On Wed, Apr 14, 2021 at 07:29:07PM +0530, Kashyap Desai wrote: > > > I tried both - 5.12.0-rc1 and 5.11.0-rc2+ and there is a same > behavior. > > > Let me also check megaraid_sas and see if anything generic or this is > > > a special case of scsi_debug. > > > > As I mentioned, it could be one generic issue wrt. SCHED_RESTART. > > shared tags might have to restart all hctx since all share same tags. > > Ming - I tried many combination on MR shared host tag driver but there is > no single instance of IO hang. > I will keep trying, but when I look at scsi_debug driver code I found > below odd settings in scsi_debug driver. > can_queue of adapter is set to 128 but queue_depth of sdev is set to 255. > > If I apply below patch, scsi_debug driver's hang is also resolved. Ideally > sdev->queue depth cannot exceed shost->can_queue. > Not sure why cmd_per_lun is 255 in scsi_debug driver which can easily > exceed can_queue. I will simulate something similar in MR driver and see > how it behaves w.r.t IO hang issue. > > diff --git a/drivers/scsi/scsi_debug.c b/drivers/scsi/scsi_debug.c > index 70165be10f00..dded762540ee 100644 > --- a/drivers/scsi/scsi_debug.c > +++ b/drivers/scsi/scsi_debug.c > @@ -218,7 +218,7 @@ static const char *sdebug_version_date = "20200710"; > */ > #define SDEBUG_CANQUEUE_WORDS 3 /* a WORD is bits in a long */ > #define SDEBUG_CANQUEUE (SDEBUG_CANQUEUE_WORDS * BITS_PER_LONG) > -#define DEF_CMD_PER_LUN 255 > +#define DEF_CMD_PER_LUN SDEBUG_CANQUEUE > > /* UA - Unit Attention; SA - Service Action; SSU - Start Stop Unit */ > #define F_D_IN 1 /* Data-in command (e.g. READ) */ > @@ -7558,6 +7558,7 @@ static int sdebug_driver_probe(struct device *dev) > sdbg_host = to_sdebug_host(dev); > > sdebug_driver_template.can_queue = sdebug_max_queue; > + sdebug_driver_template.cmd_per_lun = sdebug_max_queue; > if (!sdebug_clustering) > sdebug_driver_template.dma_boundary = PAGE_SIZE - 1; Yeah, the change makes the issue disappear. That looks scsi's restart (scsi_run_queue_async<-scsi_end_request) finally restarts the queue. Thanks, Ming