bcachefs: Fix printing of device durability

BCH_MEMBER_DURABILITY() was not present initially; a value of 0 means
use the default, nonzero means use v - 1.

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: __bch2_journal_key_to_wb -> bch2_journal_key_to_wb_slowpath

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: __journal_keys_sort() refactoring

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: wb_key_cmp -> wb_key_ref_cmp

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: track transaction durations

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: btree_trans always has stats

reserve slot 0 for unknown (when we overflow), to avoid some branches

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: Split brain detection

Use the new bch_member->seq, sb->write_time fields to detect split brain
and kick out devices when necessary.

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: bch_member->seq

Add new fields for split brain detection:

 - bch_member->seq, which tracks the sequence number of the last superblock
   write that happened to each member device

 - bch_sb->write_time, which tracks the time of the last superblock write,
   to allow detection of when two members have diverged but had the same
   number of superblock writes.

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: Fix nochanges/read_only interaction

nochanges means "we cannot issue writes at all"; it's possible to go
into a pseudo read-write mode where we pin dirty metadata in memory,
which is used for fsck in dry run mode and doing journal replay on a
read only mount, but we do not want to allow an actual read-write mount
in nochanges mode.

But we do always want to allow early read-write, during recovery - this
patch clarifies that.

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: Check journal entries for invalid keys in trans commit path

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: check_directory_structure() can now be run online

Now that we have dynamically resizable btree paths,
check_directory_structure() can check one path - inode up to the root -
in a single transaction.

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: Fix reattach_inode() for snapshots

reattach_inode() was broken w.r.t. snapshots - we'd lookup the subvolume
to look up lost+found, but if we're in an interior node snapshot that
didn't make any sense.

Instead, this adds a dirent path for creating in a specific snapshot,
skipping the subvolume; and we also make sure to create lost+found in
the root snapshot, to avoid conflicts with lost+found being created in
overlapping snapshots.

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: bch2_btree_trans_peek_slot_updates

refactoring the BTREE_ITER_WITH_UPDATES code, prep for removing the flag
and making it always-on

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: bch2_btree_trans_peek_prev_updates

bch2_btree_iter_peek_prev() now supports BTREE_ITER_WITH_UPDATES

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: bch2_btree_trans_peek_updates

refactoring the BTREE_ITER_WITH_UPDATES code, prep for removing the flag
and making it always-on

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: growable btree_paths

XXX: we're allocating memory with btree locks held - bad

We need to plumb through an error path so we can do
allocate_dropping_locks() - but we're merging this now because it fixes
a transaction path overflow caused by indirect extent fragmentation, and
the resize path is rare.

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: Fix interior update path btree_path uses

Since the btree_paths array is now about to become growable, we have to
be careful not to refer to paths by pointer across contexts where they
may be reallocated.

This fixes the remaining btree_interior_update() paths - split and
merge.

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: trans->nr_paths

Start to plumb through dynamically growable btree_paths; this patch
replaces most BTREE_ITER_MAX references with trans->nr_paths.

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: trans->updates will also be resizable

the reflink triggers are also bumping up against the maximum number of
paths in a transaction - and generating proportional numbers of updates.

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: optimize __bch2_trans_get(), kill DEBUG_TRANSACTIONS

 - Some tweaks to greatly reduce locking overhead for the list of btree
   transactions, so that it can always be enabled: leave btree_trans
   objects on the list when they're on the percpu single item freelist,
   and only check for duplicates in the same process when
   CONFIG_BCACHEFS_DEBUG is enabled

 - don't zero out the full btree_trans() unless we allocated it from
   the mempool

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: rcu protect trans->paths

Upcoming patches are going to be changing trans->paths to a
reallocatable buffer. We need to guard against use after free when it's
used by other threads; this introduces RCU protection to those paths and
changes them to check for trans->paths == NULL

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: Clean up btree_trans

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: kill btree_path.idx

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: get_unlocked_mut_path() -> btree_path_idx_t

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: bch2_btree_iter_peek_prev() no longer uses path->idx

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: bch2_path_get() no longer uses path->idx

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: trans_for_each_path_with_node() no longer uses path->idx

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: trans_for_each_path() no longer uses path->idx

path->idx is now a code smell: we should be using path_idx_t, since it's
stable across btree path reallocation.

This is also a bit faster, using the same loop counter vs. fetching
path->idx from each path we iterate over.

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: kill trans_for_each_path_from()

dead code

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: bch2_btree_path_to_text() -> btree_path_idx_t

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: struct trans_for_each_path_inorder_iter

reducing our usage of path->idx

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: btree_insert_entry -> btree_path_idx_t

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: btree_iter -> btree_path_idx_t

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: btree_path_alloc() -> btree_path_idx_t

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: bch2_btree_path_traverse() -> btree_path_idx_t

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: bch2_btree_path_make_mut() -> btree_path_idx_t

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: bch2_btree_path_set_pos() -> btree_path_idx_t

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs; bch2_path_put() -> btree_path_idx_t

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: bch2_path_get() -> btree_path_idx_t

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: minor bch2_btree_path_set_pos() optimization

bpos_eq() is cheaper than bpos_cmp()

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: Kill GFP_NOFAIL usage in readahead path

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: Convert split_devs() to darray

Bit of cleanup & modernization: also moving this code to util.c, it'll
be used by userspace as well.

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: skip journal more often in key cache reclaim

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: for_each_keylist_key() declares loop iter

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: bkey_for_each_ptr() now declares loop iter

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: kill __bch2_btree_iter_peek_upto_and_restart()

dead code

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: fsck -> bch2_trans_run()

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: vstruct_for_each() now declares loop iter

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: for_each_member_device_rcu() now declares loop iter

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: for_each_member_device() now declares loop iter

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: for_each_btree_key() now declares loop iter

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: kill for_each_btree_key_norestart()

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: kill for_each_btree_key_old_upto()

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: for_each_btree_key_upto() -> for_each_btree_key_old_upto()

And for_each_btree_key2_upto -> for_each_btree_key_upto

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: bch2_dirent_lookup() -> lockrestart_do()

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: bch2_trans_srcu_lock() should be static

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: use track_event_change() for allocator blocked stats

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: fix warning about uninitialized time_stats

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: add more verbose logging

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: better error message in btree_node_write_work()

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: simplify bch_devs_list

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: darray_for_each() now declares loop iter

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: trans_for_each_update() now declares loop iter

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: Improve the nopromote tracepoint

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: Use GFP_KERNEL for promote allocations

We already have btree locks dropped here - no need for GFP_NOFS.

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: mean and variance: fix kernel-doc for function params

Add missing function parameter descriptions in mean_and_variance.c.
The also eliminates the "Excess function parameter" warnings.

Prevents these kernel-doc warnings:

mean_and_variance.c:67: warning: Function parameter or member 's' not described in 'mean_and_variance_get_mean'
mean_and_variance.c:78: warning: Function parameter or member 's1' not described in 'mean_and_variance_get_variance'
mean_and_variance.c:94: warning: Function parameter or member 's' not described in 'mean_and_variance_get_stddev'
mean_and_variance.c:108: warning: Function parameter or member 's' not described in 'mean_and_variance_weighted_update'
mean_and_variance.c:108: warning: Function parameter or member 'x' not described in 'mean_and_variance_weighted_update'
mean_and_variance.c:108: warning: Excess function parameter 's1' description in 'mean_and_variance_weighted_update'
mean_and_variance.c:108: warning: Excess function parameter 's2' description in 'mean_and_variance_weighted_update'
mean_and_variance.c:134: warning: Function parameter or member 's' not described in 'mean_and_variance_weighted_get_mean'
mean_and_variance.c:143: warning: Function parameter or member 's' not described in 'mean_and_variance_weighted_get_variance'
mean_and_variance.c:153: warning: Function parameter or member 's' not described in 'mean_and_variance_weighted_get_stddev'

Signed-off-by: Randy Dunlap <rdunlap@infradead.org>
Cc: Kent Overstreet <kent.overstreet@linux.dev>
Cc: Brian Foster <bfoster@redhat.com>
Cc: linux-bcachefs@vger.kernel.org
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: check for failure to downgrade

With the upcoming member seq patch, it's now critical that we don't ever
write to a superblock that hasn't been version downgraded - failure to
update member seq fields will cause split brain detection to fire
erroniously.

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: Fixes for rust bindgen

bindgen doesn't seem to like u128 or DECLARE_FLEX_ARRAY(), but we can
hack around them.

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: Delete dio read alignment check

We'll typically fomat devices with the physical blocksize supported, but
the logical blocksize will be smaller.

There's no real need to be checking the blocksize at the filesystem
level, anyways - the block layer has to check this anyways.

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

MAINTAINERS: Update my email address

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: clean up some dead fallocate code

The have_reservation local variable in bch2_extent_fallocate() is
initialized to false and set to true further down in the function.
Between this two points, one branch of code checks for negative
value and one for positive, and nothing ever checks the variable
after it is set to true. Clean up some of the unnecessary logic and
code.

Signed-off-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: Make sure allocation failure errors are logged

The previous patch fixed a bug in allocation path error handling, and it
would've been noticed sooner had it been logged properly.

Generally speaking, errors that shouldn't happen in normal operation and
are being returned up the stack should be logged: the write path was
already logging IO errors, but non IO errors were missed.

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: drop extra semicolon

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: Replace zero-length array with flex-array member and use __counted_by

Fake flexible arrays (zero-length and one-element arrays) are
deprecated, and should be replaced by flexible-array members.
So, replace zero-length array with a flexible-array member in
`struct bch_ioctl_fsck_offline`.

Also annotate array `devs` with `__counted_by()` to prepare for the
coming implementation by GCC and Clang of the `__counted_by` attribute.
Flexible array members annotated with `__counted_by` can have their
accesses bounds-checked at run-time via `CONFIG_UBSAN_BOUNDS` (for
array indexing) and `CONFIG_FORTIFY_SOURCE` (for strcpy/memcpy-family
functions).

This fixes the following -Warray-bounds warnings:
fs/bcachefs/chardev.c: In function 'bch2_ioctl_fsck_offline':
fs/bcachefs/chardev.c:363:34: warning: array subscript 0 is outside array bounds of '__u64[0]' {aka 'long long unsigned int[]'} [-Warray-bounds=]
  363 |         if (copy_from_user(devs, &user_arg->devs[0], sizeof(user_arg->devs[0]) * arg.nr_devs)) {
      |                                  ^~~~~~~~~~~~~~~~~~
In file included from fs/bcachefs/chardev.c:5:
fs/bcachefs/bcachefs_ioctl.h:400:33: note: while referencing 'devs'
  400 |         __u64                   devs[0];

This results in no differences in binary output.

Signed-off-by: Gustavo A. R. Silva <gustavoars@kernel.org>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: Use array_size() in call to copy_from_user()

Use array_size() helper, instead of the open-coded version in
call to copy_from_user().

Signed-off-by: Gustavo A. R. Silva <gustavoars@kernel.org>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: qstr_eq()

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: bch_err_(fn|msg) check if should print

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: fix userspace build errors

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: Drop journal entry compaction

Previously, we dropped empty journal entries and coalesced entries that
could be - but it's not worth the overhead; we very rarely leave unused
journal entries after getting a journal reservation.

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: kill btree_trans->wb_updates

the btree write buffer path now creates a journal entry directly

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: check_root() can now be run online

check_root() is simple enough to run as one single transaction, so is
trivial to run online.

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: Inline btree write buffer sort

The sort in the btree write buffer flush path is a very hot path, and
it's particularly performance sensitive since it's single threaded and
can block every other thread on a multithreaded write workload.

It's well worth doing a sort with inlined cmp and swap functions.

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: btree write buffer now slurps keys from journal

Previosuly, the transaction commit path would have to add keys to the
btree write buffer as a separate operation, requiring additional global
synchronization.

This patch introduces a new journal entry type, which indicates that the
keys need to be copied into the btree write buffer prior to being
written out. We switch the journal entry type back to
JSET_ENTRY_btree_keys prior to write, so this is not an on disk format
change.

Flushing the btree write buffer may require pulling keys out of journal
entries yet to be written, and quiescing outstanding journal
reservations; we previously added journal->buf_lock for synchronization
with the journal write path.

We also can't put strict bounds on the number of keys in the journal
destined for the write buffer, which means we might overflow the size of
the preallocated buffer and have to reallocate - this introduces a
potentially fatal memory allocation failure. This is something we'll
have to watch for, if it becomes an issue in practice we can do
additional mitigation.

The transaction commit path no longer has to explicitly check if the
write buffer is full and wait on flushing; this is another performance
optimization. Instead, when the btree write buffer is close to full we
change the journal watermark, so that only reservations for journal
reclaim are allowed.

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: journal->buf_lock

Add a new lock for synchronizing between journal IO path and btree write
buffer flush.

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: Unwritten journal buffers are always dirty

Ensure that journal bufs that haven't been written can't be reclaimed
from the journal pin fifo, and can thus have new pins taken.

Prep work for changing the btree write buffer to pull keys from the
journal directly.

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: bch2_trans_node_add no longer uses trans_for_each_path()

In the future we'll be making trans->paths resizable and potentially
having _many_ more paths (for fsck); we need to start fixing algorithms
that walk each path in a transaction where possible.

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: Improve trans->extra_journal_entries

Instead of using a darray, we now allocate journal entries for the
transaction commit path with our normal bump allocator - with an inlined
fastpath, and using btree_transaction_stats to remember how much to
initially allocate so as to avoid transaction restarts.

This is prep work for converting write buffer updates to use this
mechanism.

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs; kill bch2_btree_key_cache_flush()

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: kill btree_path->(alloc_seq|downgrade_seq)

These were for extra info in tracepoints for debugging a specialized
issue - we do not want to bloat btree_path for this, at least in release
builds.

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: Fix snapshot.c assertion for online fsck

c->curr_recovery_pass can go backwards; this adds a non rewinding
version, c->recovery_pass_done.

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: six lock: fix typos

Fix a few typos in the six.h header file.

Signed-off-by: Randy Dunlap <rdunlap@infradead.org>
Cc: Kent Overstreet <kent.overstreet@linux.dev>
Cc: Brian Foster <bfoster@redhat.com>
Cc: linux-bcachefs@vger.kernel.org
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: reserve path idx 0 for sentinal

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: Rename for_each_btree_key2() -> for_each_btree_key()

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: Kill for_each_btree_key()

for_each_btree_key() handles transaction restarts, like
for_each_btree_key2(), but only calls bch2_trans_begin() after a
transaction restart - for_each_btree_key2() wraps every loop iteration
in a transaction.

The for_each_btree_key() behaviour is problematic when it leads to
holding the SRCU lock that prevents key cache reclaim for an unbounded
amount of time - there's no real need to keep it around.

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: continue now works in for_each_btree_key2()

continue now works as in any other loop

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: Fix bch2_read_btree()

In the debugfs code, we had an incorrect use of drop_locks_do(); on
transaction restart we don't want to restart the current loop iteration,
since we've already emitted the current key to the buffer for userspace.

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: Fix open coded set_btree_iter_dontneed()

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: BCH_IOCTL_FSCK_ONLINE

This adds a new ioctl for running fsck on a mounted, in use filesystem.

This reuses the fsck_thread code from the previous patch for running
fsck on an offline, unmounted filesystem, so that log messages for the
fsck thread are redirected to userspace.

Only one running fsck instance is allowed at a time; a new semaphore
(since the lock will be taken by one thread and released by another) is
added for this.

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: BCH_IOCTL_FSCK_OFFLINE

This adds a new ioctl for running fsck on a list of devices.

Normally, if we wish to use the kernel's implementation of fsck we'd run
it at mount time with -o fsck. This ioctl lets us run fsck without
mounting, so that userspace bcachefs-tools can transparently switch to
the kernel's implementation of fsck when appropriate - primarily if the
kernel version of bcachefs better matches the filesystem on disk.

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: bch2_run_online_recovery_passes()

Add a new helper for running online recovery passes - i.e. online fsck.
This is a subset of our normal recovery passes, and does not - for now -
use or follow c->curr_recovery_pass.

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>