blk-mq: scale depth and rq map appropriate if low on memory
authorJens Axboe <axboe@fb.com>
Wed, 10 Sep 2014 15:02:03 +0000 (09:02 -0600)
committerJens Axboe <axboe@fb.com>
Wed, 10 Sep 2014 15:02:03 +0000 (09:02 -0600)
If we are running in a kdump environment, resources are scarce.
For some SCSI setups with a huge set of shared tags, we run out
of memory allocating what the drivers is asking for. So implement
a scale back logic to reduce the tag depth for those cases, allowing
the driver to successfully load.

We should extend this to detect low memory situations, and implement
a sane fallback for those (1 queue, 64 tags, or something like that).

Tested-by: Robert Elliott <elliott@hp.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
block/blk-mq.c

index f9b85e83d9ba639e408843362a7933b006898a94..383ea0cb1f0a295463789d956aae820b334f4a55 100644 (file)
@@ -1321,6 +1321,7 @@ static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
                                continue;
                        set->ops->exit_request(set->driver_data, tags->rqs[i],
                                                hctx_idx, i);
+                       tags->rqs[i] = NULL;
                }
        }
 
@@ -1354,8 +1355,9 @@ static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
 
        INIT_LIST_HEAD(&tags->page_list);
 
-       tags->rqs = kmalloc_node(set->queue_depth * sizeof(struct request *),
-                                       GFP_KERNEL, set->numa_node);
+       tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
+                                GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
+                                set->numa_node);
        if (!tags->rqs) {
                blk_mq_free_tags(tags);
                return NULL;
@@ -1379,8 +1381,9 @@ static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
                        this_order--;
 
                do {
-                       page = alloc_pages_node(set->numa_node, GFP_KERNEL,
-                                               this_order);
+                       page = alloc_pages_node(set->numa_node,
+                               GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
+                               this_order);
                        if (page)
                                break;
                        if (!this_order--)
@@ -1404,8 +1407,10 @@ static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
                        if (set->ops->init_request) {
                                if (set->ops->init_request(set->driver_data,
                                                tags->rqs[i], hctx_idx, i,
-                                               set->numa_node))
+                                               set->numa_node)) {
+                                       tags->rqs[i] = NULL;
                                        goto fail;
+                               }
                        }
 
                        p += rq_size;
@@ -1416,7 +1421,6 @@ static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
        return tags;
 
 fail:
-       pr_warn("%s: failed to allocate requests\n", __func__);
        blk_mq_free_rq_map(set, tags, hctx_idx);
        return NULL;
 }
@@ -1936,6 +1940,61 @@ static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
        return NOTIFY_OK;
 }
 
+static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
+{
+       int i;
+
+       for (i = 0; i < set->nr_hw_queues; i++) {
+               set->tags[i] = blk_mq_init_rq_map(set, i);
+               if (!set->tags[i])
+                       goto out_unwind;
+       }
+
+       return 0;
+
+out_unwind:
+       while (--i >= 0)
+               blk_mq_free_rq_map(set, set->tags[i], i);
+
+       set->tags = NULL;
+       return -ENOMEM;
+}
+
+/*
+ * Allocate the request maps associated with this tag_set. Note that this
+ * may reduce the depth asked for, if memory is tight. set->queue_depth
+ * will be updated to reflect the allocated depth.
+ */
+static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
+{
+       unsigned int depth;
+       int err;
+
+       depth = set->queue_depth;
+       do {
+               err = __blk_mq_alloc_rq_maps(set);
+               if (!err)
+                       break;
+
+               set->queue_depth >>= 1;
+               if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
+                       err = -ENOMEM;
+                       break;
+               }
+       } while (set->queue_depth);
+
+       if (!set->queue_depth || err) {
+               pr_err("blk-mq: failed to allocate request map\n");
+               return -ENOMEM;
+       }
+
+       if (depth != set->queue_depth)
+               pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
+                                               depth, set->queue_depth);
+
+       return 0;
+}
+
 /*
  * Alloc a tag set to be associated with one or more request queues.
  * May fail with EINVAL for various error conditions. May adjust the
@@ -1944,8 +2003,6 @@ static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
  */
 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
 {
-       int i;
-
        if (!set->nr_hw_queues)
                return -EINVAL;
        if (!set->queue_depth)
@@ -1966,25 +2023,18 @@ int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
                                 sizeof(struct blk_mq_tags *),
                                 GFP_KERNEL, set->numa_node);
        if (!set->tags)
-               goto out;
+               return -ENOMEM;
 
-       for (i = 0; i < set->nr_hw_queues; i++) {
-               set->tags[i] = blk_mq_init_rq_map(set, i);
-               if (!set->tags[i])
-                       goto out_unwind;
-       }
+       if (blk_mq_alloc_rq_maps(set))
+               goto enomem;
 
        mutex_init(&set->tag_list_lock);
        INIT_LIST_HEAD(&set->tag_list);
 
        return 0;
-
-out_unwind:
-       while (--i >= 0)
-               blk_mq_free_rq_map(set, set->tags[i], i);
+enomem:
        kfree(set->tags);
        set->tags = NULL;
-out:
        return -ENOMEM;
 }
 EXPORT_SYMBOL(blk_mq_alloc_tag_set);