diff mbox

rps: Handle double list_add at __napi_schedule

Message ID 21d3b5e617ed44d217ce7f82e9fafa06.squirrel@www.codeaurora.org
State Rejected, archived
Delegated to: David Miller
Headers show

Commit Message

Subash Abhinov Kasiviswanathan June 15, 2015, 9:46 p.m. UTC
When NAPI_STATE_SCHED state is not set, enqueue_to_backlog()
will queue an IPI and add the backlog queue to the poll list. A packet
added by RPS onto the core could also add the NAPI backlog struct to the
poll list. This double addition to the list causes a crash -

2920.540304:   <2> list_add double add: new=ffffffc076ed2930,
prev=ffffffc076ed2930, next=ffffffc076ed2850.
   [<ffffffc000460dd4>] __list_add+0xcc/0xf0
2921.064962:   <2> [<ffffffc000b44880>] rps_trigger_softirq+0x1c/0x40
2921.070779:   <2> [<ffffffc000284a14>]
generic_smp_call_function_single_interrupt+0xe8/0x12c
2921.078678:   <2> [<ffffffc00020d9ac>] handle_IPI+0x8c/0x1ec
2921.083796:   <2> [<ffffffc000200714>] gic_handle_irq+0x94/0xb0

Fix this race for double addition to list by checking the NAPI state.

Acked-by: Sharat Masetty <smasetty@qti.qualcomm.com>
Signed-off-by: Subash Abhinov Kasiviswanathan <subashab@codeaurora.org>


--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Comments

Eric Dumazet June 15, 2015, 11:24 p.m. UTC | #1
On Mon, 2015-06-15 at 21:46 +0000, subashab@codeaurora.org wrote:
> When NAPI_STATE_SCHED state is not set, enqueue_to_backlog()
> will queue an IPI and add the backlog queue to the poll list. A packet
> added by RPS onto the core could also add the NAPI backlog struct to the
> poll list. This double addition to the list causes a crash -
> 
> 2920.540304:   <2> list_add double add: new=ffffffc076ed2930,
> prev=ffffffc076ed2930, next=ffffffc076ed2850.
>    [<ffffffc000460dd4>] __list_add+0xcc/0xf0
> 2921.064962:   <2> [<ffffffc000b44880>] rps_trigger_softirq+0x1c/0x40
> 2921.070779:   <2> [<ffffffc000284a14>]
> generic_smp_call_function_single_interrupt+0xe8/0x12c
> 2921.078678:   <2> [<ffffffc00020d9ac>] handle_IPI+0x8c/0x1ec
> 2921.083796:   <2> [<ffffffc000200714>] gic_handle_irq+0x94/0xb0
> 
> Fix this race for double addition to list by checking the NAPI state.
> 
> Acked-by: Sharat Masetty <smasetty@qti.qualcomm.com>
> Signed-off-by: Subash Abhinov Kasiviswanathan <subashab@codeaurora.org>
> 
> diff --git a/net/core/dev.c b/net/core/dev.c
> index 6f561de..57d6d39 100644
> --- a/net/core/dev.c
> +++ b/net/core/dev.c
> @@ -3225,7 +3225,8 @@ static void rps_trigger_softirq(void *data)
>  {
>         struct softnet_data *sd = data;
> 
> -       ____napi_schedule(sd, &sd->backlog);
> +       if (!test_bit(NAPI_STATE_SCHED, &sd->backlog.state))
> +               ____napi_schedule(sd, &sd->backlog);
>         sd->received_rps++;
>  }
> 

I can not believe how many times you tried to send RPS patches.

I do not see how this condition triggers. This code path is run billions
of time per ms on our hosts and never got a single crash.

Please describe where is the race condition you want to fix.

Your test is racy by definition.





--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
diff mbox

Patch

diff --git a/net/core/dev.c b/net/core/dev.c
index 6f561de..57d6d39 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c
@@ -3225,7 +3225,8 @@  static void rps_trigger_softirq(void *data)
 {
        struct softnet_data *sd = data;

-       ____napi_schedule(sd, &sd->backlog);
+       if (!test_bit(NAPI_STATE_SCHED, &sd->backlog.state))
+               ____napi_schedule(sd, &sd->backlog);
        sd->received_rps++;
 }