On Tue, Jun 11, 2024 at 11:00 PM Mathias Nyman
<mathias.nyman@xxxxxxxxxxxxxxx> wrote:
>
> On 11.6.2024 17.45, joswang wrote:
> > On Tue, Jun 11, 2024 at 6:01 PM Mathias Nyman
> > <mathias.nyman@xxxxxxxxxxxxxxx> wrote:
> >>
> >> On 6.6.2024 17.08, joswang wrote:
> >>> On Wed, Jun 5, 2024 at 6:31 PM Mathias Nyman
> >>> <mathias.nyman@xxxxxxxxxxxxxxx> wrote:
> >> ...
> >>>>
> >>>> I was thinking of turning XHCI_AVOID_BEI behavior into the new default, so no
> >>>> quirk flag would be needed:
> >>>>
> >>>> Currently without the quirk flag:
> >>>>
> >>>> - ISOC TRBs trigger interrupt if TRB is the last in the TD
> >>>>
> >>>> Currently with XHCI_AVOID_BEI quirk flag:
> >>>>
> >>>> - ISOC TRBs trigger interrupt if TRB is the last in the TD
> >>>> - Interrupt is additionally triggered every 32 isoc TRB (initially).
> >>>> - if more than 128 events are processed in one interrupt then the
> >>>> 32 is halved, and we trigger an interrupts every 16th isoc TRB, and so
> >>>> on, 16 -> 8...
> >>>>
> >>>> I would remove the quirk flag, and make all controllers interrupt
> >>>> behave as if it was set. i.e. interrupt at least every 32 isoc TRB
> >>>
> >>> Thank you for your detailed analysis.
> >>> Excuse me, I have a question, do you mean to set "Currently with
> >>> XHCI_AVOID_BEI quirk flag" as the default behavior?
> >>
> >> Yes, unless it causes some issues or there are strong objections
> >>
> >>>>
> >>>> Is there an actual real world case where interrupting every 32nd ISOC TRB is
> >>>> too often?
> >>>
> >>> I mean that if the XHCI_AVOID_BEI quirk flag is set, an interrupt will
> >>> be triggered every 8 TRBs, which makes the interrupts seem to be quite
> >>> frequent.
> >>> Thanks
> >>> Jos
> >>>
> >>
> >> It should start with interrupting every 32nd isoc TD, not 8th
> >>
> >> #define AVOID_BEI_INTERVAL_MAX 32
> >>
> >> ir->isoc_bei_interval = AVOID_BEI_INTERVAL_MAX;
> >>
> >> Thanks
> >> Mathias
> >
> > As you mentioned，if more than 128 events are processed in one
> > interrupt then the 32 is halved, and we trigger an interrupts every
> > 16th isoc TRB, and so on, 16 -> 8...
> > xhci_handle_events()
> > ......
> > if (event_loop++ > TRBS_PER_SEGMENT / 2) {
> > if (ir->isoc_bei_interval > AVOID_BEI_INTERVAL_MIN)
> > ir->isoc_bei_interval = ir->isoc_bei_interval / 2;
> > }
> > .......
>
> Yes, and to me this seems reasonable .
> If there are over 128 events per interrupt then we should interrupt more often.
>
> Would this work as the default for the Synopsys hosts you are working with?
>
> Thanks
> Mathias
Without enabling XHCI_AVOID_BEI quirk, that is, triggering an
interrupt once for every 32TRB:
For some USB cameras (such as high-resolution cameras, etc.), the data
volume of one frame is relatively large, and the host controller will
frequently trigger interrupts during data transmission. For some SOC
platforms (such as rockchip platforms, which also use Synopsys USB IP)
with poor performance, it is possible that the CPU cannot respond to
interrupts in time, resulting in frame loss.
Therefore, I think that using XHCI_AVOID_BEI quirk by default may
aggravate the above problem. Please help evaluate, thank you.
Thanks,
Jos Wang
