Re: [RFC net] net: stream: don't purge sk_error_queue without holding its lock

On Tue, 14 Sep 2021 09:32:09 -0700 Eric Dumazet wrote:

On 9/14/21 7:18 AM, Jakub Kicinski wrote:

quoted

On Mon, 13 Sep 2021 22:14:00 -0700 Eric Dumazet wrote:  

quoted

quoted

This should not be needed.

By definition, sk_stream_kill_queues() is only called when there is no
more references on the sockets.

So all outstanding packets must have been orphaned or freed.  

I don't see the wait anywhere, would you mind spelling it out?
My (likely flawed) understanding is that inet_sock_destruct() gets
called when refs are gone (via sk->sk_destruct).  

But tcp_disconnect() + tcp_close() seem to happily call
inet_csk_destroy_sock() -> sk_stream_kill_queues() with outstanding
sk_wmem_alloc refs.  

tcp_disconnect() should probably leave the error queue as is.

For some reason I thought your report was about inet_sock_destruct()

tcp_disconnect() has always been full of bugs, it is surprising real applications
(not fuzzers) are still trying to use it.

I think I hit it because app sets SOCK_LINGER && !sk->sk_lingertime.
I don't think the app disconnects explicitly, but "same difference".

quoted

quoted

Anyway, Linux-2.6.12-rc2 had no timestamps yet.  

I see, thanks, if some form of the patch stands perhaps:

Fixes: cb9eff097831 ("net: new user space API for time stamping of incoming and outgoing packets")
  

Except that this patch wont prevent a packet being added to sk_error_queue
right after skb_queue_purge(&sk->sk_error_queue).

Right, but then inet_sock_destruct() also purges the err queue, again.
I was afraid of regressions but we could just remove the purging 
from sk_stream_kill_queues(), and target net-next?

If you think there is a bug, it must be fixed in another way.

IMO, preventing err packets from a prior session being queued after a tcp_disconnect()
is rather hard. We should not even try (packets could be stuck for hours in a qdisc)

Indeed, we could rearrange the SOCK_DEAD check in sock_queue_err_skb()
to skip queuing and put it under the err queue lock (provided we make
sk_stream_kill_queues() take that lock as well). But seems like an
overkill. I'd lean towards the existing patch or removing the purge from
sk_stream_kill_queues(). LMK what you prefer, this is not urgent.