#include "debug.h"
#include "errcode.h"
#include "error.h"
+#include "trace.h"
#include <linux/prefetch.h>
#include <linux/sched/mm.h>
-#include <trace/events/bcachefs.h>
+#include <linux/seq_buf.h>
#define BTREE_CACHE_NOT_FREED_INCREMENT(counter) \
do { \
{
unsigned i, reserve = 16;
- if (!c->btree_roots[0].b)
+ if (!c->btree_roots_known[0].b)
reserve += 8;
- for (i = 0; i < BTREE_ID_NR; i++)
- if (c->btree_roots[i].b)
- reserve += min_t(unsigned, 1,
- c->btree_roots[i].b->c.level) * 8;
+ for (i = 0; i < btree_id_nr_alive(c); i++) {
+ struct btree_root *r = bch2_btree_id_root(c, i);
+
+ if (r->b)
+ reserve += min_t(unsigned, 1, r->b->c.level) * 8;
+ }
c->btree_cache.reserve = reserve;
}
EBUG_ON(btree_node_write_in_flight(b));
+ clear_btree_node_just_written(b);
+
kvpfree(b->data, btree_bytes(c));
b->data = NULL;
#ifdef __KERNEL__
- vfree(b->aux_data);
+ kvfree(b->aux_data);
#else
munmap(b->aux_data, btree_aux_data_bytes(b));
#endif
b->data = kvpmalloc(btree_bytes(c), gfp);
if (!b->data)
- return -ENOMEM;
+ return -BCH_ERR_ENOMEM_btree_node_mem_alloc;
#ifdef __KERNEL__
- b->aux_data = vmalloc_exec(btree_aux_data_bytes(b), gfp);
+ b->aux_data = kvmalloc(btree_aux_data_bytes(b), gfp);
#else
b->aux_data = mmap(NULL, btree_aux_data_bytes(b),
PROT_READ|PROT_WRITE|PROT_EXEC,
if (!b->aux_data) {
kvpfree(b->data, btree_bytes(c));
b->data = NULL;
- return -ENOMEM;
+ return -BCH_ERR_ENOMEM_btree_node_mem_alloc;
}
return 0;
static struct btree *__btree_node_mem_alloc(struct bch_fs *c, gfp_t gfp)
{
- struct btree *b = kzalloc(sizeof(struct btree), gfp);
+ struct btree *b;
+
+ b = kzalloc(sizeof(struct btree), gfp);
if (!b)
return NULL;
bkey_btree_ptr_init(&b->key);
- __six_lock_init(&b->c.lock, "b->c.lock", &bch2_btree_node_lock_key);
-#ifdef CONFIG_DEBUG_LOCK_ALLOC
- lockdep_set_no_check_recursion(&b->c.lock.dep_map);
-#endif
INIT_LIST_HEAD(&b->list);
INIT_LIST_HEAD(&b->write_blocked);
b->byte_order = ilog2(btree_bytes(c));
struct btree *__bch2_btree_node_mem_alloc(struct bch_fs *c)
{
struct btree_cache *bc = &c->btree_cache;
- struct btree *b = __btree_node_mem_alloc(c, GFP_KERNEL);
+ struct btree *b;
+
+ b = __btree_node_mem_alloc(c, GFP_KERNEL);
if (!b)
return NULL;
return NULL;
}
+ bch2_btree_lock_init(&b->c, 0);
+
bc->used++;
list_add(&b->list, &bc->freeable);
return b;
void bch2_btree_node_hash_remove(struct btree_cache *bc, struct btree *b)
{
int ret = rhashtable_remove_fast(&bc->table, &b->hash, bch_btree_cache_params);
+
BUG_ON(ret);
/* Cause future lookups for this node to fail: */
BTREE_CACHE_NOT_FREED_INCREMENT(read_in_flight);
else if (btree_node_write_in_flight(b))
BTREE_CACHE_NOT_FREED_INCREMENT(write_in_flight);
- return -ENOMEM;
+ return -BCH_ERR_ENOMEM_btree_node_reclaim;
}
/* XXX: waiting on IO with btree cache lock held */
if (!six_trylock_intent(&b->c.lock)) {
BTREE_CACHE_NOT_FREED_INCREMENT(lock_intent);
- return -ENOMEM;
+ return -BCH_ERR_ENOMEM_btree_node_reclaim;
}
if (!six_trylock_write(&b->c.lock)) {
* the post write cleanup:
*/
if (bch2_verify_btree_ondisk)
- bch2_btree_node_write(c, b, SIX_LOCK_intent, 0);
+ bch2_btree_node_write(c, b, SIX_LOCK_intent,
+ BTREE_WRITE_cache_reclaim);
else
- __bch2_btree_node_write(c, b, 0);
+ __bch2_btree_node_write(c, b,
+ BTREE_WRITE_cache_reclaim);
six_unlock_write(&b->c.lock);
six_unlock_intent(&b->c.lock);
six_unlock_write(&b->c.lock);
out_unlock_intent:
six_unlock_intent(&b->c.lock);
- ret = -ENOMEM;
+ ret = -BCH_ERR_ENOMEM_btree_node_reclaim;
goto out;
}
six_trylock_read(&b->c.lock)) {
list_move(&bc->live, &b->list);
mutex_unlock(&bc->lock);
- __bch2_btree_node_write(c, b, 0);
+ __bch2_btree_node_write(c, b, BTREE_WRITE_cache_reclaim);
six_unlock_read(&b->c.lock);
if (touched >= nr)
goto out_nounlock;
return btree_cache_can_free(bc);
}
-static void bch2_btree_cache_shrinker_to_text(struct printbuf *out, struct shrinker *shrink)
+static void bch2_btree_cache_shrinker_to_text(struct seq_buf *s, struct shrinker *shrink)
{
struct bch_fs *c = container_of(shrink, struct bch_fs,
btree_cache.shrink);
+ char *cbuf;
+ size_t buflen = seq_buf_get_buf(s, &cbuf);
+ struct printbuf out = PRINTBUF_EXTERN(cbuf, buflen);
- bch2_btree_cache_to_text(out, &c->btree_cache);
+ bch2_btree_cache_to_text(&out, &c->btree_cache);
+ seq_buf_commit(s, out.pos);
}
void bch2_fs_btree_cache_exit(struct bch_fs *c)
struct btree *b;
unsigned i, flags;
- if (bc->shrink.list.next)
- unregister_shrinker(&bc->shrink);
+ unregister_shrinker(&bc->shrink);
/* vfree() can allocate memory: */
flags = memalloc_nofs_save();
kvpfree(c->verify_ondisk, btree_bytes(c));
- for (i = 0; i < BTREE_ID_NR; i++)
- if (c->btree_roots[i].b)
- list_add(&c->btree_roots[i].b->list, &bc->live);
+ for (i = 0; i < btree_id_nr_alive(c); i++) {
+ struct btree_root *r = bch2_btree_id_root(c, i);
+
+ if (r->b)
+ list_add(&r->b->list, &bc->live);
+ }
list_splice(&bc->freeable, &bc->live);
while (!list_empty(&bc->freed_nonpcpu)) {
b = list_first_entry(&bc->freed_nonpcpu, struct btree, list);
list_del(&b->list);
- six_lock_pcpu_free(&b->c.lock);
+ six_lock_exit(&b->c.lock);
kfree(b);
}
for (i = 0; i < bc->reserve; i++)
if (!__bch2_btree_node_mem_alloc(c)) {
- ret = -ENOMEM;
+ ret = -BCH_ERR_ENOMEM_fs_btree_cache_init;
goto out;
}
if (!cl) {
trace_and_count(c, btree_cache_cannibalize_lock_fail, c);
- return -ENOMEM;
+ return -BCH_ERR_ENOMEM_btree_cache_cannibalize_lock;
}
closure_wait(&bc->alloc_wait, cl);
}
trace_and_count(c, btree_cache_cannibalize_lock_fail, c);
- return -EAGAIN;
+ return -BCH_ERR_btree_cache_cannibalize_lock_blocked;
success:
trace_and_count(c, btree_cache_cannibalize_lock, c);
}
}
-struct btree *bch2_btree_node_mem_alloc(struct bch_fs *c, bool pcpu_read_locks)
+struct btree *bch2_btree_node_mem_alloc(struct btree_trans *trans, bool pcpu_read_locks)
{
+ struct bch_fs *c = trans->c;
struct btree_cache *bc = &c->btree_cache;
struct list_head *freed = pcpu_read_locks
? &bc->freed_pcpu
goto got_node;
}
- b = __btree_node_mem_alloc(c, __GFP_NOWARN);
+ b = __btree_node_mem_alloc(c, GFP_NOWAIT|__GFP_NOWARN);
if (!b) {
mutex_unlock(&bc->lock);
+ bch2_trans_unlock(trans);
b = __btree_node_mem_alloc(c, GFP_KERNEL);
if (!b)
goto err;
mutex_lock(&bc->lock);
}
- if (pcpu_read_locks)
- six_lock_pcpu_alloc(&b->c.lock);
+ bch2_btree_lock_init(&b->c, pcpu_read_locks ? SIX_LOCK_INIT_PCPU : 0);
BUG_ON(!six_trylock_intent(&b->c.lock));
BUG_ON(!six_trylock_write(&b->c.lock));
mutex_unlock(&bc->lock);
- if (btree_node_data_alloc(c, b, __GFP_NOWARN|GFP_KERNEL))
- goto err;
+ if (btree_node_data_alloc(c, b, GFP_NOWAIT|__GFP_NOWARN)) {
+ bch2_trans_unlock(trans);
+ if (btree_node_data_alloc(c, b, GFP_KERNEL|__GFP_NOWARN))
+ goto err;
+ }
mutex_lock(&bc->lock);
bc->used++;
/* Try to cannibalize another cached btree node: */
if (bc->alloc_lock == current) {
b2 = btree_node_cannibalize(c);
+ clear_btree_node_just_written(b2);
bch2_btree_node_hash_remove(bc, b2);
if (b) {
mutex_unlock(&bc->lock);
memalloc_nofs_restore(flags);
- return ERR_PTR(-ENOMEM);
+ return ERR_PTR(-BCH_ERR_ENOMEM_btree_node_mem_alloc);
}
/* Slowpath, don't want it inlined into btree_iter_traverse() */
-static noinline struct btree *bch2_btree_node_fill(struct bch_fs *c,
- struct btree_trans *trans,
+static noinline struct btree *bch2_btree_node_fill(struct btree_trans *trans,
struct btree_path *path,
const struct bkey_i *k,
enum btree_id btree_id,
enum six_lock_type lock_type,
bool sync)
{
+ struct bch_fs *c = trans->c;
struct btree_cache *bc = &c->btree_cache;
struct btree *b;
u32 seq;
* Parent node must be locked, else we could read in a btree node that's
* been freed:
*/
- if (trans && !bch2_btree_node_relock(trans, path, level + 1)) {
+ if (path && !bch2_btree_node_relock(trans, path, level + 1)) {
trace_and_count(c, trans_restart_relock_parent_for_fill, trans, _THIS_IP_, path);
return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_fill_relock));
}
- b = bch2_btree_node_mem_alloc(c, level != 0);
+ b = bch2_btree_node_mem_alloc(trans, level != 0);
- if (trans && b == ERR_PTR(-ENOMEM)) {
+ if (bch2_err_matches(PTR_ERR_OR_ZERO(b), ENOMEM)) {
trans->memory_allocation_failure = true;
trace_and_count(c, trans_restart_memory_allocation_failure, trans, _THIS_IP_, path);
return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_fill_mem_alloc_fail));
if (IS_ERR(b))
return b;
+ /*
+ * Btree nodes read in from disk should not have the accessed bit set
+ * initially, so that linear scans don't thrash the cache:
+ */
+ clear_btree_node_accessed(b);
+
bkey_copy(&b->key, k);
if (bch2_btree_node_hash_insert(bc, b, level, btree_id)) {
/* raced with another fill: */
set_btree_node_read_in_flight(b);
six_unlock_write(&b->c.lock);
- seq = b->c.lock.state.seq;
+ seq = six_lock_seq(&b->c.lock);
six_unlock_intent(&b->c.lock);
/* Unlock before doing IO: */
if (trans && sync)
- bch2_trans_unlock(trans);
+ bch2_trans_unlock_noassert(trans);
bch2_btree_node_read(c, b, sync);
if (!sync)
return NULL;
- if (trans) {
+ if (path) {
int ret = bch2_trans_relock(trans) ?:
bch2_btree_path_relock_intent(trans, path);
if (ret) {
}
if (!six_relock_type(&b->c.lock, lock_type, seq)) {
- if (trans)
+ if (path)
trace_and_count(c, trans_restart_relock_after_fill, trans, _THIS_IP_, path);
return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_relock_after_fill));
}
{
if (b->c.btree_id != BTREE_NODE_ID(b->data) ||
b->c.level != BTREE_NODE_LEVEL(b->data) ||
- bpos_cmp(b->data->max_key, b->key.k.p) ||
+ !bpos_eq(b->data->max_key, b->key.k.p) ||
(b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
- bpos_cmp(b->data->min_key,
+ !bpos_eq(b->data->min_key,
bkey_i_to_btree_ptr_v2(&b->key)->v.min_key)))
btree_bad_header(c, b);
}
-/**
- * bch_btree_node_get - find a btree node in the cache and lock it, reading it
- * in from disk if necessary.
- *
- * If IO is necessary and running under generic_make_request, returns -EAGAIN.
- *
- * The btree node will have either a read or a write lock held, depending on
- * the @write parameter.
- */
-struct btree *bch2_btree_node_get(struct btree_trans *trans, struct btree_path *path,
- const struct bkey_i *k, unsigned level,
- enum six_lock_type lock_type,
- unsigned long trace_ip)
+static struct btree *__bch2_btree_node_get(struct btree_trans *trans, struct btree_path *path,
+ const struct bkey_i *k, unsigned level,
+ enum six_lock_type lock_type,
+ unsigned long trace_ip)
{
struct bch_fs *c = trans->c;
struct btree_cache *bc = &c->btree_cache;
struct btree *b;
struct bset_tree *t;
+ bool need_relock = false;
int ret;
EBUG_ON(level >= BTREE_MAX_DEPTH);
-
- b = btree_node_mem_ptr(k);
-
- /*
- * Check b->hash_val _before_ calling btree_node_lock() - this might not
- * be the node we want anymore, and trying to lock the wrong node could
- * cause an unneccessary transaction restart:
- */
- if (likely(c->opts.btree_node_mem_ptr_optimization &&
- b &&
- b->hash_val == btree_ptr_hash_val(k)))
- goto lock_node;
retry:
b = btree_cache_find(bc, k);
if (unlikely(!b)) {
* else we could read in a btree node from disk that's been
* freed:
*/
- b = bch2_btree_node_fill(c, trans, path, k, path->btree_id,
+ b = bch2_btree_node_fill(trans, path, k, path->btree_id,
level, lock_type, true);
+ need_relock = true;
/* We raced and found the btree node in the cache */
if (!b)
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:
- *
- * The other thread might be holding an intent lock on the node
- * we want, and they want to update its parent node so they're
- * going to upgrade their intent lock on the parent node to a
- * write lock.
- *
- * But if we're holding a read lock on the parent, and we're
- * trying to get the intent lock they're holding, we deadlock.
- *
- * So to avoid this we drop the read locks on parent nodes when
- * we're starting to take intent locks - and handle the race.
- *
- * The race is that they might be about to free the node we
- * want, and dropping our read lock on the parent node lets them
- * update the parent marking the node we want as freed, and then
- * free it:
- *
- * To guard against this, btree nodes are evicted from the cache
- * 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
- * the parent was modified, when the pointer to the node we want
- * was removed - and we'll bail out:
- */
if (btree_node_read_locked(path, level + 1))
btree_node_unlock(trans, path, level + 1);
trace_and_count(c, trans_restart_btree_node_reused, trans, trace_ip, path);
return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_lock_node_reused));
}
+
+ /* avoid atomic set bit if it's not needed: */
+ if (!btree_node_accessed(b))
+ set_btree_node_accessed(b);
+ }
+
+ if (unlikely(btree_node_read_in_flight(b))) {
+ u32 seq = six_lock_seq(&b->c.lock);
+
+ six_unlock_type(&b->c.lock, lock_type);
+ bch2_trans_unlock(trans);
+ need_relock = true;
+
+ bch2_btree_node_wait_on_read(b);
+
+ /*
+ * should_be_locked is not set on this path yet, so we need to
+ * relock it specifically:
+ */
+ if (!six_relock_type(&b->c.lock, lock_type, seq))
+ goto retry;
+ }
+
+ if (unlikely(need_relock)) {
+ int ret = bch2_trans_relock(trans) ?:
+ bch2_btree_path_relock_intent(trans, path);
+ if (ret) {
+ six_unlock_type(&b->c.lock, lock_type);
+ return ERR_PTR(ret);
+ }
+ }
+
+ 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);
+ }
+
+ if (unlikely(btree_node_read_error(b))) {
+ six_unlock_type(&b->c.lock, lock_type);
+ return ERR_PTR(-EIO);
+ }
+
+ EBUG_ON(b->c.btree_id != path->btree_id);
+ EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
+ btree_check_header(c, b);
+
+ return b;
+}
+
+/**
+ * bch_btree_node_get - find a btree node in the cache and lock it, reading it
+ * in from disk if necessary.
+ *
+ * The btree node will have either a read or a write lock held, depending on
+ * the @write parameter.
+ */
+struct btree *bch2_btree_node_get(struct btree_trans *trans, struct btree_path *path,
+ const struct bkey_i *k, unsigned level,
+ enum six_lock_type lock_type,
+ unsigned long trace_ip)
+{
+ struct bch_fs *c = trans->c;
+ struct btree *b;
+ struct bset_tree *t;
+ int ret;
+
+ EBUG_ON(level >= BTREE_MAX_DEPTH);
+
+ b = btree_node_mem_ptr(k);
+
+ /*
+ * Check b->hash_val _before_ calling btree_node_lock() - this might not
+ * be the node we want anymore, and trying to lock the wrong node could
+ * cause an unneccessary transaction restart:
+ */
+ if (unlikely(!c->opts.btree_node_mem_ptr_optimization ||
+ !b ||
+ b->hash_val != btree_ptr_hash_val(k)))
+ return __bch2_btree_node_get(trans, path, k, level, lock_type, trace_ip);
+
+ if (btree_node_read_locked(path, level + 1))
+ btree_node_unlock(trans, path, level + 1);
+
+ ret = btree_node_lock(trans, path, &b->c, level, lock_type, trace_ip);
+ if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
+ return ERR_PTR(ret);
+
+ BUG_ON(ret);
+
+ if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
+ b->c.level != level ||
+ race_fault())) {
+ six_unlock_type(&b->c.lock, lock_type);
+ if (bch2_btree_node_relock(trans, path, level + 1))
+ return __bch2_btree_node_get(trans, path, k, level, lock_type, trace_ip);
+
+ trace_and_count(c, trans_restart_btree_node_reused, trans, trace_ip, path);
+ return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_lock_node_reused));
}
if (unlikely(btree_node_read_in_flight(b))) {
- u32 seq = b->c.lock.state.seq;
+ u32 seq = six_lock_seq(&b->c.lock);
six_unlock_type(&b->c.lock, lock_type);
bch2_trans_unlock(trans);
}
if (!six_relock_type(&b->c.lock, lock_type, seq))
- goto retry;
+ return __bch2_btree_node_get(trans, path, k, level, lock_type, trace_ip);
}
prefetch(b->aux_data);
if (nofill)
goto out;
- b = bch2_btree_node_fill(c, NULL, NULL, k, btree_id,
+ b = bch2_btree_node_fill(trans, NULL, k, btree_id,
level, SIX_LOCK_read, true);
/* We raced and found the btree node in the cache */
goto out;
} else {
lock_node:
- ret = btree_node_lock_nopath(trans, &b->c, SIX_LOCK_read);
+ ret = btree_node_lock_nopath(trans, &b->c, SIX_LOCK_read, _THIS_IP_);
if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
return ERR_PTR(ret);
return b;
}
-int bch2_btree_node_prefetch(struct bch_fs *c,
- struct btree_trans *trans,
+int bch2_btree_node_prefetch(struct btree_trans *trans,
struct btree_path *path,
const struct bkey_i *k,
enum btree_id btree_id, unsigned level)
{
+ struct bch_fs *c = trans->c;
struct btree_cache *bc = &c->btree_cache;
struct btree *b;
if (b)
return 0;
- b = bch2_btree_node_fill(c, trans, path, k, btree_id,
+ b = bch2_btree_node_fill(trans, path, k, btree_id,
level, SIX_LOCK_read, false);
return PTR_ERR_OR_ZERO(b);
}
btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write);
if (btree_node_dirty(b)) {
- __bch2_btree_node_write(c, b, 0);
+ __bch2_btree_node_write(c, b, BTREE_WRITE_cache_reclaim);
six_unlock_write(&b->c.lock);
six_unlock_intent(&b->c.lock);
goto wait_on_io;
}
void bch2_btree_node_to_text(struct printbuf *out, struct bch_fs *c,
- struct btree *b)
+ const struct btree *b)
{
const struct bkey_format *f = &b->format;
struct bset_stats stats;
stats.failed);
}
-void bch2_btree_cache_to_text(struct printbuf *out, struct btree_cache *bc)
+void bch2_btree_cache_to_text(struct printbuf *out, const struct btree_cache *bc)
{
prt_printf(out, "nr nodes:\t\t%u\n", bc->used);
prt_printf(out, "nr dirty:\t\t%u\n", atomic_read(&bc->dirty));