Message-ID: <967c9d22-7709-4f9b-aea5-8ad35dd24311@akamai.com>
Date: Tue, 24 Jun 2025 09:57:54 -0400
From: Jason Baron <jbaron@...mai.com>
To: Paolo Abeni <pabeni@...hat.com>, Jakub Kicinski <kuba@...nel.org>
Cc: davem@...emloft.net, edumazet@...gle.com, horms@...nel.org,
        kuniyu@...gle.com, netdev@...r.kernel.org,
        Kuniyuki Iwashima <kuni1840@...il.com>
Subject: Re: [PATCH net-next v2 3/3] netlink: Fix wraparound of
 sk->sk_rmem_alloc



On 6/24/25 3:55 AM, Paolo Abeni wrote:
> !-------------------------------------------------------------------|
>    This Message Is From an External Sender
>    This message came from outside your organization.
> |-------------------------------------------------------------------!
> 
> On 6/24/25 1:35 AM, Jakub Kicinski wrote:
>> On Wed, 18 Jun 2025 23:13:02 -0700 Kuniyuki Iwashima wrote:
>>> From: Jason Baron <jbaron@...mai.com>
>>> Date: Wed, 18 Jun 2025 19:13:23 -0400
>>>> For netlink sockets, when comparing allocated rmem memory with the
>>>> rcvbuf limit, the comparison is done using signed values. This means
>>>> that if rcvbuf is near INT_MAX, then sk->sk_rmem_alloc may become
>>>> negative in the comparison with rcvbuf which will yield incorrect
>>>> results.
>>>>
>>>> This can be reproduced by using the program from SOCK_DIAG(7) with
>>>> some slight modifications. First, setting sk->sk_rcvbuf to INT_MAX
>>>> using SO_RCVBUFFORCE and then secondly running the "send_query()"
>>>> in a loop while not calling "receive_responses()". In this case,
>>>> the value of sk->sk_rmem_alloc will continuously wrap around
>>>> and thus more memory is allocated than the sk->sk_rcvbuf limit.
>>>> This will eventually fill all of memory leading to an out of memory
>>>> condition with skbs filling up the slab.
>>>>
>>>> Let's fix this in a similar manner to:
>>>> commit 5a465a0da13e ("udp: Fix multiple wraparounds of sk->sk_rmem_alloc.")
>>>>
>>>> As noted in that fix, if there are multiple threads writing to a
>>>> netlink socket it's possible to slightly exceed rcvbuf value. But as
>>>> noted this avoids an expensive 'atomic_add_return()' for the common
>>>> case.
>>>
>>> This was because UDP RX path is the fast path, but netlink isn't.
>>> Also, it's common for UDP that multiple packets for the same socket
>>> are processed concurrently, and 850cbaddb52d dropped lock_sock from
>>> the path.
>>
>> To be clear -- are you saying we should fix this differently?
>> Or perhaps that the problem doesn't exist? The change doesn't
>> seem very intrusive..
> 
> AFAICS the race is possible even with netlink as netlink_unicast() runs
> without the socket lock, too.
> 
> The point is that for UDP the scenario with multiple threads enqueuing a
> packet into the same socket is a critical path, optimizing for
> performances and allowing some memory accounting inaccuracy makes sense.
> 
> For netlink socket, that scenario looks a patological one and I think we
> should prefer accuracy instead of optimization.
> 
> Thanks,
> 
> Paolo
> 

Hi,

Since the current netlink code already allows memory allocation to 
slightly exceed rcvbuf even when there is no overflow, I didn't think we 
would want to change this especially for smaller buffers as it may cause 
a regression. I tried to preserve this behavior if you look at the 
__sock_rcvbuf_has_space() function in patch 1/3.

For larger buffers, I think accurate accounting could make sense, but I 
thought it made more sense to try and optimize this common case as 
Kuniyuki Iwashima did, consolidate code between multiple use-cases and 
not overly special case handling for large buffers.

Thanks,

-Jason

Powered by blists - more mailing lists