| Message ID | 20231012172601.22600-2-alejandro.vallejo@cloud.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | Fix AMD erratum #1485 on AMD Zen4 | expand |
On 13/10/2023 1:26 am, Alejandro Vallejo wrote: > It slightly simplifies the code that uses them at no real cost because the > code is very rarely executed. This makes it harder to confuse zen uarches > due to missing or mistaken family checks. I'm afraid I disagree. It's bogus to do a family check without a vendor check. By making this change, you're separating (spatially in code, and therefore cognitively) the two checks that it's important to be able to cross-check. > diff --git a/xen/arch/x86/include/asm/amd.h b/xen/arch/x86/include/asm/amd.h > index d862cb7972..5a40bcc2ba 100644 > --- a/xen/arch/x86/include/asm/amd.h > +++ b/xen/arch/x86/include/asm/amd.h > @@ -145,11 +145,12 @@ > * Hygon (Fam18h) but without simple model number rules. Instead, use STIBP > * as a heuristic that distinguishes the two. > * > - * The caller is required to perform the appropriate vendor/family checks > - * first. > + * The caller is required to perform the appropriate vendor check first. > */ > -#define is_zen1_uarch() (!boot_cpu_has(X86_FEATURE_AMD_STIBP)) > -#define is_zen2_uarch() boot_cpu_has(X86_FEATURE_AMD_STIBP) > +#define is_zen1_uarch() ((boot_cpu_data.x86 == 0x17 || boot_cpu_data.x86 == 0x18) && \ > + !boot_cpu_has(X86_FEATURE_AMD_STIBP)) > +#define is_zen2_uarch() (boot_cpu_data.x86 == 0x17 && \ > + boot_cpu_has(X86_FEATURE_AMD_STIBP)) What leads you to believe there aren't Hygon Zen2's ? ~Andrew
Hi, I'll just remove this patch (and amend the next) in the interest of having it committed early. That said... On Fri, Oct 13, 2023 at 10:14:45AM +0800, Andrew Cooper wrote: > On 13/10/2023 1:26 am, Alejandro Vallejo wrote: > > It slightly simplifies the code that uses them at no real cost because the > > code is very rarely executed. This makes it harder to confuse zen uarches > > due to missing or mistaken family checks. > > I'm afraid I disagree. As it stands, it's a matter of time before a bug of this form creeps up. Particularly because it reads very innocent. if (is_zen1_uarch()) fun1(); else if (is_zen2_uarch()) fun2(); else if (is_zen3_uarch()) fun3(); else if (is_zen4_uarch()) fun4(); Particularly if the bodies of each conditional are big enough that you lose track of the family you're dealing with. > > It's bogus to do a family check without a vendor check. I can get behind that. I didn't include the vendor check because by and large these macros are used in vendor-specific areas. Would that appease your concerns? Whenever the macros are used we're in glacially cold paths anyway. > By making this > change, you're separating (spatially in code, and therefore cognitively) > the two checks that it's important to be able to cross-check. IMO, It's no different from separating the heuristic from the vendor/family check. What causes definite cognitive load is (as a reader) having to remember what particular fields must be read off boot_cpu_data in order to discriminate zenN, which of them are co-familiar and so on. > > > diff --git a/xen/arch/x86/include/asm/amd.h b/xen/arch/x86/include/asm/amd.h > > index d862cb7972..5a40bcc2ba 100644 > > --- a/xen/arch/x86/include/asm/amd.h > > +++ b/xen/arch/x86/include/asm/amd.h > > @@ -145,11 +145,12 @@ > > * Hygon (Fam18h) but without simple model number rules. Instead, use STIBP > > * as a heuristic that distinguishes the two. > > * > > - * The caller is required to perform the appropriate vendor/family checks > > - * first. > > + * The caller is required to perform the appropriate vendor check first. > > */ > > -#define is_zen1_uarch() (!boot_cpu_has(X86_FEATURE_AMD_STIBP)) > > -#define is_zen2_uarch() boot_cpu_has(X86_FEATURE_AMD_STIBP) > > +#define is_zen1_uarch() ((boot_cpu_data.x86 == 0x17 || boot_cpu_data.x86 == 0x18) && \ > > + !boot_cpu_has(X86_FEATURE_AMD_STIBP)) > > +#define is_zen2_uarch() (boot_cpu_data.x86 == 0x17 && \ > > + boot_cpu_has(X86_FEATURE_AMD_STIBP)) > > What leads you to believe there aren't Hygon Zen2's ? The same argument that a Hygon zen2 supports STIBP. Having seen neither HW nor docs all that's left is divination skills :) > > ~Andrew As I said, let's move this discussion away from the errata fix (I'll remove this patch and adapt the next one to remove the family check), as it's unrelated and can be done later if I manage to convince you. Thanks, Alejandro
On Fri, Oct 13, 2023 at 12:23:43PM +0100, Alejandro Vallejo wrote: > Hi, > > I'll just remove this patch (and amend the next) in the interest of having > it committed early. That said... > > On Fri, Oct 13, 2023 at 10:14:45AM +0800, Andrew Cooper wrote: > > On 13/10/2023 1:26 am, Alejandro Vallejo wrote: > > > It slightly simplifies the code that uses them at no real cost because the > > > code is very rarely executed. This makes it harder to confuse zen uarches > > > due to missing or mistaken family checks. > > > > I'm afraid I disagree. > As it stands, it's a matter of time before a bug of this form creeps up. Particularly > because it reads very innocent. > > if (is_zen1_uarch()) > fun1(); > else if (is_zen2_uarch()) > fun2(); > else if (is_zen3_uarch()) > fun3(); > else if (is_zen4_uarch()) > fun4(); > > Particularly if the bodies of each conditional are big enough that you lose > track of the family you're dealing with. > > > > > It's bogus to do a family check without a vendor check. > I can get behind that. I didn't include the vendor check because by and > large these macros are used in vendor-specific areas. Would that appease > your concerns? Whenever the macros are used we're in glacially cold paths > anyway. > > > By making this > > change, you're separating (spatially in code, and therefore cognitively) > > the two checks that it's important to be able to cross-check. > IMO, It's no different from separating the heuristic from the vendor/family > check. What causes definite cognitive load is (as a reader) having to > remember what particular fields must be read off boot_cpu_data in order to > discriminate zenN, which of them are co-familiar and so on. > > > > > > diff --git a/xen/arch/x86/include/asm/amd.h b/xen/arch/x86/include/asm/amd.h > > > index d862cb7972..5a40bcc2ba 100644 > > > --- a/xen/arch/x86/include/asm/amd.h > > > +++ b/xen/arch/x86/include/asm/amd.h > > > @@ -145,11 +145,12 @@ > > > * Hygon (Fam18h) but without simple model number rules. Instead, use STIBP > > > * as a heuristic that distinguishes the two. > > > * > > > - * The caller is required to perform the appropriate vendor/family checks > > > - * first. > > > + * The caller is required to perform the appropriate vendor check first. > > > */ > > > -#define is_zen1_uarch() (!boot_cpu_has(X86_FEATURE_AMD_STIBP)) > > > -#define is_zen2_uarch() boot_cpu_has(X86_FEATURE_AMD_STIBP) > > > +#define is_zen1_uarch() ((boot_cpu_data.x86 == 0x17 || boot_cpu_data.x86 == 0x18) && \ > > > + !boot_cpu_has(X86_FEATURE_AMD_STIBP)) > > > +#define is_zen2_uarch() (boot_cpu_data.x86 == 0x17 && \ > > > + boot_cpu_has(X86_FEATURE_AMD_STIBP)) > > > > What leads you to believe there aren't Hygon Zen2's ? > The same argument that a Hygon zen2 supports STIBP. Having seen neither HW > nor docs all that's left is divination skills :) There is at least one reference online, from: https://www.tomshardware.com/news/amd-zen-china-x86-ip-license,39573.html "THATIC* was a single-generation technology license, and there are no additional technology licenses," Su explained, though she did not clarify if the decision not to extend the technology transfers was a direct result of the trade war. That means that the technology transfer, which provided THATIC with access to the first-gen Zen microarchitecture, will not be extended to allow the Chinese chipmaker access to AMD's Zen 2 microarchitecture." * Tianjin Haiguang Advanced Technology Investment Co. Ltd. (THATIC) So I would expect there aren't any Hygon Zen2 given those claims, but again that relies on how much you trust tomshardware.com. The article seems to have been reviewed by AMD as some changes were requested. Regards, Roger.
On 13/10/2023 7:36 pm, Roger Pau Monné wrote: > On Fri, Oct 13, 2023 at 12:23:43PM +0100, Alejandro Vallejo wrote: >>>> diff --git a/xen/arch/x86/include/asm/amd.h b/xen/arch/x86/include/asm/amd.h >>>> index d862cb7972..5a40bcc2ba 100644 >>>> --- a/xen/arch/x86/include/asm/amd.h >>>> +++ b/xen/arch/x86/include/asm/amd.h >>>> @@ -145,11 +145,12 @@ >>>> * Hygon (Fam18h) but without simple model number rules. Instead, use STIBP >>>> * as a heuristic that distinguishes the two. >>>> * >>>> - * The caller is required to perform the appropriate vendor/family checks >>>> - * first. >>>> + * The caller is required to perform the appropriate vendor check first. >>>> */ >>>> -#define is_zen1_uarch() (!boot_cpu_has(X86_FEATURE_AMD_STIBP)) >>>> -#define is_zen2_uarch() boot_cpu_has(X86_FEATURE_AMD_STIBP) >>>> +#define is_zen1_uarch() ((boot_cpu_data.x86 == 0x17 || boot_cpu_data.x86 == 0x18) && \ >>>> + !boot_cpu_has(X86_FEATURE_AMD_STIBP)) >>>> +#define is_zen2_uarch() (boot_cpu_data.x86 == 0x17 && \ >>>> + boot_cpu_has(X86_FEATURE_AMD_STIBP)) >>> What leads you to believe there aren't Hygon Zen2's ? >> The same argument that a Hygon zen2 supports STIBP. Having seen neither HW >> nor docs all that's left is divination skills :) Hygon reported/fixed Zen2 uarch issues in Xen. Something clearly exists. ~Andrew
diff --git a/xen/arch/x86/cpu/amd.c b/xen/arch/x86/cpu/amd.c index 4f27187f92..22aa8c0085 100644 --- a/xen/arch/x86/cpu/amd.c +++ b/xen/arch/x86/cpu/amd.c @@ -1056,8 +1056,7 @@ static void cf_check init_amd(struct cpuinfo_x86 *c) amd_init_ssbd(c); - if (c->x86 == 0x17) - amd_init_spectral_chicken(); + amd_init_spectral_chicken(); /* Probe for NSCB on Zen2 CPUs when not virtualised */ if (!cpu_has_hypervisor && !cpu_has_nscb && c == &boot_cpu_data && @@ -1293,7 +1292,7 @@ static int __init cf_check zen2_c6_errata_check(void) */ s_time_t delta; - if (cpu_has_hypervisor || boot_cpu_data.x86 != 0x17 || !is_zen2_uarch()) + if (cpu_has_hypervisor || !is_zen2_uarch()) return 0; /* diff --git a/xen/arch/x86/include/asm/amd.h b/xen/arch/x86/include/asm/amd.h index d862cb7972..5a40bcc2ba 100644 --- a/xen/arch/x86/include/asm/amd.h +++ b/xen/arch/x86/include/asm/amd.h @@ -145,11 +145,12 @@ * Hygon (Fam18h) but without simple model number rules. Instead, use STIBP * as a heuristic that distinguishes the two. * - * The caller is required to perform the appropriate vendor/family checks - * first. + * The caller is required to perform the appropriate vendor check first. */ -#define is_zen1_uarch() (!boot_cpu_has(X86_FEATURE_AMD_STIBP)) -#define is_zen2_uarch() boot_cpu_has(X86_FEATURE_AMD_STIBP) +#define is_zen1_uarch() ((boot_cpu_data.x86 == 0x17 || boot_cpu_data.x86 == 0x18) && \ + !boot_cpu_has(X86_FEATURE_AMD_STIBP)) +#define is_zen2_uarch() (boot_cpu_data.x86 == 0x17 && \ + boot_cpu_has(X86_FEATURE_AMD_STIBP)) struct cpuinfo_x86; int cpu_has_amd_erratum(const struct cpuinfo_x86 *, int, ...); diff --git a/xen/arch/x86/spec_ctrl.c b/xen/arch/x86/spec_ctrl.c index 6fd7d44ce4..508c454516 100644 --- a/xen/arch/x86/spec_ctrl.c +++ b/xen/arch/x86/spec_ctrl.c @@ -976,9 +976,6 @@ static bool __init has_div_vuln(void) (X86_VENDOR_AMD | X86_VENDOR_HYGON)) ) return false; - if ( boot_cpu_data.x86 != 0x17 && boot_cpu_data.x86 != 0x18 ) - return false; - return is_zen1_uarch(); }
It slightly simplifies the code that uses them at no real cost because the code is very rarely executed. This makes it harder to confuse zen uarches due to missing or mistaken family checks. No functional change intended. Signed-off-by: Alejandro Vallejo <alejandro.vallejo@cloud.com> --- xen/arch/x86/cpu/amd.c | 5 ++--- xen/arch/x86/include/asm/amd.h | 9 +++++---- xen/arch/x86/spec_ctrl.c | 3 --- 3 files changed, 7 insertions(+), 10 deletions(-)