[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-Id: <20191026114733.28111-15-pablo@netfilter.org>
Date: Sat, 26 Oct 2019 13:47:16 +0200
From: Pablo Neira Ayuso <pablo@...filter.org>
To: netfilter-devel@...r.kernel.org
Cc: davem@...emloft.net, netdev@...r.kernel.org
Subject: [PATCH 14/31] netfilter: ecache: document extension area access rules
From: Florian Westphal <fw@...len.de>
Once ct->ext gets free'd via kfree() rather than kfree_rcu we can't
access the extension area anymore without owning the conntrack.
This is a special case:
The worker is walking the pcpu dying list while holding dying list lock:
Neither ct nor ct->ext can be free'd until after the walk has completed.
Signed-off-by: Florian Westphal <fw@...len.de>
Signed-off-by: Pablo Neira Ayuso <pablo@...filter.org>
---
net/netfilter/nf_conntrack_ecache.c | 17 +++++++++++++++--
1 file changed, 15 insertions(+), 2 deletions(-)
diff --git a/net/netfilter/nf_conntrack_ecache.c b/net/netfilter/nf_conntrack_ecache.c
index 6fba74b5aaf7..0d83c159671c 100644
--- a/net/netfilter/nf_conntrack_ecache.c
+++ b/net/netfilter/nf_conntrack_ecache.c
@@ -30,6 +30,7 @@
static DEFINE_MUTEX(nf_ct_ecache_mutex);
#define ECACHE_RETRY_WAIT (HZ/10)
+#define ECACHE_STACK_ALLOC (256 / sizeof(void *))
enum retry_state {
STATE_CONGESTED,
@@ -39,11 +40,11 @@ enum retry_state {
static enum retry_state ecache_work_evict_list(struct ct_pcpu *pcpu)
{
- struct nf_conn *refs[16];
+ struct nf_conn *refs[ECACHE_STACK_ALLOC];
+ enum retry_state ret = STATE_DONE;
struct nf_conntrack_tuple_hash *h;
struct hlist_nulls_node *n;
unsigned int evicted = 0;
- enum retry_state ret = STATE_DONE;
spin_lock(&pcpu->lock);
@@ -54,10 +55,22 @@ static enum retry_state ecache_work_evict_list(struct ct_pcpu *pcpu)
if (!nf_ct_is_confirmed(ct))
continue;
+ /* This ecache access is safe because the ct is on the
+ * pcpu dying list and we hold the spinlock -- the entry
+ * cannot be free'd until after the lock is released.
+ *
+ * This is true even if ct has a refcount of 0: the
+ * cpu that is about to free the entry must remove it
+ * from the dying list and needs the lock to do so.
+ */
e = nf_ct_ecache_find(ct);
if (!e || e->state != NFCT_ECACHE_DESTROY_FAIL)
continue;
+ /* ct is in NFCT_ECACHE_DESTROY_FAIL state, this means
+ * the worker owns this entry: the ct will remain valid
+ * until the worker puts its ct reference.
+ */
if (nf_conntrack_event(IPCT_DESTROY, ct)) {
ret = STATE_CONGESTED;
break;
--
2.11.0
Powered by blists - more mailing lists