X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=libbcachefs%2Fbtree_update_leaf.c;h=c93c132dd815d963360dfe5060dd74925e3ba915;hb=1ee7dc7a55273d34358a0ee525a9e823c999ffe6;hp=a2b37dd45cf8f6d5132a9a9d224c2f5fc1a3c679;hpb=71111771690f244d13650c73d52ff601ad914d95;p=bcachefs-tools-debian diff --git a/libbcachefs/btree_update_leaf.c b/libbcachefs/btree_update_leaf.c index a2b37dd..c93c132 100644 --- a/libbcachefs/btree_update_leaf.c +++ b/libbcachefs/btree_update_leaf.c @@ -8,6 +8,7 @@ #include "btree_iter.h" #include "btree_key_cache.h" #include "btree_locking.h" +#include "btree_write_buffer.h" #include "buckets.h" #include "debug.h" #include "errcode.h" @@ -24,12 +25,28 @@ #include #include +/* + * bch2_btree_path_peek_slot() for a cached iterator might return a key in a + * different snapshot: + */ +struct bkey_s_c bch2_btree_path_peek_slot_exact(struct btree_path *path, struct bkey *u) +{ + struct bkey_s_c k = bch2_btree_path_peek_slot(path, u); + + if (k.k && bpos_eq(path->pos, k.k->p)) + return k; + + bkey_init(u); + u->p = path->pos; + return (struct bkey_s_c) { u, NULL }; +} + static void verify_update_old_key(struct btree_trans *trans, struct btree_insert_entry *i) { #ifdef CONFIG_BCACHEFS_DEBUG struct bch_fs *c = trans->c; struct bkey u; - struct bkey_s_c k = bch2_btree_path_peek_slot(i->path, &u); + struct bkey_s_c k = bch2_btree_path_peek_slot_exact(i->path, &u); if (unlikely(trans->journal_replay_not_finished)) { struct bkey_i *j_k = @@ -39,9 +56,10 @@ static void verify_update_old_key(struct btree_trans *trans, struct btree_insert k = bkey_i_to_s_c(j_k); } - i->old_k.needs_whiteout = k.k->needs_whiteout; + u = *k.k; + u.needs_whiteout = i->old_k.needs_whiteout; - BUG_ON(memcmp(&i->old_k, k.k, sizeof(struct bkey))); + BUG_ON(memcmp(&i->old_k, &u, sizeof(struct bkey))); BUG_ON(i->old_v != k.v); #endif } @@ -84,9 +102,6 @@ inline void bch2_btree_node_prep_for_write(struct btree_trans *trans, { struct bch_fs *c = trans->c; - if (path->cached) - return; - if (unlikely(btree_node_just_written(b)) && bch2_btree_post_write_cleanup(c, b)) bch2_trans_node_reinit_iter(trans, b); @@ -236,25 +251,26 @@ inline void bch2_btree_add_journal_pin(struct bch_fs *c, /** * btree_insert_key - insert a key one key into a leaf node */ -static void btree_insert_key_leaf(struct btree_trans *trans, - struct btree_insert_entry *insert) +inline void bch2_btree_insert_key_leaf(struct btree_trans *trans, + struct btree_path *path, + struct bkey_i *insert, + u64 journal_seq) { struct bch_fs *c = trans->c; - struct btree *b = insert_l(insert)->b; + struct btree *b = path_l(path)->b; struct bset_tree *t = bset_tree_last(b); struct bset *i = bset(b, t); int old_u64s = bset_u64s(t); int old_live_u64s = b->nr.live_u64s; int live_u64s_added, u64s_added; - if (unlikely(!bch2_btree_bset_insert_key(trans, insert->path, b, - &insert_l(insert)->iter, insert->k))) + if (unlikely(!bch2_btree_bset_insert_key(trans, path, b, + &path_l(path)->iter, insert))) return; - i->journal_seq = cpu_to_le64(max(trans->journal_res.seq, - le64_to_cpu(i->journal_seq))); + i->journal_seq = cpu_to_le64(max(journal_seq, le64_to_cpu(i->journal_seq))); - bch2_btree_add_journal_pin(c, b, trans->journal_res.seq); + bch2_btree_add_journal_pin(c, b, journal_seq); if (unlikely(!btree_node_dirty(b))) set_btree_node_dirty_acct(c, b); @@ -272,6 +288,12 @@ static void btree_insert_key_leaf(struct btree_trans *trans, bch2_trans_node_reinit_iter(trans, b); } +static void btree_insert_key_leaf(struct btree_trans *trans, + struct btree_insert_entry *insert) +{ + bch2_btree_insert_key_leaf(trans, insert->path, insert->k, trans->journal_res.seq); +} + /* Cached btree updates: */ /* Normal update interface: */ @@ -291,7 +313,7 @@ static inline void btree_insert_entry_checks(struct btree_trans *trans, } static noinline int -bch2_trans_journal_preres_get_cold(struct btree_trans *trans, unsigned u64s, +bch2_trans_journal_preres_get_cold(struct btree_trans *trans, unsigned flags, unsigned long trace_ip) { struct bch_fs *c = trans->c; @@ -300,7 +322,9 @@ bch2_trans_journal_preres_get_cold(struct btree_trans *trans, unsigned u64s, bch2_trans_unlock(trans); ret = bch2_journal_preres_get(&c->journal, - &trans->journal_preres, u64s, 0); + &trans->journal_preres, + trans->journal_preres_u64s, + (flags & JOURNAL_WATERMARK_MASK)); if (ret) return ret; @@ -314,17 +338,15 @@ bch2_trans_journal_preres_get_cold(struct btree_trans *trans, unsigned u64s, } static __always_inline int bch2_trans_journal_res_get(struct btree_trans *trans, - unsigned flags) + unsigned flags) { return bch2_journal_res_get(&trans->c->journal, &trans->journal_res, - trans->journal_u64s, - flags| - (trans->flags & JOURNAL_WATERMARK_MASK)); + trans->journal_u64s, flags); } #define JSET_ENTRY_LOG_U64s 4 -static void journal_transaction_name(struct btree_trans *trans) +static noinline void journal_transaction_name(struct btree_trans *trans) { struct bch_fs *c = trans->c; struct journal *j = &c->journal; @@ -349,9 +371,8 @@ static inline int btree_key_can_insert(struct btree_trans *trans, return 0; } -static int btree_key_can_insert_cached(struct btree_trans *trans, - struct btree_path *path, - unsigned u64s) +static int btree_key_can_insert_cached(struct btree_trans *trans, unsigned flags, + struct btree_path *path, unsigned u64s) { struct bch_fs *c = trans->c; struct bkey_cached *ck = (void *) path->l[0].b; @@ -363,7 +384,7 @@ static int btree_key_can_insert_cached(struct btree_trans *trans, if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) && bch2_btree_key_cache_must_wait(c) && - !(trans->flags & BTREE_INSERT_JOURNAL_RECLAIM)) + !(flags & BTREE_INSERT_JOURNAL_RECLAIM)) return -BCH_ERR_btree_insert_need_journal_reclaim; /* @@ -413,7 +434,8 @@ static int run_one_mem_trigger(struct btree_trans *trans, if (bch2_bkey_ops[old.k->type].atomic_trigger == bch2_bkey_ops[i->k->k.type].atomic_trigger && ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) { - ret = bch2_mark_key(trans, old, bkey_i_to_s_c(new), + ret = bch2_mark_key(trans, i->btree_id, i->level, + old, bkey_i_to_s_c(new), BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE|flags); } else { struct bkey _deleted = KEY(0, 0, 0); @@ -421,9 +443,11 @@ static int run_one_mem_trigger(struct btree_trans *trans, _deleted.p = i->path->pos; - ret = bch2_mark_key(trans, deleted, bkey_i_to_s_c(new), + ret = bch2_mark_key(trans, i->btree_id, i->level, + deleted, bkey_i_to_s_c(new), BTREE_TRIGGER_INSERT|flags) ?: - bch2_mark_key(trans, old, deleted, + bch2_mark_key(trans, i->btree_id, i->level, + old, deleted, BTREE_TRIGGER_OVERWRITE|flags); } @@ -573,12 +597,13 @@ static noinline int bch2_trans_commit_run_gc_triggers(struct btree_trans *trans) } static inline int -bch2_trans_commit_write_locked(struct btree_trans *trans, +bch2_trans_commit_write_locked(struct btree_trans *trans, unsigned flags, struct btree_insert_entry **stopped_at, unsigned long trace_ip) { struct bch_fs *c = trans->c; struct btree_insert_entry *i; + struct btree_write_buffered_key *wb; struct btree_trans_commit_hook *h; unsigned u64s = 0; bool marking = false; @@ -613,7 +638,7 @@ bch2_trans_commit_write_locked(struct btree_trans *trans, u64s += i->k->k.u64s; ret = !i->cached ? btree_key_can_insert(trans, insert_l(i)->b, u64s) - : btree_key_can_insert_cached(trans, i->path, u64s); + : btree_key_can_insert_cached(trans, flags, i->path, u64s); if (ret) { *stopped_at = i; return ret; @@ -623,17 +648,23 @@ bch2_trans_commit_write_locked(struct btree_trans *trans, marking = true; } + if (trans->nr_wb_updates && + trans->nr_wb_updates + c->btree_write_buffer.state.nr > c->btree_write_buffer.size) + return -BCH_ERR_btree_insert_need_flush_buffer; + /* * Don't get journal reservation until after we know insert will * succeed: */ - if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) { + if (likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY))) { ret = bch2_trans_journal_res_get(trans, + (flags & JOURNAL_WATERMARK_MASK)| JOURNAL_RES_GET_NONBLOCK); if (ret) return ret; - journal_transaction_name(trans); + if (unlikely(trans->journal_transaction_names)) + journal_transaction_name(trans); } else { trans->journal_res.seq = c->journal.replay_journal_seq; } @@ -644,7 +675,7 @@ bch2_trans_commit_write_locked(struct btree_trans *trans, */ if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG) && - !(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)) { + !(flags & BTREE_INSERT_JOURNAL_REPLAY)) { if (bch2_journal_seq_verify) trans_for_each_update(trans, i) i->k->k.version.lo = trans->journal_res.seq; @@ -657,17 +688,25 @@ bch2_trans_commit_write_locked(struct btree_trans *trans, bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas)) return -BCH_ERR_btree_insert_need_mark_replicas; + if (trans->nr_wb_updates) { + EBUG_ON(flags & BTREE_INSERT_JOURNAL_REPLAY); + + ret = bch2_btree_insert_keys_write_buffer(trans); + if (ret) + goto revert_fs_usage; + } + trans_for_each_update(trans, i) if (BTREE_NODE_TYPE_HAS_MEM_TRIGGERS & (1U << i->bkey_type)) { ret = run_one_mem_trigger(trans, i, i->flags); if (ret) - return ret; + goto fatal_err; } if (unlikely(c->gc_pos.phase)) { ret = bch2_trans_commit_run_gc_triggers(trans); if (ret) - return ret; + goto fatal_err; } if (unlikely(trans->extra_journal_entries.nr)) { @@ -679,22 +718,27 @@ bch2_trans_commit_write_locked(struct btree_trans *trans, trans->journal_res.u64s -= trans->extra_journal_entries.nr; } - if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) { - trans_for_each_update(trans, i) { - struct journal *j = &c->journal; - struct jset_entry *entry; + if (likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY))) { + struct journal *j = &c->journal; + struct jset_entry *entry; + trans_for_each_update(trans, i) { if (i->key_cache_already_flushed) continue; + if (i->flags & BTREE_UPDATE_NOJOURNAL) + continue; + verify_update_old_key(trans, i); - entry = bch2_journal_add_entry(j, &trans->journal_res, - BCH_JSET_ENTRY_overwrite, - i->btree_id, i->level, - i->old_k.u64s); - bkey_reassemble(&entry->start[0], - (struct bkey_s_c) { &i->old_k, i->old_v }); + if (trans->journal_transaction_names) { + entry = bch2_journal_add_entry(j, &trans->journal_res, + BCH_JSET_ENTRY_overwrite, + i->btree_id, i->level, + i->old_k.u64s); + bkey_reassemble(&entry->start[0], + (struct bkey_s_c) { &i->old_k, i->old_v }); + } entry = bch2_journal_add_entry(j, &trans->journal_res, BCH_JSET_ENTRY_btree_keys, @@ -703,6 +747,14 @@ bch2_trans_commit_write_locked(struct btree_trans *trans, bkey_copy(&entry->start[0], i->k); } + trans_for_each_wb_update(trans, wb) { + entry = bch2_journal_add_entry(j, &trans->journal_res, + BCH_JSET_ENTRY_btree_keys, + wb->btree, 0, + wb->k.k.u64s); + bkey_copy(&entry->start[0], &wb->k); + } + if (trans->journal_seq) *trans->journal_seq = trans->journal_res.seq; } @@ -713,13 +765,19 @@ bch2_trans_commit_write_locked(struct btree_trans *trans, if (!i->cached) btree_insert_key_leaf(trans, i); else if (!i->key_cache_already_flushed) - bch2_btree_insert_key_cached(trans, i->path, i->k); + bch2_btree_insert_key_cached(trans, flags, i->path, i->k); else { bch2_btree_key_cache_drop(trans, i->path); btree_path_set_dirty(i->path, BTREE_ITER_NEED_TRAVERSE); } } + return 0; +fatal_err: + bch2_fatal_error(c); +revert_fs_usage: + if (trans->fs_usage_deltas) + bch2_trans_fs_usage_revert(trans, trans->fs_usage_deltas); return ret; } @@ -747,7 +805,8 @@ static inline int trans_lock_write(struct btree_trans *trans) if (bch2_btree_node_lock_write(trans, i->path, &insert_l(i)->b->c)) return trans_lock_write_fail(trans, i); - bch2_btree_node_prep_for_write(trans, i->path, insert_l(i)->b); + if (!i->cached) + bch2_btree_node_prep_for_write(trans, i->path, insert_l(i)->b); } return 0; @@ -756,18 +815,22 @@ static inline int trans_lock_write(struct btree_trans *trans) static noinline void bch2_drop_overwrites_from_journal(struct btree_trans *trans) { struct btree_insert_entry *i; + struct btree_write_buffered_key *wb; trans_for_each_update(trans, i) bch2_journal_key_overwritten(trans->c, i->btree_id, i->level, i->k->k.p); + + trans_for_each_wb_update(trans, wb) + bch2_journal_key_overwritten(trans->c, wb->btree, 0, wb->k.k.p); } #ifdef CONFIG_BCACHEFS_DEBUG -static noinline int bch2_trans_commit_bkey_invalid(struct btree_trans *trans, +static noinline int bch2_trans_commit_bkey_invalid(struct btree_trans *trans, unsigned flags, struct btree_insert_entry *i, struct printbuf *err) { struct bch_fs *c = trans->c; - int rw = (trans->flags & BTREE_INSERT_JOURNAL_REPLAY) ? READ : WRITE; + int rw = (flags & BTREE_INSERT_JOURNAL_REPLAY) ? READ : WRITE; printbuf_reset(err); prt_printf(err, "invalid bkey on insert from %s -> %ps", @@ -793,26 +856,27 @@ static noinline int bch2_trans_commit_bkey_invalid(struct btree_trans *trans, /* * Get journal reservation, take write locks, and attempt to do btree update(s): */ -static inline int do_bch2_trans_commit(struct btree_trans *trans, +static inline int do_bch2_trans_commit(struct btree_trans *trans, unsigned flags, struct btree_insert_entry **stopped_at, unsigned long trace_ip) { struct bch_fs *c = trans->c; struct btree_insert_entry *i; - struct printbuf buf = PRINTBUF; int ret, u64s_delta = 0; #ifdef CONFIG_BCACHEFS_DEBUG + struct printbuf buf = PRINTBUF; + trans_for_each_update(trans, i) { - int rw = (trans->flags & BTREE_INSERT_JOURNAL_REPLAY) ? READ : WRITE; + int rw = (flags & BTREE_INSERT_JOURNAL_REPLAY) ? READ : WRITE; if (unlikely(bch2_bkey_invalid(c, bkey_i_to_s_c(i->k), i->bkey_type, rw, &buf))) - return bch2_trans_commit_bkey_invalid(trans, i, &buf); + return bch2_trans_commit_bkey_invalid(trans, flags, i, &buf); btree_insert_entry_checks(trans, i); } -#endif printbuf_exit(&buf); +#endif trans_for_each_update(trans, i) { if (i->cached) @@ -824,7 +888,7 @@ static inline int do_bch2_trans_commit(struct btree_trans *trans, if (!same_leaf_as_next(trans, i)) { if (u64s_delta <= 0) { ret = bch2_foreground_maybe_merge(trans, i->path, - i->level, trans->flags); + i->level, flags); if (unlikely(ret)) return ret; } @@ -835,11 +899,9 @@ static inline int do_bch2_trans_commit(struct btree_trans *trans, ret = bch2_journal_preres_get(&c->journal, &trans->journal_preres, trans->journal_preres_u64s, - JOURNAL_RES_GET_NONBLOCK| - (trans->flags & JOURNAL_WATERMARK_MASK)); + (flags & JOURNAL_WATERMARK_MASK)|JOURNAL_RES_GET_NONBLOCK); if (unlikely(ret == -BCH_ERR_journal_preres_get_blocked)) - ret = bch2_trans_journal_preres_get_cold(trans, - trans->journal_preres_u64s, trace_ip); + ret = bch2_trans_journal_preres_get_cold(trans, flags, trace_ip); if (unlikely(ret)) return ret; @@ -847,7 +909,7 @@ static inline int do_bch2_trans_commit(struct btree_trans *trans, if (unlikely(ret)) return ret; - ret = bch2_trans_commit_write_locked(trans, stopped_at, trace_ip); + ret = bch2_trans_commit_write_locked(trans, flags, stopped_at, trace_ip); if (!ret && unlikely(trans->journal_replay_not_finished)) bch2_drop_overwrites_from_journal(trans); @@ -886,7 +948,7 @@ static int journal_reclaim_wait_done(struct bch_fs *c) } static noinline -int bch2_trans_commit_error(struct btree_trans *trans, +int bch2_trans_commit_error(struct btree_trans *trans, unsigned flags, struct btree_insert_entry *i, int ret, unsigned long trace_ip) { @@ -894,7 +956,7 @@ int bch2_trans_commit_error(struct btree_trans *trans, switch (ret) { case -BCH_ERR_btree_insert_btree_node_full: - ret = bch2_btree_split_leaf(trans, i->path, trans->flags); + ret = bch2_btree_split_leaf(trans, i->path, flags); if (bch2_err_matches(ret, BCH_ERR_transaction_restart)) trace_and_count(c, trans_restart_btree_node_split, trans, trace_ip, i->path); break; @@ -912,13 +974,15 @@ int bch2_trans_commit_error(struct btree_trans *trans, case -BCH_ERR_journal_res_get_blocked: bch2_trans_unlock(trans); - if ((trans->flags & BTREE_INSERT_JOURNAL_RECLAIM) && - !(trans->flags & JOURNAL_WATERMARK_reserved)) { + if ((flags & BTREE_INSERT_JOURNAL_RECLAIM) && + !(flags & JOURNAL_WATERMARK_reserved)) { ret = -BCH_ERR_journal_reclaim_would_deadlock; break; } - ret = bch2_trans_journal_res_get(trans, JOURNAL_RES_GET_CHECK); + ret = bch2_trans_journal_res_get(trans, + (flags & JOURNAL_WATERMARK_MASK)| + JOURNAL_RES_GET_CHECK); if (ret) break; @@ -940,6 +1004,30 @@ int bch2_trans_commit_error(struct btree_trans *trans, if (ret) trace_and_count(c, trans_restart_journal_reclaim, trans, trace_ip); break; + case -BCH_ERR_btree_insert_need_flush_buffer: { + struct btree_write_buffer *wb = &c->btree_write_buffer; + + ret = 0; + + if (wb->state.nr > wb->size * 3 / 4) { + bch2_trans_reset_updates(trans); + bch2_trans_unlock(trans); + + mutex_lock(&wb->flush_lock); + + if (wb->state.nr > wb->size * 3 / 4) + ret = __bch2_btree_write_buffer_flush(trans, + flags|BTREE_INSERT_NOCHECK_RW, true); + else + mutex_unlock(&wb->flush_lock); + + if (!ret) { + trace_and_count(c, trans_restart_write_buffer_flush, trans, _THIS_IP_); + ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_write_buffer_flush); + } + } + break; + } default: BUG_ON(ret >= 0); break; @@ -948,20 +1036,20 @@ int bch2_trans_commit_error(struct btree_trans *trans, BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted); bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOSPC) && - !(trans->flags & BTREE_INSERT_NOWAIT) && - (trans->flags & BTREE_INSERT_NOFAIL), c, + !(flags & BTREE_INSERT_NOWAIT) && + (flags & BTREE_INSERT_NOFAIL), c, "%s: incorrectly got %s\n", __func__, bch2_err_str(ret)); return ret; } static noinline int -bch2_trans_commit_get_rw_cold(struct btree_trans *trans) +bch2_trans_commit_get_rw_cold(struct btree_trans *trans, unsigned flags) { struct bch_fs *c = trans->c; int ret; - if (likely(!(trans->flags & BTREE_INSERT_LAZY_RW)) || + if (likely(!(flags & BTREE_INSERT_LAZY_RW)) || test_bit(BCH_FS_STARTED, &c->flags)) return -BCH_ERR_erofs_trans_commit; @@ -972,7 +1060,7 @@ bch2_trans_commit_get_rw_cold(struct btree_trans *trans) if (ret) return ret; - percpu_ref_get(&c->writes); + bch2_write_ref_get(c, BCH_WRITE_REF_trans); return 0; } @@ -997,18 +1085,20 @@ do_bch2_trans_commit_to_journal_replay(struct btree_trans *trans) return ret; } -int __bch2_trans_commit(struct btree_trans *trans) +int __bch2_trans_commit(struct btree_trans *trans, unsigned flags) { struct bch_fs *c = trans->c; struct btree_insert_entry *i = NULL; + struct btree_write_buffered_key *wb; unsigned u64s; int ret = 0; if (!trans->nr_updates && + !trans->nr_wb_updates && !trans->extra_journal_entries.nr) goto out_reset; - if (trans->flags & BTREE_INSERT_GC_LOCK_HELD) + if (flags & BTREE_INSERT_GC_LOCK_HELD) lockdep_assert_held(&c->gc_lock); ret = bch2_trans_commit_run_triggers(trans); @@ -1020,13 +1110,27 @@ int __bch2_trans_commit(struct btree_trans *trans) goto out_reset; } - if (!(trans->flags & BTREE_INSERT_NOCHECK_RW) && - unlikely(!percpu_ref_tryget_live(&c->writes))) { - ret = bch2_trans_commit_get_rw_cold(trans); + if (!(flags & BTREE_INSERT_NOCHECK_RW) && + unlikely(!bch2_write_ref_tryget(c, BCH_WRITE_REF_trans))) { + ret = bch2_trans_commit_get_rw_cold(trans, flags); if (ret) goto out_reset; } + if (c->btree_write_buffer.state.nr > c->btree_write_buffer.size / 2 && + mutex_trylock(&c->btree_write_buffer.flush_lock)) { + bch2_trans_begin(trans); + bch2_trans_unlock(trans); + + ret = __bch2_btree_write_buffer_flush(trans, + flags|BTREE_INSERT_NOCHECK_RW, true); + if (!ret) { + trace_and_count(c, trans_restart_write_buffer_flush, trans, _THIS_IP_); + ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_write_buffer_flush); + } + goto out; + } + EBUG_ON(test_bit(BCH_FS_CLEAN_SHUTDOWN, &c->flags)); memset(&trans->journal_preres, 0, sizeof(trans->journal_preres)); @@ -1034,8 +1138,10 @@ int __bch2_trans_commit(struct btree_trans *trans) trans->journal_u64s = trans->extra_journal_entries.nr; trans->journal_preres_u64s = 0; - /* For journalling transaction name: */ - trans->journal_u64s += jset_u64s(JSET_ENTRY_LOG_U64s); + trans->journal_transaction_names = READ_ONCE(c->opts.journal_transaction_names); + + if (trans->journal_transaction_names) + trans->journal_u64s += jset_u64s(JSET_ENTRY_LOG_U64s); trans_for_each_update(trans, i) { EBUG_ON(!i->path->should_be_locked); @@ -1052,27 +1158,35 @@ int __bch2_trans_commit(struct btree_trans *trans) /* we're going to journal the key being updated: */ u64s = jset_u64s(i->k->k.u64s); if (i->cached && - likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) + likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY))) trans->journal_preres_u64s += u64s; + + if (i->flags & BTREE_UPDATE_NOJOURNAL) + continue; + trans->journal_u64s += u64s; /* and we're also going to log the overwrite: */ - trans->journal_u64s += jset_u64s(i->old_k.u64s); + if (trans->journal_transaction_names) + trans->journal_u64s += jset_u64s(i->old_k.u64s); } + trans_for_each_wb_update(trans, wb) + trans->journal_u64s += jset_u64s(wb->k.k.u64s); + if (trans->extra_journal_res) { ret = bch2_disk_reservation_add(c, trans->disk_res, trans->extra_journal_res, - (trans->flags & BTREE_INSERT_NOFAIL) + (flags & BTREE_INSERT_NOFAIL) ? BCH_DISK_RESERVATION_NOFAIL : 0); if (ret) goto err; } retry: - EBUG_ON(trans->restarted); + bch2_trans_verify_not_in_restart(trans); memset(&trans->journal_res, 0, sizeof(trans->journal_res)); - ret = do_bch2_trans_commit(trans, &i, _RET_IP_); + ret = do_bch2_trans_commit(trans, flags, &i, _RET_IP_); /* make sure we didn't drop or screw up locks: */ bch2_trans_verify_locks(trans); @@ -1084,22 +1198,14 @@ retry: out: bch2_journal_preres_put(&c->journal, &trans->journal_preres); - if (likely(!(trans->flags & BTREE_INSERT_NOCHECK_RW))) - percpu_ref_put(&c->writes); + if (likely(!(flags & BTREE_INSERT_NOCHECK_RW))) + bch2_write_ref_put(c, BCH_WRITE_REF_trans); out_reset: bch2_trans_reset_updates(trans); - if (trans->fs_usage_deltas) { - trans->fs_usage_deltas->used = 0; - memset((void *) trans->fs_usage_deltas + - offsetof(struct replicas_delta_list, memset_start), 0, - (void *) &trans->fs_usage_deltas->memset_end - - (void *) &trans->fs_usage_deltas->memset_start); - } - return ret; err: - ret = bch2_trans_commit_error(trans, i, ret, _RET_IP_); + ret = bch2_trans_commit_error(trans, flags, i, ret, _RET_IP_); if (ret) goto out; @@ -1152,13 +1258,97 @@ static inline int check_pos_snapshot_overwritten(struct btree_trans *trans, return __check_pos_snapshot_overwritten(trans, id, pos); } +static noinline int extent_front_merge(struct btree_trans *trans, + struct btree_iter *iter, + struct bkey_s_c k, + struct bkey_i **insert, + enum btree_update_flags flags) +{ + struct bch_fs *c = trans->c; + struct bkey_i *update; + int ret; + + update = bch2_bkey_make_mut(trans, k); + ret = PTR_ERR_OR_ZERO(update); + if (ret) + return ret; + + if (!bch2_bkey_merge(c, bkey_i_to_s(update), bkey_i_to_s_c(*insert))) + return 0; + + ret = check_pos_snapshot_overwritten(trans, iter->btree_id, k.k->p) ?: + check_pos_snapshot_overwritten(trans, iter->btree_id, (*insert)->k.p); + if (ret < 0) + return ret; + if (ret) + return 0; + + ret = bch2_btree_delete_at(trans, iter, flags); + if (ret) + return ret; + + *insert = update; + return 0; +} + +static noinline int extent_back_merge(struct btree_trans *trans, + struct btree_iter *iter, + struct bkey_i *insert, + struct bkey_s_c k) +{ + struct bch_fs *c = trans->c; + int ret; + + ret = check_pos_snapshot_overwritten(trans, iter->btree_id, insert->k.p) ?: + check_pos_snapshot_overwritten(trans, iter->btree_id, k.k->p); + if (ret < 0) + return ret; + if (ret) + return 0; + + bch2_bkey_merge(c, bkey_i_to_s(insert), k); + return 0; +} + +/* + * When deleting, check if we need to emit a whiteout (because we're overwriting + * something in an ancestor snapshot) + */ +static int need_whiteout_for_snapshot(struct btree_trans *trans, + enum btree_id btree_id, struct bpos pos) +{ + struct btree_iter iter; + struct bkey_s_c k; + u32 snapshot = pos.snapshot; + int ret; + + if (!bch2_snapshot_parent(trans->c, pos.snapshot)) + return 0; + + pos.snapshot++; + + for_each_btree_key_norestart(trans, iter, btree_id, pos, + BTREE_ITER_ALL_SNAPSHOTS| + BTREE_ITER_NOPRESERVE, k, ret) { + if (!bkey_eq(k.k->p, pos)) + break; + + if (bch2_snapshot_is_ancestor(trans->c, snapshot, + k.k->p.snapshot)) { + ret = !bkey_whiteout(k.k); + break; + } + } + bch2_trans_iter_exit(trans, &iter); + + return ret; +} int bch2_trans_update_extent(struct btree_trans *trans, struct btree_iter *orig_iter, struct bkey_i *insert, enum btree_update_flags flags) { - struct bch_fs *c = trans->c; - struct btree_iter iter, update_iter; + struct btree_iter iter; struct bpos start = bkey_start_pos(&insert->k); struct bkey_i *update; struct bkey_s_c k; @@ -1175,46 +1365,15 @@ int bch2_trans_update_extent(struct btree_trans *trans, if (!k.k) goto out; - if (bch2_bkey_maybe_mergable(k.k, &insert->k)) { - /* - * We can't merge extents if they belong to interior snapshot - * tree nodes, and there's a snapshot in which one extent is - * visible and the other is not - i.e. if visibility is - * different. - * - * Instead of checking if visibilitiy of the two extents is - * different, for now we just check if either has been - * overwritten: - */ - ret = check_pos_snapshot_overwritten(trans, btree_id, insert->k.p); - if (ret < 0) - goto err; - if (ret) - goto nomerge1; - - ret = check_pos_snapshot_overwritten(trans, btree_id, k.k->p); - if (ret < 0) - goto err; - if (ret) - goto nomerge1; - - update = bch2_bkey_make_mut(trans, k); - if ((ret = PTR_ERR_OR_ZERO(update))) - goto err; - - if (bch2_bkey_merge(c, bkey_i_to_s(update), bkey_i_to_s_c(insert))) { - ret = bch2_btree_delete_at(trans, &iter, flags); + if (bkey_eq(k.k->p, bkey_start_pos(&insert->k))) { + if (bch2_bkey_maybe_mergable(k.k, &insert->k)) { + ret = extent_front_merge(trans, &iter, k, &insert, flags); if (ret) goto err; - - insert = update; - goto next; } - } -nomerge1: - ret = 0; - if (bkey_eq(k.k->p, start)) + goto next; + } while (bkey_gt(insert->k.p, bkey_start_pos(k.k))) { bool front_split = bkey_lt(bkey_start_pos(k.k), start); @@ -1237,16 +1396,8 @@ nomerge1: bch2_cut_back(start, update); - bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p, - BTREE_ITER_NOT_EXTENTS| - BTREE_ITER_ALL_SNAPSHOTS| - BTREE_ITER_INTENT); - ret = bch2_btree_iter_traverse(&update_iter) ?: - bch2_trans_update(trans, &update_iter, update, - BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE| - flags); - bch2_trans_iter_exit(trans, &update_iter); - + ret = bch2_btree_insert_nonextent(trans, btree_id, update, + BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags); if (ret) goto err; } @@ -1260,15 +1411,8 @@ nomerge1: bch2_cut_front(start, update); bch2_cut_back(insert->k.p, update); - bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p, - BTREE_ITER_NOT_EXTENTS| - BTREE_ITER_ALL_SNAPSHOTS| - BTREE_ITER_INTENT); - ret = bch2_btree_iter_traverse(&update_iter) ?: - bch2_trans_update(trans, &update_iter, update, - BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE| - flags); - bch2_trans_iter_exit(trans, &update_iter); + ret = bch2_btree_insert_nonextent(trans, btree_id, update, + BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags); if (ret) goto err; } @@ -1280,21 +1424,15 @@ nomerge1: bkey_init(&update->k); update->k.p = k.k->p; + update->k.p.snapshot = insert->k.p.snapshot; - if (insert->k.p.snapshot != k.k->p.snapshot) { - update->k.p.snapshot = insert->k.p.snapshot; + if (insert->k.p.snapshot != k.k->p.snapshot || + (btree_type_has_snapshots(btree_id) && + need_whiteout_for_snapshot(trans, btree_id, update->k.p))) update->k.type = KEY_TYPE_whiteout; - } - - bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p, - BTREE_ITER_NOT_EXTENTS| - BTREE_ITER_INTENT); - ret = bch2_btree_iter_traverse(&update_iter) ?: - bch2_trans_update(trans, &update_iter, update, - BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE| - flags); - bch2_trans_iter_exit(trans, &update_iter); + ret = bch2_btree_insert_nonextent(trans, btree_id, update, + BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags); if (ret) goto err; } @@ -1323,22 +1461,10 @@ next: } if (bch2_bkey_maybe_mergable(&insert->k, k.k)) { - ret = check_pos_snapshot_overwritten(trans, btree_id, insert->k.p); - if (ret < 0) - goto err; + ret = extent_back_merge(trans, &iter, insert, k); if (ret) - goto nomerge2; - - ret = check_pos_snapshot_overwritten(trans, btree_id, k.k->p); - if (ret < 0) goto err; - if (ret) - goto nomerge2; - - bch2_bkey_merge(c, bkey_i_to_s(insert), k); } -nomerge2: - ret = 0; out: if (!bkey_deleted(&insert->k)) { /* @@ -1358,40 +1484,6 @@ err: return ret; } -/* - * When deleting, check if we need to emit a whiteout (because we're overwriting - * something in an ancestor snapshot) - */ -static int need_whiteout_for_snapshot(struct btree_trans *trans, - enum btree_id btree_id, struct bpos pos) -{ - struct btree_iter iter; - struct bkey_s_c k; - u32 snapshot = pos.snapshot; - int ret; - - if (!bch2_snapshot_parent(trans->c, pos.snapshot)) - return 0; - - pos.snapshot++; - - for_each_btree_key_norestart(trans, iter, btree_id, pos, - BTREE_ITER_ALL_SNAPSHOTS| - BTREE_ITER_NOPRESERVE, k, ret) { - if (!bkey_eq(k.k->p, pos)) - break; - - if (bch2_snapshot_is_ancestor(trans->c, snapshot, - k.k->p.snapshot)) { - ret = !bkey_whiteout(k.k); - break; - } - } - bch2_trans_iter_exit(trans, &iter); - - return ret; -} - static int __must_check bch2_trans_update_by_path_trace(struct btree_trans *trans, struct btree_path *path, struct bkey_i *k, enum btree_update_flags flags, @@ -1476,7 +1568,7 @@ bch2_trans_update_by_path_trace(struct btree_trans *trans, struct btree_path *pa array_insert_item(trans->updates, trans->nr_updates, i - trans->updates, n); - i->old_v = bch2_btree_path_peek_slot(path, &i->old_k).v; + i->old_v = bch2_btree_path_peek_slot_exact(path, &i->old_k).v; i->old_btree_u64s = !bkey_deleted(&i->old_k) ? i->old_k.u64s : 0; if (unlikely(trans->journal_replay_not_finished)) { @@ -1499,7 +1591,9 @@ bch2_trans_update_by_path_trace(struct btree_trans *trans, struct btree_path *pa * the key cache - but the key has to exist in the btree for that to * work: */ - if (unlikely(path->cached && bkey_deleted(&i->old_k))) + if (path->cached && + bkey_deleted(&i->old_k) && + !(flags & BTREE_UPDATE_NO_KEY_CACHE_COHERENCY)) return flush_new_cached_update(trans, path, i, flags, ip); return 0; @@ -1575,6 +1669,59 @@ int __must_check bch2_trans_update(struct btree_trans *trans, struct btree_iter return bch2_trans_update_by_path(trans, path, k, flags); } +int __must_check bch2_trans_update_buffered(struct btree_trans *trans, + enum btree_id btree, + struct bkey_i *k) +{ + struct btree_write_buffered_key *i; + int ret; + + EBUG_ON(trans->nr_wb_updates > trans->wb_updates_size); + EBUG_ON(k->k.u64s > BTREE_WRITE_BUFERED_U64s_MAX); + + trans_for_each_wb_update(trans, i) { + if (i->btree == btree && bpos_eq(i->k.k.p, k->k.p)) { + bkey_copy(&i->k, k); + return 0; + } + } + + if (!trans->wb_updates || + trans->nr_wb_updates == trans->wb_updates_size) { + struct btree_write_buffered_key *u; + + if (trans->nr_wb_updates == trans->wb_updates_size) { + struct btree_transaction_stats *s = btree_trans_stats(trans); + + BUG_ON(trans->wb_updates_size > U8_MAX / 2); + trans->wb_updates_size = max(1, trans->wb_updates_size * 2); + if (s) + s->wb_updates_size = trans->wb_updates_size; + } + + u = bch2_trans_kmalloc_nomemzero(trans, + trans->wb_updates_size * + sizeof(struct btree_write_buffered_key)); + ret = PTR_ERR_OR_ZERO(u); + if (ret) + return ret; + + if (trans->nr_wb_updates) + memcpy(u, trans->wb_updates, trans->nr_wb_updates * + sizeof(struct btree_write_buffered_key)); + trans->wb_updates = u; + } + + trans->wb_updates[trans->nr_wb_updates] = (struct btree_write_buffered_key) { + .btree = btree, + }; + + bkey_copy(&trans->wb_updates[trans->nr_wb_updates].k, k); + trans->nr_wb_updates++; + + return 0; +} + void bch2_trans_commit_hook(struct btree_trans *trans, struct btree_trans_commit_hook *h) { @@ -1582,16 +1729,33 @@ void bch2_trans_commit_hook(struct btree_trans *trans, trans->hooks = h; } -int __bch2_btree_insert(struct btree_trans *trans, - enum btree_id id, struct bkey_i *k) +int bch2_btree_insert_nonextent(struct btree_trans *trans, + enum btree_id btree, struct bkey_i *k, + enum btree_update_flags flags) +{ + struct btree_iter iter; + int ret; + + bch2_trans_iter_init(trans, &iter, btree, k->k.p, + BTREE_ITER_NOT_EXTENTS| + BTREE_ITER_INTENT); + ret = bch2_btree_iter_traverse(&iter) ?: + bch2_trans_update(trans, &iter, k, flags); + bch2_trans_iter_exit(trans, &iter); + return ret; +} + +int __bch2_btree_insert(struct btree_trans *trans, enum btree_id id, + struct bkey_i *k, enum btree_update_flags flags) { struct btree_iter iter; int ret; bch2_trans_iter_init(trans, &iter, id, bkey_start_pos(&k->k), + BTREE_ITER_CACHED| BTREE_ITER_INTENT); ret = bch2_btree_iter_traverse(&iter) ?: - bch2_trans_update(trans, &iter, k, 0); + bch2_trans_update(trans, &iter, k, flags); bch2_trans_iter_exit(trans, &iter); return ret; } @@ -1609,7 +1773,7 @@ int bch2_btree_insert(struct bch_fs *c, enum btree_id id, u64 *journal_seq, int flags) { return bch2_trans_do(c, disk_res, journal_seq, flags, - __bch2_btree_insert(&trans, id, k)); + __bch2_btree_insert(&trans, id, k, 0)); } int bch2_btree_delete_extent_at(struct btree_trans *trans, struct btree_iter *iter, @@ -1671,18 +1835,10 @@ int bch2_btree_delete_range_trans(struct btree_trans *trans, enum btree_id id, */ delete.k.p = iter.pos; - if (iter.flags & BTREE_ITER_IS_EXTENTS) { - unsigned max_sectors = - KEY_SIZE_MAX & (~0 << trans->c->block_bits); - - /* create the biggest key we can */ - bch2_key_resize(&delete.k, max_sectors); - bch2_cut_back(end, &delete); - - ret = bch2_extent_trim_atomic(trans, &iter, &delete); - if (ret) - goto err; - } + if (iter.flags & BTREE_ITER_IS_EXTENTS) + bch2_key_resize(&delete.k, + bpos_min(end, k.k->p).offset - + iter.pos.offset); ret = bch2_trans_update(trans, &iter, &delete, update_flags) ?: bch2_trans_commit(trans, &disk_res, journal_seq,