Re: [Qemu-devel] [PATCH v2 2/4] spapr/rtas: disable the decrementer interrupt when a CPU is unplugged

[David Gibson <david@gibson.dropbear.id.au>]

[-- Attachment #1: Type: text/plain, Size: 7406 bytes --]

On Wed, Oct 11, 2017 at 01:55:20PM +0200, Cédric Le Goater wrote:
> On 10/11/2017 08:45 AM, David Gibson wrote:
> > On Mon, Oct 09, 2017 at 05:49:28PM +0200, Cédric Le Goater wrote:
> >> When a CPU is stopped with the 'stop-self' RTAS call, its state
> >> 'halted' is switched to 1 and, in this case, the MSR is not taken into
> >> account anymore in the cpu_has_work() routine. Only the pending
> >> hardware interrupts are checked with their LPCR:PECE* enablement bit.
> >>
> >> If the DECR timer fires after 'stop-self' is called and before the CPU
> >> 'stop' state is reached, the nearly-dead CPU will have some work to do
> >> and the guest will crash. This case happens very frequently with the
> >> not yet upstream P9 XIVE exploitation mode. In XICS mode, the DECR is
> >> occasionally fired but after 'stop' state, so no work is to be done
> >> and the guest survives.
> >>
> >> I suspect there is a race between the QEMU mainloop triggering the
> >> timers and the TCG CPU thread but I could not quite identify the root
> >> cause. To be safe, let's disable the decrementer interrupt in the LPCR
> >> when the CPU is halted and reenable it when the CPU is restarted.
> >>
> >> Signed-off-by: Cédric Le Goater <clg@kaod.org>
> >> ---
> >>
> >> Changes in v2:
> >>
> >>  - used a new routine ppc_cpu_pvr_match() to discriminate CPU versions
> >>  - removed the LPCR:PECE* enablement bit when the CPU is initialized
> >>    if it is a secondary
> >>
> >>  hw/ppc/spapr_rtas.c         | 20 ++++++++++++++++++++
> >>  target/ppc/translate_init.c | 19 +++++++++++++++++--
> >>  2 files changed, 37 insertions(+), 2 deletions(-)
> >>
> >> diff --git a/hw/ppc/spapr_rtas.c b/hw/ppc/spapr_rtas.c
> >> index cdf0b607a0a0..dfdbf1e2c6f8 100644
> >> --- a/hw/ppc/spapr_rtas.c
> >> +++ b/hw/ppc/spapr_rtas.c
> >> @@ -46,6 +46,7 @@
> >>  #include "qemu/cutils.h"
> >>  #include "trace.h"
> >>  #include "hw/ppc/fdt.h"
> >> +#include "target/ppc/cpu-models.h"
> >>  
> >>  static void rtas_display_character(PowerPCCPU *cpu, sPAPRMachineState *spapr,
> >>                                     uint32_t token, uint32_t nargs,
> >> @@ -174,6 +175,15 @@ static void rtas_start_cpu(PowerPCCPU *cpu_, sPAPRMachineState *spapr,
> >>          kvm_cpu_synchronize_state(cs);
> >>  
> >>          env->msr = (1ULL << MSR_SF) | (1ULL << MSR_ME);
> >> +
> >> +        /* Enable DECR interrupt */
> >> +        if (ppc_cpu_pvr_match(cpu, CPU_POWERPC_LOGICAL_3_00)) {
> > 
> > Sorry, I didn't reply to your earlier mail in time.  Going via the PVR
> > in this way seems bonkers to me - I like it even less than checking
> > the mmu type.  After all, classifying a bunch of precise models (PVRs)
> > together by behaviour is kind of exactly what the CPU classes are for,
> > so using object_dynamic_case() (==instance_of) is a better idea here.
> 
> hmm, and which type should I use ? we don't have any TYPE_POWER9* we 
> could use for a object_dynamic_cast(). I don't think so ? I could use 
> the name and strcmp("power9") probably but it looks ugly.

Actually there is, but, yeah, it's a lot less obvious than I thought.
It's constructed by the POWERPC_FAILY macro and will be
"POWER9-family-powerpc64-cpu"

> The only thing we have is "CPU_POWERPC_POWER9_BASE" and it only 
> applicates to PVR.
> 
> May be I don't understand your idea.

Urgh, sorry.  This got much muckier than I thought it would be.  I
think maybe it's best to go back to the mmu type test, and later on we
can fix up both the previously existing test like that, and the new
one to something better.

> >> +            env->spr[SPR_LPCR] |= LPCR_DEE;
> >> +        } else {
> >> +            /* P7 and P8 both have same bit for DECR */
> >> +            env->spr[SPR_LPCR] |= LPCR_P8_PECE3;
> >> +        }
> >> +
> >>          env->nip = start;
> >>          env->gpr[3] = r3;
> >>          cs->halted = 0;
> > 
> > The other option I'm wondering about here is to actually add a
> > "shutdown" (or something) method to the cpu class, which does whatever
> > is necessary to put the vcpu into a quiescent state that won't be
> > woken up unless it's specifically requested.
> 
> yes. That is a good idea. 
> 
> Thanks,
> 
> C. 
> 
> 
> >> @@ -210,6 +220,16 @@ static void rtas_stop_self(PowerPCCPU *cpu, sPAPRMachineState *spapr,
> >>       * no need to bother with specific bits, we just clear it.
> >>       */
> >>      env->msr = 0;
> >> +
> >> +    /* Don't let the decremeter run on a CPU being stopped. This could
> >> +     * deliver an interrupt on a dying CPU and crash the guest.
> >> +     */
> >> +    if (ppc_cpu_pvr_match(cpu, CPU_POWERPC_LOGICAL_3_00)) {
> >> +        env->spr[SPR_LPCR] &= ~LPCR_DEE;
> >> +    } else {
> >> +        /* P7 and P8 both have same bit for DECR */
> >> +        env->spr[SPR_LPCR] &= ~LPCR_P8_PECE3;
> >> +    }
> >>  }
> >>  
> >>  static inline int sysparm_st(target_ulong addr, target_ulong len,
> >> diff --git a/target/ppc/translate_init.c b/target/ppc/translate_init.c
> >> index 0d6379fcc5b4..1a62159843e7 100644
> >> --- a/target/ppc/translate_init.c
> >> +++ b/target/ppc/translate_init.c
> >> @@ -8905,6 +8905,7 @@ void cpu_ppc_set_papr(PowerPCCPU *cpu, PPCVirtualHypervisor *vhyp)
> >>      CPUPPCState *env = &cpu->env;
> >>      ppc_spr_t *lpcr = &env->spr_cb[SPR_LPCR];
> >>      ppc_spr_t *amor = &env->spr_cb[SPR_AMOR];
> >> +    CPUState *cs = CPU(cpu);
> >>  
> >>      cpu->vhyp = vhyp;
> >>  
> >> @@ -8946,8 +8947,15 @@ void cpu_ppc_set_papr(PowerPCCPU *cpu, PPCVirtualHypervisor *vhyp)
> >>          } else {
> >>              lpcr->default_value &= ~(LPCR_UPRT | LPCR_GTSE);
> >>          }
> >> -        lpcr->default_value |= LPCR_PDEE | LPCR_HDEE | LPCR_EEE | LPCR_DEE |
> >> +        lpcr->default_value |= LPCR_PDEE | LPCR_HDEE | LPCR_EEE |
> >>                                 LPCR_OEE;
> > 
> > But I guess we'd also need a "set_papr" method to go with that.
> > 
> >> +
> >> +        /* Only let the decremeter wake up the boot CPU. The RTAS
> >> +         * command start-cpu will enable it on secondaries.
> >> +         */
> >> +        if (cs == first_cpu) {
> >> +            lpcr->default_value |= LPCR_DEE;
> >> +        }
> >>          break;
> >>      default:
> >>          /* P7 and P8 has slightly different PECE bits, mostly because P8 adds
> >> @@ -8955,7 +8963,14 @@ void cpu_ppc_set_papr(PowerPCCPU *cpu, PPCVirtualHypervisor *vhyp)
> >>           * will work as expected for both implementations
> >>           */
> >>          lpcr->default_value |= LPCR_P8_PECE0 | LPCR_P8_PECE1 | LPCR_P8_PECE2 |
> >> -                               LPCR_P8_PECE3 | LPCR_P8_PECE4;
> >> +                               LPCR_P8_PECE4;
> >> +
> >> +        /* Only let the decremeter wake up the boot CPU. The RTAS
> >> +         * command start-cpu will enable it on secondaries.
> >> +         */
> >> +        if (cs == first_cpu) {
> >> +            lpcr->default_value |= LPCR_P8_PECE3;
> >> +        }
> >>      }
> >>  
> >>      /* We should be followed by a CPU reset but update the active value
> > 
> 

-- 
David Gibson			| I'll have my music baroque, and my code
david AT gibson.dropbear.id.au	| minimalist, thank you.  NOT _the_ _other_
				| _way_ _around_!
http://www.ozlabs.org/~dgibson

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 833 bytes --]