static mirror_map_t *
vdev_mirror_map_alloc(zio_t *zio)
{
mirror_map_t *mm = NULL;
mirror_child_t *mc;
vdev_t *vd = zio->io_vd;
int c, d;
if (vd == NULL) {
dva_t *dva = zio->io_bp->blk_dva;
spa_t *spa = zio->io_spa;
c = BP_GET_NDVAS(zio->io_bp);
mm = kmem_zalloc(offsetof(mirror_map_t, mm_child[c]), KM_SLEEP);
mm->mm_children = c;
mm->mm_replacing = B_FALSE;
mm->mm_preferred = spa_get_random(c);
mm->mm_root = B_TRUE;
/*
* Check the other, lower-index DVAs to see if they're on
* the same vdev as the child we picked. If they are, use
* them since they are likely to have been allocated from
* the primary metaslab in use at the time, and hence are
* more likely to have locality with single-copy data.
*/
for (c = mm->mm_preferred, d = c - 1; d >= 0; d--) {
if (DVA_GET_VDEV(&dva[d]) == DVA_GET_VDEV(&dva[c]))
mm->mm_preferred = d;
}
for (c = 0; c < mm->mm_children; c++) {
mc = &mm->mm_child[c];
mc->mc_vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[c]));
mc->mc_offset = DVA_GET_OFFSET(&dva[c]);
}
} else {
c = vd->vdev_children;
mm = kmem_zalloc(offsetof(mirror_map_t, mm_child[c]), KM_SLEEP);
mm->mm_children = c;
mm->mm_replacing = (vd->vdev_ops == &vdev_replacing_ops ||
vd->vdev_ops == &vdev_spare_ops);
mm->mm_preferred = mm->mm_replacing ? 0 :
(zio->io_offset >> vdev_mirror_shift) % c;
mm->mm_root = B_FALSE;
for (c = 0; c < mm->mm_children; c++) {
mc = &mm->mm_child[c];
mc->mc_vd = vd->vdev_child[c];
mc->mc_offset = zio->io_offset;
}
}
zio->io_vsd = mm;
zio->io_vsd_ops = &vdev_mirror_vsd_ops;
return (mm);
}
static mirror_map_t *
vdev_mirror_map_init(zio_t *zio)
{
mirror_map_t *mm = NULL;
mirror_child_t *mc;
vdev_t *vd = zio->io_vd;
int c;
if (vd == NULL) {
dva_t *dva = zio->io_bp->blk_dva;
spa_t *spa = zio->io_spa;
mm = vdev_mirror_map_alloc(BP_GET_NDVAS(zio->io_bp), B_FALSE,
B_TRUE);
for (c = 0; c < mm->mm_children; c++) {
mc = &mm->mm_child[c];
mc->mc_vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[c]));
mc->mc_offset = DVA_GET_OFFSET(&dva[c]);
}
} else {
mm = vdev_mirror_map_alloc(vd->vdev_children,
(vd->vdev_ops == &vdev_replacing_ops ||
vd->vdev_ops == &vdev_spare_ops), B_FALSE);
for (c = 0; c < mm->mm_children; c++) {
mc = &mm->mm_child[c];
mc->mc_vd = vd->vdev_child[c];
mc->mc_offset = zio->io_offset;
}
}
zio->io_vsd = mm;
zio->io_vsd_ops = &vdev_mirror_vsd_ops;
return (mm);
}
/*
* This sync task completes (finishes) a condense, deleting the old
* mapping and replacing it with the new one.
*/
static void
spa_condense_indirect_complete_sync(void *arg, dmu_tx_t *tx)
{
spa_condensing_indirect_t *sci = arg;
spa_t *spa = dmu_tx_pool(tx)->dp_spa;
spa_condensing_indirect_phys_t *scip =
&spa->spa_condensing_indirect_phys;
vdev_t *vd = vdev_lookup_top(spa, scip->scip_vdev);
vdev_indirect_config_t *vic = &vd->vdev_indirect_config;
objset_t *mos = spa->spa_meta_objset;
vdev_indirect_mapping_t *old_mapping = vd->vdev_indirect_mapping;
uint64_t old_count = vdev_indirect_mapping_num_entries(old_mapping);
uint64_t new_count =
vdev_indirect_mapping_num_entries(sci->sci_new_mapping);
ASSERT(dmu_tx_is_syncing(tx));
ASSERT3P(vd->vdev_ops, ==, &vdev_indirect_ops);
ASSERT3P(sci, ==, spa->spa_condensing_indirect);
for (int i = 0; i < TXG_SIZE; i++) {
ASSERT(list_is_empty(&sci->sci_new_mapping_entries[i]));
}
ASSERT(vic->vic_mapping_object != 0);
ASSERT3U(vd->vdev_id, ==, scip->scip_vdev);
ASSERT(scip->scip_next_mapping_object != 0);
ASSERT(scip->scip_prev_obsolete_sm_object != 0);
/*
* Reset vdev_indirect_mapping to refer to the new object.
*/
rw_enter(&vd->vdev_indirect_rwlock, RW_WRITER);
vdev_indirect_mapping_close(vd->vdev_indirect_mapping);
vd->vdev_indirect_mapping = sci->sci_new_mapping;
rw_exit(&vd->vdev_indirect_rwlock);
sci->sci_new_mapping = NULL;
vdev_indirect_mapping_free(mos, vic->vic_mapping_object, tx);
vic->vic_mapping_object = scip->scip_next_mapping_object;
scip->scip_next_mapping_object = 0;
space_map_free_obj(mos, scip->scip_prev_obsolete_sm_object, tx);
spa_feature_decr(spa, SPA_FEATURE_OBSOLETE_COUNTS, tx);
scip->scip_prev_obsolete_sm_object = 0;
scip->scip_vdev = 0;
VERIFY0(zap_remove(mos, DMU_POOL_DIRECTORY_OBJECT,
DMU_POOL_CONDENSING_INDIRECT, tx));
spa_condensing_indirect_destroy(spa->spa_condensing_indirect);
spa->spa_condensing_indirect = NULL;
zfs_dbgmsg("finished condense of vdev %llu in txg %llu: "
"new mapping object %llu has %llu entries "
"(was %llu entries)",
vd->vdev_id, dmu_tx_get_txg(tx), vic->vic_mapping_object,
new_count, old_count);
vdev_config_dirty(spa->spa_root_vdev);
}
/*
* Mark the DVA vdev_id:offset:size as being obsolete in the given tx. This
* wrapper is provided because the DMU does not know about vdev_t's and
* cannot directly call vdev_indirect_mark_obsolete.
*/
void
spa_vdev_indirect_mark_obsolete(spa_t *spa, uint64_t vdev_id, uint64_t offset,
uint64_t size, dmu_tx_t *tx)
{
vdev_t *vd = vdev_lookup_top(spa, vdev_id);
ASSERT(dmu_tx_is_syncing(tx));
/* The DMU can only remap indirect vdevs. */
ASSERT3P(vd->vdev_ops, ==, &vdev_indirect_ops);
vdev_indirect_mark_obsolete(vd, offset, size);
}
void
zil_flush_vdevs(zilog_t *zilog)
{
spa_t *spa = zilog->zl_spa;
avl_tree_t *t = &zilog->zl_vdev_tree;
void *cookie = NULL;
zil_vdev_node_t *zv;
zio_t *zio;
ASSERT(zilog->zl_writer);
/*
* We don't need zl_vdev_lock here because we're the zl_writer,
* and all zl_get_data() callbacks are done.
*/
if (avl_numnodes(t) == 0)
return;
spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
if (vd != NULL)
zio_flush(zio, vd);
kmem_free(zv, sizeof (*zv));
}
/*
* Wait for all the flushes to complete. Not all devices actually
* support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails.
*/
(void) zio_wait(zio);
spa_config_exit(spa, SCL_STATE, FTAG);
}
/*
* Avoid inlining the function to keep vdev_mirror_io_start(), which
* is this functions only caller, as small as possible on the stack.
*/
noinline static mirror_map_t *
vdev_mirror_map_alloc(zio_t *zio)
{
mirror_map_t *mm = NULL;
mirror_child_t *mc;
vdev_t *vd = zio->io_vd;
int c, d;
if (vd == NULL) {
dva_t *dva = zio->io_bp->blk_dva;
spa_t *spa = zio->io_spa;
c = BP_GET_NDVAS(zio->io_bp);
mm = kmem_zalloc(offsetof(mirror_map_t, mm_child[c]),
KM_PUSHPAGE);
mm->mm_children = c;
mm->mm_replacing = B_FALSE;
mm->mm_preferred = spa_get_random(c);
mm->mm_root = B_TRUE;
/*
* Check the other, lower-index DVAs to see if they're on
* the same vdev as the child we picked. If they are, use
* them since they are likely to have been allocated from
* the primary metaslab in use at the time, and hence are
* more likely to have locality with single-copy data.
*/
for (c = mm->mm_preferred, d = c - 1; d >= 0; d--) {
if (DVA_GET_VDEV(&dva[d]) == DVA_GET_VDEV(&dva[c]))
mm->mm_preferred = d;
}
for (c = 0; c < mm->mm_children; c++) {
mc = &mm->mm_child[c];
mc->mc_vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[c]));
mc->mc_offset = DVA_GET_OFFSET(&dva[c]);
}
} else {
int lowest_pending = INT_MAX;
int lowest_nr = 1;
c = vd->vdev_children;
mm = kmem_zalloc(offsetof(mirror_map_t, mm_child[c]),
KM_PUSHPAGE);
mm->mm_children = c;
mm->mm_replacing = (vd->vdev_ops == &vdev_replacing_ops ||
vd->vdev_ops == &vdev_spare_ops);
mm->mm_preferred = 0;
mm->mm_root = B_FALSE;
for (c = 0; c < mm->mm_children; c++) {
mc = &mm->mm_child[c];
mc->mc_vd = vd->vdev_child[c];
mc->mc_offset = zio->io_offset;
if (mm->mm_replacing)
continue;
if (!vdev_readable(mc->mc_vd)) {
mc->mc_error = SET_ERROR(ENXIO);
mc->mc_tried = 1;
mc->mc_skipped = 1;
mc->mc_pending = INT_MAX;
continue;
}
mc->mc_pending = vdev_mirror_pending(mc->mc_vd);
if (mc->mc_pending < lowest_pending) {
lowest_pending = mc->mc_pending;
lowest_nr = 1;
} else if (mc->mc_pending == lowest_pending) {
lowest_nr++;
}
}
d = gethrtime() / (NSEC_PER_USEC * zfs_vdev_mirror_switch_us);
d = (d % lowest_nr) + 1;
for (c = 0; c < mm->mm_children; c++) {
mc = &mm->mm_child[c];
if (mm->mm_child[c].mc_pending == lowest_pending) {
if (--d == 0) {
mm->mm_preferred = c;
break;
}
}
}
}
zio->io_vsd = mm;
zio->io_vsd_ops = &vdev_mirror_vsd_ops;
return (mm);
}
static mirror_map_t *
vdev_mirror_map_init(zio_t *zio)
{
mirror_map_t *mm = NULL;
mirror_child_t *mc;
vdev_t *vd = zio->io_vd;
int c;
if (vd == NULL) {
dva_t *dva = zio->io_bp->blk_dva;
spa_t *spa = zio->io_spa;
mm = vdev_mirror_map_alloc(BP_GET_NDVAS(zio->io_bp), B_FALSE,
B_TRUE);
for (c = 0; c < mm->mm_children; c++) {
mc = &mm->mm_child[c];
mc->mc_vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[c]));
mc->mc_offset = DVA_GET_OFFSET(&dva[c]);
}
} else {
/*
* If we are resilvering, then we should handle scrub reads
* differently; we shouldn't issue them to the resilvering
* device because it might not have those blocks.
*
* We are resilvering iff:
* 1) We are a replacing vdev (ie our name is "replacing-1" or
* "spare-1" or something like that), and
* 2) The pool is currently being resilvered.
*
* We cannot simply check vd->vdev_resilver_txg, because it's
* not set in this path.
*
* Nor can we just check our vdev_ops; there are cases (such as
* when a user types "zpool replace pool odev spare_dev" and
* spare_dev is in the spare list, or when a spare device is
* automatically used to replace a DEGRADED device) when
* resilvering is complete but both the original vdev and the
* spare vdev remain in the pool. That behavior is intentional.
* It helps implement the policy that a spare should be
* automatically removed from the pool after the user replaces
* the device that originally failed.
*
* If a spa load is in progress, then spa_dsl_pool may be
* uninitialized. But we shouldn't be resilvering during a spa
* load anyway.
*/
boolean_t replacing = (vd->vdev_ops == &vdev_replacing_ops ||
vd->vdev_ops == &vdev_spare_ops) &&
spa_load_state(vd->vdev_spa) == SPA_LOAD_NONE &&
dsl_scan_resilvering(vd->vdev_spa->spa_dsl_pool);
mm = vdev_mirror_map_alloc(vd->vdev_children, replacing,
B_FALSE);
for (c = 0; c < mm->mm_children; c++) {
mc = &mm->mm_child[c];
mc->mc_vd = vd->vdev_child[c];
mc->mc_offset = zio->io_offset;
}
}
zio->io_vsd = mm;
zio->io_vsd_ops = &vdev_mirror_vsd_ops;
return (mm);
}
/* ARGSUSED */
static int
spa_condense_indirect_thread(void *arg, zthr_t *zthr)
{
spa_t *spa = arg;
vdev_t *vd;
ASSERT3P(spa->spa_condensing_indirect, !=, NULL);
spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
vd = vdev_lookup_top(spa, spa->spa_condensing_indirect_phys.scip_vdev);
ASSERT3P(vd, !=, NULL);
spa_config_exit(spa, SCL_VDEV, FTAG);
spa_condensing_indirect_t *sci = spa->spa_condensing_indirect;
spa_condensing_indirect_phys_t *scip =
&spa->spa_condensing_indirect_phys;
uint32_t *counts;
uint64_t start_index;
vdev_indirect_mapping_t *old_mapping = vd->vdev_indirect_mapping;
space_map_t *prev_obsolete_sm = NULL;
ASSERT3U(vd->vdev_id, ==, scip->scip_vdev);
ASSERT(scip->scip_next_mapping_object != 0);
ASSERT(scip->scip_prev_obsolete_sm_object != 0);
ASSERT3P(vd->vdev_ops, ==, &vdev_indirect_ops);
for (int i = 0; i < TXG_SIZE; i++) {
/*
* The list must start out empty in order for the
* _commit_sync() sync task to be properly registered
* on the first call to _commit_entry(); so it's wise
* to double check and ensure we actually are starting
* with empty lists.
*/
ASSERT(list_is_empty(&sci->sci_new_mapping_entries[i]));
}
VERIFY0(space_map_open(&prev_obsolete_sm, spa->spa_meta_objset,
scip->scip_prev_obsolete_sm_object, 0, vd->vdev_asize, 0));
space_map_update(prev_obsolete_sm);
counts = vdev_indirect_mapping_load_obsolete_counts(old_mapping);
if (prev_obsolete_sm != NULL) {
vdev_indirect_mapping_load_obsolete_spacemap(old_mapping,
counts, prev_obsolete_sm);
}
space_map_close(prev_obsolete_sm);
/*
* Generate new mapping. Determine what index to continue from
* based on the max offset that we've already written in the
* new mapping.
*/
uint64_t max_offset =
vdev_indirect_mapping_max_offset(sci->sci_new_mapping);
if (max_offset == 0) {
/* We haven't written anything to the new mapping yet. */
start_index = 0;
} else {
/*
* Pick up from where we left off. _entry_for_offset()
* returns a pointer into the vim_entries array. If
* max_offset is greater than any of the mappings
* contained in the table NULL will be returned and
* that indicates we've exhausted our iteration of the
* old_mapping.
*/
vdev_indirect_mapping_entry_phys_t *entry =
vdev_indirect_mapping_entry_for_offset_or_next(old_mapping,
max_offset);
if (entry == NULL) {
/*
* We've already written the whole new mapping.
* This special value will cause us to skip the
* generate_new_mapping step and just do the sync
* task to complete the condense.
*/
start_index = UINT64_MAX;
} else {
start_index = entry - old_mapping->vim_entries;
ASSERT3U(start_index, <,
vdev_indirect_mapping_num_entries(old_mapping));
}
}
spa_condense_indirect_generate_new_mapping(vd, counts,
start_index, zthr);
vdev_indirect_mapping_free_obsolete_counts(old_mapping, counts);
/*
* If the zthr has received a cancellation signal while running
* in generate_new_mapping() or at any point after that, then bail
* early. We don't want to complete the condense if the spa is
* shutting down.
*/
if (zthr_iscancelled(zthr))
return (0);
VERIFY0(dsl_sync_task(spa_name(spa), NULL,
spa_condense_indirect_complete_sync, sci, 0,
ZFS_SPACE_CHECK_EXTRA_RESERVED));
return (0);
thread_exit();
}
static mirror_map_t *
vdev_mirror_map_alloc(zio_t *zio)
{
mirror_map_t *mm = NULL;
mirror_child_t *mc;
vdev_t *vd = zio->io_vd;
int c, d;
if (vd == NULL) {
dva_t *dva = zio->io_bp->blk_dva;
spa_t *spa = zio->io_spa;
c = BP_GET_NDVAS(zio->io_bp);
mm = kmem_zalloc(offsetof(mirror_map_t, mm_child[c]), KM_SLEEP);
mm->mm_children = c;
mm->mm_resilvering = B_FALSE;
mm->mm_preferred = spa_get_random(c);
mm->mm_root = B_TRUE;
/*
* Check the other, lower-index DVAs to see if they're on
* the same vdev as the child we picked. If they are, use
* them since they are likely to have been allocated from
* the primary metaslab in use at the time, and hence are
* more likely to have locality with single-copy data.
*/
for (c = mm->mm_preferred, d = c - 1; d >= 0; d--) {
if (DVA_GET_VDEV(&dva[d]) == DVA_GET_VDEV(&dva[c]))
mm->mm_preferred = d;
}
for (c = 0; c < mm->mm_children; c++) {
mc = &mm->mm_child[c];
mc->mc_vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[c]));
mc->mc_offset = DVA_GET_OFFSET(&dva[c]);
}
} else {
int replacing;
c = vd->vdev_children;
mm = kmem_zalloc(offsetof(mirror_map_t, mm_child[c]), KM_SLEEP);
mm->mm_children = c;
/*
* If we are resilvering, then we should handle scrub reads
* differently; we shouldn't issue them to the resilvering
* device because it might not have those blocks.
*
* We are resilvering iff:
* 1) We are a replacing vdev (ie our name is "replacing-1" or
* "spare-1" or something like that), and
* 2) The pool is currently being resilvered.
*
* We cannot simply check vd->vdev_resilver_txg, because it's
* not set in this path.
*
* Nor can we just check our vdev_ops; there are cases (such as
* when a user types "zpool replace pool odev spare_dev" and
* spare_dev is in the spare list, or when a spare device is
* automatically used to replace a DEGRADED device) when
* resilvering is complete but both the original vdev and the
* spare vdev remain in the pool. That behavior is intentional.
* It helps implement the policy that a spare should be
* automatically removed from the pool after the user replaces
* the device that originally failed.
*/
replacing = (vd->vdev_ops == &vdev_replacing_ops ||
vd->vdev_ops == &vdev_spare_ops);
/*
* If a spa load is in progress, then spa_dsl_pool may be
* uninitialized. But we shouldn't be resilvering during a spa
* load anyway.
*/
if (replacing &&
(spa_load_state(vd->vdev_spa) == SPA_LOAD_NONE) &&
dsl_scan_resilvering(vd->vdev_spa->spa_dsl_pool)) {
mm->mm_resilvering = B_TRUE;
} else {
mm->mm_resilvering = B_FALSE;
}
mm->mm_preferred = mm->mm_resilvering ? 0 :
(zio->io_offset >> vdev_mirror_shift) % c;
mm->mm_root = B_FALSE;
for (c = 0; c < mm->mm_children; c++) {
mc = &mm->mm_child[c];
mc->mc_vd = vd->vdev_child[c];
mc->mc_offset = zio->io_offset;
}
}
zio->io_vsd = mm;
zio->io_vsd_ops = &vdev_mirror_vsd_ops;
return (mm);
}
