Message ID | 20231204143606.1806432-1-maz@kernel.org (mailing list archive) |
---|---|
Headers | show |
Series | arm64: Drop support for VPIPT i-cache policy | expand |
On Mon, Dec 04, 2023 at 02:36:03PM +0000, Marc Zyngier wrote: > ARMv8.2 introduced support for VPIPT i-caches, the V standing for > VMID-tagged. Although this looked like a reasonable idea, no > implementation has ever made it into the wild. > > Linux has supported this for over 6 years (amusingly, just as the > architecture was dropping support for AIVIVT i-caches), but we had no > way to even test it, and it is likely that this code was just > bit-rotting. > > However, in a recent breakthrough (XML drop 2023-09, tagged as > d55f5af8e09052abe92a02adf820deea2eaed717), the architecture has > finally been purged of this option, making VIPT and PIPT the only two > valid options. > > This really means this code is just dead code. Nobody will ever come > up with such an implementation, and we can just get rid of it. > > Most of the impact is on KVM, where we drop a few large comment blocks > (and a bit of code), while the core arch code loses the detection code > itself. > > * From v2: > - Fix reserved naming for RESERVED_AIVIVT > - Collected RBs from Anshuman an Zenghui > > Marc Zyngier (3): > KVM: arm64: Remove VPIPT I-cache handling > arm64: Kill detection of VPIPT i-cache policy > arm64: Rename reserved values for CTR_EL0.L1Ip For the series: Acked-by: Mark Rutland <mark.rutland@arm.com> Looking forward, we can/should probably replace __icache_flags with a single ICACHE_NOALIASING or ICACHE_PIPT cpucap, which'd get rid of a bunch of duplicated logic and make that more sound in the case of races around cpu onlining. Mark. > > arch/arm64/include/asm/cache.h | 6 ---- > arch/arm64/include/asm/kvm_mmu.h | 7 ---- > arch/arm64/kernel/cpuinfo.c | 5 --- > arch/arm64/kvm/hyp/nvhe/pkvm.c | 2 +- > arch/arm64/kvm/hyp/nvhe/tlb.c | 61 -------------------------------- > arch/arm64/kvm/hyp/vhe/tlb.c | 13 ------- > arch/arm64/tools/sysreg | 5 +-- > 7 files changed, 4 insertions(+), 95 deletions(-) > > -- > 2.39.2 > >
On Mon, 04 Dec 2023 14:44:56 +0000, Mark Rutland <mark.rutland@arm.com> wrote: > > On Mon, Dec 04, 2023 at 02:36:03PM +0000, Marc Zyngier wrote: > > ARMv8.2 introduced support for VPIPT i-caches, the V standing for > > VMID-tagged. Although this looked like a reasonable idea, no > > implementation has ever made it into the wild. > > > > Linux has supported this for over 6 years (amusingly, just as the > > architecture was dropping support for AIVIVT i-caches), but we had no > > way to even test it, and it is likely that this code was just > > bit-rotting. > > > > However, in a recent breakthrough (XML drop 2023-09, tagged as > > d55f5af8e09052abe92a02adf820deea2eaed717), the architecture has > > finally been purged of this option, making VIPT and PIPT the only two > > valid options. > > > > This really means this code is just dead code. Nobody will ever come > > up with such an implementation, and we can just get rid of it. > > > > Most of the impact is on KVM, where we drop a few large comment blocks > > (and a bit of code), while the core arch code loses the detection code > > itself. > > > > * From v2: > > - Fix reserved naming for RESERVED_AIVIVT > > - Collected RBs from Anshuman an Zenghui > > > > Marc Zyngier (3): > > KVM: arm64: Remove VPIPT I-cache handling > > arm64: Kill detection of VPIPT i-cache policy > > arm64: Rename reserved values for CTR_EL0.L1Ip > > For the series: > > Acked-by: Mark Rutland <mark.rutland@arm.com> Thanks. > Looking forward, we can/should probably replace __icache_flags with a single > ICACHE_NOALIASING or ICACHE_PIPT cpucap, which'd get rid of a bunch of > duplicated logic and make that more sound in the case of races around cpu > onlining. As long as we refuse VIPT CPUs coming up late (i.e. after we have patched the kernel to set ICACHE_PIPT), it should be doable. I guess we already have this restriction as userspace is able to probe the I-cache policy anyway. How about the patch below (tested in a guest with a bunch of hacks to expose different L1Ip values)? Thanks, M. From 8f88afb0b317213dcbf18ea460a508bfccc18568 Mon Sep 17 00:00:00 2001 From: Marc Zyngier <maz@kernel.org> Date: Mon, 4 Dec 2023 18:09:40 +0000 Subject: [PATCH] arm64: Make icache detection a cpu capability Now that we only have two icache policies, we are in a good position to make the whole detection business more robust. Let's replace __icache_flags by a single capability (ICACHE_PIPT), and use this if all CPUs are indeed PIPT. It means we can rely on existing logic to mandate that a VIPT CPU coming up late will be denied booting, which is the safe thing to do. This also leads to some nice cleanups in pKVM, and KVM as a whole can use ARM64_ICACHE_PIPT as a final cap. Suggested-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Marc Zyngier <maz@kernel.org> --- arch/arm64/include/asm/cache.h | 9 ++------- arch/arm64/include/asm/kvm_hyp.h | 1 - arch/arm64/include/asm/kvm_mmu.h | 2 +- arch/arm64/kernel/cpufeature.c | 7 +++++++ arch/arm64/kernel/cpuinfo.c | 34 -------------------------------- arch/arm64/kvm/arm.c | 1 - arch/arm64/kvm/hyp/nvhe/pkvm.c | 3 --- arch/arm64/tools/cpucaps | 1 + arch/arm64/tools/sysreg | 2 +- 9 files changed, 12 insertions(+), 48 deletions(-) diff --git a/arch/arm64/include/asm/cache.h b/arch/arm64/include/asm/cache.h index 06a4670bdb0b..8ef9522a6151 100644 --- a/arch/arm64/include/asm/cache.h +++ b/arch/arm64/include/asm/cache.h @@ -37,9 +37,9 @@ #ifndef __ASSEMBLY__ -#include <linux/bitops.h> #include <linux/kasan-enabled.h> +#include <asm/cpufeature.h> #include <asm/cputype.h> #include <asm/mte-def.h> #include <asm/sysreg.h> @@ -55,18 +55,13 @@ static inline unsigned int arch_slab_minalign(void) #define arch_slab_minalign() arch_slab_minalign() #endif -#define CTR_L1IP(ctr) SYS_FIELD_GET(CTR_EL0, L1Ip, ctr) - -#define ICACHEF_ALIASING 0 -extern unsigned long __icache_flags; - /* * Whilst the D-side always behaves as PIPT on AArch64, aliasing is * permitted in the I-cache. */ static inline int icache_is_aliasing(void) { - return test_bit(ICACHEF_ALIASING, &__icache_flags); + return !cpus_have_cap(ARM64_ICACHE_PIPT); } static inline u32 cache_type_cwg(void) diff --git a/arch/arm64/include/asm/kvm_hyp.h b/arch/arm64/include/asm/kvm_hyp.h index 145ce73fc16c..7ad70f9865fd 100644 --- a/arch/arm64/include/asm/kvm_hyp.h +++ b/arch/arm64/include/asm/kvm_hyp.h @@ -140,7 +140,6 @@ extern u64 kvm_nvhe_sym(id_aa64mmfr1_el1_sys_val); extern u64 kvm_nvhe_sym(id_aa64mmfr2_el1_sys_val); extern u64 kvm_nvhe_sym(id_aa64smfr0_el1_sys_val); -extern unsigned long kvm_nvhe_sym(__icache_flags); extern unsigned int kvm_nvhe_sym(kvm_arm_vmid_bits); #endif /* __ARM64_KVM_HYP_H__ */ diff --git a/arch/arm64/include/asm/kvm_mmu.h b/arch/arm64/include/asm/kvm_mmu.h index e3e793d0ec30..ab86c2f025cf 100644 --- a/arch/arm64/include/asm/kvm_mmu.h +++ b/arch/arm64/include/asm/kvm_mmu.h @@ -248,7 +248,7 @@ static inline void __invalidate_icache_guest_page(void *va, size_t size) * invalidation range exceeds our arbitrary limit on invadations by * cache line. */ - if (icache_is_aliasing() || size > __invalidate_icache_max_range()) + if (!cpus_have_final_cap(ARM64_ICACHE_PIPT) || size > __invalidate_icache_max_range()) icache_inval_all_pou(); else icache_inval_pou((unsigned long)va, (unsigned long)va + size); diff --git a/arch/arm64/kernel/cpufeature.c b/arch/arm64/kernel/cpufeature.c index 646591c67e7a..73a37176676e 100644 --- a/arch/arm64/kernel/cpufeature.c +++ b/arch/arm64/kernel/cpufeature.c @@ -2272,6 +2272,13 @@ static const struct arm64_cpu_capabilities arm64_features[] = { .type = ARM64_CPUCAP_SYSTEM_FEATURE, .matches = has_always, }, + { + .desc = "PIPT I-cache policy", + .capability = ARM64_ICACHE_PIPT, + .type = ARM64_CPUCAP_SYSTEM_FEATURE, + .matches = has_cpuid_feature, + ARM64_CPUID_FIELDS(CTR_EL0, L1Ip, PIPT) + }, { .desc = "GIC system register CPU interface", .capability = ARM64_HAS_GIC_CPUIF_SYSREGS, diff --git a/arch/arm64/kernel/cpuinfo.c b/arch/arm64/kernel/cpuinfo.c index 47043c0d95ec..a4ea331fb6d4 100644 --- a/arch/arm64/kernel/cpuinfo.c +++ b/arch/arm64/kernel/cpuinfo.c @@ -33,20 +33,6 @@ DEFINE_PER_CPU(struct cpuinfo_arm64, cpu_data); static struct cpuinfo_arm64 boot_cpu_data; -static inline const char *icache_policy_str(int l1ip) -{ - switch (l1ip) { - case CTR_EL0_L1Ip_VIPT: - return "VIPT"; - case CTR_EL0_L1Ip_PIPT: - return "PIPT"; - default: - return "RESERVED/UNKNOWN"; - } -} - -unsigned long __icache_flags; - static const char *const hwcap_str[] = { [KERNEL_HWCAP_FP] = "fp", [KERNEL_HWCAP_ASIMD] = "asimd", @@ -378,24 +364,6 @@ static int __init cpuinfo_regs_init(void) } device_initcall(cpuinfo_regs_init); -static void cpuinfo_detect_icache_policy(struct cpuinfo_arm64 *info) -{ - unsigned int cpu = smp_processor_id(); - u32 l1ip = CTR_L1IP(info->reg_ctr); - - switch (l1ip) { - case CTR_EL0_L1Ip_PIPT: - break; - case CTR_EL0_L1Ip_VIPT: - default: - /* Assume aliasing */ - set_bit(ICACHEF_ALIASING, &__icache_flags); - break; - } - - pr_info("Detected %s I-cache on CPU%d\n", icache_policy_str(l1ip), cpu); -} - static void __cpuinfo_store_cpu_32bit(struct cpuinfo_32bit *info) { info->reg_id_dfr0 = read_cpuid(ID_DFR0_EL1); @@ -457,8 +425,6 @@ static void __cpuinfo_store_cpu(struct cpuinfo_arm64 *info) if (id_aa64pfr0_32bit_el0(info->reg_id_aa64pfr0)) __cpuinfo_store_cpu_32bit(&info->aarch32); - - cpuinfo_detect_icache_policy(info); } void cpuinfo_store_cpu(void) diff --git a/arch/arm64/kvm/arm.c b/arch/arm64/kvm/arm.c index e5f75f1f1085..a5f71165cd5c 100644 --- a/arch/arm64/kvm/arm.c +++ b/arch/arm64/kvm/arm.c @@ -2242,7 +2242,6 @@ static void kvm_hyp_init_symbols(void) kvm_nvhe_sym(id_aa64mmfr1_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64MMFR1_EL1); kvm_nvhe_sym(id_aa64mmfr2_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64MMFR2_EL1); kvm_nvhe_sym(id_aa64smfr0_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64SMFR0_EL1); - kvm_nvhe_sym(__icache_flags) = __icache_flags; kvm_nvhe_sym(kvm_arm_vmid_bits) = kvm_arm_vmid_bits; } diff --git a/arch/arm64/kvm/hyp/nvhe/pkvm.c b/arch/arm64/kvm/hyp/nvhe/pkvm.c index b29f15418c0a..187ce5720697 100644 --- a/arch/arm64/kvm/hyp/nvhe/pkvm.c +++ b/arch/arm64/kvm/hyp/nvhe/pkvm.c @@ -12,9 +12,6 @@ #include <nvhe/pkvm.h> #include <nvhe/trap_handler.h> -/* Used by icache_is_aliasing(). */ -unsigned long __icache_flags; - /* Used by kvm_get_vttbr(). */ unsigned int kvm_arm_vmid_bits; diff --git a/arch/arm64/tools/cpucaps b/arch/arm64/tools/cpucaps index b98c38288a9d..9c6313cd6f5e 100644 --- a/arch/arm64/tools/cpucaps +++ b/arch/arm64/tools/cpucaps @@ -53,6 +53,7 @@ HAS_TLB_RANGE HAS_VIRT_HOST_EXTN HAS_WFXT HW_DBM +ICACHE_PIPT KVM_HVHE KVM_PROTECTED_MODE MISMATCHED_CACHE_TYPE diff --git a/arch/arm64/tools/sysreg b/arch/arm64/tools/sysreg index c5af75b23187..db8c96841138 100644 --- a/arch/arm64/tools/sysreg +++ b/arch/arm64/tools/sysreg @@ -2003,7 +2003,7 @@ Field 28 IDC Field 27:24 CWG Field 23:20 ERG Field 19:16 DminLine -Enum 15:14 L1Ip +UnsignedEnum 15:14 L1Ip # This was named as VPIPT in the ARM but now documented as reserved 0b00 RESERVED_VPIPT # This is named as AIVIVT in the ARM but documented as reserved
On Mon, Dec 04, 2023 at 06:26:25PM +0000, Marc Zyngier wrote: > On Mon, 04 Dec 2023 14:44:56 +0000, > Mark Rutland <mark.rutland@arm.com> wrote: > > > > On Mon, Dec 04, 2023 at 02:36:03PM +0000, Marc Zyngier wrote: > > > ARMv8.2 introduced support for VPIPT i-caches, the V standing for > > > VMID-tagged. Although this looked like a reasonable idea, no > > > implementation has ever made it into the wild. > > > > > > Linux has supported this for over 6 years (amusingly, just as the > > > architecture was dropping support for AIVIVT i-caches), but we had no > > > way to even test it, and it is likely that this code was just > > > bit-rotting. > > > > > > However, in a recent breakthrough (XML drop 2023-09, tagged as > > > d55f5af8e09052abe92a02adf820deea2eaed717), the architecture has > > > finally been purged of this option, making VIPT and PIPT the only two > > > valid options. > > > > > > This really means this code is just dead code. Nobody will ever come > > > up with such an implementation, and we can just get rid of it. > > > > > > Most of the impact is on KVM, where we drop a few large comment blocks > > > (and a bit of code), while the core arch code loses the detection code > > > itself. > > > > > > * From v2: > > > - Fix reserved naming for RESERVED_AIVIVT > > > - Collected RBs from Anshuman an Zenghui > > > > > > Marc Zyngier (3): > > > KVM: arm64: Remove VPIPT I-cache handling > > > arm64: Kill detection of VPIPT i-cache policy > > > arm64: Rename reserved values for CTR_EL0.L1Ip > > > > For the series: > > > > Acked-by: Mark Rutland <mark.rutland@arm.com> > > Thanks. > > > Looking forward, we can/should probably replace __icache_flags with a single > > ICACHE_NOALIASING or ICACHE_PIPT cpucap, which'd get rid of a bunch of > > duplicated logic and make that more sound in the case of races around cpu > > onlining. > > As long as we refuse VIPT CPUs coming up late (i.e. after we have > patched the kernel to set ICACHE_PIPT), it should be doable. I guess > we already have this restriction as userspace is able to probe the > I-cache policy anyway. > > How about the patch below (tested in a guest with a bunch of hacks to > expose different L1Ip values)? That's roughly what I was thinking; the diff looks good, minor comments below. > > Thanks, > > M. > > From 8f88afb0b317213dcbf18ea460a508bfccc18568 Mon Sep 17 00:00:00 2001 > From: Marc Zyngier <maz@kernel.org> > Date: Mon, 4 Dec 2023 18:09:40 +0000 > Subject: [PATCH] arm64: Make icache detection a cpu capability > > Now that we only have two icache policies, we are in a good position > to make the whole detection business more robust. > > Let's replace __icache_flags by a single capability (ICACHE_PIPT), > and use this if all CPUs are indeed PIPT. It means we can rely on > existing logic to mandate that a VIPT CPU coming up late will be > denied booting, which is the safe thing to do. > > This also leads to some nice cleanups in pKVM, and KVM as a whole > can use ARM64_ICACHE_PIPT as a final cap. > > Suggested-by: Mark Rutland <mark.rutland@arm.com> > Signed-off-by: Marc Zyngier <maz@kernel.org> > --- > arch/arm64/include/asm/cache.h | 9 ++------- > arch/arm64/include/asm/kvm_hyp.h | 1 - > arch/arm64/include/asm/kvm_mmu.h | 2 +- > arch/arm64/kernel/cpufeature.c | 7 +++++++ > arch/arm64/kernel/cpuinfo.c | 34 -------------------------------- > arch/arm64/kvm/arm.c | 1 - > arch/arm64/kvm/hyp/nvhe/pkvm.c | 3 --- > arch/arm64/tools/cpucaps | 1 + > arch/arm64/tools/sysreg | 2 +- > 9 files changed, 12 insertions(+), 48 deletions(-) Nice diffstat! [...] > /* > * Whilst the D-side always behaves as PIPT on AArch64, aliasing is > * permitted in the I-cache. > */ > static inline int icache_is_aliasing(void) > { > - return test_bit(ICACHEF_ALIASING, &__icache_flags); > + return !cpus_have_cap(ARM64_ICACHE_PIPT); > } It might be nicer to use alternative_has_cap_{likely,unlikely}(...) for consistency with other cap checks, though that won't matter for hyp code and I don't think the likely/unlikely part particularly matters either. [...] > -static void cpuinfo_detect_icache_policy(struct cpuinfo_arm64 *info) > -{ > - unsigned int cpu = smp_processor_id(); > - u32 l1ip = CTR_L1IP(info->reg_ctr); > - > - switch (l1ip) { > - case CTR_EL0_L1Ip_PIPT: > - break; > - case CTR_EL0_L1Ip_VIPT: > - default: > - /* Assume aliasing */ > - set_bit(ICACHEF_ALIASING, &__icache_flags); > - break; > - } > - > - pr_info("Detected %s I-cache on CPU%d\n", icache_policy_str(l1ip), cpu); > -} Not printing this for each CPU is a nice bonus! [...] > diff --git a/arch/arm64/tools/sysreg b/arch/arm64/tools/sysreg > index c5af75b23187..db8c96841138 100644 > --- a/arch/arm64/tools/sysreg > +++ b/arch/arm64/tools/sysreg > @@ -2003,7 +2003,7 @@ Field 28 IDC > Field 27:24 CWG > Field 23:20 ERG > Field 19:16 DminLine > -Enum 15:14 L1Ip > +UnsignedEnum 15:14 L1Ip > # This was named as VPIPT in the ARM but now documented as reserved > 0b00 RESERVED_VPIPT > # This is named as AIVIVT in the ARM but documented as reserved I was initially surprised by the use of UnsignedEnum, but given PIPT is 0b11, I can see that works. Otherwise, we can keep this as an enum and use a helper that checks for an exact match. Mark.
On Tue, 05 Dec 2023 11:03:16 +0000, Mark Rutland <mark.rutland@arm.com> wrote: > > On Mon, Dec 04, 2023 at 06:26:25PM +0000, Marc Zyngier wrote: > > How about the patch below (tested in a guest with a bunch of hacks to > > expose different L1Ip values)? > > That's roughly what I was thinking; the diff looks good, minor comments below. > > [...] > > > /* > > * Whilst the D-side always behaves as PIPT on AArch64, aliasing is > > * permitted in the I-cache. > > */ > > static inline int icache_is_aliasing(void) > > { > > - return test_bit(ICACHEF_ALIASING, &__icache_flags); > > + return !cpus_have_cap(ARM64_ICACHE_PIPT); > > } > > It might be nicer to use alternative_has_cap_{likely,unlikely}(...) for > consistency with other cap checks, though that won't matter for hyp code and I > don't think the likely/unlikely part particularly matters either. Right, this should be marginally better. I was initially worried of the cap not being configured yet, but since we start by assuming VIPT and only flip to PIPT once we have seen all CPUs, this stays correct by construction. > [...] > > > diff --git a/arch/arm64/tools/sysreg b/arch/arm64/tools/sysreg > > index c5af75b23187..db8c96841138 100644 > > --- a/arch/arm64/tools/sysreg > > +++ b/arch/arm64/tools/sysreg > > @@ -2003,7 +2003,7 @@ Field 28 IDC > > Field 27:24 CWG > > Field 23:20 ERG > > Field 19:16 DminLine > > -Enum 15:14 L1Ip > > +UnsignedEnum 15:14 L1Ip > > # This was named as VPIPT in the ARM but now documented as reserved > > 0b00 RESERVED_VPIPT > > # This is named as AIVIVT in the ARM but documented as reserved > > I was initially surprised by the use of UnsignedEnum, but given PIPT is 0b11, I > can see that works. Otherwise, we can keep this as an enum and use a helper > that checks for an exact match. Yeah, I hesitated either way, but maybe the helper approach is cleaner. I'll give it a go and repost this patch. Thanks, M.
On Mon, 4 Dec 2023 14:36:03 +0000, Marc Zyngier wrote: > ARMv8.2 introduced support for VPIPT i-caches, the V standing for > VMID-tagged. Although this looked like a reasonable idea, no > implementation has ever made it into the wild. > > Linux has supported this for over 6 years (amusingly, just as the > architecture was dropping support for AIVIVT i-caches), but we had no > way to even test it, and it is likely that this code was just > bit-rotting. > > [...] Applied to arm64 (for-next/rip-vpipt), thanks! [1/3] KVM: arm64: Remove VPIPT I-cache handling https://git.kernel.org/arm64/c/ced242ba9d7c [2/3] arm64: Kill detection of VPIPT i-cache policy https://git.kernel.org/arm64/c/d8e12a0d3715 [3/3] arm64: Rename reserved values for CTR_EL0.L1Ip https://git.kernel.org/arm64/c/f35c32ca6839 Cheers,