]> git.kernelconcepts.de Git - karo-tx-linux.git/commitdiff
dm: remove support for stacking dm-mq on .request_fn device(s)
authorMike Snitzer <snitzer@redhat.com>
Sat, 20 Feb 2016 19:02:49 +0000 (14:02 -0500)
committerMike Snitzer <snitzer@redhat.com>
Tue, 23 Feb 2016 03:33:46 +0000 (22:33 -0500)
Remove all fiddley code that propped up this support for a blk-mq
request-queue ontop of all .request_fn devices.

Testing has proven this niche request-based dm-mq mode to be buggy, when
testing fault tolerance with DM multipath, and there is no point trying
to preserve it.

Should help improve efficiency of pure dm-mq code and make code
maintenance less delicate.

Signed-off-by: Mike Snitzer <snitzer@redhat.com>
drivers/md/dm-mpath.c
drivers/md/dm.c

index 3ddaa11584ef80bea8416f25cf99619cef307461..7259eeba6a5892765d011d7498595196081e0fca 100644 (file)
@@ -418,7 +418,10 @@ static int __multipath_map(struct dm_target *ti, struct request *clone,
        spin_unlock_irq(&m->lock);
 
        if (clone) {
-               /* Old request-based interface: allocated clone is passed in */
+               /*
+                * Old request-based interface: allocated clone is passed in.
+                * Used by: .request_fn stacked on .request_fn path(s).
+                */
                clone->q = bdev_get_queue(bdev);
                clone->rq_disk = bdev->bd_disk;
                clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
index 8a62e43fbff505e78fa30cc9c2789469e1c6e7fe..26fedd93702e1c56cdc9d411a31445988972371f 100644 (file)
@@ -1141,11 +1141,6 @@ static void free_rq_clone(struct request *clone)
        else if (!md->queue->mq_ops)
                /* request_fn queue stacked on request_fn queue(s) */
                free_clone_request(md, clone);
-       /*
-        * NOTE: for the blk-mq queue stacked on request_fn queue(s) case:
-        * no need to call free_clone_request() because we leverage blk-mq by
-        * allocating the clone at the end of the blk-mq pdu (see: clone_rq)
-        */
 
        if (!md->queue->mq_ops)
                free_rq_tio(tio);
@@ -1866,24 +1861,18 @@ static struct request *clone_rq(struct request *rq, struct mapped_device *md,
                                struct dm_rq_target_io *tio, gfp_t gfp_mask)
 {
        /*
-        * Do not allocate a clone if tio->clone was already set
-        * (see: dm_mq_queue_rq).
+        * Create clone for use with .request_fn request_queue
         */
-       bool alloc_clone = !tio->clone;
        struct request *clone;
 
-       if (alloc_clone) {
-               clone = alloc_clone_request(md, gfp_mask);
-               if (!clone)
-                       return NULL;
-       } else
-               clone = tio->clone;
+       clone = alloc_clone_request(md, gfp_mask);
+       if (!clone)
+               return NULL;
 
        blk_rq_init(NULL, clone);
        if (setup_clone(clone, rq, tio, gfp_mask)) {
                /* -ENOMEM */
-               if (alloc_clone)
-                       free_clone_request(md, clone);
+               free_clone_request(md, clone);
                return NULL;
        }
 
@@ -2692,22 +2681,12 @@ static int dm_mq_queue_rq(struct blk_mq_hw_ctx *hctx,
         */
        tio->ti = ti;
 
-       /*
-        * Both the table and md type cannot change after initial table load
-        */
-       if (dm_get_md_type(md) == DM_TYPE_REQUEST_BASED) {
-               /* clone request is allocated at the end of the pdu */
-               tio->clone = (void *)blk_mq_rq_to_pdu(rq) + sizeof(struct dm_rq_target_io);
-               (void) clone_rq(rq, md, tio, GFP_ATOMIC);
-               queue_kthread_work(&md->kworker, &tio->work);
-       } else {
-               /* Direct call is fine since .queue_rq allows allocations */
-               if (map_request(tio, rq, md) == DM_MAPIO_REQUEUE) {
-                       /* Undo dm_start_request() before requeuing */
-                       rq_end_stats(md, rq);
-                       rq_completed(md, rq_data_dir(rq), false);
-                       return BLK_MQ_RQ_QUEUE_BUSY;
-               }
+       /* Direct call is fine since .queue_rq allows allocations */
+       if (map_request(tio, rq, md) == DM_MAPIO_REQUEUE) {
+               /* Undo dm_start_request() before requeuing */
+               rq_end_stats(md, rq);
+               rq_completed(md, rq_data_dir(rq), false);
+               return BLK_MQ_RQ_QUEUE_BUSY;
        }
 
        return BLK_MQ_RQ_QUEUE_OK;
@@ -2726,6 +2705,11 @@ static int dm_init_request_based_blk_mq_queue(struct mapped_device *md)
        struct request_queue *q;
        int err;
 
+       if (dm_get_md_type(md) == DM_TYPE_REQUEST_BASED) {
+               DMERR("request-based dm-mq may only be stacked on blk-mq device(s)");
+               return -EINVAL;
+       }
+
        md->tag_set = kzalloc(sizeof(struct blk_mq_tag_set), GFP_KERNEL);
        if (!md->tag_set)
                return -ENOMEM;
@@ -2738,10 +2722,6 @@ static int dm_init_request_based_blk_mq_queue(struct mapped_device *md)
        md->tag_set->driver_data = md;
 
        md->tag_set->cmd_size = sizeof(struct dm_rq_target_io);
-       if (md_type == DM_TYPE_REQUEST_BASED) {
-               /* put the memory for non-blk-mq clone at the end of the pdu */
-               md->tag_set->cmd_size += sizeof(struct request);
-       }
 
        err = blk_mq_alloc_tag_set(md->tag_set);
        if (err)
@@ -2758,9 +2738,6 @@ static int dm_init_request_based_blk_mq_queue(struct mapped_device *md)
        /* backfill 'mq' sysfs registration normally done in blk_register_queue */
        blk_mq_register_disk(md->disk);
 
-       if (md_type == DM_TYPE_REQUEST_BASED)
-               init_rq_based_worker_thread(md);
-
        return 0;
 
 out_tag_set: