|
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index] Re: [Xen-devel] Xen 4.5 random freeze question
I got this strange log:
(XEN) received maintenance interrupt irq=1023
And platform does not hang due to this:
+ hcr = GICH[GICH_HCR];
+ if ( hcr & GICH_HCR_UIE )
+ {
+ GICH[GICH_HCR] &= ~GICH_HCR_UIE;
+ uie_on = 1;
+ }
On Wed, Nov 19, 2014 at 6:50 PM, Stefano Stabellini
<stefano.stabellini@xxxxxxxxxxxxx> wrote:
> On Wed, 19 Nov 2014, Andrii Tseglytskyi wrote:
>> On Wed, Nov 19, 2014 at 6:13 PM, Stefano Stabellini
>> <stefano.stabellini@xxxxxxxxxxxxx> wrote:
>> > On Wed, 19 Nov 2014, Andrii Tseglytskyi wrote:
>> >> On Wed, Nov 19, 2014 at 6:01 PM, Andrii Tseglytskyi
>> >> <andrii.tseglytskyi@xxxxxxxxxxxxxxx> wrote:
>> >> > On Wed, Nov 19, 2014 at 5:41 PM, Stefano Stabellini
>> >> > <stefano.stabellini@xxxxxxxxxxxxx> wrote:
>> >> >> On Wed, 19 Nov 2014, Andrii Tseglytskyi wrote:
>> >> >>> Hi Stefano,
>> >> >>>
>> >> >>> On Wed, Nov 19, 2014 at 4:52 PM, Stefano Stabellini
>> >> >>> <stefano.stabellini@xxxxxxxxxxxxx> wrote:
>> >> >>> > On Wed, 19 Nov 2014, Andrii Tseglytskyi wrote:
>> >> >>> >> Hi Stefano,
>> >> >>> >>
>> >> >>> >> > > if ( !list_empty(¤t->arch.vgic.lr_pending) &&
>> >> >>> >> > > lr_all_full() )
>> >> >>> >> > > - GICH[GICH_HCR] |= GICH_HCR_UIE;
>> >> >>> >> > > + GICH[GICH_HCR] |= GICH_HCR_NPIE;
>> >> >>> >> > > else
>> >> >>> >> > > - GICH[GICH_HCR] &= ~GICH_HCR_UIE;
>> >> >>> >> > > + GICH[GICH_HCR] &= ~GICH_HCR_NPIE;
>> >> >>> >> > >
>> >> >>> >> > > }
>> >> >>> >> >
>> >> >>> >> > Yes, exactly
>> >> >>> >>
>> >> >>> >> I tried, hang still occurs with this change
>> >> >>> >
>> >> >>> > We need to figure out why during the hang you still have all the LRs
>> >> >>> > busy even if you are getting maintenance interrupts that should
>> >> >>> > cause
>> >> >>> > them to be cleared.
>> >> >>> >
>> >> >>>
>> >> >>> I see that I have free LRs during maintenance interrupt
>> >> >>>
>> >> >>> (XEN) gic.c:871:d0v0 maintenance interrupt
>> >> >>> (XEN) GICH_LRs (vcpu 0) mask=0
>> >> >>> (XEN) HW_LR[0]=9a015856
>> >> >>> (XEN) HW_LR[1]=0
>> >> >>> (XEN) HW_LR[2]=0
>> >> >>> (XEN) HW_LR[3]=0
>> >> >>> (XEN) Inflight irq=86 lr=0
>> >> >>> (XEN) Inflight irq=2 lr=255
>> >> >>> (XEN) Pending irq=2
>> >> >>>
>> >> >>> But I see that after I got hang - maintenance interrupts are generated
>> >> >>> continuously. Platform continues printing the same log till reboot.
>> >> >>
>> >> >> Exactly the same log? As in the one above you just pasted?
>> >> >> That is very very suspicious.
>> >> >
>> >> > Yes exactly the same log. And looks like it means that LRs are flushed
>> >> > correctly.
>> >> >
>> >> >>
>> >> >> I am thinking that we are not handling GICH_HCR_UIE correctly and
>> >> >> something we do in Xen, maybe writing to an LR register, might trigger
>> >> >> a
>> >> >> new maintenance interrupt immediately causing an infinite loop.
>> >> >>
>> >> >
>> >> > Yes, this is what I'm thinking about. Taking in account all collected
>> >> > debug info it looks like once LRs are overloaded with SGIs -
>> >> > maintenance interrupt occurs.
>> >> > And then it is not handled properly, and occurs again and again - so
>> >> > platform hangs inside its handler.
>> >> >
>> >> >> Could you please try this patch? It disable GICH_HCR_UIE immediately on
>> >> >> hypervisor entry.
>> >> >>
>> >> >
>> >> > Now trying.
>> >> >
>> >> >>
>> >> >> diff --git a/xen/arch/arm/gic.c b/xen/arch/arm/gic.c
>> >> >> index 4d2a92d..6ae8dc4 100644
>> >> >> --- a/xen/arch/arm/gic.c
>> >> >> +++ b/xen/arch/arm/gic.c
>> >> >> @@ -701,6 +701,8 @@ void gic_clear_lrs(struct vcpu *v)
>> >> >> if ( is_idle_vcpu(v) )
>> >> >> return;
>> >> >>
>> >> >> + GICH[GICH_HCR] &= ~GICH_HCR_UIE;
>> >> >> +
>> >> >> spin_lock_irqsave(&v->arch.vgic.lock, flags);
>> >> >>
>> >> >> while ((i = find_next_bit((const unsigned long *)
>> >> >> &this_cpu(lr_mask),
>> >> >> @@ -821,12 +823,8 @@ void gic_inject(void)
>> >> >>
>> >> >> gic_restore_pending_irqs(current);
>> >> >>
>> >> >> -
>> >> >> if ( !list_empty(¤t->arch.vgic.lr_pending) && lr_all_full()
>> >> >> )
>> >> >> GICH[GICH_HCR] |= GICH_HCR_UIE;
>> >> >> - else
>> >> >> - GICH[GICH_HCR] &= ~GICH_HCR_UIE;
>> >> >> -
>> >> >> }
>> >> >>
>> >> >> static void do_sgi(struct cpu_user_regs *regs, int othercpu, enum
>> >> >> gic_sgi sgi)
>> >> >
>> >>
>> >> Heh - I don't see hangs with this patch :) But also I see that
>> >> maintenance interrupt doesn't occur (and no hang as result)
>> >> Stefano - is this expected?
>> >
>> > No maintenance interrupts at all? That's strange. You should be
>> > receiving them when LRs are full and you still have interrupts pending
>> > to be added to them.
>> >
>> > You could add another printk here to see if you should be receiving
>> > them:
>> >
>> > if ( !list_empty(¤t->arch.vgic.lr_pending) && lr_all_full() )
>> > + {
>> > + gdprintk(XENLOG_DEBUG, "requesting maintenance interrupt\n");
>> > GICH[GICH_HCR] |= GICH_HCR_UIE;
>> > - else
>> > - GICH[GICH_HCR] &= ~GICH_HCR_UIE;
>> > -
>> > + }
>> > }
>> >
>>
>> Requested properly:
>>
>> (XEN) gic.c:756:d0v0 requesting maintenance interrupt
>> (XEN) gic.c:756:d0v0 requesting maintenance interrupt
>> (XEN) gic.c:756:d0v0 requesting maintenance interrupt
>> (XEN) gic.c:756:d0v0 requesting maintenance interrupt
>> (XEN) gic.c:756:d0v0 requesting maintenance interrupt
>> (XEN) gic.c:756:d0v0 requesting maintenance interrupt
>> (XEN) gic.c:756:d0v0 requesting maintenance interrupt
>>
>> But does not occur
>
> OK, let's see what's going on then by printing the irq number of the
> maintenance interrupt:
>
> diff --git a/xen/arch/arm/gic.c b/xen/arch/arm/gic.c
> index 4d2a92d..fed3167 100644
> --- a/xen/arch/arm/gic.c
> +++ b/xen/arch/arm/gic.c
> @@ -55,6 +55,7 @@ static struct {
> static DEFINE_PER_CPU(uint64_t, lr_mask);
>
> static uint8_t nr_lrs;
> +static bool uie_on;
> #define lr_all_full() (this_cpu(lr_mask) == ((1 << nr_lrs) - 1))
>
> /* The GIC mapping of CPU interfaces does not necessarily match the
> @@ -694,6 +695,7 @@ void gic_clear_lrs(struct vcpu *v)
> {
> int i = 0;
> unsigned long flags;
> + unsigned long hcr;
>
> /* The idle domain has no LRs to be cleared. Since gic_restore_state
> * doesn't write any LR registers for the idle domain they could be
> @@ -701,6 +703,13 @@ void gic_clear_lrs(struct vcpu *v)
> if ( is_idle_vcpu(v) )
> return;
>
> + hcr = GICH[GICH_HCR];
> + if ( hcr & GICH_HCR_UIE )
> + {
> + GICH[GICH_HCR] &= ~GICH_HCR_UIE;
> + uie_on = 1;
> + }
> +
> spin_lock_irqsave(&v->arch.vgic.lock, flags);
>
> while ((i = find_next_bit((const unsigned long *) &this_cpu(lr_mask),
> @@ -865,6 +873,11 @@ void gic_interrupt(struct cpu_user_regs *regs, int
> is_fiq)
> intack = GICC[GICC_IAR];
> irq = intack & GICC_IA_IRQ;
>
> + if ( uie_on )
> + {
> + uie_on = 0;
> + printk("received maintenance interrupt irq=%d\n", irq);
> + }
> if ( likely(irq >= 16 && irq < 1021) )
> {
> local_irq_enable();
--
Andrii Tseglytskyi | Embedded Dev
GlobalLogic
www.globallogic.com
_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel
|
![]() |
Lists.xenproject.org is hosted with RackSpace, monitoring our |