bcachefs: Fix lock_graph_remove_non_waiters()
authorKent Overstreet <kent.overstreet@linux.dev>
Wed, 12 Oct 2022 22:17:49 +0000 (18:17 -0400)
committerKent Overstreet <kent.overstreet@linux.dev>
Sun, 22 Oct 2023 21:09:43 +0000 (17:09 -0400)
We were removing 1 more entry than we were supposed to - oops.

Also some other simplifications and cleanups, and bring back the abort
preference code in a better fashion.

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
fs/bcachefs/btree_locking.c

index 4940b3069a76b5b122182d0eb57a0c6b49b30530..922cfc7f545095f164fdacbdc0751c7f5e9885af 100644 (file)
@@ -94,6 +94,37 @@ 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 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) {
@@ -106,40 +137,42 @@ 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;
+               }
        }
 
-       {
+       if (unlikely(!best)) {
                struct bch_fs *c = g->g->trans->c;
                struct printbuf buf = PRINTBUF;
 
@@ -162,21 +195,13 @@ 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 +209,23 @@ 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;
 
+               while (g->nr)
+                       lock_graph_up(g);
                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;
+               return btree_trans_restart(orig_trans, BCH_ERR_transaction_restart_deadlock_recursion_limit);
        }
 
-       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,
-       };
-
+       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)
@@ -266,8 +247,7 @@ int bch2_check_for_deadlock(struct btree_trans *trans, struct printbuf *cycle)
        }
 
        g.nr = 0;
-       ret = lock_graph_descend(&g, trans, cycle);
-       BUG_ON(ret);
+       lock_graph_down(&g, trans);
 next:
        if (!g.nr)
                return 0;
@@ -295,7 +275,7 @@ next:
                        b = &READ_ONCE(path->l[top->level].b)->c;
 
                        if (unlikely(IS_ERR_OR_NULL(b))) {
-                               lock_graph_remove_non_waiters(&g, cycle);
+                               BUG_ON(!lock_graph_remove_non_waiters(&g));
                                goto next;
                        }
 
@@ -321,7 +301,7 @@ next:
                                raw_spin_unlock(&b->lock.wait_lock);
 
                                if (ret)
-                                       return ret < 0 ? ret : 0;
+                                       return ret;
                                goto next;
 
                        }
@@ -331,7 +311,7 @@ next:
 
        if (g.nr > 1 && cycle)
                print_chain(cycle, &g);
-       lock_graph_pop(&g);
+       lock_graph_up(&g);
        goto next;
 }