Skip to contentSkip to conversation

tun: publish tfile after it's fully initialized

PATCH In review Stanislav Fomichev 4 participants 5 messages Jan 7, 2019 – Jan 10, 2019

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH v2 1/2 tun: publish tfile after it's fully initialized

BUG: unable to handle kernel NULL pointer dereference at 00000000000000d1
Call Trace:
 ? napi_gro_frags+0xa7/0x2c0
 tun_get_user+0xb50/0xf20
 tun_chr_write_iter+0x53/0x70
 new_sync_write+0xff/0x160
 vfs_write+0x191/0x1e0
 __x64_sys_write+0x5e/0xd0
 do_syscall_64+0x47/0xf0
 entry_SYSCALL_64_after_hwframe+0x44/0xa9

I think there is a subtle race between sending a packet via tap and attaching it:

CPU0: CPU1:
tun_chr_ioctl(TUNSETIFF)
tun_set_iff

    tun_attach
      rcu_assign_pointer(tfile->tun, tun);
                         tun_fops->write_iter()
                           tun_chr_write_iter
                             tun_napi_alloc_frags
                               napi_get_frags
                                 napi->skb = napi_alloc_skb
      tun_napi_init
        netif_napi_add
          napi->skb = NULL
                              napi->skb is NULL here
                              napi_gro_frags
                                napi_frags_skb
				  skb = napi->skb
				  skb_reset_mac_header(skb)
				  panic()

Move rcu_assign_pointer(tfile->tun) and rcu_assign_pointer(tun->tfiles) to be the last thing we do in tun_attach(); this should guarantee that when we call tun_get() we always get an initialized object.

v2 changes:
* remove extra napi_mutex locks/unlocks for napi operations

Fixes 90e33d459407 ("tun: enable napi_gro_frags() for TUN/TAP driver")
---
 drivers/net/tun.c | 11 +++++++----
 1 file changed, 7 insertions(+), 4 deletions(-)
drivers/net/tun.c +7 -4
index a4fdad475594..18656c4094b3 100644
@@ -856,10 +856,6 @@ static int tun_attach(struct tun_struct *tun, struct file *file,
 		err = 0;
 	}
 
-	rcu_assign_pointer(tfile->tun, tun);
-	rcu_assign_pointer(tun->tfiles[tun->numqueues], tfile);
-	tun->numqueues++;
-
 	if (tfile->detached) {
 		tun_enable_queue(tfile);
 	} else {
@@ -876,6 +872,13 @@ static int tun_attach(struct tun_struct *tun, struct file *file,
 	 * refcnt.
 	 */
 
+	/* Publish tfile->tun and tun->tfiles only after we've fully
+	 * initialized tfile; otherwise we risk using half-initialized
+	 * object.
+	 */
+	rcu_assign_pointer(tfile->tun, tun);
+	rcu_assign_pointer(tun->tfiles[tun->numqueues], tfile);
+	tun->numqueues++;
 out:
 	return err;
 }
Signature · 1 line
2.20.1.97.g81188d93c3-goog
#1 lore raw

PATCH v2 2/2 tun: always set skb->dev to tun->dev

While debugging previous issue I noticed that commit 90e33d459407 ("tun:
enable napi_gro_frags() for TUN/TAP driver") started conditionally (!frags) calling eth_type_trans(skb, tun->dev) for IFF_TAP case. Since eth_type_trans sets skb->dev, some skbs can now have NULL skb->dev. Fix that by always setting skb->dev unconditionally.

The syzbot fails with the following trace:

WARNING: CPU: 0 PID: 11136 at net/core/flow_dissector.c:764
 skb_flow_dissect_flow_keys_basic include/linux/skbuff.h:1240 [inline]
 skb_probe_transport_header include/linux/skbuff.h:2403 [inline]
 tun_get_user+0x2d4a/0x4250 drivers/net/tun.c:1906
 tun_chr_write_iter+0xb9/0x160 drivers/net/tun.c:1993
 call_write_iter include/linux/fs.h:1808 [inline]
 new_sync_write fs/read_write.c:474 [inline]

But I don't think there is an actual issue since we exercise flow dissector via eth_get_headlen which doesn't use skb (and hence BPF flow dissector). But let's still properly set skb->dev so we don't have any problems going forward.

Fixes 90e33d459407 ("tun: enable napi_gro_frags() for TUN/TAP driver")
---
 drivers/net/tun.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/net/tun.c +1 -1
index 18656c4094b3..2dea2fb88b62 100644
@@ -1890,6 +1890,7 @@ static ssize_t tun_get_user(struct tun_struct *tun, struct tun_file *tfile,
 		return -EINVAL;
 	}
 
+	skb->dev = tun->dev;
 	switch (tun->flags & TUN_TYPE_MASK) {
 	case IFF_TUN:
 		if (tun->flags & IFF_NO_PI) {
@@ -1911,7 +1912,6 @@ static ssize_t tun_get_user(struct tun_struct *tun, struct tun_file *tfile,
 
 		skb_reset_mac_header(skb);
 		skb->protocol = pi.proto;
-		skb->dev = tun->dev;
 		break;
 	case IFF_TAP:
 		if (!frags)
Signature · 1 line
2.20.1.97.g81188d93c3-goog
maintainer
#2 lore raw

Re: [PATCH net v2 2/2] tun: always set skb->dev to tun->dev

Quoted 16 lines · Stanislav Fomichev

While debugging previous issue I noticed that commit 90e33d459407 ("tun:
enable napi_gro_frags() for TUN/TAP driver") started conditionally (!frags) calling eth_type_trans(skb, tun->dev) for IFF_TAP case. Since eth_type_trans sets skb->dev, some skbs can now have NULL skb->dev. Fix that by always setting skb->dev unconditionally.

The syzbot fails with the following trace:

WARNING: CPU: 0 PID: 11136 at net/core/flow_dissector.c:764
 skb_flow_dissect_flow_keys_basic include/linux/skbuff.h:1240 [inline]
 skb_probe_transport_header include/linux/skbuff.h:2403 [inline]
 tun_get_user+0x2d4a/0x4250 drivers/net/tun.c:1906
 tun_chr_write_iter+0xb9/0x160 drivers/net/tun.c:1993
 call_write_iter include/linux/fs.h:1808 [inline]
 new_sync_write fs/read_write.c:474 [inline]
Quoted 3 lines

But I don't think there is an actual issue since we exercise flow dissector via eth_get_headlen which doesn't use skb (and hence BPF flow dissector).

Do you mean skb_probe_transport_header?

if frags, tun_napi_alloc_frags will return napi->skb, which has skb->dev set by napi_reuse_skb.

I don't think this is needed.

#3 lore raw

Re: [PATCH net v2 2/2] tun: always set skb->dev to tun->dev

Quoted 17 lines · Willem de Bruijn
Quoted 16 lines · Stanislav Fomichev

While debugging previous issue I noticed that commit 90e33d459407 ("tun:
enable napi_gro_frags() for TUN/TAP driver") started conditionally (!frags) calling eth_type_trans(skb, tun->dev) for IFF_TAP case. Since eth_type_trans sets skb->dev, some skbs can now have NULL skb->dev. Fix that by always setting skb->dev unconditionally.

The syzbot fails with the following trace:

WARNING: CPU: 0 PID: 11136 at net/core/flow_dissector.c:764
 skb_flow_dissect_flow_keys_basic include/linux/skbuff.h:1240 [inline]
 skb_probe_transport_header include/linux/skbuff.h:2403 [inline]
 tun_get_user+0x2d4a/0x4250 drivers/net/tun.c:1906
 tun_chr_write_iter+0xb9/0x160 drivers/net/tun.c:1993
 call_write_iter include/linux/fs.h:1808 [inline]
 new_sync_write fs/read_write.c:474 [inline]
Quoted 5 lines
Quoted 3 lines

But I don't think there is an actual issue since we exercise flow dissector via eth_get_headlen which doesn't use skb (and hence BPF flow dissector).

Do you mean skb_probe_transport_header?

I was actually thinking about possible future conversion of eth_get_headlen to be able to run bpf flow dissector, I missed the fact that we already call dissector via skb_probe_transport_header :-(

Quoted 5 lines

if frags, tun_napi_alloc_frags will return napi->skb, which has skb->dev set by napi_reuse_skb.

I don't think this is needed.

Ah, indeed, napi_alloc_skb sets proper skb->dev.

maintainer
#4 lore raw
In reply to Stanislav Fomichev

From: Stanislav Fomichev <[email protected]>
Date: Mon, 7 Jan 2019 13:38:38 -0800

Quoted 2 lines
BUG: unable to handle kernel NULL pointer dereference at 00000000000000d1
Call Trace:

...

Quoted 6 lines

I think there is a subtle race between sending a packet via tap and attaching it:

CPU0: CPU1:
tun_chr_ioctl(TUNSETIFF)

...

Quoted 10 lines

Move rcu_assign_pointer(tfile->tun) and rcu_assign_pointer(tun->tfiles) to be the last thing we do in tun_attach(); this should guarantee that when we call tun_get() we always get an initialized object.

v2 changes:
* remove extra napi_mutex locks/unlocks for napi operations

Fixes 90e33d459407 ("tun: enable napi_gro_frags() for TUN/TAP driver")

Applied and queued up for -stable.

Please, the next time you submit a patch series, provide a proper header posting.

Thank you.