]> git.baikalelectronics.ru Git - kernel.git/commitdiff
rhashtable: Fix rhashtable_remove failures
authorHerbert Xu <herbert@gondor.apana.org.au>
Sun, 15 Mar 2015 10:12:05 +0000 (21:12 +1100)
committerDavid S. Miller <davem@davemloft.net>
Mon, 16 Mar 2015 02:22:08 +0000 (22:22 -0400)
The commit 76bff2ee499f60150ad7747903abecb0b21f1479 ("rhashtable:
Free bucket tables asynchronously after rehash") causes gratuitous
failures in rhashtable_remove.

The reason is that it inadvertently introduced multiple rehashing
from the perspective of readers.  IOW it is now possible to see
more than two tables during a single RCU critical section.

Fortunately the other reader rhashtable_lookup already deals with
this correctly thanks to 62cd523211c1e03eb4ce7e0618396a526602cbac
("rhashtable: rhashtable: Move future_tbl into struct bucket_table")
so only rhashtable_remove is broken by this change.

This patch fixes this by looping over every table from the first
one to the last or until we find the element that we were trying
to delete.

Incidentally the simple test for detecting rehashing to prevent
starting another shrinking no longer works.  Since it isn't needed
anyway (the work queue and the mutex serves as a natural barrier
to unnecessary rehashes) I've simply killed the test.

Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
lib/rhashtable.c

index b916679b3e3b74a54e3e5b33a2a853f6a72e9605..c523d3a563aa8858acfba422305d07898ebc5a9f 100644 (file)
@@ -511,28 +511,25 @@ static bool __rhashtable_remove(struct rhashtable *ht,
  */
 bool rhashtable_remove(struct rhashtable *ht, struct rhash_head *obj)
 {
-       struct bucket_table *tbl, *old_tbl;
+       struct bucket_table *tbl;
        bool ret;
 
        rcu_read_lock();
 
-       old_tbl = rht_dereference_rcu(ht->tbl, ht);
-       ret = __rhashtable_remove(ht, old_tbl, obj);
+       tbl = rht_dereference_rcu(ht->tbl, ht);
 
        /* Because we have already taken (and released) the bucket
         * lock in old_tbl, if we find that future_tbl is not yet
         * visible then that guarantees the entry to still be in
-        * old_tbl if it exists.
+        * the old tbl if it exists.
         */
-       tbl = rht_dereference_rcu(old_tbl->future_tbl, ht) ?: old_tbl;
-       if (!ret && old_tbl != tbl)
-               ret = __rhashtable_remove(ht, tbl, obj);
+       while (!(ret = __rhashtable_remove(ht, tbl, obj)) &&
+              (tbl = rht_dereference_rcu(tbl->future_tbl, ht)))
+               ;
 
        if (ret) {
-               bool no_resize_running = tbl == old_tbl;
-
                atomic_dec(&ht->nelems);
-               if (no_resize_running && rht_shrink_below_30(ht, tbl))
+               if (rht_shrink_below_30(ht, tbl))
                        schedule_work(&ht->run_work);
        }