]> git.sesse.net Git - bcachefs-tools-debian/blobdiff - libbcachefs/btree_cache.c
Fixes tests which have broken since 'list' improvements
[bcachefs-tools-debian] / libbcachefs / btree_cache.c
index c950f2564f25d9e907a564fdf07843b1f3b880a6..d3addd3a89647164cbd4f0a48e69154c585dfa97 100644 (file)
@@ -1,3 +1,4 @@
+// SPDX-License-Identifier: GPL-2.0
 
 #include "bcachefs.h"
 #include "btree_cache.h"
@@ -5,20 +6,18 @@
 #include "btree_iter.h"
 #include "btree_locking.h"
 #include "debug.h"
-#include "extents.h"
 
 #include <linux/prefetch.h>
+#include <linux/sched/mm.h>
 #include <trace/events/bcachefs.h>
 
-#define DEF_BTREE_ID(kwd, val, name) name,
-
 const char * const bch2_btree_ids[] = {
-       DEFINE_BCH_BTREE_IDS()
+#define x(kwd, val, name) name,
+       BCH_BTREE_IDS()
+#undef x
        NULL
 };
 
-#undef DEF_BTREE_ID
-
 void bch2_recalc_btree_reserve(struct bch_fs *c)
 {
        unsigned i, reserve = 16;
@@ -29,7 +28,7 @@ void bch2_recalc_btree_reserve(struct bch_fs *c)
        for (i = 0; i < BTREE_ID_NR; i++)
                if (c->btree_roots[i].b)
                        reserve += min_t(unsigned, 1,
-                                        c->btree_roots[i].b->level) * 8;
+                                        c->btree_roots[i].b->c.level) * 8;
 
        c->btree_cache.reserve = reserve;
 }
@@ -63,34 +62,43 @@ static int bch2_btree_cache_cmp_fn(struct rhashtable_compare_arg *arg,
        const struct btree *b = obj;
        const u64 *v = arg->key;
 
-       return PTR_HASH(&b->key) == *v ? 0 : 1;
+       return b->hash_val == *v ? 0 : 1;
 }
 
 static const struct rhashtable_params bch_btree_cache_params = {
        .head_offset    = offsetof(struct btree, hash),
-       .key_offset     = offsetof(struct btree, key.v),
-       .key_len        = sizeof(struct bch_extent_ptr),
+       .key_offset     = offsetof(struct btree, hash_val),
+       .key_len        = sizeof(u64),
        .obj_cmpfn      = bch2_btree_cache_cmp_fn,
 };
 
-static void btree_node_data_alloc(struct bch_fs *c, struct btree *b, gfp_t gfp)
+static int __btree_node_data_alloc(struct bch_fs *c, struct btree *b, gfp_t gfp)
 {
-       struct btree_cache *bc = &c->btree_cache;
+       BUG_ON(b->data || b->aux_data);
 
        b->data = kvpmalloc(btree_bytes(c), gfp);
        if (!b->data)
-               goto err;
+               return -ENOMEM;
 
-       if (bch2_btree_keys_alloc(b, btree_page_order(c), gfp))
-               goto err;
+       if (bch2_btree_keys_alloc(b, btree_page_order(c), gfp)) {
+               kvpfree(b->data, btree_bytes(c));
+               b->data = NULL;
+               return -ENOMEM;
+       }
 
-       bc->used++;
-       list_move(&b->list, &bc->freeable);
-       return;
-err:
-       kvpfree(b->data, btree_bytes(c));
-       b->data = NULL;
-       list_move(&b->list, &bc->freed);
+       return 0;
+}
+
+static void btree_node_data_alloc(struct bch_fs *c, struct btree *b, gfp_t gfp)
+{
+       struct btree_cache *bc = &c->btree_cache;
+
+       if (!__btree_node_data_alloc(c, b, gfp)) {
+               bc->used++;
+               list_move(&b->list, &bc->freeable);
+       } else {
+               list_move(&b->list, &bc->freed);
+       }
 }
 
 static struct btree *btree_node_mem_alloc(struct bch_fs *c, gfp_t gfp)
@@ -99,8 +107,8 @@ static struct btree *btree_node_mem_alloc(struct bch_fs *c, gfp_t gfp)
        if (!b)
                return NULL;
 
-       bkey_extent_init(&b->key);
-       six_lock_init(&b->lock);
+       bkey_btree_ptr_init(&b->key);
+       six_lock_init(&b->c.lock);
        INIT_LIST_HEAD(&b->list);
        INIT_LIST_HEAD(&b->write_blocked);
 
@@ -115,11 +123,14 @@ void bch2_btree_node_hash_remove(struct btree_cache *bc, struct btree *b)
        rhashtable_remove_fast(&bc->table, &b->hash, bch_btree_cache_params);
 
        /* Cause future lookups for this node to fail: */
-       bkey_i_to_extent(&b->key)->v._data[0] = 0;
+       b->hash_val = 0;
 }
 
 int __bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b)
 {
+       BUG_ON(b->hash_val);
+       b->hash_val = btree_ptr_hash_val(&b->key);
+
        return rhashtable_lookup_insert_fast(&bc->table, &b->hash,
                                             bch_btree_cache_params);
 }
@@ -129,8 +140,8 @@ int bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b,
 {
        int ret;
 
-       b->level        = level;
-       b->btree_id     = id;
+       b->c.level      = level;
+       b->c.btree_id   = id;
 
        mutex_lock(&bc->lock);
        ret = __bch2_btree_node_hash_insert(bc, b);
@@ -145,8 +156,9 @@ __flatten
 static inline struct btree *btree_cache_find(struct btree_cache *bc,
                                     const struct bkey_i *k)
 {
-       return rhashtable_lookup_fast(&bc->table, &PTR_HASH(k),
-                                     bch_btree_cache_params);
+       u64 v = btree_ptr_hash_val(k);
+
+       return rhashtable_lookup_fast(&bc->table, &v, bch_btree_cache_params);
 }
 
 /*
@@ -160,10 +172,10 @@ static int __btree_node_reclaim(struct bch_fs *c, struct btree *b, bool flush)
 
        lockdep_assert_held(&bc->lock);
 
-       if (!six_trylock_intent(&b->lock))
+       if (!six_trylock_intent(&b->c.lock))
                return -ENOMEM;
 
-       if (!six_trylock_write(&b->lock))
+       if (!six_trylock_write(&b->c.lock))
                goto out_unlock_intent;
 
        if (btree_node_noevict(b))
@@ -172,6 +184,10 @@ static int __btree_node_reclaim(struct bch_fs *c, struct btree *b, bool flush)
        if (!btree_node_may_write(b))
                goto out_unlock;
 
+       if (btree_node_dirty(b) &&
+           test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags))
+               goto out_unlock;
+
        if (btree_node_dirty(b) ||
            btree_node_write_in_flight(b) ||
            btree_node_read_in_flight(b)) {
@@ -196,13 +212,13 @@ static int __btree_node_reclaim(struct bch_fs *c, struct btree *b, bool flush)
                btree_node_wait_on_io(b);
        }
 out:
-       if (PTR_HASH(&b->key) && !ret)
+       if (b->hash_val && !ret)
                trace_btree_node_reap(c, b);
        return ret;
 out_unlock:
-       six_unlock_write(&b->lock);
+       six_unlock_write(&b->c.lock);
 out_unlock_intent:
-       six_unlock_intent(&b->lock);
+       six_unlock_intent(&b->c.lock);
        ret = -ENOMEM;
        goto out;
 }
@@ -234,7 +250,7 @@ static unsigned long bch2_btree_cache_scan(struct shrinker *shrink,
                return SHRINK_STOP;
 
        /* Return -1 if we can't do anything right now */
-       if (sc->gfp_mask & __GFP_IO)
+       if (sc->gfp_mask & __GFP_FS)
                mutex_lock(&bc->lock);
        else if (!mutex_trylock(&bc->lock))
                return -1;
@@ -260,8 +276,8 @@ static unsigned long bch2_btree_cache_scan(struct shrinker *shrink,
                if (++i > 3 &&
                    !btree_node_reclaim(c, b)) {
                        btree_node_data_free(c, b);
-                       six_unlock_write(&b->lock);
-                       six_unlock_intent(&b->lock);
+                       six_unlock_write(&b->c.lock);
+                       six_unlock_intent(&b->c.lock);
                        freed++;
                }
        }
@@ -287,13 +303,13 @@ restart:
                        mutex_unlock(&bc->lock);
 
                        bch2_btree_node_hash_remove(bc, b);
-                       six_unlock_write(&b->lock);
-                       six_unlock_intent(&b->lock);
+                       six_unlock_write(&b->c.lock);
+                       six_unlock_intent(&b->c.lock);
 
                        if (freed >= nr)
                                goto out;
 
-                       if (sc->gfp_mask & __GFP_IO)
+                       if (sc->gfp_mask & __GFP_FS)
                                mutex_lock(&bc->lock);
                        else if (!mutex_trylock(&bc->lock))
                                goto out;
@@ -506,7 +522,9 @@ struct btree *bch2_btree_node_mem_alloc(struct bch_fs *c)
        struct btree_cache *bc = &c->btree_cache;
        struct btree *b;
        u64 start_time = local_clock();
+       unsigned flags;
 
+       flags = memalloc_nofs_save();
        mutex_lock(&bc->lock);
 
        /*
@@ -515,35 +533,47 @@ struct btree *bch2_btree_node_mem_alloc(struct bch_fs *c)
         */
        list_for_each_entry(b, &bc->freeable, list)
                if (!btree_node_reclaim(c, b))
-                       goto out_unlock;
+                       goto got_node;
 
        /*
         * We never free struct btree itself, just the memory that holds the on
         * disk node. Check the freed list before allocating a new one:
         */
        list_for_each_entry(b, &bc->freed, list)
-               if (!btree_node_reclaim(c, b)) {
-                       btree_node_data_alloc(c, b, __GFP_NOWARN|GFP_NOIO);
-                       if (b->data)
-                               goto out_unlock;
+               if (!btree_node_reclaim(c, b))
+                       goto got_node;
 
-                       six_unlock_write(&b->lock);
-                       six_unlock_intent(&b->lock);
+       b = NULL;
+got_node:
+       if (b)
+               list_del_init(&b->list);
+       mutex_unlock(&bc->lock);
+
+       if (!b) {
+               b = kzalloc(sizeof(struct btree), GFP_KERNEL);
+               if (!b)
                        goto err;
-               }
 
-       b = btree_node_mem_alloc(c, __GFP_NOWARN|GFP_NOIO);
-       if (!b)
-               goto err;
+               bkey_btree_ptr_init(&b->key);
+               six_lock_init(&b->c.lock);
+               INIT_LIST_HEAD(&b->list);
+               INIT_LIST_HEAD(&b->write_blocked);
+
+               BUG_ON(!six_trylock_intent(&b->c.lock));
+               BUG_ON(!six_trylock_write(&b->c.lock));
+       }
+
+       if (!b->data) {
+               if (__btree_node_data_alloc(c, b, __GFP_NOWARN|GFP_KERNEL))
+                       goto err;
+
+               mutex_lock(&bc->lock);
+               bc->used++;
+               mutex_unlock(&bc->lock);
+       }
 
-       BUG_ON(!six_trylock_intent(&b->lock));
-       BUG_ON(!six_trylock_write(&b->lock));
-out_unlock:
        BUG_ON(btree_node_hashed(b));
        BUG_ON(btree_node_write_in_flight(b));
-
-       list_del_init(&b->list);
-       mutex_unlock(&bc->lock);
 out:
        b->flags                = 0;
        b->written              = 0;
@@ -551,14 +581,22 @@ out:
        b->sib_u64s[0]          = 0;
        b->sib_u64s[1]          = 0;
        b->whiteout_u64s        = 0;
-       b->uncompacted_whiteout_u64s = 0;
        bch2_btree_keys_init(b, &c->expensive_debug_checks);
 
        bch2_time_stats_update(&c->times[BCH_TIME_btree_node_mem_alloc],
                               start_time);
 
+       memalloc_nofs_restore(flags);
        return b;
 err:
+       mutex_lock(&bc->lock);
+
+       if (b) {
+               list_add(&b->list, &bc->freed);
+               six_unlock_write(&b->c.lock);
+               six_unlock_intent(&b->c.lock);
+       }
+
        /* Try to cannibalize another cached btree node: */
        if (bc->alloc_lock == current) {
                b = btree_node_cannibalize(c);
@@ -572,66 +610,81 @@ err:
        }
 
        mutex_unlock(&bc->lock);
+       memalloc_nofs_restore(flags);
        return ERR_PTR(-ENOMEM);
 }
 
 /* Slowpath, don't want it inlined into btree_iter_traverse() */
 static noinline struct btree *bch2_btree_node_fill(struct bch_fs *c,
-                                                  struct btree_iter *iter,
-                                                  const struct bkey_i *k,
-                                                  unsigned level,
-                                                  enum six_lock_type lock_type)
+                               struct btree_iter *iter,
+                               const struct bkey_i *k,
+                               enum btree_id btree_id,
+                               unsigned level,
+                               enum six_lock_type lock_type,
+                               bool sync)
 {
        struct btree_cache *bc = &c->btree_cache;
        struct btree *b;
 
+       BUG_ON(level + 1 >= BTREE_MAX_DEPTH);
        /*
         * Parent node must be locked, else we could read in a btree node that's
         * been freed:
         */
-       BUG_ON(!btree_node_locked(iter, level + 1));
+       if (iter && !bch2_btree_node_relock(iter, level + 1))
+               return ERR_PTR(-EINTR);
 
        b = bch2_btree_node_mem_alloc(c);
        if (IS_ERR(b))
                return b;
 
        bkey_copy(&b->key, k);
-       if (bch2_btree_node_hash_insert(bc, b, level, iter->btree_id)) {
+       if (bch2_btree_node_hash_insert(bc, b, level, btree_id)) {
                /* raced with another fill: */
 
                /* mark as unhashed... */
-               bkey_i_to_extent(&b->key)->v._data[0] = 0;
+               b->hash_val = 0;
 
                mutex_lock(&bc->lock);
                list_add(&b->list, &bc->freeable);
                mutex_unlock(&bc->lock);
 
-               six_unlock_write(&b->lock);
-               six_unlock_intent(&b->lock);
+               six_unlock_write(&b->c.lock);
+               six_unlock_intent(&b->c.lock);
                return NULL;
        }
 
        /*
-        * If the btree node wasn't cached, we can't drop our lock on
-        * the parent until after it's added to the cache - because
-        * otherwise we could race with a btree_split() freeing the node
-        * we're trying to lock.
+        * Unlock before doing IO:
         *
-        * But the deadlock described below doesn't exist in this case,
-        * so it's safe to not drop the parent lock until here:
+        * XXX: ideally should be dropping all btree node locks here
         */
-       if (btree_node_read_locked(iter, level + 1))
+       if (iter && btree_node_read_locked(iter, level + 1))
                btree_node_unlock(iter, level + 1);
 
-       bch2_btree_node_read(c, b, true);
-       six_unlock_write(&b->lock);
+       bch2_btree_node_read(c, b, sync);
+
+       six_unlock_write(&b->c.lock);
+
+       if (!sync) {
+               six_unlock_intent(&b->c.lock);
+               return NULL;
+       }
 
        if (lock_type == SIX_LOCK_read)
-               six_lock_downgrade(&b->lock);
+               six_lock_downgrade(&b->c.lock);
 
        return b;
 }
 
+static int lock_node_check_fn(struct six_lock *lock, void *p)
+{
+       struct btree *b = container_of(lock, struct btree, c.lock);
+       const struct bkey_i *k = p;
+
+       return b->hash_val == btree_ptr_hash_val(k) ? 0 : -1;
+}
+
 /**
  * bch_btree_node_get - find a btree node in the cache and lock it, reading it
  * in from disk if necessary.
@@ -649,21 +702,21 @@ struct btree *bch2_btree_node_get(struct bch_fs *c, struct btree_iter *iter,
        struct btree *b;
        struct bset_tree *t;
 
-       /* btree_node_fill() requires parent to be locked: */
-       EBUG_ON(!btree_node_locked(iter, level + 1));
        EBUG_ON(level >= BTREE_MAX_DEPTH);
+
+       b = btree_node_mem_ptr(k);
+       if (b)
+               goto lock_node;
 retry:
-       rcu_read_lock();
        b = btree_cache_find(bc, k);
-       rcu_read_unlock();
-
        if (unlikely(!b)) {
                /*
                 * We must have the parent locked to call bch2_btree_node_fill(),
                 * else we could read in a btree node from disk that's been
                 * freed:
                 */
-               b = bch2_btree_node_fill(c, iter, k, level, lock_type);
+               b = bch2_btree_node_fill(c, iter, k, iter->btree_id,
+                                        level, lock_type, true);
 
                /* We raced and found the btree node in the cache */
                if (!b)
@@ -672,6 +725,7 @@ retry:
                if (IS_ERR(b))
                        return b;
        } else {
+lock_node:
                /*
                 * There's a potential deadlock with splits and insertions into
                 * interior nodes we have to avoid:
@@ -693,7 +747,7 @@ retry:
                 * free it:
                 *
                 * To guard against this, btree nodes are evicted from the cache
-                * when they're freed - and PTR_HASH() is zeroed out, which we
+                * when they're freed - and b->hash_val is zeroed out, which we
                 * check for after we lock the node.
                 *
                 * Then, bch2_btree_node_relock() on the parent will fail - because
@@ -703,20 +757,26 @@ retry:
                if (btree_node_read_locked(iter, level + 1))
                        btree_node_unlock(iter, level + 1);
 
-               if (!btree_node_lock(b, k->k.p, level, iter, lock_type))
+               if (!btree_node_lock(b, k->k.p, level, iter, lock_type,
+                                    lock_node_check_fn, (void *) k)) {
+                       if (b->hash_val != btree_ptr_hash_val(k))
+                               goto retry;
                        return ERR_PTR(-EINTR);
+               }
 
-               if (unlikely(PTR_HASH(&b->key) != PTR_HASH(k) ||
-                            b->level != level ||
+               if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
+                            b->c.level != level ||
                             race_fault())) {
-                       six_unlock_type(&b->lock, lock_type);
+                       six_unlock_type(&b->c.lock, lock_type);
                        if (bch2_btree_node_relock(iter, level + 1))
                                goto retry;
 
+                       trace_trans_restart_btree_node_reused(iter->trans->ip);
                        return ERR_PTR(-EINTR);
                }
        }
 
+       /* XXX: waiting on IO with btree locks held: */
        wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
                       TASK_UNINTERRUPTIBLE);
 
@@ -731,15 +791,86 @@ retry:
        }
 
        /* avoid atomic set bit if it's not needed: */
-       if (btree_node_accessed(b))
+       if (!btree_node_accessed(b))
                set_btree_node_accessed(b);
 
        if (unlikely(btree_node_read_error(b))) {
-               six_unlock_type(&b->lock, lock_type);
+               six_unlock_type(&b->c.lock, lock_type);
                return ERR_PTR(-EIO);
        }
 
-       EBUG_ON(b->btree_id != iter->btree_id ||
+       EBUG_ON(b->c.btree_id != iter->btree_id ||
+               BTREE_NODE_LEVEL(b->data) != level ||
+               bkey_cmp(b->data->max_key, k->k.p));
+
+       return b;
+}
+
+struct btree *bch2_btree_node_get_noiter(struct bch_fs *c,
+                                        const struct bkey_i *k,
+                                        enum btree_id btree_id,
+                                        unsigned level)
+{
+       struct btree_cache *bc = &c->btree_cache;
+       struct btree *b;
+       struct bset_tree *t;
+       int ret;
+
+       EBUG_ON(level >= BTREE_MAX_DEPTH);
+
+       b = btree_node_mem_ptr(k);
+       if (b)
+               goto lock_node;
+retry:
+       b = btree_cache_find(bc, k);
+       if (unlikely(!b)) {
+               b = bch2_btree_node_fill(c, NULL, k, btree_id,
+                                        level, SIX_LOCK_read, true);
+
+               /* We raced and found the btree node in the cache */
+               if (!b)
+                       goto retry;
+
+               if (IS_ERR(b))
+                       return b;
+       } else {
+lock_node:
+               ret = six_lock_read(&b->c.lock, lock_node_check_fn, (void *) k);
+               if (ret)
+                       goto retry;
+
+               if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
+                            b->c.btree_id != btree_id ||
+                            b->c.level != level)) {
+                       six_unlock_read(&b->c.lock);
+                       goto retry;
+               }
+       }
+
+       /* XXX: waiting on IO with btree locks held: */
+       wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
+                      TASK_UNINTERRUPTIBLE);
+
+       prefetch(b->aux_data);
+
+       for_each_bset(b, t) {
+               void *p = (u64 *) b->aux_data + t->aux_data_offset;
+
+               prefetch(p + L1_CACHE_BYTES * 0);
+               prefetch(p + L1_CACHE_BYTES * 1);
+               prefetch(p + L1_CACHE_BYTES * 2);
+       }
+
+       /* avoid atomic set bit if it's not needed: */
+       if (!btree_node_accessed(b))
+               set_btree_node_accessed(b);
+
+       if (unlikely(btree_node_read_error(b))) {
+               six_unlock_read(&b->c.lock);
+               return ERR_PTR(-EIO);
+       }
+
+       EBUG_ON(b->c.btree_id != btree_id ||
                BTREE_NODE_LEVEL(b->data) != level ||
                bkey_cmp(b->data->max_key, k->k.p));
 
@@ -751,150 +882,173 @@ struct btree *bch2_btree_node_get_sibling(struct bch_fs *c,
                                          struct btree *b,
                                          enum btree_node_sibling sib)
 {
+       struct btree_trans *trans = iter->trans;
        struct btree *parent;
        struct btree_node_iter node_iter;
        struct bkey_packed *k;
        BKEY_PADDED(k) tmp;
-       struct btree *ret;
-       unsigned level = b->level;
+       struct btree *ret = NULL;
+       unsigned level = b->c.level;
 
        parent = btree_iter_node(iter, level + 1);
        if (!parent)
                return NULL;
 
-       if (!bch2_btree_node_relock(iter, level + 1)) {
-               bch2_btree_iter_set_locks_want(iter, level + 2);
+       /*
+        * There's a corner case where a btree_iter might have a node locked
+        * that is just outside its current pos - when
+        * bch2_btree_iter_set_pos_same_leaf() gets to the end of the node.
+        *
+        * But the lock ordering checks in __bch2_btree_node_lock() go off of
+        * iter->pos, not the node's key: so if the iterator is marked as
+        * needing to be traversed, we risk deadlock if we don't bail out here:
+        */
+       if (iter->uptodate >= BTREE_ITER_NEED_TRAVERSE)
                return ERR_PTR(-EINTR);
+
+       if (!bch2_btree_node_relock(iter, level + 1)) {
+               ret = ERR_PTR(-EINTR);
+               goto out;
        }
 
-       node_iter = iter->l[parent->level].iter;
+       node_iter = iter->l[parent->c.level].iter;
 
        k = bch2_btree_node_iter_peek_all(&node_iter, parent);
        BUG_ON(bkey_cmp_left_packed(parent, k, &b->key.k.p));
 
-       do {
-               k = sib == btree_prev_sib
-                       ? bch2_btree_node_iter_prev_all(&node_iter, parent)
-                       : (bch2_btree_node_iter_advance(&node_iter, parent),
-                          bch2_btree_node_iter_peek_all(&node_iter, parent));
-               if (!k)
-                       return NULL;
-       } while (bkey_deleted(k));
+       k = sib == btree_prev_sib
+               ? bch2_btree_node_iter_prev(&node_iter, parent)
+               : (bch2_btree_node_iter_advance(&node_iter, parent),
+                  bch2_btree_node_iter_peek(&node_iter, parent));
+       if (!k)
+               goto out;
 
        bch2_bkey_unpack(parent, &tmp.k, k);
 
-       ret = bch2_btree_node_get(c, iter, &tmp.k, level, SIX_LOCK_intent);
+       ret = bch2_btree_node_get(c, iter, &tmp.k, level,
+                                 SIX_LOCK_intent);
 
-       if (IS_ERR(ret) && PTR_ERR(ret) == -EINTR) {
-               btree_node_unlock(iter, level);
+       if (PTR_ERR_OR_ZERO(ret) == -EINTR && !trans->nounlock) {
+               struct btree_iter *linked;
 
-               if (!bch2_btree_node_relock(iter, level + 1)) {
-                       bch2_btree_iter_set_locks_want(iter, level + 2);
-                       return ERR_PTR(-EINTR);
+               if (!bch2_btree_node_relock(iter, level + 1))
+                       goto out;
+
+               /*
+                * We might have got -EINTR because trylock failed, and we're
+                * holding other locks that would cause us to deadlock:
+                */
+               trans_for_each_iter(trans, linked)
+                       if (btree_iter_cmp(iter, linked) < 0)
+                               __bch2_btree_iter_unlock(linked);
+
+               if (sib == btree_prev_sib)
+                       btree_node_unlock(iter, level);
+
+               ret = bch2_btree_node_get(c, iter, &tmp.k, level,
+                                         SIX_LOCK_intent);
+
+               /*
+                * before btree_iter_relock() calls btree_iter_verify_locks():
+                */
+               if (btree_lock_want(iter, level + 1) == BTREE_NODE_UNLOCKED)
+                       btree_node_unlock(iter, level + 1);
+
+               if (!bch2_btree_node_relock(iter, level)) {
+                       btree_iter_set_dirty(iter, BTREE_ITER_NEED_RELOCK);
+
+                       if (!IS_ERR(ret)) {
+                               six_unlock_intent(&ret->c.lock);
+                               ret = ERR_PTR(-EINTR);
+                       }
                }
 
-               ret = bch2_btree_node_get(c, iter, &tmp.k, level, SIX_LOCK_intent);
+               bch2_trans_relock(trans);
        }
+out:
+       if (btree_lock_want(iter, level + 1) == BTREE_NODE_UNLOCKED)
+               btree_node_unlock(iter, level + 1);
 
-       if (!bch2_btree_node_relock(iter, level)) {
-               btree_iter_set_dirty(iter, BTREE_ITER_NEED_RELOCK);
+       if (PTR_ERR_OR_ZERO(ret) == -EINTR)
+               bch2_btree_iter_upgrade(iter, level + 2);
 
-               if (!IS_ERR(ret)) {
-                       six_unlock_intent(&ret->lock);
-                       ret = ERR_PTR(-EINTR);
-               }
+       BUG_ON(!IS_ERR(ret) && !btree_node_locked(iter, level));
+
+       if (!IS_ERR_OR_NULL(ret)) {
+               struct btree *n1 = ret, *n2 = b;
+
+               if (sib != btree_prev_sib)
+                       swap(n1, n2);
+
+               BUG_ON(bkey_cmp(bkey_successor(n1->key.k.p),
+                               n2->data->min_key));
        }
 
+       bch2_btree_trans_verify_locks(trans);
+
        return ret;
 }
 
-void bch2_btree_node_prefetch(struct bch_fs *c, const struct bkey_i *k,
-                             unsigned level, enum btree_id btree_id)
+void bch2_btree_node_prefetch(struct bch_fs *c, struct btree_iter *iter,
+                             const struct bkey_i *k, unsigned level)
 {
        struct btree_cache *bc = &c->btree_cache;
        struct btree *b;
 
+       BUG_ON(!btree_node_locked(iter, level + 1));
        BUG_ON(level >= BTREE_MAX_DEPTH);
 
-       rcu_read_lock();
        b = btree_cache_find(bc, k);
-       rcu_read_unlock();
-
        if (b)
                return;
 
-       b = bch2_btree_node_mem_alloc(c);
-       if (IS_ERR(b))
-               return;
-
-       bkey_copy(&b->key, k);
-       if (bch2_btree_node_hash_insert(bc, b, level, btree_id)) {
-               /* raced with another fill: */
-
-               /* mark as unhashed... */
-               bkey_i_to_extent(&b->key)->v._data[0] = 0;
-
-               mutex_lock(&bc->lock);
-               list_add(&b->list, &bc->freeable);
-               mutex_unlock(&bc->lock);
-               goto out;
-       }
-
-       bch2_btree_node_read(c, b, false);
-out:
-       six_unlock_write(&b->lock);
-       six_unlock_intent(&b->lock);
+       bch2_btree_node_fill(c, iter, k, iter->btree_id,
+                            level, SIX_LOCK_read, false);
 }
 
-int bch2_print_btree_node(struct bch_fs *c, struct btree *b,
-                         char *buf, size_t len)
+void bch2_btree_node_to_text(struct printbuf *out, struct bch_fs *c,
+                            struct btree *b)
 {
        const struct bkey_format *f = &b->format;
        struct bset_stats stats;
-       char ptrs[100];
 
        memset(&stats, 0, sizeof(stats));
 
-       bch2_val_to_text(c, BKEY_TYPE_BTREE, ptrs, sizeof(ptrs),
-                       bkey_i_to_s_c(&b->key));
        bch2_btree_keys_stats(b, &stats);
 
-       return scnprintf(buf, len,
-                        "l %u %llu:%llu - %llu:%llu:\n"
-                        "    ptrs: %s\n"
-                        "    format: u64s %u fields %u %u %u %u %u\n"
-                        "    unpack fn len: %u\n"
-                        "    bytes used %zu/%zu (%zu%% full)\n"
-                        "    sib u64s: %u, %u (merge threshold %zu)\n"
-                        "    nr packed keys %u\n"
-                        "    nr unpacked keys %u\n"
-                        "    floats %zu\n"
-                        "    failed unpacked %zu\n"
-                        "    failed prev %zu\n"
-                        "    failed overflow %zu\n",
-                        b->level,
-                        b->data->min_key.inode,
-                        b->data->min_key.offset,
-                        b->data->max_key.inode,
-                        b->data->max_key.offset,
-                        ptrs,
-                        f->key_u64s,
-                        f->bits_per_field[0],
-                        f->bits_per_field[1],
-                        f->bits_per_field[2],
-                        f->bits_per_field[3],
-                        f->bits_per_field[4],
-                        b->unpack_fn_len,
-                        b->nr.live_u64s * sizeof(u64),
-                        btree_bytes(c) - sizeof(struct btree_node),
-                        b->nr.live_u64s * 100 / btree_max_u64s(c),
-                        b->sib_u64s[0],
-                        b->sib_u64s[1],
-                        BTREE_FOREGROUND_MERGE_THRESHOLD(c),
-                        b->nr.packed_keys,
-                        b->nr.unpacked_keys,
-                        stats.floats,
-                        stats.failed_unpacked,
-                        stats.failed_prev,
-                        stats.failed_overflow);
+       pr_buf(out,
+              "l %u %llu:%llu - %llu:%llu:\n"
+              "    ptrs: ",
+              b->c.level,
+              b->data->min_key.inode,
+              b->data->min_key.offset,
+              b->data->max_key.inode,
+              b->data->max_key.offset);
+       bch2_val_to_text(out, c, bkey_i_to_s_c(&b->key));
+       pr_buf(out, "\n"
+              "    format: u64s %u fields %u %u %u %u %u\n"
+              "    unpack fn len: %u\n"
+              "    bytes used %zu/%zu (%zu%% full)\n"
+              "    sib u64s: %u, %u (merge threshold %zu)\n"
+              "    nr packed keys %u\n"
+              "    nr unpacked keys %u\n"
+              "    floats %zu\n"
+              "    failed unpacked %zu\n",
+              f->key_u64s,
+              f->bits_per_field[0],
+              f->bits_per_field[1],
+              f->bits_per_field[2],
+              f->bits_per_field[3],
+              f->bits_per_field[4],
+              b->unpack_fn_len,
+              b->nr.live_u64s * sizeof(u64),
+              btree_bytes(c) - sizeof(struct btree_node),
+              b->nr.live_u64s * 100 / btree_max_u64s(c),
+              b->sib_u64s[0],
+              b->sib_u64s[1],
+              BTREE_FOREGROUND_MERGE_THRESHOLD(c),
+              b->nr.packed_keys,
+              b->nr.unpacked_keys,
+              stats.floats,
+              stats.failed);
 }