#include "alloc.h"
#include "bkey_methods.h"
#include "btree_locking.h"
-#include "btree_update.h"
+#include "btree_update_interior.h"
#include "btree_io.h"
#include "btree_gc.h"
#include "buckets.h"
#include "journal.h"
#include "keylist.h"
#include "move.h"
+#include "replicas.h"
#include "super-io.h"
#include <linux/slab.h>
#include <linux/bitops.h>
#include <linux/freezer.h>
#include <linux/kthread.h>
+#include <linux/preempt.h>
#include <linux/rcupdate.h>
#include <trace/events/bcachefs.h>
struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
extent_for_each_ptr(e, ptr) {
- struct bch_dev *ca = c->devs[ptr->dev];
+ struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
size_t b = PTR_BUCKET_NR(ca, ptr);
if (gen_after(ca->oldest_gens[b], ptr->gen))
/*
* For runtime mark and sweep:
*/
-static u8 bch2_btree_mark_key(struct bch_fs *c, enum bkey_type type,
- struct bkey_s_c k)
+static u8 bch2_gc_mark_key(struct bch_fs *c, enum bkey_type type,
+ struct bkey_s_c k, unsigned flags)
{
+ struct gc_pos pos = { 0 };
+ struct bch_fs_usage *stats;
+ u8 ret = 0;
+
+ preempt_disable();
+ stats = this_cpu_ptr(c->usage_percpu);
switch (type) {
case BKEY_TYPE_BTREE:
- bch2_gc_mark_key(c, k, c->sb.btree_node_size, true);
- return 0;
+ bch2_mark_key(c, k, c->opts.btree_node_size, true, pos, stats,
+ 0, flags|
+ BCH_BUCKET_MARK_MAY_MAKE_UNAVAILABLE|
+ BCH_BUCKET_MARK_GC_LOCK_HELD);
+ break;
case BKEY_TYPE_EXTENTS:
- bch2_gc_mark_key(c, k, k.k->size, false);
- return bch2_btree_key_recalc_oldest_gen(c, k);
+ bch2_mark_key(c, k, k.k->size, false, pos, stats,
+ 0, flags|
+ BCH_BUCKET_MARK_MAY_MAKE_UNAVAILABLE|
+ BCH_BUCKET_MARK_GC_LOCK_HELD);
+ ret = bch2_btree_key_recalc_oldest_gen(c, k);
+ break;
default:
BUG();
}
+ preempt_enable();
+
+ return ret;
}
int bch2_btree_mark_key_initial(struct bch_fs *c, enum bkey_type type,
struct bkey_s_c k)
{
- int ret;
+ enum bch_data_type data_type = type == BKEY_TYPE_BTREE
+ ? BCH_DATA_BTREE : BCH_DATA_USER;
+ int ret = 0;
+
+ if (test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) ||
+ fsck_err_on(!bch2_bkey_replicas_marked(c, data_type, k), c,
+ "superblock not marked as containing replicas (type %u)",
+ data_type)) {
+ ret = bch2_mark_bkey_replicas(c, data_type, k);
+ if (ret)
+ return ret;
+ }
switch (k.k->type) {
case BCH_EXTENT:
const struct bch_extent_ptr *ptr;
extent_for_each_ptr(e, ptr) {
- struct bch_dev *ca = c->devs[ptr->dev];
+ struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
+ size_t b = PTR_BUCKET_NR(ca, ptr);
struct bucket *g = PTR_BUCKET(ca, ptr);
- unfixable_fsck_err_on(gen_cmp(ptr->gen, g->mark.gen) > 0, c,
- "%s ptr gen in the future: %u > %u",
- type == BKEY_TYPE_BTREE
- ? "btree" : "data",
- ptr->gen, g->mark.gen);
+ if (mustfix_fsck_err_on(!g->mark.gen_valid, c,
+ "found ptr with missing gen in alloc btree,\n"
+ "type %s gen %u",
+ bch2_data_types[data_type],
+ ptr->gen)) {
+ g->_mark.gen = ptr->gen;
+ g->_mark.gen_valid = 1;
+ set_bit(b, ca->buckets_dirty);
+ }
+
+ if (mustfix_fsck_err_on(gen_cmp(ptr->gen, g->mark.gen) > 0, c,
+ "%s ptr gen in the future: %u > %u",
+ bch2_data_types[data_type],
+ ptr->gen, g->mark.gen)) {
+ g->_mark.gen = ptr->gen;
+ g->_mark.gen_valid = 1;
+ set_bit(b, ca->buckets_dirty);
+ set_bit(BCH_FS_FIXED_GENS, &c->flags);
+ }
}
break;
max_t(u64, k.k->version.lo,
atomic64_read(&c->key_version)));
- bch2_btree_mark_key(c, type, k);
- return 0;
+ bch2_gc_mark_key(c, type, k, BCH_BUCKET_MARK_NOATOMIC);
fsck_err:
return ret;
}
-static bool btree_gc_mark_node(struct bch_fs *c, struct btree *b)
+static unsigned btree_gc_mark_node(struct bch_fs *c, struct btree *b)
{
- if (btree_node_has_ptrs(b)) {
- struct btree_node_iter iter;
- struct bkey unpacked;
- struct bkey_s_c k;
- u8 stale = 0;
+ enum bkey_type type = btree_node_type(b);
+ struct btree_node_iter iter;
+ struct bkey unpacked;
+ struct bkey_s_c k;
+ u8 stale = 0;
+ if (btree_node_has_ptrs(b))
for_each_btree_node_key_unpack(b, k, &iter,
btree_node_is_extents(b),
&unpacked) {
bch2_bkey_debugcheck(c, b, k);
- stale = max(stale, bch2_btree_mark_key(c,
- btree_node_type(b), k));
+ stale = max(stale, bch2_gc_mark_key(c, type, k, 0));
}
- if (btree_gc_rewrite_disabled(c))
- return false;
-
- if (stale > 10)
- return true;
- }
-
- if (btree_gc_always_rewrite(c))
- return true;
-
- return false;
+ return stale;
}
static inline void __gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
{
struct btree_iter iter;
struct btree *b;
- bool should_rewrite;
struct range_checks r;
unsigned depth = btree_id == BTREE_ID_EXTENTS ? 0 : 1;
- int ret;
+ unsigned max_stale;
+ int ret = 0;
/*
* if expensive_debug_checks is on, run range_checks on all leaf nodes:
btree_node_range_checks_init(&r, depth);
- for_each_btree_node(&iter, c, btree_id, POS_MIN, depth, b) {
+ __for_each_btree_node(&iter, c, btree_id, POS_MIN,
+ 0, depth, BTREE_ITER_PREFETCH, b) {
btree_node_range_checks(c, b, &r);
bch2_verify_btree_nr_keys(b);
- should_rewrite = btree_gc_mark_node(c, b);
+ max_stale = btree_gc_mark_node(c, b);
gc_pos_set(c, gc_pos_btree_node(b));
- if (should_rewrite)
- bch2_btree_node_rewrite(&iter, b, NULL);
+ if (max_stale > 32)
+ bch2_btree_node_rewrite(c, &iter,
+ b->data->keys.seq,
+ BTREE_INSERT_USE_RESERVE|
+ BTREE_INSERT_GC_LOCK_HELD);
+ else if (!btree_gc_rewrite_disabled(c) &&
+ (btree_gc_always_rewrite(c) || max_stale > 16))
+ bch2_btree_node_rewrite(c, &iter,
+ b->data->keys.seq,
+ BTREE_INSERT_NOWAIT|
+ BTREE_INSERT_GC_LOCK_HELD);
bch2_btree_iter_cond_resched(&iter);
}
mutex_lock(&c->btree_root_lock);
b = c->btree_roots[btree_id].b;
- bch2_btree_mark_key(c, BKEY_TYPE_BTREE, bkey_i_to_s_c(&b->key));
+ if (!btree_node_fake(b))
+ bch2_gc_mark_key(c, BKEY_TYPE_BTREE, bkey_i_to_s_c(&b->key), 0);
gc_pos_set(c, gc_pos_btree_root(b->btree_id));
mutex_unlock(&c->btree_root_lock);
return 0;
}
-static void bch2_mark_allocator_buckets(struct bch_fs *c)
-{
- struct bch_dev *ca;
- struct open_bucket *ob;
- size_t i, j, iter;
- unsigned ci;
-
- for_each_member_device(ca, c, ci) {
- spin_lock(&ca->freelist_lock);
-
- fifo_for_each_entry(i, &ca->free_inc, iter)
- bch2_mark_alloc_bucket(ca, &ca->buckets[i], true);
-
- for (j = 0; j < RESERVE_NR; j++)
- fifo_for_each_entry(i, &ca->free[j], iter)
- bch2_mark_alloc_bucket(ca, &ca->buckets[i], true);
-
- spin_unlock(&ca->freelist_lock);
- }
-
- for (ob = c->open_buckets;
- ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
- ob++) {
- const struct bch_extent_ptr *ptr;
-
- mutex_lock(&ob->lock);
- open_bucket_for_each_ptr(ob, ptr) {
- ca = c->devs[ptr->dev];
- bch2_mark_alloc_bucket(ca, PTR_BUCKET(ca, ptr), true);
- }
- mutex_unlock(&ob->lock);
- }
-}
-
-static void mark_metadata_sectors(struct bch_dev *ca, u64 start, u64 end,
- enum bucket_data_type type)
+static void mark_metadata_sectors(struct bch_fs *c, struct bch_dev *ca,
+ u64 start, u64 end,
+ enum bch_data_type type,
+ unsigned flags)
{
- u64 b = start >> ca->bucket_bits;
+ u64 b = sector_to_bucket(ca, start);
do {
- bch2_mark_metadata_bucket(ca, ca->buckets + b, type, true);
+ unsigned sectors =
+ min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
+
+ bch2_mark_metadata_bucket(c, ca, b, type, sectors,
+ gc_phase(GC_PHASE_SB), flags);
b++;
- } while (b < end >> ca->bucket_bits);
+ start += sectors;
+ } while (start < end);
}
-static void bch2_dev_mark_superblocks(struct bch_dev *ca)
+void bch2_mark_dev_superblock(struct bch_fs *c, struct bch_dev *ca,
+ unsigned flags)
{
struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
unsigned i;
-
- for (i = 0; i < layout->nr_superblocks; i++) {
- if (layout->sb_offset[i] == BCH_SB_SECTOR)
- mark_metadata_sectors(ca, 0, BCH_SB_SECTOR,
- BUCKET_SB);
-
- mark_metadata_sectors(ca,
- layout->sb_offset[i],
- layout->sb_offset[i] +
- (1 << layout->sb_max_size_bits),
- BUCKET_SB);
- }
-}
-
-/*
- * Mark non btree metadata - prios, journal
- */
-void bch2_mark_dev_metadata(struct bch_fs *c, struct bch_dev *ca)
-{
- unsigned i;
u64 b;
- lockdep_assert_held(&c->sb_lock);
+ if (c)
+ lockdep_assert_held(&c->sb_lock);
- bch2_dev_mark_superblocks(ca);
+ for (i = 0; i < layout->nr_superblocks; i++) {
+ u64 offset = le64_to_cpu(layout->sb_offset[i]);
- spin_lock(&c->journal.lock);
+ if (offset == BCH_SB_SECTOR)
+ mark_metadata_sectors(c, ca, 0, BCH_SB_SECTOR,
+ BCH_DATA_SB, flags);
- for (i = 0; i < ca->journal.nr; i++) {
- b = ca->journal.buckets[i];
- bch2_mark_metadata_bucket(ca, ca->buckets + b,
- BUCKET_JOURNAL, true);
+ mark_metadata_sectors(c, ca, offset,
+ offset + (1 << layout->sb_max_size_bits),
+ BCH_DATA_SB, flags);
}
- spin_unlock(&c->journal.lock);
-
- spin_lock(&ca->prio_buckets_lock);
+ if (c)
+ spin_lock(&c->journal.lock);
- for (i = 0; i < prio_buckets(ca) * 2; i++) {
- b = ca->prio_buckets[i];
- if (b)
- bch2_mark_metadata_bucket(ca, ca->buckets + b,
- BUCKET_PRIOS, true);
+ for (i = 0; i < ca->journal.nr; i++) {
+ b = ca->journal.buckets[i];
+ bch2_mark_metadata_bucket(c, ca, b, BCH_DATA_JOURNAL,
+ ca->mi.bucket_size,
+ gc_phase(GC_PHASE_SB), flags);
}
- spin_unlock(&ca->prio_buckets_lock);
+ if (c)
+ spin_unlock(&c->journal.lock);
}
-static void bch2_mark_metadata(struct bch_fs *c)
+static void bch2_mark_superblocks(struct bch_fs *c)
{
struct bch_dev *ca;
unsigned i;
mutex_lock(&c->sb_lock);
- gc_pos_set(c, gc_phase(GC_PHASE_SB_METADATA));
+ gc_pos_set(c, gc_phase(GC_PHASE_SB));
for_each_online_member(ca, c, i)
- bch2_mark_dev_metadata(c, ca);
+ bch2_mark_dev_superblock(c, ca,
+ BCH_BUCKET_MARK_MAY_MAKE_UNAVAILABLE|
+ BCH_BUCKET_MARK_GC_LOCK_HELD);
mutex_unlock(&c->sb_lock);
}
/* Also see bch2_pending_btree_node_free_insert_done() */
static void bch2_mark_pending_btree_node_frees(struct bch_fs *c)
{
+ struct gc_pos pos = { 0 };
struct bch_fs_usage stats = { 0 };
- struct btree_interior_update *as;
+ struct btree_update *as;
struct pending_btree_node_free *d;
mutex_lock(&c->btree_interior_update_lock);
for_each_pending_btree_node_free(c, as, d)
if (d->index_update_done)
- __bch2_gc_mark_key(c, bkey_i_to_s_c(&d->key),
- c->sb.btree_node_size, true,
- &stats);
+ bch2_mark_key(c, bkey_i_to_s_c(&d->key),
+ c->opts.btree_node_size, true, pos,
+ &stats, 0,
+ BCH_BUCKET_MARK_MAY_MAKE_UNAVAILABLE|
+ BCH_BUCKET_MARK_GC_LOCK_HELD);
/*
* Don't apply stats - pending deletes aren't tracked in
* bch_alloc_stats:
mutex_unlock(&c->btree_interior_update_lock);
}
-/**
- * bch_gc - recompute bucket marks and oldest_gen, rewrite btree nodes
- */
-void bch2_gc(struct bch_fs *c)
+static void bch2_mark_allocator_buckets(struct bch_fs *c)
{
struct bch_dev *ca;
- struct bucket *g;
- struct bucket_mark new;
- u64 start_time = local_clock();
- unsigned i;
- int cpu;
+ struct open_bucket *ob;
+ size_t i, j, iter;
+ unsigned ci;
- /*
- * Walk _all_ references to buckets, and recompute them:
- *
- * Order matters here:
- * - Concurrent GC relies on the fact that we have a total ordering for
- * everything that GC walks - see gc_will_visit_node(),
- * gc_will_visit_root()
- *
- * - also, references move around in the course of index updates and
- * various other crap: everything needs to agree on the ordering
- * references are allowed to move around in - e.g., we're allowed to
- * start with a reference owned by an open_bucket (the allocator) and
- * move it to the btree, but not the reverse.
- *
- * This is necessary to ensure that gc doesn't miss references that
- * move around - if references move backwards in the ordering GC
- * uses, GC could skip past them
- */
+ spin_lock(&c->freelist_lock);
+ gc_pos_set(c, gc_pos_alloc(c, NULL));
- if (test_bit(BCH_FS_GC_FAILURE, &c->flags))
- return;
+ for_each_member_device(ca, c, ci) {
+ fifo_for_each_entry(i, &ca->free_inc, iter)
+ bch2_mark_alloc_bucket(c, ca, i, true,
+ gc_pos_alloc(c, NULL),
+ BCH_BUCKET_MARK_MAY_MAKE_UNAVAILABLE|
+ BCH_BUCKET_MARK_GC_LOCK_HELD);
- trace_gc_start(c);
- /*
- * Do this before taking gc_lock - bch2_disk_reservation_get() blocks on
- * gc_lock if sectors_available goes to 0:
- */
- bch2_recalc_sectors_available(c);
- down_write(&c->gc_lock);
+ for (j = 0; j < RESERVE_NR; j++)
+ fifo_for_each_entry(i, &ca->free[j], iter)
+ bch2_mark_alloc_bucket(c, ca, i, true,
+ gc_pos_alloc(c, NULL),
+ BCH_BUCKET_MARK_MAY_MAKE_UNAVAILABLE|
+ BCH_BUCKET_MARK_GC_LOCK_HELD);
+ }
+
+ spin_unlock(&c->freelist_lock);
+
+ for (ob = c->open_buckets;
+ ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
+ ob++) {
+ spin_lock(&ob->lock);
+ if (ob->valid) {
+ gc_pos_set(c, gc_pos_alloc(c, ob));
+ ca = bch_dev_bkey_exists(c, ob->ptr.dev);
+ bch2_mark_alloc_bucket(c, ca, PTR_BUCKET_NR(ca, &ob->ptr), true,
+ gc_pos_alloc(c, ob),
+ BCH_BUCKET_MARK_MAY_MAKE_UNAVAILABLE|
+ BCH_BUCKET_MARK_GC_LOCK_HELD);
+ }
+ spin_unlock(&ob->lock);
+ }
+}
+
+static void bch2_gc_start(struct bch_fs *c)
+{
+ struct bch_dev *ca;
+ struct bucket_array *buckets;
+ struct bucket_mark new;
+ unsigned i;
+ size_t b;
+ int cpu;
lg_global_lock(&c->usage_lock);
per_cpu_ptr(c->usage_percpu, cpu);
memset(p->s, 0, sizeof(p->s));
- p->persistent_reserved = 0;
}
lg_global_unlock(&c->usage_lock);
/* Clear bucket marks: */
- for_each_member_device(ca, c, i)
- for_each_bucket(g, ca) {
- bucket_cmpxchg(g, new, ({
+ for_each_member_device(ca, c, i) {
+ down_read(&ca->bucket_lock);
+ buckets = bucket_array(ca);
+
+ for (b = buckets->first_bucket; b < buckets->nbuckets; b++) {
+ bucket_cmpxchg(buckets->b + b, new, ({
new.owned_by_allocator = 0;
new.data_type = 0;
new.cached_sectors = 0;
new.dirty_sectors = 0;
}));
- ca->oldest_gens[g - ca->buckets] = new.gen;
+ ca->oldest_gens[b] = new.gen;
}
+ up_read(&ca->bucket_lock);
+ }
+}
- /* Walk allocator's references: */
- bch2_mark_allocator_buckets(c);
+/**
+ * bch_gc - recompute bucket marks and oldest_gen, rewrite btree nodes
+ */
+void bch2_gc(struct bch_fs *c)
+{
+ struct bch_dev *ca;
+ u64 start_time = local_clock();
+ unsigned i;
+
+ /*
+ * Walk _all_ references to buckets, and recompute them:
+ *
+ * Order matters here:
+ * - Concurrent GC relies on the fact that we have a total ordering for
+ * everything that GC walks - see gc_will_visit_node(),
+ * gc_will_visit_root()
+ *
+ * - also, references move around in the course of index updates and
+ * various other crap: everything needs to agree on the ordering
+ * references are allowed to move around in - e.g., we're allowed to
+ * start with a reference owned by an open_bucket (the allocator) and
+ * move it to the btree, but not the reverse.
+ *
+ * This is necessary to ensure that gc doesn't miss references that
+ * move around - if references move backwards in the ordering GC
+ * uses, GC could skip past them
+ */
+ trace_gc_start(c);
+
+ /*
+ * Do this before taking gc_lock - bch2_disk_reservation_get() blocks on
+ * gc_lock if sectors_available goes to 0:
+ */
+ bch2_recalc_sectors_available(c);
+
+ down_write(&c->gc_lock);
+ if (test_bit(BCH_FS_GC_FAILURE, &c->flags))
+ goto out;
+
+ bch2_gc_start(c);
/* Walk btree: */
while (c->gc_pos.phase < (int) BTREE_ID_NR) {
if (ret) {
bch_err(c, "btree gc failed: %d", ret);
set_bit(BCH_FS_GC_FAILURE, &c->flags);
- up_write(&c->gc_lock);
- return;
+ goto out;
}
gc_pos_set(c, gc_phase(c->gc_pos.phase + 1));
}
- bch2_mark_metadata(c);
+ bch2_mark_superblocks(c);
bch2_mark_pending_btree_node_frees(c);
+ bch2_mark_allocator_buckets(c);
for_each_member_device(ca, c, i)
atomic_long_set(&ca->saturated_count, 0);
/* Indicates that gc is no longer in progress: */
gc_pos_set(c, gc_phase(GC_PHASE_DONE));
-
+ c->gc_count++;
+out:
up_write(&c->gc_lock);
trace_gc_end(c);
bch2_time_stats_update(&c->btree_gc_time, start_time);
*/
for_each_member_device(ca, c, i)
bch2_wake_allocator(ca);
+
+ /*
+ * At startup, allocations can happen directly instead of via the
+ * allocator thread - issue wakeup in case they blocked on gc_lock:
+ */
+ closure_wake_up(&c->freelist_wait);
}
/* Btree coalescing */
btree_keys_account_key_add(&b->nr, 0, k);
}
-static void bch2_coalesce_nodes(struct btree *old_nodes[GC_MERGE_NODES],
- struct btree_iter *iter)
+static void bch2_coalesce_nodes(struct bch_fs *c, struct btree_iter *iter,
+ struct btree *old_nodes[GC_MERGE_NODES])
{
- struct btree *parent = iter->nodes[old_nodes[0]->level + 1];
- struct bch_fs *c = iter->c;
+ struct btree *parent = btree_node_parent(iter, old_nodes[0]);
unsigned i, nr_old_nodes, nr_new_nodes, u64s = 0;
unsigned blocks = btree_blocks(c) * 2 / 3;
struct btree *new_nodes[GC_MERGE_NODES];
- struct btree_interior_update *as;
- struct btree_reserve *res;
+ struct btree_update *as;
struct keylist keylist;
struct bkey_format_state format_state;
struct bkey_format new_format;
memset(new_nodes, 0, sizeof(new_nodes));
- bch2_keylist_init(&keylist, NULL, 0);
+ bch2_keylist_init(&keylist, NULL);
/* Count keys that are not deleted */
for (i = 0; i < GC_MERGE_NODES && old_nodes[i]; i++)
DIV_ROUND_UP(u64s, nr_old_nodes - 1)) > blocks)
return;
- res = bch2_btree_reserve_get(c, parent, nr_old_nodes,
- BTREE_INSERT_NOFAIL|
- BTREE_INSERT_USE_RESERVE,
- NULL);
- if (IS_ERR(res)) {
- trace_btree_gc_coalesce_fail(c,
- BTREE_GC_COALESCE_FAIL_RESERVE_GET);
- return;
- }
-
- if (bch2_keylist_realloc(&keylist, NULL, 0,
- (BKEY_U64s + BKEY_EXTENT_U64s_MAX) * nr_old_nodes)) {
- trace_btree_gc_coalesce_fail(c,
- BTREE_GC_COALESCE_FAIL_KEYLIST_REALLOC);
- goto out;
- }
-
/* Find a format that all keys in @old_nodes can pack into */
bch2_bkey_format_init(&format_state);
if (!bch2_btree_node_format_fits(c, old_nodes[i], &new_format)) {
trace_btree_gc_coalesce_fail(c,
BTREE_GC_COALESCE_FAIL_FORMAT_FITS);
- goto out;
+ return;
}
- trace_btree_gc_coalesce(c, parent, nr_old_nodes);
+ if (bch2_keylist_realloc(&keylist, NULL, 0,
+ (BKEY_U64s + BKEY_EXTENT_U64s_MAX) * nr_old_nodes)) {
+ trace_btree_gc_coalesce_fail(c,
+ BTREE_GC_COALESCE_FAIL_KEYLIST_REALLOC);
+ return;
+ }
+
+ as = bch2_btree_update_start(c, iter->btree_id,
+ btree_update_reserve_required(c, parent) + nr_old_nodes,
+ BTREE_INSERT_NOFAIL|
+ BTREE_INSERT_USE_RESERVE,
+ NULL);
+ if (IS_ERR(as)) {
+ trace_btree_gc_coalesce_fail(c,
+ BTREE_GC_COALESCE_FAIL_RESERVE_GET);
+ bch2_keylist_free(&keylist, NULL);
+ return;
+ }
- as = bch2_btree_interior_update_alloc(c);
+ trace_btree_gc_coalesce(c, old_nodes[0]);
for (i = 0; i < nr_old_nodes; i++)
- bch2_btree_interior_update_will_free_node(c, as, old_nodes[i]);
+ bch2_btree_interior_update_will_free_node(as, old_nodes[i]);
/* Repack everything with @new_format and sort down to one bset */
for (i = 0; i < nr_old_nodes; i++)
new_nodes[i] =
- __bch2_btree_node_alloc_replacement(c, old_nodes[i],
- new_format, res);
+ __bch2_btree_node_alloc_replacement(as, old_nodes[i],
+ new_format);
/*
* Conceptually we concatenate the nodes together and slice them
bch2_btree_build_aux_trees(n);
six_unlock_write(&n->lock);
- bch2_btree_node_write(c, n, &as->cl, SIX_LOCK_intent, -1);
+ bch2_btree_node_write(c, n, SIX_LOCK_intent);
}
/*
bch2_keylist_add_in_order(&keylist, &new_nodes[i]->key);
/* Insert the newly coalesced nodes */
- bch2_btree_insert_node(parent, iter, &keylist, res, as);
+ bch2_btree_insert_node(as, parent, iter, &keylist);
BUG_ON(!bch2_keylist_empty(&keylist));
- BUG_ON(iter->nodes[old_nodes[0]->level] != old_nodes[0]);
+ BUG_ON(iter->l[old_nodes[0]->level].b != old_nodes[0]);
BUG_ON(!bch2_btree_iter_node_replace(iter, new_nodes[0]));
/* Free the old nodes and update our sliding window */
for (i = 0; i < nr_old_nodes; i++) {
- bch2_btree_node_free_inmem(iter, old_nodes[i]);
+ bch2_btree_node_free_inmem(c, old_nodes[i], iter);
six_unlock_intent(&old_nodes[i]->lock);
/*
six_unlock_intent(&new_nodes[i]->lock);
}
}
-out:
+
+ bch2_btree_update_done(as);
bch2_keylist_free(&keylist, NULL);
- bch2_btree_reserve_put(c, res);
}
static int bch2_coalesce_btree(struct bch_fs *c, enum btree_id btree_id)
*/
memset(merge, 0, sizeof(merge));
- __for_each_btree_node(&iter, c, btree_id, POS_MIN, 0, b, U8_MAX) {
+ __for_each_btree_node(&iter, c, btree_id, POS_MIN,
+ BTREE_MAX_DEPTH, 0,
+ BTREE_ITER_PREFETCH, b) {
memmove(merge + 1, merge,
sizeof(merge) - sizeof(merge[0]));
memmove(lock_seq + 1, lock_seq,
}
memset(merge + i, 0, (GC_MERGE_NODES - i) * sizeof(merge[0]));
- bch2_coalesce_nodes(merge, &iter);
+ bch2_coalesce_nodes(c, &iter, merge);
for (i = 1; i < GC_MERGE_NODES && merge[i]; i++) {
lock_seq[i] = merge[i]->lock.state.seq;
* and the nodes in our sliding window might not have the same
* parent anymore - blow away the sliding window:
*/
- if (iter.nodes[iter.level + 1] &&
+ if (btree_iter_node(&iter, iter.level + 1) &&
!btree_node_intent_locked(&iter, iter.level + 1))
memset(merge + 1, 0,
(GC_MERGE_NODES - 1) * sizeof(merge[0]));
*/
void bch2_coalesce(struct bch_fs *c)
{
- u64 start_time;
enum btree_id id;
if (test_bit(BCH_FS_GC_FAILURE, &c->flags))
down_read(&c->gc_lock);
trace_gc_coalesce_start(c);
- start_time = local_clock();
for (id = 0; id < BTREE_ID_NR; id++) {
int ret = c->btree_roots[id].b
}
}
- bch2_time_stats_update(&c->btree_coalesce_time, start_time);
trace_gc_coalesce_end(c);
up_read(&c->gc_lock);
}
set_freezable();
while (1) {
- unsigned long next = last + c->capacity / 16;
-
- while (atomic_long_read(&clock->now) < next) {
+ while (1) {
set_current_state(TASK_INTERRUPTIBLE);
if (kthread_should_stop()) {
return 0;
}
- if (atomic_read(&c->kick_gc) != last_kick) {
- __set_current_state(TASK_RUNNING);
+ if (atomic_read(&c->kick_gc) != last_kick)
break;
+
+ if (c->btree_gc_periodic) {
+ unsigned long next = last + c->capacity / 16;
+
+ if (atomic_long_read(&clock->now) >= next)
+ break;
+
+ bch2_io_clock_schedule_timeout(clock, next);
+ } else {
+ schedule();
}
- bch2_io_clock_schedule_timeout(clock, next);
try_to_freeze();
}
+ __set_current_state(TASK_RUNNING);
last = atomic_long_read(&clock->now);
last_kick = atomic_read(&c->kick_gc);
bch2_gc(c);
- if (!btree_gc_coalesce_disabled(c))
- bch2_coalesce(c);
debug_check_no_locks_held();
}
if (!c->btree_roots[id].b)
return 0;
- ret = bch2_btree_mark_key_initial(c, BKEY_TYPE_BTREE,
- bkey_i_to_s_c(&c->btree_roots[id].b->key));
+ b = c->btree_roots[id].b;
+ if (!btree_node_fake(b))
+ ret = bch2_btree_mark_key_initial(c, BKEY_TYPE_BTREE,
+ bkey_i_to_s_c(&b->key));
if (ret)
return ret;
* We have to hit every btree node before starting journal replay, in
* order for the journal seq blacklist machinery to work:
*/
- for_each_btree_node(&iter, c, id, POS_MIN, 0, b) {
+ for_each_btree_node(&iter, c, id, POS_MIN, BTREE_ITER_PREFETCH, b) {
btree_node_range_checks(c, b, &r);
if (btree_node_has_ptrs(b)) {
bch2_btree_iter_cond_resched(&iter);
}
err:
- bch2_btree_iter_unlock(&iter);
- return ret;
+ return bch2_btree_iter_unlock(&iter) ?: ret;
}
-int bch2_initial_gc(struct bch_fs *c, struct list_head *journal)
+static int __bch2_initial_gc(struct bch_fs *c, struct list_head *journal)
{
+ unsigned iter = 0;
enum btree_id id;
int ret;
+ mutex_lock(&c->sb_lock);
+ if (!bch2_sb_get_replicas(c->disk_sb.sb)) {
+ if (BCH_SB_INITIALIZED(c->disk_sb.sb))
+ bch_info(c, "building replicas info");
+ set_bit(BCH_FS_REBUILD_REPLICAS, &c->flags);
+ }
+ mutex_unlock(&c->sb_lock);
+again:
+ bch2_gc_start(c);
+
for (id = 0; id < BTREE_ID_NR; id++) {
ret = bch2_initial_gc_btree(c, id);
if (ret)
return ret;
}
- if (journal) {
- ret = bch2_journal_mark(c, journal);
- if (ret)
- return ret;
- }
+ ret = bch2_journal_mark(c, journal);
+ if (ret)
+ return ret;
- bch2_mark_metadata(c);
+ if (test_bit(BCH_FS_FIXED_GENS, &c->flags)) {
+ if (iter++ > 2) {
+ bch_info(c, "Unable to fix bucket gens, looping");
+ return -EINVAL;
+ }
+
+ bch_info(c, "Fixed gens, restarting initial mark and sweep:");
+ clear_bit(BCH_FS_FIXED_GENS, &c->flags);
+ goto again;
+ }
/*
* Skip past versions that might have possibly been used (as nonces),
if (c->sb.encryption_type)
atomic64_add(1 << 16, &c->key_version);
+ bch2_mark_superblocks(c);
+
gc_pos_set(c, gc_phase(GC_PHASE_DONE));
set_bit(BCH_FS_INITIAL_GC_DONE, &c->flags);
return 0;
}
+
+int bch2_initial_gc(struct bch_fs *c, struct list_head *journal)
+{
+ int ret;
+
+ down_write(&c->gc_lock);
+ ret = __bch2_initial_gc(c, journal);
+ up_write(&c->gc_lock);
+
+ return ret;
+}