*/
/*
- * Copyright (c) 2012, 2017 by Delphix. All rights reserved.
+ * Copyright (c) 2012, 2018 by Delphix. All rights reserved.
*/
#include <sys/zfs_context.h>
uint32_t zfs_vdev_async_write_max_active = 10;
uint32_t zfs_vdev_scrub_min_active = 1;
uint32_t zfs_vdev_scrub_max_active = 2;
+uint32_t zfs_vdev_removal_min_active = 1;
+uint32_t zfs_vdev_removal_max_active = 2;
+uint32_t zfs_vdev_initializing_min_active = 1;
+uint32_t zfs_vdev_initializing_max_active = 1;
/*
* When the pool has less than zfs_vdev_async_write_active_min_dirty_percent
* we include spans of optional I/Os to aid aggregation at the disk even when
* they aren't able to help us aggregate at this level.
*/
-int zfs_vdev_aggregation_limit = SPA_OLD_MAXBLOCKSIZE;
+int zfs_vdev_aggregation_limit = 1 << 20;
int zfs_vdev_read_gap_limit = 32 << 10;
int zfs_vdev_write_gap_limit = 4 << 10;
int zfs_vdev_queue_depth_pct = 300;
#endif
+/*
+ * When performing allocations for a given metaslab, we want to make sure that
+ * there are enough IOs to aggregate together to improve throughput. We want to
+ * ensure that there are at least 128k worth of IOs that can be aggregated, and
+ * we assume that the average allocation size is 4k, so we need the queue depth
+ * to be 32 per allocator to get good aggregation of sequential writes.
+ */
+int zfs_vdev_def_queue_depth = 32;
+
int
vdev_queue_offset_compare(const void *x1, const void *x2)
return (zfs_vdev_async_write_min_active);
case ZIO_PRIORITY_SCRUB:
return (zfs_vdev_scrub_min_active);
+ case ZIO_PRIORITY_REMOVAL:
+ return (zfs_vdev_removal_min_active);
+ case ZIO_PRIORITY_INITIALIZING:
+ return (zfs_vdev_initializing_min_active);
default:
panic("invalid priority %u", p);
return (0);
return (vdev_queue_max_async_writes(spa));
case ZIO_PRIORITY_SCRUB:
return (zfs_vdev_scrub_max_active);
+ case ZIO_PRIORITY_REMOVAL:
+ return (zfs_vdev_removal_max_active);
+ case ZIO_PRIORITY_INITIALIZING:
+ return (zfs_vdev_initializing_max_active);
default:
panic("invalid priority %u", p);
return (0);
sizeof (zio_t), offsetof(struct zio, io_queue_node));
}
- vq->vq_lastoffset = 0;
+ vq->vq_last_offset = 0;
}
void
vdev_queue_fini(vdev_t *vd)
{
vdev_queue_t *vq = &vd->vdev_queue;
- zio_priority_t p;
- for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++)
+ for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++)
avl_destroy(vdev_queue_class_tree(vq, p));
avl_destroy(&vq->vq_active_tree);
avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_READ));
vdev_queue_io_add(vdev_queue_t *vq, zio_t *zio)
{
spa_t *spa = zio->io_spa;
- spa_stats_history_t *ssh = &spa->spa_stats.io_history;
+ spa_history_kstat_t *shk = &spa->spa_stats.io_history;
ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
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) {
- mutex_enter(&ssh->lock);
- kstat_waitq_enter(ssh->kstat->ks_data);
- mutex_exit(&ssh->lock);
+ if (shk->kstat != NULL) {
+ mutex_enter(&shk->lock);
+ kstat_waitq_enter(shk->kstat->ks_data);
+ mutex_exit(&shk->lock);
}
}
vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio)
{
spa_t *spa = zio->io_spa;
- spa_stats_history_t *ssh = &spa->spa_stats.io_history;
+ spa_history_kstat_t *shk = &spa->spa_stats.io_history;
ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
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) {
- mutex_enter(&ssh->lock);
- kstat_waitq_exit(ssh->kstat->ks_data);
- mutex_exit(&ssh->lock);
+ if (shk->kstat != NULL) {
+ mutex_enter(&shk->lock);
+ kstat_waitq_exit(shk->kstat->ks_data);
+ mutex_exit(&shk->lock);
}
}
vdev_queue_pending_add(vdev_queue_t *vq, zio_t *zio)
{
spa_t *spa = zio->io_spa;
- spa_stats_history_t *ssh = &spa->spa_stats.io_history;
+ spa_history_kstat_t *shk = &spa->spa_stats.io_history;
ASSERT(MUTEX_HELD(&vq->vq_lock));
ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
vq->vq_class[zio->io_priority].vqc_active++;
avl_add(&vq->vq_active_tree, zio);
- if (ssh->kstat != NULL) {
- mutex_enter(&ssh->lock);
- kstat_runq_enter(ssh->kstat->ks_data);
- mutex_exit(&ssh->lock);
+ if (shk->kstat != NULL) {
+ mutex_enter(&shk->lock);
+ kstat_runq_enter(shk->kstat->ks_data);
+ mutex_exit(&shk->lock);
}
}
vdev_queue_pending_remove(vdev_queue_t *vq, zio_t *zio)
{
spa_t *spa = zio->io_spa;
- spa_stats_history_t *ssh = &spa->spa_stats.io_history;
+ spa_history_kstat_t *shk = &spa->spa_stats.io_history;
ASSERT(MUTEX_HELD(&vq->vq_lock));
ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
vq->vq_class[zio->io_priority].vqc_active--;
avl_remove(&vq->vq_active_tree, zio);
- if (ssh->kstat != NULL) {
- kstat_io_t *ksio = ssh->kstat->ks_data;
+ if (shk->kstat != NULL) {
+ kstat_io_t *ksio = shk->kstat->ks_data;
- mutex_enter(&ssh->lock);
+ mutex_enter(&shk->lock);
kstat_runq_exit(ksio);
if (zio->io_type == ZIO_TYPE_READ) {
ksio->reads++;
ksio->writes++;
ksio->nwritten += zio->io_size;
}
- mutex_exit(&ssh->lock);
+ mutex_exit(&shk->lock);
}
}
vdev_queue_aggregate(vdev_queue_t *vq, zio_t *zio)
{
zio_t *first, *last, *aio, *dio, *mandatory, *nio;
+ zio_link_t *zl = NULL;
uint64_t maxgap = 0;
uint64_t size;
uint64_t limit;
+ int maxblocksize;
boolean_t stretch = B_FALSE;
avl_tree_t *t = vdev_queue_type_tree(vq, zio->io_type);
enum zio_flag flags = zio->io_flags & ZIO_FLAG_AGG_INHERIT;
abd_t *abd;
- limit = MAX(MIN(zfs_vdev_aggregation_limit,
- spa_maxblocksize(vq->vq_vdev->vdev_spa)), 0);
+ maxblocksize = spa_maxblocksize(vq->vq_vdev->vdev_spa);
+ limit = MAX(MIN(zfs_vdev_aggregation_limit, maxblocksize), 0);
if (zio->io_flags & ZIO_FLAG_DONT_AGGREGATE || limit == 0)
return (NULL);
while ((dio = AVL_PREV(t, first)) != NULL &&
(dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
IO_SPAN(dio, last) <= limit &&
- IO_GAP(dio, first) <= maxgap) {
+ IO_GAP(dio, first) <= maxgap &&
+ dio->io_type == zio->io_type) {
first = dio;
if (mandatory == NULL && !(first->io_flags & ZIO_FLAG_OPTIONAL))
mandatory = first;
(dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
(IO_SPAN(first, dio) <= limit ||
(dio->io_flags & ZIO_FLAG_OPTIONAL)) &&
- IO_GAP(last, dio) <= maxgap) {
+ IO_SPAN(first, dio) <= maxblocksize &&
+ IO_GAP(last, dio) <= maxgap &&
+ dio->io_type == zio->io_type) {
last = dio;
if (!(last->io_flags & ZIO_FLAG_OPTIONAL))
mandatory = last;
return (NULL);
size = IO_SPAN(first, last);
+ ASSERT3U(size, <=, maxblocksize);
abd = abd_alloc_for_io(size, B_TRUE);
if (abd == NULL)
zio_add_child(dio, aio);
vdev_queue_io_remove(vq, dio);
+ } while (dio != last);
+
+ /*
+ * We need to drop the vdev queue's lock to avoid a deadlock that we
+ * could encounter since this I/O will complete immediately.
+ */
+ mutex_exit(&vq->vq_lock);
+ while ((dio = zio_walk_parents(aio, &zl)) != NULL) {
zio_vdev_io_bypass(dio);
zio_execute(dio);
- } while (dio != last);
+ }
+ mutex_enter(&vq->vq_lock);
return (aio);
}
}
/*
- * For LBA-ordered queues (async / scrub), issue the i/o which follows
- * the most recently issued i/o in LBA (offset) order.
+ * For LBA-ordered queues (async / scrub / initializing), issue the
+ * i/o which follows the most recently issued i/o in LBA (offset) order.
*
* For FIFO queues (sync), issue the i/o with the lowest timestamp.
*/
tree = vdev_queue_class_tree(vq, p);
vq->vq_io_search.io_timestamp = 0;
- vq->vq_io_search.io_offset = vq->vq_last_offset + 1;
- VERIFY3P(avl_find(tree, &vq->vq_io_search,
- &idx), ==, NULL);
+ vq->vq_io_search.io_offset = vq->vq_last_offset - 1;
+ VERIFY3P(avl_find(tree, &vq->vq_io_search, &idx), ==, NULL);
zio = avl_nearest(tree, idx, AVL_AFTER);
if (zio == NULL)
zio = avl_first(tree);
}
vdev_queue_pending_add(vq, zio);
- vq->vq_last_offset = zio->io_offset;
+ vq->vq_last_offset = zio->io_offset + zio->io_size;
return (zio);
}
if (zio->io_type == ZIO_TYPE_READ) {
if (zio->io_priority != ZIO_PRIORITY_SYNC_READ &&
zio->io_priority != ZIO_PRIORITY_ASYNC_READ &&
- zio->io_priority != ZIO_PRIORITY_SCRUB)
+ zio->io_priority != ZIO_PRIORITY_SCRUB &&
+ zio->io_priority != ZIO_PRIORITY_REMOVAL &&
+ zio->io_priority != ZIO_PRIORITY_INITIALIZING)
zio->io_priority = ZIO_PRIORITY_ASYNC_READ;
} else {
ASSERT(zio->io_type == ZIO_TYPE_WRITE);
if (zio->io_priority != ZIO_PRIORITY_SYNC_WRITE &&
- zio->io_priority != ZIO_PRIORITY_ASYNC_WRITE)
+ zio->io_priority != ZIO_PRIORITY_ASYNC_WRITE &&
+ zio->io_priority != ZIO_PRIORITY_REMOVAL &&
+ zio->io_priority != ZIO_PRIORITY_INITIALIZING)
zio->io_priority = ZIO_PRIORITY_ASYNC_WRITE;
}
mutex_exit(&vq->vq_lock);
}
+void
+vdev_queue_change_io_priority(zio_t *zio, zio_priority_t priority)
+{
+ vdev_queue_t *vq = &zio->io_vd->vdev_queue;
+ avl_tree_t *tree;
+
+ /*
+ * ZIO_PRIORITY_NOW is used by the vdev cache code and the aggregate zio
+ * code to issue IOs without adding them to the vdev queue. In this
+ * case, the zio is already going to be issued as quickly as possible
+ * and so it doesn't need any reprioitization to help.
+ */
+ if (zio->io_priority == ZIO_PRIORITY_NOW)
+ return;
+
+ ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
+ ASSERT3U(priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
+
+ if (zio->io_type == ZIO_TYPE_READ) {
+ if (priority != ZIO_PRIORITY_SYNC_READ &&
+ priority != ZIO_PRIORITY_ASYNC_READ &&
+ priority != ZIO_PRIORITY_SCRUB)
+ priority = ZIO_PRIORITY_ASYNC_READ;
+ } else {
+ ASSERT(zio->io_type == ZIO_TYPE_WRITE);
+ if (priority != ZIO_PRIORITY_SYNC_WRITE &&
+ priority != ZIO_PRIORITY_ASYNC_WRITE)
+ priority = ZIO_PRIORITY_ASYNC_WRITE;
+ }
+
+ mutex_enter(&vq->vq_lock);
+
+ /*
+ * If the zio is in none of the queues we can simply change
+ * the priority. If the zio is waiting to be submitted we must
+ * remove it from the queue and re-insert it with the new priority.
+ * Otherwise, the zio is currently active and we cannot change its
+ * priority.
+ */
+ tree = vdev_queue_class_tree(vq, zio->io_priority);
+ if (avl_find(tree, zio, NULL) == zio) {
+ avl_remove(vdev_queue_class_tree(vq, zio->io_priority), zio);
+ zio->io_priority = priority;
+ avl_add(vdev_queue_class_tree(vq, zio->io_priority), zio);
+ } else if (avl_find(&vq->vq_active_tree, zio, NULL) != zio) {
+ zio->io_priority = priority;
+ }
+
+ mutex_exit(&vq->vq_lock);
+}
+
/*
- * As these three methods are only used for load calculations we're not
+ * As these two methods are only used for load calculations we're not
* concerned if we get an incorrect value on 32bit platforms due to lack of
* vq_lock mutex use here, instead we prefer to keep it lock free for
* performance.
}
uint64_t
-vdev_queue_lastoffset(vdev_t *vd)
+vdev_queue_last_offset(vdev_t *vd)
{
- return (vd->vdev_queue.vq_lastoffset);
+ return (vd->vdev_queue.vq_last_offset);
}
-void
-vdev_queue_register_lastoffset(vdev_t *vd, zio_t *zio)
-{
- vd->vdev_queue.vq_lastoffset = zio->io_offset + zio->io_size;
-}
-
-#if defined(_KERNEL) && defined(HAVE_SPL)
+#if defined(_KERNEL)
module_param(zfs_vdev_aggregation_limit, int, 0644);
MODULE_PARM_DESC(zfs_vdev_aggregation_limit, "Max vdev I/O aggregation size");
MODULE_PARM_DESC(zfs_vdev_async_write_min_active,
"Min active async write I/Os per vdev");
+module_param(zfs_vdev_initializing_max_active, int, 0644);
+MODULE_PARM_DESC(zfs_vdev_initializing_max_active,
+ "Max active initializing I/Os per vdev");
+
+module_param(zfs_vdev_initializing_min_active, int, 0644);
+MODULE_PARM_DESC(zfs_vdev_initializing_min_active,
+ "Min active initializing I/Os per vdev");
+
+module_param(zfs_vdev_removal_max_active, int, 0644);
+MODULE_PARM_DESC(zfs_vdev_removal_max_active,
+ "Max active removal I/Os per vdev");
+
+module_param(zfs_vdev_removal_min_active, int, 0644);
+MODULE_PARM_DESC(zfs_vdev_removal_min_active,
+ "Min active removal I/Os per vdev");
+
module_param(zfs_vdev_scrub_max_active, int, 0644);
-MODULE_PARM_DESC(zfs_vdev_scrub_max_active, "Max active scrub I/Os per vdev");
+MODULE_PARM_DESC(zfs_vdev_scrub_max_active,
+ "Max active scrub I/Os per vdev");
module_param(zfs_vdev_scrub_min_active, int, 0644);
-MODULE_PARM_DESC(zfs_vdev_scrub_min_active, "Min active scrub I/Os per vdev");
+MODULE_PARM_DESC(zfs_vdev_scrub_min_active,
+ "Min active scrub I/Os per vdev");
module_param(zfs_vdev_sync_read_max_active, int, 0644);
MODULE_PARM_DESC(zfs_vdev_sync_read_max_active,
projects / mirror_zfs.git / blobdiff