Re: [PATCH v7 1/4] irqchip: gic: Support hierarchy irq domain.

[Yingjoe Chen <yingjoe.chen-NuS5LvNUpcJWk0Htik3J/w@public.gmane.org>]

Hi,

On Thu, 2014-11-20 at 10:07 +0000, Marc Zyngier wrote:
> On Thu, Nov 20 2014 at  4:26:10 am GMT, Jiang Liu <jiang.liu-VuQAYsv1563Yd54FQh9/CA@public.gmane.org> wrote:
> 
> Hi Jiang,
> 
> > On 2014/11/20 1:18, Marc Zyngier wrote:
> >> Hi Yingjoe,
> >> 
> >> On Wed, Nov 19 2014 at 2:14:08 pm GMT, Yingjoe Chen
> >> <yingjoe.chen-NuS5LvNUpcJWk0Htik3J/w@public.gmane.org> wrote:
> >>> +
> >>> +static const struct irq_domain_ops gic_irq_domain_hierarchy_ops = {
> >>> +	.xlate = gic_irq_domain_xlate,
> >>> +	.alloc = gic_irq_domain_alloc,
> >>> +	.free = irq_domain_free_irqs_top,
> >> 
> >> I'm convinced that irq_domain_free_irqs_top is the wrong function to
> >> call here, because you're calling it from the bottom, not the top-level
> >> (it has no parent).
> >> 
> >> I cannot verify this with your code as I don't a working platform with
> >> GICv2m, but if I enable something similar on GICv3, it dies a very
> >> painful way:
> >> 
> >> Unable to handle kernel NULL pointer dereference at virtual address 00000018
> >> pgd = ffffffc03d059000
> >> [00000018] *pgd=0000000081356003, *pud=0000000081356003, *pmd=0000000000000000
> >> Internal error: Oops: 96000006 [#1] SMP
> >> Modules linked in:
> >> CPU: 4 PID: 1052 Comm: sh Not tainted 3.18.0-rc4+ #3311
> >> task: ffffffc03e320000 ti: ffffffc001390000 task.ti: ffffffc001390000
> >> PC is at irq_domain_free_irqs_recursive+0x1c/0x80
> >> LR is at irq_domain_free_irqs_common+0x88/0x9c
> >> pc : [<ffffffc0000ed790>] lr : [<ffffffc0000ede20>] pstate: 60000145
> >> [...]
> >> [<ffffffc0000ed790>] irq_domain_free_irqs_recursive+0x1c/0x80
> >> [<ffffffc0000ede1c>] irq_domain_free_irqs_common+0x84/0x9c
> >> [<ffffffc0000ede98>] irq_domain_free_irqs_top+0x64/0x7c  <-- gic_domain.free()
> >> [<ffffffc0000ed798>] irq_domain_free_irqs_recursive+0x24/0x80
> >> [<ffffffc0000ee468>] irq_domain_free_irqs_parent+0x14/0x20
> >> [<ffffffc0003500b8>] its_irq_domain_free+0xc8/0x250
> >> [<ffffffc0000ed798>] irq_domain_free_irqs_recursive+0x24/0x80
> >> [<ffffffc0000ede1c>] irq_domain_free_irqs_common+0x84/0x9c
> >> [<ffffffc0000ede98>] irq_domain_free_irqs_top+0x64/0x7c
> >> [<ffffffc0000ef518>] msi_domain_free+0x70/0x88
> >> [<ffffffc0000ed798>] irq_domain_free_irqs_recursive+0x24/0x80
> >> [<ffffffc0000ee3ac>] irq_domain_free_irqs+0x108/0x17c
> >> [<ffffffc0000efb68>] msi_domain_free_irqs+0x28/0x4c
> >> [<ffffffc000369cac>] free_msi_irqs+0xb4/0x1c0
> >> [<ffffffc00036adec>] pci_disable_msix+0x3c/0x4c
> >> [...]
> >> 
> >> and I cannot see how this could work on the standard GIC either.
> >> 
> >> Thomas, Jiang: could you please confirm or infirm my suspicions? My
> >> understanding is that irq_domain_free_irqs_top can only be called from
> >> the top-level domain.
> > Hi Marc,
> > 	It indicates that irq_domain_free_irqs_top() is not a good name.
> > We have:
> > 1) irq_domain_set_hwirq_and_chip() to set irq_chip and chip_data
> > 2) irq_domain_set_info() to set irq_chip, chip_data, flow_handler and
> >    handler_data;
> > 3) irq_domain_reset_irq_data() resets irq_chip and chip_data.
> > 4) irq_domain_free_irqs_common() resets irq_chip, chip_data and calls
> >    parent domain's domain_ops.free() callback.
> > 5) irq_domain_free_irqs_top() resets irq_chip, chip_data, flow handler,
> >    handler_data and call parent domain's domain_ops.free() callback.
> 
> Yes, and this "call parent domain's free callback" is where the problem
> lies. Here, it is called from the innermost domain, with no parent.
> 
> > So there two possible improvements here:
> > 1) Rename irq_domain_free_irqs_top() with better name, any suggestions?
> >    It's named as is because it's always called by the outer-most
> >    irqdomains on x86.
> > 2) Change irq_domain_free_irqs_common() and irq_domain_free_irqs_top()
> >    to call parent domain's domain_ops.free() callback only if parent
> >    exists. By this way, they could be used for inner-most irqdomains.
> > If OK, I will respin a version 4 patch set based on tip/irq/irqdomain.
> > Thoughts?
> 
> Checking the parent is probably a safe solution (this is not a hot path
> anyway). I don't care much about the name though, and I the only thing I
> can think of is irq_domain_free_irqs_reset_flow, which looks so bad it's
> not even funny. I'll let the matter rest in your capable hands! ;-)

I've applied Jiang's "irqdomain: Enhance irq_domain_free_irqs_common()
to support parentless irqdomain" patch and it did fix the crash.

Thanks Jiang, Marc

Joe.C


--
To unsubscribe from this list: send the line "unsubscribe devicetree" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html