Re: [Xenomai-help] Slow hard drive access in xenomai

[Philippe Gerum <rpm@xenomai.org>]

On Mon, 2010-10-25 at 21:40 +0200, Jan Kiszka wrote:
> Am 25.10.2010 21:03, Peter Pastor wrote:
> > Hey Jan,
> > 
> > I did not apply any ubuntu patch for kernel 2.6.35 (since I do not have
> > one).  Also, good to know that I should not use xenomai patches together
> > with ubuntu patches.
> > 
> > Anyway, the problem occurred as well with the kernel 2.6.35 (see attached
> > dmesg_bad_2.6.35)
> > I also attached the config.
> > 
> 
> ...
> 
> > [ 5751.714643] irq 16: nobody cared (try booting with the "irqpoll" option)
> > [ 5751.714649] Pid: 0, comm: swapper Tainted: P            2.6.35-ipipe-2.5.4-slim #2
> > [ 5751.714653] Call Trace:
> > [ 5751.714655]  <IRQ>  [<ffffffff8108bb56>] __report_bad_irq+0x26/0xa0
> > [ 5751.714668]  [<ffffffff8108bd5c>] note_interrupt+0x18c/0x1d0
> > [ 5751.714672]  [<ffffffff8108c77d>] handle_fasteoi_irq+0xcd/0x100
> > [ 5751.714677]  [<ffffffff8100656d>] handle_irq+0x1d/0x30
> > [ 5751.714681]  [<ffffffff81005a40>] do_IRQ+0x70/0x100
> > [ 5751.714685]  [<ffffffff81092147>] __ipipe_sync_stage+0x207/0x20d
> > [ 5751.714689]  [<ffffffff810059d0>] ? do_IRQ+0x0/0x100
> > [ 5751.714692]  [<ffffffff8109214d>] ? __xirq_end+0x0/0x9c
> > [ 5751.714696]  [<ffffffff810059d0>] ? do_IRQ+0x0/0x100
> > [ 5751.714700]  [<ffffffff810926a3>] __ipipe_walk_pipeline+0x113/0x120
> > [ 5751.714706]  [<ffffffff81024414>] __ipipe_handle_irq+0x124/0x310
> > [ 5751.714708]  [<ffffffff8108bf10>] ? __ipipe_ack_fasteoi_irq+0x0/0x10
> > [ 5751.714712]  [<ffffffff814f78d3>] common_interrupt+0x13/0x2c
> > [ 5751.714713]  <EOI>  [<ffffffff810249d6>] ? __ipipe_halt_root+0x26/0x40
> > [ 5751.714718]  [<ffffffff81061191>] ? atomic_notifier_call_chain+0x11/0x20
> > [ 5751.714722]  [<ffffffff8100cbd5>] default_idle+0x45/0x50
> > [ 5751.714725]  [<ffffffff8100198a>] cpu_idle+0x7a/0xd0
> > [ 5751.714728]  [<ffffffff814f14a1>] start_secondary+0x1c1/0x1c5
> > [ 5751.714730] handlers:
> > [ 5751.714730] [<ffffffff8136ed60>] (usb_hcd_irq+0x0/0xb0)
> > [ 5751.714735] [<ffffffffa00bac30>] (mpt_interrupt+0x0/0xa00 [mptbase])
> > [ 5751.714747] Disabling IRQ #16
> 
> I'm not yet sure, but a first thought: We have a shared fasteoi IRQ
> here, and we are on SMP. Compared to vanilla, the fasteoi flow of ipipe
> looks so much different to me ATM that I tend to believe two cores end
> up having this IRQ queued at the same time. One runs first and handles
> all triggers, the second bails out like above.
> 
> Philippe, we _end_ fasteoi in the ipipe ack path. Do we mask them prior
> to this? What prevents a second IRQ arriving after this early eoi?

All fasteoi handlers are supposed to mask+ack when the pipeline is
enabled, to avoid interrupt storm due to the deferral we may introduce
in the irq delivery. I do see this in the regular ioapic chip
descriptor, but this is lacking with interrupt remap. I guess we could
have a problem with Intel IOMMUs.

> 
> Jan
> 

-- 
Philippe.