c/mirror_zfs
1
0
Fork 0
mirror of https://git.proxmox.com/git/mirror_zfs.git synced 2024-12-25 18:59:33 +03:00
Code Issues Packages Projects Releases Wiki Activity

5313 Allow I/Os to be aggregated across ZIO priority classes

Browse Source 
Reviewed by: Andriy Gapon <avg@FreeBSD.org>
Reviewed by: Will Andrews <willa@SpectraLogic.com>
Reviewed by: Matt Ahrens <mahrens@delphix.com>
Reviewed by: George Wilson <george@delphix.com>
Approved by: Robert Mustacchi <rm@joyent.com>

References:
  https://www.illumos.org/issues/5313
  https://github.com/illumos/illumos-gate/commit/fe319232

Ported-by: DHE <git@dehacked.net>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #3280
This commit is contained in:
Justin T. Gibbs 2015-04-11 14:51:06 -04:00 committed by Brian Behlendorf
parent 0bf8501ae1
commit ec8501ee12
3 changed files with 51 additions and 31 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
include/sys/vdev_impl.h
View File

@ -113,6 +113,8 @@ struct vdev_queue {
vdev_t *vq_vdev; vdev_t *vq_vdev;
vdev_queue_class_t vq_class[ZIO_PRIORITY_NUM_QUEUEABLE]; vdev_queue_class_t vq_class[ZIO_PRIORITY_NUM_QUEUEABLE];
avl_tree_t vq_active_tree; avl_tree_t vq_active_tree;
avl_tree_t vq_read_offset_tree;
avl_tree_t vq_write_offset_tree;
uint64_t vq_last_offset; uint64_t vq_last_offset;
hrtime_t vq_io_complete_ts; /* time last i/o completed */ hrtime_t vq_io_complete_ts; /* time last i/o completed */
hrtime_t vq_io_delta_ts; hrtime_t vq_io_delta_ts;

1
include/sys/zio.h
View File

@ -427,6 +427,7 @@ struct zio {
hrtime_t io_delta; /* vdev queue service delta */ hrtime_t io_delta; /* vdev queue service delta */
uint64_t io_delay; /* vdev disk service delta (ticks) */ uint64_t io_delay; /* vdev disk service delta (ticks) */
avl_node_t io_queue_node; avl_node_t io_queue_node;
avl_node_t io_offset_node;
/* Internal pipeline state */ /* Internal pipeline state */
enum zio_flag io_flags; enum zio_flag io_flags;

79
module/zfs/vdev_queue.c
View File

@ -190,6 +190,22 @@ vdev_queue_offset_compare(const void *x1, const void *x2)
return (0); return (0);
} }
static inline avl_tree_t *
vdev_queue_class_tree(vdev_queue_t *vq, zio_priority_t p)
{
return (&vq->vq_class[p].vqc_queued_tree);
}
static inline avl_tree_t *
vdev_queue_type_tree(vdev_queue_t *vq, zio_type_t t)
{
ASSERT(t == ZIO_TYPE_READ || t == ZIO_TYPE_WRITE);
if (t == ZIO_TYPE_READ)
return (&vq->vq_read_offset_tree);
else
return (&vq->vq_write_offset_tree);
}
int int
vdev_queue_timestamp_compare(const void *x1, const void *x2) vdev_queue_timestamp_compare(const void *x1, const void *x2)
{ {
@ -303,7 +319,7 @@ vdev_queue_class_to_issue(vdev_queue_t *vq)
/* find a queue that has not reached its minimum # outstanding i/os */ /* find a queue that has not reached its minimum # outstanding i/os */
for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
if (avl_numnodes(&vq->vq_class[p].vqc_queued_tree) > 0 && if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 &&
vq->vq_class[p].vqc_active < vq->vq_class[p].vqc_active <
vdev_queue_class_min_active(p)) vdev_queue_class_min_active(p))
return (p); return (p);
@ -314,7 +330,7 @@ vdev_queue_class_to_issue(vdev_queue_t *vq)
* maximum # outstanding i/os. * maximum # outstanding i/os.
*/ */
for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
if (avl_numnodes(&vq->vq_class[p].vqc_queued_tree) > 0 && if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 &&
vq->vq_class[p].vqc_active < vq->vq_class[p].vqc_active <
vdev_queue_class_max_active(spa, p)) vdev_queue_class_max_active(spa, p))
return (p); return (p);
@ -335,20 +351,27 @@ vdev_queue_init(vdev_t *vd)
avl_create(&vq->vq_active_tree, vdev_queue_offset_compare, avl_create(&vq->vq_active_tree, vdev_queue_offset_compare,
sizeof (zio_t), offsetof(struct zio, io_queue_node)); sizeof (zio_t), offsetof(struct zio, io_queue_node));
avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_READ),
vdev_queue_offset_compare, sizeof (zio_t),
offsetof(struct zio, io_offset_node));
avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE),
vdev_queue_offset_compare, sizeof (zio_t),
offsetof(struct zio, io_offset_node));
for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
int (*compfn) (const void *, const void *);
/* /*
* The synchronous i/o queues are FIFO rather than LBA ordered. * The synchronous i/o queues are dispatched in FIFO rather
* This provides more consistent latency for these i/os, and * than LBA order. This provides more consistent latency for
* they tend to not be tightly clustered anyway so there is * these i/os.
* little to no throughput loss.
*/ */
boolean_t fifo = (p == ZIO_PRIORITY_SYNC_READ || if (p == ZIO_PRIORITY_SYNC_READ || p == ZIO_PRIORITY_SYNC_WRITE)
p == ZIO_PRIORITY_SYNC_WRITE); compfn = vdev_queue_timestamp_compare;
avl_create(&vq->vq_class[p].vqc_queued_tree, else
fifo ? vdev_queue_timestamp_compare : compfn = vdev_queue_offset_compare;
vdev_queue_offset_compare, avl_create(vdev_queue_class_tree(vq, p), compfn,
sizeof (zio_t), offsetof(struct zio, io_queue_node)); sizeof (zio_t), offsetof(struct zio, io_queue_node));
} }
} }
@ -359,8 +382,10 @@ vdev_queue_fini(vdev_t *vd)
zio_priority_t p; zio_priority_t p;
for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++)
avl_destroy(&vq->vq_class[p].vqc_queued_tree); avl_destroy(vdev_queue_class_tree(vq, p));
avl_destroy(&vq->vq_active_tree); avl_destroy(&vq->vq_active_tree);
avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_READ));
avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE));
mutex_destroy(&vq->vq_lock); mutex_destroy(&vq->vq_lock);
} }
@ -372,7 +397,8 @@ vdev_queue_io_add(vdev_queue_t *vq, zio_t *zio)
spa_stats_history_t *ssh = &spa->spa_stats.io_history; spa_stats_history_t *ssh = &spa->spa_stats.io_history;
ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
avl_add(&vq->vq_class[zio->io_priority].vqc_queued_tree, zio); avl_add(vdev_queue_class_tree(vq, zio->io_priority), zio);
avl_add(vdev_queue_type_tree(vq, zio->io_type), zio);
if (ssh->kstat != NULL) { if (ssh->kstat != NULL) {
mutex_enter(&ssh->lock); mutex_enter(&ssh->lock);
@ -388,7 +414,8 @@ vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio)
spa_stats_history_t *ssh = &spa->spa_stats.io_history; spa_stats_history_t *ssh = &spa->spa_stats.io_history;
ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
avl_remove(&vq->vq_class[zio->io_priority].vqc_queued_tree, zio); avl_remove(vdev_queue_class_tree(vq, zio->io_priority), zio);
avl_remove(vdev_queue_type_tree(vq, zio->io_type), zio);
if (ssh->kstat != NULL) { if (ssh->kstat != NULL) {
mutex_enter(&ssh->lock); mutex_enter(&ssh->lock);
@ -472,8 +499,7 @@ vdev_queue_aggregate(vdev_queue_t *vq, zio_t *zio)
uint64_t maxgap = 0; uint64_t maxgap = 0;
uint64_t size; uint64_t size;
boolean_t stretch = B_FALSE; boolean_t stretch = B_FALSE;
vdev_queue_class_t *vqc = &vq->vq_class[zio->io_priority]; avl_tree_t *t = vdev_queue_type_tree(vq, zio->io_type);
avl_tree_t *t = &vqc->vqc_queued_tree;
enum zio_flag flags = zio->io_flags & ZIO_FLAG_AGG_INHERIT; enum zio_flag flags = zio->io_flags & ZIO_FLAG_AGG_INHERIT;
if (zio->io_flags & ZIO_FLAG_DONT_AGGREGATE) if (zio->io_flags & ZIO_FLAG_DONT_AGGREGATE)
@ -486,15 +512,6 @@ vdev_queue_aggregate(vdev_queue_t *vq, zio_t *zio)
zfs_vdev_aggregation_limit = zfs_vdev_aggregation_limit =
MIN(zfs_vdev_aggregation_limit, SPA_MAXBLOCKSIZE); MIN(zfs_vdev_aggregation_limit, SPA_MAXBLOCKSIZE);
/*
* The synchronous i/o queues are not sorted by LBA, so we can't
* find adjacent i/os. These i/os tend to not be tightly clustered,
* or too large to aggregate, so this has little impact on performance.
*/
if (zio->io_priority == ZIO_PRIORITY_SYNC_READ ||
zio->io_priority == ZIO_PRIORITY_SYNC_WRITE)
return (NULL);
first = last = zio; first = last = zio;
if (zio->io_type == ZIO_TYPE_READ) if (zio->io_type == ZIO_TYPE_READ)
@ -627,7 +644,7 @@ vdev_queue_io_to_issue(vdev_queue_t *vq)
zio_t *zio, *aio; zio_t *zio, *aio;
zio_priority_t p; zio_priority_t p;
avl_index_t idx; avl_index_t idx;
vdev_queue_class_t *vqc; avl_tree_t *tree;
again: again:
ASSERT(MUTEX_HELD(&vq->vq_lock)); ASSERT(MUTEX_HELD(&vq->vq_lock));
@ -645,14 +662,14 @@ again:
* *
* For FIFO queues (sync), issue the i/o with the lowest timestamp. * For FIFO queues (sync), issue the i/o with the lowest timestamp.
*/ */
vqc = &vq->vq_class[p]; tree = vdev_queue_class_tree(vq, p);
vq->vq_io_search.io_timestamp = 0; vq->vq_io_search.io_timestamp = 0;
vq->vq_io_search.io_offset = vq->vq_last_offset + 1; vq->vq_io_search.io_offset = vq->vq_last_offset + 1;
VERIFY3P(avl_find(&vqc->vqc_queued_tree, &vq->vq_io_search, VERIFY3P(avl_find(tree, &vq->vq_io_search,
&idx), ==, NULL); &idx), ==, NULL);
zio = avl_nearest(&vqc->vqc_queued_tree, idx, AVL_AFTER); zio = avl_nearest(tree, idx, AVL_AFTER);
if (zio == NULL) if (zio == NULL)
zio = avl_first(&vqc->vqc_queued_tree); zio = avl_first(tree);
ASSERT3U(zio->io_priority, ==, p); ASSERT3U(zio->io_priority, ==, p);
aio = vdev_queue_aggregate(vq, zio); aio = vdev_queue_aggregate(vq, zio);