| Message ID | 20241126201539.477448-2-patryk.wlazlyn@linux.intel.com (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | SRF: Fix offline CPU preventing pc6 entry | expand |
Hello Patryk, On Tue, Nov 26, 2024 at 09:15:37PM +0100, Patryk Wlazlyn wrote: > The MWAIT instruction needs different hints on different CPUs to reach > specific idle states. The current hint calculation* in mwait_play_dead() > code works in practice on current Intel hardware, but it fails on a > recent one, Intel's Sierra Forest and possibly some future ones. Those > newer CPUs' power efficiency suffers when the CPU is put offline. > > * The current algorithm for looking up the mwait hint for the deepest > cstate, in mwait_play_dead() code works by inspecting CPUID leaf 0x5 and > calculates the mwait hint based on the number of reported substates. > This approach depends on the hints associated with them to be continuous > in the range [0, NUM_SUBSTATES-1]. This continuity is not documented and > is not met on the recent Intel platforms. > > For example, Intel's Sierra Forest report two cstates with two substates > each in cpuid leaf 0x5: > > Name* target cstate target subcstate (mwait hint) > =========================================================== > C1 0x00 0x00 > C1E 0x00 0x01 > > -- 0x10 ---- > > C6S 0x20 0x22 > C6P 0x20 0x23 > > -- 0x30 ---- > > /* No more (sub)states all the way down to the end. */ > =========================================================== > > * Names of the cstates are not included in the CPUID leaf 0x5, they are > taken from the product specific documentation. > > Notice that hints 0x20 and 0x21 are skipped entirely for the target > cstate 0x20 (C6), being a cause of the problem for the current cpuid > leaf 0x5 algorithm. > > Allow cpuidle code to call mwait play dead loop with a known hint for > the deepest idle state on a given platform, skipping the cpuid based > calculation. Apologies for what may appear as bikeshedding, after this patch, the cpuidle code still won't call any mwait based play dead loop since the support for enter_dead for FFh based idle states in acpi_idle and intel_idle only gets added in Patches 2 and 3. Does it make sense to split this Patch 1 into 2 patches : 1/4 and 4/4 1/4 just introduces the mwait_play_dead_with_hint() helper which will be used by patches 2 and 3. 4/4 get rids of the of logic to find the deepest state from mwait_play_dead() and modifies native_play_dead() to call cpuidle_play_dead() followed by hlt_play_dead() thus removing any reference to mwait_play_dead(). Optionally you can even rename mwait_play_dead_with_hints() to mwait_play_dead(). That way the changelog that you have for this patch can be used in 4/4 since with the addition of play_dead support for FFh states in both acpi_idle and intel_idle via patches 2 and 3, the logic to find the deepest ffh state in mwait_play_dead() is no longer required. Thoughts ? -- Thanks and Regards gautham. > > Signed-off-by: Patryk Wlazlyn <patryk.wlazlyn@linux.intel.com> > --- > arch/x86/include/asm/smp.h | 3 +++ > arch/x86/kernel/smpboot.c | 46 +++++--------------------------------- > 2 files changed, 8 insertions(+), 41 deletions(-) > > diff --git a/arch/x86/include/asm/smp.h b/arch/x86/include/asm/smp.h > index ca073f40698f..633b4a4aec6b 100644 > --- a/arch/x86/include/asm/smp.h > +++ b/arch/x86/include/asm/smp.h > @@ -114,6 +114,7 @@ void wbinvd_on_cpu(int cpu); > int wbinvd_on_all_cpus(void); > > void smp_kick_mwait_play_dead(void); > +void mwait_play_dead(unsigned long hint); > > void native_smp_send_reschedule(int cpu); > void native_send_call_func_ipi(const struct cpumask *mask); > @@ -164,6 +165,8 @@ static inline struct cpumask *cpu_llc_shared_mask(int cpu) > { > return (struct cpumask *)cpumask_of(0); > } > + > +static inline void mwait_play_dead(unsigned long eax_hint) { } > #endif /* CONFIG_SMP */ > > #ifdef CONFIG_DEBUG_NMI_SELFTEST > diff --git a/arch/x86/kernel/smpboot.c b/arch/x86/kernel/smpboot.c > index b5a8f0891135..5dc143e1d6af 100644 > --- a/arch/x86/kernel/smpboot.c > +++ b/arch/x86/kernel/smpboot.c > @@ -1276,45 +1276,9 @@ void play_dead_common(void) > * We need to flush the caches before going to sleep, lest we have > * dirty data in our caches when we come back up. > */ > -static inline void mwait_play_dead(void) > +void __noreturn mwait_play_dead(unsigned long eax_hint) > { > struct mwait_cpu_dead *md = this_cpu_ptr(&mwait_cpu_dead); > - unsigned int eax, ebx, ecx, edx; > - unsigned int highest_cstate = 0; > - unsigned int highest_subcstate = 0; > - int i; > - > - if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD || > - boot_cpu_data.x86_vendor == X86_VENDOR_HYGON) > - return; > - if (!this_cpu_has(X86_FEATURE_MWAIT)) > - return; > - if (!this_cpu_has(X86_FEATURE_CLFLUSH)) > - return; > - if (__this_cpu_read(cpu_info.cpuid_level) < CPUID_MWAIT_LEAF) > - return; > - > - eax = CPUID_MWAIT_LEAF; > - ecx = 0; > - native_cpuid(&eax, &ebx, &ecx, &edx); > - > - /* > - * eax will be 0 if EDX enumeration is not valid. > - * Initialized below to cstate, sub_cstate value when EDX is valid. > - */ > - if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED)) { > - eax = 0; > - } else { > - edx >>= MWAIT_SUBSTATE_SIZE; > - for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) { > - if (edx & MWAIT_SUBSTATE_MASK) { > - highest_cstate = i; > - highest_subcstate = edx & MWAIT_SUBSTATE_MASK; > - } > - } > - eax = (highest_cstate << MWAIT_SUBSTATE_SIZE) | > - (highest_subcstate - 1); > - } > > /* Set up state for the kexec() hack below */ > md->status = CPUDEAD_MWAIT_WAIT; > @@ -1335,7 +1299,7 @@ static inline void mwait_play_dead(void) > mb(); > __monitor(md, 0, 0); > mb(); > - __mwait(eax, 0); > + __mwait(eax_hint, 0); > > if (READ_ONCE(md->control) == CPUDEAD_MWAIT_KEXEC_HLT) { > /* > @@ -1407,9 +1371,9 @@ void native_play_dead(void) > play_dead_common(); > tboot_shutdown(TB_SHUTDOWN_WFS); > > - mwait_play_dead(); > - if (cpuidle_play_dead()) > - hlt_play_dead(); > + /* Below returns only on error. */ > + cpuidle_play_dead(); > + hlt_play_dead(); > } > > #else /* ... !CONFIG_HOTPLUG_CPU */ > -- > 2.47.1 >
> Hello Patryk, > > Apologies for what may appear as bikeshedding, after this patch, the > cpuidle code still won't call any mwait based play dead loop since the > support for enter_dead for FFh based idle states in acpi_idle and > intel_idle only gets added in Patches 2 and 3. > > Does it make sense to split this Patch 1 into 2 patches : 1/4 and 4/4 > > 1/4 just introduces the mwait_play_dead_with_hint() helper which will > be used by patches 2 and 3. > > 4/4 get rids of the of logic to find the deepest state from > mwait_play_dead() and modifies native_play_dead() to call > cpuidle_play_dead() followed by hlt_play_dead() thus removing any > reference to mwait_play_dead(). Optionally you can even rename > mwait_play_dead_with_hints() to mwait_play_dead(). > > That way the changelog that you have for this patch can be used in 4/4 > since with the addition of play_dead support for FFh states in both > acpi_idle and intel_idle via patches 2 and 3, the logic to find the > deepest ffh state in mwait_play_dead() is no longer required. > > Thoughts ? Yeah, makes sense. I just wanted to simplify, but at some point crossed my mind that submitting it like you suggested may be better. I am going to split that if I don't see any objections.
On Wed, Nov 27, 2024 at 9:33 AM Patryk Wlazlyn <patryk.wlazlyn@linux.intel.com> wrote: > > > Hello Patryk, > > > > Apologies for what may appear as bikeshedding, after this patch, the > > cpuidle code still won't call any mwait based play dead loop since the > > support for enter_dead for FFh based idle states in acpi_idle and > > intel_idle only gets added in Patches 2 and 3. > > > > Does it make sense to split this Patch 1 into 2 patches : 1/4 and 4/4 > > > > 1/4 just introduces the mwait_play_dead_with_hint() helper which will > > be used by patches 2 and 3. > > > > 4/4 get rids of the of logic to find the deepest state from > > mwait_play_dead() and modifies native_play_dead() to call > > cpuidle_play_dead() followed by hlt_play_dead() thus removing any > > reference to mwait_play_dead(). Optionally you can even rename > > mwait_play_dead_with_hints() to mwait_play_dead(). > > > > That way the changelog that you have for this patch can be used in 4/4 > > since with the addition of play_dead support for FFh states in both > > acpi_idle and intel_idle via patches 2 and 3, the logic to find the > > deepest ffh state in mwait_play_dead() is no longer required. > > > > Thoughts ? > > Yeah, makes sense. I just wanted to simplify, but at some point crossed my mind > that submitting it like you suggested may be better. Not just it may be better, but this is the only way to do it or people who apply the first patch without the other two patches will likely have problems (and that may easily happen during bisection, for example). Things should always work between consecutive patches in a patch series. > I am going to split that if I don't see any objections. Yes, please!
diff --git a/arch/x86/include/asm/smp.h b/arch/x86/include/asm/smp.h index ca073f40698f..633b4a4aec6b 100644 --- a/arch/x86/include/asm/smp.h +++ b/arch/x86/include/asm/smp.h @@ -114,6 +114,7 @@ void wbinvd_on_cpu(int cpu); int wbinvd_on_all_cpus(void); void smp_kick_mwait_play_dead(void); +void mwait_play_dead(unsigned long hint); void native_smp_send_reschedule(int cpu); void native_send_call_func_ipi(const struct cpumask *mask); @@ -164,6 +165,8 @@ static inline struct cpumask *cpu_llc_shared_mask(int cpu) { return (struct cpumask *)cpumask_of(0); } + +static inline void mwait_play_dead(unsigned long eax_hint) { } #endif /* CONFIG_SMP */ #ifdef CONFIG_DEBUG_NMI_SELFTEST diff --git a/arch/x86/kernel/smpboot.c b/arch/x86/kernel/smpboot.c index b5a8f0891135..5dc143e1d6af 100644 --- a/arch/x86/kernel/smpboot.c +++ b/arch/x86/kernel/smpboot.c @@ -1276,45 +1276,9 @@ void play_dead_common(void) * We need to flush the caches before going to sleep, lest we have * dirty data in our caches when we come back up. */ -static inline void mwait_play_dead(void) +void __noreturn mwait_play_dead(unsigned long eax_hint) { struct mwait_cpu_dead *md = this_cpu_ptr(&mwait_cpu_dead); - unsigned int eax, ebx, ecx, edx; - unsigned int highest_cstate = 0; - unsigned int highest_subcstate = 0; - int i; - - if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD || - boot_cpu_data.x86_vendor == X86_VENDOR_HYGON) - return; - if (!this_cpu_has(X86_FEATURE_MWAIT)) - return; - if (!this_cpu_has(X86_FEATURE_CLFLUSH)) - return; - if (__this_cpu_read(cpu_info.cpuid_level) < CPUID_MWAIT_LEAF) - return; - - eax = CPUID_MWAIT_LEAF; - ecx = 0; - native_cpuid(&eax, &ebx, &ecx, &edx); - - /* - * eax will be 0 if EDX enumeration is not valid. - * Initialized below to cstate, sub_cstate value when EDX is valid. - */ - if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED)) { - eax = 0; - } else { - edx >>= MWAIT_SUBSTATE_SIZE; - for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) { - if (edx & MWAIT_SUBSTATE_MASK) { - highest_cstate = i; - highest_subcstate = edx & MWAIT_SUBSTATE_MASK; - } - } - eax = (highest_cstate << MWAIT_SUBSTATE_SIZE) | - (highest_subcstate - 1); - } /* Set up state for the kexec() hack below */ md->status = CPUDEAD_MWAIT_WAIT; @@ -1335,7 +1299,7 @@ static inline void mwait_play_dead(void) mb(); __monitor(md, 0, 0); mb(); - __mwait(eax, 0); + __mwait(eax_hint, 0); if (READ_ONCE(md->control) == CPUDEAD_MWAIT_KEXEC_HLT) { /* @@ -1407,9 +1371,9 @@ void native_play_dead(void) play_dead_common(); tboot_shutdown(TB_SHUTDOWN_WFS); - mwait_play_dead(); - if (cpuidle_play_dead()) - hlt_play_dead(); + /* Below returns only on error. */ + cpuidle_play_dead(); + hlt_play_dead(); } #else /* ... !CONFIG_HOTPLUG_CPU */
The MWAIT instruction needs different hints on different CPUs to reach specific idle states. The current hint calculation* in mwait_play_dead() code works in practice on current Intel hardware, but it fails on a recent one, Intel's Sierra Forest and possibly some future ones. Those newer CPUs' power efficiency suffers when the CPU is put offline. * The current algorithm for looking up the mwait hint for the deepest cstate, in mwait_play_dead() code works by inspecting CPUID leaf 0x5 and calculates the mwait hint based on the number of reported substates. This approach depends on the hints associated with them to be continuous in the range [0, NUM_SUBSTATES-1]. This continuity is not documented and is not met on the recent Intel platforms. For example, Intel's Sierra Forest report two cstates with two substates each in cpuid leaf 0x5: Name* target cstate target subcstate (mwait hint) =========================================================== C1 0x00 0x00 C1E 0x00 0x01 -- 0x10 ---- C6S 0x20 0x22 C6P 0x20 0x23 -- 0x30 ---- /* No more (sub)states all the way down to the end. */ =========================================================== * Names of the cstates are not included in the CPUID leaf 0x5, they are taken from the product specific documentation. Notice that hints 0x20 and 0x21 are skipped entirely for the target cstate 0x20 (C6), being a cause of the problem for the current cpuid leaf 0x5 algorithm. Allow cpuidle code to call mwait play dead loop with a known hint for the deepest idle state on a given platform, skipping the cpuid based calculation. Signed-off-by: Patryk Wlazlyn <patryk.wlazlyn@linux.intel.com> --- arch/x86/include/asm/smp.h | 3 +++ arch/x86/kernel/smpboot.c | 46 +++++--------------------------------- 2 files changed, 8 insertions(+), 41 deletions(-)