]> git.sesse.net Git - bcachefs-tools-debian/blobdiff - libbcachefs/btree_locking.c
Update bcachefs sources to 8dbfede1d9 fixup! bcachefs: More info on check_bucket_ref...
[bcachefs-tools-debian] / libbcachefs / btree_locking.c
index 9a525d34c7f46a28a8f59291a70435437fe6a50b..80398e497bd53a3a546a77c45bafd522c065ad4d 100644 (file)
@@ -94,6 +94,43 @@ static noinline void print_chain(struct printbuf *out, struct lock_graph *g)
        prt_newline(out);
 }
 
+static void lock_graph_up(struct lock_graph *g)
+{
+       closure_put(&g->g[--g->nr].trans->ref);
+}
+
+static noinline void lock_graph_pop_all(struct lock_graph *g)
+{
+       while (g->nr)
+               lock_graph_up(g);
+}
+
+static void lock_graph_down(struct lock_graph *g, struct btree_trans *trans)
+{
+       closure_get(&trans->ref);
+
+       g->g[g->nr++] = (struct trans_waiting_for_lock) {
+               .trans          = trans,
+               .node_want      = trans->locking,
+               .lock_want      = trans->locking_wait.lock_want,
+       };
+}
+
+static bool lock_graph_remove_non_waiters(struct lock_graph *g)
+{
+       struct trans_waiting_for_lock *i;
+
+       for (i = g->g + 1; i < g->g + g->nr; i++)
+               if (i->trans->locking != i->node_want ||
+                   i->trans->locking_wait.start_time != i[-1].lock_start_time) {
+                       while (g->g + g->nr > i)
+                               lock_graph_up(g);
+                       return true;
+               }
+
+       return false;
+}
+
 static int abort_lock(struct lock_graph *g, struct trans_waiting_for_lock *i)
 {
        if (i == g->g) {
@@ -106,44 +143,45 @@ static int abort_lock(struct lock_graph *g, struct trans_waiting_for_lock *i)
        }
 }
 
-static noinline int break_cycle(struct lock_graph *g)
+static int btree_trans_abort_preference(struct btree_trans *trans)
 {
-       struct trans_waiting_for_lock *i;
-
-       /*
-        * We'd like to prioritize aborting transactions that have done less
-        * work - but it appears breaking cycles by telling other transactions
-        * to abort may still be buggy:
-        */
-#if 0
-       for (i = g->g; i < g->g + g->nr; i++) {
-               if (i->trans->lock_may_not_fail ||
-                   i->trans->locking_wait.lock_want == SIX_LOCK_write)
-                       continue;
+       if (trans->lock_may_not_fail)
+               return 0;
+       if (trans->locking_wait.lock_want == SIX_LOCK_write)
+               return 1;
+       if (!trans->in_traverse_all)
+               return 2;
+       return 3;
+}
 
-               return abort_lock(g, i);
-       }
+static noinline int break_cycle(struct lock_graph *g, struct printbuf *cycle)
+{
+       struct trans_waiting_for_lock *i, *abort = NULL;
+       unsigned best = 0, pref;
+       int ret;
 
-       for (i = g->g; i < g->g + g->nr; i++) {
-               if (i->trans->lock_may_not_fail ||
-                   !i->trans->in_traverse_all)
-                       continue;
+       if (lock_graph_remove_non_waiters(g))
+               return 0;
 
-               return abort_lock(g, i);
+       /* Only checking, for debugfs: */
+       if (cycle) {
+               print_cycle(cycle, g);
+               ret = -1;
+               goto out;
        }
-#endif
-       for (i = g->g; i < g->g + g->nr; i++) {
-               if (i->trans->lock_may_not_fail)
-                       continue;
 
-               return abort_lock(g, i);
+       for (i = g->g; i < g->g + g->nr; i++) {
+               pref = btree_trans_abort_preference(i->trans);
+               if (pref > best) {
+                       abort = i;
+                       best = pref;
+               }
        }
 
-       {
-               struct bch_fs *c = g->g->trans->c;
+       if (unlikely(!best)) {
                struct printbuf buf = PRINTBUF;
 
-               bch_err(c, "cycle of nofail locks");
+               prt_printf(&buf, bch2_fmt(g->g->trans->c, "cycle of nofail locks"));
 
                for (i = g->g; i < g->g + g->nr; i++) {
                        struct btree_trans *trans = i->trans;
@@ -153,7 +191,7 @@ static noinline int break_cycle(struct lock_graph *g)
                        prt_printf(&buf, "backtrace:");
                        prt_newline(&buf);
                        printbuf_indent_add(&buf, 2);
-                       bch2_prt_backtrace(&buf, trans->locking_wait.task);
+                       bch2_prt_task_backtrace(&buf, trans->locking_wait.task);
                        printbuf_indent_sub(&buf, 2);
                        prt_newline(&buf);
                }
@@ -162,21 +200,13 @@ static noinline int break_cycle(struct lock_graph *g)
                printbuf_exit(&buf);
                BUG();
        }
-}
-
-static void lock_graph_pop(struct lock_graph *g)
-{
-       closure_put(&g->g[--g->nr].trans->ref);
-}
-
-static void lock_graph_pop_above(struct lock_graph *g, struct trans_waiting_for_lock *above,
-                                struct printbuf *cycle)
-{
-       if (g->nr > 1 && cycle)
-               print_chain(cycle, g);
 
-       while (g->g + g->nr > above)
-               lock_graph_pop(g);
+       ret = abort_lock(g, abort);
+out:
+       if (ret)
+               while (g->nr)
+                       lock_graph_up(g);
+       return ret;
 }
 
 static int lock_graph_descend(struct lock_graph *g, struct btree_trans *trans,
@@ -184,67 +214,27 @@ static int lock_graph_descend(struct lock_graph *g, struct btree_trans *trans,
 {
        struct btree_trans *orig_trans = g->g->trans;
        struct trans_waiting_for_lock *i;
-       int ret = 0;
-
-       for (i = g->g; i < g->g + g->nr; i++) {
-               if (i->trans->locking != i->node_want) {
-                       lock_graph_pop_above(g, i - 1, cycle);
-                       return 0;
-               }
-
-               if (i->trans == trans) {
-                       if (cycle) {
-                               /* Only checking: */
-                               print_cycle(cycle, g);
-                               ret = -1;
-                       } else {
-                               ret = break_cycle(g);
-                       }
 
-                       if (ret)
-                               goto deadlock;
-                       /*
-                        * If we didn't abort (instead telling another
-                        * transaction to abort), keep checking:
-                        */
-               }
-       }
+       for (i = g->g; i < g->g + g->nr; i++)
+               if (i->trans == trans)
+                       return break_cycle(g, cycle);
 
        if (g->nr == ARRAY_SIZE(g->g)) {
                if (orig_trans->lock_may_not_fail)
                        return 0;
 
-               trace_and_count(trans->c, trans_restart_would_deadlock_recursion_limit, trans, _RET_IP_);
-               ret = btree_trans_restart(orig_trans, BCH_ERR_transaction_restart_deadlock_recursion_limit);
-               goto deadlock;
-       }
+               while (g->nr)
+                       lock_graph_up(g);
 
-       closure_get(&trans->ref);
+               if (cycle)
+                       return 0;
 
-       g->g[g->nr++] = (struct trans_waiting_for_lock) {
-               .trans          = trans,
-               .node_want      = trans->locking,
-               .lock_want      = trans->locking_wait.lock_want,
-       };
+               trace_and_count(trans->c, trans_restart_would_deadlock_recursion_limit, trans, _RET_IP_);
+               return btree_trans_restart(orig_trans, BCH_ERR_transaction_restart_deadlock_recursion_limit);
+       }
 
+       lock_graph_down(g, trans);
        return 0;
-deadlock:
-       lock_graph_pop_above(g, g->g, cycle);
-       return ret;
-}
-
-static noinline void lock_graph_remove_non_waiters(struct lock_graph *g,
-                                                  struct printbuf *cycle)
-{
-       struct trans_waiting_for_lock *i;
-
-       for (i = g->g + 1; i < g->g + g->nr; i++)
-               if (i->trans->locking != i->node_want ||
-                   i->trans->locking_wait.start_time != i[-1].lock_start_time) {
-                       lock_graph_pop_above(g, i - 1, cycle);
-                       return;
-               }
-       BUG();
 }
 
 static bool lock_type_conflicts(enum six_lock_type t1, enum six_lock_type t2)
@@ -261,13 +251,15 @@ int bch2_check_for_deadlock(struct btree_trans *trans, struct printbuf *cycle)
        int ret;
 
        if (trans->lock_must_abort) {
+               if (cycle)
+                       return -1;
+
                trace_and_count(trans->c, trans_restart_would_deadlock, trans, _RET_IP_);
                return btree_trans_restart(trans, BCH_ERR_transaction_restart_would_deadlock);
        }
 
        g.nr = 0;
-       ret = lock_graph_descend(&g, trans, cycle);
-       BUG_ON(ret);
+       lock_graph_down(&g, trans);
 next:
        if (!g.nr)
                return 0;
@@ -294,8 +286,26 @@ next:
 
                        b = &READ_ONCE(path->l[top->level].b)->c;
 
-                       if (unlikely(IS_ERR_OR_NULL(b))) {
-                               lock_graph_remove_non_waiters(&g, cycle);
+                       if (IS_ERR_OR_NULL(b)) {
+                               /*
+                                * If we get here, it means we raced with the
+                                * other thread updating its btree_path
+                                * structures - which means it can't be blocked
+                                * waiting on a lock:
+                                */
+                               if (!lock_graph_remove_non_waiters(&g)) {
+                                       /*
+                                        * If lock_graph_remove_non_waiters()
+                                        * didn't do anything, it must be
+                                        * because we're being called by debugfs
+                                        * checking for lock cycles, which
+                                        * invokes us on btree_transactions that
+                                        * aren't actually waiting on anything.
+                                        * Just bail out:
+                                        */
+                                       lock_graph_pop_all(&g);
+                               }
+
                                goto next;
                        }
 
@@ -321,7 +331,7 @@ next:
                                raw_spin_unlock(&b->lock.wait_lock);
 
                                if (ret)
-                                       return ret < 0 ? ret : 0;
+                                       return ret;
                                goto next;
 
                        }
@@ -331,7 +341,7 @@ next:
 
        if (g.nr > 1 && cycle)
                print_chain(cycle, &g);
-       lock_graph_pop(&g);
+       lock_graph_up(&g);
        goto next;
 }
 
@@ -356,7 +366,8 @@ int __bch2_btree_node_lock_write(struct btree_trans *trans, struct btree_path *p
         * locked:
         */
        six_lock_readers_add(&b->lock, -readers);
-       ret = __btree_node_lock_nopath(trans, b, SIX_LOCK_write, lock_may_not_fail);
+       ret = __btree_node_lock_nopath(trans, b, SIX_LOCK_write,
+                                      lock_may_not_fail, _RET_IP_);
        six_lock_readers_add(&b->lock, readers);
 
        if (ret)
@@ -428,7 +439,7 @@ bool __bch2_btree_node_relock(struct btree_trans *trans,
                return true;
        }
 fail:
-       if (trace)
+       if (trace && !trans->notrace_relock_fail)
                trace_and_count(trans->c, btree_path_relock_fail, trans, _RET_IP_, path, level);
        return false;
 }
@@ -525,6 +536,17 @@ bool bch2_btree_path_relock_norestart(struct btree_trans *trans,
        return btree_path_get_locks(trans, path, false);
 }
 
+int __bch2_btree_path_relock(struct btree_trans *trans,
+                       struct btree_path *path, unsigned long trace_ip)
+{
+       if (!bch2_btree_path_relock_norestart(trans, path, trace_ip)) {
+               trace_and_count(trans->c, trans_restart_relock_path, trans, trace_ip, path);
+               return btree_trans_restart(trans, BCH_ERR_transaction_restart_relock_path);
+       }
+
+       return 0;
+}
+
 __flatten
 bool bch2_btree_path_upgrade_norestart(struct btree_trans *trans,
                        struct btree_path *path, unsigned long trace_ip)
@@ -625,7 +647,7 @@ int bch2_trans_relock(struct btree_trans *trans)
        struct btree_path *path;
 
        if (unlikely(trans->restarted))
-               return - ((int) trans->restarted);
+               return -((int) trans->restarted);
 
        trans_for_each_path(trans, path)
                if (path->should_be_locked &&
@@ -636,6 +658,21 @@ int bch2_trans_relock(struct btree_trans *trans)
        return 0;
 }
 
+int bch2_trans_relock_notrace(struct btree_trans *trans)
+{
+       struct btree_path *path;
+
+       if (unlikely(trans->restarted))
+               return -((int) trans->restarted);
+
+       trans_for_each_path(trans, path)
+               if (path->should_be_locked &&
+                   !bch2_btree_path_relock_norestart(trans, path, _RET_IP_)) {
+                       return btree_trans_restart(trans, BCH_ERR_transaction_restart_relock);
+               }
+       return 0;
+}
+
 void bch2_trans_unlock(struct btree_trans *trans)
 {
        struct btree_path *path;