Skip to contentSkip to conversation

Improve zoned storage write performance

PATCH In review Bart Van Assche 4 participants 25 messages Jun 27, 2022 – Jun 29, 2022

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

v3 cover Improve zoned storage write performance

Hi Jens,

Measurements have shown that limiting the queue depth to one per sequential zone has a significant negative performance impact for zoned UFS devices. Hence this patch series that increases the queue depth for write commands for sequential zones when using the mq-deadline scheduler.

Please consider this patch series for kernel v5.20.

Thanks,

Bart.

Changes compared to v2:
- Included the patches again that enable write pipelining for NVMe ZNS SSDs.
- Disabled merging in the null_blk test script. As a result, the test script now shows a 14x improvement with pipelining enabled.
- Renamed blk_rq_is_zoned_seq_write() into blk_rq_is_seq_zone_write().

Changes compared to v1:
- Left out the patch for the UFS driver.
- Included patches for the null_blk driver.

Bart Van Assche (8):
block: Document blk_queue_zone_is_seq() and blk_rq_zone_is_seq()
block: Introduce the blk_rq_is_seq_zone_write() function
block: Introduce a request queue flag for pipelining zoned writes block/mq-deadline: Only use zone locking if necessary block/null_blk: Refactor null_queue_rq() block/null_blk: Add support for pipelining zoned writes
nvme: Make the number of retries command specific
nvme: Enable pipelining of zoned writes

 block/blk-zoned.c                 | 17 +++--------------
 block/mq-deadline.c               | 18 ++++++++++++------
 drivers/block/null_blk/main.c     | 30 ++++++++++++++++++++----------
 drivers/block/null_blk/null_blk.h |  3 +++
 drivers/block/null_blk/zoned.c    | 13 ++++++++++++-
 drivers/nvme/host/core.c          |  9 ++++++++-
 drivers/nvme/host/nvme.h          |  1 +
 drivers/nvme/host/zns.c           |  2 ++
 include/linux/blk-mq.h            | 30 ++++++++++++++++++++++++++++++
 include/linux/blkdev.h            | 16 ++++++++++++++++
 10 files changed, 107 insertions(+), 32 deletions(-)
#1 lore raw

PATCH v3 1/8 block: Document blk_queue_zone_is_seq() and blk_rq_zone_is_seq()

Since it is nontrivial to figure out how blk_queue_zone_is_seq() and blk_rq_zone_is_seq() handle sequential write preferred zones, document this.

---
 include/linux/blk-mq.h | 7 +++++++
 include/linux/blkdev.h | 9 +++++++++
 2 files changed, 16 insertions(+)
include/linux/blk-mq.h +7 -0
index e2d9daf7e8dd..909d47e34b7c 100644
@@ -1124,6 +1124,13 @@ static inline unsigned int blk_rq_zone_no(struct request *rq)
 	return blk_queue_zone_no(rq->q, blk_rq_pos(rq));
 }
 
+/**
+ * blk_rq_zone_is_seq() - Whether a request is for a sequential zone.
+ * @rq: Request pointer.
+ *
+ * Return: true if and only if blk_rq_pos(@rq) refers either to a sequential
+ * write required or a sequential write preferred zone.
+ */
 static inline unsigned int blk_rq_zone_is_seq(struct request *rq)
 {
 	return blk_queue_zone_is_seq(rq->q, blk_rq_pos(rq));
include/linux/blkdev.h +9 -0
index b2d42201bd5d..6491250ba286 100644
@@ -687,6 +687,15 @@ static inline unsigned int blk_queue_zone_no(struct request_queue *q,
 	return sector >> ilog2(q->limits.chunk_sectors);
 }
 
+/**
+ * blk_queue_zone_is_seq() - Whether a logical block is in a sequential zone.
+ * @q: Request queue pointer.
+ * @sector: Offset from start of block device in 512 byte units.
+ *
+ * Return: true if and only if @q is associated with a zoned block device and
+ * @sector refers either to a sequential write required or a sequential write
+ * preferred zone.
+ */
 static inline bool blk_queue_zone_is_seq(struct request_queue *q,
 					 sector_t sector)
 {
#3 lore raw

PATCH v3 2/8 block: Introduce the blk_rq_is_seq_zone_write() function

In reply to Bart Van Assche

Introduce a function that makes it easy to verify whether a write request is for a sequential write required or sequential write preferred zone. Simplify blk_req_needs_zone_write_lock() by using the new function.

---
 block/blk-zoned.c      | 14 +-------------
 include/linux/blk-mq.h | 25 ++++++++++++++++++++++++-
 2 files changed, 25 insertions(+), 14 deletions(-)
block/blk-zoned.c +1 -13
index 38cd840d8838..9da8cf1bb378 100644
@@ -57,19 +57,7 @@ EXPORT_SYMBOL_GPL(blk_zone_cond_str);
  */
 bool blk_req_needs_zone_write_lock(struct request *rq)
 {
-	if (!rq->q->seq_zones_wlock)
-		return false;
-
-	if (blk_rq_is_passthrough(rq))
-		return false;
-
-	switch (req_op(rq)) {
-	case REQ_OP_WRITE_ZEROES:
-	case REQ_OP_WRITE:
-		return blk_rq_zone_is_seq(rq);
-	default:
-		return false;
-	}
+	return rq->q->seq_zones_wlock && blk_rq_is_seq_zone_write(rq);
 }
 EXPORT_SYMBOL_GPL(blk_req_needs_zone_write_lock);
 
include/linux/blk-mq.h +24 -1
index 909d47e34b7c..ccfcac9db879 100644
@@ -1129,13 +1129,31 @@ static inline unsigned int blk_rq_zone_no(struct request *rq)
  * @rq: Request pointer.
  *
  * Return: true if and only if blk_rq_pos(@rq) refers either to a sequential
- * write required or a sequential write preferred zone.
+ * write required or to a sequential write preferred zone.
  */
 static inline unsigned int blk_rq_zone_is_seq(struct request *rq)
 {
 	return blk_queue_zone_is_seq(rq->q, blk_rq_pos(rq));
 }
 
+/**
+ * blk_rq_is_seq_zone_write() - Whether @rq is a write request for a sequential zone.
+ * @rq: Request to examine.
+ *
+ * In this context sequential zone means either a sequential write required or
+ * to a sequential write preferred zone.
+ */
+static inline bool blk_rq_is_seq_zone_write(struct request *rq)
+{
+	switch (req_op(rq)) {
+	case REQ_OP_WRITE:
+	case REQ_OP_WRITE_ZEROES:
+		return blk_rq_zone_is_seq(rq);
+	default:
+		return false;
+	}
+}
+
 bool blk_req_needs_zone_write_lock(struct request *rq);
 bool blk_req_zone_write_trylock(struct request *rq);
 void __blk_req_zone_write_lock(struct request *rq);
@@ -1166,6 +1184,11 @@ static inline bool blk_req_can_dispatch_to_zone(struct request *rq)
 	return !blk_req_zone_is_write_locked(rq);
 }
 #else /* CONFIG_BLK_DEV_ZONED */
+static inline bool blk_rq_is_seq_zone_write(struct request *rq)
+{
+	return false;
+}
+
 static inline bool blk_req_needs_zone_write_lock(struct request *rq)
 {
 	return false;
Reviewed-by on v3 2/8
#4 lore raw

Re: [PATCH v3 2/8] block: Introduce the blk_rq_is_seq_zone_write() function

Quoted 7 lines · Bart Van Assche

Introduce a function that makes it easy to verify whether a write request is for a sequential write required or sequential write preferred zone. Simplify blk_req_needs_zone_write_lock() by using the new function.

---

Looks good.

-ck

#5 lore raw

PATCH v3 3/8 block: Introduce a request queue flag for pipelining zoned writes

In reply to Bart Van Assche

Writes in sequential write required zones must happen at the write pointer. Even if the submitter of the write commands (e.g. a filesystem) submits writes for sequential write required zones in order, the block layer or the storage controller may reorder these write commands.

The zone locking mechanism in the mq-deadline I/O scheduler serializes write commands for sequential zones. Some but not all storage controllers require this serialization. Introduce a new flag such that block drivers can request pipelining of writes for sequential write required zones.

An example of a storage controller standard that requires write serialization is AHCI (Advanced Host Controller Interface). Submitting commands to AHCI controllers happens by writing a bit pattern into a register. Each set bit corresponds to an active command. This mechanism does not preserve command ordering information.

---
 include/linux/blkdev.h | 7 +++++++
 1 file changed, 7 insertions(+)
include/linux/blkdev.h +7 -0
index 6491250ba286..7ee59e507fbf 100644
@@ -580,6 +580,8 @@ struct request_queue {
 #define QUEUE_FLAG_HCTX_ACTIVE	28	/* at least one blk-mq hctx is active */
 #define QUEUE_FLAG_NOWAIT       29	/* device supports NOWAIT */
 #define QUEUE_FLAG_SQ_SCHED     30	/* single queue style io dispatch */
+/* Writes for sequential write required zones may be pipelined. */
+#define QUEUE_FLAG_PIPELINE_ZONED_WRITES 31
 
 #define QUEUE_FLAG_MQ_DEFAULT	((1 << QUEUE_FLAG_IO_STAT) |		\
 				 (1 << QUEUE_FLAG_SAME_COMP) |		\
@@ -623,6 +625,11 @@ bool blk_queue_flag_test_and_set(unsigned int flag, struct request_queue *q);
 #define blk_queue_nowait(q)	test_bit(QUEUE_FLAG_NOWAIT, &(q)->queue_flags)
 #define blk_queue_sq_sched(q)	test_bit(QUEUE_FLAG_SQ_SCHED, &(q)->queue_flags)
 
+static inline bool blk_queue_pipeline_zoned_writes(struct request_queue *q)
+{
+	return test_bit(QUEUE_FLAG_PIPELINE_ZONED_WRITES, &(q)->queue_flags);
+}
+
 extern void blk_set_pm_only(struct request_queue *q);
 extern void blk_clear_pm_only(struct request_queue *q);
 
#6 lore raw

Re: [PATCH v3 3/8] block: Introduce a request queue flag for pipelining zoned writes

Quoted 8 lines
  #define QUEUE_FLAG_SQ_SCHED     30	/* single queue style io dispatch */
include/linux/blkdev.h +2 -0
+/* Writes for sequential write required zones may be pipelined. */
+#define QUEUE_FLAG_PIPELINE_ZONED_WRITES 31
 
  #define QUEUE_FLAG_MQ_DEFAULT	((1 << QUEUE_FLAG_IO_STAT) |		\
  				 (1 << QUEUE_FLAG_SAME_COMP) |		\

include/linux/blkdev.h bool blk_queue_flag_test_and_set(unsigned int flag, struct request_queue *q); · line 632

Quoted 8 lines
include/linux/blkdev.h +5 -0
@@ -623,6 +625,11 @@ bool blk_queue_flag_test_and_set(unsigned int flag, struct request_queue *q);
  #define blk_queue_nowait(q)	test_bit(QUEUE_FLAG_NOWAIT, &(q)->queue_flags)
  #define blk_queue_sq_sched(q)	test_bit(QUEUE_FLAG_SQ_SCHED, &(q)->queue_flags)
 
+static inline bool blk_queue_pipeline_zoned_writes(struct request_queue *q)
+{
+	return test_bit(QUEUE_FLAG_PIPELINE_ZONED_WRITES, &(q)->queue_flags);
+}
+

From the comments that I've received, when introducing a new helper or
a flag should be a part of the patch that actually uses it, is there a specific reason that this has made as a separate patch ?

-ck

#7 lore raw

Re: [PATCH v3 3/8] block: Introduce a request queue flag for pipelining zoned writes

Quoted 3 lines · Chaitanya Kulkarni

From the comments that I've received, when introducing a new helper or a flag should be a part of the patch that actually uses it, is there a specific reason that this has made as a separate patch ?

Hi Chaitanya,

This patch is a separate patch to make it easy to review this patch series. I'm concerned that if I would combine this patch with the next patch that the result would be harder to verify than two separate patches.

Thanks,

Bart.

#8 lore raw

PATCH v3 4/8 block/mq-deadline: Only use zone locking if necessary

In reply to Bart Van Assche

Measurements have shown that limiting the queue depth to one for zoned writes has a significant negative performance impact on zoned UFS devices. Hence this patch that disables zone locking from the mq-deadline scheduler for storage controllers that support pipelining zoned writes. This patch is based on the following assumptions:
- Applications submit write requests to sequential write required zones in order.
- The I/O priority of all pipelined write requests is the same per zone.
- Pipelined zoned write requests are submitted to a single hardware queue per zone.
- If such write requests get reordered by the software or hardware queue mechanism, nr_requests - 1 retries are sufficient to reorder the write requests.
- It happens infrequently that zoned write requests are reordered by the block layer.
- Either no I/O scheduler is used or an I/O scheduler is used that submits write requests per zone in LBA order.

---
 block/blk-zoned.c   |  3 ++-
 block/mq-deadline.c | 18 ++++++++++++------
 2 files changed, 14 insertions(+), 7 deletions(-)
block/blk-zoned.c +2 -1
index 9da8cf1bb378..63c730a18ac4 100644
@@ -513,7 +513,8 @@ static int blk_revalidate_zone_cb(struct blk_zone *zone, unsigned int idx,
 		break;
 	case BLK_ZONE_TYPE_SEQWRITE_REQ:
 	case BLK_ZONE_TYPE_SEQWRITE_PREF:
-		if (!args->seq_zones_wlock) {
+		if (!blk_queue_pipeline_zoned_writes(q) &&
+		    !args->seq_zones_wlock) {
 			args->seq_zones_wlock =
 				blk_alloc_zone_bitmap(q->node, args->nr_zones);
 			if (!args->seq_zones_wlock)
block/mq-deadline.c +12 -6
index 1a9e835e816c..aaef07a55984 100644
@@ -292,7 +292,8 @@ deadline_fifo_request(struct deadline_data *dd, struct dd_per_prio *per_prio,
 		return NULL;
 
 	rq = rq_entry_fifo(per_prio->fifo_list[data_dir].next);
-	if (data_dir == DD_READ || !blk_queue_is_zoned(rq->q))
+	if (data_dir == DD_READ || !blk_queue_is_zoned(rq->q) ||
+	    blk_queue_pipeline_zoned_writes(rq->q))
 		return rq;
 
 	/*
@@ -326,7 +327,8 @@ deadline_next_request(struct deadline_data *dd, struct dd_per_prio *per_prio,
 	if (!rq)
 		return NULL;
 
-	if (data_dir == DD_READ || !blk_queue_is_zoned(rq->q))
+	if (data_dir == DD_READ || !blk_queue_is_zoned(rq->q) ||
+	    blk_queue_pipeline_zoned_writes(rq->q))
 		return rq;
 
 	/*
@@ -445,8 +447,9 @@ static struct request *__dd_dispatch_request(struct deadline_data *dd,
 	}
 
 	/*
-	 * For a zoned block device, if we only have writes queued and none of
-	 * them can be dispatched, rq will be NULL.
+	 * For a zoned block device that requires write serialization, if we
+	 * only have writes queued and none of them can be dispatched, rq will
+	 * be NULL.
 	 */
 	if (!rq)
 		return NULL;
@@ -719,6 +722,8 @@ static void dd_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
 	u8 ioprio_class = IOPRIO_PRIO_CLASS(ioprio);
 	struct dd_per_prio *per_prio;
 	enum dd_prio prio;
+	bool pipelined_seq_write = blk_queue_pipeline_zoned_writes(q) &&
+		blk_rq_is_seq_zone_write(rq);
 	LIST_HEAD(free);
 
 	lockdep_assert_held(&dd->lock);
@@ -743,7 +748,7 @@ static void dd_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
 
 	trace_block_rq_insert(rq);
 
-	if (at_head) {
+	if (at_head && !pipelined_seq_write) {
 		list_add(&rq->queuelist, &per_prio->dispatch);
 		rq->fifo_time = jiffies;
 	} else {
@@ -823,7 +828,8 @@ static void dd_finish_request(struct request *rq)
 
 	atomic_inc(&per_prio->stats.completed);
 
-	if (blk_queue_is_zoned(q)) {
+	if (blk_queue_is_zoned(rq->q) &&
+	    !blk_queue_pipeline_zoned_writes(q)) {
 		unsigned long flags;
 
 		spin_lock_irqsave(&dd->zone_lock, flags);
#9 lore raw

PATCH v3 5/8 block/null_blk: Refactor null_queue_rq()

In reply to Bart Van Assche

Introduce a local variable for the expression bd->rq since that expression occurs multiple times. This patch does not change any functionality.

---
 drivers/block/null_blk/main.c | 21 +++++++++++----------
 1 file changed, 11 insertions(+), 10 deletions(-)
drivers/block/null_blk/main.c +11 -10
index 6b67088f4ea7..fd68e6f4637f 100644
@@ -1609,10 +1609,11 @@ static enum blk_eh_timer_return null_timeout_rq(struct request *rq, bool res)
 static blk_status_t null_queue_rq(struct blk_mq_hw_ctx *hctx,
 			 const struct blk_mq_queue_data *bd)
 {
-	struct nullb_cmd *cmd = blk_mq_rq_to_pdu(bd->rq);
+	struct request *rq = bd->rq;
+	struct nullb_cmd *cmd = blk_mq_rq_to_pdu(rq);
 	struct nullb_queue *nq = hctx->driver_data;
-	sector_t nr_sectors = blk_rq_sectors(bd->rq);
-	sector_t sector = blk_rq_pos(bd->rq);
+	sector_t nr_sectors = blk_rq_sectors(rq);
+	sector_t sector = blk_rq_pos(rq);
 	const bool is_poll = hctx->type == HCTX_TYPE_POLL;
 
 	might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
@@ -1621,14 +1622,14 @@ static blk_status_t null_queue_rq(struct blk_mq_hw_ctx *hctx,
 		hrtimer_init(&cmd->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
 		cmd->timer.function = null_cmd_timer_expired;
 	}
-	cmd->rq = bd->rq;
+	cmd->rq = rq;
 	cmd->error = BLK_STS_OK;
 	cmd->nq = nq;
-	cmd->fake_timeout = should_timeout_request(bd->rq);
+	cmd->fake_timeout = should_timeout_request(rq);
 
-	blk_mq_start_request(bd->rq);
+	blk_mq_start_request(rq);
 
-	if (should_requeue_request(bd->rq)) {
+	if (should_requeue_request(rq)) {
 		/*
 		 * Alternate between hitting the core BUSY path, and the
 		 * driver driven requeue path
@@ -1637,21 +1638,21 @@ static blk_status_t null_queue_rq(struct blk_mq_hw_ctx *hctx,
 		if (nq->requeue_selection & 1)
 			return BLK_STS_RESOURCE;
 		else {
-			blk_mq_requeue_request(bd->rq, true);
+			blk_mq_requeue_request(rq, true);
 			return BLK_STS_OK;
 		}
 	}
 
 	if (is_poll) {
 		spin_lock(&nq->poll_lock);
-		list_add_tail(&bd->rq->queuelist, &nq->poll_list);
+		list_add_tail(&rq->queuelist, &nq->poll_list);
 		spin_unlock(&nq->poll_lock);
 		return BLK_STS_OK;
 	}
 	if (cmd->fake_timeout)
 		return BLK_STS_OK;
 
-	return null_handle_cmd(cmd, sector, nr_sectors, req_op(bd->rq));
+	return null_handle_cmd(cmd, sector, nr_sectors, req_op(rq));
 }
 
 static void cleanup_queue(struct nullb_queue *nq)
#11 lore raw

PATCH v3 6/8 block/null_blk: Add support for pipelining zoned writes

In reply to Bart Van Assche

Add a new configfs attribute for enabling pipelining of zoned writes. If that attribute has been set, retry zoned writes that are not aligned with the write pointer. The test script below reports 244 K IOPS with no I/O scheduler, 5.79 K IOPS with mq-deadline and pipelining disabled and 79.6 K IOPS with mq-deadline and pipelining enabled. This shows that pipelining results in about 14 times more IOPS for this particular test case.

    #!/bin/bash

    for mode in "none 0" "mq-deadline 0" "mq-deadline 1"; do
        set +e
        for d in /sys/kernel/config/nullb/*; do
            [ -d "$d" ] && rmdir "$d"
        done
        modprobe -r null_blk
        set -e
        read -r iosched pipelining <<<"$mode"
        modprobe null_blk nr_devices=0
        (
            cd /sys/kernel/config/nullb
            mkdir nullb0
            cd nullb0
            params=(
                completion_nsec=100000
                hw_queue_depth=64
                irqmode=2                # NULL_IRQ_TIMER
                max_sectors=$((4096/512))
                memory_backed=1
                pipeline_zoned_writes="${pipelining}"
                size=1
                submit_queues=1
                zone_size=1
                zoned=1
                power=1
            )
            for p in "${params[@]}"; do
                echo "${p//*=}" > "${p//=*}"
            done
        )
        udevadm settle
        dev=/dev/nullb0
        [ -b "${dev}" ]
        params=(
            --direct=1
            --filename="${dev}"
            --iodepth=64
            --iodepth_batch=16
            --ioengine=io_uring
            --ioscheduler="${iosched}"
            --gtod_reduce=1
            --hipri=0
            --name=nullb0
            --runtime=30
            --rw=write
            --time_based=1
            --zonemode=zbd
        )
        fio "${params[@]}"
    done
---
 drivers/block/null_blk/main.c     |  9 +++++++++
 drivers/block/null_blk/null_blk.h |  3 +++
 drivers/block/null_blk/zoned.c    | 13 ++++++++++++-
 3 files changed, 24 insertions(+), 1 deletion(-)
drivers/block/null_blk/main.c +9 -0
index fd68e6f4637f..de9ed9a4ca10 100644
@@ -408,6 +408,7 @@ NULLB_DEVICE_ATTR(zone_capacity, ulong, NULL);
 NULLB_DEVICE_ATTR(zone_nr_conv, uint, NULL);
 NULLB_DEVICE_ATTR(zone_max_open, uint, NULL);
 NULLB_DEVICE_ATTR(zone_max_active, uint, NULL);
+NULLB_DEVICE_ATTR(pipeline_zoned_writes, bool, NULL);
 NULLB_DEVICE_ATTR(virt_boundary, bool, NULL);
 
 static ssize_t nullb_device_power_show(struct config_item *item, char *page)
@@ -531,6 +532,7 @@ static struct configfs_attribute *nullb_device_attrs[] = {
 	&nullb_device_attr_zone_nr_conv,
 	&nullb_device_attr_zone_max_open,
 	&nullb_device_attr_zone_max_active,
+	&nullb_device_attr_pipeline_zoned_writes,
 	&nullb_device_attr_virt_boundary,
 	NULL,
 };
@@ -1626,6 +1628,11 @@ static blk_status_t null_queue_rq(struct blk_mq_hw_ctx *hctx,
 	cmd->error = BLK_STS_OK;
 	cmd->nq = nq;
 	cmd->fake_timeout = should_timeout_request(rq);
+	if (!(rq->rq_flags & RQF_DONTPREP)) {
+		rq->rq_flags |= RQF_DONTPREP;
+		cmd->retries = 0;
+		cmd->max_attempts = rq->q->nr_requests;
+	}
 
 	blk_mq_start_request(rq);
 
@@ -2042,6 +2049,8 @@ static int null_add_dev(struct nullb_device *dev)
 	nullb->q->queuedata = nullb;
 	blk_queue_flag_set(QUEUE_FLAG_NONROT, nullb->q);
 	blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, nullb->q);
+	if (dev->pipeline_zoned_writes)
+		blk_queue_flag_set(QUEUE_FLAG_PIPELINE_ZONED_WRITES, nullb->q);
 
 	mutex_lock(&lock);
 	nullb->index = ida_simple_get(&nullb_indexes, 0, 0, GFP_KERNEL);
drivers/block/null_blk/null_blk.h +3 -0
index 8359b43842f2..bbe2cb17bdbd 100644
@@ -23,6 +23,8 @@ struct nullb_cmd {
 	unsigned int tag;
 	blk_status_t error;
 	bool fake_timeout;
+	u16 retries;
+	u16 max_attempts;
 	struct nullb_queue *nq;
 	struct hrtimer timer;
 };
@@ -112,6 +114,7 @@ struct nullb_device {
 	bool memory_backed; /* if data is stored in memory */
 	bool discard; /* if support discard */
 	bool zoned; /* if device is zoned */
+	bool pipeline_zoned_writes;
 	bool virt_boundary; /* virtual boundary on/off for the device */
 };
 
drivers/block/null_blk/zoned.c +12 -1
index 2fdd7b20c224..7138954dc01c 100644
@@ -403,7 +403,18 @@ static blk_status_t null_zone_write(struct nullb_cmd *cmd, sector_t sector,
 		else
 			cmd->bio->bi_iter.bi_sector = sector;
 	} else if (sector != zone->wp) {
-		ret = BLK_STS_IOERR;
+		/*
+		 * In case of a misaligned write and if pipelining of zoned
+		 * writes has been enabled, request the block layer to retry
+		 * until the maximum number of attempts has been reached. If
+		 * the maximum number of attempts has been reached, fail the
+		 * misaligned write.
+		 */
+		if (dev->pipeline_zoned_writes &&
+		    ++cmd->retries < cmd->max_attempts)
+			ret = BLK_STS_DEV_RESOURCE;
+		else
+			ret = BLK_STS_IOERR;
 		goto unlock;
 	}
 
#12 lore raw

Re: [PATCH v3 6/8] block/null_blk: Add support for pipelining zoned writes

Quoted 51 lines · Bart Van Assche

Add a new configfs attribute for enabling pipelining of zoned writes. If that attribute has been set, retry zoned writes that are not aligned with the write pointer. The test script below reports 244 K IOPS with no I/O scheduler, 5.79 K IOPS with mq-deadline and pipelining disabled and 79.6 K IOPS with mq-deadline and pipelining enabled. This shows that pipelining results in about 14 times more IOPS for this particular test case.

     #!/bin/bash

     for mode in "none 0" "mq-deadline 0" "mq-deadline 1"; do
         set +e
         for d in /sys/kernel/config/nullb/*; do
             [ -d "$d" ] && rmdir "$d"
         done
         modprobe -r null_blk
         set -e
         read -r iosched pipelining <<<"$mode"
         modprobe null_blk nr_devices=0
         (
             cd /sys/kernel/config/nullb
             mkdir nullb0
             cd nullb0
             params=(
                 completion_nsec=100000
                 hw_queue_depth=64
                 irqmode=2                # NULL_IRQ_TIMER
                 max_sectors=$((4096/512))
                 memory_backed=1
                 pipeline_zoned_writes="${pipelining}"
                 size=1
                 submit_queues=1
                 zone_size=1
                 zoned=1
                 power=1
             )
             for p in "${params[@]}"; do
                 echo "${p//*=}" > "${p//=*}"
             done
         )
         udevadm settle
         dev=/dev/nullb0
         [ -b "${dev}" ]
         params=(
             --direct=1
             --filename="${dev}"
             --iodepth=64
             --iodepth_batch=16
             --ioengine=io_uring
             --ioscheduler="${iosched}"
             --gtod_reduce=1
             --hipri=0
Quoted 9 lines
             --name=nullb0
             --runtime=30
             --rw=write
             --time_based=1
             --zonemode=zbd
         )
         fio "${params[@]}"
     done

Is it possible to move this into the blktests ?

-ck

#13 lore raw

Re: [PATCH v3 6/8] block/null_blk: Add support for pipelining zoned writes

Quoted 1 line · Chaitanya Kulkarni

Is it possible to move this into the blktests ?

That sounds like a good idea to me. I will look into that after this patch series has been merged.

Thanks,

Bart.

#14 lore raw

PATCH v3 7/8 nvme: Make the number of retries command specific

In reply to Bart Van Assche

Add support for specifying the number of retries per NVMe command.

Cc Christoph Hellwig <[email protected]>
Cc Keith Busch <[email protected]>
Cc Sagi Grimberg <[email protected]>
Cc Chaitanya Kulkarni <[email protected]>
---
 drivers/nvme/host/core.c | 3 ++-
 drivers/nvme/host/nvme.h | 1 +
 2 files changed, 3 insertions(+), 1 deletion(-)
drivers/nvme/host/core.c +2 -1
index b3d9c29aba1e..df9ac7fab9b8 100644
@@ -339,7 +339,7 @@ static inline enum nvme_disposition nvme_decide_disposition(struct request *req)
 
 	if (blk_noretry_request(req) ||
 	    (nvme_req(req)->status & NVME_SC_DNR) ||
-	    nvme_req(req)->retries >= nvme_max_retries)
+	    nvme_req(req)->retries >= nvme_req(req)->max_retries)
 		return COMPLETE;
 
 	if (req->cmd_flags & REQ_NVME_MPATH) {
@@ -632,6 +632,7 @@ static inline void nvme_clear_nvme_request(struct request *req)
 {
 	nvme_req(req)->status = 0;
 	nvme_req(req)->retries = 0;
+	nvme_req(req)->max_retries = nvme_max_retries;
 	nvme_req(req)->flags = 0;
 	req->rq_flags |= RQF_DONTPREP;
 }
drivers/nvme/host/nvme.h +1 -0
index 0da94b233fed..ca415cd9571e 100644
@@ -160,6 +160,7 @@ struct nvme_request {
 	union nvme_result	result;
 	u8			genctr;
 	u8			retries;
+	u8			max_retries;
 	u8			flags;
 	u16			status;
 	struct nvme_ctrl	*ctrl;
#15 lore raw

Re: [PATCH v3 7/8] nvme: Make the number of retries command specific

drivers/nvme/host/nvme.h struct nvme_request { · line 166

Quoted 13 lines
drivers/nvme/host/nvme.h +1 -0
index 0da94b233fed..ca415cd9571e 100644
@@ -160,6 +160,7 @@ struct nvme_request {
  	union nvme_result	result;
  	u8			genctr;
  	u8			retries;
+	u8			max_retries;
  	u8			flags;
  	u16			status;
  	struct nvme_ctrl	*ctrl;

If I understand correctly then per command max_retries count is only needed for zoned devices.

why not make struct nvme_request->max_retries field and subsequent code configurable under CONFIG_BLK_DEV_ZONED ?

That will avoid increasing size of the nvme_request for !CONFIG_BLK_DEV_ZONED case where per command nvme_request->max_retries has no use.

-ck

#16 lore raw

Re: [PATCH v3 7/8] nvme: Make the number of retries command specific

Quoted 17 lines · Chaitanya Kulkarni

drivers/nvme/host/nvme.h struct nvme_request { · line 166

Quoted 13 lines
drivers/nvme/host/nvme.h +1 -0
index 0da94b233fed..ca415cd9571e 100644
@@ -160,6 +160,7 @@ struct nvme_request {
   	union nvme_result	result;
   	u8			genctr;
   	u8			retries;
+	u8			max_retries;
   	u8			flags;
   	u16			status;
   	struct nvme_ctrl	*ctrl;

If I understand correctly then per command max_retries count is only needed for zoned devices.

Quoted 6 lines

why not make struct nvme_request->max_retries field and subsequent code configurable under CONFIG_BLK_DEV_ZONED ?

That will avoid increasing size of the nvme_request for !CONFIG_BLK_DEV_ZONED case where per command nvme_request->max_retries has no use.

Hi Chaitanya,

Thanks for the review.

We may disagree about whether or not this patch increases the size of struct nvme_request. I think the new member fills an existing hole and hence does not increase the size of struct nvme_request :-)

Bart.

#17 lore raw

PATCH v3 8/8 nvme: Enable pipelining of zoned writes

In reply to Bart Van Assche

Enabling pipelining for zoned writes. Increase the number of retries for zoned writes to the maximum number of outstanding commands per hardware queue.

Cc Christoph Hellwig <[email protected]>
Cc Keith Busch <[email protected]>
Cc Sagi Grimberg <[email protected]>
Cc Chaitanya Kulkarni <[email protected]>
---
 drivers/nvme/host/core.c | 6 ++++++
 drivers/nvme/host/zns.c  | 2 ++
 2 files changed, 8 insertions(+)
drivers/nvme/host/core.c +6 -0
index df9ac7fab9b8..3f71d3d82d5e 100644
@@ -854,6 +854,12 @@ static inline blk_status_t nvme_setup_rw(struct nvme_ns *ns,
 	if (req->cmd_flags & REQ_RAHEAD)
 		dsmgmt |= NVME_RW_DSM_FREQ_PREFETCH;
 
+	if (blk_queue_pipeline_zoned_writes(req->q) &&
+	    blk_rq_is_seq_zone_write(req))
+		nvme_req(req)->max_retries =
+			min(0UL + type_max(typeof(nvme_req(req)->max_retries)),
+			    nvme_req(req)->max_retries + req->q->nr_requests);
+
 	cmnd->rw.opcode = op;
 	cmnd->rw.flags = 0;
 	cmnd->rw.nsid = cpu_to_le32(ns->head->ns_id);
drivers/nvme/host/zns.c +2 -0
index 9f81beb4df4e..0b10db3b8d3a 100644
@@ -54,6 +54,8 @@ int nvme_update_zone_info(struct nvme_ns *ns, unsigned lbaf)
 	struct nvme_id_ns_zns *id;
 	int status;
 
+	blk_queue_flag_set(QUEUE_FLAG_PIPELINE_ZONED_WRITES, ns->queue);
+
 	/* Driver requires zone append support */
 	if ((le32_to_cpu(log->iocs[nvme_cmd_zone_append]) &
 			NVME_CMD_EFFECTS_CSUPP)) {
maintainer
#18 lore raw

Re: [PATCH v3 8/8] nvme: Enable pipelining of zoned writes

drivers/nvme/host/core.c static inline blk_status_t nvme_setup_rw(struct nvme_ns *ns, · line 861

Quoted 9 lines · Bart Van Assche
drivers/nvme/host/core.c +5 -0
@@ -854,6 +854,12 @@ static inline blk_status_t nvme_setup_rw(struct nvme_ns *ns,
 	if (req->cmd_flags & REQ_RAHEAD)
 		dsmgmt |= NVME_RW_DSM_FREQ_PREFETCH;
 
+	if (blk_queue_pipeline_zoned_writes(req->q) &&
+	    blk_rq_is_seq_zone_write(req))
+		nvme_req(req)->max_retries =
+			min(0UL + type_max(typeof(nvme_req(req)->max_retries)),
+			    nvme_req(req)->max_retries + req->q->nr_requests);

I can't make much sense of what the above is trying to accomplish. This reevaluates max_retries every time the request is retried, and the new max_retries is based on the previous max_retries?

#19 lore raw

Re: [PATCH v3 8/8] nvme: Enable pipelining of zoned writes

Quoted 5 lines · Keith Busch
Quoted 4 lines · Bart Van Assche
drivers/nvme/host/core.c +0 -0
@@ -854,6 +854,12 @@ static inline blk_status_t nvme_setup_rw(struct nvme_ns *ns,
  	if (req->cmd_flags & REQ_RAHEAD)
  		dsmgmt |= NVME_RW_DSM_FREQ_PREFETCH;
 
Quoted 9 lines

drivers/nvme/host/core.c static inline blk_status_t nvme_setup_rw(struct nvme_ns *ns, · line 861

Quoted 5 lines
drivers/nvme/host/core.c +5 -0
+	if (blk_queue_pipeline_zoned_writes(req->q) &&
+	    blk_rq_is_seq_zone_write(req))
+		nvme_req(req)->max_retries =
+			min(0UL + type_max(typeof(nvme_req(req)->max_retries)),
+			    nvme_req(req)->max_retries + req->q->nr_requests);

I can't make much sense of what the above is trying to accomplish. This reevaluates max_retries every time the request is retried, and the new max_retries is based on the previous max_retries?

I also had a hard time quickly reading the code but that maybe just me.

Perhaps a well documented comment or a helper with comment explaining the logic will be helpful here ?

-ck

#20 lore raw

Re: [PATCH v3 8/8] nvme: Enable pipelining of zoned writes

In reply to Keith Busch
Quoted 5 lines · Keith Busch
Quoted 4 lines · Bart Van Assche
(quoted diff) +0 -0
@@ -854,6 +854,12 @@ static inline blk_status_t nvme_setup_rw(struct nvme_ns *ns,
  	if (req->cmd_flags & REQ_RAHEAD)
  		dsmgmt |= NVME_RW_DSM_FREQ_PREFETCH;
 
Quoted 9 lines

Review static inline blk_status_t nvme_setup_rw(struct nvme_ns *ns, · line 861

Quoted 5 lines
(quoted diff) +5 -0
+	if (blk_queue_pipeline_zoned_writes(req->q) &&
+	    blk_rq_is_seq_zone_write(req))
+		nvme_req(req)->max_retries =
+			min(0UL + type_max(typeof(nvme_req(req)->max_retries)),
+			    nvme_req(req)->max_retries + req->q->nr_requests);

I can't make much sense of what the above is trying to accomplish. This reevaluates max_retries every time the request is retried, and the new max_retries is based on the previous max_retries?

Hi Keith,

Thanks for your question. It made me realize that the above code should be moved. How about replacing patch 8/8 with the (untested) patch below?

Thanks,

Bart.

Subject: [PATCH] nvme: Enable pipelining of zoned writes

Enabling pipelining for zoned writes. Increase the number of retries for zoned writes to the maximum number of outstanding commands per hardware queue.

drivers/nvme/host/core.c +6 -1
index aeeaca1c3197..6a0b824a343f 100644
@@ -630,10 +630,15 @@ EXPORT_SYMBOL_NS_GPL(nvme_put_ns, NVME_TARGET_PASSTHRU);
 
 static inline void nvme_clear_nvme_request(struct request *req)
 {
+	u32 max_retries = nvme_max_retries;
+
 	nvme_req(req)->status = 0;
 	nvme_req(req)->retries = 0;
 	nvme_req(req)->flags = 0;
-	nvme_req(req)->max_retries = nvme_max_retries;
+	if (blk_queue_pipeline_zoned_writes(req->q) &&
+	    blk_rq_is_seq_zone_write(req))
+		max_retries += req->q->nr_requests;
+	nvme_req(req)->max_retries = min(0xffU, max_retries);
 	req->rq_flags |= RQF_DONTPREP;
 }
 
drivers/nvme/host/zns.c +2 -0
index 9f81beb4df4e..0b10db3b8d3a 100644
@@ -54,6 +54,8 @@ int nvme_update_zone_info(struct nvme_ns *ns, unsigned lbaf)
 	struct nvme_id_ns_zns *id;
 	int status;
 
+	blk_queue_flag_set(QUEUE_FLAG_PIPELINE_ZONED_WRITES, ns->queue);
+
 	/* Driver requires zone append support */
 	if ((le32_to_cpu(log->iocs[nvme_cmd_zone_append]) &
 			NVME_CMD_EFFECTS_CSUPP)) {
#21 lore raw

Re: [PATCH v3 8/8] nvme: Enable pipelining of zoned writes

In reply to Bart Van Assche
Quoted 3 lines · Bart Van Assche

Enabling pipelining for zoned writes. Increase the number of retries for zoned writes to the maximum number of outstanding commands per hardware queue.

How is this supposed to work? NVMe controllers are completely free to reorder. It also doesn't make sense as all zoned writes in Linux either use zone append or block layer based zone locking.

#22 lore raw

Re: [PATCH v3 8/8] nvme: Enable pipelining of zoned writes

Quoted 8 lines · Christoph Hellwig
Quoted 3 lines · Bart Van Assche

Enabling pipelining for zoned writes. Increase the number of retries for zoned writes to the maximum number of outstanding commands per hardware queue.

How is this supposed to work? NVMe controllers are completely free to reorder. It also doesn't make sense as all zoned writes in Linux either use zone append or block layer based zone locking.

Agreed that the NVMe specification allows to reorder outstanding commands. Are there any NVMe controllers that do this if multiple zoned write commands are outstanding for a single zone? I do not expect that an NVMe controller would reorder write commands in such a way that an I/O error is introduced.

Regarding zoned writes in Linux, Android devices use F2FS. F2FS submits regular writes (REQ_OP_WRITE) to zoned devices. Patch 4/8 from this patch series disables zone locking in the mq-deadline scheduler if the block driver has set the QUEUE_FLAG_PIPELINE_ZONED_WRITES flag.

Thanks,

Bart.

#23 lore raw

Re: [PATCH v3 8/8] nvme: Enable pipelining of zoned writes

Quoted 4 lines · Bart Van Assche

Regarding zoned writes in Linux, Android devices use F2FS. F2FS submits regular writes (REQ_OP_WRITE) to zoned devices. Patch 4/8 from this patch series disables zone locking in the mq-deadline scheduler if the block driver has set the QUEUE_FLAG_PIPELINE_ZONED_WRITES flag.

Please fix f2fs to use zone appends rather than working around the stack.

#24 lore raw

Re: [PATCH v3 8/8] nvme: Enable pipelining of zoned writes

In reply to Bart Van Assche

[look like my previously reply accidentally lost the first half,
so here is is]

Quoted 5 lines · Bart Van Assche

Agreed that the NVMe specification allows to reorder outstanding commands. Are there any NVMe controllers that do this if multiple zoned write commands are outstanding for a single zone? I do not expect that an NVMe controller would reorder write commands in such a way that an I/O error is introduced.

NVMe not only allows reordering, but actually requires it due to the round robin command arbitrary. Moreoever once you are on IP based transports there is plenty of reordering that can and will go on before even reaching the controller.