Merge tag 'for-linus' of git://git.kernel.org/pub/scm/virt/kvm/kvm
[sfrench/cifs-2.6.git] / block / blk-flush.c
index 3cb5e9e7108adc8473296d8fe762bb16338d17f4..20badd7b9d1b664f534a791b83435534f94a70ec 100644 (file)
@@ -28,7 +28,7 @@
  *
  * The actual execution of flush is double buffered.  Whenever a request
  * needs to execute PRE or POSTFLUSH, it queues at
- * q->flush_queue[q->flush_pending_idx].  Once certain criteria are met, a
+ * fq->flush_queue[fq->flush_pending_idx].  Once certain criteria are met, a
  * flush is issued and the pending_idx is toggled.  When the flush
  * completes, all the requests which were pending are proceeded to the next
  * step.  This allows arbitrary merging of different types of FLUSH/FUA
@@ -91,7 +91,8 @@ enum {
        FLUSH_PENDING_TIMEOUT   = 5 * HZ,
 };
 
-static bool blk_kick_flush(struct request_queue *q);
+static bool blk_kick_flush(struct request_queue *q,
+                          struct blk_flush_queue *fq);
 
 static unsigned int blk_flush_policy(unsigned int fflags, struct request *rq)
 {
@@ -126,8 +127,6 @@ static void blk_flush_restore_request(struct request *rq)
        /* make @rq a normal request */
        rq->cmd_flags &= ~REQ_FLUSH_SEQ;
        rq->end_io = rq->flush.saved_end_io;
-
-       blk_clear_rq_complete(rq);
 }
 
 static bool blk_flush_queue_rq(struct request *rq, bool add_front)
@@ -150,6 +149,7 @@ static bool blk_flush_queue_rq(struct request *rq, bool add_front)
 /**
  * blk_flush_complete_seq - complete flush sequence
  * @rq: FLUSH/FUA request being sequenced
+ * @fq: flush queue
  * @seq: sequences to complete (mask of %REQ_FSEQ_*, can be zero)
  * @error: whether an error occurred
  *
@@ -157,16 +157,17 @@ static bool blk_flush_queue_rq(struct request *rq, bool add_front)
  * completion and trigger the next step.
  *
  * CONTEXT:
- * spin_lock_irq(q->queue_lock or q->mq_flush_lock)
+ * spin_lock_irq(q->queue_lock or fq->mq_flush_lock)
  *
  * RETURNS:
  * %true if requests were added to the dispatch queue, %false otherwise.
  */
-static bool blk_flush_complete_seq(struct request *rq, unsigned int seq,
-                                  int error)
+static bool blk_flush_complete_seq(struct request *rq,
+                                  struct blk_flush_queue *fq,
+                                  unsigned int seq, int error)
 {
        struct request_queue *q = rq->q;
-       struct list_head *pending = &q->flush_queue[q->flush_pending_idx];
+       struct list_head *pending = &fq->flush_queue[fq->flush_pending_idx];
        bool queued = false, kicked;
 
        BUG_ON(rq->flush.seq & seq);
@@ -182,12 +183,12 @@ static bool blk_flush_complete_seq(struct request *rq, unsigned int seq,
        case REQ_FSEQ_POSTFLUSH:
                /* queue for flush */
                if (list_empty(pending))
-                       q->flush_pending_since = jiffies;
+                       fq->flush_pending_since = jiffies;
                list_move_tail(&rq->flush.list, pending);
                break;
 
        case REQ_FSEQ_DATA:
-               list_move_tail(&rq->flush.list, &q->flush_data_in_flight);
+               list_move_tail(&rq->flush.list, &fq->flush_data_in_flight);
                queued = blk_flush_queue_rq(rq, true);
                break;
 
@@ -202,7 +203,7 @@ static bool blk_flush_complete_seq(struct request *rq, unsigned int seq,
                list_del_init(&rq->flush.list);
                blk_flush_restore_request(rq);
                if (q->mq_ops)
-                       blk_mq_end_io(rq, error);
+                       blk_mq_end_request(rq, error);
                else
                        __blk_end_request_all(rq, error);
                break;
@@ -211,7 +212,7 @@ static bool blk_flush_complete_seq(struct request *rq, unsigned int seq,
                BUG();
        }
 
-       kicked = blk_kick_flush(q);
+       kicked = blk_kick_flush(q, fq);
        return kicked | queued;
 }
 
@@ -222,17 +223,18 @@ static void flush_end_io(struct request *flush_rq, int error)
        bool queued = false;
        struct request *rq, *n;
        unsigned long flags = 0;
+       struct blk_flush_queue *fq = blk_get_flush_queue(q, flush_rq->mq_ctx);
 
        if (q->mq_ops) {
-               spin_lock_irqsave(&q->mq_flush_lock, flags);
-               q->flush_rq->tag = -1;
+               spin_lock_irqsave(&fq->mq_flush_lock, flags);
+               flush_rq->tag = -1;
        }
 
-       running = &q->flush_queue[q->flush_running_idx];
-       BUG_ON(q->flush_pending_idx == q->flush_running_idx);
+       running = &fq->flush_queue[fq->flush_running_idx];
+       BUG_ON(fq->flush_pending_idx == fq->flush_running_idx);
 
        /* account completion of the flush request */
-       q->flush_running_idx ^= 1;
+       fq->flush_running_idx ^= 1;
 
        if (!q->mq_ops)
                elv_completed_request(q, flush_rq);
@@ -242,7 +244,7 @@ static void flush_end_io(struct request *flush_rq, int error)
                unsigned int seq = blk_flush_cur_seq(rq);
 
                BUG_ON(seq != REQ_FSEQ_PREFLUSH && seq != REQ_FSEQ_POSTFLUSH);
-               queued |= blk_flush_complete_seq(rq, seq, error);
+               queued |= blk_flush_complete_seq(rq, fq, seq, error);
        }
 
        /*
@@ -256,71 +258,81 @@ static void flush_end_io(struct request *flush_rq, int error)
         * directly into request_fn may confuse the driver.  Always use
         * kblockd.
         */
-       if (queued || q->flush_queue_delayed) {
+       if (queued || fq->flush_queue_delayed) {
                WARN_ON(q->mq_ops);
                blk_run_queue_async(q);
        }
-       q->flush_queue_delayed = 0;
+       fq->flush_queue_delayed = 0;
        if (q->mq_ops)
-               spin_unlock_irqrestore(&q->mq_flush_lock, flags);
+               spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
 }
 
 /**
  * blk_kick_flush - consider issuing flush request
  * @q: request_queue being kicked
+ * @fq: flush queue
  *
  * Flush related states of @q have changed, consider issuing flush request.
  * Please read the comment at the top of this file for more info.
  *
  * CONTEXT:
- * spin_lock_irq(q->queue_lock or q->mq_flush_lock)
+ * spin_lock_irq(q->queue_lock or fq->mq_flush_lock)
  *
  * RETURNS:
  * %true if flush was issued, %false otherwise.
  */
-static bool blk_kick_flush(struct request_queue *q)
+static bool blk_kick_flush(struct request_queue *q, struct blk_flush_queue *fq)
 {
-       struct list_head *pending = &q->flush_queue[q->flush_pending_idx];
+       struct list_head *pending = &fq->flush_queue[fq->flush_pending_idx];
        struct request *first_rq =
                list_first_entry(pending, struct request, flush.list);
+       struct request *flush_rq = fq->flush_rq;
 
        /* C1 described at the top of this file */
-       if (q->flush_pending_idx != q->flush_running_idx || list_empty(pending))
+       if (fq->flush_pending_idx != fq->flush_running_idx || list_empty(pending))
                return false;
 
        /* C2 and C3 */
-       if (!list_empty(&q->flush_data_in_flight) &&
+       if (!list_empty(&fq->flush_data_in_flight) &&
            time_before(jiffies,
-                       q->flush_pending_since + FLUSH_PENDING_TIMEOUT))
+                       fq->flush_pending_since + FLUSH_PENDING_TIMEOUT))
                return false;
 
        /*
         * Issue flush and toggle pending_idx.  This makes pending_idx
         * different from running_idx, which means flush is in flight.
         */
-       q->flush_pending_idx ^= 1;
+       fq->flush_pending_idx ^= 1;
 
-       blk_rq_init(q, q->flush_rq);
-       if (q->mq_ops)
-               blk_mq_clone_flush_request(q->flush_rq, first_rq);
+       blk_rq_init(q, flush_rq);
+
+       /*
+        * Borrow tag from the first request since they can't
+        * be in flight at the same time.
+        */
+       if (q->mq_ops) {
+               flush_rq->mq_ctx = first_rq->mq_ctx;
+               flush_rq->tag = first_rq->tag;
+       }
 
-       q->flush_rq->cmd_type = REQ_TYPE_FS;
-       q->flush_rq->cmd_flags = WRITE_FLUSH | REQ_FLUSH_SEQ;
-       q->flush_rq->rq_disk = first_rq->rq_disk;
-       q->flush_rq->end_io = flush_end_io;
+       flush_rq->cmd_type = REQ_TYPE_FS;
+       flush_rq->cmd_flags = WRITE_FLUSH | REQ_FLUSH_SEQ;
+       flush_rq->rq_disk = first_rq->rq_disk;
+       flush_rq->end_io = flush_end_io;
 
-       return blk_flush_queue_rq(q->flush_rq, false);
+       return blk_flush_queue_rq(flush_rq, false);
 }
 
 static void flush_data_end_io(struct request *rq, int error)
 {
        struct request_queue *q = rq->q;
+       struct blk_flush_queue *fq = blk_get_flush_queue(q, NULL);
 
        /*
         * After populating an empty queue, kick it to avoid stall.  Read
         * the comment in flush_end_io().
         */
-       if (blk_flush_complete_seq(rq, REQ_FSEQ_DATA, error))
+       if (blk_flush_complete_seq(rq, fq, REQ_FSEQ_DATA, error))
                blk_run_queue_async(q);
 }
 
@@ -328,20 +340,20 @@ static void mq_flush_data_end_io(struct request *rq, int error)
 {
        struct request_queue *q = rq->q;
        struct blk_mq_hw_ctx *hctx;
-       struct blk_mq_ctx *ctx;
+       struct blk_mq_ctx *ctx = rq->mq_ctx;
        unsigned long flags;
+       struct blk_flush_queue *fq = blk_get_flush_queue(q, ctx);
 
-       ctx = rq->mq_ctx;
        hctx = q->mq_ops->map_queue(q, ctx->cpu);
 
        /*
         * After populating an empty queue, kick it to avoid stall.  Read
         * the comment in flush_end_io().
         */
-       spin_lock_irqsave(&q->mq_flush_lock, flags);
-       if (blk_flush_complete_seq(rq, REQ_FSEQ_DATA, error))
+       spin_lock_irqsave(&fq->mq_flush_lock, flags);
+       if (blk_flush_complete_seq(rq, fq, REQ_FSEQ_DATA, error))
                blk_mq_run_hw_queue(hctx, true);
-       spin_unlock_irqrestore(&q->mq_flush_lock, flags);
+       spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
 }
 
 /**
@@ -361,6 +373,7 @@ void blk_insert_flush(struct request *rq)
        struct request_queue *q = rq->q;
        unsigned int fflags = q->flush_flags;   /* may change, cache */
        unsigned int policy = blk_flush_policy(fflags, rq);
+       struct blk_flush_queue *fq = blk_get_flush_queue(q, rq->mq_ctx);
 
        /*
         * @policy now records what operations need to be done.  Adjust
@@ -378,7 +391,7 @@ void blk_insert_flush(struct request *rq)
         */
        if (!policy) {
                if (q->mq_ops)
-                       blk_mq_end_io(rq, 0);
+                       blk_mq_end_request(rq, 0);
                else
                        __blk_end_bidi_request(rq, 0, 0, 0);
                return;
@@ -411,14 +424,14 @@ void blk_insert_flush(struct request *rq)
        if (q->mq_ops) {
                rq->end_io = mq_flush_data_end_io;
 
-               spin_lock_irq(&q->mq_flush_lock);
-               blk_flush_complete_seq(rq, REQ_FSEQ_ACTIONS & ~policy, 0);
-               spin_unlock_irq(&q->mq_flush_lock);
+               spin_lock_irq(&fq->mq_flush_lock);
+               blk_flush_complete_seq(rq, fq, REQ_FSEQ_ACTIONS & ~policy, 0);
+               spin_unlock_irq(&fq->mq_flush_lock);
                return;
        }
        rq->end_io = flush_data_end_io;
 
-       blk_flush_complete_seq(rq, REQ_FSEQ_ACTIONS & ~policy, 0);
+       blk_flush_complete_seq(rq, fq, REQ_FSEQ_ACTIONS & ~policy, 0);
 }
 
 /**
@@ -474,7 +487,43 @@ int blkdev_issue_flush(struct block_device *bdev, gfp_t gfp_mask,
 }
 EXPORT_SYMBOL(blkdev_issue_flush);
 
-void blk_mq_init_flush(struct request_queue *q)
+struct blk_flush_queue *blk_alloc_flush_queue(struct request_queue *q,
+               int node, int cmd_size)
+{
+       struct blk_flush_queue *fq;
+       int rq_sz = sizeof(struct request);
+
+       fq = kzalloc_node(sizeof(*fq), GFP_KERNEL, node);
+       if (!fq)
+               goto fail;
+
+       if (q->mq_ops) {
+               spin_lock_init(&fq->mq_flush_lock);
+               rq_sz = round_up(rq_sz + cmd_size, cache_line_size());
+       }
+
+       fq->flush_rq = kzalloc_node(rq_sz, GFP_KERNEL, node);
+       if (!fq->flush_rq)
+               goto fail_rq;
+
+       INIT_LIST_HEAD(&fq->flush_queue[0]);
+       INIT_LIST_HEAD(&fq->flush_queue[1]);
+       INIT_LIST_HEAD(&fq->flush_data_in_flight);
+
+       return fq;
+
+ fail_rq:
+       kfree(fq);
+ fail:
+       return NULL;
+}
+
+void blk_free_flush_queue(struct blk_flush_queue *fq)
 {
-       spin_lock_init(&q->mq_flush_lock);
+       /* bio based request queue hasn't flush queue */
+       if (!fq)
+               return;
+
+       kfree(fq->flush_rq);
+       kfree(fq);
 }