Skip to contentSkip to conversation

scsi_debug: add new defer_type for mq_poll

PATCH In review Douglas Gilbert 2 participants 7 messages Feb 9, 2021 – Feb 12, 2021 SCSI

lore
Share & tools
lore
#0 lore raw

PATCH v3 1/1 scsi_debug: add new defer_type for mq_poll

Add a new sdeb_defer_type enumeration: SDEB_DEFER_POLL for requests that have REQ_HIPRI set in cmd_flags field. It is expected that these requests will be polled via the mq_poll entry point which is driven by calls to blk_poll() in the block layer. Therefore timer events are not 'wired up' in the normal fashion.

There are still cases with short delays (e.g. < 10 microseconds) where by the time the command response processing occurs, the delay is already exceeded in which case the code calls scsi_done() directly. In such cases there is no window for mq_poll() to be called.

Add 'mq_polls' counter that increments on each scsi_done() called via the mq_poll entry point. Can be used to show (with 'cat /proc/scsi/scsi_debug/<host_id>') that blk_poll() is causing completions rather than some other mechanism.

This patch is against the 5.12/scsi-staging branch which includes
a98d6bdf181eb71bf4686666eaf47c642a061642
scsi_debug : iouring iopoll support
which it alters. So this patch is a fix of that patch.

Changes since version 2 [sent 20210206 to linux-scsi list]
- the sdebug_blk_mq_poll() callback didn't cope with the

    uncommon case where sqcp->sd_dp is NULL. Fix.

Changes since version 1 [sent 20210201 to linux-scsi list]
- harden SDEB_DEFER_POLL which broke under testing
- add mq_polls counter for debug output

---
 drivers/scsi/scsi_debug.c | 141 +++++++++++++++++++++++++-------------
 1 file changed, 94 insertions(+), 47 deletions(-)
drivers/scsi/scsi_debug.c +94 -47
index 746eec521f79..b9ec3a47fbf1 100644
@@ -322,17 +322,19 @@ struct sdeb_store_info {
 	container_of(d, struct sdebug_host_info, dev)
 
 enum sdeb_defer_type {SDEB_DEFER_NONE = 0, SDEB_DEFER_HRT = 1,
-		      SDEB_DEFER_WQ = 2};
+		      SDEB_DEFER_WQ = 2, SDEB_DEFER_POLL = 3};
 
 struct sdebug_defer {
 	struct hrtimer hrt;
 	struct execute_work ew;
+	ktime_t cmpl_ts;/* time since boot to complete this cmd */
 	int sqa_idx;	/* index of sdebug_queue array */
 	int qc_idx;	/* index of sdebug_queued_cmd array within sqa_idx */
 	int hc_idx;	/* hostwide tag index */
 	int issuing_cpu;
 	bool init_hrt;
 	bool init_wq;
+	bool init_poll;
 	bool aborted;	/* true when blk_abort_request() already called */
 	enum sdeb_defer_type defer_t;
 };
@@ -357,6 +359,7 @@ static atomic_t sdebug_completions;  /* count of deferred completions */
 static atomic_t sdebug_miss_cpus;    /* submission + completion cpus differ */
 static atomic_t sdebug_a_tsf;	     /* 'almost task set full' counter */
 static atomic_t sdeb_inject_pending;
+static atomic_t sdeb_mq_poll_count;  /* bumped when mq_poll returns > 0 */
 
 struct opcode_info_t {
 	u8 num_attached;	/* 0 if this is it (i.e. a leaf); use 0xff */
@@ -4730,7 +4733,6 @@ static void sdebug_q_cmd_complete(struct sdebug_defer *sd_dp)
 	struct scsi_cmnd *scp;
 	struct sdebug_dev_info *devip;
 
-	sd_dp->defer_t = SDEB_DEFER_NONE;
 	if (unlikely(aborted))
 		sd_dp->aborted = false;
 	qc_idx = sd_dp->qc_idx;
@@ -4745,6 +4747,7 @@ static void sdebug_q_cmd_complete(struct sdebug_defer *sd_dp)
 		return;
 	}
 	spin_lock_irqsave(&sqp->qc_lock, iflags);
+	sd_dp->defer_t = SDEB_DEFER_NONE;
 	sqcp = &sqp->qc_arr[qc_idx];
 	scp = sqcp->a_cmnd;
 	if (unlikely(scp == NULL)) {
@@ -5517,40 +5520,66 @@ static int schedule_resp(struct scsi_cmnd *cmnd, struct sdebug_dev_info *devip,
 				kt -= d;
 			}
 		}
-		if (!sd_dp->init_hrt) {
-			sd_dp->init_hrt = true;
-			sqcp->sd_dp = sd_dp;
-			hrtimer_init(&sd_dp->hrt, CLOCK_MONOTONIC,
-				     HRTIMER_MODE_REL_PINNED);
-			sd_dp->hrt.function = sdebug_q_cmd_hrt_complete;
-			sd_dp->sqa_idx = sqp - sdebug_q_arr;
-			sd_dp->qc_idx = k;
+		sd_dp->cmpl_ts = ktime_add(ns_to_ktime(ns_from_boot), kt);
+		if (cmnd->request->cmd_flags & REQ_HIPRI) {
+			spin_lock_irqsave(&sqp->qc_lock, iflags);
+			if (!sd_dp->init_poll) {
+				sd_dp->init_poll = true;
+				sqcp->sd_dp = sd_dp;
+				sd_dp->sqa_idx = sqp - sdebug_q_arr;
+				sd_dp->qc_idx = k;
+			}
+			sd_dp->defer_t = SDEB_DEFER_POLL;
+			spin_unlock_irqrestore(&sqp->qc_lock, iflags);
+		} else {
+			if (!sd_dp->init_hrt) {
+				sd_dp->init_hrt = true;
+				sqcp->sd_dp = sd_dp;
+				hrtimer_init(&sd_dp->hrt, CLOCK_MONOTONIC,
+					     HRTIMER_MODE_REL_PINNED);
+				sd_dp->hrt.function = sdebug_q_cmd_hrt_complete;
+				sd_dp->sqa_idx = sqp - sdebug_q_arr;
+				sd_dp->qc_idx = k;
+			}
+			sd_dp->defer_t = SDEB_DEFER_HRT;
+			/* schedule the invocation of scsi_done() for a later time */
+			hrtimer_start(&sd_dp->hrt, kt, HRTIMER_MODE_REL_PINNED);
 		}
 		if (sdebug_statistics)
 			sd_dp->issuing_cpu = raw_smp_processor_id();
-		sd_dp->defer_t = SDEB_DEFER_HRT;
-		/* schedule the invocation of scsi_done() for a later time */
-		hrtimer_start(&sd_dp->hrt, kt, HRTIMER_MODE_REL_PINNED);
 	} else {	/* jdelay < 0, use work queue */
-		if (!sd_dp->init_wq) {
-			sd_dp->init_wq = true;
-			sqcp->sd_dp = sd_dp;
-			sd_dp->sqa_idx = sqp - sdebug_q_arr;
-			sd_dp->qc_idx = k;
-			INIT_WORK(&sd_dp->ew.work, sdebug_q_cmd_wq_complete);
-		}
-		if (sdebug_statistics)
-			sd_dp->issuing_cpu = raw_smp_processor_id();
-		sd_dp->defer_t = SDEB_DEFER_WQ;
 		if (unlikely((sdebug_opts & SDEBUG_OPT_CMD_ABORT) &&
 			     atomic_read(&sdeb_inject_pending)))
 			sd_dp->aborted = true;
-		schedule_work(&sd_dp->ew.work);
-		if (unlikely((sdebug_opts & SDEBUG_OPT_CMD_ABORT) &&
-			     atomic_read(&sdeb_inject_pending))) {
+		sd_dp->cmpl_ts = ns_to_ktime(ns_from_boot);
+		if (cmnd->request->cmd_flags & REQ_HIPRI) {
+			spin_lock_irqsave(&sqp->qc_lock, iflags);
+			if (!sd_dp->init_poll) {
+				sd_dp->init_poll = true;
+				sqcp->sd_dp = sd_dp;
+				sd_dp->sqa_idx = sqp - sdebug_q_arr;
+				sd_dp->qc_idx = k;
+			}
+			sd_dp->defer_t = SDEB_DEFER_POLL;
+			spin_unlock_irqrestore(&sqp->qc_lock, iflags);
+		} else {
+			if (!sd_dp->init_wq) {
+				sd_dp->init_wq = true;
+				sqcp->sd_dp = sd_dp;
+				sd_dp->sqa_idx = sqp - sdebug_q_arr;
+				sd_dp->qc_idx = k;
+				INIT_WORK(&sd_dp->ew.work, sdebug_q_cmd_wq_complete);
+			}
+			sd_dp->defer_t = SDEB_DEFER_WQ;
+			schedule_work(&sd_dp->ew.work);
+		}
+		if (sdebug_statistics)
+			sd_dp->issuing_cpu = raw_smp_processor_id();
+		if (unlikely(sd_dp->aborted)) {
 			sdev_printk(KERN_INFO, sdp, "abort request tag %d\n", cmnd->request->tag);
 			blk_abort_request(cmnd->request);
 			atomic_set(&sdeb_inject_pending, 0);
+			sd_dp->aborted = false;
 		}
 	}
 	if (unlikely((SDEBUG_OPT_Q_NOISE & sdebug_opts) && scsi_result == device_qfull_result))
@@ -5779,11 +5808,12 @@ static int scsi_debug_show_info(struct seq_file *m, struct Scsi_Host *host)
 		   dix_reads, dix_writes, dif_errors);
 	seq_printf(m, "usec_in_jiffy=%lu, statistics=%d\n", TICK_NSEC / 1000,
 		   sdebug_statistics);
-	seq_printf(m, "cmnd_count=%d, completions=%d, %s=%d, a_tsf=%d\n",
+	seq_printf(m, "cmnd_count=%d, completions=%d, %s=%d, a_tsf=%d, mq_polls=%d\n",
 		   atomic_read(&sdebug_cmnd_count),
 		   atomic_read(&sdebug_completions),
 		   "miss_cpus", atomic_read(&sdebug_miss_cpus),
-		   atomic_read(&sdebug_a_tsf));
+		   atomic_read(&sdebug_a_tsf),
+		   atomic_read(&sdeb_mq_poll_count));
 
 	seq_printf(m, "submit_queues=%d\n", submit_queues);
 	for (j = 0, sqp = sdebug_q_arr; j < submit_queues; ++j, ++sqp) {
@@ -7246,52 +7276,68 @@ static int sdebug_map_queues(struct Scsi_Host *shost)
 
 static int sdebug_blk_mq_poll(struct Scsi_Host *shost, unsigned int queue_num)
 {
-	int qc_idx;
-	int retiring = 0;
+	bool first;
+	bool retiring = false;
+	int num_entries = 0;
+	unsigned int qc_idx = 0;
 	unsigned long iflags;
+	ktime_t kt_from_boot = ktime_get_boottime();
 	struct sdebug_queue *sqp;
 	struct sdebug_queued_cmd *sqcp;
 	struct scsi_cmnd *scp;
 	struct sdebug_dev_info *devip;
-	int num_entries = 0;
+	struct sdebug_defer *sd_dp;
 
 	sqp = sdebug_q_arr + queue_num;
+	spin_lock_irqsave(&sqp->qc_lock, iflags);
 
-	do {
-		spin_lock_irqsave(&sqp->qc_lock, iflags);
-		qc_idx = find_first_bit(sqp->in_use_bm, sdebug_max_queue);
-		if (unlikely((qc_idx < 0) || (qc_idx >= sdebug_max_queue)))
-			goto out;
+	for (first = true; first || qc_idx + 1 < sdebug_max_queue; )   {
+		if (first) {
+			qc_idx = find_first_bit(sqp->in_use_bm, sdebug_max_queue);
+			first = false;
+		} else {
+			qc_idx = find_next_bit(sqp->in_use_bm, sdebug_max_queue, qc_idx + 1);
+		}
+		if (unlikely(qc_idx >= sdebug_max_queue))
+			break;
 
 		sqcp = &sqp->qc_arr[qc_idx];
+		sd_dp = sqcp->sd_dp;
+		if (unlikely(!sd_dp))
+			continue;
 		scp = sqcp->a_cmnd;
 		if (unlikely(scp == NULL)) {
-			pr_err("scp is NULL, queue_num=%d, qc_idx=%d from %s\n",
+			pr_err("scp is NULL, queue_num=%d, qc_idx=%u from %s\n",
 			       queue_num, qc_idx, __func__);
-			goto out;
+			break;
 		}
+		if (sd_dp->defer_t == SDEB_DEFER_POLL) {
+			if (kt_from_boot < sd_dp->cmpl_ts)
+				continue;
+
+		} else		/* ignoring non REQ_HIPRI requests */
+			continue;
 		devip = (struct sdebug_dev_info *)scp->device->hostdata;
 		if (likely(devip))
 			atomic_dec(&devip->num_in_q);
 		else
 			pr_err("devip=NULL from %s\n", __func__);
 		if (unlikely(atomic_read(&retired_max_queue) > 0))
-			retiring = 1;
+			retiring = true;
 
 		sqcp->a_cmnd = NULL;
 		if (unlikely(!test_and_clear_bit(qc_idx, sqp->in_use_bm))) {
-			pr_err("Unexpected completion sqp %p queue_num=%d qc_idx=%d from %s\n",
+			pr_err("Unexpected completion sqp %p queue_num=%d qc_idx=%u from %s\n",
 				sqp, queue_num, qc_idx, __func__);
-			goto out;
+			break;
 		}
-
 		if (unlikely(retiring)) {	/* user has reduced max_queue */
 			int k, retval;
 
 			retval = atomic_read(&retired_max_queue);
 			if (qc_idx >= retval) {
 				pr_err("index %d too large\n", retval);
-				goto out;
+				break;
 			}
 			k = find_last_bit(sqp->in_use_bm, retval);
 			if ((k < sdebug_max_queue) || (k == retval))
@@ -7299,17 +7345,18 @@ static int sdebug_blk_mq_poll(struct Scsi_Host *shost, unsigned int queue_num)
 			else
 				atomic_set(&retired_max_queue, k + 1);
 		}
+		sd_dp->defer_t = SDEB_DEFER_NONE;
 		spin_unlock_irqrestore(&sqp->qc_lock, iflags);
 		scp->scsi_done(scp); /* callback to mid level */
+		spin_lock_irqsave(&sqp->qc_lock, iflags);
 		num_entries++;
-	} while (1);
-
-out:
+	}
 	spin_unlock_irqrestore(&sqp->qc_lock, iflags);
+	if (num_entries > 0)
+		atomic_add(num_entries, &sdeb_mq_poll_count);
 	return num_entries;
 }
 
-
 static int scsi_debug_queuecommand(struct Scsi_Host *shost,
 				   struct scsi_cmnd *scp)
 {
Signature · 1 line
2.25.1
#1 lore raw
Quoted 19 lines · Douglas Gilbert

Add a new sdeb_defer_type enumeration: SDEB_DEFER_POLL for requests that have REQ_HIPRI set in cmd_flags field. It is expected that these requests will be polled via the mq_poll entry point which is driven by calls to blk_poll() in the block layer. Therefore timer events are not 'wired up' in the normal fashion.

There are still cases with short delays (e.g. < 10 microseconds) where by the time the command response processing occurs, the delay is already exceeded in which case the code calls scsi_done() directly. In such cases there is no window for mq_poll() to be called.

Add 'mq_polls' counter that increments on each scsi_done() called via the mq_poll entry point. Can be used to show (with 'cat /proc/scsi/scsi_debug/<host_id>') that blk_poll() is causing completions rather than some other mechanism.

Quoted 8 lines

This patch is against the 5.12/scsi-staging branch which includes
a98d6bdf181eb71bf4686666eaf47c642a061642
scsi_debug : iouring iopoll support
which it alters. So this patch is a fix of that patch.

Changes since version 2 [sent 20210206 to linux-scsi list]
- the sdebug_blk_mq_poll() callback didn't cope with the

    uncommon case where sqcp->sd_dp is NULL. Fix.

Hi Doug, I tried this patch on top of below iouring patch series - "[v3,1/4] add io_uring with IOPOLL support in scsi layer"

After applying patch, I am seeing an IO hang issue - I see io_wqe_worker goes into a tight loop.

18210 root 20 0 1316348 3552 2376 R 99.1 0.0 0:24.09
fio 18303 root 20 0
0 0 0 D 78.8 0.0 0:01.75 io_wqe_worker-0

           18219 root      20   0       0      0      0 R  71.7   0.0

0:06.59 io_wqe_worker-0

I used below command -
insmod drivers/scsi/scsi_debug.ko dev_size_mb=1024 sector_size=512 add_host=24 per_host_store=1 ndelay=10000 host_max_queue=32 submit_queues=76 num_parts=1 poll_queues=8

and here is my fio script -
[global]
ioengine=io_uring
hipri=1
direct=1
runtime=30s
rw=randread
norandommap
bs=4k
iodepth=64

[seqprecon]
filename=/dev/sdd
[seqprecon]
filename=/dev/sde
[seqprecon]
filename=/dev/sdf

I will ask Martin to pick all the patches from "[v3,1/4] add io_uring with IOPOLL support in the scsi layer" except scsi_debug. We can work on scsi_debug and send standalone patch.

Kashyap

Quoted 275 lines

Changes since version 1 [sent 20210201 to linux-scsi list]
- harden SDEB_DEFER_POLL which broke under testing
- add mq_polls counter for debug output

---
 drivers/scsi/scsi_debug.c | 141 +++++++++++++++++++++++++-------------
 1 file changed, 94 insertions(+), 47 deletions(-)
drivers/scsi/scsi_debug.c +94 -48
index 746eec521f79..b9ec3a47fbf1 100644
@@ -322,17 +322,19 @@ struct sdeb_store_info {
        container_of(d, struct sdebug_host_info, dev)
 
 enum sdeb_defer_type {SDEB_DEFER_NONE = 0, SDEB_DEFER_HRT = 1,
-                     SDEB_DEFER_WQ = 2};
+                     SDEB_DEFER_WQ = 2, SDEB_DEFER_POLL = 3};
 
 struct sdebug_defer {
        struct hrtimer hrt;
        struct execute_work ew;
+       ktime_t cmpl_ts;/* time since boot to complete this cmd */
        int sqa_idx;    /* index of sdebug_queue array */
        int qc_idx;     /* index of sdebug_queued_cmd array within sqa_idx */
        int hc_idx;     /* hostwide tag index */
        int issuing_cpu;
        bool init_hrt;
        bool init_wq;
+       bool init_poll;
        bool aborted;   /* true when blk_abort_request() already called */
        enum sdeb_defer_type defer_t;
 };
@@ -357,6 +359,7 @@ static atomic_t sdebug_completions;  /* count of deferred completions */
 static atomic_t sdebug_miss_cpus;    /* submission + completion cpus differ */
 static atomic_t sdebug_a_tsf;       /* 'almost task set full' counter */
 static atomic_t sdeb_inject_pending;
+static atomic_t sdeb_mq_poll_count;  /* bumped when mq_poll returns > 0 */
 
 struct opcode_info_t {
        u8 num_attached;        /* 0 if this is it (i.e. a leaf); use 0xff */
@@ -4730,7 +4733,6 @@ static void sdebug_q_cmd_complete(struct sdebug_defer *sd_dp)
        struct scsi_cmnd *scp;
        struct sdebug_dev_info *devip;
 
-       sd_dp->defer_t = SDEB_DEFER_NONE;
        if (unlikely(aborted))
                sd_dp->aborted = false;
        qc_idx = sd_dp->qc_idx;
@@ -4745,6 +4747,7 @@ static void sdebug_q_cmd_complete(struct sdebug_defer *sd_dp)
                return;
        }
        spin_lock_irqsave(&sqp->qc_lock, iflags);
+       sd_dp->defer_t = SDEB_DEFER_NONE;
        sqcp = &sqp->qc_arr[qc_idx];
        scp = sqcp->a_cmnd;
        if (unlikely(scp == NULL)) {
@@ -5517,40 +5520,66 @@ static int schedule_resp(struct scsi_cmnd *cmnd, struct sdebug_dev_info *devip,
                                kt -= d;
                        }
                }
-               if (!sd_dp->init_hrt) {
-                       sd_dp->init_hrt = true;
-                       sqcp->sd_dp = sd_dp;
-                       hrtimer_init(&sd_dp->hrt, CLOCK_MONOTONIC,
-                                    HRTIMER_MODE_REL_PINNED);
-                       sd_dp->hrt.function = sdebug_q_cmd_hrt_complete;
-                       sd_dp->sqa_idx = sqp - sdebug_q_arr;
-                       sd_dp->qc_idx = k;
+               sd_dp->cmpl_ts = ktime_add(ns_to_ktime(ns_from_boot), kt);
+               if (cmnd->request->cmd_flags & REQ_HIPRI) {
+                       spin_lock_irqsave(&sqp->qc_lock, iflags);
+                       if (!sd_dp->init_poll) {
+                               sd_dp->init_poll = true;
+                               sqcp->sd_dp = sd_dp;
+                               sd_dp->sqa_idx = sqp - sdebug_q_arr;
+                               sd_dp->qc_idx = k;
+                       }
+                       sd_dp->defer_t = SDEB_DEFER_POLL;
+                       spin_unlock_irqrestore(&sqp->qc_lock, iflags);
+               } else {
+                       if (!sd_dp->init_hrt) {
+                               sd_dp->init_hrt = true;
+                               sqcp->sd_dp = sd_dp;
+                               hrtimer_init(&sd_dp->hrt, CLOCK_MONOTONIC,
+                                            HRTIMER_MODE_REL_PINNED);
+                               sd_dp->hrt.function = sdebug_q_cmd_hrt_complete;
+                               sd_dp->sqa_idx = sqp - sdebug_q_arr;
+                               sd_dp->qc_idx = k;
+                       }
+                       sd_dp->defer_t = SDEB_DEFER_HRT;
+                       /* schedule the invocation of scsi_done() for a later time */
+                       hrtimer_start(&sd_dp->hrt, kt, HRTIMER_MODE_REL_PINNED);
                }
                if (sdebug_statistics)
                        sd_dp->issuing_cpu = raw_smp_processor_id();
-               sd_dp->defer_t = SDEB_DEFER_HRT;
-               /* schedule the invocation of scsi_done() for a later time */
-               hrtimer_start(&sd_dp->hrt, kt, HRTIMER_MODE_REL_PINNED);
        } else {        /* jdelay < 0, use work queue */
-               if (!sd_dp->init_wq) {
-                       sd_dp->init_wq = true;
-                       sqcp->sd_dp = sd_dp;
-                       sd_dp->sqa_idx = sqp - sdebug_q_arr;
-                       sd_dp->qc_idx = k;
-                       INIT_WORK(&sd_dp->ew.work, sdebug_q_cmd_wq_complete);
-               }
-               if (sdebug_statistics)
-                       sd_dp->issuing_cpu = raw_smp_processor_id();
-               sd_dp->defer_t = SDEB_DEFER_WQ;
                if (unlikely((sdebug_opts & SDEBUG_OPT_CMD_ABORT) &&
                             atomic_read(&sdeb_inject_pending)))
                        sd_dp->aborted = true;
-               schedule_work(&sd_dp->ew.work);
-               if (unlikely((sdebug_opts & SDEBUG_OPT_CMD_ABORT) &&
-                            atomic_read(&sdeb_inject_pending))) {
+               sd_dp->cmpl_ts = ns_to_ktime(ns_from_boot);
+               if (cmnd->request->cmd_flags & REQ_HIPRI) {
+                       spin_lock_irqsave(&sqp->qc_lock, iflags);
+                       if (!sd_dp->init_poll) {
+                               sd_dp->init_poll = true;
+                               sqcp->sd_dp = sd_dp;
+                               sd_dp->sqa_idx = sqp - sdebug_q_arr;
+                               sd_dp->qc_idx = k;
+                       }
+                       sd_dp->defer_t = SDEB_DEFER_POLL;
+                       spin_unlock_irqrestore(&sqp->qc_lock, iflags);
+               } else {
+                       if (!sd_dp->init_wq) {
+                               sd_dp->init_wq = true;
+                               sqcp->sd_dp = sd_dp;
+                               sd_dp->sqa_idx = sqp - sdebug_q_arr;
+                               sd_dp->qc_idx = k;
+                               INIT_WORK(&sd_dp->ew.work, sdebug_q_cmd_wq_complete);
+                       }
+                       sd_dp->defer_t = SDEB_DEFER_WQ;
+                       schedule_work(&sd_dp->ew.work);
+               }
+               if (sdebug_statistics)
+                       sd_dp->issuing_cpu = raw_smp_processor_id();
+               if (unlikely(sd_dp->aborted)) {
                        sdev_printk(KERN_INFO, sdp, "abort request tag %d\n", cmnd->request->tag);
                        blk_abort_request(cmnd->request);
                        atomic_set(&sdeb_inject_pending, 0);
+                       sd_dp->aborted = false;
                }
        }
        if (unlikely((SDEBUG_OPT_Q_NOISE & sdebug_opts) && scsi_result == device_qfull_result))
@@ -5779,11 +5808,12 @@ static int scsi_debug_show_info(struct seq_file *m, struct Scsi_Host *host)
                   dix_reads, dix_writes, dif_errors);
        seq_printf(m, "usec_in_jiffy=%lu, statistics=%d\n", TICK_NSEC / 1000,
                   sdebug_statistics);
-       seq_printf(m, "cmnd_count=%d, completions=%d, %s=%d, a_tsf=%d\n",
+       seq_printf(m, "cmnd_count=%d, completions=%d, %s=%d, a_tsf=%d, mq_polls=%d\n",
                   atomic_read(&sdebug_cmnd_count),
                   atomic_read(&sdebug_completions),
                   "miss_cpus", atomic_read(&sdebug_miss_cpus),
-                  atomic_read(&sdebug_a_tsf));
+                  atomic_read(&sdebug_a_tsf),
+                  atomic_read(&sdeb_mq_poll_count));
 
        seq_printf(m, "submit_queues=%d\n", submit_queues);
        for (j = 0, sqp = sdebug_q_arr; j < submit_queues; ++j, ++sqp) {
@@ -7246,52 +7276,68 @@ static int sdebug_map_queues(struct Scsi_Host *shost)
 
 static int sdebug_blk_mq_poll(struct Scsi_Host *shost, unsigned int queue_num)
 {
-       int qc_idx;
-       int retiring = 0;
+       bool first;
+       bool retiring = false;
+       int num_entries = 0;
+       unsigned int qc_idx = 0;
        unsigned long iflags;
+       ktime_t kt_from_boot = ktime_get_boottime();
        struct sdebug_queue *sqp;
        struct sdebug_queued_cmd *sqcp;
        struct scsi_cmnd *scp;
        struct sdebug_dev_info *devip;
-       int num_entries = 0;
+       struct sdebug_defer *sd_dp;
 
        sqp = sdebug_q_arr + queue_num;
+       spin_lock_irqsave(&sqp->qc_lock, iflags);
 
-       do {
-               spin_lock_irqsave(&sqp->qc_lock, iflags);
-               qc_idx = find_first_bit(sqp->in_use_bm, sdebug_max_queue);
-               if (unlikely((qc_idx < 0) || (qc_idx >= sdebug_max_queue)))
-                       goto out;
+       for (first = true; first || qc_idx + 1 < sdebug_max_queue; )   {
+               if (first) {
+                       qc_idx = find_first_bit(sqp->in_use_bm, sdebug_max_queue);
+                       first = false;
+               } else {
+                       qc_idx = find_next_bit(sqp->in_use_bm, sdebug_max_queue, qc_idx + 1);
+               }
+               if (unlikely(qc_idx >= sdebug_max_queue))
+                       break;
 
                sqcp = &sqp->qc_arr[qc_idx];
+               sd_dp = sqcp->sd_dp;
+               if (unlikely(!sd_dp))
+                       continue;
                scp = sqcp->a_cmnd;
                if (unlikely(scp == NULL)) {
-                       pr_err("scp is NULL, queue_num=%d, qc_idx=%d from %s\n",
+                       pr_err("scp is NULL, queue_num=%d, qc_idx=%u from %s\n",
                               queue_num, qc_idx, __func__);
-                       goto out;
+                       break;
                }
+               if (sd_dp->defer_t == SDEB_DEFER_POLL) {
+                       if (kt_from_boot < sd_dp->cmpl_ts)
+                               continue;
+
+               } else          /* ignoring non REQ_HIPRI requests */
+                       continue;
                devip = (struct sdebug_dev_info *)scp->device->hostdata;
                if (likely(devip))
                        atomic_dec(&devip->num_in_q);
                else
                        pr_err("devip=NULL from %s\n", __func__);
                if (unlikely(atomic_read(&retired_max_queue) > 0))
-                       retiring = 1;
+                       retiring = true;
 
                sqcp->a_cmnd = NULL;
                if (unlikely(!test_and_clear_bit(qc_idx, sqp->in_use_bm))) {
-                       pr_err("Unexpected completion sqp %p queue_num=%d qc_idx=%d from %s\n",
+                       pr_err("Unexpected completion sqp %p queue_num=%d qc_idx=%u from %s\n",
                                sqp, queue_num, qc_idx, __func__);
-                       goto out;
+                       break;
                }
-
                if (unlikely(retiring)) {       /* user has reduced max_queue */
                        int k, retval;
 
                        retval = atomic_read(&retired_max_queue);
                        if (qc_idx >= retval) {
                                pr_err("index %d too large\n", retval);
-                               goto out;
+                               break;
                        }
                        k = find_last_bit(sqp->in_use_bm, retval);
                        if ((k < sdebug_max_queue) || (k == retval))
@@ -7299,17 +7345,18 @@ static int sdebug_blk_mq_poll(struct Scsi_Host *shost, unsigned int queue_num)
                        else
                                atomic_set(&retired_max_queue, k + 1);
                }
+               sd_dp->defer_t = SDEB_DEFER_NONE;
                spin_unlock_irqrestore(&sqp->qc_lock, iflags);
                scp->scsi_done(scp); /* callback to mid level */
+               spin_lock_irqsave(&sqp->qc_lock, iflags);
                num_entries++;
-       } while (1);
-
-out:
+       }
        spin_unlock_irqrestore(&sqp->qc_lock, iflags);
+       if (num_entries > 0)
+               atomic_add(num_entries, &sdeb_mq_poll_count);
        return num_entries;
 }
 
-
 static int scsi_debug_queuecommand(struct Scsi_Host *shost,
                                   struct scsi_cmnd *scp)
 {
--

2.25.1

#2 lore raw
Quoted 66 lines · Kashyap Desai
Quoted 19 lines · Douglas Gilbert

Add a new sdeb_defer_type enumeration: SDEB_DEFER_POLL for requests that have REQ_HIPRI set in cmd_flags field. It is expected that these requests will be polled via the mq_poll entry point which is driven by calls to blk_poll() in the block layer. Therefore timer events are not 'wired up' in the normal fashion.

There are still cases with short delays (e.g. < 10 microseconds) where by the time the command response processing occurs, the delay is already exceeded in which case the code calls scsi_done() directly. In such cases there is no window for mq_poll() to be called.

Add 'mq_polls' counter that increments on each scsi_done() called via the mq_poll entry point. Can be used to show (with 'cat /proc/scsi/scsi_debug/<host_id>') that blk_poll() is causing completions rather than some other mechanism.

Quoted 8 lines

This patch is against the 5.12/scsi-staging branch which includes
a98d6bdf181eb71bf4686666eaf47c642a061642
scsi_debug : iouring iopoll support
which it alters. So this patch is a fix of that patch.

Changes since version 2 [sent 20210206 to linux-scsi list]
- the sdebug_blk_mq_poll() callback didn't cope with the

     uncommon case where sqcp->sd_dp is NULL. Fix.

Hi Doug, I tried this patch on top of below iouring patch series - "[v3,1/4] add io_uring with IOPOLL support in scsi layer"

After applying patch, I am seeing an IO hang issue - I see io_wqe_worker goes into a tight loop.

18210 root 20 0 1316348 3552 2376 R 99.1 0.0 0:24.09
fio 18303 root 20 0
0 0 0 D 78.8 0.0 0:01.75 io_wqe_worker-0

            18219 root      20   0       0      0      0 R  71.7   0.0

0:06.59 io_wqe_worker-0

I used below command -
insmod drivers/scsi/scsi_debug.ko dev_size_mb=1024 sector_size=512 add_host=24 per_host_store=1 ndelay=10000 host_max_queue=32 submit_queues=76 num_parts=1 poll_queues=8

and here is my fio script -
[global]
ioengine=io_uring
hipri=1
direct=1
runtime=30s
rw=randread
norandommap
bs=4k
iodepth=64

[seqprecon]
filename=/dev/sdd
[seqprecon]
filename=/dev/sde
[seqprecon]
filename=/dev/sdf

Quoted 3 lines

I will ask Martin to pick all the patches from "[v3,1/4] add io_uring with IOPOLL support in the scsi layer" except scsi_debug. We can work on scsi_debug and send standalone patch.

Hi Kashyap,
Yes, you are correct, that script certainly fails when sdd, sde and sdf are scsi_debug devices.

On closer examination, IMO there is a fundamental misunderstanding about iopoll on one of our parts. I am assuming that all these things are closely coupled:
1) a request with REQ_HIPRI set in its cmd_flags
2) the associated LLD receiving mq_poll callbacks
3) _something_ in the upper layers calls blk_poll()

And it is point 3) that is missing in the case of that fio script shown above. To be able to use sd devices, I believe the sd driver needs to be altered to issue blk_poll() calls when it sees that it is waiting for completion of request/commands that have REQ_HIPRI set. And that is not happening when I run the above fio script.

My scsi_debug patch counts the number of times mq_poll returns a value

Quoted 1 line

0 (i.e. it completed a request) and this is what I see:

# cat /proc/scsi/scsi_debug/3
scsi_debug adapter driver, version 0190 [20200710]
num_tgts=1, shared (ram) size=1024 MB, opts=0x0, every_nth=0
delay=-9999, ndelay=10000, max_luns=1, sector_size=4096 bytes
cylinders=128, heads=64, sectors=32, command aborts=98
RESETs: device=5, target=5, bus=5, host=2
dix_reads=0, dix_writes=0, dif_errors=0
usec_in_jiffy=4000, statistics=0
cmnd_count=0, completions=0, miss_cpus=0, a_tsf=0, mq_polls=0
submit_queues=4
  queue 0:
  queue 1:
  queue 2:
  queue 3:
this host_no=3
....

The fact that 'mq_poll=0' tells me that blk_poll() was never called.

Assuming the above fio script works when sdd, sde snd sdf are MR devices suggests to me your LLD treats mq_poll callbacks as optional. The way I have coded the sg an scsi_debug drivers is that when REQ_HIPRI is given, then blk_poll() calls are _essential_ and said request will not complete without them.

This code is from drivers/nvme/host/core.c :

static void nvme_execute_rq_polled(struct request_queue *q,
struct gendisk *bd_disk, struct request *rq, int at_head)

{
        DECLARE_COMPLETION_ONSTACK(wait);

        WARN_ON_ONCE(!test_bit(QUEUE_FLAG_POLL, &q->queue_flags));

        rq->cmd_flags |= REQ_HIPRI;
        rq->end_io_data = &wait;
        blk_execute_rq_nowait(q, bd_disk, rq, at_head, nvme_end_sync_rq);

        while (!completion_done(&wait)) {
                blk_poll(q, request_to_qc_t(rq->mq_hctx, rq), true);
                cond_resched();
        }
}

For requests with REQ_HIPRI set, the sg driver (with blk_poll patch) does an async version of the synchronous procedure shown in that nvme function.

Doug Gilbert

Quoted 274 lines
Quoted 274 lines

Changes since version 1 [sent 20210201 to linux-scsi list]
- harden SDEB_DEFER_POLL which broke under testing
- add mq_polls counter for debug output

---
  drivers/scsi/scsi_debug.c | 141 +++++++++++++++++++++++++-------------
  1 file changed, 94 insertions(+), 47 deletions(-)
drivers/scsi/scsi_debug.c +94 -48
index 746eec521f79..b9ec3a47fbf1 100644
@@ -322,17 +322,19 @@ struct sdeb_store_info {
         container_of(d, struct sdebug_host_info, dev)
 
  enum sdeb_defer_type {SDEB_DEFER_NONE = 0, SDEB_DEFER_HRT = 1,
-                     SDEB_DEFER_WQ = 2};
+                     SDEB_DEFER_WQ = 2, SDEB_DEFER_POLL = 3};
 
  struct sdebug_defer {
         struct hrtimer hrt;
         struct execute_work ew;
+       ktime_t cmpl_ts;/* time since boot to complete this cmd */
         int sqa_idx;    /* index of sdebug_queue array */
         int qc_idx;     /* index of sdebug_queued_cmd array within sqa_idx */
         int hc_idx;     /* hostwide tag index */
         int issuing_cpu;
         bool init_hrt;
         bool init_wq;
+       bool init_poll;
         bool aborted;   /* true when blk_abort_request() already called */
         enum sdeb_defer_type defer_t;
  };
@@ -357,6 +359,7 @@ static atomic_t sdebug_completions;  /* count of deferred completions */
  static atomic_t sdebug_miss_cpus;    /* submission + completion cpus differ */
  static atomic_t sdebug_a_tsf;       /* 'almost task set full' counter */
  static atomic_t sdeb_inject_pending;
+static atomic_t sdeb_mq_poll_count;  /* bumped when mq_poll returns > 0 */
 
  struct opcode_info_t {
         u8 num_attached;        /* 0 if this is it (i.e. a leaf); use 0xff */
@@ -4730,7 +4733,6 @@ static void sdebug_q_cmd_complete(struct sdebug_defer *sd_dp)
         struct scsi_cmnd *scp;
         struct sdebug_dev_info *devip;
 
-       sd_dp->defer_t = SDEB_DEFER_NONE;
         if (unlikely(aborted))
                 sd_dp->aborted = false;
         qc_idx = sd_dp->qc_idx;
@@ -4745,6 +4747,7 @@ static void sdebug_q_cmd_complete(struct sdebug_defer *sd_dp)
                 return;
         }
         spin_lock_irqsave(&sqp->qc_lock, iflags);
+       sd_dp->defer_t = SDEB_DEFER_NONE;
         sqcp = &sqp->qc_arr[qc_idx];
         scp = sqcp->a_cmnd;
         if (unlikely(scp == NULL)) {
@@ -5517,40 +5520,66 @@ static int schedule_resp(struct scsi_cmnd *cmnd, struct sdebug_dev_info *devip,
                                 kt -= d;
                         }
                 }
-               if (!sd_dp->init_hrt) {
-                       sd_dp->init_hrt = true;
-                       sqcp->sd_dp = sd_dp;
-                       hrtimer_init(&sd_dp->hrt, CLOCK_MONOTONIC,
-                                    HRTIMER_MODE_REL_PINNED);
-                       sd_dp->hrt.function = sdebug_q_cmd_hrt_complete;
-                       sd_dp->sqa_idx = sqp - sdebug_q_arr;
-                       sd_dp->qc_idx = k;
+               sd_dp->cmpl_ts = ktime_add(ns_to_ktime(ns_from_boot), kt);
+               if (cmnd->request->cmd_flags & REQ_HIPRI) {
+                       spin_lock_irqsave(&sqp->qc_lock, iflags);
+                       if (!sd_dp->init_poll) {
+                               sd_dp->init_poll = true;
+                               sqcp->sd_dp = sd_dp;
+                               sd_dp->sqa_idx = sqp - sdebug_q_arr;
+                               sd_dp->qc_idx = k;
+                       }
+                       sd_dp->defer_t = SDEB_DEFER_POLL;
+                       spin_unlock_irqrestore(&sqp->qc_lock, iflags);
+               } else {
+                       if (!sd_dp->init_hrt) {
+                               sd_dp->init_hrt = true;
+                               sqcp->sd_dp = sd_dp;
+                               hrtimer_init(&sd_dp->hrt, CLOCK_MONOTONIC,
+                                            HRTIMER_MODE_REL_PINNED);
+                               sd_dp->hrt.function = sdebug_q_cmd_hrt_complete;
+                               sd_dp->sqa_idx = sqp - sdebug_q_arr;
+                               sd_dp->qc_idx = k;
+                       }
+                       sd_dp->defer_t = SDEB_DEFER_HRT;
+                       /* schedule the invocation of scsi_done() for a later time */
+                       hrtimer_start(&sd_dp->hrt, kt, HRTIMER_MODE_REL_PINNED);
                 }
                 if (sdebug_statistics)
                         sd_dp->issuing_cpu = raw_smp_processor_id();
-               sd_dp->defer_t = SDEB_DEFER_HRT;
-               /* schedule the invocation of scsi_done() for a later time */
-               hrtimer_start(&sd_dp->hrt, kt, HRTIMER_MODE_REL_PINNED);
         } else {        /* jdelay < 0, use work queue */
-               if (!sd_dp->init_wq) {
-                       sd_dp->init_wq = true;
-                       sqcp->sd_dp = sd_dp;
-                       sd_dp->sqa_idx = sqp - sdebug_q_arr;
-                       sd_dp->qc_idx = k;
-                       INIT_WORK(&sd_dp->ew.work, sdebug_q_cmd_wq_complete);
-               }
-               if (sdebug_statistics)
-                       sd_dp->issuing_cpu = raw_smp_processor_id();
-               sd_dp->defer_t = SDEB_DEFER_WQ;
                 if (unlikely((sdebug_opts & SDEBUG_OPT_CMD_ABORT) &&
                              atomic_read(&sdeb_inject_pending)))
                         sd_dp->aborted = true;
-               schedule_work(&sd_dp->ew.work);
-               if (unlikely((sdebug_opts & SDEBUG_OPT_CMD_ABORT) &&
-                            atomic_read(&sdeb_inject_pending))) {
+               sd_dp->cmpl_ts = ns_to_ktime(ns_from_boot);
+               if (cmnd->request->cmd_flags & REQ_HIPRI) {
+                       spin_lock_irqsave(&sqp->qc_lock, iflags);
+                       if (!sd_dp->init_poll) {
+                               sd_dp->init_poll = true;
+                               sqcp->sd_dp = sd_dp;
+                               sd_dp->sqa_idx = sqp - sdebug_q_arr;
+                               sd_dp->qc_idx = k;
+                       }
+                       sd_dp->defer_t = SDEB_DEFER_POLL;
+                       spin_unlock_irqrestore(&sqp->qc_lock, iflags);
+               } else {
+                       if (!sd_dp->init_wq) {
+                               sd_dp->init_wq = true;
+                               sqcp->sd_dp = sd_dp;
+                               sd_dp->sqa_idx = sqp - sdebug_q_arr;
+                               sd_dp->qc_idx = k;
+                               INIT_WORK(&sd_dp->ew.work, sdebug_q_cmd_wq_complete);
+                       }
+                       sd_dp->defer_t = SDEB_DEFER_WQ;
+                       schedule_work(&sd_dp->ew.work);
+               }
+               if (sdebug_statistics)
+                       sd_dp->issuing_cpu = raw_smp_processor_id();
+               if (unlikely(sd_dp->aborted)) {
                         sdev_printk(KERN_INFO, sdp, "abort request tag %d\n", cmnd->request->tag);
                         blk_abort_request(cmnd->request);
                         atomic_set(&sdeb_inject_pending, 0);
+                       sd_dp->aborted = false;
                 }
         }
         if (unlikely((SDEBUG_OPT_Q_NOISE & sdebug_opts) && scsi_result == device_qfull_result))
@@ -5779,11 +5808,12 @@ static int scsi_debug_show_info(struct seq_file *m, struct Scsi_Host *host)
                    dix_reads, dix_writes, dif_errors);
         seq_printf(m, "usec_in_jiffy=%lu, statistics=%d\n", TICK_NSEC / 1000,
                    sdebug_statistics);
-       seq_printf(m, "cmnd_count=%d, completions=%d, %s=%d, a_tsf=%d\n",
+       seq_printf(m, "cmnd_count=%d, completions=%d, %s=%d, a_tsf=%d, mq_polls=%d\n",
                    atomic_read(&sdebug_cmnd_count),
                    atomic_read(&sdebug_completions),
                    "miss_cpus", atomic_read(&sdebug_miss_cpus),
-                  atomic_read(&sdebug_a_tsf));
+                  atomic_read(&sdebug_a_tsf),
+                  atomic_read(&sdeb_mq_poll_count));
 
         seq_printf(m, "submit_queues=%d\n", submit_queues);
         for (j = 0, sqp = sdebug_q_arr; j < submit_queues; ++j, ++sqp) {
@@ -7246,52 +7276,68 @@ static int sdebug_map_queues(struct Scsi_Host *shost)
 
  static int sdebug_blk_mq_poll(struct Scsi_Host *shost, unsigned int queue_num)
  {
-       int qc_idx;
-       int retiring = 0;
+       bool first;
+       bool retiring = false;
+       int num_entries = 0;
+       unsigned int qc_idx = 0;
         unsigned long iflags;
+       ktime_t kt_from_boot = ktime_get_boottime();
         struct sdebug_queue *sqp;
         struct sdebug_queued_cmd *sqcp;
         struct scsi_cmnd *scp;
         struct sdebug_dev_info *devip;
-       int num_entries = 0;
+       struct sdebug_defer *sd_dp;
 
         sqp = sdebug_q_arr + queue_num;
+       spin_lock_irqsave(&sqp->qc_lock, iflags);
 
-       do {
-               spin_lock_irqsave(&sqp->qc_lock, iflags);
-               qc_idx = find_first_bit(sqp->in_use_bm, sdebug_max_queue);
-               if (unlikely((qc_idx < 0) || (qc_idx >= sdebug_max_queue)))
-                       goto out;
+       for (first = true; first || qc_idx + 1 < sdebug_max_queue; )   {
+               if (first) {
+                       qc_idx = find_first_bit(sqp->in_use_bm, sdebug_max_queue);
+                       first = false;
+               } else {
+                       qc_idx = find_next_bit(sqp->in_use_bm, sdebug_max_queue, qc_idx + 1);
+               }
+               if (unlikely(qc_idx >= sdebug_max_queue))
+                       break;
 
                 sqcp = &sqp->qc_arr[qc_idx];
+               sd_dp = sqcp->sd_dp;
+               if (unlikely(!sd_dp))
+                       continue;
                 scp = sqcp->a_cmnd;
                 if (unlikely(scp == NULL)) {
-                       pr_err("scp is NULL, queue_num=%d, qc_idx=%d from %s\n",
+                       pr_err("scp is NULL, queue_num=%d, qc_idx=%u from %s\n",
                                queue_num, qc_idx, __func__);
-                       goto out;
+                       break;
                 }
+               if (sd_dp->defer_t == SDEB_DEFER_POLL) {
+                       if (kt_from_boot < sd_dp->cmpl_ts)
+                               continue;
+
+               } else          /* ignoring non REQ_HIPRI requests */
+                       continue;
                 devip = (struct sdebug_dev_info *)scp->device->hostdata;
                 if (likely(devip))
                         atomic_dec(&devip->num_in_q);
                 else
                         pr_err("devip=NULL from %s\n", __func__);
                 if (unlikely(atomic_read(&retired_max_queue) > 0))
-                       retiring = 1;
+                       retiring = true;
 
                 sqcp->a_cmnd = NULL;
                 if (unlikely(!test_and_clear_bit(qc_idx, sqp->in_use_bm))) {
-                       pr_err("Unexpected completion sqp %p queue_num=%d qc_idx=%d from %s\n",
+                       pr_err("Unexpected completion sqp %p queue_num=%d qc_idx=%u from %s\n",
                                 sqp, queue_num, qc_idx, __func__);
-                       goto out;
+                       break;
                 }
-
                 if (unlikely(retiring)) {       /* user has reduced max_queue */
                         int k, retval;
 
                         retval = atomic_read(&retired_max_queue);
                         if (qc_idx >= retval) {
                                 pr_err("index %d too large\n", retval);
-                               goto out;
+                               break;
                         }
                         k = find_last_bit(sqp->in_use_bm, retval);
                         if ((k < sdebug_max_queue) || (k == retval))
@@ -7299,17 +7345,18 @@ static int sdebug_blk_mq_poll(struct Scsi_Host *shost, unsigned int queue_num)
                         else
                                 atomic_set(&retired_max_queue, k + 1);
                 }
+               sd_dp->defer_t = SDEB_DEFER_NONE;
                 spin_unlock_irqrestore(&sqp->qc_lock, iflags);
                 scp->scsi_done(scp); /* callback to mid level */
+               spin_lock_irqsave(&sqp->qc_lock, iflags);
                 num_entries++;
-       } while (1);
-
-out:
+       }
         spin_unlock_irqrestore(&sqp->qc_lock, iflags);
+       if (num_entries > 0)
+               atomic_add(num_entries, &sdeb_mq_poll_count);
         return num_entries;
  }
 
-
  static int scsi_debug_queuecommand(struct Scsi_Host *shost,
                                    struct scsi_cmnd *scp)
  {
--

2.25.1

#3 lore raw
In reply to Kashyap Desai
Quoted 47 lines · Kashyap Desai
Quoted 19 lines · Douglas Gilbert

Add a new sdeb_defer_type enumeration: SDEB_DEFER_POLL for requests that have REQ_HIPRI set in cmd_flags field. It is expected that these requests will be polled via the mq_poll entry point which is driven by calls to blk_poll() in the block layer. Therefore timer events are not 'wired up' in the normal fashion.

There are still cases with short delays (e.g. < 10 microseconds) where by the time the command response processing occurs, the delay is already exceeded in which case the code calls scsi_done() directly. In such cases there is no window for mq_poll() to be called.

Add 'mq_polls' counter that increments on each scsi_done() called via the mq_poll entry point. Can be used to show (with 'cat /proc/scsi/scsi_debug/<host_id>') that blk_poll() is causing completions rather than some other mechanism.

Quoted 8 lines

This patch is against the 5.12/scsi-staging branch which includes
a98d6bdf181eb71bf4686666eaf47c642a061642
scsi_debug : iouring iopoll support
which it alters. So this patch is a fix of that patch.

Changes since version 2 [sent 20210206 to linux-scsi list]
- the sdebug_blk_mq_poll() callback didn't cope with the

     uncommon case where sqcp->sd_dp is NULL. Fix.

Hi Doug, I tried this patch on top of below iouring patch series - "[v3,1/4] add io_uring with IOPOLL support in scsi layer"

After applying patch, I am seeing an IO hang issue - I see io_wqe_worker goes into a tight loop.

18210 root 20 0 1316348 3552 2376 R 99.1 0.0 0:24.09
fio 18303 root 20 0
0 0 0 D 78.8 0.0 0:01.75 io_wqe_worker-0

            18219 root      20   0       0      0      0 R  71.7   0.0

0:06.59 io_wqe_worker-0

I used below command -
insmod drivers/scsi/scsi_debug.ko dev_size_mb=1024 sector_size=512 add_host=24 per_host_store=1 ndelay=10000 host_max_queue=32 submit_queues=76 num_parts=1 poll_queues=8

Quoted 10 lines

and here is my fio script -
[global]
ioengine=io_uring
hipri=1
direct=1
runtime=30s
rw=randread
norandommap
bs=4k
iodepth=64

Kashyap,
There is another issue here, namely iodepth > host_max_queue (64 > 32) and in my setup that is not handled well. So there is a problem with scsi_debug *** or the mid-level in handling this case.

If you have modified the sd driver to call blk_poll() then perhaps you could try the above test again with a reduced iodepth.

Doug Gilbert

*** the scsi_debug should either yield a TASK SET FULL status or return
SCSI_MLQUEUE_HOST_BUSY from queuecommand() when it has run out of slots.

Quoted 13 lines

[seqprecon]
filename=/dev/sdd
[seqprecon]
filename=/dev/sde
[seqprecon]
filename=/dev/sdf

I will ask Martin to pick all the patches from "[v3,1/4] add io_uring with IOPOLL support in the scsi layer" except scsi_debug. We can work on scsi_debug and send standalone patch.

Kashyap

Tested-by
#4 lore raw
Quoted 9 lines

Kashyap,
There is another issue here, namely iodepth > host_max_queue (64 > 32) and in my setup that is not handled well. So there is a problem with scsi_debug
*** or the mid-level in handling this case.

If you have modified the sd driver to call blk_poll() then perhaps you could try
the above test again with a reduced iodepth.

Doug -

I debug the issue and found the fix. We need to remove below code to handle this new defer type for mq poll.
Earlier I did not used " REQ_HIPRI" to handle mq poll, but now you are managing it through REQ_HIPRI check, it is safe and required to have below patch.

drivers/scsi/scsi_debug.c +0 -0
index b9ec3a47fbf1..c50de49a2c2f 100644
@@ -5437,13 +5437,6 @@ static int schedule_resp(struct scsi_cmnd *cmnd,

struct sdebug_dev_info *devip,

        sd_dp = sqcp->sd_dp;
        spin_unlock_irqrestore(&sqp->qc_lock, iflags);

- /* Do not complete IO from default completion path.

-        * Let it to be on queue.
-        * Completion should happen from mq_poll interface.
-        */
-       if ((sqp - sdebug_q_arr) >= (submit_queues - poll_queues))
-               return 0;
-
        if (!sd_dp) {
                sd_dp = kzalloc(sizeof(*sd_dp), GFP_ATOMIC);
                if (!sd_dp) {

On top of your v3 patch + above fix, I am able to run IO and there are no issues. There is no issue with higher QD as well, so we are good to go.

So how shall we progress now ? Shall I ask Martin to pick all the patches from " [PATCH v3 0/4] io_uring iopoll in scsi layer" since this patch has dependency on my series.

Kashyap

Quoted 20 lines

Doug Gilbert

*** the scsi_debug should either yield a TASK SET FULL status or return
SCSI_MLQUEUE_HOST_BUSY from queuecommand() when it has run out of slots.

Quoted 13 lines

[seqprecon]
filename=/dev/sdd
[seqprecon]
filename=/dev/sde
[seqprecon]
filename=/dev/sdf

I will ask Martin to pick all the patches from "[v3,1/4] add io_uring with IOPOLL support in the scsi layer" except scsi_debug. We can work on scsi_debug and send standalone patch.

Kashyap

#5 lore raw
Quoted 40 lines · Kashyap Desai
Quoted 9 lines

Kashyap,
There is another issue here, namely iodepth > host_max_queue (64 > 32) and in my setup that is not handled well. So there is a problem with scsi_debug
*** or the mid-level in handling this case.

If you have modified the sd driver to call blk_poll() then perhaps you could try
the above test again with a reduced iodepth.

Doug -

I debug the issue and found the fix. We need to remove below code to handle this new defer type for mq poll.
Earlier I did not used " REQ_HIPRI" to handle mq poll, but now you are managing it through REQ_HIPRI check, it is safe and required to have below patch.

drivers/scsi/scsi_debug.c +0 -7
index b9ec3a47fbf1..c50de49a2c2f 100644
@@ -5437,13 +5437,6 @@ static int schedule_resp(struct scsi_cmnd *cmnd,
 struct sdebug_dev_info *devip,
         sd_dp = sqcp->sd_dp;
         spin_unlock_irqrestore(&sqp->qc_lock, iflags);
 
-       /* Do not complete IO from default completion path.
-        * Let it to be on queue.
-        * Completion should happen from mq_poll interface.
-        */
-       if ((sqp - sdebug_q_arr) >= (submit_queues - poll_queues))
-               return 0;
-
         if (!sd_dp) {
                 sd_dp = kzalloc(sizeof(*sd_dp), GFP_ATOMIC);
                 if (!sd_dp) {
Quoted 8 lines

On top of your v3 patch + above fix, I am able to run IO and there are no issues. There is no issue with higher QD as well, so we are good to go.

So how shall we progress now ? Shall I ask Martin to pick all the patches from " [PATCH v3 0/4] io_uring iopoll in scsi layer" since this patch has dependency on my series.

Kashyap,
Good catch!

You could roll all the scsi_debug stuff into one patch and put both our sign-offs on it, or present it as two patches:
- first one: your original plus this fix
- second one: mine with your tested-by on it

Since this is new functionality for the scsi sub-system, it is unlikely that all the corner cases have been visited. However, with what we have in place for the next cycle, others have something reasonably solid to build on.

Doug Gilbert

Quoted 17 lines
Quoted 17 lines

*** the scsi_debug should either yield a TASK SET FULL status or return
SCSI_MLQUEUE_HOST_BUSY from queuecommand() when it has run out of slots.

Quoted 13 lines

[seqprecon]
filename=/dev/sdd
[seqprecon]
filename=/dev/sde
[seqprecon]
filename=/dev/sdf

I will ask Martin to pick all the patches from "[v3,1/4] add io_uring with IOPOLL support in the scsi layer" except scsi_debug. We can work on scsi_debug and send standalone patch.

Kashyap

#6 lore raw
Quoted 4 lines
Quoted 3 lines

So how shall we progress now ? Shall I ask Martin to pick all the patches from " [PATCH v3 0/4] io_uring iopoll in scsi layer" since this patch has dependency on my series.

Quoted 8 lines

Kashyap,
Good catch!

You could roll all the scsi_debug stuff into one patch and put both our sign-
offs on it, or present it as two patches:
- first one: your original plus this fix
- second one: mine with your tested-by on it

Doug - I will include your patch "scsi_debug: add new defer_type for mq_poll" (including the recent fix) as separate patch in my series and provide Tested-by tag for the same.

Quoted 6 lines

Since this is new functionality for the scsi sub-system, it is unlikely that all the
corner cases have been visited. However, with what we have in place for the
next cycle, others have something reasonably solid to build on.

Agree.