Re: [PATCH v8 20/29] s390: Introduce Start Arm Execution instruction

[Steffen Eiden]

On Mon, Sep 28, 2026 at 05:53:33PM +0200, Ilya Leoshkevich wrote:
> 
> 
> On 9/18/26 15:30, Steffen Eiden wrote:
> > The Start Arm Execution (SAE) instruction is the centerpiece for
> > executing arm64 (KVM) guests on s390. Its purpose is, similar to SIE, to
> > enable accelerated execution of arm64 virtual machines. SAE expects the
> > physical address of a control block as the only argument.
> > 
> > The host is responsible to save & restore
> >   - GPRs 0-13
> >   - access register 0-15
>              ^^^^^^^^
>              registers>   - breaking event register (BEAR)
> >   - vector/floating point registers
> > between SAE executions to guarantee host consistency.
> > 
> > GPRs and BEAR are save and restores in the asm functions. The other
>                     ^^^^^^^^^^^^^^^^^
>                     saved and restored
> 
> > register are handled in within C code. Access registers are handled in a
>   ^^^^^^^^             ^^
>   registers            not needed probably?
> 
> > later patch and SVEs will be handled when they are introduced in a
> > future series. Most arm64 registers are handled by a satellite block
> > called save_area. Some registers, frequently used by hypervisors, are
> > placed into the SAE control block itself.
> > 
> > Enlighten asm/kvm_host_types.h for the new header variant. The new
> > header is chosen instead of asm/kvm_host_s390_types.h if KVM_S390_ARM64
> > is defined.
> > 
> > Co-developed-by: Andreas Grapentin <gra@xxxxxxxxxxxxx>
> > Signed-off-by: Andreas Grapentin <gra@xxxxxxxxxxxxx>
> > Co-developed-by: Nina Schoetterl-Glausch <nsg@xxxxxxxxxxxxx>
> > Signed-off-by: Nina Schoetterl-Glausch <nsg@xxxxxxxxxxxxx>
> > Signed-off-by: Steffen Eiden <seiden@xxxxxxxxxxxxx>
> > ---
> >   arch/s390/include/asm/asm-prototypes.h       |   1 +
> >   arch/s390/include/asm/kvm_host_arm64_types.h | 124 +++++++++++++++++++
> >   arch/s390/include/asm/kvm_host_types.h       |   4 +
> >   arch/s390/include/asm/sae.h                  |  58 +++++++++
> >   arch/s390/include/asm/stacktrace.h           |   5 +
> >   arch/s390/kernel/asm-offsets.c               |   1 +
> >   arch/s390/kernel/entry.S                     |  22 ++++
> >   arch/s390/tools/opcodes.txt                  |   3 +
> >   8 files changed, 218 insertions(+)
> >   create mode 100644 arch/s390/include/asm/kvm_host_arm64_types.h
> >   create mode 100644 arch/s390/include/asm/sae.h
> 
> [...]
>    > +#if IS_ENABLED(CONFIG_KVM)
> > +/*
> > + * __sae64a calling convention:
> > + * %r2 pointer to sae control block physical address
>           ^^^^^^^^^^ I guess it's just a physical address and not a
>                      pointer to it?
> 
yes. Fixing.


>                      ^^^ sometimes it's called "sae", sometimes "SAE" -
>                          it would be nice to have some consistency> + */
Thanks, I will convert everything to SAE


> > +SYM_FUNC_START(__sae64a)
> > +	stmg	%r6,%r14,__SF_GPRS(%r15)	# store kernel registers
> 
> Strictly speaking we don't need to save %r14, but I guess you do this
> for unwinding?
yes. Also, some later (unsent) patch needs this and before I forget this, let's
do it now. 

> 
> > +	STBEAR	__SF_SAE_BEAR(%r15)		# save breaking event address register
> > +	.insn	rre,0xb9a50000,%r2,0		# Start Arm Execution
> > +	LBEAR	__SF_SAE_BEAR(%r15)		# restore breaking event address register
> > +	lmg	%r6,%r14,__SF_GPRS(%r15)	# restore kernel registers
> > +	xgr	%r0,%r0				# clear guest registers to
> > +	xgr	%r1,%r1				# prevent speculative use
> > +	xgr	%r2,%r2
> > +	xgr	%r3,%r3
> > +	xgr	%r4,%r4
> > +	xgr	%r5,%r5
> > +	BR_EX	%r14
> > +SYM_FUNC_END(__sae64a)
> > +EXPORT_SYMBOL(__sae64a)
> > +#endif
> [...]