#include "btree_iter.h"
#include "btree_locking.h"
#include "buckets.h"
+#include "error.h"
#include "extents.h"
#include "journal.h"
#include "journal_reclaim.h"
#include "keylist.h"
+#include "recovery.h"
#include "replicas.h"
#include "super-io.h"
#include <linux/random.h>
#include <trace/events/bcachefs.h>
+static void bch2_btree_insert_node(struct btree_update *, struct btree_trans *,
+ struct btree_path *, struct btree *,
+ struct keylist *, unsigned);
+static void bch2_btree_update_add_new_node(struct btree_update *, struct btree *);
+
/* Debug code: */
/*
struct bkey_s_c k;
struct bkey_s_c_btree_ptr_v2 bp;
struct bkey unpacked;
+ char buf1[100], buf2[100];
BUG_ON(!b->c.level);
- if (!test_bit(BCH_FS_BTREE_INTERIOR_REPLAY_DONE, &c->flags))
+ if (!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags))
return;
bch2_btree_node_iter_init_from_start(&iter, b);
break;
bp = bkey_s_c_to_btree_ptr_v2(k);
- BUG_ON(bkey_cmp(next_node, bp.v->min_key));
+ if (bpos_cmp(next_node, bp.v->min_key)) {
+ bch2_dump_btree_node(c, b);
+ panic("expected next min_key %s got %s\n",
+ (bch2_bpos_to_text(&PBUF(buf1), next_node), buf1),
+ (bch2_bpos_to_text(&PBUF(buf2), bp.v->min_key), buf2));
+ }
bch2_btree_node_iter_advance(&iter, b);
if (bch2_btree_node_iter_end(&iter)) {
- BUG_ON(bkey_cmp(k.k->p, b->key.k.p));
+ if (bpos_cmp(k.k->p, b->key.k.p)) {
+ bch2_dump_btree_node(c, b);
+ panic("expected end %s got %s\n",
+ (bch2_bpos_to_text(&PBUF(buf1), b->key.k.p), buf1),
+ (bch2_bpos_to_text(&PBUF(buf2), k.k->p), buf2));
+ }
break;
}
- next_node = bkey_successor(k.k->p);
+ next_node = bpos_successor(k.k->p);
}
#endif
}
struct bset_tree *t;
struct bkey uk;
- bch2_bkey_format_add_pos(s, b->data->min_key);
-
for_each_bset(b, t)
bset_tree_for_each_key(b, t, k)
- if (!bkey_whiteout(k)) {
+ if (!bkey_deleted(k)) {
uk = bkey_unpack_key(b, k);
bch2_bkey_format_add_key(s, &uk);
}
struct bkey_format_state s;
bch2_bkey_format_init(&s);
+ bch2_bkey_format_add_pos(&s, b->data->min_key);
+ bch2_bkey_format_add_pos(&s, b->data->max_key);
__bch2_btree_calc_format(&s, b);
return bch2_bkey_format_done(&s);
clear_btree_node_noevict(b);
- bch2_btree_node_hash_remove(&c->btree_cache, b);
-
mutex_lock(&c->btree_cache.lock);
list_move(&b->list, &c->btree_cache.freeable);
mutex_unlock(&c->btree_cache.lock);
}
-void bch2_btree_node_free_never_inserted(struct bch_fs *c, struct btree *b)
+static void bch2_btree_node_free_inmem(struct btree_trans *trans,
+ struct btree *b)
{
- struct open_buckets ob = b->ob;
+ struct bch_fs *c = trans->c;
+ struct btree_path *path;
- b->ob.nr = 0;
+ trans_for_each_path(trans, path)
+ BUG_ON(path->l[b->c.level].b == b &&
+ path->l[b->c.level].lock_seq == b->c.lock.state.seq);
- clear_btree_node_dirty(b);
+ six_lock_write(&b->c.lock, NULL, NULL);
- btree_node_lock_type(c, b, SIX_LOCK_write);
+ bch2_btree_node_hash_remove(&c->btree_cache, b);
__btree_node_free(c, b);
- six_unlock_write(&b->c.lock);
-
- bch2_open_buckets_put(c, &ob);
-}
-
-void bch2_btree_node_free_inmem(struct bch_fs *c, struct btree *b,
- struct btree_iter *iter)
-{
- struct btree_iter *linked;
-
- trans_for_each_iter(iter->trans, linked)
- BUG_ON(linked->l[b->c.level].b == b);
- six_lock_write(&b->c.lock, NULL, NULL);
- __btree_node_free(c, b);
six_unlock_write(&b->c.lock);
six_unlock_intent(&b->c.lock);
}
{
struct write_point *wp;
struct btree *b;
- BKEY_PADDED(k) tmp;
+ __BKEY_PADDED(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
struct open_buckets ob = { .nr = 0 };
struct bch_devs_list devs_have = (struct bch_devs_list) { 0 };
unsigned nr_reserve;
enum alloc_reserve alloc_reserve;
- if (flags & BTREE_INSERT_USE_ALLOC_RESERVE) {
+ if (flags & BTREE_INSERT_USE_RESERVE) {
nr_reserve = 0;
- alloc_reserve = RESERVE_ALLOC;
- } else if (flags & BTREE_INSERT_USE_RESERVE) {
- nr_reserve = BTREE_NODE_RESERVE / 2;
- alloc_reserve = RESERVE_BTREE;
+ alloc_reserve = RESERVE_BTREE_MOVINGGC;
} else {
nr_reserve = BTREE_NODE_RESERVE;
- alloc_reserve = RESERVE_NONE;
+ alloc_reserve = RESERVE_BTREE;
}
mutex_lock(&c->btree_reserve_cache_lock);
mutex_unlock(&c->btree_reserve_cache_lock);
retry:
- wp = bch2_alloc_sectors_start(c, c->opts.foreground_target, 0,
+ wp = bch2_alloc_sectors_start(c,
+ c->opts.metadata_target ?:
+ c->opts.foreground_target,
+ 0,
writepoint_ptr(&c->btree_write_point),
&devs_have,
res->nr_replicas,
if (IS_ERR(wp))
return ERR_CAST(wp);
- if (wp->sectors_free < c->opts.btree_node_size) {
+ if (wp->sectors_free < btree_sectors(c)) {
struct open_bucket *ob;
unsigned i;
open_bucket_for_each(c, &wp->ptrs, ob, i)
- if (ob->sectors_free < c->opts.btree_node_size)
+ if (ob->sectors_free < btree_sectors(c))
ob->sectors_free = 0;
bch2_alloc_sectors_done(c, wp);
goto retry;
}
- if (c->sb.features & (1ULL << BCH_FEATURE_btree_ptr_v2))
- bkey_btree_ptr_v2_init(&tmp.k);
- else
- bkey_btree_ptr_init(&tmp.k);
-
- bch2_alloc_sectors_append_ptrs(c, wp, &tmp.k, c->opts.btree_node_size);
+ bkey_btree_ptr_v2_init(&tmp.k);
+ bch2_alloc_sectors_append_ptrs(c, wp, &tmp.k, btree_sectors(c), false);
bch2_open_bucket_get(c, wp, &ob);
bch2_alloc_sectors_done(c, wp);
b = as->prealloc_nodes[--as->nr_prealloc_nodes];
set_btree_node_accessed(b);
- set_btree_node_dirty(b);
+ set_btree_node_dirty(c, b);
set_btree_node_need_write(b);
bch2_bset_init_first(b, &b->data->keys);
b->c.level = level;
b->c.btree_id = as->btree_id;
+ b->version_ondisk = c->sb.version;
memset(&b->nr, 0, sizeof(b->nr));
b->data->magic = cpu_to_le64(bset_magic(c));
+ memset(&b->data->_ptr, 0, sizeof(b->data->_ptr));
b->data->flags = 0;
SET_BTREE_NODE_ID(b->data, as->btree_id);
SET_BTREE_NODE_LEVEL(b->data, level);
- b->data->ptr = bch2_bkey_ptrs_c(bkey_i_to_s_c(&b->key)).start->ptr;
if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(&b->key);
bp->v.mem_ptr = 0;
bp->v.seq = b->data->keys.seq;
bp->v.sectors_written = 0;
- bp->v.sectors = cpu_to_le16(c->opts.btree_node_size);
}
- if (c->sb.features & (1ULL << BCH_FEATURE_new_extent_overwrite))
- SET_BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data, true);
-
- if (btree_node_is_extents(b) &&
- !BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data)) {
- set_btree_node_old_extent_overwrite(b);
- set_btree_node_need_rewrite(b);
- }
+ SET_BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data, true);
bch2_btree_build_aux_trees(b);
struct btree *b = bch2_btree_node_alloc(as, level);
btree_set_min(b, POS_MIN);
- btree_set_max(b, POS_MAX);
+ btree_set_max(b, SPOS_MAX);
b->data->format = bch2_btree_calc_format(b);
btree_node_set_format(b, b->data->format);
goto err_free;
}
- ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(&b->key));
- if (ret)
- goto err_free;
-
as->prealloc_nodes[as->nr_prealloc_nodes++] = b;
}
{
struct bch_fs *c = as->c;
+ if (as->took_gc_lock)
+ up_read(&c->gc_lock);
+ as->took_gc_lock = false;
+
bch2_journal_preres_put(&c->journal, &as->journal_preres);
bch2_journal_pin_drop(&c->journal, &as->journal);
bch2_disk_reservation_put(c, &as->disk_res);
bch2_btree_reserve_put(as);
+ bch2_time_stats_update(&c->times[BCH_TIME_btree_interior_update_total],
+ as->start_time);
+
mutex_lock(&c->btree_interior_update_lock);
list_del(&as->unwritten_list);
list_del(&as->list);
- mutex_unlock(&c->btree_interior_update_lock);
closure_debug_destroy(&as->cl);
mempool_free(as, &c->btree_interior_update_pool);
+ /*
+ * Have to do the wakeup with btree_interior_update_lock still held,
+ * since being on btree_interior_update_list is our ref on @c:
+ */
closure_wake_up(&c->btree_interior_update_wait);
+
+ mutex_unlock(&c->btree_interior_update_lock);
}
static void btree_update_will_delete_key(struct btree_update *as,
trans->journal_pin = &as->journal;
for_each_keylist_key(&as->new_keys, k) {
- ret = bch2_trans_mark_key(trans, bkey_i_to_s_c(k),
- 0, 0, BTREE_TRIGGER_INSERT);
+ ret = bch2_trans_mark_key(trans,
+ bkey_s_c_null,
+ bkey_i_to_s_c(k),
+ BTREE_TRIGGER_INSERT);
if (ret)
return ret;
}
for_each_keylist_key(&as->old_keys, k) {
- ret = bch2_trans_mark_key(trans, bkey_i_to_s_c(k),
- 0, 0, BTREE_TRIGGER_OVERWRITE);
+ ret = bch2_trans_mark_key(trans,
+ bkey_i_to_s_c(k),
+ bkey_s_c_null,
+ BTREE_TRIGGER_OVERWRITE);
if (ret)
return ret;
}
{
struct bch_fs *c = as->c;
struct btree *b = as->b;
+ struct btree_trans trans;
u64 journal_seq = 0;
unsigned i;
int ret;
+ /*
+ * If we're already in an error state, it might be because a btree node
+ * was never written, and we might be trying to free that same btree
+ * node here, but it won't have been marked as allocated and we'll see
+ * spurious disk usage inconsistencies in the transactional part below
+ * if we don't skip it:
+ */
+ ret = bch2_journal_error(&c->journal);
+ if (ret)
+ goto err;
+
+ BUG_ON(!journal_pin_active(&as->journal));
+
+ /*
+ * Wait for any in flight writes to finish before we free the old nodes
+ * on disk:
+ */
+ for (i = 0; i < as->nr_old_nodes; i++) {
+ struct btree *old = as->old_nodes[i];
+ __le64 seq;
+
+ six_lock_read(&old->c.lock, NULL, NULL);
+ seq = old->data ? old->data->keys.seq : 0;
+ six_unlock_read(&old->c.lock);
+
+ if (seq == as->old_nodes_seq[i])
+ wait_on_bit_io(&old->flags, BTREE_NODE_write_in_flight_inner,
+ TASK_UNINTERRUPTIBLE);
+ }
+
/*
* We did an update to a parent node where the pointers we added pointed
* to child nodes that weren't written yet: now, the child nodes have
* journal reclaim does btree updates when flushing bkey_cached entries,
* which may require allocations as well.
*/
- ret = bch2_trans_do(c, &as->disk_res, &journal_seq,
- BTREE_INSERT_NOFAIL|
- BTREE_INSERT_USE_RESERVE|
- BTREE_INSERT_USE_ALLOC_RESERVE|
- BTREE_INSERT_NOCHECK_RW|
- BTREE_INSERT_JOURNAL_RECLAIM|
- BTREE_INSERT_JOURNAL_RESERVED,
- btree_update_nodes_written_trans(&trans, as));
- BUG_ON(ret && !bch2_journal_error(&c->journal));
-
+ bch2_trans_init(&trans, c, 0, 512);
+ ret = __bch2_trans_do(&trans, &as->disk_res, &journal_seq,
+ BTREE_INSERT_NOFAIL|
+ BTREE_INSERT_NOCHECK_RW|
+ BTREE_INSERT_JOURNAL_RECLAIM|
+ BTREE_INSERT_JOURNAL_RESERVED,
+ btree_update_nodes_written_trans(&trans, as));
+ bch2_trans_exit(&trans);
+
+ bch2_fs_fatal_err_on(ret && !bch2_journal_error(&c->journal), c,
+ "error %i in btree_update_nodes_written()", ret);
+err:
if (b) {
/*
* @b is the node we did the final insert into:
list_del(&as->write_blocked_list);
- if (!ret && as->b == b) {
+ /*
+ * Node might have been freed, recheck under
+ * btree_interior_update_lock:
+ */
+ if (as->b == b) {
struct bset *i = btree_bset_last(b);
BUG_ON(!b->c.level);
BUG_ON(!btree_node_dirty(b));
- i->journal_seq = cpu_to_le64(
- max(journal_seq,
- le64_to_cpu(i->journal_seq)));
-
- bch2_btree_add_journal_pin(c, b, journal_seq);
+ if (!ret) {
+ i->journal_seq = cpu_to_le64(
+ max(journal_seq,
+ le64_to_cpu(i->journal_seq)));
+
+ bch2_btree_add_journal_pin(c, b, journal_seq);
+ } else {
+ /*
+ * If we didn't get a journal sequence number we
+ * can't write this btree node, because recovery
+ * won't know to ignore this write:
+ */
+ set_btree_node_never_write(b);
+ }
}
mutex_unlock(&c->btree_interior_update_lock);
child->b = NULL;
child->mode = BTREE_INTERIOR_UPDATING_AS;
- /*
- * When we write a new btree root, we have to drop our journal pin
- * _before_ the new nodes are technically reachable; see
- * btree_update_nodes_written().
- *
- * This goes for journal pins that are recursively blocked on us - so,
- * just transfer the journal pin to the new interior update so
- * btree_update_nodes_written() can drop it.
- */
bch2_journal_pin_copy(&c->journal, &as->journal, &child->journal, NULL);
- bch2_journal_pin_drop(&c->journal, &child->journal);
}
static void btree_update_updated_root(struct btree_update *as, struct btree *b)
* And it adds @b to the list of @as's new nodes, so that we can update sector
* counts in bch2_btree_update_nodes_written:
*/
-void bch2_btree_update_add_new_node(struct btree_update *as, struct btree *b)
+static void bch2_btree_update_add_new_node(struct btree_update *as, struct btree *b)
{
struct bch_fs *c = as->c;
closure_put(&as->cl);
}
-void bch2_btree_update_get_open_buckets(struct btree_update *as, struct btree *b)
+static void bch2_btree_update_get_open_buckets(struct btree_update *as, struct btree *b)
{
while (b->ob.nr)
as->open_buckets[as->nr_open_buckets++] =
* nodes and thus outstanding btree_updates - redirect @b's
* btree_updates to point to this btree_update:
*/
-void bch2_btree_interior_update_will_free_node(struct btree_update *as,
+static void bch2_btree_interior_update_will_free_node(struct btree_update *as,
struct btree *b)
{
struct bch_fs *c = as->c;
closure_wake_up(&c->btree_interior_update_wait);
}
- clear_btree_node_dirty(b);
+ clear_btree_node_dirty(c, b);
clear_btree_node_need_write(b);
/*
btree_update_drop_new_node(c, b);
btree_update_will_delete_key(as, &b->key);
+
+ as->old_nodes[as->nr_old_nodes] = b;
+ as->old_nodes_seq[as->nr_old_nodes] = b->data->keys.seq;
+ as->nr_old_nodes++;
}
-void bch2_btree_update_done(struct btree_update *as)
+static void bch2_btree_update_done(struct btree_update *as)
{
+ struct bch_fs *c = as->c;
+ u64 start_time = as->start_time;
+
BUG_ON(as->mode == BTREE_INTERIOR_NO_UPDATE);
+ if (as->took_gc_lock)
+ up_read(&as->c->gc_lock);
+ as->took_gc_lock = false;
+
bch2_btree_reserve_put(as);
- continue_at(&as->cl, btree_update_set_nodes_written, system_freezable_wq);
+ continue_at(&as->cl, btree_update_set_nodes_written,
+ as->c->btree_interior_update_worker);
+
+ bch2_time_stats_update(&c->times[BCH_TIME_btree_interior_update_foreground],
+ start_time);
}
-struct btree_update *
-bch2_btree_update_start(struct btree_trans *trans, enum btree_id id,
- unsigned nr_nodes, unsigned flags,
- struct closure *cl)
+static struct btree_update *
+bch2_btree_update_start(struct btree_trans *trans, struct btree_path *path,
+ unsigned level, unsigned nr_nodes, unsigned flags)
{
struct bch_fs *c = trans->c;
struct btree_update *as;
+ struct closure cl;
+ u64 start_time = local_clock();
int disk_res_flags = (flags & BTREE_INSERT_NOFAIL)
? BCH_DISK_RESERVATION_NOFAIL : 0;
- int journal_flags = (flags & BTREE_INSERT_JOURNAL_RESERVED)
- ? JOURNAL_RES_GET_RECLAIM : 0;
+ int journal_flags = 0;
int ret = 0;
+ BUG_ON(!path->should_be_locked);
+
+ if (flags & BTREE_INSERT_JOURNAL_RESERVED)
+ journal_flags |= JOURNAL_RES_GET_RESERVED;
+
+ closure_init_stack(&cl);
+retry:
+
/*
- * This check isn't necessary for correctness - it's just to potentially
- * prevent us from doing a lot of work that'll end up being wasted:
+ * XXX: figure out how far we might need to split,
+ * instead of locking/reserving all the way to the root:
*/
- ret = bch2_journal_error(&c->journal);
- if (ret)
+ if (!bch2_btree_path_upgrade(trans, path, U8_MAX)) {
+ trace_trans_restart_iter_upgrade(trans->fn, _RET_IP_,
+ path->btree_id, &path->pos);
+ ret = btree_trans_restart(trans);
return ERR_PTR(ret);
+ }
+
+ if (flags & BTREE_INSERT_GC_LOCK_HELD)
+ lockdep_assert_held(&c->gc_lock);
+ else if (!down_read_trylock(&c->gc_lock)) {
+ bch2_trans_unlock(trans);
+ down_read(&c->gc_lock);
+ if (!bch2_trans_relock(trans)) {
+ up_read(&c->gc_lock);
+ return ERR_PTR(-EINTR);
+ }
+ }
as = mempool_alloc(&c->btree_interior_update_pool, GFP_NOIO);
memset(as, 0, sizeof(*as));
closure_init(&as->cl, NULL);
as->c = c;
+ as->start_time = start_time;
as->mode = BTREE_INTERIOR_NO_UPDATE;
- as->btree_id = id;
+ as->took_gc_lock = !(flags & BTREE_INSERT_GC_LOCK_HELD);
+ as->btree_id = path->btree_id;
INIT_LIST_HEAD(&as->list);
INIT_LIST_HEAD(&as->unwritten_list);
INIT_LIST_HEAD(&as->write_blocked_list);
bch2_keylist_init(&as->new_keys, as->_new_keys);
bch2_keylist_init(&as->parent_keys, as->inline_keys);
+ mutex_lock(&c->btree_interior_update_lock);
+ list_add_tail(&as->list, &c->btree_interior_update_list);
+ mutex_unlock(&c->btree_interior_update_lock);
+
+ /*
+ * We don't want to allocate if we're in an error state, that can cause
+ * deadlock on emergency shutdown due to open buckets getting stuck in
+ * the btree_reserve_cache after allocator shutdown has cleared it out.
+ * This check needs to come after adding us to the btree_interior_update
+ * list but before calling bch2_btree_reserve_get, to synchronize with
+ * __bch2_fs_read_only().
+ */
+ ret = bch2_journal_error(&c->journal);
+ if (ret)
+ goto err;
+
ret = bch2_journal_preres_get(&c->journal, &as->journal_preres,
BTREE_UPDATE_JOURNAL_RES,
journal_flags|JOURNAL_RES_GET_NONBLOCK);
if (ret == -EAGAIN) {
- if (flags & BTREE_INSERT_NOUNLOCK)
- return ERR_PTR(-EINTR);
-
bch2_trans_unlock(trans);
+ if (flags & BTREE_INSERT_JOURNAL_RECLAIM) {
+ bch2_btree_update_free(as);
+ btree_trans_restart(trans);
+ return ERR_PTR(ret);
+ }
+
ret = bch2_journal_preres_get(&c->journal, &as->journal_preres,
BTREE_UPDATE_JOURNAL_RES,
journal_flags);
- if (ret)
- return ERR_PTR(ret);
+ if (ret) {
+ trace_trans_restart_journal_preres_get(trans->fn, _RET_IP_);
+ goto err;
+ }
if (!bch2_trans_relock(trans)) {
ret = -EINTR;
}
ret = bch2_disk_reservation_get(c, &as->disk_res,
- nr_nodes * c->opts.btree_node_size,
+ nr_nodes * btree_sectors(c),
c->opts.metadata_replicas,
disk_res_flags);
if (ret)
goto err;
- ret = bch2_btree_reserve_get(as, nr_nodes, flags, cl);
+ ret = bch2_btree_reserve_get(as, nr_nodes, flags, &cl);
if (ret)
goto err;
- mutex_lock(&c->btree_interior_update_lock);
- list_add_tail(&as->list, &c->btree_interior_update_list);
- mutex_unlock(&c->btree_interior_update_lock);
+ bch2_journal_pin_add(&c->journal,
+ atomic64_read(&c->journal.seq),
+ &as->journal, NULL);
return as;
err:
bch2_btree_update_free(as);
+
+ if (ret == -EAGAIN) {
+ bch2_trans_unlock(trans);
+ closure_sync(&cl);
+ ret = -EINTR;
+ }
+
+ if (ret == -EINTR && bch2_trans_relock(trans))
+ goto retry;
+
return ERR_PTR(ret);
}
list_del_init(&b->list);
mutex_unlock(&c->btree_cache.lock);
+ if (b->c.level)
+ six_lock_pcpu_alloc(&b->c.lock);
+ else
+ six_lock_pcpu_free(&b->c.lock);
+
mutex_lock(&c->btree_root_lock);
BUG_ON(btree_node_root(c, b) &&
(b->c.level < btree_node_root(c, b)->c.level ||
* is nothing new to be done. This just guarantees that there is a
* journal write.
*/
-static void bch2_btree_set_root(struct btree_update *as, struct btree *b,
- struct btree_iter *iter)
+static void bch2_btree_set_root(struct btree_update *as,
+ struct btree_trans *trans,
+ struct btree_path *path,
+ struct btree *b)
{
struct bch_fs *c = as->c;
struct btree *old;
* Ensure no one is using the old root while we switch to the
* new root:
*/
- bch2_btree_node_lock_write(old, iter);
+ bch2_btree_node_lock_write(trans, path, old);
bch2_btree_set_root_inmem(c, b);
* an intent lock on the new root, and any updates that would
* depend on the new root would have to update the new root.
*/
- bch2_btree_node_unlock_write(old, iter);
+ bch2_btree_node_unlock_write(trans, path, old);
}
/* Interior node updates: */
-static void bch2_insert_fixup_btree_ptr(struct btree_update *as, struct btree *b,
- struct btree_iter *iter,
- struct bkey_i *insert,
- struct btree_node_iter *node_iter)
+static void bch2_insert_fixup_btree_ptr(struct btree_update *as,
+ struct btree_trans *trans,
+ struct btree_path *path,
+ struct btree *b,
+ struct btree_node_iter *node_iter,
+ struct bkey_i *insert)
{
+ struct bch_fs *c = as->c;
struct bkey_packed *k;
+ const char *invalid;
+
+ BUG_ON(insert->k.type == KEY_TYPE_btree_ptr_v2 &&
+ !btree_ptr_sectors_written(insert));
+
+ if (unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags)))
+ bch2_journal_key_overwritten(c, b->c.btree_id, b->c.level, insert->k.p);
+
+ invalid = bch2_bkey_invalid(c, bkey_i_to_s_c(insert), btree_node_type(b)) ?:
+ bch2_bkey_in_btree_node(b, bkey_i_to_s_c(insert));
+ if (invalid) {
+ char buf[160];
+
+ bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(insert));
+ bch2_fs_inconsistent(c, "inserting invalid bkey %s: %s", buf, invalid);
+ dump_stack();
+ }
BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
ARRAY_SIZE(as->journal_entries));
bkey_iter_pos_cmp(b, k, &insert->k.p) < 0)
bch2_btree_node_iter_advance(node_iter, b);
- bch2_btree_bset_insert_key(iter, b, node_iter, insert);
- set_btree_node_dirty(b);
+ bch2_btree_bset_insert_key(trans, path, b, node_iter, insert);
+ set_btree_node_dirty(c, b);
set_btree_node_need_write(b);
}
+static void
+__bch2_btree_insert_keys_interior(struct btree_update *as,
+ struct btree_trans *trans,
+ struct btree_path *path,
+ struct btree *b,
+ struct btree_node_iter node_iter,
+ struct keylist *keys)
+{
+ struct bkey_i *insert = bch2_keylist_front(keys);
+ struct bkey_packed *k;
+
+ BUG_ON(btree_node_type(b) != BKEY_TYPE_btree);
+
+ while ((k = bch2_btree_node_iter_prev_all(&node_iter, b)) &&
+ (bkey_cmp_left_packed(b, k, &insert->k.p) >= 0))
+ ;
+
+ while (!bch2_keylist_empty(keys)) {
+ bch2_insert_fixup_btree_ptr(as, trans, path, b,
+ &node_iter, bch2_keylist_front(keys));
+ bch2_keylist_pop_front(keys);
+ }
+}
+
/*
* Move keys from n1 (original replacement node, now lower node) to n2 (higher
* node)
*/
static struct btree *__btree_split_node(struct btree_update *as,
- struct btree *n1,
- struct btree_iter *iter)
+ struct btree *n1)
{
+ struct bkey_format_state s;
size_t nr_packed = 0, nr_unpacked = 0;
struct btree *n2;
struct bset *set1, *set2;
- struct bkey_packed *k, *prev = NULL;
+ struct bkey_packed *k, *set2_start, *set2_end, *out, *prev = NULL;
+ struct bpos n1_pos;
n2 = bch2_btree_node_alloc(as, n1->c.level);
bch2_btree_update_add_new_node(as, n2);
SET_BTREE_NODE_SEQ(n2->data, BTREE_NODE_SEQ(n1->data));
n2->key.k.p = n1->key.k.p;
- btree_node_set_format(n2, n2->data->format);
-
set1 = btree_bset_first(n1);
set2 = btree_bset_first(n2);
*/
k = set1->start;
while (1) {
- struct bkey_packed *n = bkey_next_skip_noops(k, vstruct_last(set1));
+ struct bkey_packed *n = bkey_next(k);
if (n == vstruct_last(set1))
break;
}
BUG_ON(!prev);
+ set2_start = k;
+ set2_end = vstruct_last(set1);
- btree_set_max(n1, bkey_unpack_pos(n1, prev));
- btree_set_min(n2, bkey_successor(n1->key.k.p));
-
- set2->u64s = cpu_to_le16((u64 *) vstruct_end(set1) - (u64 *) k);
- set1->u64s = cpu_to_le16(le16_to_cpu(set1->u64s) - le16_to_cpu(set2->u64s));
-
+ set1->u64s = cpu_to_le16((u64 *) set2_start - set1->_data);
set_btree_bset_end(n1, n1->set);
- set_btree_bset_end(n2, n2->set);
-
- n2->nr.live_u64s = le16_to_cpu(set2->u64s);
- n2->nr.bset_u64s[0] = le16_to_cpu(set2->u64s);
- n2->nr.packed_keys = n1->nr.packed_keys - nr_packed;
- n2->nr.unpacked_keys = n1->nr.unpacked_keys - nr_unpacked;
n1->nr.live_u64s = le16_to_cpu(set1->u64s);
n1->nr.bset_u64s[0] = le16_to_cpu(set1->u64s);
n1->nr.packed_keys = nr_packed;
n1->nr.unpacked_keys = nr_unpacked;
+ n1_pos = bkey_unpack_pos(n1, prev);
+ if (as->c->sb.version < bcachefs_metadata_version_snapshot)
+ n1_pos.snapshot = U32_MAX;
+
+ btree_set_max(n1, n1_pos);
+ btree_set_min(n2, bpos_successor(n1->key.k.p));
+
+ bch2_bkey_format_init(&s);
+ bch2_bkey_format_add_pos(&s, n2->data->min_key);
+ bch2_bkey_format_add_pos(&s, n2->data->max_key);
+
+ for (k = set2_start; k != set2_end; k = bkey_next(k)) {
+ struct bkey uk = bkey_unpack_key(n1, k);
+ bch2_bkey_format_add_key(&s, &uk);
+ }
+
+ n2->data->format = bch2_bkey_format_done(&s);
+ btree_node_set_format(n2, n2->data->format);
+
+ out = set2->start;
+ memset(&n2->nr, 0, sizeof(n2->nr));
+
+ for (k = set2_start; k != set2_end; k = bkey_next(k)) {
+ BUG_ON(!bch2_bkey_transform(&n2->format, out, bkey_packed(k)
+ ? &n1->format : &bch2_bkey_format_current, k));
+ out->format = KEY_FORMAT_LOCAL_BTREE;
+ btree_keys_account_key_add(&n2->nr, 0, out);
+ out = bkey_next(out);
+ }
+
+ set2->u64s = cpu_to_le16((u64 *) out - set2->_data);
+ set_btree_bset_end(n2, n2->set);
+
BUG_ON(!set1->u64s);
BUG_ON(!set2->u64s);
- memcpy_u64s(set2->start,
- vstruct_end(set1),
- le16_to_cpu(set2->u64s));
-
btree_node_reset_sib_u64s(n1);
btree_node_reset_sib_u64s(n2);
* nodes that were coalesced, and thus in the middle of a child node post
* coalescing:
*/
-static void btree_split_insert_keys(struct btree_update *as, struct btree *b,
- struct btree_iter *iter,
+static void btree_split_insert_keys(struct btree_update *as,
+ struct btree_trans *trans,
+ struct btree_path *path,
+ struct btree *b,
struct keylist *keys)
{
struct btree_node_iter node_iter;
struct bkey_packed *src, *dst, *n;
struct bset *i;
- BUG_ON(btree_node_type(b) != BKEY_TYPE_BTREE);
-
bch2_btree_node_iter_init(&node_iter, b, &k->k.p);
- while (!bch2_keylist_empty(keys)) {
- k = bch2_keylist_front(keys);
-
- bch2_insert_fixup_btree_ptr(as, b, iter, k, &node_iter);
- bch2_keylist_pop_front(keys);
- }
+ __bch2_btree_insert_keys_interior(as, trans, path, b, node_iter, keys);
/*
* We can't tolerate whiteouts here - with whiteouts there can be
i = btree_bset_first(b);
src = dst = i->start;
while (src != vstruct_last(i)) {
- n = bkey_next_skip_noops(src, vstruct_last(i));
+ n = bkey_next(src);
if (!bkey_deleted(src)) {
memmove_u64s_down(dst, src, src->u64s);
dst = bkey_next(dst);
src = n;
}
+ /* Also clear out the unwritten whiteouts area: */
+ b->whiteout_u64s = 0;
+
i->u64s = cpu_to_le16((u64 *) dst - i->_data);
set_btree_bset_end(b, b->set);
btree_node_interior_verify(as->c, b);
}
-static void btree_split(struct btree_update *as, struct btree *b,
- struct btree_iter *iter, struct keylist *keys,
- unsigned flags)
+static void btree_split(struct btree_update *as, struct btree_trans *trans,
+ struct btree_path *path, struct btree *b,
+ struct keylist *keys, unsigned flags)
{
struct bch_fs *c = as->c;
- struct btree *parent = btree_node_parent(iter, b);
+ struct btree *parent = btree_node_parent(path, b);
struct btree *n1, *n2 = NULL, *n3 = NULL;
u64 start_time = local_clock();
BUG_ON(!parent && (b != btree_node_root(c, b)));
- BUG_ON(!btree_node_intent_locked(iter, btree_node_root(c, b)->c.level));
+ BUG_ON(!btree_node_intent_locked(path, btree_node_root(c, b)->c.level));
bch2_btree_interior_update_will_free_node(as, b);
bch2_btree_update_add_new_node(as, n1);
if (keys)
- btree_split_insert_keys(as, n1, iter, keys);
+ btree_split_insert_keys(as, trans, path, n1, keys);
if (bset_u64s(&n1->set[0]) > BTREE_SPLIT_THRESHOLD(c)) {
trace_btree_split(c, b);
- n2 = __btree_split_node(as, n1, iter);
+ n2 = __btree_split_node(as, n1);
bch2_btree_build_aux_trees(n2);
bch2_btree_build_aux_trees(n1);
six_unlock_write(&n2->c.lock);
six_unlock_write(&n1->c.lock);
+ bch2_btree_node_write(c, n1, SIX_LOCK_intent);
bch2_btree_node_write(c, n2, SIX_LOCK_intent);
/*
n3->sib_u64s[0] = U16_MAX;
n3->sib_u64s[1] = U16_MAX;
- btree_split_insert_keys(as, n3, iter, &as->parent_keys);
+ btree_split_insert_keys(as, trans, path, n3, &as->parent_keys);
bch2_btree_node_write(c, n3, SIX_LOCK_intent);
}
bch2_btree_build_aux_trees(n1);
six_unlock_write(&n1->c.lock);
+ bch2_btree_node_write(c, n1, SIX_LOCK_intent);
+
if (parent)
bch2_keylist_add(&as->parent_keys, &n1->key);
}
- bch2_btree_node_write(c, n1, SIX_LOCK_intent);
-
/* New nodes all written, now make them visible: */
if (parent) {
/* Split a non root node */
- bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
+ bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
} else if (n3) {
- bch2_btree_set_root(as, n3, iter);
+ bch2_btree_set_root(as, trans, path, n3);
} else {
/* Root filled up but didn't need to be split */
- bch2_btree_set_root(as, n1, iter);
+ bch2_btree_set_root(as, trans, path, n1);
}
bch2_btree_update_get_open_buckets(as, n1);
if (n3)
bch2_btree_update_get_open_buckets(as, n3);
- /* Successful split, update the iterator to point to the new nodes: */
+ /* Successful split, update the path to point to the new nodes: */
six_lock_increment(&b->c.lock, SIX_LOCK_intent);
- bch2_btree_iter_node_drop(iter, b);
if (n3)
- bch2_btree_iter_node_replace(iter, n3);
+ bch2_trans_node_add(trans, n3);
if (n2)
- bch2_btree_iter_node_replace(iter, n2);
- bch2_btree_iter_node_replace(iter, n1);
+ bch2_trans_node_add(trans, n2);
+ bch2_trans_node_add(trans, n1);
/*
* The old node must be freed (in memory) _before_ unlocking the new
* node after another thread has locked and updated the new node, thus
* seeing stale data:
*/
- bch2_btree_node_free_inmem(c, b, iter);
+ bch2_btree_node_free_inmem(trans, b);
if (n3)
six_unlock_intent(&n3->c.lock);
six_unlock_intent(&n2->c.lock);
six_unlock_intent(&n1->c.lock);
- bch2_btree_trans_verify_locks(iter->trans);
+ bch2_trans_verify_locks(trans);
- bch2_time_stats_update(&c->times[BCH_TIME_btree_node_split],
+ bch2_time_stats_update(&c->times[n2
+ ? BCH_TIME_btree_node_split
+ : BCH_TIME_btree_node_compact],
start_time);
}
static void
-bch2_btree_insert_keys_interior(struct btree_update *as, struct btree *b,
- struct btree_iter *iter, struct keylist *keys)
+bch2_btree_insert_keys_interior(struct btree_update *as,
+ struct btree_trans *trans,
+ struct btree_path *path,
+ struct btree *b,
+ struct keylist *keys)
{
- struct btree_iter *linked;
- struct btree_node_iter node_iter;
- struct bkey_i *insert = bch2_keylist_front(keys);
- struct bkey_packed *k;
-
- /* Don't screw up @iter's position: */
- node_iter = iter->l[b->c.level].iter;
+ struct btree_path *linked;
- /*
- * btree_split(), btree_gc_coalesce() will insert keys before
- * the iterator's current position - they know the keys go in
- * the node the iterator points to:
- */
- while ((k = bch2_btree_node_iter_prev_all(&node_iter, b)) &&
- (bkey_cmp_packed(b, k, &insert->k) >= 0))
- ;
-
- for_each_keylist_key(keys, insert)
- bch2_insert_fixup_btree_ptr(as, b, iter, insert, &node_iter);
+ __bch2_btree_insert_keys_interior(as, trans, path, b,
+ path->l[b->c.level].iter, keys);
btree_update_updated_node(as, b);
- trans_for_each_iter_with_node(iter->trans, b, linked)
+ trans_for_each_path_with_node(trans, b, linked)
bch2_btree_node_iter_peek(&linked->l[b->c.level].iter, b);
- bch2_btree_trans_verify_iters(iter->trans, b);
+ bch2_trans_verify_paths(trans);
}
/**
* If a split occurred, this function will return early. This can only happen
* for leaf nodes -- inserts into interior nodes have to be atomic.
*/
-void bch2_btree_insert_node(struct btree_update *as, struct btree *b,
- struct btree_iter *iter, struct keylist *keys,
- unsigned flags)
+static void bch2_btree_insert_node(struct btree_update *as, struct btree_trans *trans,
+ struct btree_path *path, struct btree *b,
+ struct keylist *keys, unsigned flags)
{
struct bch_fs *c = as->c;
int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s);
int old_live_u64s = b->nr.live_u64s;
int live_u64s_added, u64s_added;
- BUG_ON(!btree_node_intent_locked(iter, btree_node_root(c, b)->c.level));
+ lockdep_assert_held(&c->gc_lock);
+ BUG_ON(!btree_node_intent_locked(path, btree_node_root(c, b)->c.level));
BUG_ON(!b->c.level);
BUG_ON(!as || as->b);
bch2_verify_keylist_sorted(keys);
- if (as->must_rewrite)
- goto split;
-
- bch2_btree_node_lock_for_insert(c, b, iter);
+ bch2_btree_node_lock_for_insert(trans, path, b);
if (!bch2_btree_node_insert_fits(c, b, bch2_keylist_u64s(keys))) {
- bch2_btree_node_unlock_write(b, iter);
+ bch2_btree_node_unlock_write(trans, path, b);
goto split;
}
- bch2_btree_insert_keys_interior(as, b, iter, keys);
+ btree_node_interior_verify(c, b);
+
+ bch2_btree_insert_keys_interior(as, trans, path, b, keys);
live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_u64s;
if (u64s_added > live_u64s_added &&
bch2_maybe_compact_whiteouts(c, b))
- bch2_btree_iter_reinit_node(iter, b);
+ bch2_trans_node_reinit_iter(trans, b);
- bch2_btree_node_unlock_write(b, iter);
+ bch2_btree_node_unlock_write(trans, path, b);
btree_node_interior_verify(c, b);
-
- /*
- * when called from the btree_split path the new nodes aren't added to
- * the btree iterator yet, so the merge path's unlock/wait/relock dance
- * won't work:
- */
- bch2_foreground_maybe_merge(c, iter, b->c.level,
- flags|BTREE_INSERT_NOUNLOCK);
return;
split:
- btree_split(as, b, iter, keys, flags);
+ btree_split(as, trans, path, b, keys, flags);
}
-int bch2_btree_split_leaf(struct bch_fs *c, struct btree_iter *iter,
+int bch2_btree_split_leaf(struct btree_trans *trans,
+ struct btree_path *path,
unsigned flags)
{
- struct btree_trans *trans = iter->trans;
- struct btree *b = iter_l(iter)->b;
+ struct bch_fs *c = trans->c;
+ struct btree *b = path_l(path)->b;
struct btree_update *as;
- struct closure cl;
+ unsigned l;
int ret = 0;
- struct btree_insert_entry *i;
-
- /*
- * We already have a disk reservation and open buckets pinned; this
- * allocation must not block:
- */
- trans_for_each_update(trans, i)
- if (btree_node_type_needs_gc(i->iter->btree_id))
- flags |= BTREE_INSERT_USE_RESERVE;
- closure_init_stack(&cl);
-
- /* Hack, because gc and splitting nodes doesn't mix yet: */
- if (!(flags & BTREE_INSERT_GC_LOCK_HELD) &&
- !down_read_trylock(&c->gc_lock)) {
- if (flags & BTREE_INSERT_NOUNLOCK) {
- trace_transaction_restart_ip(trans->ip, _THIS_IP_);
- return -EINTR;
- }
-
- bch2_trans_unlock(trans);
- down_read(&c->gc_lock);
-
- if (!bch2_trans_relock(trans))
- ret = -EINTR;
- }
-
- /*
- * XXX: figure out how far we might need to split,
- * instead of locking/reserving all the way to the root:
- */
- if (!bch2_btree_iter_upgrade(iter, U8_MAX)) {
- trace_trans_restart_iter_upgrade(trans->ip);
- ret = -EINTR;
- goto out;
- }
-
- as = bch2_btree_update_start(trans, iter->btree_id,
- btree_update_reserve_required(c, b), flags,
- !(flags & BTREE_INSERT_NOUNLOCK) ? &cl : NULL);
- if (IS_ERR(as)) {
- ret = PTR_ERR(as);
- if (ret == -EAGAIN) {
- BUG_ON(flags & BTREE_INSERT_NOUNLOCK);
- bch2_trans_unlock(trans);
- ret = -EINTR;
-
- trace_transaction_restart_ip(trans->ip, _THIS_IP_);
- }
- goto out;
- }
+ as = bch2_btree_update_start(trans, path, path->level,
+ btree_update_reserve_required(c, b), flags);
+ if (IS_ERR(as))
+ return PTR_ERR(as);
- btree_split(as, b, iter, NULL, flags);
+ btree_split(as, trans, path, b, NULL, flags);
bch2_btree_update_done(as);
- /*
- * We haven't successfully inserted yet, so don't downgrade all the way
- * back to read locks;
- */
- __bch2_btree_iter_downgrade(iter, 1);
-out:
- if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
- up_read(&c->gc_lock);
- closure_sync(&cl);
+ for (l = path->level + 1; btree_path_node(path, l) && !ret; l++)
+ ret = bch2_foreground_maybe_merge(trans, path, l, flags);
+
return ret;
}
-void __bch2_foreground_maybe_merge(struct bch_fs *c,
- struct btree_iter *iter,
- unsigned level,
- unsigned flags,
- enum btree_node_sibling sib)
+int __bch2_foreground_maybe_merge(struct btree_trans *trans,
+ struct btree_path *path,
+ unsigned level,
+ unsigned flags,
+ enum btree_node_sibling sib)
{
- struct btree_trans *trans = iter->trans;
+ struct bch_fs *c = trans->c;
+ struct btree_path *sib_path = NULL;
struct btree_update *as;
struct bkey_format_state new_s;
struct bkey_format new_f;
struct bkey_i delete;
struct btree *b, *m, *n, *prev, *next, *parent;
- struct closure cl;
+ struct bpos sib_pos;
size_t sib_u64s;
+ u64 start_time = local_clock();
int ret = 0;
- BUG_ON(!btree_node_locked(iter, level));
-
- closure_init_stack(&cl);
-retry:
- BUG_ON(!btree_node_locked(iter, level));
+ BUG_ON(!path->should_be_locked);
+ BUG_ON(!btree_node_locked(path, level));
- b = iter->l[level].b;
+ b = path->l[level].b;
- parent = btree_node_parent(iter, b);
- if (!parent)
- goto out;
+ if ((sib == btree_prev_sib && !bpos_cmp(b->data->min_key, POS_MIN)) ||
+ (sib == btree_next_sib && !bpos_cmp(b->data->max_key, SPOS_MAX))) {
+ b->sib_u64s[sib] = U16_MAX;
+ return 0;
+ }
- if (b->sib_u64s[sib] > BTREE_FOREGROUND_MERGE_THRESHOLD(c))
- goto out;
+ sib_pos = sib == btree_prev_sib
+ ? bpos_predecessor(b->data->min_key)
+ : bpos_successor(b->data->max_key);
- /* XXX: can't be holding read locks */
- m = bch2_btree_node_get_sibling(c, iter, b, sib);
- if (IS_ERR(m)) {
- ret = PTR_ERR(m);
+ sib_path = bch2_path_get(trans, path->btree_id, sib_pos,
+ U8_MAX, level, BTREE_ITER_INTENT, _THIS_IP_);
+ ret = bch2_btree_path_traverse(trans, sib_path, false);
+ if (ret)
goto err;
- }
- /* NULL means no sibling: */
- if (!m) {
+ sib_path->should_be_locked = true;
+
+ m = sib_path->l[level].b;
+
+ if (btree_node_parent(path, b) !=
+ btree_node_parent(sib_path, m)) {
b->sib_u64s[sib] = U16_MAX;
goto out;
}
next = m;
}
+ if (bkey_cmp(bpos_successor(prev->data->max_key), next->data->min_key)) {
+ char buf1[100], buf2[100];
+
+ bch2_bpos_to_text(&PBUF(buf1), prev->data->max_key);
+ bch2_bpos_to_text(&PBUF(buf2), next->data->min_key);
+ bch_err(c,
+ "btree topology error in btree merge:\n"
+ " prev ends at %s\n"
+ " next starts at %s",
+ buf1, buf2);
+ bch2_topology_error(c);
+ ret = -EIO;
+ goto err;
+ }
+
bch2_bkey_format_init(&new_s);
- __bch2_btree_calc_format(&new_s, b);
- __bch2_btree_calc_format(&new_s, m);
+ bch2_bkey_format_add_pos(&new_s, prev->data->min_key);
+ __bch2_btree_calc_format(&new_s, prev);
+ __bch2_btree_calc_format(&new_s, next);
+ bch2_bkey_format_add_pos(&new_s, next->data->max_key);
new_f = bch2_bkey_format_done(&new_s);
sib_u64s = btree_node_u64s_with_format(b, &new_f) +
}
sib_u64s = min(sib_u64s, btree_max_u64s(c));
+ sib_u64s = min(sib_u64s, (size_t) U16_MAX - 1);
b->sib_u64s[sib] = sib_u64s;
- if (b->sib_u64s[sib] > BTREE_FOREGROUND_MERGE_THRESHOLD(c)) {
- six_unlock_intent(&m->c.lock);
+ if (b->sib_u64s[sib] > c->btree_foreground_merge_threshold)
goto out;
- }
-
- /* We're changing btree topology, doesn't mix with gc: */
- if (!(flags & BTREE_INSERT_GC_LOCK_HELD) &&
- !down_read_trylock(&c->gc_lock))
- goto err_cycle_gc_lock;
- if (!bch2_btree_iter_upgrade(iter, U8_MAX)) {
- ret = -EINTR;
- goto err_unlock;
- }
-
- as = bch2_btree_update_start(trans, iter->btree_id,
+ parent = btree_node_parent(path, b);
+ as = bch2_btree_update_start(trans, path, level,
btree_update_reserve_required(c, parent) + 1,
flags|
BTREE_INSERT_NOFAIL|
- BTREE_INSERT_USE_RESERVE,
- !(flags & BTREE_INSERT_NOUNLOCK) ? &cl : NULL);
- if (IS_ERR(as)) {
- ret = PTR_ERR(as);
- goto err_unlock;
- }
+ BTREE_INSERT_USE_RESERVE);
+ ret = PTR_ERR_OR_ZERO(as);
+ if (ret)
+ goto err;
trace_btree_merge(c, b);
bch2_btree_build_aux_trees(n);
six_unlock_write(&n->c.lock);
+ bch2_btree_node_write(c, n, SIX_LOCK_intent);
+
bkey_init(&delete.k);
delete.k.p = prev->key.k.p;
bch2_keylist_add(&as->parent_keys, &delete);
bch2_keylist_add(&as->parent_keys, &n->key);
- bch2_btree_node_write(c, n, SIX_LOCK_intent);
+ bch2_trans_verify_paths(trans);
+
+ bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
- bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
+ bch2_trans_verify_paths(trans);
bch2_btree_update_get_open_buckets(as, n);
six_lock_increment(&b->c.lock, SIX_LOCK_intent);
- bch2_btree_iter_node_drop(iter, b);
- bch2_btree_iter_node_drop(iter, m);
+ six_lock_increment(&m->c.lock, SIX_LOCK_intent);
- bch2_btree_iter_node_replace(iter, n);
+ bch2_trans_node_add(trans, n);
- bch2_btree_trans_verify_iters(trans, n);
+ bch2_trans_verify_paths(trans);
- bch2_btree_node_free_inmem(c, b, iter);
- bch2_btree_node_free_inmem(c, m, iter);
+ bch2_btree_node_free_inmem(trans, b);
+ bch2_btree_node_free_inmem(trans, m);
six_unlock_intent(&n->c.lock);
bch2_btree_update_done(as);
- if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
- up_read(&c->gc_lock);
+ bch2_time_stats_update(&c->times[BCH_TIME_btree_node_merge], start_time);
out:
- bch2_btree_trans_verify_locks(trans);
-
- /*
- * Don't downgrade locks here: we're called after successful insert,
- * and the caller will downgrade locks after a successful insert
- * anyways (in case e.g. a split was required first)
- *
- * And we're also called when inserting into interior nodes in the
- * split path, and downgrading to read locks in there is potentially
- * confusing:
- */
- closure_sync(&cl);
- return;
-
-err_cycle_gc_lock:
- six_unlock_intent(&m->c.lock);
-
- if (flags & BTREE_INSERT_NOUNLOCK)
- goto out;
-
- bch2_trans_unlock(trans);
-
- down_read(&c->gc_lock);
- up_read(&c->gc_lock);
- ret = -EINTR;
- goto err;
-
-err_unlock:
- six_unlock_intent(&m->c.lock);
- if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
- up_read(&c->gc_lock);
err:
- BUG_ON(ret == -EAGAIN && (flags & BTREE_INSERT_NOUNLOCK));
-
- if ((ret == -EAGAIN || ret == -EINTR) &&
- !(flags & BTREE_INSERT_NOUNLOCK)) {
- bch2_trans_unlock(trans);
- closure_sync(&cl);
- ret = bch2_btree_iter_traverse(iter);
- if (ret)
- goto out;
-
- goto retry;
- }
-
- goto out;
+ bch2_path_put(trans, sib_path, true);
+ bch2_trans_verify_locks(trans);
+ return ret;
}
-static int __btree_node_rewrite(struct bch_fs *c, struct btree_iter *iter,
- struct btree *b, unsigned flags,
- struct closure *cl)
+/**
+ * bch_btree_node_rewrite - Rewrite/move a btree node
+ */
+int bch2_btree_node_rewrite(struct btree_trans *trans,
+ struct btree_iter *iter,
+ struct btree *b,
+ unsigned flags)
{
- struct btree *n, *parent = btree_node_parent(iter, b);
+ struct bch_fs *c = trans->c;
+ struct btree *n, *parent;
struct btree_update *as;
+ int ret;
+
+ flags |= BTREE_INSERT_NOFAIL;
- as = bch2_btree_update_start(iter->trans, iter->btree_id,
+ parent = btree_node_parent(iter->path, b);
+ as = bch2_btree_update_start(trans, iter->path, b->c.level,
(parent
? btree_update_reserve_required(c, parent)
: 0) + 1,
- flags, cl);
- if (IS_ERR(as)) {
+ flags);
+ ret = PTR_ERR_OR_ZERO(as);
+ if (ret) {
trace_btree_gc_rewrite_node_fail(c, b);
- return PTR_ERR(as);
+ goto out;
}
bch2_btree_interior_update_will_free_node(as, b);
if (parent) {
bch2_keylist_add(&as->parent_keys, &n->key);
- bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
+ bch2_btree_insert_node(as, trans, iter->path, parent,
+ &as->parent_keys, flags);
} else {
- bch2_btree_set_root(as, n, iter);
+ bch2_btree_set_root(as, trans, iter->path, n);
}
bch2_btree_update_get_open_buckets(as, n);
six_lock_increment(&b->c.lock, SIX_LOCK_intent);
- bch2_btree_iter_node_drop(iter, b);
- bch2_btree_iter_node_replace(iter, n);
- bch2_btree_node_free_inmem(c, b, iter);
+ bch2_trans_node_add(trans, n);
+ bch2_btree_node_free_inmem(trans, b);
six_unlock_intent(&n->c.lock);
bch2_btree_update_done(as);
- return 0;
+out:
+ bch2_btree_path_downgrade(iter->path);
+ return ret;
}
-/**
- * bch_btree_node_rewrite - Rewrite/move a btree node
- *
- * Returns 0 on success, -EINTR or -EAGAIN on failure (i.e.
- * btree_check_reserve() has to wait)
- */
-int bch2_btree_node_rewrite(struct bch_fs *c, struct btree_iter *iter,
- __le64 seq, unsigned flags)
+struct async_btree_rewrite {
+ struct bch_fs *c;
+ struct work_struct work;
+ enum btree_id btree_id;
+ unsigned level;
+ struct bpos pos;
+ __le64 seq;
+};
+
+static int async_btree_node_rewrite_trans(struct btree_trans *trans,
+ struct async_btree_rewrite *a)
{
- struct btree_trans *trans = iter->trans;
- struct closure cl;
+ struct btree_iter iter;
struct btree *b;
int ret;
- flags |= BTREE_INSERT_NOFAIL;
+ bch2_trans_node_iter_init(trans, &iter, a->btree_id, a->pos,
+ BTREE_MAX_DEPTH, a->level, 0);
+ b = bch2_btree_iter_peek_node(&iter);
+ ret = PTR_ERR_OR_ZERO(b);
+ if (ret)
+ goto out;
- closure_init_stack(&cl);
+ if (!b || b->data->keys.seq != a->seq)
+ goto out;
- bch2_btree_iter_upgrade(iter, U8_MAX);
+ ret = bch2_btree_node_rewrite(trans, &iter, b, 0);
+out :
+ bch2_trans_iter_exit(trans, &iter);
- if (!(flags & BTREE_INSERT_GC_LOCK_HELD)) {
- if (!down_read_trylock(&c->gc_lock)) {
- bch2_trans_unlock(trans);
- down_read(&c->gc_lock);
- }
- }
+ return ret;
+}
- while (1) {
- ret = bch2_btree_iter_traverse(iter);
- if (ret)
- break;
+void async_btree_node_rewrite_work(struct work_struct *work)
+{
+ struct async_btree_rewrite *a =
+ container_of(work, struct async_btree_rewrite, work);
+ struct bch_fs *c = a->c;
+
+ bch2_trans_do(c, NULL, NULL, 0,
+ async_btree_node_rewrite_trans(&trans, a));
+ percpu_ref_put(&c->writes);
+ kfree(a);
+}
- b = bch2_btree_iter_peek_node(iter);
- if (!b || b->data->keys.seq != seq)
- break;
+void bch2_btree_node_rewrite_async(struct bch_fs *c, struct btree *b)
+{
+ struct async_btree_rewrite *a;
- ret = __btree_node_rewrite(c, iter, b, flags, &cl);
- if (ret != -EAGAIN &&
- ret != -EINTR)
- break;
+ if (!percpu_ref_tryget(&c->writes))
+ return;
- bch2_trans_unlock(trans);
- closure_sync(&cl);
+ a = kmalloc(sizeof(*a), GFP_NOFS);
+ if (!a) {
+ percpu_ref_put(&c->writes);
+ return;
}
- bch2_btree_iter_downgrade(iter);
-
- if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
- up_read(&c->gc_lock);
+ a->c = c;
+ a->btree_id = b->c.btree_id;
+ a->level = b->c.level;
+ a->pos = b->key.k.p;
+ a->seq = b->data->keys.seq;
- closure_sync(&cl);
- return ret;
+ INIT_WORK(&a->work, async_btree_node_rewrite_work);
+ queue_work(c->btree_interior_update_worker, &a->work);
}
-static void __bch2_btree_node_update_key(struct bch_fs *c,
- struct btree_update *as,
- struct btree_iter *iter,
- struct btree *b, struct btree *new_hash,
- struct bkey_i *new_key)
+static int __bch2_btree_node_update_key(struct btree_trans *trans,
+ struct btree_iter *iter,
+ struct btree *b, struct btree *new_hash,
+ struct bkey_i *new_key,
+ bool skip_triggers)
{
+ struct bch_fs *c = trans->c;
+ struct btree_iter iter2 = { NULL };
struct btree *parent;
+ u64 journal_entries[BKEY_BTREE_PTR_U64s_MAX];
int ret;
- btree_update_will_delete_key(as, &b->key);
- btree_update_will_add_key(as, new_key);
+ if (!skip_triggers) {
+ ret = bch2_trans_mark_key(trans,
+ bkey_s_c_null,
+ bkey_i_to_s_c(new_key),
+ BTREE_TRIGGER_INSERT);
+ if (ret)
+ return ret;
+
+ ret = bch2_trans_mark_key(trans,
+ bkey_i_to_s_c(&b->key),
+ bkey_s_c_null,
+ BTREE_TRIGGER_OVERWRITE);
+ if (ret)
+ return ret;
+ }
+
+ if (new_hash) {
+ bkey_copy(&new_hash->key, new_key);
+ ret = bch2_btree_node_hash_insert(&c->btree_cache,
+ new_hash, b->c.level, b->c.btree_id);
+ BUG_ON(ret);
+ }
- parent = btree_node_parent(iter, b);
+ parent = btree_node_parent(iter->path, b);
if (parent) {
- if (new_hash) {
- bkey_copy(&new_hash->key, new_key);
- ret = bch2_btree_node_hash_insert(&c->btree_cache,
- new_hash, b->c.level, b->c.btree_id);
- BUG_ON(ret);
- }
+ bch2_trans_copy_iter(&iter2, iter);
- bch2_keylist_add(&as->parent_keys, new_key);
- bch2_btree_insert_node(as, parent, iter, &as->parent_keys, 0);
+ iter2.path = bch2_btree_path_make_mut(trans, iter2.path,
+ iter2.flags & BTREE_ITER_INTENT,
+ _THIS_IP_);
- if (new_hash) {
- mutex_lock(&c->btree_cache.lock);
- bch2_btree_node_hash_remove(&c->btree_cache, new_hash);
+ BUG_ON(iter2.path->level != b->c.level);
+ BUG_ON(bpos_cmp(iter2.path->pos, new_key->k.p));
- bch2_btree_node_hash_remove(&c->btree_cache, b);
+ btree_node_unlock(iter2.path, iter2.path->level);
+ path_l(iter2.path)->b = BTREE_ITER_NO_NODE_UP;
+ iter2.path->level++;
- bkey_copy(&b->key, new_key);
- ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
- BUG_ON(ret);
- mutex_unlock(&c->btree_cache.lock);
- } else {
- bkey_copy(&b->key, new_key);
- }
+ ret = bch2_btree_iter_traverse(&iter2) ?:
+ bch2_trans_update(trans, &iter2, new_key, BTREE_TRIGGER_NORUN);
+ if (ret)
+ goto err;
} else {
BUG_ON(btree_node_root(c, b) != b);
- bch2_btree_node_lock_write(b, iter);
- bkey_copy(&b->key, new_key);
+ trans->extra_journal_entries = (void *) &journal_entries[0];
+ trans->extra_journal_entry_u64s =
+ journal_entry_set((void *) &journal_entries[0],
+ BCH_JSET_ENTRY_btree_root,
+ b->c.btree_id, b->c.level,
+ new_key, new_key->k.u64s);
+ }
- if (btree_ptr_hash_val(&b->key) != b->hash_val) {
- mutex_lock(&c->btree_cache.lock);
- bch2_btree_node_hash_remove(&c->btree_cache, b);
+ ret = bch2_trans_commit(trans, NULL, NULL,
+ BTREE_INSERT_NOFAIL|
+ BTREE_INSERT_NOCHECK_RW|
+ BTREE_INSERT_JOURNAL_RECLAIM|
+ BTREE_INSERT_JOURNAL_RESERVED);
+ if (ret)
+ goto err;
- ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
- BUG_ON(ret);
- mutex_unlock(&c->btree_cache.lock);
- }
+ bch2_btree_node_lock_write(trans, iter->path, b);
- btree_update_updated_root(as, b);
- bch2_btree_node_unlock_write(b, iter);
+ if (new_hash) {
+ mutex_lock(&c->btree_cache.lock);
+ bch2_btree_node_hash_remove(&c->btree_cache, new_hash);
+ bch2_btree_node_hash_remove(&c->btree_cache, b);
+
+ bkey_copy(&b->key, new_key);
+ ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
+ BUG_ON(ret);
+ mutex_unlock(&c->btree_cache.lock);
+ } else {
+ bkey_copy(&b->key, new_key);
}
- bch2_btree_update_done(as);
+ bch2_btree_node_unlock_write(trans, iter->path, b);
+out:
+ bch2_trans_iter_exit(trans, &iter2);
+ return ret;
+err:
+ if (new_hash) {
+ mutex_lock(&c->btree_cache.lock);
+ bch2_btree_node_hash_remove(&c->btree_cache, b);
+ mutex_unlock(&c->btree_cache.lock);
+ }
+ goto out;
}
-int bch2_btree_node_update_key(struct bch_fs *c, struct btree_iter *iter,
- struct btree *b,
- struct bkey_i *new_key)
+int bch2_btree_node_update_key(struct btree_trans *trans, struct btree_iter *iter,
+ struct btree *b, struct bkey_i *new_key,
+ bool skip_triggers)
{
- struct btree *parent = btree_node_parent(iter, b);
- struct btree_update *as = NULL;
+ struct bch_fs *c = trans->c;
struct btree *new_hash = NULL;
+ struct btree_path *path = iter->path;
struct closure cl;
- int ret;
-
- closure_init_stack(&cl);
+ int ret = 0;
- if (!bch2_btree_iter_upgrade(iter, U8_MAX))
+ if (!btree_node_intent_locked(path, b->c.level) &&
+ !bch2_btree_path_upgrade(trans, path, b->c.level + 1)) {
+ btree_trans_restart(trans);
return -EINTR;
-
- if (!down_read_trylock(&c->gc_lock)) {
- bch2_trans_unlock(iter->trans);
- down_read(&c->gc_lock);
-
- if (!bch2_trans_relock(iter->trans)) {
- ret = -EINTR;
- goto err;
- }
}
+ closure_init_stack(&cl);
+
/*
* check btree_ptr_hash_val() after @b is locked by
* btree_iter_traverse():
*/
if (btree_ptr_hash_val(new_key) != b->hash_val) {
- /* bch2_btree_reserve_get will unlock */
ret = bch2_btree_cache_cannibalize_lock(c, &cl);
if (ret) {
- bch2_trans_unlock(iter->trans);
- up_read(&c->gc_lock);
+ bch2_trans_unlock(trans);
closure_sync(&cl);
- down_read(&c->gc_lock);
-
- if (!bch2_trans_relock(iter->trans)) {
- ret = -EINTR;
- goto err;
- }
+ if (!bch2_trans_relock(trans))
+ return -EINTR;
}
new_hash = bch2_btree_node_mem_alloc(c);
}
-retry:
- as = bch2_btree_update_start(iter->trans, iter->btree_id,
- parent ? btree_update_reserve_required(c, parent) : 0,
- BTREE_INSERT_NOFAIL|
- BTREE_INSERT_USE_RESERVE|
- BTREE_INSERT_USE_ALLOC_RESERVE,
- &cl);
-
- if (IS_ERR(as)) {
- ret = PTR_ERR(as);
- if (ret == -EAGAIN)
- ret = -EINTR;
-
- if (ret == -EINTR) {
- bch2_trans_unlock(iter->trans);
- up_read(&c->gc_lock);
- closure_sync(&cl);
- down_read(&c->gc_lock);
-
- if (bch2_trans_relock(iter->trans))
- goto retry;
- }
-
- goto err;
- }
- ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(new_key));
- if (ret)
- goto err_free_update;
+ path->intent_ref++;
+ ret = __bch2_btree_node_update_key(trans, iter, b, new_hash,
+ new_key, skip_triggers);
+ --path->intent_ref;
- __bch2_btree_node_update_key(c, as, iter, b, new_hash, new_key);
-
- bch2_btree_iter_downgrade(iter);
-err:
if (new_hash) {
mutex_lock(&c->btree_cache.lock);
list_move(&new_hash->list, &c->btree_cache.freeable);
six_unlock_write(&new_hash->c.lock);
six_unlock_intent(&new_hash->c.lock);
}
- up_read(&c->gc_lock);
closure_sync(&cl);
+ bch2_btree_cache_cannibalize_unlock(c);
+ return ret;
+}
+
+int bch2_btree_node_update_key_get_iter(struct btree_trans *trans,
+ struct btree *b, struct bkey_i *new_key,
+ bool skip_triggers)
+{
+ struct btree_iter iter;
+ int ret;
+
+ bch2_trans_node_iter_init(trans, &iter, b->c.btree_id, b->key.k.p,
+ BTREE_MAX_DEPTH, b->c.level,
+ BTREE_ITER_INTENT);
+ ret = bch2_btree_iter_traverse(&iter);
+ if (ret)
+ goto out;
+
+ /* has node been freed? */
+ if (iter.path->l[b->c.level].b != b) {
+ /* node has been freed: */
+ BUG_ON(!btree_node_dying(b));
+ goto out;
+ }
+
+ BUG_ON(!btree_node_hashed(b));
+
+ ret = bch2_btree_node_update_key(trans, &iter, b, new_key, skip_triggers);
+out:
+ bch2_trans_iter_exit(trans, &iter);
return ret;
-err_free_update:
- bch2_btree_update_free(as);
- goto err;
}
/* Init code: */
b->c.btree_id = id;
bkey_btree_ptr_init(&b->key);
- b->key.k.p = POS_MAX;
+ b->key.k.p = SPOS_MAX;
*((u64 *) bkey_i_to_btree_ptr(&b->key)->v.start) = U64_MAX - id;
bch2_bset_init_first(b, &b->data->keys);
b->data->flags = 0;
btree_set_min(b, POS_MIN);
- btree_set_max(b, POS_MAX);
+ btree_set_max(b, SPOS_MAX);
b->data->format = bch2_btree_calc_format(b);
btree_node_set_format(b, b->data->format);