On Fri, Jan 12, 2024 at 02:34:52PM -0500, Ewan Milne wrote: > On Fri, Jan 12, 2024 at 7:43 AM Ming Lei <ming.lei@xxxxxxxxxx> wrote: > > > > On Fri, Jan 12, 2024 at 12:12:57PM +0100, Hannes Reinecke wrote: > > > On 1/12/24 08:00, Ming Lei wrote: > > > > Inside scsi_eh_wakeup(), scsi_host_busy() is called & checked with host lock > > > > every time for deciding if error handler kthread needs to be waken up. > > > > > > > > This way can be too heavy in case of recovery, such as: > > > > > > > > - N hardware queues > > > > - queue depth is M for each hardware queue > > > > - each scsi_host_busy() iterates over (N * M) tag/requests > > > > > > > > If recovery is triggered in case that all requests are in-flight, each > > > > scsi_eh_wakeup() is strictly serialized, when scsi_eh_wakeup() is called > > > > for the last in-flight request, scsi_host_busy() has been run for (N * M - 1) > > > > times, and request has been iterated for (N*M - 1) * (N * M) times. > > > > > > > > If both N and M are big enough, hard lockup can be triggered on acquiring > > > > host lock, and it is observed on mpi3mr(128 hw queues, queue depth 8169). > > > > > > > > Fix the issue by calling scsi_host_busy() outside host lock, and we > > > > don't need host lock for getting busy count because host lock never > > > > covers that. > > > > > > > Can you share details for the hard lockup? > > > I do agree that scsi_host_busy() is an expensive operation, so it > > > might not be ideal to call it under a spin lock. > > > But I wonder where the lockup comes in here. > > > Care to explain? > > > > Recovery happens when there is N * M inflight requests, then scsi_dec_host_busy() > > can be called for each inflight request/scmnd from irq context. > > > > host lock serializes every scsi_eh_wakeup(). > > > > Given each hardware queue has its own irq handler, so there could be one > > request, scsi_dec_host_busy() is called and the host lock is spinned until > > it is released from scsi_dec_host_busy() for all requests from all other > > hardware queues. > > > > The spin time can be long enough to trigger the hard lockup if N and M > > is big enough, and the total wait time can be: > > > > (N - 1) * M * time_taken_in_scsi_host_busy(). > > > > Meantime the same story happens on scsi_eh_inc_host_failed() which is > > called from softirq context, so host lock spin can be much more worse. > > > > It is observed on mpi3mr with 128(N) hw queues and 8169(M) queue depth. > > > > > > > > And if it leads to a lockup, aren't other instances calling scsi_host_busy() > > > under a spinlock affected, as well? > > > > It is only possible when it is called in per-command situation. > > > > > > Thanks, > > Ming > > > > I can't see why this wouldn't work, or cause a problem with a lost wakeup, > but the cost of iterating to obtain the host_busy value is still being paid, > just outside the host_lock. If this has triggered a hard lockup, should > we revisit the algorithm, e.g. are we still delaying EH wakeup for a noticeable > amount of time? SCSI EH is designed to start handling until all in-flight commands are failed, so it waits until all requests are failed first. > O(n^2) algorithms in the kernel don't seem like the best idea. It is actually O(n) because each hardware queue handles request in parallel. It is degraded to O(n^2) or O(n * m) just because of shared host lock. Single or N scsi_host_busy() won't take too long without host lock, what matters is actually the per-host lock spin time which can be accumulated as too big. > > In any case... > Reviewed-by: Ewan D. Milne <emilne@xxxxxxxxxx> Thanks for the review! -- Ming
