writeback: use a per-sb counter to drain inode wb switches at umount

Tracking in-flight inode wb switches with a single global counter
(isw_nr_in_flight) plus a synchronize_rcu() based wait in
cgroup_writeback_umount() forces every umount to take a global hit
whenever any other superblock on the system has wb switches in flight,
even if the superblock being unmounted has none of its own.

Replace the global synchronize_rcu()/flush_workqueue() pair with a
per-sb counter, s_isw_nr_in_flight, plus three small helpers:

  - cgroup_writeback_pin(sb)   - increment counter
  - cgroup_writeback_unpin(sb) - decrement and wake drainer if last
  - cgroup_writeback_drain(sb) - wait for counter to reach zero

The wiring is:

  - inode_prepare_wbs_switch() pins before checking SB_ACTIVE and
    grabbing the inode; failure paths unpin before returning.  A
    lockless SB_ACTIVE check at the top of the function lets us skip
    the atomic_inc/smp_mb dance once SB_ACTIVE has been cleared (it
    is monotonic and never set back).
  - process_inode_switch_wbs() unpins after the matching iput().
  - cgroup_writeback_umount() drains the per-sb counter via
    wait_var_event().

The smp_mb() pair between inode_prepare_wbs_switch() and
cgroup_writeback_umount() keeps the SB_ACTIVE / counter ordering:
either the umounter sees a non-zero counter and waits, or the
switcher sees SB_ACTIVE cleared and aborts before grabbing the
inode.

The global isw_nr_in_flight is left in place, since it is still used
to throttle in-flight switches via WB_FRN_MAX_IN_FLIGHT.

The rcu_read_lock() extension in inode_switch_wbs() and
cleanup_offline_cgwb() that the race fix added is no longer needed
and is reverted; the synchronize_rcu() that the race fix added to
cgroup_writeback_umount() is dropped as well.

The following numbers were measured on a 16 vCPU QEMU guest with 4
background superblocks each churning "create memcg -> write 1 MiB ->
rmdir memcg" to keep the global isw_nr_in_flight non-zero.  Latencies
are wall-clock around umount(8); only the target sb's umount is
measured.

Target sb runs its own cgwb churn:

                              p50      p95      p99      max
  global synchronize_rcu()   67.6 ms  88.3 ms  88.3 ms  96.8 ms
  per-sb counter (this)       7.9 ms  10.0 ms  10.0 ms  10.1 ms

Idle target umount latency under cross-sb cgwb-switch pressure:

                              p50      p95      p99      max
  global synchronize_rcu()   62.7 ms  95.4 ms 108.1 ms 108.6 ms
  per-sb counter (this)       5.3 ms   6.9 ms   7.4 ms   7.4 ms
  no-pressure baseline        4.9 ms   5.9 ms   6.3 ms   6.7 ms

8 concurrent umounts of idle sbs under the same pressure:

                              p50      p95      max
  global synchronize_rcu()   61.3 ms  99.5 ms 113.7 ms
  per-sb counter (this)       8.1 ms   9.1 ms   9.5 ms

In-kernel cgroup_writeback_umount() time across the same run
(bpftrace, ~340 calls covering all scenarios):

  global synchronize_rcu()    12371 ms total (~36 ms / call)
  per-sb counter (this)        1.37 ms total ( ~4 us / call)

Suggested-by: Christian Brauner <brauner@kernel.org>
Link: https://lore.kernel.org/r/177910456953.488929.2169908940676707307.b4-review@b4
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Baokun Li <libaokun@linux.alibaba.com>
Link: https://patch.msgid.link/20260521095016.2791354-4-libaokun@linux.alibaba.com
Acked-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>

0 files changed, 0 insertions, 0 deletions