Re: ipc-msg broken again on 3.11-rc7?

[Manfred Spraul]

Hi,

[forgot to cc everyone, thus I'll summarize some mails...]
On 09/02/2013 06:58 AM, Vineet Gupta wrote:
> On 08/31/2013 11:20 PM, Linus Torvalds wrote:
> > Vineet, actual patch for what Davidlohr suggests attached. Can you try it?
> >
> >               Linus
> Apologies for late in getting back to this - I was away from my computer for a bit.
>
> Unfortunately, with a quick test, this patch doesn't help.
> FWIW, this is latest mainline (.config attached).
>
> Let me know what diagnostics I can add to help with this.

msgctl08 is a bulk message send/receive test. I had to look at it once 
before, then it was a broken hardware:
https://lkml.org/lkml/2008/6/12/365
This can be ruled out, because it works with 3.10.

msgctl08 uses pairs of threads: one thread does msgsnd(), the other one 
msgrcv().
There is no synchronization, i.e. the msgsnd() can race ahead until the 
kernel buffer is full and then a block with msgrcv() follows or it could 
be pairs of alternating msgsnd()/msgrcv() operations.
No special features are used: each pair of threads has it's own message 
queues, all messages have type=1.

Vineet ran strace - and just before the signal from killing msgctl08, 
there are only msgsnd()/msgrcv() calls.
Vineet:
a) could you run strace tomorrow again, with '-ttt' as an additional 
option? I don't see where exactly it hangs.
b) Could you check that it is not just a performance regression?
    Does ./msgctl08 1000 16 hang, too?

In ipc/msg.c, I haven't seen any obvious reason why it should hang.
The only race I spotted so far is this one:
>       for (;;) {
>                 struct msg_sender s;
>
>                 err = -EACCES;
>                 if (ipcperms(ns, &msq->q_perm, S_IWUGO))
>                         goto out_unlock1;
>
>                 err = security_msg_queue_msgsnd(msq, msg, msgflg);
>                 if (err)
>                         goto out_unlock1;
>
>                 if (msgsz + msq->q_cbytes <= msq->q_qbytes &&
>                                 1 + msq->q_qnum <= msq->q_qbytes) {
>                         break;
>                 }
[snip]
>         if (!pipelined_send(msq, msg)) {
>                 /* no one is waiting for this message, enqueue it */
>                 list_add_tail(&msg->m_list, &msq->q_messages);
>                 msq->q_cbytes += msgsz;
>                 msq->q_qnum++;
>                 atomic_add(msgsz, &ns->msg_bytes);

The access to msq->q_cbytes is not protected. Thus two parallel msgsnd() 
calls could succeed, even if both together brings the queue length above 
the limit.
But it can't explain why 3.11-rc7 hangs: As explained above, msgctl08 
uses one queue for each thread pair.

--
    Manfred

--
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/