Hi Ming,

On Tue, Mar 24, 2020 at 4:29 PM Baolin Wang [off-list ref] wrote:

Hi Ming,

On Mon, Mar 23, 2020 at 5:58 PM Ming Lei [off-list ref] wrote:

quoted

On Mon, Mar 23, 2020 at 05:13:27PM +0800, Baolin Wang wrote:

quoted

On Mon, Mar 23, 2020 at 4:29 PM Ming Lei [off-list ref] wrote:

quoted

On Mon, Mar 23, 2020 at 04:22:38PM +0800, Baolin Wang wrote:

quoted

On Mon, Mar 23, 2020 at 3:27 PM Ming Lei [off-list ref] wrote:

quoted

On Mon, Mar 23, 2020 at 01:36:34PM +0800, Baolin Wang wrote:

quoted

On Mon, Mar 23, 2020 at 11:44 AM Ming Lei [off-list ref] wrote:

quoted

On Fri, Mar 20, 2020 at 06:27:41PM +0800, Baolin Wang wrote:

quoted

Hi Ming,

On Wed, Mar 18, 2020 at 6:26 PM Baolin Wang [off-list ref] wrote:

quoted

Hi Ming,

On Wed, Mar 18, 2020 at 6:01 PM Ming Lei [off-list ref] wrote:

quoted

On Mon, Mar 16, 2020 at 06:01:19PM +0800, Baolin Wang wrote:

quoted

As we know, some SD/MMC host controllers can support packed request,
that means we can package several requests to host controller at one
time to improve performence. So the hardware driver expects the blk-mq
can dispatch a batch of requests at one time, and driver can use bd.last
to indicate if it is the last request in the batch to help to combine
requests as much as possible.

Thus we should add batch requests setting from the block driver to tell
the scheduler how many requests can be dispatched in a batch, as well
as changing the scheduler to dispatch more than one request if setting
the maximum batch requests number.

I feel this batch dispatch style is more complicated, and some other
drivers(virtio blk/scsi) still may get benefit if we can pass real 'last' flag in
.queue_rq().

So what about the following way by extending .commit_rqs() to this usage?
And you can do whatever batch processing in .commit_rqs() which will be
guaranteed to be called if BLK_MQ_F_FORCE_COMMIT_RQS is set by driver.

I'm very appreciated for your good suggestion, which is much simpler than mine.
It seems to solve my problem, and I will try it on my platform to see
if it can work and give you the feadback. Thanks again.

I tried your approach on my platform, but met some problems, see below.

quoted

quoted

diff --git a/block/blk-mq-sched.c b/block/blk-mq-sched.c
index 856356b1619e..cd2bbe56f83f 100644
--- a/block/blk-mq-sched.c
+++ b/block/blk-mq-sched.c

@@ -85,11 +85,12 @@ void blk_mq_sched_restart(struct blk_mq_hw_ctx *hctx)
  * its queue by itself in its completion handler, so we don't need to
  * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
  */
-static void blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
+static bool blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
 {
        struct request_queue *q = hctx->queue;
        struct elevator_queue *e = q->elevator;
        LIST_HEAD(rq_list);
+       bool ret = false;

        do {
                struct request *rq;

@@ -112,7 +113,10 @@ static void blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
                 * in blk_mq_dispatch_rq_list().
                 */
                list_add(&rq->queuelist, &rq_list);
-       } while (blk_mq_dispatch_rq_list(q, &rq_list, true));
+               ret = blk_mq_dispatch_rq_list(q, &rq_list, true);
+       } while (ret);
+
+       return ret;
 }

 static struct blk_mq_ctx *blk_mq_next_ctx(struct blk_mq_hw_ctx *hctx,

@@ -131,11 +135,12 @@ static struct blk_mq_ctx *blk_mq_next_ctx(struct blk_mq_hw_ctx *hctx,
  * its queue by itself in its completion handler, so we don't need to
  * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
  */
-static void blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx *hctx)
+static bool blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx *hctx)
 {
        struct request_queue *q = hctx->queue;
        LIST_HEAD(rq_list);
        struct blk_mq_ctx *ctx = READ_ONCE(hctx->dispatch_from);
+       bool ret = false;

        do {
                struct request *rq;

@@ -161,10 +166,12 @@ static void blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx *hctx)

                /* round robin for fair dispatch */
                ctx = blk_mq_next_ctx(hctx, rq->mq_ctx);
-
-       } while (blk_mq_dispatch_rq_list(q, &rq_list, true));
+               ret = blk_mq_dispatch_rq_list(q, &rq_list, true);
+       } while (ret);

        WRITE_ONCE(hctx->dispatch_from, ctx);
+
+       return ret;
 }

 void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)

@@ -173,6 +180,7 @@ void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
        struct elevator_queue *e = q->elevator;
        const bool has_sched_dispatch = e && e->type->ops.dispatch_request;
        LIST_HEAD(rq_list);
+       bool dispatch_ret;

        /* RCU or SRCU read lock is needed before checking quiesced flag */
        if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)))

@@ -206,20 +214,26 @@ void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
         */
        if (!list_empty(&rq_list)) {
                blk_mq_sched_mark_restart_hctx(hctx);
-               if (blk_mq_dispatch_rq_list(q, &rq_list, false)) {
+               dispatch_ret = blk_mq_dispatch_rq_list(q, &rq_list, false);
+               if (dispatch_ret) {
                        if (has_sched_dispatch)
-                               blk_mq_do_dispatch_sched(hctx);
+                               dispatch_ret = blk_mq_do_dispatch_sched(hctx);

If we dispatched a request successfully by blk_mq_dispatch_rq_list(),
and got dispatch_ret = true now. Then we will try to dispatch more
reuqests from scheduler by blk_mq_do_dispatch_sched(), but if now no
more requests in scheduler, then we will got dispatch_ret = false. In

'dispatch_ret' always holds result of the last blk_mq_do_dispatch_sched().
When any one request has been dispatched successfully, 'dispatch_ret'
is true. New request is always added to list before calling
blk_mq_do_dispatch_sched(), so once blk_mq_do_dispatch_sched() returns
false, it means that .commit_rqs() has been called.

Not really, if no requests int the IO cheduler, we will break the loop
in blk_mq_do_dispatch_sched() and return false without calling
.commit_rqs().

If there isn't any request to dispatch, false is returned. Otherwise,
always return the return value of last 'blk_mq_dispatch_rq_list'.

quoted

So in this case, blk_mq_do_dispatch_sched() will return 'false', which
overlapped the return value of 'true' from blk_mq_dispatch_rq_list(),
and did not call .commit_rqs(). Then the IO processing will be stuck.

See below.

quoted

static void blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
{
        struct request_queue *q = hctx->queue;
        struct elevator_queue *e = q->elevator;
        LIST_HEAD(rq_list);
        bool ret = false;

The above initialization is just done once.

quoted

       do {
              struct request *rq;

              if (e->type->ops.has_work && !e->type->ops.has_work(hctx))
                     break;

              .......

                            ret = blk_mq_dispatch_rq_list(q, list, ...);

list includes one request, so blk_mq_dispatch_rq_list() won't return
false in case of no request in list.

quoted

       } while (ret);

       return ret;

'ret' is always updated by return value of last blk_mq_dispatch_rq_list()
if at least one request is dispatched. So if it becomes false, the loop
breaks, that means .commit_rqs() has been called cause 'list' does
include one request for blk_mq_dispatch_rq_list(). Otherwise, true is
still returned.

Sorry for my confusing description, let me try again to describe the problem.
When I try to mount the block device, I got the IO stuck with your
patch, and I did some debugging. I found we missed calling
commit_rqs() for one case:

void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)

@@ -173,6 +180,7 @@ void blk_mq_sched_dispatch_requests(struct

blk_mq_hw_ctx *hctx)
        struct elevator_queue *e = q->elevator;
        const bool has_sched_dispatch = e && e->type->ops.dispatch_request;
        LIST_HEAD(rq_list);
+       bool dispatch_ret;

        /* RCU or SRCU read lock is needed before checking quiesced flag */
        if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)))

@@ -206,20 +214,26 @@ void blk_mq_sched_dispatch_requests(struct

blk_mq_hw_ctx *hctx)
         */
        if (!list_empty(&rq_list)) {
                blk_mq_sched_mark_restart_hctx(hctx);
-               if (blk_mq_dispatch_rq_list(q, &rq_list, false)) {
+               dispatch_ret = blk_mq_dispatch_rq_list(q, &rq_list, false);

Suppose we dispatch one request to block driver, and return 'true' here.

+               if (dispatch_ret) {
                        if (has_sched_dispatch)
-                               blk_mq_do_dispatch_sched(hctx);
+                               dispatch_ret = blk_mq_do_dispatch_sched(hctx);

Then we will continue to try to dispatch more requests from IO
scheduler, but if there are no requests in IO scheduler now, it will
return 'false' here, and set dispatch_ret as false.

                        else
-                               blk_mq_do_dispatch_ctx(hctx);
+                               dispatch_ret = blk_mq_do_dispatch_ctx(hctx);

OK, this one is an issue, but it can be fixed simply by not updating
'dispatch_ret' for the following dispatch, something like the below
way:

        if (dispatch_ret) {
                if (has_sched_dispatch)
                        blk_mq_do_dispatch_sched(hctx);
                else
                        blk_mq_do_dispatch_ctx(hctx);
        }

Yes, this can work.

But I found your patch will drop some performance comparing with my
method in patch 1/2. My method can fetch several requests from IO
scheduler and dispatch them to block driver at one time, but in your
patch we still need dispatch request one by one, which will drop some
performance I think.
What do you think? Thanks.

Please run your test and see if performance drop can be observed.

From my testing (using the same fio configuration in cover letter), I
found your method will drop some performance from below data.

My original patches:
Sequential read: 229.6MiB/s
Random read:180.8MiB/s
Sequential write: 172MiB/s
Random write:169.2MiB/s

Your patches:
Sequential read: 209MiB/s
Random read:177MiB/s
Sequential write: 148MiB/s
Random write:147MiB/s

After some optimiziton and I did more testing, I did not found any
performance issue with your patch comparing with my old method. Sorry
for noise in my last email.

So will you send out a formal patch? If yes, please add my test-by
tag. Thanks for your help.
Tested-by: Baolin Wang <baolin.wang7@gmail.com>

-- 
Baolin Wang