The total_bytes_pinned counter is completely broken when accounting
delayed refs:

- If two drops for the same extent are merged, we will decrement
  total_bytes_pinned twice but only increment it once.
- If an add is merged into a drop or vice versa, we will decrement the
  total_bytes_pinned counter but never increment it.
- If multiple references to an extent are dropped, we will account it
  multiple times, potentially vastly over-estimating the number of bytes
  that will be freed by a commit and doing unnecessary work when we're
  close to ENOSPC.

The last issue is relatively minor, but the first two make the
total_bytes_pinned counter leak or underflow very often. These
accounting issues were introduced in b150a4f1 ("Btrfs: use a percpu
to keep track of possibly pinned bytes"), but they were papered over by
zeroing out the counter on every commit until d288db5d ("Btrfs: fix
race of using total_bytes_pinned").

We need to make sure that an extent is accounted as pinned exactly once
if and only if we will drop references to it when when the transaction
is committed. Ideally we would only add to total_bytes_pinned when the
*last* reference is dropped, but this information isn't readily
available for data extents. Again, this over-estimation can lead to
extra commits when we're close to ENOSPC, but it's not as bad as before.

The fix implemented here is to increment total_bytes_pinned when the
total refmod count for an extent goes negative and decrement it if the
refmod count goes back to non-negative or after we've run all of the
delayed refs for that extent.
Signed-off-by: Omar Sandoval <osandov@fb.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Reviewed-by: Liu Bo <bo.li.liu@oracle.com>
Signed-off-by: David Sterba <dsterba@suse.com>