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);
}
static int
zil_dva_compare(const void *x1, const void *x2)
{
const dva_t *dva1 = x1;
const dva_t *dva2 = x2;
if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2))
return (-1);
if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2))
return (1);
if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2))
return (-1);
if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2))
return (1);
return (0);
}
/*
* Set the external verifier for a gang block based on <vdev, offset, txg>,
* a tuple which is guaranteed to be unique for the life of the pool.
*/
static void
zio_checksum_gang_verifier(zio_cksum_t *zcp, blkptr_t *bp)
{
dva_t *dva = BP_IDENTITY(bp);
uint64_t txg = BP_PHYSICAL_BIRTH(bp);
ASSERT(BP_IS_GANG(bp));
ZIO_SET_CHECKSUM(zcp, DVA_GET_VDEV(dva), DVA_GET_OFFSET(dva), txg, 0);
}
/*
* 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);
}
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);
}