The IOCB_DONTCACHE writeback path in generic_write_sync() calls
filemap_flush_range() on every write, submitting writeback inline in
the writer's context. Perf lock contention profiling shows the
performance problem is not lock contention but the writeback submission
work itself — walking the page tree and submitting I/O blocks the writer
for milliseconds, inflating p99.9 latency from 23ms (buffered) to 93ms
(dontcache).
Replace the inline filemap_flush_range() call with a flusher kick that
drains dirty pages in the background. This moves writeback submission
completely off the writer's hot path.
To avoid flushing unrelated buffered dirty data, add a dedicated
WB_start_dontcache bit and wb_check_start_dontcache() handler that uses
the per-wb WB_DONTCACHE_DIRTY counter to determine how many pages to
write back. The flusher writes back that many pages from the oldest dirty
inodes (not restricted to dontcache-specific inodes). This helps
preserve I/O batching while limiting the scope of expedited writeback.
Like WB_start_all, the WB_start_dontcache bit coalesces multiple
DONTCACHE writes into a single flusher wakeup without per-write
allocations. Use test_and_clear_bit to atomically consume the kick
request before reading the dirty counter and starting writeback, so that
concurrent DONTCACHE writes during writeback can re-set the bit and
schedule a follow-up flusher run.
Read the dirty counter with wb_stat_sum() (aggregating per-CPU batches)
rather than wb_stat() (which reads only the global counter) to ensure
small writes below the percpu batch threshold are visible to the flusher.
In filemap_dontcache_kick_writeback(), set the WB_start_dontcache bit
inside the unlocked_inode_to_wb_begin/end section for correct cgroup
writeback domain targeting, but defer the wb_wakeup() call until after
the section ends, since wb_wakeup() uses spin_unlock_irq() which would
unconditionally re-enable interrupts while the i_pages xa_lock may still
be held under irqsave during a cgroup writeback switch. Pin the wb with
wb_get() inside the RCU critical section before calling wb_wakeup()
outside it, since cgroup bdi_writeback structures are RCU-freed and the
wb pointer could become invalid after unlocked_inode_to_wb_end() drops
the RCU read lock.
Also add WB_REASON_DONTCACHE as a new writeback reason for tracing
visibility.
dontcache-bench results (same host, T6F_SKL_1920GBF, 251 GiB RAM,
xfs on NVMe, fio io_uring):
Buffered and direct I/O paths are unaffected by this patchset. All
improvements are confined to the dontcache path:
Single-stream throughput (MB/s):
Before After Change
seq-write/dontcache 298 897 +201%
rand-write/dontcache 131 236 +80%
Tail latency improvements (seq-write/dontcache):
p99: 135,266 us -> 23,986 us (-82%)
p99.9: 8,925,479 us -> 28,443 us (-99.7%)
Multi-writer (4 jobs, sequential write):
Before After Change
dontcache aggregate (MB/s) 2,529 4,532 +79%
dontcache p99 (us) 8,553 1,002 -88%
dontcache p99.9 (us) 109,314 1,057 -99%
Dontcache multi-writer throughput now matches buffered (4,532 vs
4,616 MB/s).
32-file write (Axboe test):
Before After Change
dontcache aggregate (MB/s) 1,548 3,499 +126%
dontcache p99 (us) 10,170 602 -94%
Peak dirty pages (MB) 1,837 213 -88%
Dontcache now reaches 81% of buffered throughput (was 35%).
Competing writers (dontcache vs buffered, separate files):
Before After
buffered writer 868 433 MB/s
dontcache writer 415 433 MB/s
Aggregate 1,284 866 MB/s
Previously the buffered writer starved the dontcache writer 2:1.
With per-bdi_writeback tracking, both writers now receive equal
bandwidth. The aggregate matches the buffered-vs-buffered baseline
(863 MB/s), indicating fair sharing regardless of I/O mode.
The dontcache writer's p99.9 latency collapsed from 119 ms to
33 ms (-73%), eliminating the severe periodic stalls seen in the
baseline. Both writers now share identical latency profiles,
matching the buffered-vs-buffered pattern.
The per-bdi_writeback dirty tracking dramatically reduces peak dirty
pages in dontcache workloads, with the 32-file test dropping from
1.8 GB to 213 MB. Dontcache sequential write throughput triples and
multi-writer throughput reaches parity with buffered I/O, with tail
latencies collapsing by 1-2 orders of magnitude.
Assisted-by: Claude:claude-opus-4-6
Signed-off-by: Jeff Layton <jlayton@kernel.org>
Link: https://patch.msgid.link/20260511-dontcache-v7-3-2848ddce8090@kernel.org
Reviewed-by: Jan Kara <jack@suse.cz>
Reviewed-by: Ritesh Harjani (IBM) <ritesh.list@gmail.com>
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Jeff Layton
e1bf796284