[PATCH v4 00/12] dm: support IO polling

[Changes since v3]
- newly add patch 7 and patch 11, as a new optimization improving
performance of multiple polling processes. Now performance of multiple
polling processes can be as scalable as single polling process (~30%).
Refer to the following [Performance] chapter for more details.


[Intention]
Bio-based polling (e.g., for dm/md devices) is one indispensable part of
high performance IO stack. As far as I know, dm (e.g., dm-stripe) is
widely used in database, splicing several NVMe disks as one whole disk,
in hope of achieving better performance. With this patch set, io_uring
could be used upon dm devices.


[Optimizations]
Basically, there are three paths for IO polling.

1. fastpath (patch 9/10)
The polling routine will go into this path when bio submitted to dm
device is not split.

In this case, there will be only one bio submitted to only one polling
hw queue of one underlying mq device, and thus we don't need to track
all split bios or iterate through all polling hw queues. The pointer to
the polling hw queue the bio submitted to is returned here as the
returned cookie. In this case, the polling routine will call
mq_ops->poll() directly with the hw queue converted from the input
cookie.


- One process reading dm-linear (mapping to three underlying NVMe devices,
with one polling hw queue per NVMe device).

(ioengine=io_uring, iodepth=128, numjobs=1, rw=randread, sqthread_poll=0
direct=1, bs=4k)

	    	 | IOPS (IRQ mode) | IOPS (iopoll=1 mode) | diff
---------------- | --------------- | -------------------- | ----
with patchset    |	      212k |		     284k | ~32%


- Three processes reading dm-linear (mapping to three underlying NVMe
devices, with one polling hw queue per NVMe device).

(ioengine=io_uring, iodepth=128, numjobs=3, rw=randread, sqthread_poll=0
direct=1, bs=4k)

	    	 | IOPS (IRQ mode) | IOPS (iopoll=1 mode) | diff
---------------- | --------------- | -------------------- | ----
with patchset    |	      615k |		     735k | ~16%


- Three processes reading dm-linear (mapping to three underlying NVMe
devices, with three polling hw queues per NVMe device), with every
process pinned to one CPU and mapped to one exclusive hw queue.

(ioengine=io_uring, iodepth=128, numjobs=3, rw=randread, sqthread_poll=0
direct=1, bs=4k)

	    	 | IOPS (IRQ mode) | IOPS (iopoll=1 mode) | diff
---------------- | --------------- | -------------------- | ----
with patchset    |	      631k |		     833k | ~32%



2. sub-fastpath (patch 7/11)

The polling routine will go into this path when bio submitted to dm
device gets split and enqueued into multiple hw queues, while the IO
submission process has not been migrated to another CPU.

In this case, the IO submission routine will return the CPU number on
which the IO submission happened as the returned cookie, while the
polling routine will only iterate and poll on hw queues that this CPU
number maps, instead of iterating *all* hw queues.

This optimization can dramatically reduce cache ping-pong and thus
improve the polling performance, when multiple hw queues in polling mode
per device could be reserved when there are multiple polling processes.

- Three processes reading dm-stripe (mapping to three underlying NVMe
devices, with three polling hw queues per NVMe device), with every
process pinned to one CPU and mapped to one exclusive hw queue.

(ioengine=io_uring, iodepth=128, numjobs=3, rw=randread, sqthread_poll=0
direct=1, bs=12k(4k for every NVMe device))

	    	 | IOPS (IRQ mode) | IOPS (iopoll=1 mode) | diff
---------------- | --------------- | -------------------- | ----
with patchset    |	      307k |		     412k | ~34%


3. default path

It will fall back to iterating all hw queues in polling mode, once bio
submitted to dm device gets split and enqueued into multiple hw queues,
and the IO process has ever been migrated to another CPU during the IO
submission phase.


[Remained Issue]
It has been mentioned in patch 4 that, users could change the state of
the underlying devices through '/sys/block/<dev>/io_poll', bypassing
the dm device above. Thus it can cause a situation where QUEUE_FLAG_POLL
is still set for the request_queue of dm device, while one of the
underlying mq device may has cleared this flag.

In this case, it will pass the 'test_bit(QUEUE_FLAG_POLL, &q->queue_flags)'
check in blk_poll(), while the input cookie may actually points to a hw
queue in IRQ mode since patch 11. Thus for this hw queue (in IRQ mode),
the bio-based polling routine will handle this hw queue acquiring
'spin_lock(&nvmeq->cq_poll_lock)' (refer
drivers/nvme/host/pci.c:nvme_poll), which is not adequate since this hw
queue may also be accessed in IRQ context. In other words,
spin_lock_irq() should be used here.

I have not come up one simple way to fix it. I don't want to do sanity
check (e.g., the type of the hw queue is HCTX_TYPE_POLL or not) in the
IO path (submit_bio()/blk_poll()), i.e., fast path.

We'd better fix it in the control path, i.e., dm could be aware of the
change when attribute (e.g., support io_poll or not) of one of the
underlying devices changed at runtime.




[Changes since v2]

Patchset v2 caches all hw queues (in polling mode) of underlying mq
devices in dm layer. The polling routine actually iterates through all
these cached hw queues.

However, mq may change the queue mapping at runtime (e.g., NVMe RESET
command), thus the cached hw queues in dm layer may be out-of-date. Thus
patchset v3 falls back to the implementation of the very first RFC
version, in which the mq layer needs to export one interface iterating
all polling hw queues (patch 5), and the bio-based polling routine just
calls this interface to iterate all polling hw queues.

Besides, several new optimization is proposed.


- patch 1,2,7
same as v2, untouched

- patch 3
Considering advice from Christoph Hellwig, while refactoring blk_poll(),
split mq and bio-based polling routine from the very beginning. Now
blk_poll() is just a simple entry. blk_bio_poll() is simply copied from
blk_mq_poll(), while the loop structure is some sort of duplication
though.

- patch 4
This patch is newly added to support turning on/off polling through
'/sys/block/<dev>/queue/io_poll' dynamiclly for bio-based devices.
Patchset v2 implemented this functionality by added one new queue flag,
which is not preferred since the queue flag resource is quite short of
nowadays.

- patch 5
This patch is newly added, preparing for the following bio-based
polling. The following bio-based polling will call this helper function,
accounting on the corresponding hw queue.

- patch 6
It's from the very first RFC version, preparing for the following
bio-based polling.

- patch 8
One fixing patch needed by the following bio-based polling. It's
actually a v2 of [1]. I had sent the v2 singly in-reply-to [1], though
it has not been visible on the mailing list maybe due to the delay.

- patch 9
It's from the very first RFC version.

- patch 10
This patch is newly added. Patchset v2 had ever proposed one
optimization that, skipping the **busy** hw queues during the iteration
phase. Back upon that time, one flag of 'atomic_t' is specifically
maintained in dm layer, representing if the corresponding hw queue is
busy or not. The idea is inherited, while the implementation changes.
Now @nvmeq->cq_poll_lock is used directly here, no need for extra flag
anymore.

This optimization can significantly reduce the competition for one hw
queue between multiple polling instances. Following statistics is the
test result when 3 threads concurrently randread (bs=4k, direct=1) one
dm-linear device, which is built upon 3 nvme devices, with one polling
hw queue per nvme device.

	    | IOPS (IRQ mode) | IOPS (iopoll=1 mode) | diff
----------- | --------------- | -------------------- | ----
without opt | 		 318k |		 	256k | ~-20%
with opt    |		 314k |		 	354k | ~13%
							

- patch 11
This is another newly added optimizatin for bio-based polling.

One intuitive insight is that, when the original bio submitted to dm
device doesn't get split, then the bio gets enqueued into only one hw
queue of one of the underlying mq devices. In this case, we no longer
need to track all split bios, and one cookie (for the only split bio)
is enough. It is implemented by returning the pointer to the
corresponding hw queue in this case.

It should be safe by directly returning the pointer to the hw queue,
since 'struct blk_mq_hw_ctx' won't be freed during the whole lifetime of
'struct request_queue'. Even when the number of hw queues may decrease
when NVMe RESET happens, the 'struct request_queue' structure of decreased
hw queues won't be freed, instead it's buffered into
&q->unused_hctx_list list.

Though this optimization seems quite intuitive, the performance test
shows that it does no benefit nor harm to the performance, while 3
threads concurrently randreading (bs=4k, direct=1) one dm-linear
device, which is built upon 3 nvme devices, with one polling hw queue
per nvme device.

I'm not sure why it doesn't work, maybe because the number of devices,
or the depth of the devcice stack is to low in my test case?


changes since v1:
- patch 1,2,4 is the same as v1 and have already been reviewed
- patch 3 is refactored a bit on the basis of suggestions from
Mike Snitzer.
- patch 5 is newly added and introduces one new queue flag
representing if the queue is capable of IO polling. This mainly
simplifies the logic in queue_poll_store().
- patch 6 implements the core mechanism supporting IO polling.
The sanity check checking if the dm device supports IO polling is
also folded into this patch, and the queue flag will be cleared if
it doesn't support, in case of table reloading.




Jeffle Xu (12):
  block: move definition of blk_qc_t to types.h
  block: add queue_to_disk() to get gendisk from request_queue
  block: add poll method to support bio-based IO polling
  block: add poll_capable method to support bio-based IO polling
  blk-mq: extract one helper function polling hw queue
  blk-mq: add iterator for polling hw queues
  blk-mq: add one helper function getting hw queue
  dm: always return BLK_QC_T_NONE for bio-based device
  nvme/pci: don't wait for locked polling queue
  block: fastpath for bio-based polling
  block: sub-fastpath for bio-based polling
  dm: support IO polling for bio-based dm device

 block/blk-core.c              | 112 +++++++++++++++++++++++++++++++++-
 block/blk-mq.c                |  37 ++++-------
 block/blk-sysfs.c             |  14 ++++-
 drivers/md/dm-table.c         |  26 ++++++++
 drivers/md/dm.c               | 102 ++++++++++++++++++++++++++-----
 drivers/nvme/host/pci.c       |   4 +-
 include/linux/blk-mq.h        |  23 +++++++
 include/linux/blk_types.h     |  66 +++++++++++++++++++-
 include/linux/blkdev.h        |   4 ++
 include/linux/device-mapper.h |   1 +
 include/linux/fs.h            |   2 +-
 include/linux/types.h         |   3 +
 include/trace/events/kyber.h  |   6 +-
 13 files changed, 350 insertions(+), 50 deletions(-)

-- 
2.27.0

So that kiocb.ki_cookie can be defined as blk_qc_t, which will enforce
the encapsulation.

Signed-off-by: Jeffle Xu <jefflexu@linux.alibaba.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Mike Snitzer <redacted>
---
 include/linux/blk_types.h | 2 +-
 include/linux/fs.h        | 2 +-
 include/linux/types.h     | 3 +++
 3 files changed, 5 insertions(+), 2 deletions(-)

diff --git a/include/linux/blk_types.h b/include/linux/blk_types.h
index 866f74261b3b..2e05244fc16d 100644
--- a/include/linux/blk_types.h
+++ b/include/linux/blk_types.h

@@ -532,7 +532,7 @@ static inline int op_stat_group(unsigned int op)
 	return op_is_write(op);
 }
 
-typedef unsigned int blk_qc_t;
+/* Macros for blk_qc_t */
 #define BLK_QC_T_NONE		-1U
 #define BLK_QC_T_SHIFT		16
 #define BLK_QC_T_INTERNAL	(1U << 31)

diff --git a/include/linux/fs.h b/include/linux/fs.h
index fd47deea7c17..04b687150736 100644
--- a/include/linux/fs.h
+++ b/include/linux/fs.h

@@ -330,7 +330,7 @@ struct kiocb {
 	u16			ki_hint;
 	u16			ki_ioprio; /* See linux/ioprio.h */
 	union {
-		unsigned int		ki_cookie; /* for ->iopoll */
+		blk_qc_t		ki_cookie; /* for ->iopoll */
 		struct wait_page_queue	*ki_waitq; /* for async buffered IO */
 	};
 

diff --git a/include/linux/types.h b/include/linux/types.h
index a147977602b5..da5ca7e1bea9 100644
--- a/include/linux/types.h
+++ b/include/linux/types.h

@@ -125,6 +125,9 @@ typedef s64			int64_t;
 typedef u64 sector_t;
 typedef u64 blkcnt_t;
 
+/* cookie used for IO polling */
+typedef unsigned int blk_qc_t;
+
 /*
  * The type of an index into the pagecache.
  */

-- 
2.27.0

Sometimes we need to get the corresponding gendisk from request_queue.

It is preferred that block drivers store private data in
gendisk->private_data rather than request_queue->queuedata, e.g. see:
commit c4a59c4e5db3 ("dm: stop using ->queuedata").

So if only request_queue is given, we need to get its corresponding
gendisk to get the private data stored in that gendisk.

Signed-off-by: Jeffle Xu <jefflexu@linux.alibaba.com>
Reviewed-by: Mike Snitzer <redacted>
---
 include/linux/blkdev.h       | 2 ++
 include/trace/events/kyber.h | 6 +++---
 2 files changed, 5 insertions(+), 3 deletions(-)

diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index f94ee3089e01..2a802e011a95 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h

@@ -687,6 +687,8 @@ static inline bool blk_account_rq(struct request *rq)
 	dma_map_page_attrs(dev, (bv)->bv_page, (bv)->bv_offset, (bv)->bv_len, \
 	(dir), (attrs))
 
+#define queue_to_disk(q)	(dev_to_disk(kobj_to_dev((q)->kobj.parent)))
+
 static inline bool queue_is_mq(struct request_queue *q)
 {
 	return q->mq_ops;

diff --git a/include/trace/events/kyber.h b/include/trace/events/kyber.h
index c0e7d24ca256..f9802562edf6 100644
--- a/include/trace/events/kyber.h
+++ b/include/trace/events/kyber.h

@@ -30,7 +30,7 @@ TRACE_EVENT(kyber_latency,
 	),
 
 	TP_fast_assign(
-		__entry->dev		= disk_devt(dev_to_disk(kobj_to_dev(q->kobj.parent)));
+		__entry->dev		= disk_devt(queue_to_disk(q));
 		strlcpy(__entry->domain, domain, sizeof(__entry->domain));
 		strlcpy(__entry->type, type, sizeof(__entry->type));
 		__entry->percentile	= percentile;

@@ -59,7 +59,7 @@ TRACE_EVENT(kyber_adjust,
 	),
 
 	TP_fast_assign(
-		__entry->dev		= disk_devt(dev_to_disk(kobj_to_dev(q->kobj.parent)));
+		__entry->dev		= disk_devt(queue_to_disk(q));
 		strlcpy(__entry->domain, domain, sizeof(__entry->domain));
 		__entry->depth		= depth;
 	),

@@ -81,7 +81,7 @@ TRACE_EVENT(kyber_throttled,
 	),
 
 	TP_fast_assign(
-		__entry->dev		= disk_devt(dev_to_disk(kobj_to_dev(q->kobj.parent)));
+		__entry->dev		= disk_devt(queue_to_disk(q));
 		strlcpy(__entry->domain, domain, sizeof(__entry->domain));
 	),
 

-- 
2.27.0

This method can be used to check if bio-based device supports IO polling
or not. For mq devices, checking for hw queue in polling mode is
adequate, while the sanity check shall be implementation specific for
bio-based devices. For example, dm device needs to check if all
underlying devices are capable of IO polling.

Though bio-based device may have done the sanity check during the
device initialization phase, cacheing the result of this sanity check
(such as by cacheing in the queue_flags) may not work. Because for dm
devices, users could change the state of the underlying devices through
'/sys/block/<dev>/io_poll', bypassing the dm device above. In this case,
the cached result of the very beginning sanity check could be
out-of-date. Thus the sanity check needs to be done every time 'io_poll'
is to be modified.

Signed-off-by: Jeffle Xu <jefflexu@linux.alibaba.com>
---
 block/blk-sysfs.c      | 14 +++++++++++---
 include/linux/blkdev.h |  1 +
 2 files changed, 12 insertions(+), 3 deletions(-)

diff --git a/block/blk-sysfs.c b/block/blk-sysfs.c
index b513f1683af0..f11fedefc37d 100644
--- a/block/blk-sysfs.c
+++ b/block/blk-sysfs.c

@@ -420,9 +420,17 @@ static ssize_t queue_poll_store(struct request_queue *q, const char *page,
 	unsigned long poll_on;
 	ssize_t ret;
 
-	if (!q->tag_set || q->tag_set->nr_maps <= HCTX_TYPE_POLL ||
-	    !q->tag_set->map[HCTX_TYPE_POLL].nr_queues)
-		return -EINVAL;
+	if (queue_is_mq(q)) {
+		if (!q->tag_set || q->tag_set->nr_maps <= HCTX_TYPE_POLL ||
+		    !q->tag_set->map[HCTX_TYPE_POLL].nr_queues)
+			return -EINVAL;
+	} else {
+		struct gendisk *disk = queue_to_disk(q);
+
+		if (!disk->fops->poll_capable ||
+		    !disk->fops->poll_capable(disk))
+			return -EINVAL;
+	}
 
 	ret = queue_var_store(&poll_on, page, count);
 	if (ret < 0)

diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index ebe16f55cba4..8a84088642ce 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h

@@ -1854,6 +1854,7 @@ static inline void blk_ksm_unregister(struct request_queue *q) { }
 struct block_device_operations {
 	blk_qc_t (*submit_bio) (struct bio *bio);
 	int (*poll)(struct request_queue *q, blk_qc_t cookie);
+	bool (*poll_capable)(struct gendisk *disk);
 	int (*open) (struct block_device *, fmode_t);
 	void (*release) (struct gendisk *, fmode_t);
 	int (*rw_page)(struct block_device *, sector_t, struct page *, unsigned int);

-- 
2.27.0

Currently the returned cookie of bio-based device is not used at all.

Cookie of bio-based device will be refactored in the following patch.

Signed-off-by: Jeffle Xu <jefflexu@linux.alibaba.com>
Reviewed-by: Mike Snitzer <redacted>
---
 drivers/md/dm.c | 26 ++++++++++----------------
 1 file changed, 10 insertions(+), 16 deletions(-)

diff --git a/drivers/md/dm.c b/drivers/md/dm.c
index 8a3d73efb9dd..c2945c90745e 100644
--- a/drivers/md/dm.c
+++ b/drivers/md/dm.c

@@ -1252,14 +1252,13 @@ void dm_accept_partial_bio(struct bio *bio, unsigned n_sectors)
 }
 EXPORT_SYMBOL_GPL(dm_accept_partial_bio);
 
-static blk_qc_t __map_bio(struct dm_target_io *tio)
+static void __map_bio(struct dm_target_io *tio)
 {
 	int r;
 	sector_t sector;
 	struct bio *clone = &tio->clone;
 	struct dm_io *io = tio->io;
 	struct dm_target *ti = tio->ti;
-	blk_qc_t ret = BLK_QC_T_NONE;
 
 	clone->bi_end_io = clone_endio;
 

@@ -1278,7 +1277,7 @@ static blk_qc_t __map_bio(struct dm_target_io *tio)
 	case DM_MAPIO_REMAPPED:
 		/* the bio has been remapped so dispatch it */
 		trace_block_bio_remap(clone, bio_dev(io->orig_bio), sector);
-		ret = submit_bio_noacct(clone);
+		submit_bio_noacct(clone);
 		break;
 	case DM_MAPIO_KILL:
 		free_tio(tio);

@@ -1292,8 +1291,6 @@ static blk_qc_t __map_bio(struct dm_target_io *tio)
 		DMWARN("unimplemented target map return value: %d", r);
 		BUG();
 	}
-
-	return ret;
 }
 
 static void bio_setup_sector(struct bio *bio, sector_t sector, unsigned len)

@@ -1380,7 +1377,7 @@ static void alloc_multiple_bios(struct bio_list *blist, struct clone_info *ci,
 	}
 }
 
-static blk_qc_t __clone_and_map_simple_bio(struct clone_info *ci,
+static void __clone_and_map_simple_bio(struct clone_info *ci,
 					   struct dm_target_io *tio, unsigned *len)
 {
 	struct bio *clone = &tio->clone;

@@ -1391,7 +1388,7 @@ static blk_qc_t __clone_and_map_simple_bio(struct clone_info *ci,
 	if (len)
 		bio_setup_sector(clone, ci->sector, *len);
 
-	return __map_bio(tio);
+	__map_bio(tio);
 }
 
 static void __send_duplicate_bios(struct clone_info *ci, struct dm_target *ti,

@@ -1405,7 +1402,7 @@ static void __send_duplicate_bios(struct clone_info *ci, struct dm_target *ti,
 
 	while ((bio = bio_list_pop(&blist))) {
 		tio = container_of(bio, struct dm_target_io, clone);
-		(void) __clone_and_map_simple_bio(ci, tio, len);
+		__clone_and_map_simple_bio(ci, tio, len);
 	}
 }
 

@@ -1450,7 +1447,7 @@ static int __clone_and_map_data_bio(struct clone_info *ci, struct dm_target *ti,
 		free_tio(tio);
 		return r;
 	}
-	(void) __map_bio(tio);
+	__map_bio(tio);
 
 	return 0;
 }

@@ -1565,11 +1562,10 @@ static void init_clone_info(struct clone_info *ci, struct mapped_device *md,
 /*
  * Entry point to split a bio into clones and submit them to the targets.
  */
-static blk_qc_t __split_and_process_bio(struct mapped_device *md,
+static void __split_and_process_bio(struct mapped_device *md,
 					struct dm_table *map, struct bio *bio)
 {
 	struct clone_info ci;
-	blk_qc_t ret = BLK_QC_T_NONE;
 	int error = 0;
 
 	init_clone_info(&ci, md, map, bio);

@@ -1613,7 +1609,7 @@ static blk_qc_t __split_and_process_bio(struct mapped_device *md,
 
 				bio_chain(b, bio);
 				trace_block_split(b, bio->bi_iter.bi_sector);
-				ret = submit_bio_noacct(bio);
+				submit_bio_noacct(bio);
 				break;
 			}
 		}

@@ -1621,13 +1617,11 @@ static blk_qc_t __split_and_process_bio(struct mapped_device *md,
 
 	/* drop the extra reference count */
 	dec_pending(ci.io, errno_to_blk_status(error));
-	return ret;
 }
 
 static blk_qc_t dm_submit_bio(struct bio *bio)
 {
 	struct mapped_device *md = bio->bi_disk->private_data;
-	blk_qc_t ret = BLK_QC_T_NONE;
 	int srcu_idx;
 	struct dm_table *map;
 

@@ -1657,10 +1651,10 @@ static blk_qc_t dm_submit_bio(struct bio *bio)
 	if (is_abnormal_io(bio))
 		blk_queue_split(&bio);
 
-	ret = __split_and_process_bio(md, map, bio);
+	__split_and_process_bio(md, map, bio);
 out:
 	dm_put_live_table(md, srcu_idx);
-	return ret;
+	return BLK_QC_T_NONE;
 }
 
 /*-----------------------------------------------------------------

-- 
2.27.0

Offer one fastpath for bio-based polling when bio submitted to dm
device is not split.

In this case, there will be only one bio submitted to only one polling
hw queue of one underlying mq device, and thus we don't need to track
all split bios or iterate through all polling hw queues. The pointer to
the polling hw queue the bio submitted to is returned here as the
returned cookie. In this case, the polling routine will call
mq_ops->poll() directly with the hw queue converted from the input
cookie.

If the original bio submitted to dm device is split to multiple bios and
thus submitted to multiple polling hw queues, the polling routine will
fall back to iterating all hw queues (in polling mode) of all underlying
mq devices.

Signed-off-by: Jeffle Xu <jefflexu@linux.alibaba.com>
---
 block/blk-core.c          | 38 ++++++++++++++++++++++++++++++++++++--
 include/linux/blk_types.h |  8 ++++++++
 include/linux/types.h     |  2 +-
 3 files changed, 45 insertions(+), 3 deletions(-)

diff --git a/block/blk-core.c b/block/blk-core.c
index 37aa513da5f2..82349e3c2a24 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c

@@ -956,11 +956,19 @@ static blk_qc_t __submit_bio(struct bio *bio)
  * bio_list_on_stack[0] contains bios submitted by the current ->submit_bio.
  * bio_list_on_stack[1] contains bios that were submitted before the current
  *	->submit_bio_bio, but that haven't been processed yet.
+ *
+ * Return:
+ *   - BLK_QC_T_NONE, no need for IO polling.
+ *   - BLK_QC_T_BIO_POLL_ALL, @bio gets split and enqueued into multi hw queues.
+ *   - Otherwise, @bio is not split, returning the pointer to the corresponding
+ *     hw queue that the bio enqueued into as the returned cookie.
  */
 static blk_qc_t __submit_bio_noacct(struct bio *bio)
 {
 	struct bio_list bio_list_on_stack[2];
 	blk_qc_t ret = BLK_QC_T_NONE;
+	struct request_queue *top_q = bio->bi_disk->queue;
+	bool poll_on = test_bit(QUEUE_FLAG_POLL, &top_q->queue_flags);
 
 	BUG_ON(bio->bi_next);
 

@@ -968,6 +976,7 @@ static blk_qc_t __submit_bio_noacct(struct bio *bio)
 	current->bio_list = bio_list_on_stack;
 
 	do {
+		blk_qc_t cookie;
 		struct request_queue *q = bio->bi_disk->queue;
 		struct bio_list lower, same;
 

@@ -980,7 +989,23 @@ static blk_qc_t __submit_bio_noacct(struct bio *bio)
 		bio_list_on_stack[1] = bio_list_on_stack[0];
 		bio_list_init(&bio_list_on_stack[0]);
 
-		ret = __submit_bio(bio);
+		cookie = __submit_bio(bio);
+
+		if (poll_on && blk_qc_t_valid(cookie)) {
+			unsigned int queue_num = blk_qc_t_to_queue_num(cookie);
+			struct blk_mq_hw_ctx *hctx = q->queue_hw_ctx[queue_num];
+
+			cookie = (blk_qc_t)hctx;
+
+			if (!blk_qc_t_valid(ret)) {
+				/* set initial value */
+				ret = cookie;
+			} else if (ret != cookie) {
+				/* bio gets split and enqueued to multi hctxs */
+				ret = BLK_QC_T_BIO_POLL_ALL;
+				poll_on = false;
+			}
+		}
 
 		/*
 		 * Sort new bios into those for a lower level and those for the

@@ -1003,6 +1028,7 @@ static blk_qc_t __submit_bio_noacct(struct bio *bio)
 	} while ((bio = bio_list_pop(&bio_list_on_stack[0])));
 
 	current->bio_list = NULL;
+
 	return ret;
 }
 

@@ -1142,7 +1168,15 @@ static int blk_bio_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
 	do {
 		int ret;
 
-		ret = disk->fops->poll(q, cookie);
+		if (unlikely(blk_qc_t_is_poll_multi(cookie)))
+			ret = disk->fops->poll(q, cookie);
+		else {
+			struct blk_mq_hw_ctx *hctx = (struct blk_mq_hw_ctx *)cookie;
+			struct request_queue *target_q = hctx->queue;
+
+			ret = blk_mq_poll_hctx(target_q, hctx);
+		}
+
 		if (ret > 0) {
 			__set_current_state(TASK_RUNNING);
 			return ret;

diff --git a/include/linux/blk_types.h b/include/linux/blk_types.h
index 2e05244fc16d..6f27446ebada 100644
--- a/include/linux/blk_types.h
+++ b/include/linux/blk_types.h

@@ -557,6 +557,14 @@ static inline bool blk_qc_t_is_internal(blk_qc_t cookie)
 	return (cookie & BLK_QC_T_INTERNAL) != 0;
 }
 
+/* Macros for blk_qc_t used for bio-based polling */
+#define BLK_QC_T_BIO_POLL_ALL	1U
+
+static inline bool blk_qc_t_is_poll_multi(blk_qc_t cookie)
+{
+	return cookie & BLK_QC_T_BIO_POLL_ALL;
+}
+
 struct blk_rq_stat {
 	u64 mean;
 	u64 min;

diff --git a/include/linux/types.h b/include/linux/types.h
index da5ca7e1bea9..f6301014a459 100644
--- a/include/linux/types.h
+++ b/include/linux/types.h

@@ -126,7 +126,7 @@ typedef u64 sector_t;
 typedef u64 blkcnt_t;
 
 /* cookie used for IO polling */
-typedef unsigned int blk_qc_t;
+typedef uintptr_t blk_qc_t;
 
 /*
  * The type of an index into the pagecache.

-- 
2.27.0

IO polling is enabled when all underlying target devices are capable
of IO polling. The sanity check supports the stacked device model, in
which one dm device may be build upon another dm device. In this case,
the mapped device will check if the underlying dm target device
supports IO polling.

Signed-off-by: Jeffle Xu <jefflexu@linux.alibaba.com>
---
 drivers/md/dm-table.c         | 26 ++++++++++++
 drivers/md/dm.c               | 78 +++++++++++++++++++++++++++++++++++
 include/linux/device-mapper.h |  1 +
 3 files changed, 105 insertions(+)

diff --git a/drivers/md/dm-table.c b/drivers/md/dm-table.c
index 2bc256d61550..0210ec7a3dee 100644
--- a/drivers/md/dm-table.c
+++ b/drivers/md/dm-table.c

@@ -1768,6 +1768,19 @@ static int device_requires_stable_pages(struct dm_target *ti,
 	return blk_queue_stable_writes(q);
 }
 
+static int device_not_poll_capable(struct dm_target *ti, struct dm_dev *dev,
+				   sector_t start, sector_t len, void *data)
+{
+	struct request_queue *q = bdev_get_queue(dev->bdev);
+
+	return !test_bit(QUEUE_FLAG_POLL, &q->queue_flags);
+}
+
+int dm_table_supports_poll(struct dm_table *t)
+{
+	return !dm_table_any_dev_attr(t, device_not_poll_capable, NULL);
+}
+
 void dm_table_set_restrictions(struct dm_table *t, struct request_queue *q,
 			       struct queue_limits *limits)
 {

@@ -1864,6 +1877,19 @@ void dm_table_set_restrictions(struct dm_table *t, struct request_queue *q,
 #endif
 
 	blk_queue_update_readahead(q);
+
+	/*
+	 * Check for request-based device is remained to
+	 * dm_mq_init_request_queue()->blk_mq_init_allocated_queue().
+	 * For bio-based device, only set QUEUE_FLAG_POLL when all underlying
+	 * devices supporting polling.
+	 */
+	if (__table_type_bio_based(t->type)) {
+		if (dm_table_supports_poll(t))
+			blk_queue_flag_set(QUEUE_FLAG_POLL, q);
+		else
+			blk_queue_flag_clear(QUEUE_FLAG_POLL, q);
+	}
 }
 
 unsigned int dm_table_get_num_targets(struct dm_table *t)

diff --git a/drivers/md/dm.c b/drivers/md/dm.c
index c2945c90745e..5f358efda50f 100644
--- a/drivers/md/dm.c
+++ b/drivers/md/dm.c

@@ -1657,6 +1657,82 @@ static blk_qc_t dm_submit_bio(struct bio *bio)
 	return BLK_QC_T_NONE;
 }
 
+struct dm_poll_data {
+	blk_qc_t cookie;
+	int count;
+};
+
+static void dm_poll_one_md(struct mapped_device *md, struct dm_poll_data *pdata);
+
+static int dm_poll_one_dev(struct dm_target *ti, struct dm_dev *dev,
+				sector_t start, sector_t len, void *data)
+{
+	struct dm_poll_data *pdata = data;
+	struct request_queue *q = bdev_get_queue(dev->bdev);
+	struct blk_mq_hw_ctx *hctx;
+	int i, cpu;
+
+	if (queue_is_mq(q)) {
+		if (!percpu_ref_tryget(&q->q_usage_counter))
+			return 0;
+
+		if (blk_qc_t_is_poll_cpu(pdata->cookie)) {
+			cpu = blk_qc_t_to_cpu(pdata->cookie);
+			hctx = blk_mq_get_hctx(q, cpu, REQ_HIPRI);
+			pdata->count += blk_mq_poll_hctx(q, hctx);
+		} else {
+			queue_for_each_poll_hw_ctx(q, hctx, i)
+				pdata->count += blk_mq_poll_hctx(q, hctx);
+		}
+
+		percpu_ref_put(&q->q_usage_counter);
+	} else
+		dm_poll_one_md(dev->bdev->bd_disk->private_data, pdata);
+
+	return 0;
+}
+
+static void dm_poll_one_md(struct mapped_device *md, struct dm_poll_data *pdata)
+{
+	int i, srcu_idx;
+	struct dm_table *t;
+	struct dm_target *ti;
+
+	t = dm_get_live_table(md, &srcu_idx);
+
+	for (i = 0; i < dm_table_get_num_targets(t); i++) {
+		ti = dm_table_get_target(t, i);
+		ti->type->iterate_devices(ti, dm_poll_one_dev, pdata);
+	}
+
+	dm_put_live_table(md, srcu_idx);
+}
+
+static int dm_bio_poll(struct request_queue *q, blk_qc_t cookie)
+{
+	struct gendisk *disk = queue_to_disk(q);
+	struct mapped_device *md = disk->private_data;
+	struct dm_poll_data pdata = {
+		.cookie = cookie,
+	};
+
+	dm_poll_one_md(md, &pdata);
+	return pdata.count;
+}
+
+static bool dm_bio_poll_capable(struct gendisk *disk)
+{
+	int ret, srcu_idx;
+	struct mapped_device *md = disk->private_data;
+	struct dm_table *t;
+
+	t = dm_get_live_table(md, &srcu_idx);
+	ret = dm_table_supports_poll(t);
+	dm_put_live_table(md, srcu_idx);
+
+	return ret;
+}
+
 /*-----------------------------------------------------------------
  * An IDR is used to keep track of allocated minor numbers.
  *---------------------------------------------------------------*/

@@ -3049,6 +3125,8 @@ static const struct pr_ops dm_pr_ops = {
 };
 
 static const struct block_device_operations dm_blk_dops = {
+	.poll = dm_bio_poll,
+	.poll_capable = dm_bio_poll_capable,
 	.submit_bio = dm_submit_bio,
 	.open = dm_blk_open,
 	.release = dm_blk_close,

diff --git a/include/linux/device-mapper.h b/include/linux/device-mapper.h
index 61a66fb8ebb3..6a9de3fd0087 100644
--- a/include/linux/device-mapper.h
+++ b/include/linux/device-mapper.h

@@ -515,6 +515,7 @@ unsigned int dm_table_get_num_targets(struct dm_table *t);
 fmode_t dm_table_get_mode(struct dm_table *t);
 struct mapped_device *dm_table_get_md(struct dm_table *t);
 const char *dm_table_device_name(struct dm_table *t);
+int dm_table_supports_poll(struct dm_table *t);
 
 /*
  * Trigger an event.

-- 
2.27.0

Offer one sub-fastpath for bio-based polling when bio submitted to dm
device gets split and enqueued into multiple hw queues, while the IO
submission process has not been migrated to another CPU.

In this case, the IO submission routine will return the CPU number on
which the IO submission happened as the returned cookie, while the
polling routine will only iterate and poll on hw queues that this CPU
number maps, instead of iterating *all* hw queues.

This optimization can dramatically reduce cache ping-pong and thus
improve the polling performance, when multiple hw queues in polling mode
per device could be reserved when there are multiple polling processes.

It will fall back to iterating all hw queues in polling mode, once the
process has ever been migrated to another CPU during the IO submission
phase.

Signed-off-by: Jeffle Xu <jefflexu@linux.alibaba.com>
---
 block/blk-core.c          | 24 ++++++++++++++--
 include/linux/blk_types.h | 60 +++++++++++++++++++++++++++++++++++++--
 2 files changed, 79 insertions(+), 5 deletions(-)

diff --git a/block/blk-core.c b/block/blk-core.c
index 82349e3c2a24..e13219908ef8 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c

@@ -959,7 +959,10 @@ static blk_qc_t __submit_bio(struct bio *bio)
  *
  * Return:
  *   - BLK_QC_T_NONE, no need for IO polling.
- *   - BLK_QC_T_BIO_POLL_ALL, @bio gets split and enqueued into multi hw queues.
+ *   - When @bio gets split and enqueued into multi hw queues, return the
+ *     corresponding CPU number with BLK_QC_T_BIO_POLL_CPU flag set when the
+ *     current process has not been migrated to another CPU,
+ *     BLK_QC_T_BIO_POLL_ALL otherwise.
  *   - Otherwise, @bio is not split, returning the pointer to the corresponding
  *     hw queue that the bio enqueued into as the returned cookie.
  */

@@ -968,13 +971,17 @@ static blk_qc_t __submit_bio_noacct(struct bio *bio)
 	struct bio_list bio_list_on_stack[2];
 	blk_qc_t ret = BLK_QC_T_NONE;
 	struct request_queue *top_q = bio->bi_disk->queue;
-	bool poll_on = test_bit(QUEUE_FLAG_POLL, &top_q->queue_flags);
+	bool orig_poll_on, poll_on;
+	u64 old_nr_migrations;
 
 	BUG_ON(bio->bi_next);
 
 	bio_list_init(&bio_list_on_stack[0]);
 	current->bio_list = bio_list_on_stack;
 
+	orig_poll_on = poll_on = test_bit(QUEUE_FLAG_POLL, &top_q->queue_flags);
+	old_nr_migrations = READ_ONCE(current->se.nr_migrations);
+
 	do {
 		blk_qc_t cookie;
 		struct request_queue *q = bio->bi_disk->queue;

@@ -1002,7 +1009,7 @@ static blk_qc_t __submit_bio_noacct(struct bio *bio)
 				ret = cookie;
 			} else if (ret != cookie) {
 				/* bio gets split and enqueued to multi hctxs */
-				ret = BLK_QC_T_BIO_POLL_ALL;
+				ret = blk_qc_t_get();
 				poll_on = false;
 			}
 		}

@@ -1029,6 +1036,17 @@ static blk_qc_t __submit_bio_noacct(struct bio *bio)
 
 	current->bio_list = NULL;
 
+	/*
+	 * For cases when bio gets split and enqueued into multi hctxs, return
+	 * the corresponding CPU number when current process has not been
+	 * migrated to another CPU. Return BLK_QC_T_BIO_POLL_ALL otherwise,
+	 * falling back to iterating and polling on all hw queues, since split
+	 * bios are submitted to different CPUs in this case.
+	 */
+	if (orig_poll_on != poll_on &&
+	    old_nr_migrations != READ_ONCE(current->se.nr_migrations))
+		ret = BLK_QC_T_BIO_POLL_ALL;
+
 	return ret;
 }
 

diff --git a/include/linux/blk_types.h b/include/linux/blk_types.h
index 6f27446ebada..f8d34f02378d 100644
--- a/include/linux/blk_types.h
+++ b/include/linux/blk_types.h

@@ -557,12 +557,68 @@ static inline bool blk_qc_t_is_internal(blk_qc_t cookie)
 	return (cookie & BLK_QC_T_INTERNAL) != 0;
 }
 
-/* Macros for blk_qc_t used for bio-based polling */
+/*
+ * Macros for blk_qc_t used for bio-based polling.
+ *
+ * 1. When @bio is not split, the returned cookie is actually the pointer to the
+ *    corresponding hw queue that the bio is enqueued into, thus the low two
+ *    bits could be reused as other flags.
+ *    +-----------------------------------------+
+ *    |        struct blk_mq_hw_ctx*        |0|0|
+ *    +-----------------------------------------+
+ *
+ * 2. When @bio gets split and enqueued into multi hw queues, and current
+ *    process has not been migrated to another CPU, the returned cookie actually
+ *    stores the corresponding CPU number on which the IO submission happened,
+ *    Also with BLK_QC_T_BIO_POLL_CPU flag set.
+ *    In this case, bio-based polling could only iterate and poll on hw queues
+ *    that this CPU number maps.
+ *    +-----------------------------------------+
+ *    |         CPU         |               |1| |
+ *    +-----------------------------------------+
+ *                                           ^
+ *                                           BLK_QC_T_BIO_POLL_CPU
+ *
+ * 3. When @bio gets split and enqueued into multi hw queues, and current
+ *    process has ever been migrated to another CPU, the returned cookie is just
+ *    BLK_QC_T_BIO_POLL_ALL flag.
+ *    In this case, bio-based polling should iterate and poll on all hw queues
+ *    in polling mode.
+ *    +-----------------------------------------+
+ *    |                                     | |1|
+ *    +-----------------------------------------+
+ *                                             ^
+ *                                             BLK_QC_T_BIO_POLL_ALL
+ *
+ * 4. Otherwise, return BLK_QC_T_NONE as the cookie.
+ *    No need for IO polling in this case.
+ *    +-----------------------------------------+
+ *    |               BLK_QC_T_NONE             |
+ *    +-----------------------------------------+
+ */
+
+/* The low two bits of cookie for bio-based polling are reused for flags. */
 #define BLK_QC_T_BIO_POLL_ALL	1U
+#define BLK_QC_T_BIO_POLL_CPU	2U
 
 static inline bool blk_qc_t_is_poll_multi(blk_qc_t cookie)
 {
-	return cookie & BLK_QC_T_BIO_POLL_ALL;
+	return cookie & (BLK_QC_T_BIO_POLL_ALL | BLK_QC_T_BIO_POLL_CPU);
+}
+
+static inline bool blk_qc_t_is_poll_cpu(blk_qc_t cookie)
+{
+	return cookie & BLK_QC_T_BIO_POLL_CPU;
+}
+
+static inline blk_qc_t blk_qc_t_get(void)
+{
+	return (raw_smp_processor_id() << BLK_QC_T_SHIFT) | BLK_QC_T_BIO_POLL_CPU;
+}
+
+static inline int blk_qc_t_to_cpu(blk_qc_t cookie)
+{
+	return cookie >> BLK_QC_T_SHIFT;
 }
 
 struct blk_rq_stat {

-- 
2.27.0

Add one helper function getting hw queue mapping to specific CPU, and of
specific type.

Signed-off-by: Jeffle Xu <jefflexu@linux.alibaba.com>
---
 block/blk-mq.c         | 10 ++++++++++
 include/linux/blk-mq.h |  2 ++
 2 files changed, 12 insertions(+)

diff --git a/block/blk-mq.c b/block/blk-mq.c
index b4de2b37b826..734aa1338b69 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c

@@ -3876,6 +3876,16 @@ unsigned int blk_mq_rq_cpu(struct request *rq)
 }
 EXPORT_SYMBOL(blk_mq_rq_cpu);
 
+
+struct blk_mq_hw_ctx *blk_mq_get_hctx(struct request_queue *q, int cpu,
+				      unsigned int flags)
+{
+	struct blk_mq_ctx *ctx = __blk_mq_get_ctx(q, cpu);
+
+	return blk_mq_map_queue(q, flags, ctx);
+}
+EXPORT_SYMBOL(blk_mq_get_hctx);
+
 static int __init blk_mq_init(void)
 {
 	int i;

diff --git a/include/linux/blk-mq.h b/include/linux/blk-mq.h
index 0c511555ec16..9b6a5c1fcb2d 100644
--- a/include/linux/blk-mq.h
+++ b/include/linux/blk-mq.h

@@ -630,5 +630,7 @@ blk_qc_t blk_mq_submit_bio(struct bio *bio);
 int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, bool spin);
 void blk_mq_hctx_set_fq_lock_class(struct blk_mq_hw_ctx *hctx,
 		struct lock_class_key *key);
+struct blk_mq_hw_ctx *blk_mq_get_hctx(struct request_queue *q, int cpu,
+		unsigned int flags);
 
 #endif

-- 
2.27.0

Add one helper function for iterating all hardware queues in polling
mode.

Signed-off-by: Jeffle Xu <jefflexu@linux.alibaba.com>
---
 include/linux/blk-mq.h | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/include/linux/blk-mq.h b/include/linux/blk-mq.h
index cf1910c6d5ae..0c511555ec16 100644
--- a/include/linux/blk-mq.h
+++ b/include/linux/blk-mq.h

@@ -574,6 +574,13 @@ static inline void *blk_mq_rq_to_pdu(struct request *rq)
 	for ((i) = 0; (i) < (q)->nr_hw_queues &&			\
 	     ({ hctx = (q)->queue_hw_ctx[i]; 1; }); (i)++)
 
+#define queue_for_each_poll_hw_ctx(q, hctx, i)				\
+	for ((i) = 0; ((q)->tag_set->nr_maps > HCTX_TYPE_POLL) &&	\
+	     (i) < (q)->tag_set->map[HCTX_TYPE_POLL].nr_queues &&	\
+	     ({ int __idx = (q)->tag_set->map[HCTX_TYPE_POLL].queue_offset + (i); \
+	     hctx = (q)->queue_hw_ctx[__idx]; 1; }); \
+	     (i)++)
+
 #define hctx_for_each_ctx(hctx, ctx, i)					\
 	for ((i) = 0; (i) < (hctx)->nr_ctx &&				\
 	     ({ ctx = (hctx)->ctxs[(i)]; 1; }); (i)++)

-- 
2.27.0

There's no sense waiting for the hw queue when it currently has been
locked by another polling instance. The polling instance currently
occupying the hw queue will help reap the completion events.

It shall be safe to surrender the hw queue, as long as we could reapply
for polling later. For Synchronous polling, blk_poll() will reapply for
polling, since @spin is always True in this case. While For asynchronous
polling, i.e. io_uring itself will reapply for polling when the previous
polling returns 0.

Besides, it shall do no harm to the polling performance of mq devices.

Signed-off-by: Jeffle Xu <jefflexu@linux.alibaba.com>
---
 drivers/nvme/host/pci.c | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/drivers/nvme/host/pci.c b/drivers/nvme/host/pci.c
index 81e6389b2042..db164fcd04e2 100644
--- a/drivers/nvme/host/pci.c
+++ b/drivers/nvme/host/pci.c

@@ -1106,7 +1106,9 @@ static int nvme_poll(struct blk_mq_hw_ctx *hctx)
 	if (!nvme_cqe_pending(nvmeq))
 		return 0;
 
-	spin_lock(&nvmeq->cq_poll_lock);
+	if (!spin_trylock(&nvmeq->cq_poll_lock))
+		return 0;
+
 	found = nvme_process_cq(nvmeq);
 	spin_unlock(&nvmeq->cq_poll_lock);
 

-- 
2.27.0

->poll_fn was introduced in commit ea435e1b9392 ("block: add a poll_fn
callback to struct request_queue") to support bio-based queues such as
nvme multipath, but was later removed in commit 529262d56dbe ("block:
remove ->poll_fn").

Given commit c62b37d96b6e ("block: move ->make_request_fn to struct
block_device_operations") restore the possibility of bio-based IO
polling support by adding an ->poll method to gendisk->fops.

Make blk_mq_poll() specific to blk-mq, while blk_bio_poll() is
originally a copy from blk_mq_poll(), and is specific to bio-based
polling. Currently hybrid polling is not supported by bio-based polling.

Signed-off-by: Jeffle Xu <jefflexu@linux.alibaba.com>
---
 block/blk-core.c       | 58 ++++++++++++++++++++++++++++++++++++++++++
 block/blk-mq.c         | 22 +---------------
 include/linux/blk-mq.h |  1 +
 include/linux/blkdev.h |  1 +
 4 files changed, 61 insertions(+), 21 deletions(-)

diff --git a/block/blk-core.c b/block/blk-core.c
index 7663a9b94b80..37aa513da5f2 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c

@@ -1132,6 +1132,64 @@ blk_qc_t submit_bio(struct bio *bio)
 }
 EXPORT_SYMBOL(submit_bio);
 
+
+static int blk_bio_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
+{
+	long state;
+	struct gendisk *disk = queue_to_disk(q);
+
+	state = current->state;
+	do {
+		int ret;
+
+		ret = disk->fops->poll(q, cookie);
+		if (ret > 0) {
+			__set_current_state(TASK_RUNNING);
+			return ret;
+		}
+
+		if (signal_pending_state(state, current))
+			__set_current_state(TASK_RUNNING);
+
+		if (current->state == TASK_RUNNING)
+			return 1;
+		if (ret < 0 || !spin)
+			break;
+		cpu_relax();
+	} while (!need_resched());
+
+	__set_current_state(TASK_RUNNING);
+	return 0;
+}
+
+/**
+ * blk_poll - poll for IO completions
+ * @q:  the queue
+ * @cookie: cookie passed back at IO submission time
+ * @spin: whether to spin for completions
+ *
+ * Description:
+ *    Poll for completions on the passed in queue. Returns number of
+ *    completed entries found. If @spin is true, then blk_poll will continue
+ *    looping until at least one completion is found, unless the task is
+ *    otherwise marked running (or we need to reschedule).
+ */
+int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
+{
+	if (!blk_qc_t_valid(cookie) ||
+	    !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
+		return 0;
+
+	if (current->plug)
+		blk_flush_plug_list(current->plug, false);
+
+	if (queue_is_mq(q))
+		return blk_mq_poll(q, cookie, spin);
+	else
+		return blk_bio_poll(q, cookie, spin);
+}
+EXPORT_SYMBOL_GPL(blk_poll);
+
 /**
  * blk_cloned_rq_check_limits - Helper function to check a cloned request
  *                              for the new queue limits

diff --git a/block/blk-mq.c b/block/blk-mq.c
index f285a9123a8b..c18bc23834a0 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c

@@ -3826,30 +3826,11 @@ static bool blk_mq_poll_hybrid(struct request_queue *q,
 	return blk_mq_poll_hybrid_sleep(q, rq);
 }
 
-/**
- * blk_poll - poll for IO completions
- * @q:  the queue
- * @cookie: cookie passed back at IO submission time
- * @spin: whether to spin for completions
- *
- * Description:
- *    Poll for completions on the passed in queue. Returns number of
- *    completed entries found. If @spin is true, then blk_poll will continue
- *    looping until at least one completion is found, unless the task is
- *    otherwise marked running (or we need to reschedule).
- */
-int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
+int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
 {
 	struct blk_mq_hw_ctx *hctx;
 	long state;
 
-	if (!blk_qc_t_valid(cookie) ||
-	    !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
-		return 0;
-
-	if (current->plug)
-		blk_flush_plug_list(current->plug, false);
-
 	hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
 
 	/*

@@ -3891,7 +3872,6 @@ int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
 	__set_current_state(TASK_RUNNING);
 	return 0;
 }
-EXPORT_SYMBOL_GPL(blk_poll);
 
 unsigned int blk_mq_rq_cpu(struct request *rq)
 {

diff --git a/include/linux/blk-mq.h b/include/linux/blk-mq.h
index d705b174d346..adc789d5888e 100644
--- a/include/linux/blk-mq.h
+++ b/include/linux/blk-mq.h

@@ -607,6 +607,7 @@ static inline void blk_rq_bio_prep(struct request *rq, struct bio *bio,
 }
 
 blk_qc_t blk_mq_submit_bio(struct bio *bio);
+int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, bool spin);
 void blk_mq_hctx_set_fq_lock_class(struct blk_mq_hw_ctx *hctx,
 		struct lock_class_key *key);
 

diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index 2a802e011a95..ebe16f55cba4 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h

@@ -1853,6 +1853,7 @@ static inline void blk_ksm_unregister(struct request_queue *q) { }
 
 struct block_device_operations {
 	blk_qc_t (*submit_bio) (struct bio *bio);
+	int (*poll)(struct request_queue *q, blk_qc_t cookie);
 	int (*open) (struct block_device *, fmode_t);
 	void (*release) (struct gendisk *, fmode_t);
 	int (*rw_page)(struct block_device *, sector_t, struct page *, unsigned int);

-- 
2.27.0

Extract the logic of polling one hw queue and related statistics
handling out as the helper function.

Signed-off-by: Jeffle Xu <jefflexu@linux.alibaba.com>
---
 block/blk-mq.c         |  5 +----
 include/linux/blk-mq.h | 13 +++++++++++++
 2 files changed, 14 insertions(+), 4 deletions(-)

diff --git a/block/blk-mq.c b/block/blk-mq.c
index c18bc23834a0..b4de2b37b826 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c

@@ -3850,11 +3850,8 @@ int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
 	do {
 		int ret;
 
-		hctx->poll_invoked++;
-
-		ret = q->mq_ops->poll(hctx);
+		ret = blk_mq_poll_hctx(q, hctx);
 		if (ret > 0) {
-			hctx->poll_success++;
 			__set_current_state(TASK_RUNNING);
 			return ret;
 		}

diff --git a/include/linux/blk-mq.h b/include/linux/blk-mq.h
index adc789d5888e..cf1910c6d5ae 100644
--- a/include/linux/blk-mq.h
+++ b/include/linux/blk-mq.h

@@ -606,6 +606,19 @@ static inline void blk_rq_bio_prep(struct request *rq, struct bio *bio,
 		rq->rq_disk = bio->bi_disk;
 }
 
+static inline int blk_mq_poll_hctx(struct request_queue *q,
+				   struct blk_mq_hw_ctx *hctx)
+{
+	int ret;
+
+	hctx->poll_invoked++;
+	ret = q->mq_ops->poll(hctx);
+	if (ret > 0)
+		hctx->poll_success++;
+
+	return ret;
+}
+
 blk_qc_t blk_mq_submit_bio(struct bio *bio);
 int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, bool spin);
 void blk_mq_hctx_set_fq_lock_class(struct blk_mq_hw_ctx *hctx,

-- 
2.27.0

On 2/20/21 7:06 PM, Jeffle Xu wrote:

[Changes since v3]
- newly add patch 7 and patch 11, as a new optimization improving
performance of multiple polling processes. Now performance of multiple
polling processes can be as scalable as single polling process (~30%).
Refer to the following [Performance] chapter for more details.

Hi Mike, would please evaluate this new version patch set? I think this
mechanism is near maturity, since multi-thread performance is as
scalable as single-thread (~30%) now.


Thanks
Jeffle

[Intention]
Bio-based polling (e.g., for dm/md devices) is one indispensable part of
high performance IO stack. As far as I know, dm (e.g., dm-stripe) is
widely used in database, splicing several NVMe disks as one whole disk,
in hope of achieving better performance. With this patch set, io_uring
could be used upon dm devices.


[Optimizations]
Basically, there are three paths for IO polling.

1. fastpath (patch 9/10)
The polling routine will go into this path when bio submitted to dm
device is not split.

In this case, there will be only one bio submitted to only one polling
hw queue of one underlying mq device, and thus we don't need to track
all split bios or iterate through all polling hw queues. The pointer to
the polling hw queue the bio submitted to is returned here as the
returned cookie. In this case, the polling routine will call
mq_ops->poll() directly with the hw queue converted from the input
cookie.


- One process reading dm-linear (mapping to three underlying NVMe devices,
with one polling hw queue per NVMe device).

(ioengine=io_uring, iodepth=128, numjobs=1, rw=randread, sqthread_poll=0
direct=1, bs=4k)

	    	 | IOPS (IRQ mode) | IOPS (iopoll=1 mode) | diff
---------------- | --------------- | -------------------- | ----
with patchset    |	      212k |		     284k | ~32%


- Three processes reading dm-linear (mapping to three underlying NVMe
devices, with one polling hw queue per NVMe device).

(ioengine=io_uring, iodepth=128, numjobs=3, rw=randread, sqthread_poll=0
direct=1, bs=4k)

	    	 | IOPS (IRQ mode) | IOPS (iopoll=1 mode) | diff
---------------- | --------------- | -------------------- | ----
with patchset    |	      615k |		     735k | ~16%


- Three processes reading dm-linear (mapping to three underlying NVMe
devices, with three polling hw queues per NVMe device), with every
process pinned to one CPU and mapped to one exclusive hw queue.

(ioengine=io_uring, iodepth=128, numjobs=3, rw=randread, sqthread_poll=0
direct=1, bs=4k)

	    	 | IOPS (IRQ mode) | IOPS (iopoll=1 mode) | diff
---------------- | --------------- | -------------------- | ----
with patchset    |	      631k |		     833k | ~32%



2. sub-fastpath (patch 7/11)

The polling routine will go into this path when bio submitted to dm
device gets split and enqueued into multiple hw queues, while the IO
submission process has not been migrated to another CPU.

In this case, the IO submission routine will return the CPU number on
which the IO submission happened as the returned cookie, while the
polling routine will only iterate and poll on hw queues that this CPU
number maps, instead of iterating *all* hw queues.

This optimization can dramatically reduce cache ping-pong and thus
improve the polling performance, when multiple hw queues in polling mode
per device could be reserved when there are multiple polling processes.

- Three processes reading dm-stripe (mapping to three underlying NVMe
devices, with three polling hw queues per NVMe device), with every
process pinned to one CPU and mapped to one exclusive hw queue.

(ioengine=io_uring, iodepth=128, numjobs=3, rw=randread, sqthread_poll=0
direct=1, bs=12k(4k for every NVMe device))

	    	 | IOPS (IRQ mode) | IOPS (iopoll=1 mode) | diff
---------------- | --------------- | -------------------- | ----
with patchset    |	      307k |		     412k | ~34%


3. default path

It will fall back to iterating all hw queues in polling mode, once bio
submitted to dm device gets split and enqueued into multiple hw queues,
and the IO process has ever been migrated to another CPU during the IO
submission phase.


[Remained Issue]
It has been mentioned in patch 4 that, users could change the state of
the underlying devices through '/sys/block/<dev>/io_poll', bypassing
the dm device above. Thus it can cause a situation where QUEUE_FLAG_POLL
is still set for the request_queue of dm device, while one of the
underlying mq device may has cleared this flag.

In this case, it will pass the 'test_bit(QUEUE_FLAG_POLL, &q->queue_flags)'
check in blk_poll(), while the input cookie may actually points to a hw
queue in IRQ mode since patch 11. Thus for this hw queue (in IRQ mode),
the bio-based polling routine will handle this hw queue acquiring
'spin_lock(&nvmeq->cq_poll_lock)' (refer
drivers/nvme/host/pci.c:nvme_poll), which is not adequate since this hw
queue may also be accessed in IRQ context. In other words,
spin_lock_irq() should be used here.

I have not come up one simple way to fix it. I don't want to do sanity
check (e.g., the type of the hw queue is HCTX_TYPE_POLL or not) in the
IO path (submit_bio()/blk_poll()), i.e., fast path.

We'd better fix it in the control path, i.e., dm could be aware of the
change when attribute (e.g., support io_poll or not) of one of the
underlying devices changed at runtime.




[Changes since v2]

Patchset v2 caches all hw queues (in polling mode) of underlying mq
devices in dm layer. The polling routine actually iterates through all
these cached hw queues.

However, mq may change the queue mapping at runtime (e.g., NVMe RESET
command), thus the cached hw queues in dm layer may be out-of-date. Thus
patchset v3 falls back to the implementation of the very first RFC
version, in which the mq layer needs to export one interface iterating
all polling hw queues (patch 5), and the bio-based polling routine just
calls this interface to iterate all polling hw queues.

Besides, several new optimization is proposed.


- patch 1,2,7
same as v2, untouched

- patch 3
Considering advice from Christoph Hellwig, while refactoring blk_poll(),
split mq and bio-based polling routine from the very beginning. Now
blk_poll() is just a simple entry. blk_bio_poll() is simply copied from
blk_mq_poll(), while the loop structure is some sort of duplication
though.

- patch 4
This patch is newly added to support turning on/off polling through
'/sys/block/<dev>/queue/io_poll' dynamiclly for bio-based devices.
Patchset v2 implemented this functionality by added one new queue flag,
which is not preferred since the queue flag resource is quite short of
nowadays.

- patch 5
This patch is newly added, preparing for the following bio-based
polling. The following bio-based polling will call this helper function,
accounting on the corresponding hw queue.

- patch 6
It's from the very first RFC version, preparing for the following
bio-based polling.

- patch 8
One fixing patch needed by the following bio-based polling. It's
actually a v2 of [1]. I had sent the v2 singly in-reply-to [1], though
it has not been visible on the mailing list maybe due to the delay.

- patch 9
It's from the very first RFC version.

- patch 10
This patch is newly added. Patchset v2 had ever proposed one
optimization that, skipping the **busy** hw queues during the iteration
phase. Back upon that time, one flag of 'atomic_t' is specifically
maintained in dm layer, representing if the corresponding hw queue is
busy or not. The idea is inherited, while the implementation changes.
Now @nvmeq->cq_poll_lock is used directly here, no need for extra flag
anymore.

This optimization can significantly reduce the competition for one hw
queue between multiple polling instances. Following statistics is the
test result when 3 threads concurrently randread (bs=4k, direct=1) one
dm-linear device, which is built upon 3 nvme devices, with one polling
hw queue per nvme device.

	    | IOPS (IRQ mode) | IOPS (iopoll=1 mode) | diff
----------- | --------------- | -------------------- | ----
without opt | 		 318k |		 	256k | ~-20%
with opt    |		 314k |		 	354k | ~13%
							

- patch 11
This is another newly added optimizatin for bio-based polling.

One intuitive insight is that, when the original bio submitted to dm
device doesn't get split, then the bio gets enqueued into only one hw
queue of one of the underlying mq devices. In this case, we no longer
need to track all split bios, and one cookie (for the only split bio)
is enough. It is implemented by returning the pointer to the
corresponding hw queue in this case.

It should be safe by directly returning the pointer to the hw queue,
since 'struct blk_mq_hw_ctx' won't be freed during the whole lifetime of
'struct request_queue'. Even when the number of hw queues may decrease
when NVMe RESET happens, the 'struct request_queue' structure of decreased
hw queues won't be freed, instead it's buffered into
&q->unused_hctx_list list.

Though this optimization seems quite intuitive, the performance test
shows that it does no benefit nor harm to the performance, while 3
threads concurrently randreading (bs=4k, direct=1) one dm-linear
device, which is built upon 3 nvme devices, with one polling hw queue
per nvme device.

I'm not sure why it doesn't work, maybe because the number of devices,
or the depth of the devcice stack is to low in my test case?


changes since v1:
- patch 1,2,4 is the same as v1 and have already been reviewed
- patch 3 is refactored a bit on the basis of suggestions from
Mike Snitzer.
- patch 5 is newly added and introduces one new queue flag
representing if the queue is capable of IO polling. This mainly
simplifies the logic in queue_poll_store().
- patch 6 implements the core mechanism supporting IO polling.
The sanity check checking if the dm device supports IO polling is
also folded into this patch, and the queue flag will be cleared if
it doesn't support, in case of table reloading.




Jeffle Xu (12):
  block: move definition of blk_qc_t to types.h
  block: add queue_to_disk() to get gendisk from request_queue
  block: add poll method to support bio-based IO polling
  block: add poll_capable method to support bio-based IO polling
  blk-mq: extract one helper function polling hw queue
  blk-mq: add iterator for polling hw queues
  blk-mq: add one helper function getting hw queue
  dm: always return BLK_QC_T_NONE for bio-based device
  nvme/pci: don't wait for locked polling queue
  block: fastpath for bio-based polling
  block: sub-fastpath for bio-based polling
  dm: support IO polling for bio-based dm device

 block/blk-core.c              | 112 +++++++++++++++++++++++++++++++++-
 block/blk-mq.c                |  37 ++++-------
 block/blk-sysfs.c             |  14 ++++-
 drivers/md/dm-table.c         |  26 ++++++++
 drivers/md/dm.c               | 102 ++++++++++++++++++++++++++-----
 drivers/nvme/host/pci.c       |   4 +-
 include/linux/blk-mq.h        |  23 +++++++
 include/linux/blk_types.h     |  66 +++++++++++++++++++-
 include/linux/blkdev.h        |   4 ++
 include/linux/device-mapper.h |   1 +
 include/linux/fs.h            |   2 +-
 include/linux/types.h         |   3 +
 include/trace/events/kyber.h  |   6 +-
 13 files changed, 350 insertions(+), 50 deletions(-)

-- 
Thanks,
Jeffle

On Mon, Feb 22 2021 at 10:55pm -0500,
JeffleXu [off-list ref] wrote:

On 2/20/21 7:06 PM, Jeffle Xu wrote:

quoted

[Changes since v3]
- newly add patch 7 and patch 11, as a new optimization improving
performance of multiple polling processes. Now performance of multiple
polling processes can be as scalable as single polling process (~30%).
Refer to the following [Performance] chapter for more details.

Hi Mike, would please evaluate this new version patch set? I think this
mechanism is near maturity, since multi-thread performance is as
scalable as single-thread (~30%) now.

OK, can do. But first I think you need to repost with a v5 that
addresses Mikulas' v3 feedback:

https://listman.redhat.com/archives/dm-devel/2021-February/msg00254.html
https://listman.redhat.com/archives/dm-devel/2021-February/msg00255.html

Mike

On 2/24/21 4:54 AM, Mike Snitzer wrote:

On Mon, Feb 22 2021 at 10:55pm -0500,
JeffleXu [off-list ref] wrote:

quoted

On 2/20/21 7:06 PM, Jeffle Xu wrote:

quoted

[Changes since v3]
- newly add patch 7 and patch 11, as a new optimization improving
performance of multiple polling processes. Now performance of multiple
polling processes can be as scalable as single polling process (~30%).
Refer to the following [Performance] chapter for more details.

Hi Mike, would please evaluate this new version patch set? I think this
mechanism is near maturity, since multi-thread performance is as
scalable as single-thread (~30%) now.

OK, can do. But first I think you need to repost with a v5 that
addresses Mikulas' v3 feedback:

https://listman.redhat.com/archives/dm-devel/2021-February/msg00254.html
https://listman.redhat.com/archives/dm-devel/2021-February/msg00255.html

Will do. Besides I will also rebase to 5.12 in the next version.

-- 
Thanks,
Jeffle