Date:   Tue, 25 Jun 2019 17:59:55 -0400
From:   Willem de Bruijn <willemdebruijn.kernel@...il.com>
To:     Neil Horman <nhorman@...driver.com>
Cc:     Network Development <netdev@...r.kernel.org>,
        Matteo Croce <mcroce@...hat.com>,
        "David S. Miller" <davem@...emloft.net>
Subject: Re: [PATCH v3 net] af_packet: Block execution of tasks waiting for
 transmit to complete in AF_PACKET

On Tue, Jun 25, 2019 at 12:20 PM Neil Horman <nhorman@...driver.com> wrote:
>
> On Tue, Jun 25, 2019 at 09:37:17AM -0400, Willem de Bruijn wrote:
> > On Tue, Jun 25, 2019 at 7:03 AM Neil Horman <nhorman@...driver.com> wrote:
> > >
> > > On Mon, Jun 24, 2019 at 06:15:29PM -0400, Willem de Bruijn wrote:
> > > > > > > +               if (need_wait && !packet_next_frame(po, &po->tx_ring, TP_STATUS_SEND_REQUEST)) {
> > > > > > > +                       po->wait_on_complete = 1;
> > > > > > > +                       timeo = sock_sndtimeo(&po->sk, msg->msg_flags & MSG_DONTWAIT);
> > > > > >
> > > > > > This resets timeout on every loop. should only set above the loop once.
> > > > > >
> > > > > I explained exactly why I did that in the change log.  Its because I reuse the
> > > > > timeout variable to get the return value of the wait_for_complete call.
> > > > > Otherwise I need to add additional data to the stack, which I don't want to do.
> > > > > Sock_sndtimeo is an inline function and really doesn't add any overhead to this
> > > > > path, so I see no reason not to reuse the variable.
> > > >
> > > > The issue isn't the reuse. It is that timeo is reset to sk_sndtimeo
> > > > each time. Whereas wait_for_common and its variants return the
> > > > number of jiffies left in case the loop needs to sleep again later.
> > > >
> > > > Reading sock_sndtimeo once and passing it to wait_.. repeatedly is a
> > > > common pattern across the stack.
> > > >
> > > But those patterns are unique to those situations.  For instance, in
> > > tcp_sendmsg_locked, we aquire the value of the socket timeout, and use that to
> > > wait for the entire message send operation to complete, which consists of
> > > potentially several blocking operations (waiting for the tcp connection to be
> > > established, waiting for socket memory, etc).  In that situation we want to wait
> > > for all of those operations to complete to send a single message, and fail if
> > > they exceed the timeout in aggregate.  The semantics are different with
> > > AF_PACKET.  In this use case, the message is in effect empty, and just used to
> > > pass some control information.  tpacket_snd, sends as many frames from the
> > > memory mapped buffer as possible, and on each iteration we want to wait for the
> > > specified timeout for those frames to complete sending.  I think resetting the
> > > timeout on each wait instance is the right way to go here.
> >
> > I disagree. If a SO_SNDTIMEO of a given time is set, the thread should
> > not wait beyond that. Else what is the point of passing a specific
> > duration in the syscall?
> >
> I can appreciate that, but you said yourself that you wanted to minimize this
> change.  The current behavior of AF_PACKET is to:
> a) check for their being no more packets ready to send
> b) if (a) is false we schedule() (nominally allowing other tasks to run)
> c) when (b) is complete, check for additional frames to send, and exit if there
> are not, otherwise, continue sending and waiting
>
> In that model, a single task calling sendmsg can run in the kernel indefinately,
> if userspace continues to fill the memory mapped buffer
>
> Given that model, resetting the timeout on each call to wait_for_completion
> keeps that behavior.  In fact, if we allow the timeout value to get continuously
> decremented, it will be possible for a call to sendmsg in this protocol to
> _always_ return -ETIMEDOUT (if we keep the buffer full in userspace long enough,
> then the sending task will eventually decrement timeo to zero, and force an
> -ETIMEDOUT call).
>
> I'm convinced that resetting the timeout here is the right way to go.

It upholds the contract, but extends when userspace appends to the
ring. Okay, yes, that makes sense.

> > Btw, we can always drop the timeo and go back to unconditional (bar
> > signals) waiting.
> >
> We could, but it would be nice to implement the timeout feature here if
> possible.
>
> > >
> > > > > > > @@ -2728,6 +2755,11 @@ static int tpacket_snd(struct packet_sock *po, struct msghdr *msg)
> > > > > > >                         err = net_xmit_errno(err);
> > > > > > >                         if (err && __packet_get_status(po, ph) ==
> > > > > > >                                    TP_STATUS_AVAILABLE) {
> > > > > > > +                               /* re-init completion queue to avoid subsequent fallthrough
> > > > > > > +                                * on a future thread calling wait_on_complete_interruptible_timeout
> > > > > > > +                                */
> > > > > > > +                               po->wait_on_complete = 0;
> > > > > >
> > > > > > If setting where sleeping, no need for resetting if a failure happens
> > > > > > between those blocks.
> > > > > >
> > > > > > > +                               init_completion(&po->skb_completion);
> > > > > >
> > > > > > no need to reinit between each use?
> > > > > >
> > > > > I explained exactly why I did this in the comment above.  We have to set
> > > > > wait_for_complete prior to calling transmit, so as to ensure that we call
> > > > > wait_for_completion before we exit the loop. However, in this error case, we
> > > > > exit the loop prior to calling wait_for_complete, so we need to reset the
> > > > > completion variable and the wait_for_complete flag.  Otherwise we will be in a
> > > > > case where, on the next entrace to this loop we will have a completion variable
> > > > > with completion->done > 0, meaning the next wait will be a fall through case,
> > > > > which we don't want.
> > > >
> > > > By moving back to the point where schedule() is called, hopefully this
> > > > complexity automatically goes away. Same as my comment to the line
> > > > immediately above.
> > > >
> > > Its going to change what the complexity is, actually.  I was looking at this
> > > last night, and I realized that your assertion that we could remove
> > > packet_next_frame came at a cost.  This is because we need to determine if we
> > > have to wait before we call po->xmit, but we need to actually do the wait after
> > > po->xmit
> >
> > Why? The existing method using schedule() doesn't.
> >
> Because the existing method using schedule doesn't have to rely on an
> asynchronous event to wake the sending task back up.  Specifically, we need to
> set some state that indicates tpacket_destruct_skb should call complete, and
> since tpacket_destruct_skb is called asynchronously in a separate execution
> context, we need to ensure there is no race condition whereby we execute
> tpacket_destruct_skb before we set the state that we also use to determine if we
> should call wait_for_complete.  ie:
> 1) tpacket_send_skb calls po->xmit
> 2) tpacket_send_skb loops around and checks to see if there are more packets to
> send.  If not and if need_wait is set we call wait_for_complete
>
> If tpacket_destruct_skb is called after (2), we're fine.  But if its called
> between (1) and (2), then tpacket_destruct_skb won't call complete (because
> wait_on_complete isn't set yet), causing a hang.
>
> Because you wanted to remove packet_next_frame, we have no way to determine if
> we're done sending frames prior to calling po->xmit, which is the point past
> which tpacket_destruct_skb might be called, so now I have to find an alternate
> state condition to key off of.
>
>
> > Can we not just loop while sending and sleep immediately when
> > __packet_get_status returns !TP_STATUS_AVAILABLE?
> >
> > I don't understand the need to probe the next packet to send instead
> > of the current.
> >
> See above, we can definately check the return of ph at the top of the loop, and
> sleep if its NULL, but in so doing we cant use po->wait_on_complete as a gating
> variable because we've already called po->xmit.  Once we call that function, if
> we haven't already set po->wait_on_complete to true, its possible
> tpacket_destruct_skb will already have been called before we get to the point
> where we check wait_for_completion.  That means we will wait on a completion
> variable that never gets completed, so I need to find a new way to track weather
> or not we are waiting.  Its entirely doable, I'm sure, its just not as straight
> forward as your making it out to be.
>
> > This seems to be the crux of the disagreement. My guess is that it has
> > to do with setting wait_on_complete, but I don't see the problem. It
> > can be set immediately before going to sleep.
> >
> The reason has to do with the fact that
> tpacket_destruct_skb can be called from ksoftirq context (i.e. it will be called
> asynchrnously, not from the thread running in tpacket_snd).  Condsider this
> flow, assuming we do as you suggest, and just set po->wait_on_complete=1
> immediately prior to calling wait_for_complete_interruptible_timeout:
>
> 1) tpacket_snd gets a frame, builds the skb for transmission
> 2) tpakcket_snd calls po->xmit(skb), sending the frame to the driver
> 3) tpacket_snd, iterates through back to the top of the loop, and determines
> that we need to wait for the previous packet to complete
> 4) The driver gets a tx completion interrupt, interrupting the cpu on which we
> are executing, and calls kfree_skb_irq on the skb we just transmitted
> 5) the ksoftirq runs on the cpu, calling kfree_skb on our skb
> 6) (5) calls skb->destruct (which is tpakcet_skb_destruct)
> 7) tpacket_skb_destruct executes, sees that po->wait_on_completion is zero, and
> skips calling complete()
> 8) the thread running tpacket_snd gets scheduled back on the cpu and now sets
> po->wait_on_complete, then calls wait_for_completion_interruptible_timeout, but
> since the skb we are waiting to complete has already been freed, we will never
> get the completion notification, and so we will wait for the maximal timeout,
> which is absolutely not what we want.
>
> Interestingly (Ironically), that control flow will never happen in the use case
> I'm trying to fix here, because its SCHED_FIFO, and will run until such time as
> we call wait_for_completion_interuptible_timeout, so in this case we're safe.
> But in the nominal case, where the sending task is acturally SCHED_OTHER, the
> above can aboslutely happen, and thats very broken.
>
> > I don't meant to draw this out, btw, or add to your workload. If you
> > prefer, I can take a stab at my (admittedly hand-wavy) suggestion.
> >
> No, I have another method in flight that I'm testing now, it removes the
> po->wait_on_complete variable entirely, checking instead to make sure that
> we've:
> a) sent at least one frame
> and
> b) that we have a positive pending count
> and
> c) that we need to wait
> and
> d) that we have no more frames to send
>
> This is why I was saying that your idea can be done, but it trades one form of
> complexity for another.

Thanks for the detailed explanation of the races.

I was thinking this through, but just see V4. Will take a look at that.

Powered by blists - more mailing lists