It's easy to get scsi_debug to error on throughput testing when we have multiple hosts: #lsscsi [7:0:0:0] disk Linux scsi_debug 0191 [0:0:0:0] disk Linux scsi_debug 0191 # fio --filename=/dev/sda --filename=/dev/sdb --direct=1 --rw=read --bs=4k --iodepth=256 --runtime=60 --numjobs=40 --time_based --name=jpg --eta-newline=1 --readonly --ioengine=io_uring --hipri --exitall_on_error jpg: (g=0): rw=read, bs=(R) 4096B-4096B, (W) 4096B-4096B, (T) 4096B-4096B, ioengine=io_uring, iodepth=256 ... fio-3.28 Starting 40 processes [ 27.521809] hrtimer: interrupt took 33067 ns [ 27.904660] sd 7:0:0:0: [sdb] tag#171 FAILED Result: hostbyte=DID_ABORT driverbyte=DRIVER_OK cmd_age=0s [ 27.904660] sd 0:0:0:0: [sda] tag#58 FAILED Result: hostbyte=DID_ABORT driverbyte=DRIVER_OK cmd_age=0s fio: io_u error [ 27.904667] sd 0:0:0:0: [sda] tag#58 CDB: Read(10) 28 00 00 00 27 00 00 01 18 00 on file /dev/sda[ 27.904670] sd 0:0:0:0: [sda] tag#62 FAILED Result: hostbyte=DID_ABORT driverbyte=DRIVER_OK cmd_age=0s The issue is related to how the driver manages submit queues and tags. A single array of submit queues - sdebug_q_arr - with its own set of tags is shared among all hosts. As such, for occasions when we have more than one host it is possible to overload the submit queues and run out of tags. Running out of tags makes the driver error in the submission path. We could change the driver error code in this scenario to request SCSI ml to requeue, but that is not a proper fix. A solution to solving this multi-host problem could be by making the submit queues per host. A downside there is that we make the driver considerably more complex again. There is another separate issue that we may reduce the host submit queue depth, sdebug_max_queue, dynamically causing the host to be overloaded. How many IOs which the host may be sent is fixed at can_queue at init time, which is the same initial value for sdebug_max_queue. So reducing sdebug_max_queue means that the host may be sent more IOs than it is configured to handle, causing overloading. About submit queues, they are not really required in the driver at all. blk-mq and the SCSI ml already manages tags for us, so that the LLD does not have to. The submit queues hold per-IO deferred command info in sdebug_defer, but this could be dynamically allocated and assigned on a per-IO basis. This series removes the scsi_debug submit queue concept and uses pre-existing APIs to manage and examine tags, like scsi_block_requests() and blk_mq_tagset_busy_iter(). Using standard APIs makes the driver more maintainable and extensible in future. The last patch is an RFC as the LLD should not be calling sbitmap_queue_resize() for the host tags. This series is based on my earlier series https://lore.kernel.org/linux-scsi/20230313093114.1498305-1-john.g.garry@xxxxxxxxxx/T/#m80f1de854ac590ce79c577e307ee2ba94a3534dd John Garry (9): scsi: scsi_debug: Don't iter all hosts in clear_luns_changed_on_target() scsi: scsi_debug: Change host list lock to a mutex scsi: scsi_debug: Protect block_unblock_all_queues() with mutex scsi: scsi_debug: Use scsi_block_requests() to block queues scsi: scsi_debug: Dynamically allocate sdebug_queued_cmd scsi: scsi_debug: Use blk_mq_tagset_busy_iter() in sdebug_blk_mq_poll() scsi: scsi_debug: Use blk_mq_tagset_busy_iter() in stop_all_queued() scsi: scsi_debug: Use scsi_host_busy() in delay_store() and ndelay_store() scsi: scsi_debug: Drop sdebug_queue concept drivers/scsi/scsi_debug.c | 767 +++++++++++++++++--------------------- 1 file changed, 342 insertions(+), 425 deletions(-) -- 2.35.3
