Re: [PATCH v7 10/21] RISC-V: KVM: Handle MMIO exits for VCPU

From: Anup Patel
Date: Wed Oct 09 2019 - 00:58:46 EST


On Wed, Oct 9, 2019 at 4:14 AM Palmer Dabbelt <palmer@xxxxxxxxxx> wrote:
>
> On Mon, 23 Sep 2019 04:12:17 PDT (-0700), pbonzini@xxxxxxxxxx wrote:
> > On 04/09/19 18:15, Anup Patel wrote:
> >> + unsigned long guest_sstatus =
> >> + vcpu->arch.guest_context.sstatus | SR_MXR;
> >> + unsigned long guest_hstatus =
> >> + vcpu->arch.guest_context.hstatus | HSTATUS_SPRV;
> >> + unsigned long guest_vsstatus, old_stvec, tmp;
> >> +
> >> + guest_sstatus = csr_swap(CSR_SSTATUS, guest_sstatus);
> >> + old_stvec = csr_swap(CSR_STVEC, (ulong)&__kvm_riscv_unpriv_trap);
> >> +
> >> + if (read_insn) {
> >> + guest_vsstatus = csr_read_set(CSR_VSSTATUS, SR_MXR);
> >
> > Is this needed? IIUC SSTATUS.MXR encompasses a wider set of permissions:
> >
> > The HS-level MXR bit makes any executable page readable. {\tt
> > vsstatus}.MXR makes readable those pages marked executable at the VS
> > translation level, but only if readable at the guest-physical
> > translation level.
> >
> > So it should be enough to set SSTATUS.MXR=1 I think. But you also
> > shouldn't set SSTATUS.MXR=1 in the !read_insn case.
> >
> > Also, you can drop the irq save/restore (which is already a save/restore
> > of SSTATUS) since you already write 0 to SSTATUS.SIE in your csr_swap.
> > Perhaps add a BUG_ON(guest_sstatus & SR_SIE) before the csr_swap?
> >
> >> + asm volatile ("\n"
> >> + "csrrw %[hstatus], " STR(CSR_HSTATUS) ", %[hstatus]\n"
> >> + "li %[tilen], 4\n"
> >> + "li %[tscause], 0\n"
> >> + "lhu %[val], (%[addr])\n"
> >> + "andi %[tmp], %[val], 3\n"
> >> + "addi %[tmp], %[tmp], -3\n"
> >> + "bne %[tmp], zero, 2f\n"
> >> + "lhu %[tmp], 2(%[addr])\n"
> >> + "sll %[tmp], %[tmp], 16\n"
> >> + "add %[val], %[val], %[tmp]\n"
> >> + "2: csrw " STR(CSR_HSTATUS) ", %[hstatus]"
> >> + : [hstatus] "+&r"(guest_hstatus), [val] "=&r" (val),
> >> + [tmp] "=&r" (tmp), [tilen] "+&r" (tilen),
> >> + [tscause] "+&r" (tscause)
> >> + : [addr] "r" (addr));
> >> + csr_write(CSR_VSSTATUS, guest_vsstatus);
> >
> >>
> >> +#ifndef CONFIG_RISCV_ISA_C
> >> + "li %[tilen], 4\n"
> >> +#else
> >> + "li %[tilen], 2\n"
> >> +#endif
> >
> > Can you use an assembler directive to force using a non-compressed
> > format for ld and lw? This would get rid of tilen, which is costing 6
> > bytes (if I did the RVC math right) in order to save two. :)
> >
> > Paolo
> >
> >> + "li %[tscause], 0\n"
> >> +#ifdef CONFIG_64BIT
> >> + "ld %[val], (%[addr])\n"
> >> +#else
> >> + "lw %[val], (%[addr])\n"
> >> +#endif
> To: anup@xxxxxxxxxxxxxx
> CC: pbonzini@xxxxxxxxxx
> CC: Anup Patel <Anup.Patel@xxxxxxx>
> CC: Paul Walmsley <paul.walmsley@xxxxxxxxxx>
> CC: rkrcmar@xxxxxxxxxx
> CC: daniel.lezcano@xxxxxxxxxx
> CC: tglx@xxxxxxxxxxxxx
> CC: graf@xxxxxxxxxx
> CC: Atish Patra <Atish.Patra@xxxxxxx>
> CC: Alistair Francis <Alistair.Francis@xxxxxxx>
> CC: Damien Le Moal <Damien.LeMoal@xxxxxxx>
> CC: Christoph Hellwig <hch@xxxxxxxxxxxxx>
> CC: kvm@xxxxxxxxxxxxxxx
> CC: linux-riscv@xxxxxxxxxxxxxxxxxxx
> CC: linux-kernel@xxxxxxxxxxxxxxx
> Subject: Re: [PATCH v7 10/21] RISC-V: KVM: Handle MMIO exits for VCPU
> In-Reply-To: <CAAhSdy1-1yxMnjzppmUBxtSOAuwWaPtNZwW+QH1O7LAnEVP8pg@xxxxxxxxxxxxxx>
>
> On Mon, 23 Sep 2019 06:09:43 PDT (-0700), anup@xxxxxxxxxxxxxx wrote:
> > On Mon, Sep 23, 2019 at 4:42 PM Paolo Bonzini <pbonzini@xxxxxxxxxx> wrote:
> >>
> >> On 04/09/19 18:15, Anup Patel wrote:
> >> > + unsigned long guest_sstatus =
> >> > + vcpu->arch.guest_context.sstatus | SR_MXR;
> >> > + unsigned long guest_hstatus =
> >> > + vcpu->arch.guest_context.hstatus | HSTATUS_SPRV;
> >> > + unsigned long guest_vsstatus, old_stvec, tmp;
> >> > +
> >> > + guest_sstatus = csr_swap(CSR_SSTATUS, guest_sstatus);
> >> > + old_stvec = csr_swap(CSR_STVEC, (ulong)&__kvm_riscv_unpriv_trap);
> >> > +
> >> > + if (read_insn) {
> >> > + guest_vsstatus = csr_read_set(CSR_VSSTATUS, SR_MXR);
> >>
> >> Is this needed? IIUC SSTATUS.MXR encompasses a wider set of permissions:
> >>
> >> The HS-level MXR bit makes any executable page readable. {\tt
> >> vsstatus}.MXR makes readable those pages marked executable at the VS
> >> translation level, but only if readable at the guest-physical
> >> translation level.
> >>
> >> So it should be enough to set SSTATUS.MXR=1 I think. But you also
> >> shouldn't set SSTATUS.MXR=1 in the !read_insn case.
> >
> > I was being overly cautious here. Initially, I thought SSTATUS.MXR
> > applies only to Stage2 and VSSTATUS.MXR applies only to Stage1.
> >
> > I agree with you. The HS-mode should only need to set SSTATUS.MXR.
> >
> >>
> >> Also, you can drop the irq save/restore (which is already a save/restore
> >> of SSTATUS) since you already write 0 to SSTATUS.SIE in your csr_swap.
> >> Perhaps add a BUG_ON(guest_sstatus & SR_SIE) before the csr_swap?
> >
> > I had already dropped irq save/restore in v7 series and having BUG_ON()
> > on guest_sstatus here would be better.
> >
> >>
> >> > + asm volatile ("\n"
> >> > + "csrrw %[hstatus], " STR(CSR_HSTATUS) ", %[hstatus]\n"
> >> > + "li %[tilen], 4\n"
> >> > + "li %[tscause], 0\n"
> >> > + "lhu %[val], (%[addr])\n"
> >> > + "andi %[tmp], %[val], 3\n"
> >> > + "addi %[tmp], %[tmp], -3\n"
> >> > + "bne %[tmp], zero, 2f\n"
> >> > + "lhu %[tmp], 2(%[addr])\n"
> >> > + "sll %[tmp], %[tmp], 16\n"
> >> > + "add %[val], %[val], %[tmp]\n"
> >> > + "2: csrw " STR(CSR_HSTATUS) ", %[hstatus]"
> >> > + : [hstatus] "+&r"(guest_hstatus), [val] "=&r" (val),
> >> > + [tmp] "=&r" (tmp), [tilen] "+&r" (tilen),
> >> > + [tscause] "+&r" (tscause)
> >> > + : [addr] "r" (addr));
> >> > + csr_write(CSR_VSSTATUS, guest_vsstatus);
> >>
> >> >
> >> > +#ifndef CONFIG_RISCV_ISA_C
> >> > + "li %[tilen], 4\n"
> >> > +#else
> >> > + "li %[tilen], 2\n"
> >> > +#endif
> >>
> >> Can you use an assembler directive to force using a non-compressed
> >> format for ld and lw? This would get rid of tilen, which is costing 6
> >> bytes (if I did the RVC math right) in order to save two. :)
> >
> > I tried looking for it but could not find any assembler directive
> > to selectively turn-off instruction compression.
> >
> >>
> >> Paolo
> >>
> >> > + "li %[tscause], 0\n"
> >> > +#ifdef CONFIG_64BIT
> >> > + "ld %[val], (%[addr])\n"
> >> > +#else
> >> > + "lw %[val], (%[addr])\n"
> >> > +#endif
> >
> > Regards,
> > Anup
> To: pbonzini@xxxxxxxxxx
> CC: anup@xxxxxxxxxxxxxx
> CC: Anup Patel <Anup.Patel@xxxxxxx>
> CC: Paul Walmsley <paul.walmsley@xxxxxxxxxx>
> CC: rkrcmar@xxxxxxxxxx
> CC: daniel.lezcano@xxxxxxxxxx
> CC: tglx@xxxxxxxxxxxxx
> CC: graf@xxxxxxxxxx
> CC: Atish Patra <Atish.Patra@xxxxxxx>
> CC: Alistair Francis <Alistair.Francis@xxxxxxx>
> CC: Damien Le Moal <Damien.LeMoal@xxxxxxx>
> CC: Christoph Hellwig <hch@xxxxxxxxxxxxx>
> CC: kvm@xxxxxxxxxxxxxxx
> CC: linux-riscv@xxxxxxxxxxxxxxxxxxx
> CC: linux-kernel@xxxxxxxxxxxxxxx
> Subject: Re: [PATCH v7 10/21] RISC-V: KVM: Handle MMIO exits for VCPU
> In-Reply-To: <45fc3ee5-0f68-4e94-cfb3-0727ca52628f@xxxxxxxxxx>
>
> On Mon, 23 Sep 2019 06:33:14 PDT (-0700), pbonzini@xxxxxxxxxx wrote:
> > On 23/09/19 15:09, Anup Patel wrote:
> >>>> +#ifndef CONFIG_RISCV_ISA_C
> >>>> + "li %[tilen], 4\n"
> >>>> +#else
> >>>> + "li %[tilen], 2\n"
> >>>> +#endif
> >>>
> >>> Can you use an assembler directive to force using a non-compressed
> >>> format for ld and lw? This would get rid of tilen, which is costing 6
> >>> bytes (if I did the RVC math right) in order to save two. :)
> >>
> >> I tried looking for it but could not find any assembler directive
> >> to selectively turn-off instruction compression.
> >
> > ".option norvc"?
> >
> > Paolo
> To: anup@xxxxxxxxxxxxxx
> CC: pbonzini@xxxxxxxxxx
> CC: Anup Patel <Anup.Patel@xxxxxxx>
> CC: Paul Walmsley <paul.walmsley@xxxxxxxxxx>
> CC: rkrcmar@xxxxxxxxxx
> CC: daniel.lezcano@xxxxxxxxxx
> CC: tglx@xxxxxxxxxxxxx
> CC: graf@xxxxxxxxxx
> CC: Atish Patra <Atish.Patra@xxxxxxx>
> CC: Alistair Francis <Alistair.Francis@xxxxxxx>
> CC: Damien Le Moal <Damien.LeMoal@xxxxxxx>
> CC: Christoph Hellwig <hch@xxxxxxxxxxxxx>
> CC: kvm@xxxxxxxxxxxxxxx
> CC: linux-riscv@xxxxxxxxxxxxxxxxxxx
> CC: linux-kernel@xxxxxxxxxxxxxxx
> Subject: Re: [PATCH v7 10/21] RISC-V: KVM: Handle MMIO exits for VCPU
> In-Reply-To: <CAAhSdy29gi2d9c9tumtO68QbB=_+yUYp+ikN3dQ-wa2e-Lesfw@xxxxxxxxxxxxxx>
>
> On Mon, 23 Sep 2019 22:07:43 PDT (-0700), anup@xxxxxxxxxxxxxx wrote:
> > On Mon, Sep 23, 2019 at 7:03 PM Paolo Bonzini <pbonzini@xxxxxxxxxx> wrote:
> >>
> >> On 23/09/19 15:09, Anup Patel wrote:
> >> >>> +#ifndef CONFIG_RISCV_ISA_C
> >> >>> + "li %[tilen], 4\n"
> >> >>> +#else
> >> >>> + "li %[tilen], 2\n"
> >> >>> +#endif
> >> >>
> >> >> Can you use an assembler directive to force using a non-compressed
> >> >> format for ld and lw? This would get rid of tilen, which is costing 6
> >> >> bytes (if I did the RVC math right) in order to save two. :)
> >> >
> >> > I tried looking for it but could not find any assembler directive
> >> > to selectively turn-off instruction compression.
> >>
> >> ".option norvc"?
> >
> > Thanks for the hint. I will try ".option norvc"
>
> It should be something like
>
> .option push
> .option norvc
> ld ...
> .option pop
>
> which preserves C support for the rest of the file.

I have done exactly same thing in v8 patch series sent-out
last week.

Thanks,
Anup

>
> >
> > Regards,
> > Anup
> >
> >>
> >> Paolo