When a device is removed from a bcachefs volume, the associated
content is removed from the various btrees. The alloc tree uses the
key cache, so when keys are removed the deletes exist in cache for a
period of time until reclaim comes along and flushes outstanding
updates.

When a device is re-added to the bcachefs volume, the add process
re-adds some of these previously deleted keys. When marking device
superblock locations on device add, the keys will likely refer to
some of the same alloc keys that were just removed. The memory
triggers for these key updates are responsible for further updates,
such as bch2_mark_alloc() calling into bch2_dev_usage_update() to
update per-device usage accounting.

When a new key is added to key cache, the trans update path also
flushes the key to the backing btree for coherency reasons for tree
walks.

With all of this context, if a device is removed and re-added
quickly enough such that some key deletes from the remove are still
pending a key cache flush, the trans update path can view this as
addition of a new key because the old key in the insert entry refers
to a deleted key. However the deleted cached key has not been filled
by absence of a btree key, but rather refers to an explicit deletion
of an existing key that occurred during device removal.

The trans update path adds a new update to flush the key and tags
the original (cached) update to skip running the memory triggers.
This results in running triggers on the non-cached update instead,
which in turn will perform accounting updates based on incoherent
values. For example, bch2_dev_usage_update() subtracts the the old
alloc key dirty sector count in the non-cached btree key from the
newly initialized (i.e. zeroed) per device counters, leading to
underflow and accounting corruption.

There are at least a few ways to avoid this problem, the simplest of
which may be to run triggers against the cached update rather than
the non-cached update. If the key only needs to be flushed when the
key is not present in the tree, however, then this still performs an
unnecessary update. We could potentially use the cached key dirty
state to determine whether the delete is a dirty, cached update vs.
a clean cache fill, but this may require transmitting key cache
dirty state across layers, which adds complexity and seems to be of
limited value. Instead, update flush_new_cached_update() to handle
this by simply checking for the key in the btree and only perform
the flush when a backing key is not present.
Signed-off-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>