On Thu, Jun 04, 2015 at 11:46:59AM +0200, Mason wrote:
> On 03/06/2015 23:54, Felipe Balbi wrote:
>
> > On Wed, Jun 03, 2015 at 02:41:39PM -0700, Stephen Boyd wrote:
> >>>>>> AM43xx, even though it's a single processor A9, it still has TWD and global
> >>>>>> timer. I was doing some profiling with RT v4.0 and latency is 3.5x lower just
> >>>>>> by switching from gptimer to twd/global.
> >>>>>>
> >>>>>> The only problem is that currently, is_smp() check prevents me from using twd
> >>>>>> with AM43xx (that's why it's commented below, for testing purposes).
> >>>>>>
> >>>>>> In the hopes that we can start a, hopefully, small thread around the subject,
> >>>>>> I'm sending this HACK which I used to get TWD and global timer enabled so I
> >>>>>> could measure latencies with cyclictest.
> >>>>>>
> >>>>>> Is it so that TWD shouldn't be available on UP integrations of ARM's Cortex-A
> >>>>>> processors ?
> >>>>>>
> >>>>>>
> >>>>> I wondered about this recently when looking at something unrelated
> >>>>> and noticed that the check had been introduced as part of
> >>>>> 904464b91eca8 ("ARM: 7655/1: smp_twd: make twd_local_timer_of_register()
> >>>>> no-op for nosmp").
> >>>>>
> >>>>> I suspect this was just the wrong fix at the time, and that the
> >>>>> real culprit is either alloc_percpu() or request_percpu_irq()
> >>>>> getting called too early on a machine without SMP support.
> >>>>>
> >>>>> Possibly the problem is already resolved independently, if you
> >>>>> didn't run into it.
> >>>> no, no splats, nothing at all. See [1]
> >>>>
> >>>> [1] http://hastebin.com/helekubutu
> >>> Adding Shawn
> >>>
> >>
> >> Mason was also interested in doing this. See [2]. From what I could tell
> >> back then, commit 904464b91eca8 was working around the local timer APIs that
> >> no longer exist.
> >>
> >> [2] http://thread.gmane.org/gmane.linux.ports.arm.kernel/389931/focus=392348
> >
> > A lot of good information on that thread, thanks. Seems like getting
> > twd/global timer working would also have some effect on context
> > switching, perhaps ?
>
> Hello,
>
> In my case, I need to support two platforms:
>
> single core Cortex A9 MPCore
> dual core Cortex A9 MPCore
>
> However, as the MPCore moniker implies, even the single core platform
> is "SMP capable". (I think this only means an SCU is available?)
>
> Thus, I worked around the issue by using the same SMP kernel for both
> platforms; which is why I didn't push any patch.
>
> Also, check /proc/timer_list for a "Broadcast device". If you don't
> define one, the TWD timers are set to periodic mode, with hrtimers
> disabled.
Yeah, I have a broadcast device however Linux is picking a gp timer
instead of A9's global timer. Now I think I managed to get Linux to
choose global, but my device won't boot anymore :-p Debugging that. I'm
speculating global timer IRQs aren't firing or aren't wired up properly
on this particular SoC, still to confirm.
--
balbi
Attachment:
signature.asc
Description: Digital signature
