Pre-allocate vdev I/O buffers

The vdev queue layer may require a small number of buffers
when attempting to create aggregate I/O requests.  Rather than
attempting to allocate them from the global zio buffers, which
is slow under memory pressure, it makes sense to pre-allocate
them because...

1) These buffers are short lived.  They are only required for
the life of a single I/O at which point they can be used by
the next I/O.

2) The maximum number of concurrent buffers needed by a vdev is
small.  It's roughly limited by the zfs_vdev_max_pending tunable
which defaults to 10.

By keeping a small list of these buffer per-vdev we can ensure
one is always available when we need it.  This significantly
reduces contention on the vq->vq_lock, because we no longer
need to perform a slow allocation under this lock.  This is
particularly important when memory is already low on the system.

It would probably be wise to extend the use of these buffers beyond
aggregate I/O and in to the raidz implementation.  The inability
to quickly allocate buffer for the parity stripes could result in
similiar problems.

Signed-off-by: Brian Behlendorf <[email protected]>

4 files changed, 65 insertions, 2 deletions

diff --git a/include/sys/vdev_impl.h b/include/sys/vdev_impl.h
index 1df61a587..8862d9bc9 100644
--- a/include/sys/vdev_impl.h
+++ b/include/sys/vdev_impl.h
@@ -49,6 +49,7 @@ extern "C" {
  * Forward declarations that lots of things need.
  */
 typedef struct vdev_queue vdev_queue_t;
+typedef struct vdev_io vdev_io_t;
 typedef struct vdev_cache vdev_cache_t;
 typedef struct vdev_cache_entry vdev_cache_entry_t;
 
@@ -102,9 +103,15 @@ struct vdev_queue {
 	avl_tree_t	vq_read_tree;
 	avl_tree_t	vq_write_tree;
 	avl_tree_t	vq_pending_tree;
+	list_t		vq_io_list;
 	kmutex_t	vq_lock;
 };
 
+struct vdev_io {
+	char		vi_buffer[SPA_MAXBLOCKSIZE]; /* Must be first */
+	list_node_t	vi_node;
+};
+
 /*
  * Virtual device descriptor
  */
diff --git a/include/sys/zio.h b/include/sys/zio.h
index c0da4e2d7..4f20cab65 100644
--- a/include/sys/zio.h
+++ b/include/sys/zio.h
@@ -494,6 +494,8 @@ extern void *zio_buf_alloc(size_t size);
 extern void zio_buf_free(void *buf, size_t size);
 extern void *zio_data_buf_alloc(size_t size);
 extern void zio_data_buf_free(void *buf, size_t size);
+extern void *zio_vdev_alloc(void);
+extern void zio_vdev_free(void *buf);
 
 extern void zio_resubmit_stage_async(void *);
 
diff --git a/module/zfs/vdev_queue.c b/module/zfs/vdev_queue.c
index aacc55c49..7ba638952 100644
--- a/module/zfs/vdev_queue.c
+++ b/module/zfs/vdev_queue.c
@@ -106,6 +106,7 @@ void
 vdev_queue_init(vdev_t *vd)
 {
 	vdev_queue_t *vq = &vd->vdev_queue;
+	int i;
 
 	mutex_init(&vq->vq_lock, NULL, MUTEX_DEFAULT, NULL);
 
@@ -120,18 +121,36 @@ vdev_queue_init(vdev_t *vd)
 
 	avl_create(&vq->vq_pending_tree, vdev_queue_offset_compare,
 	    sizeof (zio_t), offsetof(struct zio, io_offset_node));
+
+	/*
+	 * A list of buffers which can be used for aggregate I/O, this
+	 * avoids the need to allocate them on demand when memory is low.
+	 */
+	list_create(&vq->vq_io_list, sizeof (vdev_io_t),
+	    offsetof(vdev_io_t, vi_node));
+
+	for (i = 0; i < zfs_vdev_max_pending; i++)
+		list_insert_tail(&vq->vq_io_list, zio_vdev_alloc());
 }
 
 void
 vdev_queue_fini(vdev_t *vd)
 {
 	vdev_queue_t *vq = &vd->vdev_queue;
+	vdev_io_t *vi;
 
 	avl_destroy(&vq->vq_deadline_tree);
 	avl_destroy(&vq->vq_read_tree);
 	avl_destroy(&vq->vq_write_tree);
 	avl_destroy(&vq->vq_pending_tree);
 
+	while ((vi = list_head(&vq->vq_io_list)) != NULL) {
+		list_remove(&vq->vq_io_list, vi);
+		zio_vdev_free(vi);
+	}
+
+	list_destroy(&vq->vq_io_list);
+
 	mutex_destroy(&vq->vq_lock);
 }
 
@@ -152,6 +171,8 @@ vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio)
 static void
 vdev_queue_agg_io_done(zio_t *aio)
 {
+	vdev_queue_t *vq = &aio->io_vd->vdev_queue;
+	vdev_io_t *vi = aio->io_data;
 	zio_t *pio;
 
 	while ((pio = zio_walk_parents(aio)) != NULL)
@@ -159,7 +180,9 @@ vdev_queue_agg_io_done(zio_t *aio)
 			bcopy((char *)aio->io_data + (pio->io_offset -
 			    aio->io_offset), pio->io_data, pio->io_size);
 
-	zio_buf_free(aio->io_data, aio->io_size);
+	mutex_enter(&vq->vq_lock);
+	list_insert_tail(&vq->vq_io_list, vi);
+	mutex_exit(&vq->vq_lock);
 }
 
 /*
@@ -176,6 +199,7 @@ vdev_queue_io_to_issue(vdev_queue_t *vq, uint64_t pending_limit)
 {
 	zio_t *fio, *lio, *aio, *dio, *nio, *mio;
 	avl_tree_t *t;
+	vdev_io_t *vi;
 	int flags;
 	uint64_t maxspan = zfs_vdev_aggregation_limit;
 	uint64_t maxgap;
@@ -194,6 +218,12 @@ again:
 	flags = fio->io_flags & ZIO_FLAG_AGG_INHERIT;
 	maxgap = (t == &vq->vq_read_tree) ? zfs_vdev_read_gap_limit : 0;
 
+	vi = list_head(&vq->vq_io_list);
+	if (vi == NULL) {
+		vi = zio_vdev_alloc();
+		list_insert_head(&vq->vq_io_list, vi);
+	}
+
 	if (!(flags & ZIO_FLAG_DONT_AGGREGATE)) {
 		/*
 		 * We can aggregate I/Os that are sufficiently adjacent and of
@@ -283,9 +313,10 @@ again:
 	if (fio != lio) {
 		uint64_t size = IO_SPAN(fio, lio);
 		ASSERT(size <= zfs_vdev_aggregation_limit);
+		ASSERT(vi != NULL);
 
 		aio = zio_vdev_delegated_io(fio->io_vd, fio->io_offset,
-		    zio_buf_alloc(size), size, fio->io_type, ZIO_PRIORITY_AGG,
+		    vi, size, fio->io_type, ZIO_PRIORITY_AGG,
 		    flags | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE,
 		    vdev_queue_agg_io_done, NULL);
 
@@ -313,6 +344,7 @@ again:
 		} while (dio != lio);
 
 		avl_add(&vq->vq_pending_tree, aio);
+		list_remove(&vq->vq_io_list, vi);
 
 		return (aio);
 	}
diff --git a/module/zfs/zio.c b/module/zfs/zio.c
index 206ed9a93..fe2bdc867 100644
--- a/module/zfs/zio.c
+++ b/module/zfs/zio.c
@@ -73,6 +73,7 @@ char *zio_type_name[ZIO_TYPES] = {
  */
 kmem_cache_t *zio_cache;
 kmem_cache_t *zio_link_cache;
+kmem_cache_t *zio_vdev_cache;
 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
 int zio_bulk_flags = 0;
@@ -141,6 +142,8 @@ zio_init(void)
 	    zio_cons, zio_dest, NULL, NULL, NULL, KMC_KMEM);
 	zio_link_cache = kmem_cache_create("zio_link_cache",
 	    sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, KMC_KMEM);
+	zio_vdev_cache = kmem_cache_create("zio_vdev_cache", sizeof(vdev_io_t),
+	    PAGESIZE, NULL, NULL, NULL, NULL, NULL, KMC_VMEM);
 
 	/*
 	 * For small buffers, we want a cache for each multiple of
@@ -230,6 +233,7 @@ zio_fini(void)
 		zio_data_buf_cache[c] = NULL;
 	}
 
+	kmem_cache_destroy(zio_vdev_cache);
 	kmem_cache_destroy(zio_link_cache);
 	kmem_cache_destroy(zio_cache);
 
@@ -295,6 +299,24 @@ zio_data_buf_free(void *buf, size_t size)
 }
 
 /*
+ * Dedicated I/O buffers to ensure that memory fragmentation never prevents
+ * or significantly delays the issuing of a zio.   These buffers are used
+ * to aggregate I/O and could be used for raidz stripes.
+ */
+void *
+zio_vdev_alloc(void)
+{
+	return (kmem_cache_alloc(zio_vdev_cache, KM_PUSHPAGE));
+}
+
+void
+zio_vdev_free(void *buf)
+{
+	kmem_cache_free(zio_vdev_cache, buf);
+
+}
+
+/*
  * ==========================================================================
  * Push and pop I/O transform buffers
  * ==========================================================================