OSDN Git Service

blk-mq: fix issue with shared tag queue re-running
authorJens Axboe <axboe@kernel.dk>
Thu, 9 Nov 2017 15:32:43 +0000 (08:32 -0700)
committerJens Axboe <axboe@kernel.dk>
Sat, 11 Nov 2017 02:53:25 +0000 (19:53 -0700)
This patch attempts to make the case of hctx re-running on driver tag
failure more robust. Without this patch, it's pretty easy to trigger a
stall condition with shared tags. An example is using null_blk like
this:

modprobe null_blk queue_mode=2 nr_devices=4 shared_tags=1 submit_queues=1 hw_queue_depth=1

which sets up 4 devices, sharing the same tag set with a depth of 1.
Running a fio job ala:

[global]
bs=4k
rw=randread
norandommap
direct=1
ioengine=libaio
iodepth=4

[nullb0]
filename=/dev/nullb0
[nullb1]
filename=/dev/nullb1
[nullb2]
filename=/dev/nullb2
[nullb3]
filename=/dev/nullb3

will inevitably end with one or more threads being stuck waiting for a
scheduler tag. That IO is then stuck forever, until someone else
triggers a run of the queue.

Ensure that we always re-run the hardware queue, if the driver tag we
were waiting for got freed before we added our leftover request entries
back on the dispatch list.

Reviewed-by: Bart Van Assche <bart.vanassche@wdc.com>
Tested-by: Bart Van Assche <bart.vanassche@wdc.com>
Reviewed-by: Ming Lei <ming.lei@redhat.com>
Reviewed-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
block/blk-mq-debugfs.c
block/blk-mq.c
include/linux/blk-mq.h

index 7f4a1ba..bb7f084 100644 (file)
@@ -179,7 +179,6 @@ static const char *const hctx_state_name[] = {
        HCTX_STATE_NAME(STOPPED),
        HCTX_STATE_NAME(TAG_ACTIVE),
        HCTX_STATE_NAME(SCHED_RESTART),
-       HCTX_STATE_NAME(TAG_WAITING),
        HCTX_STATE_NAME(START_ON_RUN),
 };
 #undef HCTX_STATE_NAME
index 3d759bb..fed8165 100644 (file)
@@ -998,49 +998,64 @@ done:
        return rq->tag != -1;
 }
 
-static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, int flags,
-                               void *key)
+static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
+                               int flags, void *key)
 {
        struct blk_mq_hw_ctx *hctx;
 
        hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
 
-       list_del(&wait->entry);
-       clear_bit_unlock(BLK_MQ_S_TAG_WAITING, &hctx->state);
+       list_del_init(&wait->entry);
        blk_mq_run_hw_queue(hctx, true);
        return 1;
 }
 
-static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx *hctx)
+static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx **hctx,
+                                    struct request *rq)
 {
+       struct blk_mq_hw_ctx *this_hctx = *hctx;
+       wait_queue_entry_t *wait = &this_hctx->dispatch_wait;
        struct sbq_wait_state *ws;
 
+       if (!list_empty_careful(&wait->entry))
+               return false;
+
+       spin_lock(&this_hctx->lock);
+       if (!list_empty(&wait->entry)) {
+               spin_unlock(&this_hctx->lock);
+               return false;
+       }
+
+       ws = bt_wait_ptr(&this_hctx->tags->bitmap_tags, this_hctx);
+       add_wait_queue(&ws->wait, wait);
+
        /*
-        * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait.
-        * The thread which wins the race to grab this bit adds the hardware
-        * queue to the wait queue.
+        * It's possible that a tag was freed in the window between the
+        * allocation failure and adding the hardware queue to the wait
+        * queue.
         */
-       if (test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state) ||
-           test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING, &hctx->state))
+       if (!blk_mq_get_driver_tag(rq, hctx, false)) {
+               spin_unlock(&this_hctx->lock);
                return false;
-
-       init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
-       ws = bt_wait_ptr(&hctx->tags->bitmap_tags, hctx);
+       }
 
        /*
-        * As soon as this returns, it's no longer safe to fiddle with
-        * hctx->dispatch_wait, since a completion can wake up the wait queue
-        * and unlock the bit.
+        * We got a tag, remove ourselves from the wait queue to ensure
+        * someone else gets the wakeup.
         */
-       add_wait_queue(&ws->wait, &hctx->dispatch_wait);
+       spin_lock_irq(&ws->wait.lock);
+       list_del_init(&wait->entry);
+       spin_unlock_irq(&ws->wait.lock);
+       spin_unlock(&this_hctx->lock);
        return true;
 }
 
 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list,
-               bool got_budget)
+                            bool got_budget)
 {
        struct blk_mq_hw_ctx *hctx;
        struct request *rq, *nxt;
+       bool no_tag = false;
        int errors, queued;
 
        if (list_empty(list))
@@ -1060,22 +1075,15 @@ bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list,
                if (!blk_mq_get_driver_tag(rq, &hctx, false)) {
                        /*
                         * The initial allocation attempt failed, so we need to
-                        * rerun the hardware queue when a tag is freed.
+                        * rerun the hardware queue when a tag is freed. The
+                        * waitqueue takes care of that. If the queue is run
+                        * before we add this entry back on the dispatch list,
+                        * we'll re-run it below.
                         */
-                       if (!blk_mq_dispatch_wait_add(hctx)) {
-                               if (got_budget)
-                                       blk_mq_put_dispatch_budget(hctx);
-                               break;
-                       }
-
-                       /*
-                        * It's possible that a tag was freed in the window
-                        * between the allocation failure and adding the
-                        * hardware queue to the wait queue.
-                        */
-                       if (!blk_mq_get_driver_tag(rq, &hctx, false)) {
+                       if (!blk_mq_dispatch_wait_add(&hctx, rq)) {
                                if (got_budget)
                                        blk_mq_put_dispatch_budget(hctx);
+                               no_tag = true;
                                break;
                        }
                }
@@ -1140,10 +1148,10 @@ bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list,
                 * it is no longer set that means that it was cleared by another
                 * thread and hence that a queue rerun is needed.
                 *
-                * If TAG_WAITING is set that means that an I/O scheduler has
-                * been configured and another thread is waiting for a driver
-                * tag. To guarantee fairness, do not rerun this hardware queue
-                * but let the other thread grab the driver tag.
+                * If 'no_tag' is set, that means that we failed getting
+                * a driver tag with an I/O scheduler attached. If our dispatch
+                * waitqueue is no longer active, ensure that we run the queue
+                * AFTER adding our entries back to the list.
                 *
                 * If no I/O scheduler has been configured it is possible that
                 * the hardware queue got stopped and restarted before requests
@@ -1155,8 +1163,8 @@ bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list,
                 *   returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
                 *   and dm-rq.
                 */
-               if (!blk_mq_sched_needs_restart(hctx) &&
-                   !test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state))
+               if (!blk_mq_sched_needs_restart(hctx) ||
+                   (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
                        blk_mq_run_hw_queue(hctx, true);
        }
 
@@ -2020,6 +2028,9 @@ static int blk_mq_init_hctx(struct request_queue *q,
 
        hctx->nr_ctx = 0;
 
+       init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
+       INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
+
        if (set->ops->init_hctx &&
            set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
                goto free_bitmap;
index 6746415..4ae987c 100644 (file)
@@ -35,7 +35,7 @@ struct blk_mq_hw_ctx {
        struct blk_mq_ctx       **ctxs;
        unsigned int            nr_ctx;
 
-       wait_queue_entry_t              dispatch_wait;
+       wait_queue_entry_t      dispatch_wait;
        atomic_t                wait_index;
 
        struct blk_mq_tags      *tags;
@@ -181,8 +181,7 @@ enum {
        BLK_MQ_S_STOPPED        = 0,
        BLK_MQ_S_TAG_ACTIVE     = 1,
        BLK_MQ_S_SCHED_RESTART  = 2,
-       BLK_MQ_S_TAG_WAITING    = 3,
-       BLK_MQ_S_START_ON_RUN   = 4,
+       BLK_MQ_S_START_ON_RUN   = 3,
 
        BLK_MQ_MAX_DEPTH        = 10240,