[PATCH 0/2] scsi: mpi3mr: Stop IRQ polling when no reply is ready

[Usama Arif <usama.arif@linux.dev>]

Once a reply queue has more than 8 I/Os pending, mpi3mr's IRQ thread
polls it every ~20us until nothing is pending or it has processed
max_host_ios replies. On busy HDDs that means about 45 seconds of
polling at a time, with nearly every wakeup finding nothing. On Meta's
HDD storage hosts, these threads used about 15% of the non-idle CPU
on polling.

Patch 2 goes back to interrupts once a poll finds the queue empty while
owning it. It relies on patch 1: an interrupt that finds the queue owned
by another context now hands it to the thread, instead of leaving a
reply until the next interrupt or a command timeout.

On a machine configured like those hosts (36 SATA HDDs behind a SAS40xx
HBA), HDD random reads now wake the threads twice per I/O instead of 236
times, and a co-located CPU-bound task loses 0.6% instead of 30%.
Drive-cache reads still reach 270K IOPS, and a single reply queue whose
CPU is saturated loses about 5%. With a kthread injected to hold a reply
queue, replies waited up to 50ms, and timed out under light load,
without patch 1, and at most 233us with it.

Based on mkp/scsi 7.4/scsi-staging (f09d2c7485b32).
 
Usama Arif (2):
  scsi: mpi3mr: Poll a reply queue that an interrupt found busy
  scsi: mpi3mr: Stop IRQ polling when no reply is ready

 drivers/scsi/mpi3mr/mpi3mr.h    |  2 +-
 drivers/scsi/mpi3mr/mpi3mr_fw.c | 65 ++++++++++++++++++++++++---------
 drivers/scsi/mpi3mr/mpi3mr_os.c |  2 +-
 3 files changed, 50 insertions(+), 19 deletions(-)


base-commit: f09d2c7485b32adb82336d0d748935c8237a649e
-- 
2.53.0-Meta