lkml.org 
[lkml]   [2021]   [Jun]   [30]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [PATCH] tcp: Do not reset the icsk_ca_initialized in tcp_init_transfer.
From
Date
On 6/29/2021 11:59 PM, Neal Cardwell wrote:
> On Tue, Jun 29, 2021 at 8:58 AM Eric Dumazet <edumazet@google.com> wrote:
> From my perspective, the bug was introduced when that 8919a9b31eb4
> commit introduced icsk_ca_initialized and set icsk_ca_initialized to 0
> in tcp_init_transfer(), missing the possibility that a process could
> call setsockopt(TCP_CONGESTION) in state TCP_SYN_SENT (i.e. after the
> connect() or TFO open sendmsg()), which would call
> tcp_init_congestion_control(). The 8919a9b31eb4 commit did not intend
> to reset any initialization that the user had already explicitly made;
> it just missed the possibility of that particular sequence (which
> syzkaller managed to find!).
>
>> Although I am not sure what happens at accept() time when the listener
>> socket is cloned.
>
> It seems that for listener sockets, they cannot initialize their CC
> module state, because there is no way for them to reach
> tcp_init_congestion_control(), since:
>
> (a) tcp_set_congestion_control() -> tcp_reinit_congestion_control()
> will not call tcp_init_congestion_control() on a socket in CLOSE or
> LISTEN
>
> (b) tcp_init_transfer() -> tcp_init_congestion_control() can only
> happen for established sockets and successful TFO SYN_RECV sockets
Is this what was mentioned in this commit ce69e563b325(tcp: make sure
listeners don't initialize congestion-control state)

> --
> [PATCH] tcp: fix tcp_init_transfer() to not reset icsk_ca_initialized
>
> This commit fixes a bug (found by syzkaller) that could cause spurious
> double-initializations for congestion control modules, which could cause memory
> leaks orother problems for congestion control modules (like CDG) that allocate
> memory in their init functions.
>
> The buggy scenario constructed by syzkaller was something like:
>
> (1) create a TCP socket
> (2) initiate a TFO connect via sendto()
> (3) while socket is in TCP_SYN_SENT, call setsockopt(TCP_CONGESTION),
> which calls:
> tcp_set_congestion_control() ->
> tcp_reinit_congestion_control() ->
> tcp_init_congestion_control()
> (4) receive ACK, connection is established, call tcp_init_transfer(),
> set icsk_ca_initialized=0 (without first calling cc->release()),
> call tcp_init_congestion_control() again.
>
> Note that in this sequence tcp_init_congestion_control() is called twice
> without a cc->release() call in between. Thus, for CC modules that allocate
> memory in their init() function, e.g, CDG, a memory leak may occur. The
> syzkaller tool managed to find a reproducer that triggered such a leak in CDG.
>
> The bug was introduced when that 8919a9b31eb4 commit introduced
> icsk_ca_initialized and set icsk_ca_initialized to 0 in tcp_init_transfer(),
> missing the possibility for a sequence like the one above, where a process
> could call setsockopt(TCP_CONGESTION) in state TCP_SYN_SENT (i.e. after the
> connect() or TFO open sendmsg()), which would call
> tcp_init_congestion_control(). The 8919a9b31eb4 commit did not intend to reset
> any initialization that the user had already explicitly made; it just missed
> the possibility of that particular sequence (which syzkaller managed to find).

Could I use your commit message when I resubmit patch?

Thank you.

\
 
 \ /
  Last update: 2021-06-30 20:27    [W:0.068 / U:0.128 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site