netfilter: conntrack: serialize hash resizes and cleanups
authorEric Dumazet <edumazet@google.com>
Fri, 17 Sep 2021 22:15:56 +0000 (15:15 -0700)
committerPablo Neira Ayuso <pablo@netfilter.org>
Tue, 21 Sep 2021 01:46:56 +0000 (03:46 +0200)
Syzbot was able to trigger the following warning [1]

No repro found by syzbot yet but I was able to trigger similar issue
by having 2 scripts running in parallel, changing conntrack hash sizes,
and:

for j in `seq 1 1000` ; do unshare -n /bin/true >/dev/null ; done

It would take more than 5 minutes for net_namespace structures
to be cleaned up.

This is because nf_ct_iterate_cleanup() has to restart everytime
a resize happened.

By adding a mutex, we can serialize hash resizes and cleanups
and also make get_next_corpse() faster by skipping over empty
buckets.

Even without resizes in the picture, this patch considerably
speeds up network namespace dismantles.

[1]
INFO: task syz-executor.0:8312 can't die for more than 144 seconds.
task:syz-executor.0  state:R  running task     stack:25672 pid: 8312 ppid:  6573 flags:0x00004006
Call Trace:
 context_switch kernel/sched/core.c:4955 [inline]
 __schedule+0x940/0x26f0 kernel/sched/core.c:6236
 preempt_schedule_common+0x45/0xc0 kernel/sched/core.c:6408
 preempt_schedule_thunk+0x16/0x18 arch/x86/entry/thunk_64.S:35
 __local_bh_enable_ip+0x109/0x120 kernel/softirq.c:390
 local_bh_enable include/linux/bottom_half.h:32 [inline]
 get_next_corpse net/netfilter/nf_conntrack_core.c:2252 [inline]
 nf_ct_iterate_cleanup+0x15a/0x450 net/netfilter/nf_conntrack_core.c:2275
 nf_conntrack_cleanup_net_list+0x14c/0x4f0 net/netfilter/nf_conntrack_core.c:2469
 ops_exit_list+0x10d/0x160 net/core/net_namespace.c:171
 setup_net+0x639/0xa30 net/core/net_namespace.c:349
 copy_net_ns+0x319/0x760 net/core/net_namespace.c:470
 create_new_namespaces+0x3f6/0xb20 kernel/nsproxy.c:110
 unshare_nsproxy_namespaces+0xc1/0x1f0 kernel/nsproxy.c:226
 ksys_unshare+0x445/0x920 kernel/fork.c:3128
 __do_sys_unshare kernel/fork.c:3202 [inline]
 __se_sys_unshare kernel/fork.c:3200 [inline]
 __x64_sys_unshare+0x2d/0x40 kernel/fork.c:3200
 do_syscall_x64 arch/x86/entry/common.c:50 [inline]
 do_syscall_64+0x35/0xb0 arch/x86/entry/common.c:80
 entry_SYSCALL_64_after_hwframe+0x44/0xae
RIP: 0033:0x7f63da68e739
RSP: 002b:00007f63d7c05188 EFLAGS: 00000246 ORIG_RAX: 0000000000000110
RAX: ffffffffffffffda RBX: 00007f63da792f80 RCX: 00007f63da68e739
RDX: 0000000000000000 RSI: 0000000000000000 RDI: 0000000040000000
RBP: 00007f63da6e8cc4 R08: 0000000000000000 R09: 0000000000000000
R10: 0000000000000000 R11: 0000000000000246 R12: 00007f63da792f80
R13: 00007fff50b75d3f R14: 00007f63d7c05300 R15: 0000000000022000

Showing all locks held in the system:
1 lock held by khungtaskd/27:
 #0: ffffffff8b980020 (rcu_read_lock){....}-{1:2}, at: debug_show_all_locks+0x53/0x260 kernel/locking/lockdep.c:6446
2 locks held by kworker/u4:2/153:
 #0: ffff888010c69138 ((wq_completion)events_unbound){+.+.}-{0:0}, at: arch_atomic64_set arch/x86/include/asm/atomic64_64.h:34 [inline]
 #0: ffff888010c69138 ((wq_completion)events_unbound){+.+.}-{0:0}, at: arch_atomic_long_set include/linux/atomic/atomic-long.h:41 [inline]
 #0: ffff888010c69138 ((wq_completion)events_unbound){+.+.}-{0:0}, at: atomic_long_set include/linux/atomic/atomic-instrumented.h:1198 [inline]
 #0: ffff888010c69138 ((wq_completion)events_unbound){+.+.}-{0:0}, at: set_work_data kernel/workqueue.c:634 [inline]
 #0: ffff888010c69138 ((wq_completion)events_unbound){+.+.}-{0:0}, at: set_work_pool_and_clear_pending kernel/workqueue.c:661 [inline]
 #0: ffff888010c69138 ((wq_completion)events_unbound){+.+.}-{0:0}, at: process_one_work+0x896/0x1690 kernel/workqueue.c:2268
 #1: ffffc9000140fdb0 ((kfence_timer).work){+.+.}-{0:0}, at: process_one_work+0x8ca/0x1690 kernel/workqueue.c:2272
1 lock held by systemd-udevd/2970:
1 lock held by in:imklog/6258:
 #0: ffff88807f970ff0 (&f->f_pos_lock){+.+.}-{3:3}, at: __fdget_pos+0xe9/0x100 fs/file.c:990
3 locks held by kworker/1:6/8158:
1 lock held by syz-executor.0/8312:
2 locks held by kworker/u4:13/9320:
1 lock held by syz-executor.5/10178:
1 lock held by syz-executor.4/10217:

Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: syzbot <syzkaller@googlegroups.com>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
net/netfilter/nf_conntrack_core.c

index 97b91d62589d5fc1d37090ffa6a7cd794a1920fb..770a63103c7a4240b8559a97f707588d569beba8 100644 (file)
@@ -74,6 +74,9 @@ static __read_mostly struct kmem_cache *nf_conntrack_cachep;
 static DEFINE_SPINLOCK(nf_conntrack_locks_all_lock);
 static __read_mostly bool nf_conntrack_locks_all;
 
+/* serialize hash resizes and nf_ct_iterate_cleanup */
+static DEFINE_MUTEX(nf_conntrack_mutex);
+
 #define GC_SCAN_INTERVAL       (120u * HZ)
 #define GC_SCAN_MAX_DURATION   msecs_to_jiffies(10)
 
@@ -2263,28 +2266,31 @@ get_next_corpse(int (*iter)(struct nf_conn *i, void *data),
        spinlock_t *lockp;
 
        for (; *bucket < nf_conntrack_htable_size; (*bucket)++) {
+               struct hlist_nulls_head *hslot = &nf_conntrack_hash[*bucket];
+
+               if (hlist_nulls_empty(hslot))
+                       continue;
+
                lockp = &nf_conntrack_locks[*bucket % CONNTRACK_LOCKS];
                local_bh_disable();
                nf_conntrack_lock(lockp);
-               if (*bucket < nf_conntrack_htable_size) {
-                       hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[*bucket], hnnode) {
-                               if (NF_CT_DIRECTION(h) != IP_CT_DIR_REPLY)
-                                       continue;
-                               /* All nf_conn objects are added to hash table twice, one
-                                * for original direction tuple, once for the reply tuple.
-                                *
-                                * Exception: In the IPS_NAT_CLASH case, only the reply
-                                * tuple is added (the original tuple already existed for
-                                * a different object).
-                                *
-                                * We only need to call the iterator once for each
-                                * conntrack, so we just use the 'reply' direction
-                                * tuple while iterating.
-                                */
-                               ct = nf_ct_tuplehash_to_ctrack(h);
-                               if (iter(ct, data))
-                                       goto found;
-                       }
+               hlist_nulls_for_each_entry(h, n, hslot, hnnode) {
+                       if (NF_CT_DIRECTION(h) != IP_CT_DIR_REPLY)
+                               continue;
+                       /* All nf_conn objects are added to hash table twice, one
+                        * for original direction tuple, once for the reply tuple.
+                        *
+                        * Exception: In the IPS_NAT_CLASH case, only the reply
+                        * tuple is added (the original tuple already existed for
+                        * a different object).
+                        *
+                        * We only need to call the iterator once for each
+                        * conntrack, so we just use the 'reply' direction
+                        * tuple while iterating.
+                        */
+                       ct = nf_ct_tuplehash_to_ctrack(h);
+                       if (iter(ct, data))
+                               goto found;
                }
                spin_unlock(lockp);
                local_bh_enable();
@@ -2302,26 +2308,20 @@ found:
 static void nf_ct_iterate_cleanup(int (*iter)(struct nf_conn *i, void *data),
                                  void *data, u32 portid, int report)
 {
-       unsigned int bucket = 0, sequence;
+       unsigned int bucket = 0;
        struct nf_conn *ct;
 
        might_sleep();
 
-       for (;;) {
-               sequence = read_seqcount_begin(&nf_conntrack_generation);
-
-               while ((ct = get_next_corpse(iter, data, &bucket)) != NULL) {
-                       /* Time to push up daises... */
+       mutex_lock(&nf_conntrack_mutex);
+       while ((ct = get_next_corpse(iter, data, &bucket)) != NULL) {
+               /* Time to push up daises... */
 
-                       nf_ct_delete(ct, portid, report);
-                       nf_ct_put(ct);
-                       cond_resched();
-               }
-
-               if (!read_seqcount_retry(&nf_conntrack_generation, sequence))
-                       break;
-               bucket = 0;
+               nf_ct_delete(ct, portid, report);
+               nf_ct_put(ct);
+               cond_resched();
        }
+       mutex_unlock(&nf_conntrack_mutex);
 }
 
 struct iter_data {
@@ -2557,8 +2557,10 @@ int nf_conntrack_hash_resize(unsigned int hashsize)
        if (!hash)
                return -ENOMEM;
 
+       mutex_lock(&nf_conntrack_mutex);
        old_size = nf_conntrack_htable_size;
        if (old_size == hashsize) {
+               mutex_unlock(&nf_conntrack_mutex);
                kvfree(hash);
                return 0;
        }
@@ -2598,6 +2600,8 @@ int nf_conntrack_hash_resize(unsigned int hashsize)
        nf_conntrack_all_unlock();
        local_bh_enable();
 
+       mutex_unlock(&nf_conntrack_mutex);
+
        synchronize_net();
        kvfree(old_hash);
        return 0;