8dBjqvTCB+AD/XTWSFCWWhHB4
KyQZTbjRD81xmVNbKjASazp0EA6Ahwc=
=gIsD
-----END PGP SIGNATURE-----
Merge tag 'vfs-6.12.folio' of gitolite.kernel.org:pub/scm/linux/kernel/git/vfs/vfs
Pull vfs folio updates from Christian Brauner:
"This contains work to port write_begin and write_end to rely on folios
for various filesystems.
This converts ocfs2, vboxfs, orangefs, jffs2, hostfs, fuse, f2fs,
ecryptfs, ntfs3, nilfs2, reiserfs, minixfs, qnx6, sysv, ufs, and
squashfs.
After this series lands a bunch of the filesystems in this list do not
mention struct page anymore"
* tag 'vfs-6.12.folio' of gitolite.kernel.org:pub/scm/linux/kernel/git/vfs/vfs: (61 commits)
Squashfs: Ensure all readahead pages have been used
Squashfs: Rewrite and update squashfs_readahead_fragment() to not use page->index
Squashfs: Update squashfs_readpage_block() to not use page->index
Squashfs: Update squashfs_readahead() to not use page->index
Squashfs: Update page_actor to not use page->index
jffs2: Use a folio in jffs2_garbage_collect_dnode()
jffs2: Convert jffs2_do_readpage_nolock to take a folio
buffer: Convert __block_write_begin() to take a folio
ocfs2: Convert ocfs2_write_zero_page to use a folio
fs: Convert aops->write_begin to take a folio
fs: Convert aops->write_end to take a folio
vboxsf: Use a folio in vboxsf_write_end()
orangefs: Convert orangefs_write_begin() to use a folio
orangefs: Convert orangefs_write_end() to use a folio
jffs2: Convert jffs2_write_begin() to use a folio
jffs2: Convert jffs2_write_end() to use a folio
hostfs: Convert hostfs_write_end() to use a folio
fuse: Convert fuse_write_begin() to use a folio
fuse: Convert fuse_write_end() to use a folio
f2fs: Convert f2fs_write_begin() to use a folio
...
Same as the recent change for __bch2_read(); also, kill now unnecessary
btree_trans_too_many_iters() calls.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
the standard vfs inode hash table suffers from painful lock contention -
this is long overdue
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
We had a report of data corruption on nixos when building installer
images.
https://github.com/NixOS/nixpkgs/pull/321055#issuecomment-2184131334
It seems that writes are being dropped, but only when issued by QEMU,
and possibly only in snapshot mode. It's undetermined if it's write
calls are being dropped or dirty folios.
Further testing, via minimizing the original patch to just the change
that skips the inode lock on non appends/truncates, reveals that it
really is just not taking the inode lock that causes the corruption: it
has nothing to do with the other logic changes for preserving write
atomicity in corner cases.
It's also kernel config dependent: it doesn't reproduce with the minimal
kernel config that ktest uses, but it does reproduce with nixos's distro
config. Bisection the kernel config initially pointer the finger at page
migration or compaction, but it appears that was erroneous; we haven't
yet determined what kernel config option actually triggers it.
Sadly it appears this will have to be reverted since we're getting too
close to release and my plate is full, but we'd _really_ like to fully
debug it.
My suspicion is that this patch is exposing a preexisting bug - the
inode lock actually covers very little in IO paths, and we have a
different lock (the pagecache add lock) that guards against races with
truncate here.
Fixes: 7e64c86cdc ("bcachefs: Buffered write path now can avoid the inode lock")
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
Convert all callers from working on a page to working on one page
of a folio (support for working on an entire folio can come later).
Removes a lot of folio->page->folio conversions.
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Signed-off-by: Christian Brauner <brauner@kernel.org>
Most callers have a folio, and most implementations operate on a folio,
so remove the conversion from folio->page->folio to fit through this
interface.
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Signed-off-by: Christian Brauner <brauner@kernel.org>
This adds a new helper, bch2_folio_reservation_get_partial(), which
reserves as many blocks as possible and may return partial success.
__bch2_buffered_write() is switched to the new helper - this fixes
fstests generic/275, the write until -ENOSPC test.
generic/230 now fails: this appears to be a test bug, where xfs_io isn't
looping after a partial write to get the error code.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
Set the preferred folio order in the fgp_flags by calling
fgf_set_order(). Page cache will try to allocate large folio of the
preferred order whenever possible instead of allocating multiple 0 order
folios.
This improves the buffered write performance up to 1.25x with default
mount options and up to 1.57x when mounted with no_data_io option with
the following fio workload:
fio --name=bcachefs --filename=/mnt/test --size=100G \
--ioengine=io_uring --iodepth=16 --rw=write --bs=128k
Signed-off-by: Pankaj Raghav <p.raghav@samsung.com>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
Use FGP_WRITEBEGIN to avoid repeating the individual FGP flags before
starting a buffered write.
Signed-off-by: Pankaj Raghav <p.raghav@samsung.com>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
copy_page_from_iter_atomic() will be removed at some point.
Also fixup a comment for folios.
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
Common code doesn't test the error flag, so we don't need to set it in
bcachefs. We can use folio_end_read() to combine the setting (or not)
of the uptodate flag and clearing the lock flag.
Cc: Kent Overstreet <kent.overstreet@linux.dev>
Cc: Brian Foster <bfoster@redhat.com>
Cc: linux-bcachefs@vger.kernel.org
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
Combine iter/update/trigger/str_hash flags into a single enum, and
x-macroize them for a to_text() function later.
These flags are all for a specific iter/key/update context, so it makes
sense to group them together - iter/update/trigger flags were already
given distinct bits, this cleans up and unifies that handling.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
Non append, non extending buffered writes can now avoid taking the inode
lock.
To ensure atomicity of writes w.r.t. other writes, we lock every folio
that we'll be writing to, and if this fails we fall back to taking the
inode lock.
Extensive comments are provided as to corner cases.
Link: https://lore.kernel.org/linux-fsdevel/Zdkxfspq3urnrM6I@bombadil.infradead.org/
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
Recently, we fixed our __GFP_NOFAIL usage in the readahead path, but the
easy one in read_single_folio() (where wa can return an error) was
missed - oops.
Fixes:
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
Now we can print out filesystem flags in sysfs, useful for debugging
various "what's my filesystem doing" issues.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
The bcachefs folio writeback code includes a bio full check as well
as a fixed size check to determine when to split off and submit
writeback I/O. The inclusive check of the latter against the limit
means that writeback can submit slightly prematurely. This is not a
functional problem, but results in unnecessarily split I/Os and
extent merging.
This can be observed with a buffered write sized exactly to the
current maximum value (1MB) and with key_merging_disabled=1. The
latter prevents the merge from the second write such that a
subsequent check of the extent list shows a 1020k extent followed by
a contiguous 4k extent.
The purpose for the fixed size check is also undocumented and
somewhat obscure. Lift this check into a new helper that wraps the
bio check, fix the comparison logic, and add a comment to document
the purpose and how we might improve on this in the future.
Signed-off-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
We're using more stack than we'd like in a number of functions, and
btree_trans is the biggest object that we stack allocate.
But we have to do a heap allocatation to initialize it anyways, so
there's no real downside to heap allocating the entire thing.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
More reorganization, this splits up io.c into
- io_read.c
- io_misc.c - fallocate, fpunch, truncate
- io_write.c
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
In __bch2_buffered_write, if we fail to write to an entire !uptodate
folio, we have to back out the write, bail out and retry.
But we were missing an iov_iter_revert() call, so the data written to
the folio was lost and the rest of the write shifted to the wrong
offset.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
fs-io.c is too big - time for some reorganization
- fs-dio.c: direct io
- fs-pagecache.c: pagecache data structures (bch_folio), utility code
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
