ntfs: read WOF chunks outside the decompression lock

WOF decompression uses four module-global workspaces, one per compression
format, each with a static mutex.  ntfs_read_wof_compressed_block() takes
that mutex once and holds it across the whole chunk loop, so both block
reads run inside it:

  mutex_lock(ws->lock);
  for each chunk {
          parse_wof_chunk_table(..., ws->input, ...);   /* reads disk */
          ntfs_read_wof_chunk(..., ws->input, ...);     /* reads disk */
          decompress into ws->output;
  }
  mutex_unlock(ws->lock);

Readers of system-compressed files then serialise system-wide on the disk
waits, not just on the decompressor scratch the lock exists for.  One
reader sleeping in submit_bio_wait() blocks all the rest.

The waits dominate.  Reading an 8 MiB xpress4k file (2048 chunks at a 48%
compressed ratio, so 2048 acquisitions and 4096 block reads) and timing
ws->lock against the part of it spent in ntfs_bdev_read():

  backing store                held     of that in I/O   held after
  virtio, host page cache      348 ms   321 ms  (92%)    24.6 ms
  virtio, throttled 100 MB/s   978 ms   948 ms  (96%)    36.6 ms

The page-cache row is a lower bound, having no seek cost at all, and the
share still grows with slower storage because only the wait scales while
decompression stays near 26 ms.

The reads are inside the lock only because they land in ws->input, a
buffer shared through the workspace.  Nothing else requires it:
parse_wof_chunk_table() and ntfs_read_wof_chunk() already take the buffer
as a parameter and both set *chunk_mem to a pointer inside it, so a
caller-owned buffer works unchanged.

Allocate that buffer per call, do both reads without the lock, and take
the lock only around decompression, which is the step needing ws->output
and ws->scratch.  squashfs is arranged this way already: its
squashfs_decompress() is handed a bio that has been read, and locks only
for the CPU work.

Block reads are unchanged in number, they just no longer run under the
lock, and hold time stops tracking device speed.

This also unnests two per-inode locks from the global one, runlist->lock
taken by both reads and base_ni->mrec_lock taken for a resident stream.
A resident chunk needs no I/O at all, yet used to queue behind a reader
blocked in submit_bio_wait() and then take mrec_lock inside the global
mutex.

The buffer is 4608 bytes for xpress4k and at most 33280 for lzx32k.  This
path already does GFP_NOFS allocations per call in ntfs_attr_iget(), and
in ntfs_attr_get_search_ctx() for a resident stream, so one more does not
change how it behaves under memory pressure.  The workspace keeps output
and scratch, 4 KiB to 32 KiB and 6224 bytes (xpress) or 10240 (lzx), and
its "already allocated" test moves from ws->input to ws->output.

The lock is now taken per chunk rather than per call, which differs only
for a folio spanning several chunks: a few more uncontended mutex
operations in exchange for not holding it across the reads between them.

Verified under QEMU against an uncompressed copy of the same data, on an
8 MiB file and a 100000 byte one, the latter covering the tail chunk that
is not a full comp_unit.

Signed-off-by: Zhan Xusheng <zhanxusheng@xiaomi.com>
Signed-off-by: Namjae Jeon <linkinjeon@kernel.org>

0 files changed, 0 insertions, 0 deletions