inet: frags: rework rhashtable dismantle
authorEric Dumazet <edumazet@google.com>
Fri, 24 May 2019 16:03:40 +0000 (09:03 -0700)
committerDavid S. Miller <davem@davemloft.net>
Sun, 26 May 2019 21:08:05 +0000 (14:08 -0700)
syszbot found an interesting use-after-free [1] happening
while IPv4 fragment rhashtable was destroyed at netns dismantle.

While no insertions can possibly happen at the time a dismantling
netns is destroying this rhashtable, timers can still fire and
attempt to remove elements from this rhashtable.

This is forbidden, since rhashtable_free_and_destroy() has
no synchronization against concurrent inserts and deletes.

Add a new fqdir->dead flag so that timers do not attempt
a rhashtable_remove_fast() operation.

We also have to respect an RCU grace period before starting
the rhashtable_free_and_destroy() from process context,
thus we use rcu_work infrastructure.

This is a refinement of a prior rough attempt to fix this bug :
https://marc.info/?l=linux-netdev&m=153845936820900&w=2

Since the rhashtable cleanup is now deferred to a work queue,
netns dismantles should be slightly faster.

[1]
BUG: KASAN: use-after-free in __read_once_size include/linux/compiler.h:194 [inline]
BUG: KASAN: use-after-free in rhashtable_last_table+0x162/0x180 lib/rhashtable.c:212
Read of size 8 at addr ffff8880a6497b70 by task kworker/0:0/5

CPU: 0 PID: 5 Comm: kworker/0:0 Not tainted 5.2.0-rc1+ #2
Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011
Workqueue: events rht_deferred_worker
Call Trace:
 __dump_stack lib/dump_stack.c:77 [inline]
 dump_stack+0x172/0x1f0 lib/dump_stack.c:113
 print_address_description.cold+0x7c/0x20d mm/kasan/report.c:188
 __kasan_report.cold+0x1b/0x40 mm/kasan/report.c:317
 kasan_report+0x12/0x20 mm/kasan/common.c:614
 __asan_report_load8_noabort+0x14/0x20 mm/kasan/generic_report.c:132
 __read_once_size include/linux/compiler.h:194 [inline]
 rhashtable_last_table+0x162/0x180 lib/rhashtable.c:212
 rht_deferred_worker+0x111/0x2030 lib/rhashtable.c:411
 process_one_work+0x989/0x1790 kernel/workqueue.c:2269
 worker_thread+0x98/0xe40 kernel/workqueue.c:2415
 kthread+0x354/0x420 kernel/kthread.c:255
 ret_from_fork+0x24/0x30 arch/x86/entry/entry_64.S:352

Allocated by task 32687:
 save_stack+0x23/0x90 mm/kasan/common.c:71
 set_track mm/kasan/common.c:79 [inline]
 __kasan_kmalloc mm/kasan/common.c:489 [inline]
 __kasan_kmalloc.constprop.0+0xcf/0xe0 mm/kasan/common.c:462
 kasan_kmalloc+0x9/0x10 mm/kasan/common.c:503
 __do_kmalloc_node mm/slab.c:3620 [inline]
 __kmalloc_node+0x4e/0x70 mm/slab.c:3627
 kmalloc_node include/linux/slab.h:590 [inline]
 kvmalloc_node+0x68/0x100 mm/util.c:431
 kvmalloc include/linux/mm.h:637 [inline]
 kvzalloc include/linux/mm.h:645 [inline]
 bucket_table_alloc+0x90/0x480 lib/rhashtable.c:178
 rhashtable_init+0x3f4/0x7b0 lib/rhashtable.c:1057
 inet_frags_init_net include/net/inet_frag.h:109 [inline]
 ipv4_frags_init_net+0x182/0x410 net/ipv4/ip_fragment.c:683
 ops_init+0xb3/0x410 net/core/net_namespace.c:130
 setup_net+0x2d3/0x740 net/core/net_namespace.c:316
 copy_net_ns+0x1df/0x340 net/core/net_namespace.c:439
 create_new_namespaces+0x400/0x7b0 kernel/nsproxy.c:107
 unshare_nsproxy_namespaces+0xc2/0x200 kernel/nsproxy.c:206
 ksys_unshare+0x440/0x980 kernel/fork.c:2692
 __do_sys_unshare kernel/fork.c:2760 [inline]
 __se_sys_unshare kernel/fork.c:2758 [inline]
 __x64_sys_unshare+0x31/0x40 kernel/fork.c:2758
 do_syscall_64+0xfd/0x680 arch/x86/entry/common.c:301
 entry_SYSCALL_64_after_hwframe+0x49/0xbe

Freed by task 7:
 save_stack+0x23/0x90 mm/kasan/common.c:71
 set_track mm/kasan/common.c:79 [inline]
 __kasan_slab_free+0x102/0x150 mm/kasan/common.c:451
 kasan_slab_free+0xe/0x10 mm/kasan/common.c:459
 __cache_free mm/slab.c:3432 [inline]
 kfree+0xcf/0x220 mm/slab.c:3755
 kvfree+0x61/0x70 mm/util.c:460
 bucket_table_free+0x69/0x150 lib/rhashtable.c:108
 rhashtable_free_and_destroy+0x165/0x8b0 lib/rhashtable.c:1155
 inet_frags_exit_net+0x3d/0x50 net/ipv4/inet_fragment.c:152
 ipv4_frags_exit_net+0x73/0x90 net/ipv4/ip_fragment.c:695
 ops_exit_list.isra.0+0xaa/0x150 net/core/net_namespace.c:154
 cleanup_net+0x3fb/0x960 net/core/net_namespace.c:553
 process_one_work+0x989/0x1790 kernel/workqueue.c:2269
 worker_thread+0x98/0xe40 kernel/workqueue.c:2415
 kthread+0x354/0x420 kernel/kthread.c:255
 ret_from_fork+0x24/0x30 arch/x86/entry/entry_64.S:352

The buggy address belongs to the object at ffff8880a6497b40
 which belongs to the cache kmalloc-1k of size 1024
The buggy address is located 48 bytes inside of
 1024-byte region [ffff8880a6497b40ffff8880a6497f40)
The buggy address belongs to the page:
page:ffffea0002992580 refcount:1 mapcount:0 mapping:ffff8880aa400ac0 index:0xffff8880a64964c0 compound_mapcount: 0
flags: 0x1fffc0000010200(slab|head)
raw: 01fffc0000010200 ffffea0002916e88 ffffea000218fe08 ffff8880aa400ac0
raw: ffff8880a64964c0 ffff8880a6496040 0000000100000005 0000000000000000
page dumped because: kasan: bad access detected

Memory state around the buggy address:
 ffff8880a6497a00: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb
 ffff8880a6497a80: fb fb fb fb fb fb fb fb fc fc fc fc fc fc fc fc
>ffff8880a6497b00: fc fc fc fc fc fc fc fc fb fb fb fb fb fb fb fb
                                                             ^
 ffff8880a6497b80: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb
 ffff8880a6497c00: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb

Fixes: 648700f76b03 ("inet: frags: use rhashtables for reassembly units")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: syzbot <syzkaller@googlegroups.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
include/net/inet_frag.h
net/ipv4/inet_fragment.c

index 5f754c6..002f23c 100644 (file)
@@ -13,11 +13,13 @@ struct fqdir {
        int                     max_dist;
        struct inet_frags       *f;
        struct net              *net;
+       bool                    dead;
 
        struct rhashtable       rhashtable ____cacheline_aligned_in_smp;
 
        /* Keep atomic mem on separate cachelines in structs that include it */
        atomic_long_t           mem ____cacheline_aligned_in_smp;
+       struct rcu_work         destroy_rwork;
 };
 
 /**
@@ -26,11 +28,13 @@ struct fqdir {
  * @INET_FRAG_FIRST_IN: first fragment has arrived
  * @INET_FRAG_LAST_IN: final fragment has arrived
  * @INET_FRAG_COMPLETE: frag queue has been processed and is due for destruction
+ * @INET_FRAG_HASH_DEAD: inet_frag_kill() has not removed fq from rhashtable
  */
 enum {
        INET_FRAG_FIRST_IN      = BIT(0),
        INET_FRAG_LAST_IN       = BIT(1),
        INET_FRAG_COMPLETE      = BIT(2),
+       INET_FRAG_HASH_DEAD     = BIT(3),
 };
 
 struct frag_v4_compare_key {
index b4432f2..6ca9523 100644 (file)
@@ -124,34 +124,49 @@ void inet_frags_fini(struct inet_frags *f)
 }
 EXPORT_SYMBOL(inet_frags_fini);
 
+/* called from rhashtable_free_and_destroy() at netns_frags dismantle */
 static void inet_frags_free_cb(void *ptr, void *arg)
 {
        struct inet_frag_queue *fq = ptr;
+       int count;
 
-       /* If we can not cancel the timer, it means this frag_queue
-        * is already disappearing, we have nothing to do.
-        * Otherwise, we own a refcount until the end of this function.
-        */
-       if (!del_timer(&fq->timer))
-               return;
+       count = del_timer_sync(&fq->timer) ? 1 : 0;
 
        spin_lock_bh(&fq->lock);
        if (!(fq->flags & INET_FRAG_COMPLETE)) {
                fq->flags |= INET_FRAG_COMPLETE;
-               refcount_dec(&fq->refcnt);
+               count++;
+       } else if (fq->flags & INET_FRAG_HASH_DEAD) {
+               count++;
        }
        spin_unlock_bh(&fq->lock);
 
-       inet_frag_put(fq);
+       if (refcount_sub_and_test(count, &fq->refcnt))
+               inet_frag_destroy(fq);
 }
 
-void fqdir_exit(struct fqdir *fqdir)
+static void fqdir_rwork_fn(struct work_struct *work)
 {
-       fqdir->high_thresh = 0; /* prevent creation of new frags */
+       struct fqdir *fqdir = container_of(to_rcu_work(work),
+                                          struct fqdir, destroy_rwork);
 
        rhashtable_free_and_destroy(&fqdir->rhashtable, inet_frags_free_cb, NULL);
        kfree(fqdir);
 }
+
+void fqdir_exit(struct fqdir *fqdir)
+{
+       fqdir->high_thresh = 0; /* prevent creation of new frags */
+
+       /* paired with READ_ONCE() in inet_frag_kill() :
+        * We want to prevent rhashtable_remove_fast() calls
+        */
+       smp_store_release(&fqdir->dead, true);
+
+       INIT_RCU_WORK(&fqdir->destroy_rwork, fqdir_rwork_fn);
+       queue_rcu_work(system_wq, &fqdir->destroy_rwork);
+
+}
 EXPORT_SYMBOL(fqdir_exit);
 
 void inet_frag_kill(struct inet_frag_queue *fq)
@@ -163,8 +178,18 @@ void inet_frag_kill(struct inet_frag_queue *fq)
                struct fqdir *fqdir = fq->fqdir;
 
                fq->flags |= INET_FRAG_COMPLETE;
-               rhashtable_remove_fast(&fqdir->rhashtable, &fq->node, fqdir->f->rhash_params);
-               refcount_dec(&fq->refcnt);
+               rcu_read_lock();
+               /* This READ_ONCE() is paired with smp_store_release()
+                * in inet_frags_exit_net().
+                */
+               if (!READ_ONCE(fqdir->dead)) {
+                       rhashtable_remove_fast(&fqdir->rhashtable, &fq->node,
+                                              fqdir->f->rhash_params);
+                       refcount_dec(&fq->refcnt);
+               } else {
+                       fq->flags |= INET_FRAG_HASH_DEAD;
+               }
+               rcu_read_unlock();
        }
 }
 EXPORT_SYMBOL(inet_frag_kill);