On Fri Mar 24, 2023 at 12:07 AM AEST, Thomas Huth wrote:
> On 20/03/2023 08.03, Nicholas Piggin wrote:
> > The VPA is a(n optional) memory structure shared between the hypervisor
> > and operating system, defined by PAPR. This test defines the structure
> > and adds registration, deregistration, and a few simple sanity tests.
> >
> > Signed-off-by: Nicholas Piggin <npig...@gmail.com>
> > ---
> > lib/linux/compiler.h | 2 +
> > lib/powerpc/asm/hcall.h | 1 +
> > lib/ppc64/asm/vpa.h | 62 ++++++++++++++++++++++++++++
> > powerpc/Makefile.ppc64 | 2 +-
> > powerpc/spapr_vpa.c | 90 +++++++++++++++++++++++++++++++++++++++++
>
> Please add the new test to powerpc/unittests.cfg, otherwise it won't get
> picked up by the run_tests.sh script.
Ah good point.
> > diff --git a/lib/linux/compiler.h b/lib/linux/compiler.h
> > index 6f565e4..c9d205e 100644
> > --- a/lib/linux/compiler.h
> > +++ b/lib/linux/compiler.h
> > @@ -45,7 +45,9 @@
> >
> > #define barrier() asm volatile("" : : : "memory")
> >
> > +#ifndef __always_inline
> > #define __always_inline inline __attribute__((always_inline))
> > +#endif
>
> What's this change good for? ... it doesn't seem to be related to this patch?
Some header ordering issue I forgot about, thanks for reminding. I think it
should be split it out. See /usr/include/<arch>/sys/cdefs.h:
/* Forces a function to be always inlined. */
#if __GNUC_PREREQ (3,2) || __glibc_has_attribute (__always_inline__)
/* The Linux kernel defines __always_inline in stddef.h (283d7573), and
it conflicts with this definition. Therefore undefine it first to
allow either header to be included first. */
# undef __always_inline
# define __always_inline __inline __attribute__ ((__always_inline__))
#else
# undef __always_inline
# define __always_inline __inline
#endif
> > diff --git a/lib/ppc64/asm/vpa.h b/lib/ppc64/asm/vpa.h
> > new file mode 100644
> > index 0000000..11dde01
> > --- /dev/null
> > +++ b/lib/ppc64/asm/vpa.h
> > @@ -0,0 +1,62 @@
> > +#ifndef _ASMPOWERPC_VPA_H_
> > +#define _ASMPOWERPC_VPA_H_
> > +/*
> > + * This work is licensed under the terms of the GNU LGPL, version 2.
> > + */
> > +
> > +#ifndef __ASSEMBLY__
> > +
> > +struct vpa {
> > + uint32_t descriptor;
> > + uint16_t size;
> > + uint8_t reserved1[3];
> > + uint8_t status;
>
> Where does this status field come from? ... My LoPAPR only says that there
> are 18 "reserved" bytes in total here.
Hmm, I'm not sure why that was left out of LoPAPR, Linux has been using
it for a long time. It basically just tells you if you are on a
dedicated or shared partition (hard partitioned or timesliced CPUs).
Possibly an oversight.
>
> > + uint8_t reserved2[14];
> > + uint32_t fru_node_id;
> > + uint32_t fru_proc_id;
> > + uint8_t reserved3[56];
> > + uint8_t vhpn_change_counters[8];
> > + uint8_t reserved4[80];
> > + uint8_t cede_latency;
> > + uint8_t maintain_ebb;
> > + uint8_t reserved5[6];
> > + uint8_t dtl_enable_mask;
> > + uint8_t dedicated_cpu_donate;
> > + uint8_t maintain_fpr;
> > + uint8_t maintain_pmc;
> > + uint8_t reserved6[28];
> > + uint64_t idle_estimate_purr;
> > + uint8_t reserved7[28];
> > + uint16_t maintain_nr_slb;
> > + uint8_t idle;
> > + uint8_t maintain_vmx;
> > + uint32_t vp_dispatch_count;
> > + uint32_t vp_dispatch_dispersion;
> > + uint64_t vp_fault_count;
> > + uint64_t vp_fault_tb;
> > + uint64_t purr_exprop_idle;
> > + uint64_t spurr_exprop_idle;
> > + uint64_t purr_exprop_busy;
> > + uint64_t spurr_exprop_busy;
> > + uint64_t purr_donate_idle;
> > + uint64_t spurr_donate_idle;
> > + uint64_t purr_donate_busy;
> > + uint64_t spurr_donate_busy;
> > + uint64_t vp_wait3_tb;
> > + uint64_t vp_wait2_tb;
> > + uint64_t vp_wait1_tb;
> > + uint64_t purr_exprop_adjunct_busy;
> > + uint64_t spurr_exprop_adjunct_busy;
>
> The above two fields are also marked as "reserved" in my LoPAPR ... which
> version did you use?
>
> > + uint32_t supervisor_pagein_count;
> > + uint8_t reserved8[4];
> > + uint64_t purr_exprop_adjunct_idle;
> > + uint64_t spurr_exprop_adjunct_idle;
> > + uint64_t adjunct_insns_executed;
>
> dito for the above three lines... I guess my LoPAPR is too old...
Ah, I'm guessing the "adjunct" option isn't relevant to Linux/KVM so it
was probably left out (it's much older than LoPAPR).
Generally LoPAPR is still pretty up to date, but we should do better at
keeping it current IMO. I've made some more noises about that, but
can't make any promises here.
> > + uint8_t reserved9[120];
> > + uint64_t dtl_index;
> > + uint8_t reserved10[96];
> > +};
> > +
> > +#endif /* __ASSEMBLY__ */
> > +
> > +#endif /* _ASMPOWERPC_VPA_H_ */
> > diff --git a/powerpc/Makefile.ppc64 b/powerpc/Makefile.ppc64
> > index ea68447..b0ed2b1 100644
> > --- a/powerpc/Makefile.ppc64
> > +++ b/powerpc/Makefile.ppc64
> > @@ -19,7 +19,7 @@ reloc.o = $(TEST_DIR)/reloc64.o
> > OBJDIRS += lib/ppc64
> >
> > # ppc64 specific tests
> > -tests =
> > +tests = $(TEST_DIR)/spapr_vpa.elf
> >
> > include $(SRCDIR)/$(TEST_DIR)/Makefile.common
> >
> > diff --git a/powerpc/spapr_vpa.c b/powerpc/spapr_vpa.c
> > new file mode 100644
> > index 0000000..45688fe
> > --- /dev/null
> > +++ b/powerpc/spapr_vpa.c
> > @@ -0,0 +1,90 @@
> > +/*
> > + * Test sPAPR hypervisor calls (aka. h-calls)
>
> Adjust to "Test sPAPR H_REGISTER_VPA hypervisor call" ?
Yes.
> > + rc = hcall(H_REGISTER_VPA, 5ULL << 45, cpuid, vpa);
> > + report(rc == H_SUCCESS, "VPA deregistered");
> > +
> > + disp_count1 = be32_to_cpu(vpa->vp_dispatch_count);
> > + report(disp_count1 % 2 == 1, "Dispatch count is odd after deregister");
> > +}
>
> Now that was a very tame amount of tests ;-)
Yeah it was just a start. I was going to add a few more scheduling
type ones if I can improve SMP support as well.
> I'd suggest to add some more:
>
> - Check hcall(H_REGISTER_VPA, 0, ...);
> - Check hcall(H_REGISTER_VPA, ..., bad-cpu-id, ...)
> - Check hcall(H_REGISTER_VPA, ..., ..., unaligned-address)
> - Check hcall(H_REGISTER_VPA, ..., ..., illegal-address)
> - Check registration with vpa->size being too small
> - Check registration where the vpa crosses the 4k boundary
>
> What do you think?
Good idea.
> > +int main(int argc, char **argv)
> > +{
> > + struct vpa *vpa;
> > +
> > + vpa = memalign(4096, sizeof(*vpa));
> > +
> > + memset(vpa, 0, sizeof(*vpa));
> > +
> > + vpa->size = cpu_to_be16(sizeof(*vpa));
> > +
> > + report_prefix_push("vpa");
>
> This lacks the corresponding report_prefix_pop() later.
Got it.
Thanks,
Nick
