Btree nodes are log structured; thus, we need to emit whiteouts when
we're deleting a key that's been written out to disk.
k->needs_whiteout tracks whether a key will need a whiteout when it's
deleted, and this requires some careful handling; e.g. the key we're
deleting may not have been written out to disk, but it may have
overwritten a key that was - thus we need to carry this flag around on
overwrites.
Invariants:
There may be multiple key for the same position in a given node (because
of overwrites), but only one of them will be a live (non deleted) key,
and only one key for a given position will have the needs_whiteout flag
set.
Additionally, we don't want to carry around whiteouts that need to be
written in the main searchable part of a btree node - btree_iter_peek()
will have to skip past them, and this can lead to an O(n^2) issues when
doing sequential deletions (e.g. inode rm/truncate). So there's a
separate region in the btree node buffer for unwritten whiteouts; these
are merge sorted with the rest of the keys we're writing in the btree
node write path.
The unwritten whiteouts was a later optimization that bch2_sort_keys()
didn't take into account; the unwritten whiteouts area means that we
never have deleted keys with needs_whiteout set in the main searchable
part of a btree node.
That means we can simplify and optimize some sort paths, and eliminate
an assertion that syzbot found:
- Unless we're in the btree node write path, it's always ok to drop
whiteouts when sorting
- When sorting for a btree node write, we drop the whiteout if it's not
from the unwritten whiteouts area, or if it's overwritten by a real
key at the same position.
This completely eliminates some tricky logic for propagating the
needs_whiteout flag: syzbot was able to hit the assertion that checked
that there shouldn't be more than one key at the same pos with
needs_whiteout set, likely due to a combination of flipping on
needs_whiteout on all written keys (they need whiteouts if overwritten),
combined with not always dropping unneeded whiteouts, and the tricky
logic in the sort path for preserving needs_whiteout that wasn't really
needed.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
bch2_write_super() was looping over online devices multiple times -
dropping and retaking io_ref each time.
This meant it could race with device removal; it could increment the
sequence number on a device but fail to write it - and then if the
device was re-added, it would get confused the next time around thinking
a superblock write was silently dropped.
Fix this by taking io_ref once, and stashing pointers to online devices
in a darray.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
Ancient versions of bcachefs produced packed formats that could
represent keys that our in memory format cannot represent;
bformat_needs_redo() has some tricky shifts to check for this sort of
overflow.
Reported-by: syzbot+594427aebfefeebe91c6@syzkaller.appspotmail.com
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
filefrag (and potentially other utilities that call fiemap) sometimes
pass ULONG_MAX as the length. fiemap_prep clamps excessively large
lengths - but the calculation of end can overflow if it occurs before
calling fiemap_prep. When this happens, filefrag assumes it has read to
the end and exits.
Signed-off-by: Reed Riley <reed@riley.engineer>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
The bucket_gens array is a single array allocation (one byte per
bucket), and kernel allocations are still limited to INT_MAX.
Check this limit to avoid failing the bucket_gens array allocation.
Reported-by: syzbot+b29f436493184ea42e2b@syzkaller.appspotmail.com
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
bch2_get_next_dev() and bch2_get_next_online_dev() iterate over devices,
dropping and taking refs as they go; we can't access the previous device
(for ca->dev_idx) after we've dropped our ref to it, unless we take
rcu_read_lock() first.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
bch2_dev_lookup() is supposed to take a ref on the device it returns, but
for_each_member_device() takes refs as it iterates,
for_each_member_device_rcu() does not.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
We're using mutex_lock() inside a wait_event() conditional -
prepare_to_wait() has already flipped task state, so potentially
blocking ops need annotation.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
