[PATCH] powerpc/64s: Fix hypercall entry clobbering r12 input

[Nicholas Piggin]

On Tue, 18 Jul 2017 10:52:47 -0700
Ram Pai <linuxram at us.ibm.com> wrote:

> On Tue, Jul 18, 2017 at 03:32:44PM +1000, Nicholas Piggin wrote:
> > A previous optimisation incorrectly assumed the PAPR hcall does
> > not use r12, and clobbers it upon entry. In fact it is used as
> > an input. This can result in KVM guests crashing (observed with
> > PR KVM).
> > 
> > Instead of using r12 to save r13, tihs patch saves r13 in ctr.
> > This is more costly, but not as slow as using the SPRG.
> > 
> > Fixes: acd7d8cef0153 ("powerpc/64s: Optimize hypercall/syscall entry")
> > Signed-off-by: Nicholas Piggin <npiggin at gmail.com>
> > 
> > ---
> > One brown paper bag please.
> > 
> >  arch/powerpc/kernel/exceptions-64s.S | 28 ++++++++++++++--------------
> >  1 file changed, 14 insertions(+), 14 deletions(-)
> > 
> > diff --git a/arch/powerpc/kernel/exceptions-64s.S b/arch/powerpc/kernel/exceptions-64s.S
> > index d8b4ceed7a74..c174e7db7594 100644
> > --- a/arch/powerpc/kernel/exceptions-64s.S
> > +++ b/arch/powerpc/kernel/exceptions-64s.S
> > @@ -824,7 +824,7 @@ EXC_COMMON(trap_0b_common, 0xb00, unknown_exception)
> >   * r3 volatile parameter and return value for status
> >   * r4-r10 volatile input and output value
> >   * r11 volatile hypercall number and output value
> > - * r12 volatile
> > + * r12 volatile input and output value
> >   * r13-r31 nonvolatile
> >   * LR nonvolatile
> >   * CTR volatile
> > @@ -834,25 +834,26 @@ EXC_COMMON(trap_0b_common, 0xb00, unknown_exception)
> >   * Other registers nonvolatile
> >   *
> >   * The intersection of volatile registers that don't contain possible
> > - * inputs is: r12, cr0, xer, ctr. We may use these as scratch regs
> > - * upon entry without saving.
> > + * inputs is: cr0, xer, ctr. We may use these as scratch regs upon entry
> > + * without saving, though xer is not a good idea to use, as hardware may
> > + * interpret some bits so it may be costly to change them.
> >   */
> >  #ifdef CONFIG_KVM_BOOK3S_64_HANDLER
> >  	/*
> >  	 * There is a little bit of juggling to get syscall and hcall
> > -	 * working well. Save r10 in ctr to be restored in case it is a
> > -	 * hcall.
> > +	 * working well. Save r13 in ctr to avoid using SPRG scratch
> > +	 * register.
> >  	 *
> >  	 * Userspace syscalls have already saved the PPR, hcalls must save
> >  	 * it before setting HMT_MEDIUM.
> >  	 */
> >  #define SYSCALL_KVMTEST							\
> > -	mr	r12,r13;						\
> > +	mtctr	r13;							\
> >  	GET_PACA(r13);							\
> > -	mtctr	r10;							\
> > +	std	r10,PACA_EXGEN+EX_R10(r13);				\
> >  	KVMTEST_PR(0xc00); /* uses r10, branch to do_kvm_0xc00_system_call */ \
> >  	HMT_MEDIUM;							\
> > -	mr	r9,r12;							\
> > +	mfctr	r9;
> > 
> >  #else
> >  #define SYSCALL_KVMTEST							\
> > @@ -935,8 +936,8 @@ EXC_VIRT_END(system_call, 0x4c00, 0x100)
> >  	 * This is a hcall, so register convention is as above, with these
> >  	 * differences:
> >  	 * r13 = PACA
> > -	 * r12 = orig r13
> > -	 * ctr = orig r10
> > +	 * ctr = orig r13
> > +	 * orig r10 saved in PACA
> >  	 */
> >  TRAMP_KVM_BEGIN(do_kvm_0xc00)
> >  	 /*
> > @@ -944,14 +945,13 @@ TRAMP_KVM_BEGIN(do_kvm_0xc00)
> >  	  * HMT_MEDIUM. That allows the KVM code to save that value into the
> >  	  * guest state (it is the guest's PPR value).
> >  	  */
> > -	OPT_GET_SPR(r0, SPRN_PPR, CPU_FTR_HAS_PPR)
> > +	OPT_GET_SPR(r10, SPRN_PPR, CPU_FTR_HAS_PPR)
> >  	HMT_MEDIUM
> > -	OPT_SAVE_REG_TO_PACA(PACA_EXGEN+EX_PPR, r0, CPU_FTR_HAS_PPR)
> > +	OPT_SAVE_REG_TO_PACA(PACA_EXGEN+EX_PPR, r10, CPU_FTR_HAS_PPR)
> >  	mfctr	r10  
> 
> 	^^^^ is this needed anymore? orig r10 is anyway saved in paca.
> 	     contents-of-ctr is that of orig-r13. So does it
> 	     serve any purpose?

Yes, look below :)

We have to retain the existing KVM interrupt calling convention, so we
have to save orig-r13 into SCRATCH0.

Aside: we could actually special-case the hcall entry/exit code in KVM the
same way that we special-case the syscall in Linux -- taking advantage of
volatile/clobbered registers etc. rather than go via the generic interrupt
entry/exit that saves everything and has to use SPRGs etc.

I'm not up to that point yet, and hcalls hopefully will become less
frequent with radix and xive etc, but it may be something to look at
eventually.

> 
> 
> > -	SET_SCRATCH0(r12)
> > +	SET_SCRATCH0(r10)
> >  	std	r9,PACA_EXGEN+EX_R9(r13)
> >  	mfcr	r9
> > -	std	r10,PACA_EXGEN+EX_R10(r13)
> >  	KVM_HANDLER(PACA_EXGEN, EXC_STD, 0xc00)
> >  #endif
> > 
> > -- 
> > 2.11.0  
>