[PATCH v12 5/8] arm64: kvm: Save/restore MTE registers
Steven Price
steven.price at arm.com
Thu May 20 08:21:57 PDT 2021
On 20/05/2021 10:46, Marc Zyngier wrote:
> On Wed, 19 May 2021 14:04:20 +0100,
> Steven Price <steven.price at arm.com> wrote:
>>
>> On 17/05/2021 18:17, Marc Zyngier wrote:
>>> On Mon, 17 May 2021 13:32:36 +0100,
>>> Steven Price <steven.price at arm.com> wrote:
>>>>
>>>> Define the new system registers that MTE introduces and context switch
>>>> them. The MTE feature is still hidden from the ID register as it isn't
>>>> supported in a VM yet.
>>>>
>>>> Signed-off-by: Steven Price <steven.price at arm.com>
>>>> ---
>>>> arch/arm64/include/asm/kvm_host.h | 6 ++
>>>> arch/arm64/include/asm/kvm_mte.h | 66 ++++++++++++++++++++++
>>>> arch/arm64/include/asm/sysreg.h | 3 +-
>>>> arch/arm64/kernel/asm-offsets.c | 3 +
>>>> arch/arm64/kvm/hyp/entry.S | 7 +++
>>>> arch/arm64/kvm/hyp/include/hyp/sysreg-sr.h | 21 +++++++
>>>> arch/arm64/kvm/sys_regs.c | 22 ++++++--
>>>> 7 files changed, 123 insertions(+), 5 deletions(-)
>>>> create mode 100644 arch/arm64/include/asm/kvm_mte.h
>>>>
>>>> diff --git a/arch/arm64/include/asm/kvm_host.h b/arch/arm64/include/asm/kvm_host.h
>>>> index afaa5333f0e4..309e36cc1b42 100644
>>>> --- a/arch/arm64/include/asm/kvm_host.h
>>>> +++ b/arch/arm64/include/asm/kvm_host.h
>>>> @@ -208,6 +208,12 @@ enum vcpu_sysreg {
>>>> CNTP_CVAL_EL0,
>>>> CNTP_CTL_EL0,
>>>>
>>>> + /* Memory Tagging Extension registers */
>>>> + RGSR_EL1, /* Random Allocation Tag Seed Register */
>>>> + GCR_EL1, /* Tag Control Register */
>>>> + TFSR_EL1, /* Tag Fault Status Register (EL1) */
>>>> + TFSRE0_EL1, /* Tag Fault Status Register (EL0) */
>>>> +
>>>> /* 32bit specific registers. Keep them at the end of the range */
>>>> DACR32_EL2, /* Domain Access Control Register */
>>>> IFSR32_EL2, /* Instruction Fault Status Register */
>>>> diff --git a/arch/arm64/include/asm/kvm_mte.h b/arch/arm64/include/asm/kvm_mte.h
>>>> new file mode 100644
>>>> index 000000000000..6541c7d6ce06
>>>> --- /dev/null
>>>> +++ b/arch/arm64/include/asm/kvm_mte.h
>>>> @@ -0,0 +1,66 @@
>>>> +/* SPDX-License-Identifier: GPL-2.0 */
>>>> +/*
>>>> + * Copyright (C) 2020 ARM Ltd.
>>>> + */
>>>> +#ifndef __ASM_KVM_MTE_H
>>>> +#define __ASM_KVM_MTE_H
>>>> +
>>>> +#ifdef __ASSEMBLY__
>>>> +
>>>> +#include <asm/sysreg.h>
>>>> +
>>>> +#ifdef CONFIG_ARM64_MTE
>>>> +
>>>> +.macro mte_switch_to_guest g_ctxt, h_ctxt, reg1
>>>> +alternative_if_not ARM64_MTE
>>>> + b .L__skip_switch\@
>>>> +alternative_else_nop_endif
>>>> + mrs \reg1, hcr_el2
>>>> + and \reg1, \reg1, #(HCR_ATA)
>>>> + cbz \reg1, .L__skip_switch\@
>>>> +
>>>> + mrs_s \reg1, SYS_RGSR_EL1
>>>> + str \reg1, [\h_ctxt, #CPU_RGSR_EL1]
>>>> + mrs_s \reg1, SYS_GCR_EL1
>>>> + str \reg1, [\h_ctxt, #CPU_GCR_EL1]
>>>> +
>>>> + ldr \reg1, [\g_ctxt, #CPU_RGSR_EL1]
>>>> + msr_s SYS_RGSR_EL1, \reg1
>>>> + ldr \reg1, [\g_ctxt, #CPU_GCR_EL1]
>>>> + msr_s SYS_GCR_EL1, \reg1
>>>> +
>>>> +.L__skip_switch\@:
>>>> +.endm
>>>> +
>>>> +.macro mte_switch_to_hyp g_ctxt, h_ctxt, reg1
>>>> +alternative_if_not ARM64_MTE
>>>> + b .L__skip_switch\@
>>>> +alternative_else_nop_endif
>>>> + mrs \reg1, hcr_el2
>>>> + and \reg1, \reg1, #(HCR_ATA)
>>>> + cbz \reg1, .L__skip_switch\@
>>>> +
>>>> + mrs_s \reg1, SYS_RGSR_EL1
>>>> + str \reg1, [\g_ctxt, #CPU_RGSR_EL1]
>>>> + mrs_s \reg1, SYS_GCR_EL1
>>>> + str \reg1, [\g_ctxt, #CPU_GCR_EL1]
>>>> +
>>>> + ldr \reg1, [\h_ctxt, #CPU_RGSR_EL1]
>>>> + msr_s SYS_RGSR_EL1, \reg1
>>>> + ldr \reg1, [\h_ctxt, #CPU_GCR_EL1]
>>>> + msr_s SYS_GCR_EL1, \reg1
>>>
>>> What is the rational for not having any synchronisation here? It is
>>> quite uncommon to allocate memory at EL2, but VHE can perform all kind
>>> of tricks.
>>
>> I don't follow. This is part of the __guest_exit path and there's an ISB
>> at the end of that - is that not sufficient? I don't see any possibility
>> for allocating memory before that. What am I missing?
>
> Which ISB? We have a few in the SError handling code, but that's
> conditioned on not having RAS. With any RAS-enabled CPU, we return to
> C code early, since we don't need any extra synchronisation (see the
> comment about the absence of ISB on this path).
Ah, I clearly didn't read the code (or comment) carefully enough -
indeed with RAS we're potentially skipping the ISB.
> I would really like to ensure that we return to C code in the exact
> state we left it.
Agreed, I'll stick an ISB at the end of mte_switch_to_hyp. Although
there's clearly room for optimisation here as ptrauth_switch_to_hyp has
a similar ISB.
>>
>>>> +
>>>> +.L__skip_switch\@:
>>>> +.endm
>>>> +
>>>> +#else /* CONFIG_ARM64_MTE */
>>>> +
>>>> +.macro mte_switch_to_guest g_ctxt, h_ctxt, reg1
>>>> +.endm
>>>> +
>>>> +.macro mte_switch_to_hyp g_ctxt, h_ctxt, reg1
>>>> +.endm
>>>> +
>>>> +#endif /* CONFIG_ARM64_MTE */
>>>> +#endif /* __ASSEMBLY__ */
>>>> +#endif /* __ASM_KVM_MTE_H */
>>>> diff --git a/arch/arm64/include/asm/sysreg.h b/arch/arm64/include/asm/sysreg.h
>>>> index 65d15700a168..347ccac2341e 100644
>>>> --- a/arch/arm64/include/asm/sysreg.h
>>>> +++ b/arch/arm64/include/asm/sysreg.h
>>>> @@ -651,7 +651,8 @@
>>>>
>>>> #define INIT_SCTLR_EL2_MMU_ON \
>>>> (SCTLR_ELx_M | SCTLR_ELx_C | SCTLR_ELx_SA | SCTLR_ELx_I | \
>>>> - SCTLR_ELx_IESB | SCTLR_ELx_WXN | ENDIAN_SET_EL2 | SCTLR_EL2_RES1)
>>>> + SCTLR_ELx_IESB | SCTLR_ELx_WXN | ENDIAN_SET_EL2 | \
>>>> + SCTLR_ELx_ITFSB | SCTLR_EL2_RES1)
>>>>
>>>> #define INIT_SCTLR_EL2_MMU_OFF \
>>>> (SCTLR_EL2_RES1 | ENDIAN_SET_EL2)
>>>> diff --git a/arch/arm64/kernel/asm-offsets.c b/arch/arm64/kernel/asm-offsets.c
>>>> index 0cb34ccb6e73..6b489a8462f0 100644
>>>> --- a/arch/arm64/kernel/asm-offsets.c
>>>> +++ b/arch/arm64/kernel/asm-offsets.c
>>>> @@ -111,6 +111,9 @@ int main(void)
>>>> DEFINE(VCPU_WORKAROUND_FLAGS, offsetof(struct kvm_vcpu, arch.workaround_flags));
>>>> DEFINE(VCPU_HCR_EL2, offsetof(struct kvm_vcpu, arch.hcr_el2));
>>>> DEFINE(CPU_USER_PT_REGS, offsetof(struct kvm_cpu_context, regs));
>>>> + DEFINE(CPU_RGSR_EL1, offsetof(struct kvm_cpu_context, sys_regs[RGSR_EL1]));
>>>> + DEFINE(CPU_GCR_EL1, offsetof(struct kvm_cpu_context, sys_regs[GCR_EL1]));
>>>> + DEFINE(CPU_TFSRE0_EL1, offsetof(struct kvm_cpu_context, sys_regs[TFSRE0_EL1]));
>>>
>>> TFSRE0_EL1 is never accessed from assembly code. Leftover from a
>>> previous version?
>>
>> Indeed, I will drop it.
>>
>>>> DEFINE(CPU_APIAKEYLO_EL1, offsetof(struct kvm_cpu_context, sys_regs[APIAKEYLO_EL1]));
>>>> DEFINE(CPU_APIBKEYLO_EL1, offsetof(struct kvm_cpu_context, sys_regs[APIBKEYLO_EL1]));
>>>> DEFINE(CPU_APDAKEYLO_EL1, offsetof(struct kvm_cpu_context, sys_regs[APDAKEYLO_EL1]));
>>>> diff --git a/arch/arm64/kvm/hyp/entry.S b/arch/arm64/kvm/hyp/entry.S
>>>> index e831d3dfd50d..435346ea1504 100644
>>>> --- a/arch/arm64/kvm/hyp/entry.S
>>>> +++ b/arch/arm64/kvm/hyp/entry.S
>>>> @@ -13,6 +13,7 @@
>>>> #include <asm/kvm_arm.h>
>>>> #include <asm/kvm_asm.h>
>>>> #include <asm/kvm_mmu.h>
>>>> +#include <asm/kvm_mte.h>
>>>> #include <asm/kvm_ptrauth.h>
>>>>
>>>> .text
>>>> @@ -51,6 +52,9 @@ alternative_else_nop_endif
>>>>
>>>> add x29, x0, #VCPU_CONTEXT
>>>>
>>>> + // mte_switch_to_guest(g_ctxt, h_ctxt, tmp1)
>>>> + mte_switch_to_guest x29, x1, x2
>>>> +
>>>> // Macro ptrauth_switch_to_guest format:
>>>> // ptrauth_switch_to_guest(guest cxt, tmp1, tmp2, tmp3)
>>>> // The below macro to restore guest keys is not implemented in C code
>>>> @@ -142,6 +146,9 @@ SYM_INNER_LABEL(__guest_exit, SYM_L_GLOBAL)
>>>> // when this feature is enabled for kernel code.
>>>> ptrauth_switch_to_hyp x1, x2, x3, x4, x5
>>>>
>>>> + // mte_switch_to_hyp(g_ctxt, h_ctxt, reg1)
>>>> + mte_switch_to_hyp x1, x2, x3
>>>> +
>>>> // Restore hyp's sp_el0
>>>> restore_sp_el0 x2, x3
>>>>
>>>> diff --git a/arch/arm64/kvm/hyp/include/hyp/sysreg-sr.h b/arch/arm64/kvm/hyp/include/hyp/sysreg-sr.h
>>>> index cce43bfe158f..de7e14c862e6 100644
>>>> --- a/arch/arm64/kvm/hyp/include/hyp/sysreg-sr.h
>>>> +++ b/arch/arm64/kvm/hyp/include/hyp/sysreg-sr.h
>>>> @@ -14,6 +14,7 @@
>>>> #include <asm/kvm_asm.h>
>>>> #include <asm/kvm_emulate.h>
>>>> #include <asm/kvm_hyp.h>
>>>> +#include <asm/kvm_mmu.h>
>>>>
>>>> static inline void __sysreg_save_common_state(struct kvm_cpu_context *ctxt)
>>>> {
>>>> @@ -26,6 +27,16 @@ static inline void __sysreg_save_user_state(struct kvm_cpu_context *ctxt)
>>>> ctxt_sys_reg(ctxt, TPIDRRO_EL0) = read_sysreg(tpidrro_el0);
>>>> }
>>>>
>>>> +static inline bool ctxt_has_mte(struct kvm_cpu_context *ctxt)
>>>> +{
>>>> + struct kvm_vcpu *vcpu = ctxt->__hyp_running_vcpu;
>>>> +
>>>> + if (!vcpu)
>>>> + vcpu = container_of(ctxt, struct kvm_vcpu, arch.ctxt);
>>>> +
>>>> + return kvm_has_mte(kern_hyp_va(vcpu->kvm));
>>>> +}
>>>> +
>>>> static inline void __sysreg_save_el1_state(struct kvm_cpu_context *ctxt)
>>>> {
>>>> ctxt_sys_reg(ctxt, CSSELR_EL1) = read_sysreg(csselr_el1);
>>>> @@ -46,6 +57,11 @@ static inline void __sysreg_save_el1_state(struct kvm_cpu_context *ctxt)
>>>> ctxt_sys_reg(ctxt, PAR_EL1) = read_sysreg_par();
>>>> ctxt_sys_reg(ctxt, TPIDR_EL1) = read_sysreg(tpidr_el1);
>>>>
>>>> + if (ctxt_has_mte(ctxt)) {
>>>> + ctxt_sys_reg(ctxt, TFSR_EL1) = read_sysreg_el1(SYS_TFSR);
>>>> + ctxt_sys_reg(ctxt, TFSRE0_EL1) = read_sysreg_s(SYS_TFSRE0_EL1);
>>>> + }
>>>
>>> I remember suggesting that this is slightly heavier than necessary.
>>>
>>> On nVHE, TFSRE0_EL1 could be moved to load/put, as we never run
>>> userspace with a vcpu loaded. The same holds of course for VHE, but we
>>> also can move TFSR_EL1 to load/put, as the host uses TFSR_EL2.
>>>
>>> Do you see any issue with that?
>>
>> The comment[1] I made before was:
>
> Ah, I totally missed this email (or can't remember reading it, which
> amounts to the same thing). Apologies for that.
>
>> For TFSR_EL1 + VHE I believe it is synchronised only on vcpu_load/put -
>> __sysreg_save_el1_state() is called from kvm_vcpu_load_sysregs_vhe().
>>
>> TFSRE0_EL1 potentially could be improved. I have to admit I was unsure
>> if it should be in __sysreg_save_user_state() instead. However AFAICT
>> that is called at the same time as __sysreg_save_el1_state() and there's
>> no optimisation for nVHE. And given it's an _EL1 register this seemed
>> like the logic place.
>>
>> Am I missing something here? Potentially there are other registers to be
>> optimised (TPIDRRO_EL0 looks like a possiblity), but IMHO that doesn't
>> belong in this series.
>>
>> For VHE TFSR_EL1 is already only saved/restored on load/put
>> (__sysreg_save_el1_state() is called from kvm_vcpu_put_sysregs_vhe()).
>>
>> TFSRE0_EL1 could be moved, but I'm not sure where it should live as I
>> mentioned above.
>
> Yeah, this looks fine, please ignore my rambling.
No problem!
Thanks,
Steve
More information about the linux-arm-kernel
mailing list