btrfs: inhibit extent buffer writeback to prevent COW amplification

Inhibit writeback on COW'd extent buffers for the lifetime of the
transaction handle, preventing background writeback from setting
BTRFS_HEADER_FLAG_WRITTEN and causing unnecessary re-COW.

COW amplification occurs when background writeback flushes an extent
buffer that a transaction handle is still actively modifying. When
lock_extent_buffer_for_io() transitions a buffer from dirty to
writeback, it sets BTRFS_HEADER_FLAG_WRITTEN, marking the block as
having been persisted to disk at its current bytenr. Once WRITTEN is
set, should_cow_block() must either COW the block again or overwrite
it in place, both of which are unnecessary overhead when the buffer
is still being modified by the same handle that allocated it. By
inhibiting background writeback on actively-used buffers, WRITTEN is
never set while a transaction handle holds a reference to the buffer,
avoiding this overhead entirely.

Add an atomic_t writeback_inhibitors counter to struct extent_buffer,
which fits in an existing 6-byte hole without increasing struct size.
When a buffer is COW'd in btrfs_force_cow_block(), call
btrfs_inhibit_eb_writeback() to store the eb in the transaction
handle's writeback_inhibited_ebs xarray (keyed by eb->start), take a
reference, and increment writeback_inhibitors. The function handles
dedup (same eb inhibited twice by the same handle) and replacement
(different eb at the same logical address). Allocation failure is
graceful: the buffer simply falls back to the pre-existing behavior
where it may be written back and re-COW'd.

Also inhibit writeback in should_cow_block() when COW is skipped,
so that every transaction handle that reuses an already-COW'd buffer
also inhibits its writeback. Without this, if handle A COWs a block
and inhibits it, and handle B later reuses the same block without
inhibiting, handle A's uninhibit on end_transaction leaves the buffer
unprotected while handle B is still using it. This ensures all handles
that access a COW'd buffer contribute to the inhibitor count, and the
buffer remains protected until the last handle releases it.

In lock_extent_buffer_for_io(), when writeback_inhibitors is non-zero
and the writeback mode is WB_SYNC_NONE, skip the buffer. WB_SYNC_NONE
is used by the VM flusher threads for background and periodic
writeback, which are the only paths that cause COW amplification by
opportunistically writing out dirty extent buffers mid-transaction.
Skipping these is safe because the buffers remain dirty in the page
cache and will be written out at transaction commit time.

WB_SYNC_ALL must always proceed regardless of writeback_inhibitors.
This is required for correctness in the fsync path: btrfs_sync_log()
writes log tree blocks via filemap_fdatawrite_range() (WB_SYNC_ALL)
while the transaction handle that inhibited those same blocks is still
active. Without the WB_SYNC_ALL bypass, those inhibited log tree
blocks would be silently skipped, resulting in an incomplete log on
disk and corruption on replay. btrfs_write_and_wait_transaction()
also uses WB_SYNC_ALL via filemap_fdatawrite_range(); for that path,
inhibitors are already cleared beforehand, but the bypass ensures
correctness regardless.

Uninhibit in __btrfs_end_transaction() before atomic_dec(num_writers)
to prevent a race where the committer proceeds while buffers are still
inhibited. Also uninhibit in btrfs_commit_transaction() before writing
and in cleanup_transaction() for the error path.

Reviewed-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: Sun YangKai <sunk67188@gmail.com>
Reviewed-by: Boris Burkov <boris@bur.io>
Signed-off-by: Leo Martins <loemra.dev@gmail.com>
Signed-off-by: David Sterba <dsterba@suse.com>