Re: [PATCH v2 3/5] ALSA: xen-front: Implement Xen event channel handling

[Takashi Iwai <tiwai@suse.de>]

On Wed, 25 Apr 2018 10:26:34 +0200,
Oleksandr Andrushchenko wrote:
> 
> On 04/24/2018 07:23 PM, Oleksandr Andrushchenko wrote:
> > On 04/24/2018 06:02 PM, Takashi Iwai wrote:
> >> On Tue, 24 Apr 2018 16:58:43 +0200,
> >> Oleksandr Andrushchenko wrote:
> >>> On 04/24/2018 05:35 PM, Takashi Iwai wrote:
> >>>> On Tue, 24 Apr 2018 16:29:15 +0200,
> >>>> Oleksandr Andrushchenko wrote:
> >>>>> On 04/24/2018 05:20 PM, Takashi Iwai wrote:
> >>>>>> On Mon, 16 Apr 2018 08:24:51 +0200,
> >>>>>> Oleksandr Andrushchenko wrote:
> >>>>>>> +static irqreturn_t evtchnl_interrupt_req(int irq, void *dev_id)
> >>>>>>> +{
> >>>>>>> +    struct xen_snd_front_evtchnl *channel = dev_id;
> >>>>>>> +    struct xen_snd_front_info *front_info = channel->front_info;
> >>>>>>> +    struct xensnd_resp *resp;
> >>>>>>> +    RING_IDX i, rp;
> >>>>>>> +    unsigned long flags;
> >>>>>>> +
> >>>>>>> +    if (unlikely(channel->state != EVTCHNL_STATE_CONNECTED))
> >>>>>>> +        return IRQ_HANDLED;
> >>>>>>> +
> >>>>>>> +    spin_lock_irqsave(&front_info->io_lock, flags);
> >>>>>>> +
> >>>>>>> +again:
> >>>>>>> +    rp = channel->u.req.ring.sring->rsp_prod;
> >>>>>>> +    /* ensure we see queued responses up to rp */
> >>>>>>> +    rmb();
> >>>>>>> +
> >>>>>>> +    for (i = channel->u.req.ring.rsp_cons; i != rp; i++) {
> >>>>>> I'm not familiar with Xen stuff in general, but through a quick
> >>>>>> glance, this kind of code worries me a bit.
> >>>>>>
> >>>>>> If channel->u.req.ring.rsp_cons has a bogus number, this may
> >>>>>> lead to a
> >>>>>> very long loop, no?  Better to have a sanity check of the ring
> >>>>>> buffer
> >>>>>> size.
> >>>>> In this loop I have:
> >>>>> resp = RING_GET_RESPONSE(&channel->u.req.ring, i);
> >>>>> and the RING_GET_RESPONSE macro is designed in the way that
> >>>>> it wraps around when *i* in the question gets bigger than
> >>>>> the ring size:
> >>>>>
> >>>>> #define RING_GET_REQUEST(_r, _idx)                    \
> >>>>>       (&((_r)->sring->ring[((_idx) & (RING_SIZE(_r) - 1))].req))
> >>>>>
> >>>>> So, even if the counter has a bogus number it will not last long
> >>>> Hm, this prevents from accessing outside the ring buffer, but does it
> >>>> change the loop behavior?
> >>> no, it doesn't
> >>>> Suppose channel->u.req.ring_rsp_cons = 1, and rp = 0, the loop below
> >>>> would still consume the whole 32bit counts, no?
> >>>>
> >>>>     for (i = channel->u.req.ring.rsp_cons; i != rp; i++) {
> >>>>         resp = RING_GET_RESPONSE(&channel->u.req.ring, i);
> >>>>         ...
> >>>>     }
> >>> You are right here and the comment is totally valid.
> >>> I'll put an additional check like here [1] and here [2]
> >>> Will this address your comment?
> >> Yep, this kind of sanity checks should work.
> >>
> > Great, will implement the checks this way then
> Well, after thinking a bit more on that and chatting on #xendevel IRC
> with Juergen (he is on CC list), it seems that the way the code is now
> it is all fine without the checks: the assumption here is that
> the backend is trusted to always write sane values to the ring counters,
> thus no overflow checks on frontend side are required.
> Even if I implement the checks then I have no means to recover, but
> just print
> an error message and bail out not handling any responses.
> This is probably why the checks [1] and [2] are only implemented for the
> backend side and there are no such macros for the frontend side.
> 
> Takashi, please let me know if the above sounds reasonable and
> addresses your comments.

If it's guaranteed to work, that's OK.
But maybe it's worth to comment for readers.


thanks,

Takashi