linux/fs/bcachefs/btree_locking.h

425 lines
12 KiB
C
Raw Normal View History

/* SPDX-License-Identifier: GPL-2.0 */
#ifndef _BCACHEFS_BTREE_LOCKING_H
#define _BCACHEFS_BTREE_LOCKING_H
/*
* Only for internal btree use:
*
* The btree iterator tracks what locks it wants to take, and what locks it
* currently has - here we have wrappers for locking/unlocking btree nodes and
* updating the iterator state
*/
#include "btree_iter.h"
#include "six.h"
void bch2_btree_lock_init(struct btree_bkey_cached_common *, enum six_lock_init_flags);
#ifdef CONFIG_LOCKDEP
void bch2_assert_btree_nodes_not_locked(void);
#else
static inline void bch2_assert_btree_nodes_not_locked(void) {}
#endif
void bch2_trans_unlock_noassert(struct btree_trans *);
static inline bool is_btree_node(struct btree_path *path, unsigned l)
{
return l < BTREE_MAX_DEPTH && !IS_ERR_OR_NULL(path->l[l].b);
}
static inline struct btree_transaction_stats *btree_trans_stats(struct btree_trans *trans)
{
return trans->fn_idx < ARRAY_SIZE(trans->c->btree_transaction_stats)
? &trans->c->btree_transaction_stats[trans->fn_idx]
: NULL;
}
/* matches six lock types */
enum btree_node_locked_type {
BTREE_NODE_UNLOCKED = -1,
BTREE_NODE_READ_LOCKED = SIX_LOCK_read,
BTREE_NODE_INTENT_LOCKED = SIX_LOCK_intent,
BTREE_NODE_WRITE_LOCKED = SIX_LOCK_write,
};
static inline int btree_node_locked_type(struct btree_path *path,
unsigned level)
{
return BTREE_NODE_UNLOCKED + ((path->nodes_locked >> (level << 1)) & 3);
}
static inline bool btree_node_write_locked(struct btree_path *path, unsigned l)
{
return btree_node_locked_type(path, l) == BTREE_NODE_WRITE_LOCKED;
}
static inline bool btree_node_intent_locked(struct btree_path *path, unsigned l)
{
return btree_node_locked_type(path, l) == BTREE_NODE_INTENT_LOCKED;
}
static inline bool btree_node_read_locked(struct btree_path *path, unsigned l)
{
return btree_node_locked_type(path, l) == BTREE_NODE_READ_LOCKED;
}
static inline bool btree_node_locked(struct btree_path *path, unsigned level)
{
return btree_node_locked_type(path, level) != BTREE_NODE_UNLOCKED;
}
static inline void mark_btree_node_locked_noreset(struct btree_path *path,
unsigned level,
enum btree_node_locked_type type)
{
/* relying on this to avoid a branch */
BUILD_BUG_ON(SIX_LOCK_read != 0);
BUILD_BUG_ON(SIX_LOCK_intent != 1);
path->nodes_locked &= ~(3U << (level << 1));
path->nodes_locked |= (type + 1) << (level << 1);
}
static inline void mark_btree_node_unlocked(struct btree_path *path,
unsigned level)
{
EBUG_ON(btree_node_write_locked(path, level));
mark_btree_node_locked_noreset(path, level, BTREE_NODE_UNLOCKED);
}
static inline void mark_btree_node_locked(struct btree_trans *trans,
struct btree_path *path,
unsigned level,
enum btree_node_locked_type type)
{
mark_btree_node_locked_noreset(path, level, (enum btree_node_locked_type) type);
#ifdef CONFIG_BCACHEFS_LOCK_TIME_STATS
path->l[level].lock_taken_time = local_clock();
#endif
}
static inline enum six_lock_type __btree_lock_want(struct btree_path *path, int level)
{
return level < path->locks_want
? SIX_LOCK_intent
: SIX_LOCK_read;
}
static inline enum btree_node_locked_type
btree_lock_want(struct btree_path *path, int level)
{
if (level < path->level)
return BTREE_NODE_UNLOCKED;
if (level < path->locks_want)
return BTREE_NODE_INTENT_LOCKED;
if (level == path->level)
return BTREE_NODE_READ_LOCKED;
return BTREE_NODE_UNLOCKED;
}
static void btree_trans_lock_hold_time_update(struct btree_trans *trans,
struct btree_path *path, unsigned level)
{
#ifdef CONFIG_BCACHEFS_LOCK_TIME_STATS
__bch2_time_stats_update(&btree_trans_stats(trans)->lock_hold_times,
path->l[level].lock_taken_time,
local_clock());
#endif
}
/* unlock: */
static inline void btree_node_unlock(struct btree_trans *trans,
struct btree_path *path, unsigned level)
{
int lock_type = btree_node_locked_type(path, level);
EBUG_ON(level >= BTREE_MAX_DEPTH);
if (lock_type != BTREE_NODE_UNLOCKED) {
six_unlock_type(&path->l[level].b->c.lock, lock_type);
btree_trans_lock_hold_time_update(trans, path, level);
}
mark_btree_node_unlocked(path, level);
}
static inline int btree_path_lowest_level_locked(struct btree_path *path)
{
return __ffs(path->nodes_locked) >> 1;
}
static inline int btree_path_highest_level_locked(struct btree_path *path)
{
return __fls(path->nodes_locked) >> 1;
}
static inline void __bch2_btree_path_unlock(struct btree_trans *trans,
struct btree_path *path)
{
btree_path_set_dirty(path, BTREE_ITER_NEED_RELOCK);
while (path->nodes_locked)
btree_node_unlock(trans, path, btree_path_lowest_level_locked(path));
}
/*
* Updates the saved lock sequence number, so that bch2_btree_node_relock() will
* succeed:
*/
static inline void
bch2_btree_node_unlock_write_inlined(struct btree_trans *trans, struct btree_path *path,
struct btree *b)
{
struct btree_path *linked;
unsigned i;
EBUG_ON(path->l[b->c.level].b != b);
EBUG_ON(path->l[b->c.level].lock_seq != six_lock_seq(&b->c.lock));
EBUG_ON(btree_node_locked_type(path, b->c.level) != SIX_LOCK_write);
mark_btree_node_locked_noreset(path, b->c.level, BTREE_NODE_INTENT_LOCKED);
trans_for_each_path_with_node(trans, b, linked, i)
linked->l[b->c.level].lock_seq++;
six_unlock_write(&b->c.lock);
}
void bch2_btree_node_unlock_write(struct btree_trans *,
struct btree_path *, struct btree *);
bcachefs: Deadlock cycle detector We've outgrown our own deadlock avoidance strategy. The btree iterator API provides an interface where the user doesn't need to concern themselves with lock ordering - different btree iterators can be traversed in any order. Without special care, this will lead to deadlocks. Our previous strategy was to define a lock ordering internally, and whenever we attempt to take a lock and trylock() fails, we'd check if the current btree transaction is holding any locks that cause a lock ordering violation. If so, we'd issue a transaction restart, and then bch2_trans_begin() would re-traverse all previously used iterators, but in the correct order. That approach had some issues, though. - Sometimes we'd issue transaction restarts unnecessarily, when no deadlock would have actually occured. Lock ordering restarts have become our primary cause of transaction restarts, on some workloads totally 20% of actual transaction commits. - To avoid deadlock or livelock, we'd often have to take intent locks when we only wanted a read lock: with the lock ordering approach, it is actually illegal to hold _any_ read lock while blocking on an intent lock, and this has been causing us unnecessary lock contention. - It was getting fragile - the various lock ordering rules are not trivial, and we'd been seeing occasional livelock issues related to this machinery. So, since bcachefs is already a relational database masquerading as a filesystem, we're stealing the next traditional database technique and switching to a cycle detector for avoiding deadlocks. When we block taking a btree lock, after adding ourself to the waitlist but before sleeping, we do a DFS of btree transactions waiting on other btree transactions, starting with the current transaction and walking our held locks, and transactions blocking on our held locks. If we find a cycle, we emit a transaction restart. Occasionally (e.g. the btree split path) we can not allow the lock() operation to fail, so if necessary we'll tell another transaction that it has to fail. Result: trans_restart_would_deadlock events are reduced by a factor of 10 to 100, and we'll be able to delete a whole bunch of grotty, fragile code. Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
2022-08-22 20:23:47 +03:00
int bch2_six_check_for_deadlock(struct six_lock *lock, void *p);
/* lock: */
bcachefs: Deadlock cycle detector We've outgrown our own deadlock avoidance strategy. The btree iterator API provides an interface where the user doesn't need to concern themselves with lock ordering - different btree iterators can be traversed in any order. Without special care, this will lead to deadlocks. Our previous strategy was to define a lock ordering internally, and whenever we attempt to take a lock and trylock() fails, we'd check if the current btree transaction is holding any locks that cause a lock ordering violation. If so, we'd issue a transaction restart, and then bch2_trans_begin() would re-traverse all previously used iterators, but in the correct order. That approach had some issues, though. - Sometimes we'd issue transaction restarts unnecessarily, when no deadlock would have actually occured. Lock ordering restarts have become our primary cause of transaction restarts, on some workloads totally 20% of actual transaction commits. - To avoid deadlock or livelock, we'd often have to take intent locks when we only wanted a read lock: with the lock ordering approach, it is actually illegal to hold _any_ read lock while blocking on an intent lock, and this has been causing us unnecessary lock contention. - It was getting fragile - the various lock ordering rules are not trivial, and we'd been seeing occasional livelock issues related to this machinery. So, since bcachefs is already a relational database masquerading as a filesystem, we're stealing the next traditional database technique and switching to a cycle detector for avoiding deadlocks. When we block taking a btree lock, after adding ourself to the waitlist but before sleeping, we do a DFS of btree transactions waiting on other btree transactions, starting with the current transaction and walking our held locks, and transactions blocking on our held locks. If we find a cycle, we emit a transaction restart. Occasionally (e.g. the btree split path) we can not allow the lock() operation to fail, so if necessary we'll tell another transaction that it has to fail. Result: trans_restart_would_deadlock events are reduced by a factor of 10 to 100, and we'll be able to delete a whole bunch of grotty, fragile code. Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
2022-08-22 20:23:47 +03:00
static inline int __btree_node_lock_nopath(struct btree_trans *trans,
struct btree_bkey_cached_common *b,
enum six_lock_type type,
bool lock_may_not_fail,
unsigned long ip)
bcachefs: Deadlock cycle detector We've outgrown our own deadlock avoidance strategy. The btree iterator API provides an interface where the user doesn't need to concern themselves with lock ordering - different btree iterators can be traversed in any order. Without special care, this will lead to deadlocks. Our previous strategy was to define a lock ordering internally, and whenever we attempt to take a lock and trylock() fails, we'd check if the current btree transaction is holding any locks that cause a lock ordering violation. If so, we'd issue a transaction restart, and then bch2_trans_begin() would re-traverse all previously used iterators, but in the correct order. That approach had some issues, though. - Sometimes we'd issue transaction restarts unnecessarily, when no deadlock would have actually occured. Lock ordering restarts have become our primary cause of transaction restarts, on some workloads totally 20% of actual transaction commits. - To avoid deadlock or livelock, we'd often have to take intent locks when we only wanted a read lock: with the lock ordering approach, it is actually illegal to hold _any_ read lock while blocking on an intent lock, and this has been causing us unnecessary lock contention. - It was getting fragile - the various lock ordering rules are not trivial, and we'd been seeing occasional livelock issues related to this machinery. So, since bcachefs is already a relational database masquerading as a filesystem, we're stealing the next traditional database technique and switching to a cycle detector for avoiding deadlocks. When we block taking a btree lock, after adding ourself to the waitlist but before sleeping, we do a DFS of btree transactions waiting on other btree transactions, starting with the current transaction and walking our held locks, and transactions blocking on our held locks. If we find a cycle, we emit a transaction restart. Occasionally (e.g. the btree split path) we can not allow the lock() operation to fail, so if necessary we'll tell another transaction that it has to fail. Result: trans_restart_would_deadlock events are reduced by a factor of 10 to 100, and we'll be able to delete a whole bunch of grotty, fragile code. Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
2022-08-22 20:23:47 +03:00
{
int ret;
trans->lock_may_not_fail = lock_may_not_fail;
trans->lock_must_abort = false;
trans->locking = b;
bcachefs: Deadlock cycle detector We've outgrown our own deadlock avoidance strategy. The btree iterator API provides an interface where the user doesn't need to concern themselves with lock ordering - different btree iterators can be traversed in any order. Without special care, this will lead to deadlocks. Our previous strategy was to define a lock ordering internally, and whenever we attempt to take a lock and trylock() fails, we'd check if the current btree transaction is holding any locks that cause a lock ordering violation. If so, we'd issue a transaction restart, and then bch2_trans_begin() would re-traverse all previously used iterators, but in the correct order. That approach had some issues, though. - Sometimes we'd issue transaction restarts unnecessarily, when no deadlock would have actually occured. Lock ordering restarts have become our primary cause of transaction restarts, on some workloads totally 20% of actual transaction commits. - To avoid deadlock or livelock, we'd often have to take intent locks when we only wanted a read lock: with the lock ordering approach, it is actually illegal to hold _any_ read lock while blocking on an intent lock, and this has been causing us unnecessary lock contention. - It was getting fragile - the various lock ordering rules are not trivial, and we'd been seeing occasional livelock issues related to this machinery. So, since bcachefs is already a relational database masquerading as a filesystem, we're stealing the next traditional database technique and switching to a cycle detector for avoiding deadlocks. When we block taking a btree lock, after adding ourself to the waitlist but before sleeping, we do a DFS of btree transactions waiting on other btree transactions, starting with the current transaction and walking our held locks, and transactions blocking on our held locks. If we find a cycle, we emit a transaction restart. Occasionally (e.g. the btree split path) we can not allow the lock() operation to fail, so if necessary we'll tell another transaction that it has to fail. Result: trans_restart_would_deadlock events are reduced by a factor of 10 to 100, and we'll be able to delete a whole bunch of grotty, fragile code. Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
2022-08-22 20:23:47 +03:00
ret = six_lock_ip_waiter(&b->lock, type, &trans->locking_wait,
bch2_six_check_for_deadlock, trans, ip);
bcachefs: Deadlock cycle detector We've outgrown our own deadlock avoidance strategy. The btree iterator API provides an interface where the user doesn't need to concern themselves with lock ordering - different btree iterators can be traversed in any order. Without special care, this will lead to deadlocks. Our previous strategy was to define a lock ordering internally, and whenever we attempt to take a lock and trylock() fails, we'd check if the current btree transaction is holding any locks that cause a lock ordering violation. If so, we'd issue a transaction restart, and then bch2_trans_begin() would re-traverse all previously used iterators, but in the correct order. That approach had some issues, though. - Sometimes we'd issue transaction restarts unnecessarily, when no deadlock would have actually occured. Lock ordering restarts have become our primary cause of transaction restarts, on some workloads totally 20% of actual transaction commits. - To avoid deadlock or livelock, we'd often have to take intent locks when we only wanted a read lock: with the lock ordering approach, it is actually illegal to hold _any_ read lock while blocking on an intent lock, and this has been causing us unnecessary lock contention. - It was getting fragile - the various lock ordering rules are not trivial, and we'd been seeing occasional livelock issues related to this machinery. So, since bcachefs is already a relational database masquerading as a filesystem, we're stealing the next traditional database technique and switching to a cycle detector for avoiding deadlocks. When we block taking a btree lock, after adding ourself to the waitlist but before sleeping, we do a DFS of btree transactions waiting on other btree transactions, starting with the current transaction and walking our held locks, and transactions blocking on our held locks. If we find a cycle, we emit a transaction restart. Occasionally (e.g. the btree split path) we can not allow the lock() operation to fail, so if necessary we'll tell another transaction that it has to fail. Result: trans_restart_would_deadlock events are reduced by a factor of 10 to 100, and we'll be able to delete a whole bunch of grotty, fragile code. Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
2022-08-22 20:23:47 +03:00
WRITE_ONCE(trans->locking, NULL);
WRITE_ONCE(trans->locking_wait.start_time, 0);
return ret;
}
static inline int __must_check
btree_node_lock_nopath(struct btree_trans *trans,
struct btree_bkey_cached_common *b,
enum six_lock_type type,
unsigned long ip)
{
return __btree_node_lock_nopath(trans, b, type, false, ip);
}
static inline void btree_node_lock_nopath_nofail(struct btree_trans *trans,
struct btree_bkey_cached_common *b,
enum six_lock_type type)
{
int ret = __btree_node_lock_nopath(trans, b, type, true, _THIS_IP_);
BUG_ON(ret);
}
/*
* Lock a btree node if we already have it locked on one of our linked
* iterators:
*/
static inline bool btree_node_lock_increment(struct btree_trans *trans,
struct btree_bkey_cached_common *b,
unsigned level,
enum btree_node_locked_type want)
{
struct btree_path *path;
unsigned i;
trans_for_each_path(trans, path, i)
if (&path->l[level].b->c == b &&
btree_node_locked_type(path, level) >= want) {
six_lock_increment(&b->lock, (enum six_lock_type) want);
return true;
}
return false;
}
static inline int btree_node_lock(struct btree_trans *trans,
struct btree_path *path,
struct btree_bkey_cached_common *b,
unsigned level,
enum six_lock_type type,
unsigned long ip)
{
int ret = 0;
EBUG_ON(level >= BTREE_MAX_DEPTH);
if (likely(six_trylock_type(&b->lock, type)) ||
btree_node_lock_increment(trans, b, level, (enum btree_node_locked_type) type) ||
!(ret = btree_node_lock_nopath(trans, b, type, btree_path_ip_allocated(path)))) {
#ifdef CONFIG_BCACHEFS_LOCK_TIME_STATS
path->l[b->level].lock_taken_time = local_clock();
#endif
}
return ret;
}
int __bch2_btree_node_lock_write(struct btree_trans *, struct btree_path *,
struct btree_bkey_cached_common *b, bool);
bcachefs: Deadlock cycle detector We've outgrown our own deadlock avoidance strategy. The btree iterator API provides an interface where the user doesn't need to concern themselves with lock ordering - different btree iterators can be traversed in any order. Without special care, this will lead to deadlocks. Our previous strategy was to define a lock ordering internally, and whenever we attempt to take a lock and trylock() fails, we'd check if the current btree transaction is holding any locks that cause a lock ordering violation. If so, we'd issue a transaction restart, and then bch2_trans_begin() would re-traverse all previously used iterators, but in the correct order. That approach had some issues, though. - Sometimes we'd issue transaction restarts unnecessarily, when no deadlock would have actually occured. Lock ordering restarts have become our primary cause of transaction restarts, on some workloads totally 20% of actual transaction commits. - To avoid deadlock or livelock, we'd often have to take intent locks when we only wanted a read lock: with the lock ordering approach, it is actually illegal to hold _any_ read lock while blocking on an intent lock, and this has been causing us unnecessary lock contention. - It was getting fragile - the various lock ordering rules are not trivial, and we'd been seeing occasional livelock issues related to this machinery. So, since bcachefs is already a relational database masquerading as a filesystem, we're stealing the next traditional database technique and switching to a cycle detector for avoiding deadlocks. When we block taking a btree lock, after adding ourself to the waitlist but before sleeping, we do a DFS of btree transactions waiting on other btree transactions, starting with the current transaction and walking our held locks, and transactions blocking on our held locks. If we find a cycle, we emit a transaction restart. Occasionally (e.g. the btree split path) we can not allow the lock() operation to fail, so if necessary we'll tell another transaction that it has to fail. Result: trans_restart_would_deadlock events are reduced by a factor of 10 to 100, and we'll be able to delete a whole bunch of grotty, fragile code. Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
2022-08-22 20:23:47 +03:00
static inline int __btree_node_lock_write(struct btree_trans *trans,
struct btree_path *path,
struct btree_bkey_cached_common *b,
bool lock_may_not_fail)
{
EBUG_ON(&path->l[b->level].b->c != b);
EBUG_ON(path->l[b->level].lock_seq != six_lock_seq(&b->lock));
EBUG_ON(!btree_node_intent_locked(path, b->level));
/*
* six locks are unfair, and read locks block while a thread wants a
* write lock: thus, we need to tell the cycle detector we have a write
* lock _before_ taking the lock:
*/
mark_btree_node_locked_noreset(path, b->level, BTREE_NODE_WRITE_LOCKED);
return likely(six_trylock_write(&b->lock))
bcachefs: Deadlock cycle detector We've outgrown our own deadlock avoidance strategy. The btree iterator API provides an interface where the user doesn't need to concern themselves with lock ordering - different btree iterators can be traversed in any order. Without special care, this will lead to deadlocks. Our previous strategy was to define a lock ordering internally, and whenever we attempt to take a lock and trylock() fails, we'd check if the current btree transaction is holding any locks that cause a lock ordering violation. If so, we'd issue a transaction restart, and then bch2_trans_begin() would re-traverse all previously used iterators, but in the correct order. That approach had some issues, though. - Sometimes we'd issue transaction restarts unnecessarily, when no deadlock would have actually occured. Lock ordering restarts have become our primary cause of transaction restarts, on some workloads totally 20% of actual transaction commits. - To avoid deadlock or livelock, we'd often have to take intent locks when we only wanted a read lock: with the lock ordering approach, it is actually illegal to hold _any_ read lock while blocking on an intent lock, and this has been causing us unnecessary lock contention. - It was getting fragile - the various lock ordering rules are not trivial, and we'd been seeing occasional livelock issues related to this machinery. So, since bcachefs is already a relational database masquerading as a filesystem, we're stealing the next traditional database technique and switching to a cycle detector for avoiding deadlocks. When we block taking a btree lock, after adding ourself to the waitlist but before sleeping, we do a DFS of btree transactions waiting on other btree transactions, starting with the current transaction and walking our held locks, and transactions blocking on our held locks. If we find a cycle, we emit a transaction restart. Occasionally (e.g. the btree split path) we can not allow the lock() operation to fail, so if necessary we'll tell another transaction that it has to fail. Result: trans_restart_would_deadlock events are reduced by a factor of 10 to 100, and we'll be able to delete a whole bunch of grotty, fragile code. Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
2022-08-22 20:23:47 +03:00
? 0
: __bch2_btree_node_lock_write(trans, path, b, lock_may_not_fail);
bcachefs: Deadlock cycle detector We've outgrown our own deadlock avoidance strategy. The btree iterator API provides an interface where the user doesn't need to concern themselves with lock ordering - different btree iterators can be traversed in any order. Without special care, this will lead to deadlocks. Our previous strategy was to define a lock ordering internally, and whenever we attempt to take a lock and trylock() fails, we'd check if the current btree transaction is holding any locks that cause a lock ordering violation. If so, we'd issue a transaction restart, and then bch2_trans_begin() would re-traverse all previously used iterators, but in the correct order. That approach had some issues, though. - Sometimes we'd issue transaction restarts unnecessarily, when no deadlock would have actually occured. Lock ordering restarts have become our primary cause of transaction restarts, on some workloads totally 20% of actual transaction commits. - To avoid deadlock or livelock, we'd often have to take intent locks when we only wanted a read lock: with the lock ordering approach, it is actually illegal to hold _any_ read lock while blocking on an intent lock, and this has been causing us unnecessary lock contention. - It was getting fragile - the various lock ordering rules are not trivial, and we'd been seeing occasional livelock issues related to this machinery. So, since bcachefs is already a relational database masquerading as a filesystem, we're stealing the next traditional database technique and switching to a cycle detector for avoiding deadlocks. When we block taking a btree lock, after adding ourself to the waitlist but before sleeping, we do a DFS of btree transactions waiting on other btree transactions, starting with the current transaction and walking our held locks, and transactions blocking on our held locks. If we find a cycle, we emit a transaction restart. Occasionally (e.g. the btree split path) we can not allow the lock() operation to fail, so if necessary we'll tell another transaction that it has to fail. Result: trans_restart_would_deadlock events are reduced by a factor of 10 to 100, and we'll be able to delete a whole bunch of grotty, fragile code. Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
2022-08-22 20:23:47 +03:00
}
static inline int __must_check
bch2_btree_node_lock_write(struct btree_trans *trans,
struct btree_path *path,
struct btree_bkey_cached_common *b)
{
bcachefs: Deadlock cycle detector We've outgrown our own deadlock avoidance strategy. The btree iterator API provides an interface where the user doesn't need to concern themselves with lock ordering - different btree iterators can be traversed in any order. Without special care, this will lead to deadlocks. Our previous strategy was to define a lock ordering internally, and whenever we attempt to take a lock and trylock() fails, we'd check if the current btree transaction is holding any locks that cause a lock ordering violation. If so, we'd issue a transaction restart, and then bch2_trans_begin() would re-traverse all previously used iterators, but in the correct order. That approach had some issues, though. - Sometimes we'd issue transaction restarts unnecessarily, when no deadlock would have actually occured. Lock ordering restarts have become our primary cause of transaction restarts, on some workloads totally 20% of actual transaction commits. - To avoid deadlock or livelock, we'd often have to take intent locks when we only wanted a read lock: with the lock ordering approach, it is actually illegal to hold _any_ read lock while blocking on an intent lock, and this has been causing us unnecessary lock contention. - It was getting fragile - the various lock ordering rules are not trivial, and we'd been seeing occasional livelock issues related to this machinery. So, since bcachefs is already a relational database masquerading as a filesystem, we're stealing the next traditional database technique and switching to a cycle detector for avoiding deadlocks. When we block taking a btree lock, after adding ourself to the waitlist but before sleeping, we do a DFS of btree transactions waiting on other btree transactions, starting with the current transaction and walking our held locks, and transactions blocking on our held locks. If we find a cycle, we emit a transaction restart. Occasionally (e.g. the btree split path) we can not allow the lock() operation to fail, so if necessary we'll tell another transaction that it has to fail. Result: trans_restart_would_deadlock events are reduced by a factor of 10 to 100, and we'll be able to delete a whole bunch of grotty, fragile code. Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
2022-08-22 20:23:47 +03:00
return __btree_node_lock_write(trans, path, b, false);
}
void bch2_btree_node_lock_write_nofail(struct btree_trans *,
struct btree_path *,
struct btree_bkey_cached_common *);
/* relock: */
bool bch2_btree_path_relock_norestart(struct btree_trans *, struct btree_path *);
int __bch2_btree_path_relock(struct btree_trans *,
struct btree_path *, unsigned long);
static inline int bch2_btree_path_relock(struct btree_trans *trans,
struct btree_path *path, unsigned long trace_ip)
{
return btree_node_locked(path, path->level)
? 0
: __bch2_btree_path_relock(trans, path, trace_ip);
}
bool __bch2_btree_node_relock(struct btree_trans *, struct btree_path *, unsigned, bool trace);
static inline bool bch2_btree_node_relock(struct btree_trans *trans,
struct btree_path *path, unsigned level)
{
EBUG_ON(btree_node_locked(path, level) &&
!btree_node_write_locked(path, level) &&
btree_node_locked_type(path, level) != __btree_lock_want(path, level));
return likely(btree_node_locked(path, level)) ||
(!IS_ERR_OR_NULL(path->l[level].b) &&
__bch2_btree_node_relock(trans, path, level, true));
}
static inline bool bch2_btree_node_relock_notrace(struct btree_trans *trans,
struct btree_path *path, unsigned level)
{
EBUG_ON(btree_node_locked(path, level) &&
!btree_node_write_locked(path, level) &&
btree_node_locked_type(path, level) != __btree_lock_want(path, level));
return likely(btree_node_locked(path, level)) ||
(!IS_ERR_OR_NULL(path->l[level].b) &&
__bch2_btree_node_relock(trans, path, level, false));
}
/* upgrade */
bool bch2_btree_path_upgrade_noupgrade_sibs(struct btree_trans *,
struct btree_path *, unsigned,
struct get_locks_fail *);
bool __bch2_btree_path_upgrade(struct btree_trans *,
struct btree_path *, unsigned,
struct get_locks_fail *);
static inline int bch2_btree_path_upgrade(struct btree_trans *trans,
struct btree_path *path,
unsigned new_locks_want)
{
struct get_locks_fail f;
unsigned old_locks_want = path->locks_want;
new_locks_want = min(new_locks_want, BTREE_MAX_DEPTH);
if (path->locks_want < new_locks_want
? __bch2_btree_path_upgrade(trans, path, new_locks_want, &f)
: path->uptodate == BTREE_ITER_UPTODATE)
return 0;
trace_and_count(trans->c, trans_restart_upgrade, trans, _THIS_IP_, path,
old_locks_want, new_locks_want, &f);
return btree_trans_restart(trans, BCH_ERR_transaction_restart_upgrade);
}
/* misc: */
static inline void btree_path_set_should_be_locked(struct btree_path *path)
{
EBUG_ON(!btree_node_locked(path, path->level));
EBUG_ON(path->uptodate);
path->should_be_locked = true;
}
static inline void __btree_path_set_level_up(struct btree_trans *trans,
struct btree_path *path,
unsigned l)
{
btree_node_unlock(trans, path, l);
path->l[l].b = ERR_PTR(-BCH_ERR_no_btree_node_up);
}
static inline void btree_path_set_level_up(struct btree_trans *trans,
struct btree_path *path)
{
__btree_path_set_level_up(trans, path, path->level++);
btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
}
/* debug */
struct six_lock_count bch2_btree_node_lock_counts(struct btree_trans *,
struct btree_path *,
struct btree_bkey_cached_common *b,
unsigned);
int bch2_check_for_deadlock(struct btree_trans *, struct printbuf *);
#ifdef CONFIG_BCACHEFS_DEBUG
void bch2_btree_path_verify_locks(struct btree_path *);
void bch2_trans_verify_locks(struct btree_trans *);
#else
static inline void bch2_btree_path_verify_locks(struct btree_path *path) {}
static inline void bch2_trans_verify_locks(struct btree_trans *trans) {}
#endif
#endif /* _BCACHEFS_BTREE_LOCKING_H */