Skip to content

Commit

Permalink
Use a struct to organize metaslab-group-allocator fields
Browse files Browse the repository at this point in the history 
Each metaslab group (of which there is one per top-level vdev) has
several (4, by default) "metaslab group allocators".  Each "allocator"
has its own metaslab that it prefers to allocate from (the "primary"
allocator), and each can perform allocations concurrently with the other
allocators.  In addition to the primary metaslab, there are several
other fields that need to be tracked separately for each allocator.
These are currently stored as several arrays in the metaslab_group_t,
each array indexed by allocator number.

This change organizes all the metaslab-group-allocator-specific fields
into a new struct, metaslab_group_allocator_t.  The metaslab_group_t now
needs only one array indexed by the allocator number - which contains
the metaslab_group_allocator_t's.

Reviewed-by: Paul Dagnelie <pcd@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
Closes openzfs#10213
  • Loading branch information
@ahrens @jsai20
ahrens authored and jsai20 committed Mar 30, 2021
1 parent f8ad2ff commit 83c4274
Show file tree
Hide file tree
Showing 3 changed files with 62 additions and 57 deletions.
15 changes: 11 additions & 4 deletions include/sys/metaslab_impl.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -203,6 +203,16 @@ struct metaslab_class {
multilist_t *mc_metaslab_txg_list;
};

/*
* Per-allocator data structure.
*/
typedef struct metaslab_group_allocator {
uint64_t mga_cur_max_alloc_queue_depth;
zfs_refcount_t mga_alloc_queue_depth;
metaslab_t *mga_primary;
metaslab_t *mga_secondary;
} metaslab_group_allocator_t;

/*
* Metaslab groups encapsulate all the allocatable regions (i.e. metaslabs)
* of a top-level vdev. They are linked together to form a circular linked
Expand All @@ -214,8 +224,6 @@ struct metaslab_class {
*/
struct metaslab_group {
kmutex_t mg_lock;
metaslab_t **mg_primaries;
metaslab_t **mg_secondaries;
avl_tree_t mg_metaslab_tree;
uint64_t mg_aliquot;
boolean_t mg_allocatable; /* can we allocate? */
Expand Down Expand Up @@ -263,9 +271,8 @@ struct metaslab_group {
* groups are unable to handle their share of allocations.
*/
uint64_t mg_max_alloc_queue_depth;
uint64_t *mg_cur_max_alloc_queue_depth;
zfs_refcount_t *mg_alloc_queue_depth;
int mg_allocators;
metaslab_group_allocator_t *mg_allocator; /* array */
/*
* A metalab group that can no longer allocate the minimum block
* size will set mg_no_free_space. Once a metaslab group is out
Expand Down
95 changes: 46 additions & 49 deletions module/zfs/metaslab.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -814,10 +814,6 @@ metaslab_group_create(metaslab_class_t *mc, vdev_t *vd, int allocators)
mutex_init(&mg->mg_lock, NULL, MUTEX_DEFAULT, NULL);
mutex_init(&mg->mg_ms_disabled_lock, NULL, MUTEX_DEFAULT, NULL);
cv_init(&mg->mg_ms_disabled_cv, NULL, CV_DEFAULT, NULL);
mg->mg_primaries = kmem_zalloc(allocators * sizeof (metaslab_t *),
KM_SLEEP);
mg->mg_secondaries = kmem_zalloc(allocators * sizeof (metaslab_t *),
KM_SLEEP);
avl_create(&mg->mg_metaslab_tree, metaslab_compare,
sizeof (metaslab_t), offsetof(metaslab_t, ms_group_node));
mg->mg_vd = vd;
Expand All @@ -827,13 +823,11 @@ metaslab_group_create(metaslab_class_t *mc, vdev_t *vd, int allocators)
mg->mg_no_free_space = B_TRUE;
mg->mg_allocators = allocators;

mg->mg_alloc_queue_depth = kmem_zalloc(allocators *
sizeof (zfs_refcount_t), KM_SLEEP);
mg->mg_cur_max_alloc_queue_depth = kmem_zalloc(allocators *
sizeof (uint64_t), KM_SLEEP);
mg->mg_allocator = kmem_zalloc(allocators *
sizeof (metaslab_group_allocator_t), KM_SLEEP);
for (int i = 0; i < allocators; i++) {
zfs_refcount_create_tracked(&mg->mg_alloc_queue_depth[i]);
mg->mg_cur_max_alloc_queue_depth[i] = 0;
metaslab_group_allocator_t *mga = &mg->mg_allocator[i];
zfs_refcount_create_tracked(&mga->mga_alloc_queue_depth);
}

mg->mg_taskq = taskq_create("metaslab_group_taskq", metaslab_load_pct,
Expand All @@ -856,21 +850,16 @@ metaslab_group_destroy(metaslab_group_t *mg)

taskq_destroy(mg->mg_taskq);
avl_destroy(&mg->mg_metaslab_tree);
kmem_free(mg->mg_primaries, mg->mg_allocators * sizeof (metaslab_t *));
kmem_free(mg->mg_secondaries, mg->mg_allocators *
sizeof (metaslab_t *));
mutex_destroy(&mg->mg_lock);
mutex_destroy(&mg->mg_ms_disabled_lock);
cv_destroy(&mg->mg_ms_disabled_cv);

for (int i = 0; i < mg->mg_allocators; i++) {
zfs_refcount_destroy(&mg->mg_alloc_queue_depth[i]);
mg->mg_cur_max_alloc_queue_depth[i] = 0;
metaslab_group_allocator_t *mga = &mg->mg_allocator[i];
zfs_refcount_destroy(&mga->mga_alloc_queue_depth);
}
kmem_free(mg->mg_alloc_queue_depth, mg->mg_allocators *
sizeof (zfs_refcount_t));
kmem_free(mg->mg_cur_max_alloc_queue_depth, mg->mg_allocators *
sizeof (uint64_t));
kmem_free(mg->mg_allocator, mg->mg_allocators *
sizeof (metaslab_group_allocator_t));

kmem_free(mg, sizeof (metaslab_group_t));
}
Expand Down Expand Up @@ -951,14 +940,15 @@ metaslab_group_passivate(metaslab_group_t *mg)
spa_config_enter(spa, locks & ~(SCL_ZIO - 1), spa, RW_WRITER);
metaslab_group_alloc_update(mg);
for (int i = 0; i < mg->mg_allocators; i++) {
metaslab_t *msp = mg->mg_primaries[i];
metaslab_group_allocator_t *mga = &mg->mg_allocator[i];
metaslab_t *msp = mga->mga_primary;
if (msp != NULL) {
mutex_enter(&msp->ms_lock);
metaslab_passivate(msp,
metaslab_weight_from_range_tree(msp));
mutex_exit(&msp->ms_lock);
}
msp = mg->mg_secondaries[i];
msp = mga->mga_secondary;
if (msp != NULL) {
mutex_enter(&msp->ms_lock);
metaslab_passivate(msp,
Expand Down Expand Up @@ -1218,9 +1208,9 @@ metaslab_group_allocatable(metaslab_group_t *mg, metaslab_group_t *rotor,
* regardless of the mg_allocatable or throttle settings.
*/
if (mg->mg_allocatable) {
metaslab_group_t *mgp;
metaslab_group_allocator_t *mga = &mg->mg_allocator[allocator];
int64_t qdepth;
uint64_t qmax = mg->mg_cur_max_alloc_queue_depth[allocator];
uint64_t qmax = mga->mga_cur_max_alloc_queue_depth;

if (!mc->mc_alloc_throttle_enabled)
return (B_TRUE);
Expand All @@ -1239,8 +1229,7 @@ metaslab_group_allocatable(metaslab_group_t *mg, metaslab_group_t *rotor,
*/
qmax = qmax * (4 + d) / 4;

qdepth = zfs_refcount_count(
&mg->mg_alloc_queue_depth[allocator]);
qdepth = zfs_refcount_count(&mga->mga_alloc_queue_depth);

/*
* If this metaslab group is below its qmax or it's
Expand All @@ -1258,11 +1247,14 @@ metaslab_group_allocatable(metaslab_group_t *mg, metaslab_group_t *rotor,
* racy since we can't hold the locks for all metaslab
* groups at the same time when we make this check.
*/
for (mgp = mg->mg_next; mgp != rotor; mgp = mgp->mg_next) {
qmax = mgp->mg_cur_max_alloc_queue_depth[allocator];
for (metaslab_group_t *mgp = mg->mg_next;
mgp != rotor; mgp = mgp->mg_next) {
metaslab_group_allocator_t *mgap =
&mgp->mg_allocator[allocator];
qmax = mgap->mga_cur_max_alloc_queue_depth;
qmax = qmax * (4 + d) / 4;
qdepth = zfs_refcount_count(
&mgp->mg_alloc_queue_depth[allocator]);
qdepth =
zfs_refcount_count(&mgap->mga_alloc_queue_depth);

/*
* If there is another metaslab group that
Expand Down Expand Up @@ -3205,6 +3197,7 @@ static int
metaslab_activate_allocator(metaslab_group_t *mg, metaslab_t *msp,
int allocator, uint64_t activation_weight)
{
metaslab_group_allocator_t *mga = &mg->mg_allocator[allocator];
ASSERT(MUTEX_HELD(&msp->ms_lock));

/*
Expand All @@ -3219,16 +3212,16 @@ metaslab_activate_allocator(metaslab_group_t *mg, metaslab_t *msp,
return (0);
}

metaslab_t **arr = (activation_weight == METASLAB_WEIGHT_PRIMARY ?
mg->mg_primaries : mg->mg_secondaries);
metaslab_t **mspp = (activation_weight == METASLAB_WEIGHT_PRIMARY ?
&mga->mga_primary : &mga->mga_secondary);

mutex_enter(&mg->mg_lock);
if (arr[allocator] != NULL) {
if (*mspp != NULL) {
mutex_exit(&mg->mg_lock);
return (EEXIST);
}

arr[allocator] = msp;
*mspp = msp;
ASSERT3S(msp->ms_allocator, ==, -1);
msp->ms_allocator = allocator;
msp->ms_primary = (activation_weight == METASLAB_WEIGHT_PRIMARY);
Expand All @@ -3237,7 +3230,6 @@ metaslab_activate_allocator(metaslab_group_t *mg, metaslab_t *msp,
msp->ms_activation_weight = msp->ms_weight;
metaslab_group_sort_impl(mg, msp,
msp->ms_weight | activation_weight);

mutex_exit(&mg->mg_lock);

return (0);
Expand Down Expand Up @@ -3337,14 +3329,15 @@ metaslab_passivate_allocator(metaslab_group_t *mg, metaslab_t *msp,
ASSERT3S(0, <=, msp->ms_allocator);
ASSERT3U(msp->ms_allocator, <, mg->mg_allocators);

metaslab_group_allocator_t *mga = &mg->mg_allocator[msp->ms_allocator];
if (msp->ms_primary) {
ASSERT3P(mg->mg_primaries[msp->ms_allocator], ==, msp);
ASSERT3P(mga->mga_primary, ==, msp);
ASSERT(msp->ms_weight & METASLAB_WEIGHT_PRIMARY);
mg->mg_primaries[msp->ms_allocator] = NULL;
mga->mga_primary = NULL;
} else {
ASSERT3P(mg->mg_secondaries[msp->ms_allocator], ==, msp);
ASSERT3P(mga->mga_secondary, ==, msp);
ASSERT(msp->ms_weight & METASLAB_WEIGHT_SECONDARY);
mg->mg_secondaries[msp->ms_allocator] = NULL;
mga->mga_secondary = NULL;
}
msp->ms_allocator = -1;
metaslab_group_sort_impl(mg, msp, weight);
Expand Down Expand Up @@ -4493,22 +4486,24 @@ metaslab_group_alloc_increment(spa_t *spa, uint64_t vdev, void *tag, int flags,
if (!mg->mg_class->mc_alloc_throttle_enabled)
return;

(void) zfs_refcount_add(&mg->mg_alloc_queue_depth[allocator], tag);
metaslab_group_allocator_t *mga = &mg->mg_allocator[allocator];
(void) zfs_refcount_add(&mga->mga_alloc_queue_depth, tag);
}

static void
metaslab_group_increment_qdepth(metaslab_group_t *mg, int allocator)
{
metaslab_group_allocator_t *mga = &mg->mg_allocator[allocator];
uint64_t max = mg->mg_max_alloc_queue_depth;
uint64_t cur = mg->mg_cur_max_alloc_queue_depth[allocator];
uint64_t cur = mga->mga_cur_max_alloc_queue_depth;
while (cur < max) {
if (atomic_cas_64(&mg->mg_cur_max_alloc_queue_depth[allocator],
if (atomic_cas_64(&mga->mga_cur_max_alloc_queue_depth,
cur, cur + 1) == cur) {
atomic_inc_64(
&mg->mg_class->mc_alloc_max_slots[allocator]);
return;
}
cur = mg->mg_cur_max_alloc_queue_depth[allocator];
cur = mga->mga_cur_max_alloc_queue_depth;
}
}

Expand All @@ -4524,7 +4519,8 @@ metaslab_group_alloc_decrement(spa_t *spa, uint64_t vdev, void *tag, int flags,
if (!mg->mg_class->mc_alloc_throttle_enabled)
return;

(void) zfs_refcount_remove(&mg->mg_alloc_queue_depth[allocator], tag);
metaslab_group_allocator_t *mga = &mg->mg_allocator[allocator];
(void) zfs_refcount_remove(&mga->mga_alloc_queue_depth, tag);
if (io_complete)
metaslab_group_increment_qdepth(mg, allocator);
}
Expand All @@ -4540,8 +4536,8 @@ metaslab_group_alloc_verify(spa_t *spa, const blkptr_t *bp, void *tag,
for (int d = 0; d < ndvas; d++) {
uint64_t vdev = DVA_GET_VDEV(&dva[d]);
metaslab_group_t *mg = vdev_lookup_top(spa, vdev)->vdev_mg;
VERIFY(zfs_refcount_not_held(
&mg->mg_alloc_queue_depth[allocator], tag));
metaslab_group_allocator_t *mga = &mg->mg_allocator[allocator];
VERIFY(zfs_refcount_not_held(&mga->mga_alloc_queue_depth, tag));
}
#endif
}
Expand Down Expand Up @@ -4716,6 +4712,7 @@ metaslab_group_alloc_normal(metaslab_group_t *mg, zio_alloc_list_t *zal,
*/
if (mg->mg_ms_ready < mg->mg_allocators * 3)
allocator = 0;
metaslab_group_allocator_t *mga = &mg->mg_allocator[allocator];

ASSERT3U(mg->mg_vd->vdev_ms_count, >=, 2);

Expand All @@ -4737,8 +4734,8 @@ metaslab_group_alloc_normal(metaslab_group_t *mg, zio_alloc_list_t *zal,
mutex_enter(&mg->mg_lock);

if (activation_weight == METASLAB_WEIGHT_PRIMARY &&
mg->mg_primaries[allocator] != NULL) {
msp = mg->mg_primaries[allocator];
mga->mga_primary != NULL) {
msp = mga->mga_primary;

/*
* Even though we don't hold the ms_lock for the
Expand All @@ -4753,8 +4750,8 @@ metaslab_group_alloc_normal(metaslab_group_t *mg, zio_alloc_list_t *zal,
was_active = B_TRUE;
ASSERT(msp->ms_weight & METASLAB_ACTIVE_MASK);
} else if (activation_weight == METASLAB_WEIGHT_SECONDARY &&
mg->mg_secondaries[allocator] != NULL) {
msp = mg->mg_secondaries[allocator];
mga->mga_secondary != NULL) {
msp = mga->mga_secondary;

/*
* See comment above about the similar assertions
Expand Down
9 changes: 5 additions & 4 deletions module/zfs/spa.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -8720,13 +8720,14 @@ spa_sync_adjust_vdev_max_queue_depth(spa_t *spa)
* allocations look at mg_max_alloc_queue_depth, and async
* allocations all happen from spa_sync().
*/
for (int i = 0; i < spa->spa_alloc_count; i++)
for (int i = 0; i < mg->mg_allocators; i++) {
ASSERT0(zfs_refcount_count(
&(mg->mg_alloc_queue_depth[i])));
&(mg->mg_allocator[i].mga_alloc_queue_depth)));
}
mg->mg_max_alloc_queue_depth = max_queue_depth;

for (int i = 0; i < spa->spa_alloc_count; i++) {
mg->mg_cur_max_alloc_queue_depth[i] =
for (int i = 0; i < mg->mg_allocators; i++) {
mg->mg_allocator[i].mga_cur_max_alloc_queue_depth =
zfs_vdev_def_queue_depth;
}
slots_per_allocator += zfs_vdev_def_queue_depth;
Expand Down

0 comments on commit 83c4274

Please sign in to comment.