Skip to contentSkip to conversation

nfsd: don't halt scanning the DRC LRU list when there's an RC_INPROG entry

PATCH Applied Jeff Layton 2 participants 2 messages Jun 25, 2014

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 nfsd: don't halt scanning the DRC LRU list when there's an RC_INPROG entry

The patch below has already made it into mainline as commit 1b19453d1c6abcfa7c312ba6c9f11a277568fc94. I think we need it in stable as well. In addition to making the cache pruner more efficient, it also fixes a logic bug in prune_cache_entries. If num_drc_entries is greater than the max, then we can end up freeing the entry even if it's still RC_INPROG. This patch ensures that RC_INPROG entries are always skipped and that prevents a use-after-free.

Please apply this to any stable kernel, v3.9 or above. Original patch description follows.

See:

     https://bugzilla.kernel.org/show_bug.cgi?id=77031

Thanks to Tomas Kasparek for the bug report.

--------------------------[snip]----------------------------

Currently, the DRC cache pruner will stop scanning the list when it hits an entry that is RC_INPROG. It's possible however for a call to take a *very* long time. In that case, we don't want it to block other entries from being pruned if they are expired or we need to trim the cache to get back under the limit.

Fix the DRC cache pruner to just ignore RC_INPROG entries.

---
 fs/nfsd/nfscache.c | 17 ++++++++---------
 1 file changed, 8 insertions(+), 9 deletions(-)
fs/nfsd/nfscache.c +8 -9
index f8f060ffbf4f..6040da8830ff 100644
@@ -224,13 +224,6 @@ hash_refile(struct svc_cacherep *rp)
 	hlist_add_head(&rp->c_hash, cache_hash + hash_32(rp->c_xid, maskbits));
 }
 
-static inline bool
-nfsd_cache_entry_expired(struct svc_cacherep *rp)
-{
-	return rp->c_state != RC_INPROG &&
-	       time_after(jiffies, rp->c_timestamp + RC_EXPIRE);
-}
-
 /*
  * Walk the LRU list and prune off entries that are older than RC_EXPIRE.
  * Also prune the oldest ones when the total exceeds the max number of entries.
@@ -242,8 +235,14 @@ prune_cache_entries(void)
 	long freed = 0;
 
 	list_for_each_entry_safe(rp, tmp, &lru_head, c_lru) {
-		if (!nfsd_cache_entry_expired(rp) &&
-		    num_drc_entries <= max_drc_entries)
+		/*
+		 * Don't free entries attached to calls that are still
+		 * in-progress, but do keep scanning the list.
+		 */
+		if (rp->c_state == RC_INPROG)
+			continue;
+		if (num_drc_entries <= max_drc_entries &&
+		    time_before(jiffies, rp->c_timestamp + RC_EXPIRE))
 			break;
 		nfsd_reply_cache_free_locked(rp);
 		freed++;
Signature · 1 line
1.9.3
Applied
#1 lore raw
Quoted 10 lines · Jeff Layton

The patch below has already made it into mainline as commit 1b19453d1c6abcfa7c312ba6c9f11a277568fc94. I think we need it in stable as well. In addition to making the cache pruner more efficient, it also fixes a logic bug in prune_cache_entries. If num_drc_entries is greater than the max, then we can end up freeing the entry even if it's still RC_INPROG. This patch ensures that RC_INPROG entries are always skipped and that prevents a use-after-free.

Please apply this to any stable kernel, v3.9 or above. Original patch description follows.

Pushed to 3.12 too. Thanks.

Signature · 2 lines
js
suse labs