[PATCH v4 4/5] target-arm: kvm64 fix save/restore of SPSR regs

[Christoffer Dall]

On Tue, Mar 17, 2015 at 04:18:16PM +0000, Peter Maydell wrote:
> On 16 March 2015 at 11:01, Alex Bennée <alex.bennee at linaro.org> wrote:
> > From: Christoffer Dall <christoffer.dall at linaro.org>
> >
> > The current code was negatively indexing the cpu state array and not
> > synchronizing banked spsr register state with the current mode's spsr
> > state, causing occasional failures with migration.
> >
> > Some munging is done to take care of the aarch64 mapping and also to
> > ensure the most current value of the spsr is updated to the banked
> > registers (relevant for KVM<->TCG migration).
> >
> > Signed-off-by: Christoffer Dall <christoffer.dall at linaro.org>
> > Signed-off-by: Alex Bennée <alex.bennee at linaro.org>
> >
> > ---
> > v2 (ajb)
> >   - minor tweaks and clarifications
> > v3
> >   - Use the correct bank index function for setting/getting env->spsr
> >   - only deal with spsrs in elevated exception levels
> > v4
> >  - try and make commentary clearer
> >  - ensure env->banked_spsr[0] = env->spsr before we sync
> >
> > diff --git a/target-arm/kvm64.c b/target-arm/kvm64.c
> > index 8fd0c8d..7ddb1b1 100644
> > --- a/target-arm/kvm64.c
> > +++ b/target-arm/kvm64.c
> > @@ -140,6 +140,7 @@ int kvm_arch_put_registers(CPUState *cs, int level)
> >      uint64_t val;
> >      int i;
> >      int ret;
> > +    unsigned int el;
> >
> >      ARMCPU *cpu = ARM_CPU(cs);
> >      CPUARMState *env = &cpu->env;
> > @@ -206,9 +207,29 @@ int kvm_arch_put_registers(CPUState *cs, int level)
> >          return ret;
> >      }
> >
> > +    /* Saved Program State Registers
> > +     *
> > +     * Before we restore from the banked_spsr[] array we need to
> > +     * ensure that any modifications to env->spsr are correctly
> > +     * reflected and map aarch64 exception levels if required.
> 
> "AArch64". Not entirely sure what the "map exception levels"
> is saying.
> 
> > +     */
> > +    el = arm_current_el(env);
> > +    if (el > 0) {
> > +        if (is_a64(env)) {
> > +            g_assert(el == 1);
> > +            env->banked_spsr[0] = env->spsr;
> 
> If we're in AArch64 mode then I don't believe
> env->spsr has valid content at this point. The live
> copy is in env->banked_sprsr[] for all cases.
> 

ah, ok, so we can just get rid of that one.

> > +            /* QEMUs AARCH64 EL1 SPSR is in bank 0, so map it to
> > +             * KVM_SPSR_SVC for syncing to KVM */
> 
> "QEMU's" and "AArch64", but this is just a bug in our
> implementation -- the mapping between AArch64 SPSR_EL1
> and AArch32 SPSR_svc is architecturally mandated.
> We should be using banked_spsr[1] for our regdef for
> SPSR_EL1 in helper.c.
> 
> Also the "*/" should be on its own line.
> 
> > +            env->banked_spsr[1] = env->banked_spsr[0];
> 
> What is going on here? Architecturally, AArch64 SPSR_EL1
> is mapped to AArch32 SPSR_svc, which is env->banked_spsr[1].
> env->banked_spsr[0] would be the SPSR for USR and SYS, except
> they don't have an SPSR (you cannot take exceptions into them).
> I think QEMU just has a banked_spsr[0] for convenience of
> implementation and it should never actually be used.
> Is this code just working around the QEMU SPSR_EL1 bug?
> 

no, the idea was just to let the loop below just work.  We accomplish
this by taking the AArch64 spsr (which is stored in banked_spsr[0]) and
copying it into banked_spsr[1], which is not used in AArch64 and we can
do the QEMU's spsr are indexed at kvm-index + 1.



> > +            i = bank_number(env->uncached_cpsr & CPSR_M);
> > +            env->banked_spsr[i] = env->spsr;
> > +        }
> > +    }
> > +
> >      for (i = 0; i < KVM_NR_SPSR; i++) {
> >          reg.id = AARCH64_CORE_REG(spsr[i]);
> > -        reg.addr = (uintptr_t) &env->banked_spsr[i - 1];
> > +        reg.addr = (uintptr_t) &env->banked_spsr[i+1];
> 
> Coding style demands spaces around the "+" operator.
> 
> Note that this code is implicitly relying on the
> ordering of register banks defined by the bank_number()
> function, which is a bit icky.

right, I thought you wrote this code with some deeper intention of doing
it this way so I tried to stick with the general idea - but now I
actually looked at git blame and realized that you didn't even write it.

Given all this churn around this, probably it's much cleaner to get rid
of the loop and have an explicit sync for each architecturally
implemented register, i.e. the SPSR_EL1 and the various mode-specific
AArch32 SPSR registers?

> 
> >          ret = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &reg);
> >          if (ret) {
> >              return ret;
> > @@ -254,6 +275,7 @@ int kvm_arch_get_registers(CPUState *cs)
> >      struct kvm_one_reg reg;
> >      uint64_t val;
> >      uint32_t fpr;
> > +    unsigned int el;
> >      int i;
> >      int ret;
> >
> > @@ -326,15 +348,34 @@ int kvm_arch_get_registers(CPUState *cs)
> >          return ret;
> >      }
> >
> > +    /* Fetch the SPSR registers
> > +     *
> > +     * KVM has an array of state indexed for all the possible aarch32
> > +     * privilege levels. These map onto QEMUs aarch32 banks 1 - 4.
> 
> QEMU's. AArch32. Also, you mean "1 - 5".
> 
> > +     */
> >      for (i = 0; i < KVM_NR_SPSR; i++) {
> >          reg.id = AARCH64_CORE_REG(spsr[i]);
> > -        reg.addr = (uintptr_t) &env->banked_spsr[i - 1];
> > +        reg.addr = (uintptr_t) &env->banked_spsr[i+1];
> 
> Spaces again.
> 
> >          ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &reg);
> >          if (ret) {
> >              return ret;
> >          }
> >      }
> >
> > +    el = arm_current_el(env);
> > +    if (el > 0) {
> > +        if (is_a64(env)) {
> > +            g_assert(el == 1);
> > +            /* KVM_SPSR_SVC holds the AARCH64 EL1 SPSR which QEMU
> > +             * keeps in bank 0 so copy it across. */
> > +            env->banked_spsr[0] = env->banked_spsr[1];
> > +            i = aarch64_banked_spsr_index(el);
> 
> More workarounds for a bug we should just fix, I think.
> 

again, this is just for the loop above to be generic, and then fix
things up afterwards so that things work both for is_a64() and
!is_a64().

Thanks,
-Christoffer