[PATCH v7 0/1] block/blk-mq: use atomic_t for quiesce_depth to avoid lock contention on RT

[Ionut Nechita (Wind River)]

From: Ionut Nechita <ionut.nechita@xxxxxxxxxxxxx>

Hi Jens,

This is v7 of the fix for the PREEMPT_RT performance regression caused by
commit 6bda857bcbb86 ("block: fix ordering between checking
QUEUE_FLAG_QUIESCED request adding").

Changes since v6 (May 6):
- Reader-side barrier in blk_mq_run_hw_queue() changed from smp_rmb() to
  smp_mb().  The race closed by commit 6bda857bcbb86 is a store-buffer
  pattern: one CPU inserts a request and then reads the quiesce state,
  another CPU unquiesces and then reads "has pending work".  A full
  barrier is needed on *both* sides, not just a read barrier on the
  reader, so smp_mb() now pairs with the existing writer-side
  smp_mb__after_atomic().  Thanks to Bart Van Assche for pointing out
  that smp_rmb() was insufficient.
- Rewrote the in-code comments and the commit message to spell out which
  ordering the removed q->queue_lock acquisitions provided and how it is
  preserved:
    * blk_mq_quiesce_queue_nowait(): the lock only published the state
      change; the actual visibility/drain guarantee comes from the
      synchronize_srcu()/synchronize_rcu() in blk_mq_wait_quiesce_done()
      that every caller invokes.  The smp_mb__after_atomic() is kept so
      the helper stays self-contained for the few callers that defer
      that wait.
    * blk_mq_unquiesce_queue(): write side of the store-buffer pattern,
      a full barrier between the quiesce_depth store and the
      blk_mq_hctx_has_pending() load (atomic_dec_if_positive() already
      orders the decrement on success; the barrier is spelled out for
      clarity).
    * blk_mq_run_hw_queue(): read side, a full barrier between the
      request insert and the quiesce-state re-check.
- Also note in the changelog that this is the memory-barrier alternative
  commit 6bda857bcbb86's own changelog described (and rejected as
  "harder to maintain"), and that making quiesce_depth atomic_t turns
  the lockless blk_queue_quiesced() read into a clean atomic load
  instead of a plain-int read racing with a spin_lock-protected update.
- Rebased on linux-next (next-20260505).  No other code changes; the
  atomic_t conversion and removal of QUEUE_FLAG_QUIESCED are unchanged
  from v6.

Sebastian's Reviewed-by is carried over: the approach (atomic counter +
barrier instead of the spinlock) is the one he suggested and reviewed;
the only functional change in v7 is upgrading the reader-side barrier to
a full one.

Changes since v5 (Mar 3):
- Rewrote the memory-ordering comments per Bart Van Assche's review.
- Rebased on top of linux-next.  No code-generation changes.

The problem: on PREEMPT_RT, the spinlock_t q->queue_lock that commit
6bda857bcbb86 added to blk_mq_run_hw_queue() converts to a sleeping
rt_mutex.  blk_mq_run_hw_queue() runs from every MSI-X IRQ thread and
hits that lock on the common "nothing pending" path, so all IRQ threads
serialise and go to D-state.  On a Broadcom/LSI MegaRAID 12GSAS/PCIe
Secure SAS39xx (megaraid_sas, 128 MSI-X vectors, 120 hw queues),
throughput drops from 640 MB/s to 153 MB/s.

The fix takes the memory-barrier alternative and folds the quiesce
indicator into quiesce_depth itself: quiesce_depth becomes atomic_t,
QUEUE_FLAG_QUIESCED goes away, and no lock is left on the dispatch hot
path.

v6: https://lore.kernel.org/linux-block/cover.1778048987.git.ionut.nechita@xxxxxxxxxxxxx/

Ionut Nechita (1):
  block/blk-mq: use atomic_t for quiesce_depth to avoid lock contention
    on RT

 block/blk-core.c       |  1 +
 block/blk-mq-debugfs.c |  1 -
 block/blk-mq.c         | 69 ++++++++++++++++++++++++++----------------
 include/linux/blkdev.h |  9 ++++--
 4 files changed, 50 insertions(+), 30 deletions(-)

-- 
2.54.0