Persistent cursors for inode allocation.
A free inodes btree would add substantial overhead to inode allocation
and freeing - a "next num to allocate" cursor is always going to be
faster.
We just need it to be persistent, to avoid scanning the inodes btree
from the start on startup.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
This adds a new inode field, bi_depth, for directory inodes: this allows
us to make the check_directory_structure pass much more efficient.
Currently, to ensure the filesystem is fully connect and has no loops,
for every directory we follow backpointers until we find the root. But
by adding a depth counter, it sufficies to only check the parent of each
directory, and check that the parent's bi_depth is smaller.
(fsck doesn't require that bi_depth = parent->bi_depth + 1; if a rename
causes bi_depth off, but the chain to the root is still strictly
decreasing, then the algorithm still works and there's no need for fsck
to fixup the bi_depth fields).
We've already checked backpointers, so we know that every directory
(excluding the root)has a valid parent: if bi_depth is always
decreasing, every chain must terminate, and terminate at the root
directory.
bi_depth will not necessarily be correct when fsck runs, due to
directory renames - we can't change bi_depth on every child directory
when renaming a directory. That's ok; fsck will silently fix the
bi_depth field as needed, and future fsck runs will be much faster.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
Now that bch2_move_get_io_opts() re-propagates changed inode io options
to bch_extent_rebalance, we can properly suport changing IO path options
for reflinked data.
Changing a per-file IO path option, either via the xattr interface or
via the BCHFS_IOC_REINHERIT_ATTRS ioctl, will now trigger a scan (the
inode number is marked as needing a scan, via
bch2_set_rebalance_needs_scan()), and rebalance will use
bch2_move_data(), which will walk the inode number and pick up the new
options.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
Previously, io path option changes on a file would be picked up
automatically and applied to existing data - but not for reflinked data,
as we had no way of doing this safely. A user may have had permission to
copy (and reflink) a given file, but not write to it, and if so they
shouldn't be allowed to change e.g. nr_replicas or other options.
This uses the incompat feature mechanism in the previous patch to add a
new incompatible flag to bch_reflink_p, indicating whether a given
reflink pointer may propagate io path option changes back to the
indirect extent.
In this initial patch we're only setting it for the source extents.
We'd like to set it for the destination in a reflink copy, when the user
has write access to the source, but that requires mnt_idmap which is not
curretly plumbed up to remap_file_range.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
We've been getting away from feature bits: they don't have any kind of
ordering, and thus it's possible for people to enable weird combinations
of features that were never tested or intended to be run.
Much better to just give every new feature, compatible or incompatible,
a version number.
Additionally, we probably won't ever rev the major version number: major
version numbers represent incompatible versions, but that doesn't really
fit with how we actually roll out incompatible features - we need a
better way of rolling out incompatible features.
So, this patch adds two new superblock fields:
- BCH_SB_VERSION_INCOMPAT
- BCH_SB_VERSION_INCOMPAT_ALLOWED
BCH_SB_VERSION_INCOMPAT_ALLOWED indicates that incompatible features up
to version number x are allowed to be used without user prompting, but
it does not by itself deny old versions from mounting.
BCH_SB_VERSION_INCOMPAT does deny old versions from mounting, and must
be <= BCH_SB_VERSION_INCOMPAT_ALLOWED.
BCH_SB_VERSION_INCOMPAT will only be set when a codepath attempts to use
an incompatible feature, so as to not unnecessarily break compatibility
with old versions.
bch2_request_incompat_feature() is the new interface to check if an
incompatible feature may be used.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
The backpointers passes, check_backpointers_to_extents() and
check_extents_to_backpointers() are the most expensive fsck passes.
Now that we're running the same check and repair code when using a
backpointer at runtime (via bch2_backpointer_get_key()) that fsck does,
there's no reason fsck needs to - except to verify that the filesystem
really has no errors in debug mode.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
Continuing on with the self healing theme, we should be running any
check and repair code at runtime that we can - instead of declaring the
filesystemt inconsistent.
This will also let us skip running the backpointers -> extents fsck pass
except in debug mode.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
Instead of walking every extent and every backpointer it points to,
first sum up backpointers in each bucket and check for mismatches, and
only look for missing backpointers if mismatches were detected, and only
check extents in those buckets.
This is a major fsck scalability improvement, since the two backpointers
passes (backpointers -> extents and extents -> backpointers) are the
most expensive fsck passes by far.
Additionally, to speed up the upgrade for backpointer bucket gens, or in
situations when we have to rebuild alloc info, add a special case for
when no backpointers are found in a bucket - don't check each individual
backpointer (in particular, avoiding the write buffer flushes), just
recreate them.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
In an upcoming patch bch2_backpointer_get_key() will be repairing when
it finds a dangling backpointer; it will need to flush the btree write
buffer before it can definitively say there's an error.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
bch_backpointer no longer contains the bucket_offset field, it's just a
direct LBA mapping (with low bits to account for compressed extent
splitting), so we don't need to refer to the device to construct it
anymore.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
Fix sort order for disk accounting keys, in order to fix a regression on
mount times.
The typetag is now the most significant byte of the key, meaning disk
accounting keys of the same type now sort together.
This lets us skip over disk accounting keys that aren't mirrored in
memory when reading accounting at startup, instead of having them
interleaved with other counter types.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
New on disk format version: backpointers new include the generation
number of the bucket they refer to, and the obsolete bucket_offset field
(no longer needed because we no longer store backpointers in alloc keys)
is gone.
This is an expensive forced upgrade - hopefully the last; we have to run
the extents_to_backpointers recovery pass to regenerate backpointers.
It's a forced incompatible upgrade because the alternative would've been
permamently making backpointers bigger, and as one of the biggest btrees
(along with the extents btree) that's not an ideal option.
It's worth it though, because this allows us to make the
check_extents_to_backpointers pass drastically cheaper: an upcoming
patch changes it to sum up backpointers in a bucket and check the sum
against the sector counts for that bucket, only looking for missing
backpointers if they don't match (and then only for specific buckets).
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
Switch to generating a private list of interior nodes to delete, instead
of using the equivalence class in the global data structure.
This eliminates possible races with snapshot creation, and is much
cleaner - it'll let us delete a lot of janky code for calculating and
maintaining the equivalence classes.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
When deleting dead snapshots, we move keys from redundant interior
snapshot nodes to child nodes - unless there's already a key, in which
case the ancestor key is deleted.
Previously, we tracked via equiv_seen whether the child snapshot had a
key, but this was tricky w.r.t. transaction restarts, and not
transactionally safe w.r.t. updates in the child snapshot.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
This breaks when the trigger is inserting updates for the same btree, as
the inode trigger now does.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
Originally, we ran insert triggers before overwrite so that if an extent
was being moved (by fallocate insert/collapse range), the bucket sector
count wouldn't hit 0 partway through, and so we don't trigger state
changes caused by that too soon.
But this is better solved by just moving the data type change to the
alloc trigger itself, where it's already called.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
Normally, whitouts (KEY_TYPE_whitout) are filtered from btree lookups,
since they exist only to represent deletions of keys in ancestor
snapshots - except, they should not be filtered in
BTREE_ITER_all_snapshots mode, so that e.g. snapshot deletion can clean
them up.
This means that that the key cache has to store whiteouts, and key cache
fills cannot filter them.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
In BTREE_ITER_all_snapshots mode, we're required to only return keys
where the snapshot field matches the iterator position -
BTREE_ITER_filter_snapshots requires pulling keys into the key cache
from ancestor snapshots, so we have to check for that.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
Clang 18 and newer warns (or errors with CONFIG_WERROR=y):
fs/bcachefs/str_hash.c:164:2: error: label followed by a declaration is a C23 extension [-Werror,-Wc23-extensions]
164 | struct bch_inode_unpacked inode;
| ^
In Clang 17 and prior, this is an unconditional hard error:
fs/bcachefs/str_hash.c:164:2: error: expected expression
164 | struct bch_inode_unpacked inode;
| ^
fs/bcachefs/str_hash.c:165:30: error: use of undeclared identifier 'inode'
165 | ret = bch2_inode_unpack(k, &inode);
| ^
fs/bcachefs/str_hash.c:169:55: error: use of undeclared identifier 'inode'
169 | struct bch_hash_info hash2 = bch2_hash_info_init(c, &inode);
| ^
fs/bcachefs/str_hash.c:171:40: error: use of undeclared identifier 'inode'
171 | ret = repair_inode_hash_info(trans, &inode);
| ^
Add an empty statement between the label and the declaration to fix the
warning/error without disturbing the code too much.
Fixes: 2519d3b0d656 ("bcachefs: bch2_str_hash_check_key() now checks inode hash info")
Reported-by: kernel test robot <lkp@intel.com>
Closes: https://lore.kernel.org/oe-kbuild-all/202412092339.QB7hffGC-lkp@intel.com/
Signed-off-by: Nathan Chancellor <nathan@kernel.org>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
bch2_snapshot_equiv() is going away; convert users that just wanted to
know if the snapshot exists to something better
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
Versions of the same inode in different snapshots must have the same
hash info; this is critical for lookups to work correctly.
We're going to be running the str_hash checks online, at readdir or
xattr list time, so we now need str_hash_check_key() to check for inode
hash seed mismatches, since it won't be run right after check_inodes().
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
Bkey validation checks that inodes are well-formed and unpack
successfully, so an unpack error should always indicate memory
corruption or some other kind of hardware bug - but these are still
errors we can recover from.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
Since we added per-inode counters there's now far too many counters to
show in one shot - if we want this in the future, it'll have to be in
debugfs.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
We can't hold mark_lock while calling fsck_err() - that's a deadlock,
mark_lock is meant to be a leaf node lock.
It's also unnecessary for gc_bucket() and bucket_gen(); rcu suffices
since the bucket_gens array describes its size, and we can't race with
device removal or resize during gc/fsck since that takes state lock.
Reported-by: syzbot+38641fcbda1aaffefdd4@syzkaller.appspotmail.com
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
This fixes a deadlock during journal replay when btree node read errors
kick off a ton of rewrites: we don't want them competing with journal
replay.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
For each bucket we track when the bucket became nonempty and when it
became empty again: if we can ensure that there will be no journal
flushes in the range [nonempty, empty) (possibly because they occured at
the same journal sequence number), then it's safe to reuse the bucket
without waiting for a journal commit.
This is a major performance optimization for erasure coding, where
writes are initially replicated, but the extra replicas are quickly
dropped: if those buckets are reused and overwritten without issuing a
cache flush to the underlying device, then they only cost bus bandwidth.
But there's a tricky corner case when there's multiple empty -> nonempty
-> empty transitions in quick succession, i.e. when data is getting
overwritten immediately as it's being written.
If this happens and the previous empty transition hasn't been flushed,
we need to continue tracking the previous nonempty transition - not
start a new one.
Fixing this means we now need to track both the nonempty and empty
transitions in bch_alloc_v4.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
Harder to screw up if we're explicit about the range, and more correct
as journal reservations can be outstanding on multiple journal entries
simultaneously.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
This lets us print the exact location in the journal if it was found in
the journal, or correctly print if it was found in the superblock.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
Fix an O(n^2) issue when we find many overlapping (overwritten) btree
nodes - especially when one node overwrites many smaller nodes.
This was discovered to be an issue with the bcachefs
merge_torture_flakey test - if we had a large btree that was then
emptied, the number of difficult overwrites can be unbounded.
Cc: Kuan-Wei Chiu <visitorckw@gmail.com>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
Check open buckets and buckets waiting for journal commit before doing
other expensive lookups.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
