2018-12-27 21:53:30

by David Wang

[permalink] [raw]
Subject: [PATCH v3] Optimize C3 entry on Centaur CPUs

For new Centaur CPUs the ucode will take care of the preservation of cache coherence
between CPU cores in C-states regardless of how deep the C-states are. So, it is not
necessary to flush the caches in software befor entering C3. And this useless operation
will cause performance drop for the cores which share some caches with the idling core.

Signed-off-by: David Wang <[email protected]>
Reviewed-by: Thomas Gleixner <[email protected]>

Changes from v2 to v3:
*1, Replace "c->x86_mask" with "c->x86_stepping".

Changes from v1 to v2:
* 1, Add some Family/Model/Stepping contrains to let this patch only apply
* to new centaur CPUs.
* 2, The arbiter disable/enable operations maybe needed for old VIA/Centaur
* platform. So, delete "flags->bm_control=0" in patch v1.

---
arch/x86/kernel/acpi/cstate.c | 12 ++++++++++++
1 file changed, 12 insertions(+)

diff --git a/arch/x86/kernel/acpi/cstate.c b/arch/x86/kernel/acpi/cstate.c
index 158ad14..ed17640 100644
--- a/arch/x86/kernel/acpi/cstate.c
+++ b/arch/x86/kernel/acpi/cstate.c
@@ -51,6 +51,18 @@ void acpi_processor_power_init_bm_check(struct acpi_processor_flags *flags,
if (c->x86_vendor == X86_VENDOR_INTEL &&
(c->x86 > 0xf || (c->x86 == 6 && c->x86_model >= 0x0f)))
flags->bm_control = 0;
+ /*
+ * For all recent Centaur CPUs, the ucode will make sure that each
+ * core can keep cache coherence with each other while entering C3
+ * type state. So, set bm_check to 1 to indicate that the kernel
+ * need not execute a cache flush operation (WBINVD) when entering
+ * C3 type state.
+ */
+ if (c->x86_vendor == X86_VENDOR_CENTAUR) {
+ if (c->x86 > 6 || (c->x86 == 6 && c->x86_model == 0x0f &&
+ c->x86_stepping >= 0x0e))
+ flags->bm_check = 1;
+ }
}
EXPORT_SYMBOL(acpi_processor_power_init_bm_check);

--
1.9.1



2018-12-31 18:56:01

by Pavel Machek

[permalink] [raw]
Subject: Re: [PATCH v3] Optimize C3 entry on Centaur CPUs

On Thu 2018-12-27 16:41:50, David Wang wrote:
> For new Centaur CPUs the ucode will take care of the preservation of cache coherence
> between CPU cores in C-states regardless of how deep the C-states are. So, it is not
> necessary to flush the caches in software befor entering C3. And this useless operation
> will cause performance drop for the cores which share some caches with the idling core.
>
> Signed-off-by: David Wang <[email protected]>
> Reviewed-by: Thomas Gleixner <[email protected]>

Acked-by: Pavel Machek <[email protected]>

--
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html

Attachments:
(No filename) (699.00 B)
 signature.asc (188.00 B)
Digital signature Download all attachments
Subject: [tip:x86/cpu] x86/power: Optimize C3 entry on Centaur CPUs

Commit-ID: 987ddbe4870b53623d76ac64044c55a13e368113
Gitweb: https://git.kernel.org/tip/987ddbe4870b53623d76ac64044c55a13e368113
Author: David Wang <[email protected]>
AuthorDate: Thu, 27 Dec 2018 16:41:50 +0800
Committer: Ingo Molnar <[email protected]>
CommitDate: Fri, 19 Apr 2019 19:28:06 +0200

x86/power: Optimize C3 entry on Centaur CPUs

For new Centaur CPUs the ucode will take care of the preservation of cache coherence
between CPU cores in C-states regardless of how deep the C-states are. So, it is not
necessary to flush the caches in software befor entering C3. This useless operation
will cause performance drop for the cores which share some caches with the idling core.

Signed-off-by: David Wang <[email protected]>
Reviewed-by: Thomas Gleixner <[email protected]>
Acked-by: Pavel Machek <[email protected]>
Cc: Linus Torvalds <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Cc: [email protected]
Cc: [email protected]
Cc: [email protected]
Cc: [email protected]
Cc: [email protected]
Cc: [email protected]
Cc: [email protected]
Link: http://lkml.kernel.org/r/[email protected]
[ Tidy up the comment. ]
Signed-off-by: Ingo Molnar <[email protected]>
---
arch/x86/kernel/acpi/cstate.c | 12 ++++++++++++
1 file changed, 12 insertions(+)

diff --git a/arch/x86/kernel/acpi/cstate.c b/arch/x86/kernel/acpi/cstate.c
index 158ad1483c43..cb6e076a6d39 100644
--- a/arch/x86/kernel/acpi/cstate.c
+++ b/arch/x86/kernel/acpi/cstate.c
@@ -51,6 +51,18 @@ void acpi_processor_power_init_bm_check(struct acpi_processor_flags *flags,
if (c->x86_vendor == X86_VENDOR_INTEL &&
(c->x86 > 0xf || (c->x86 == 6 && c->x86_model >= 0x0f)))
flags->bm_control = 0;
+ /*
+ * For all recent Centaur CPUs, the ucode will make sure that each
+ * core can keep cache coherence with each other while entering C3
+ * type state. So, set bm_check to 1 to indicate that the kernel
+ * doesn't need to execute a cache flush operation (WBINVD) when
+ * entering C3 type state.
+ */
+ if (c->x86_vendor == X86_VENDOR_CENTAUR) {
+ if (c->x86 > 6 || (c->x86 == 6 && c->x86_model == 0x0f &&
+ c->x86_stepping >= 0x0e))
+ flags->bm_check = 1;
+ }
}
EXPORT_SYMBOL(acpi_processor_power_init_bm_check);