Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]


Groups > linux.kernel > #1732303 > unrolled thread

[PATCH] net/packet: fix race condition between fanout_add and __unregister_prot_hook

Started bynixiaoming <nixiaoming@huawei.com>
First post2017-09-14 16:30 +0200
Last post2017-09-15 19:50 +0200
Articles 5 — 3 participants

Back to article view | Back to linux.kernel


Contents

  [PATCH] net/packet: fix race condition between fanout_add and __unregister_prot_hook nixiaoming <nixiaoming@huawei.com> - 2017-09-14 16:30 +0200
    Re: [PATCH] net/packet: fix race condition between fanout_add and __unregister_prot_hook Willem de Bruijn <willemdebruijn.kernel@gmail.com> - 2017-09-14 16:40 +0200
      Re: [PATCH] net/packet: fix race condition between fanout_add and __unregister_prot_hook Willem de Bruijn <willemdebruijn.kernel@gmail.com> - 2017-09-15 19:50 +0200
        Re: [PATCH] net/packet: fix race condition between fanout_add and __unregister_prot_hook Cong Wang <xiyou.wangcong@gmail.com> - 2017-09-15 20:20 +0200
      Re: [PATCH] net/packet: fix race condition between fanout_add and __unregister_prot_hook Cong Wang <xiyou.wangcong@gmail.com> - 2017-09-15 19:50 +0200

#1732303 — [PATCH] net/packet: fix race condition between fanout_add and __unregister_prot_hook

Fromnixiaoming <nixiaoming@huawei.com>
Date2017-09-14 16:30 +0200
Subject[PATCH] net/packet: fix race condition between fanout_add and __unregister_prot_hook
Message-ID<upDjr-6by-7@gated-at.bofh.it>
From: l00219569 <lisimin@huawei.com>

If fanout_add is preempted after running po-> fanout = match
and before running __fanout_link,
it will cause BUG_ON when __unregister_prot_hook call __fanout_unlink

so, we need add mutex_lock(&fanout_mutex) to __unregister_prot_hook
or add spin_lock(&po->bind_lock) before po-> fanout = match

this is a patch for add po->bind_lock in fanout_add

test on linux 4.1.12:
./trinity -c setsockopt -C 2 -X &

BUG: failure at net/packet/af_packet.c:1414/__fanout_unlink()!
Kernel panic - not syncing: BUG!
CPU: 2 PID: 2271 Comm: trinity-c0 Tainted: G        W  O    4.1.12 #1
Hardware name: Hisilicon PhosphorHi1382 FPGA (DT)
Call trace:
[<ffffffc000209414>] dump_backtrace+0x0/0xf8
[<ffffffc00020952c>] show_stack+0x20/0x28
[<ffffffc000635574>] dump_stack+0xac/0xe4
[<ffffffc000633fb8>] panic+0xf8/0x268
[<ffffffc0005fa778>] __unregister_prot_hook+0xa0/0x144
[<ffffffc0005fba48>] packet_set_ring+0x280/0x5b4
[<ffffffc0005fc33c>] packet_setsockopt+0x320/0x950
[<ffffffc000554a04>] SyS_setsockopt+0xa4/0xd4

Signed-off-by: nixiaoming <nixiaoming@huawei.com>
Tested-by: wudesheng <dede.wu@huawei.com>
---
 net/packet/af_packet.c | 11 ++++++++---
 1 file changed, 8 insertions(+), 3 deletions(-)

diff --git a/net/packet/af_packet.c b/net/packet/af_packet.c
index 54a18a8..7a52a3b 100644
--- a/net/packet/af_packet.c
+++ b/net/packet/af_packet.c
@@ -1446,12 +1446,16 @@ static int fanout_add(struct sock *sk, u16 id, u16 type_flags)
 	default:
 		return -EINVAL;
 	}
-
-	if (!po->running)
+	spin_lock(&po->bind_lock);
+	if (!po->running) {
+		spin_unlock(&po->bind_lock);
 		return -EINVAL;
+	}
 
-	if (po->fanout)
+	if (po->fanout) {
+		spin_unlock(&po->bind_lock);
 		return -EALREADY;
+	}
 
 	mutex_lock(&fanout_mutex);
 	match = NULL;
@@ -1501,6 +1505,7 @@ static int fanout_add(struct sock *sk, u16 id, u16 type_flags)
 	}
 out:
 	mutex_unlock(&fanout_mutex);
+	spin_unlock(&po->bind_lock);
 	return err;
 }
 
-- 
2.10.1

[toc] | [next] | [standalone]


#1732317

FromWillem de Bruijn <willemdebruijn.kernel@gmail.com>
Date2017-09-14 16:40 +0200
Message-ID<upDt8-6ey-21@gated-at.bofh.it>
In reply to#1732303
On Thu, Sep 14, 2017 at 10:07 AM, nixiaoming <nixiaoming@huawei.com> wrote:
> From: l00219569 <lisimin@huawei.com>
>
> If fanout_add is preempted after running po-> fanout = match
> and before running __fanout_link,
> it will cause BUG_ON when __unregister_prot_hook call __fanout_unlink
>
> so, we need add mutex_lock(&fanout_mutex) to __unregister_prot_hook

The packet socket code has no shortage of locks, so there are many
ways to avoid the race condition between fanout_add and packet_set_ring.

Another option would be to lock the socket when calling fanout_add:

    -               return fanout_add(sk, val & 0xffff, val >> 16);
    +               lock_sock(sk);
    +               ret = fanout_add(sk, val & 0xffff, val >> 16);
    +               release_sock(sk);
    +               return ret;

But, for consistency, and to be able to continue to make sense of the
locking policy, we should use the most appropriate lock. This
is po->bind_lock, as it ensures atomicity between testing whether
a protocol hook is active through po->running and the actual existence
of that hook on the protocol hook list.

fanout_mutex protects the fanout object's list. Taking that on
__unregister_prot_hook even in the case where fanout is not
used (and __dev_remove_pack is called) complicates locking
in this already complicated code.

> or add spin_lock(&po->bind_lock) before po-> fanout = match
>
> this is a patch for add po->bind_lock in fanout_add
>
> test on linux 4.1.12:
> ./trinity -c setsockopt -C 2 -X &

Thanks for testing!

>
> BUG: failure at net/packet/af_packet.c:1414/__fanout_unlink()!
> Kernel panic - not syncing: BUG!
> CPU: 2 PID: 2271 Comm: trinity-c0 Tainted: G        W  O    4.1.12 #1
> Hardware name: Hisilicon PhosphorHi1382 FPGA (DT)
> Call trace:
> [<ffffffc000209414>] dump_backtrace+0x0/0xf8
> [<ffffffc00020952c>] show_stack+0x20/0x28
> [<ffffffc000635574>] dump_stack+0xac/0xe4
> [<ffffffc000633fb8>] panic+0xf8/0x268
> [<ffffffc0005fa778>] __unregister_prot_hook+0xa0/0x144
> [<ffffffc0005fba48>] packet_set_ring+0x280/0x5b4
> [<ffffffc0005fc33c>] packet_setsockopt+0x320/0x950
> [<ffffffc000554a04>] SyS_setsockopt+0xa4/0xd4
>
> Signed-off-by: nixiaoming <nixiaoming@huawei.com>
> Tested-by: wudesheng <dede.wu@huawei.com>
> ---
>  net/packet/af_packet.c | 11 ++++++++---
>  1 file changed, 8 insertions(+), 3 deletions(-)
>
> diff --git a/net/packet/af_packet.c b/net/packet/af_packet.c
> index 54a18a8..7a52a3b 100644
> --- a/net/packet/af_packet.c
> +++ b/net/packet/af_packet.c
> @@ -1446,12 +1446,16 @@ static int fanout_add(struct sock *sk, u16 id, u16 type_flags)
>         default:
>                 return -EINVAL;
>         }
> -
> -       if (!po->running)
> +       spin_lock(&po->bind_lock);
> +       if (!po->running) {
> +               spin_unlock(&po->bind_lock);
>                 return -EINVAL;
> +       }
>
> -       if (po->fanout)
> +       if (po->fanout) {
> +               spin_unlock(&po->bind_lock);
>                 return -EALREADY;
> +       }
>
>         mutex_lock(&fanout_mutex);
>         match = NULL;
> @@ -1501,6 +1505,7 @@ static int fanout_add(struct sock *sk, u16 id, u16 type_flags)
>         }
>  out:
>         mutex_unlock(&fanout_mutex);
> +       spin_unlock(&po->bind_lock);

This function can call kzalloc with GFP_KERNEL, which may sleep. It is
not correct to sleep while holding a spinlock. Which is why I take the lock
later and test po->running again.

I will clean up that patch and send it for review.

[toc] | [prev] | [next] | [standalone]


#1732984

FromWillem de Bruijn <willemdebruijn.kernel@gmail.com>
Date2017-09-15 19:50 +0200
Message-ID<uq2Ux-69v-11@gated-at.bofh.it>
In reply to#1732317
On Fri, Sep 15, 2017 at 1:41 PM, Cong Wang <xiyou.wangcong@gmail.com> wrote:
> On Thu, Sep 14, 2017 at 7:35 AM, Willem de Bruijn
> <willemdebruijn.kernel@gmail.com> wrote:
>> On Thu, Sep 14, 2017 at 10:07 AM, nixiaoming <nixiaoming@huawei.com> wrote:
>>> From: l00219569 <lisimin@huawei.com>
>>>
>>> If fanout_add is preempted after running po-> fanout = match
>>> and before running __fanout_link,
>>> it will cause BUG_ON when __unregister_prot_hook call __fanout_unlink
>>>
>>> so, we need add mutex_lock(&fanout_mutex) to __unregister_prot_hook
>>
>> The packet socket code has no shortage of locks, so there are many
>> ways to avoid the race condition between fanout_add and packet_set_ring.
>>
>> Another option would be to lock the socket when calling fanout_add:
>>
>>     -               return fanout_add(sk, val & 0xffff, val >> 16);
>>     +               lock_sock(sk);
>>     +               ret = fanout_add(sk, val & 0xffff, val >> 16);
>>     +               release_sock(sk);
>>     +               return ret;
>>
>
> I don't think this is an option, because __unregister_prot_hook()
> can be called without lock_sock(), for example in packet_notifier().
>
>
>> But, for consistency, and to be able to continue to make sense of the
>> locking policy, we should use the most appropriate lock. This
>> is po->bind_lock, as it ensures atomicity between testing whether
>> a protocol hook is active through po->running and the actual existence
>> of that hook on the protocol hook list.
>
> Yeah, register_prot_hook() and unregister_prot_hook() already assume
> bind_lock.
>
> [...]
>
>>>  out:
>>>         mutex_unlock(&fanout_mutex);
>>> +       spin_unlock(&po->bind_lock);
>>
>> This function can call kzalloc with GFP_KERNEL, which may sleep. It is
>> not correct to sleep while holding a spinlock. Which is why I take the lock
>> later and test po->running again.
>
>
> Right, no need to mention the mutex_unlock() before the spin_unlock()
> is clearly wrong.
>
>
>>
>> I will clean up that patch and send it for review.
>
> How about the following patch?
>
>
> diff --git a/net/packet/af_packet.c b/net/packet/af_packet.c
> index c26172995511..f5c696a548ed 100644
> --- a/net/packet/af_packet.c
> +++ b/net/packet/af_packet.c
> @@ -1754,10 +1754,14 @@ static int fanout_add(struct sock *sk, u16 id,
> u16 type_flags)
>             match->prot_hook.dev == po->prot_hook.dev) {
>                 err = -ENOSPC;
>                 if (refcount_read(&match->sk_ref) < PACKET_FANOUT_MAX) {
> +                       spin_lock(&po->bind_lock);
>                         __dev_remove_pack(&po->prot_hook);
> -                       po->fanout = match;
> -                       refcount_set(&match->sk_ref,
> refcount_read(&match->sk_ref) + 1);
> -                       __fanout_link(sk, po);
> +                       if (po->running) {
> +                               refcount_set(&match->sk_ref,
> refcount_read(&match->sk_ref) + 1);
> +                               po->fanout = match;
> +                               __fanout_link(sk, po);
> +                       }
> +                       spin_unlock(&po->bind_lock);
>                         err = 0;
>                 }
>         }

In case of failure we also need to unlink and free match. I
sent the following:

http://patchwork.ozlabs.org/patch/813945/

[toc] | [prev] | [next] | [standalone]


#1733004

FromCong Wang <xiyou.wangcong@gmail.com>
Date2017-09-15 20:20 +0200
Message-ID<uq3nA-6EQ-11@gated-at.bofh.it>
In reply to#1732984
On Fri, Sep 15, 2017 at 10:46 AM, Willem de Bruijn
<willemdebruijn.kernel@gmail.com> wrote:
>
> In case of failure we also need to unlink and free match. I
> sent the following:
>
> http://patchwork.ozlabs.org/patch/813945/

Ah, will take a look.

[toc] | [prev] | [next] | [standalone]


#1732988

FromCong Wang <xiyou.wangcong@gmail.com>
Date2017-09-15 19:50 +0200
Message-ID<uq2Ux-69v-13@gated-at.bofh.it>
In reply to#1732317
On Thu, Sep 14, 2017 at 7:35 AM, Willem de Bruijn
<willemdebruijn.kernel@gmail.com> wrote:
> On Thu, Sep 14, 2017 at 10:07 AM, nixiaoming <nixiaoming@huawei.com> wrote:
>> From: l00219569 <lisimin@huawei.com>
>>
>> If fanout_add is preempted after running po-> fanout = match
>> and before running __fanout_link,
>> it will cause BUG_ON when __unregister_prot_hook call __fanout_unlink
>>
>> so, we need add mutex_lock(&fanout_mutex) to __unregister_prot_hook
>
> The packet socket code has no shortage of locks, so there are many
> ways to avoid the race condition between fanout_add and packet_set_ring.
>
> Another option would be to lock the socket when calling fanout_add:
>
>     -               return fanout_add(sk, val & 0xffff, val >> 16);
>     +               lock_sock(sk);
>     +               ret = fanout_add(sk, val & 0xffff, val >> 16);
>     +               release_sock(sk);
>     +               return ret;
>

I don't think this is an option, because __unregister_prot_hook()
can be called without lock_sock(), for example in packet_notifier().


> But, for consistency, and to be able to continue to make sense of the
> locking policy, we should use the most appropriate lock. This
> is po->bind_lock, as it ensures atomicity between testing whether
> a protocol hook is active through po->running and the actual existence
> of that hook on the protocol hook list.

Yeah, register_prot_hook() and unregister_prot_hook() already assume
bind_lock.

[...]

>>  out:
>>         mutex_unlock(&fanout_mutex);
>> +       spin_unlock(&po->bind_lock);
>
> This function can call kzalloc with GFP_KERNEL, which may sleep. It is
> not correct to sleep while holding a spinlock. Which is why I take the lock
> later and test po->running again.


Right, no need to mention the mutex_unlock() before the spin_unlock()
is clearly wrong.


>
> I will clean up that patch and send it for review.

How about the following patch?


diff --git a/net/packet/af_packet.c b/net/packet/af_packet.c
index c26172995511..f5c696a548ed 100644
--- a/net/packet/af_packet.c
+++ b/net/packet/af_packet.c
@@ -1754,10 +1754,14 @@ static int fanout_add(struct sock *sk, u16 id,
u16 type_flags)
            match->prot_hook.dev == po->prot_hook.dev) {
                err = -ENOSPC;
                if (refcount_read(&match->sk_ref) < PACKET_FANOUT_MAX) {
+                       spin_lock(&po->bind_lock);
                        __dev_remove_pack(&po->prot_hook);
-                       po->fanout = match;
-                       refcount_set(&match->sk_ref,
refcount_read(&match->sk_ref) + 1);
-                       __fanout_link(sk, po);
+                       if (po->running) {
+                               refcount_set(&match->sk_ref,
refcount_read(&match->sk_ref) + 1);
+                               po->fanout = match;
+                               __fanout_link(sk, po);
+                       }
+                       spin_unlock(&po->bind_lock);
                        err = 0;
                }
        }

[toc] | [prev] | [standalone]


Back to top | Article view | linux.kernel


csiph-web