Re: [PATCH v2] sunrpc: fix waitqueue_active without memory barrier in sunrpc

From: J. Bruce Fields
Date: Wed Oct 14 2015 - 12:00:49 EST


On Wed, Oct 14, 2015 at 03:57:13AM +0000, Kosuke Tatsukawa wrote:
> J. Bruce Fields wrote:
> > On Mon, Oct 12, 2015 at 10:41:06AM +0000, Kosuke Tatsukawa wrote:
> >> J. Bruce Fields wrote:
> >> > On Fri, Oct 09, 2015 at 06:29:44AM +0000, Kosuke Tatsukawa wrote:
> >> >> Neil Brown wrote:
> >> >> > Kosuke Tatsukawa <tatsu@xxxxxxxxxxxxx> writes:
> >> >> >
> >> >> >> There are several places in net/sunrpc/svcsock.c which calls
> >> >> >> waitqueue_active() without calling a memory barrier. Add a memory
> >> >> >> barrier just as in wq_has_sleeper().
> >> >> >>
> >> >> >> I found this issue when I was looking through the linux source code
> >> >> >> for places calling waitqueue_active() before wake_up*(), but without
> >> >> >> preceding memory barriers, after sending a patch to fix a similar
> >> >> >> issue in drivers/tty/n_tty.c (Details about the original issue can be
> >> >> >> found here: https://lkml.org/lkml/2015/9/28/849).
> >> >> >
> >> >> > hi,
> >> >> > this feels like the wrong approach to the problem. It requires extra
> >> >> > 'smb_mb's to be spread around which are hard to understand as easy to
> >> >> > forget.
> >> >> >
> >> >> > A quick look seems to suggest that (nearly) every waitqueue_active()
> >> >> > will need an smb_mb. Could we just put the smb_mb() inside
> >> >> > waitqueue_active()??
> >> >> <snip>
> >> >>
> >> >> There are around 200 occurrences of waitqueue_active() in the kernel
> >> >> source, and most of the places which use it before wake_up are either
> >> >> protected by some spin lock, or already has a memory barrier or some
> >> >> kind of atomic operation before it.
> >> >>
> >> >> Simply adding smp_mb() to waitqueue_active() would incur extra cost in
> >> >> many cases and won't be a good idea.
> >> >>
> >> >> Another way to solve this problem is to remove the waitqueue_active(),
> >> >> making the code look like this;
> >> >> if (wq)
> >> >> wake_up_interruptible(wq);
> >> >> This also fixes the problem because the spinlock in the wake_up*() acts
> >> >> as a memory barrier and prevents the code from being reordered by the
> >> >> CPU (and it also makes the resulting code is much simpler).
> >> >
> >> > I might not care which we did, except I don't have the means to test
> >> > this quickly, and I guess this is some of our most frequently called
> >> > code.
> >> >
> >> > I suppose your patch is the most conservative approach, as the
> >> > alternative is a spinlock/unlock in wake_up_interruptible, which I
> >> > assume is necessarily more expensive than an smp_mb().
> >> >
> >> > As far as I can tell it's been this way since forever. (Well, since a
> >> > 2002 patch "NFSD: TCP: rationalise locking in RPC server routines" which
> >> > removed some spinlocks from the data_ready routines.)
> >> >
> >> > I don't understand what the actual race is yet (which code exactly is
> >> > missing the wakeup in this case? nfsd threads seem to instead get
> >> > woken up by the wake_up_process() in svc_xprt_do_enqueue().)
> >>
> >> Thank you for the reply. I tried looking into this.
> >>
> >> The callbacks in net/sunrpc/svcsock.c are set up in svc_tcp_init() and
> >> svc_udp_init(), which are both called from svc_setup_socket().
> >> svc_setup_socket() is called (indirectly) from lockd, nfsd, and nfsv4
> >> callback port related code.
> >>
> >> Maybe I'm wrong, but there might not be any kernel code that is using
> >> the socket's wait queue in this case.
> >
> > As Trond points out there are probably waiters internal to the
> > networking code.
>
> Trond and Bruce, thank you for the comment. I was able to find the call
> to the wait function that was called from nfsd.
>
> sk_stream_wait_connect() and sk_stream_wait_memory() were called from
> either do_tcp_sendpages() or tcp_sendmsg() called from within
> svc_send(). sk_stream_wait_connect() shouldn't be called at this point,
> because the socket has already been used to receive the rpc request.
>
> On the wake_up side, sk_write_space() is called from the following
> locations. The relevant ones seems to be preceded by atomic_sub or a
> memory barrier.
> + ksocknal_write_space [drivers/staging/lustre/lnet/klnds/socklnd/socklnd_lib.c:633]
> + atm_pop_raw [net/atm/raw.c:40]
> + sock_setsockopt [net/core/sock.c:740]
> + sock_wfree [net/core/sock.c:1630]
> Preceded by atomic_sub in sock_wfree()
> + ccid3_hc_tx_packet_recv [net/dccp/ccids/ccid3.c:442]
> + do_tcp_sendpages [net/ipv4/tcp.c:1008]
> + tcp_sendmsg [net/ipv4/tcp.c:1300]
> + do_tcp_setsockopt [net/ipv4/tcp.c:2597]
> + tcp_new_space [net/ipv4/tcp_input.c:4885]
> Preceded by smp_mb__after_atomic in tcp_check_space()
> + llc_conn_state_process [net/llc/llc_conn.c:148]
> + pipe_rcv_status [net/phonet/pep.c:312]
> + pipe_do_rcv [net/phonet/pep.c:440]
> + pipe_start_flow_control [net/phonet/pep.c:554]
> + svc_sock_setbufsize [net/sunrpc/svcsock.c:45]
>
> sk_state_change() calls related to TCP/IP were called from the following
> places.
> + inet_shutdown [net/ipv4/af_inet.c:825]
> This shouldn't be called when waiting
> + tcp_done [net/ipv4/tcp.c:3078]
> spin_lock*/spin_unlock* is called in lock_timer_base
> + tcp_fin [net/ipv4/tcp_input.c:4031]
> atomic_long_sub is called from sk_memory_allocated_sub called within
> sk_mem_reclaim
> + tcp_finish_connect [net/ipv4/tcp_input.c:5415]
> This shoudn't be called when waiting
> + tcp_rcv_state_process [net/ipv4/tcp_input.c:5807,5880]
> The socket shouldn't be in TCP_SYN_RECV nor TCP_FIN_WAIT1 states when
> waiting
>
> I think the wait queue won't be used for being woken up by
> svc_{tcp,udp}_data_ready, because nfsd doesn't read from a socket.

Looking, well, I guess kernel_recvmsg() does read from a socket, but I
assume calling with MSG_DONTWAIT means that it doesn't block.

> So with the current implementation, it seems there shouldn't be any
> problems even if the memory barrier is missing.

Thanks for the detailed investigation.

I think it would be worth adding a comment if that might help someone
having to reinvestigate this again some day.

--b.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/