Skip to contentSkip to conversation

Add ACPI C4 support

PATCH Applied Mario Limonciello (AMD) 2 participants 6 messages Sep 15, 2025 – Sep 22, 2025

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

cover Add ACPI C4 support

Some AMD systems have introduced support for ACPI C4. Systems that support ACPI C4 will enter CPU OFF at ACPI C3 and VDD OFF at ACPI C4.

VDD OFF is a requirement for s0i3, and thus if C4 isn't supported then a system can't enter s0i3.

Add support for ACPI C4 to the processor idle driver, following similar semantics of ACPI C3.

Mario Limonciello (AMD) (2):
x86/acpi/cstate: Remove open coded check for cpu_feature_enabled()
ACPI: processor: Add support for ACPI C4 state

 arch/x86/kernel/acpi/cstate.c |  2 +-
 drivers/acpi/processor_idle.c | 10 +++++++---
 include/acpi/actypes.h        |  5 +++--
 3 files changed, 11 insertions(+), 6 deletions(-)
Signature · 1 line
2.43.0
#1 lore raw

PATCH 1/2 x86/acpi/cstate: Remove open coded check for cpu_feature_enabled()

acpi_processor_power_init_bm_check() has an open coded implementation of cpu_feature_enabled() to detect X86_FEATURE_ZEN.
Switch to using cpu_feature_enabled().

---
 arch/x86/kernel/acpi/cstate.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
arch/x86/kernel/acpi/cstate.c +1 -1
index 8698d66563ed6..0281703da5e26 100644
@@ -89,7 +89,7 @@ void acpi_processor_power_init_bm_check(struct acpi_processor_flags *flags,
 		 */
 		flags->bm_control = 0;
 	}
-	if (c->x86_vendor == X86_VENDOR_AMD && c->x86 >= 0x17) {
+	if (cpu_feature_enabled(X86_FEATURE_ZEN)) {
 		/*
 		 * For all AMD Zen or newer CPUs that support C3, caches
 		 * should not be flushed by software while entering C3
Signature · 1 line
2.43.0
maintainer Applied
#2 lore raw

Re: [PATCH 1/2] x86/acpi/cstate: Remove open coded check for cpu_feature_enabled()

Quoted 15 lines · Mario Limonciello (AMD)

acpi_processor_power_init_bm_check() has an open coded implementation of cpu_feature_enabled() to detect X86_FEATURE_ZEN.
Switch to using cpu_feature_enabled().

---
 arch/x86/kernel/acpi/cstate.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
arch/x86/kernel/acpi/cstate.c +0 -0
index 8698d66563ed6..0281703da5e26 100644

arch/x86/kernel/acpi/cstate.c void acpi_processor_power_init_bm_check(struct acpi_processor_flags *flags,

Quoted 9 lines
arch/x86/kernel/acpi/cstate.c +1 -2
@@ -89,7 +89,7 @@ void acpi_processor_power_init_bm_check(struct acpi_processor_flags *flags,
                 */
                flags->bm_control = 0;
        }
-       if (c->x86_vendor == X86_VENDOR_AMD && c->x86 >= 0x17) {
+       if (cpu_feature_enabled(X86_FEATURE_ZEN)) {
                /*
                 * For all AMD Zen or newer CPUs that support C3, caches
                 * should not be flushed by software while entering C3
--

Applied as 6.18 material, thanks!

#3 lore raw

PATCH 2/2 ACPI: processor: Add support for ACPI C4 state

In reply to Mario Limonciello (AMD)

The _CST object described in the ACPI specification [1] allows processors to support any arbitrary number of power states that are described by a package describing the register used, the type of C-state, latency and the power consumption.

Currently the Linux kernel supports up to ACPI C3, and if a system supports any further states they are ignored. This causes problems on some AMD hardware which can support up to ACPI C4.

AMD systems that support up to C3 will enter CPUOFF and VDDOFF when entering C3. Systems that support up to C4 will enter CPUOFF when going to C3 and will enter VDDOFF when entering into C4.

The existing semantics for bus mastering around C3 are also valid for C4, so instead of hardcoding to C3, map then >= C3. In the case of s2idle detect the deepest C-state supported and enter the deepest.

---
 drivers/acpi/processor_idle.c | 10 +++++++---
 include/acpi/actypes.h        |  5 +++--
 2 files changed, 10 insertions(+), 5 deletions(-)
drivers/acpi/processor_idle.c +7 -3
index 5dacf41d7cc0a..537b0119535ea 100644
@@ -444,6 +444,10 @@ static int acpi_processor_power_verify(struct acpi_processor *pr)
 		case ACPI_STATE_C3:
 			acpi_processor_power_verify_c3(pr, cx);
 			break;
+		case ACPI_STATE_C4:
+			if (!cx->latency || !cx->address)
+				break;
+			cx->valid = 1;
 		}
 		if (!cx->valid)
 			continue;
@@ -685,7 +689,7 @@ static int __cpuidle acpi_idle_enter(struct cpuidle_device *dev,
 		return -EINVAL;
 
 	if (cx->type != ACPI_STATE_C1) {
-		if (cx->type == ACPI_STATE_C3 && pr->flags.bm_check)
+		if (cx->type >= ACPI_STATE_C3 && pr->flags.bm_check)
 			return acpi_idle_enter_bm(drv, pr, cx, index);
 
 		/* C2 to C1 demotion. */
@@ -708,7 +712,7 @@ static int __cpuidle acpi_idle_enter_s2idle(struct cpuidle_device *dev,
 {
 	struct acpi_processor_cx *cx = per_cpu(acpi_cstate[index], dev->cpu);
 
-	if (cx->type == ACPI_STATE_C3) {
+	if (cx->index == drv->state_count - 1) {
 		struct acpi_processor *pr = __this_cpu_read(processors);
 
 		if (unlikely(!pr))
@@ -754,7 +758,7 @@ static int acpi_processor_setup_cpuidle_cx(struct acpi_processor *pr,
 		if (lapic_timer_needs_broadcast(pr, cx))
 			state->flags |= CPUIDLE_FLAG_TIMER_STOP;
 
-		if (cx->type == ACPI_STATE_C3) {
+		if (cx->type >= ACPI_STATE_C3) {
 			state->flags |= CPUIDLE_FLAG_TLB_FLUSHED;
 			if (pr->flags.bm_check)
 				state->flags |= CPUIDLE_FLAG_RCU_IDLE;
include/acpi/actypes.h +3 -2
index 8fe893d776dde..6c9f472af482c 100644
@@ -600,8 +600,9 @@ typedef u64 acpi_integer;
 #define ACPI_STATE_C1                   (u8) 1
 #define ACPI_STATE_C2                   (u8) 2
 #define ACPI_STATE_C3                   (u8) 3
-#define ACPI_C_STATES_MAX               ACPI_STATE_C3
-#define ACPI_C_STATE_COUNT              4
+#define ACPI_STATE_C4                   (u8) 4
+#define ACPI_C_STATES_MAX               ACPI_STATE_C4
+#define ACPI_C_STATE_COUNT              5
 
 /*
  * Sleep type invalid value
Signature · 1 line
2.43.0
#4 lore raw

Re: [PATCH 2/2] ACPI: processor: Add support for ACPI C4 state

Quoted 72 lines · Mario Limonciello (AMD)

The _CST object described in the ACPI specification [1] allows processors to support any arbitrary number of power states that are described by a package describing the register used, the type of C-state, latency and the power consumption.

Currently the Linux kernel supports up to ACPI C3, and if a system supports any further states they are ignored. This causes problems on some AMD hardware which can support up to ACPI C4.

AMD systems that support up to C3 will enter CPUOFF and VDDOFF when entering C3. Systems that support up to C4 will enter CPUOFF when going to C3 and will enter VDDOFF when entering into C4.

The existing semantics for bus mastering around C3 are also valid for C4, so instead of hardcoding to C3, map then >= C3. In the case of s2idle detect the deepest C-state supported and enter the deepest.

---
 drivers/acpi/processor_idle.c | 10 +++++++---
 include/acpi/actypes.h        |  5 +++--
 2 files changed, 10 insertions(+), 5 deletions(-)
drivers/acpi/processor_idle.c +7 -3
index 5dacf41d7cc0a..537b0119535ea 100644
@@ -444,6 +444,10 @@ static int acpi_processor_power_verify(struct acpi_processor *pr)
                case ACPI_STATE_C3:
                        acpi_processor_power_verify_c3(pr, cx);
                        break;
+               case ACPI_STATE_C4:
+                       if (!cx->latency || !cx->address)
+                               break;
+                       cx->valid = 1;
                }
                if (!cx->valid)
                        continue;
@@ -685,7 +689,7 @@ static int __cpuidle acpi_idle_enter(struct cpuidle_device *dev,
                return -EINVAL;
 
        if (cx->type != ACPI_STATE_C1) {
-               if (cx->type == ACPI_STATE_C3 && pr->flags.bm_check)
+               if (cx->type >= ACPI_STATE_C3 && pr->flags.bm_check)
                        return acpi_idle_enter_bm(drv, pr, cx, index);
 
                /* C2 to C1 demotion. */
@@ -708,7 +712,7 @@ static int __cpuidle acpi_idle_enter_s2idle(struct cpuidle_device *dev,
 {
        struct acpi_processor_cx *cx = per_cpu(acpi_cstate[index], dev->cpu);
 
-       if (cx->type == ACPI_STATE_C3) {
+       if (cx->index == drv->state_count - 1) {
                struct acpi_processor *pr = __this_cpu_read(processors);
 
                if (unlikely(!pr))
@@ -754,7 +758,7 @@ static int acpi_processor_setup_cpuidle_cx(struct acpi_processor *pr,
                if (lapic_timer_needs_broadcast(pr, cx))
                        state->flags |= CPUIDLE_FLAG_TIMER_STOP;
 
-               if (cx->type == ACPI_STATE_C3) {
+               if (cx->type >= ACPI_STATE_C3) {
                        state->flags |= CPUIDLE_FLAG_TLB_FLUSHED;
                        if (pr->flags.bm_check)
                                state->flags |= CPUIDLE_FLAG_RCU_IDLE;
include/acpi/actypes.h +0 -0
index 8fe893d776dde..6c9f472af482c 100644

include/acpi/actypes.h typedef u64 acpi_integer; · line 605

Quoted 8 lines
include/acpi/actypes.h +3 -2
@@ -600,8 +600,9 @@ typedef u64 acpi_integer;
 #define ACPI_STATE_C1                   (u8) 1
 #define ACPI_STATE_C2                   (u8) 2
 #define ACPI_STATE_C3                   (u8) 3
-#define ACPI_C_STATES_MAX               ACPI_STATE_C3
-#define ACPI_C_STATE_COUNT              4
+#define ACPI_STATE_C4                   (u8) 4
+#define ACPI_C_STATES_MAX               ACPI_STATE_C4
+#define ACPI_C_STATE_COUNT              5

Don't you need to update ACPICA accordingly?

include/acpi/actypes.h typedef u64 acpi_integer;

Quoted 4 lines
include/acpi/actypes.h +0 -1
 /*
  * Sleep type invalid value
--

The current point in the cycle is kind of unsuitable for changes like this. I'd prefer to revisit it after 6.18-rc1.

#5 lore raw

Re: [PATCH 2/2] ACPI: processor: Add support for ACPI C4 state

Quoted 76 lines · Rafael J. Wysocki
Quoted 74 lines · Mario Limonciello (AMD)

The _CST object described in the ACPI specification [1] allows processors to support any arbitrary number of power states that are described by a package describing the register used, the type of C-state, latency and the power consumption.

Currently the Linux kernel supports up to ACPI C3, and if a system supports any further states they are ignored. This causes problems on some AMD hardware which can support up to ACPI C4.

AMD systems that support up to C3 will enter CPUOFF and VDDOFF when entering C3. Systems that support up to C4 will enter CPUOFF when going to C3 and will enter VDDOFF when entering into C4.

The existing semantics for bus mastering around C3 are also valid for C4, so instead of hardcoding to C3, map then >= C3. In the case of s2idle detect the deepest C-state supported and enter the deepest.

---
  drivers/acpi/processor_idle.c | 10 +++++++---
  include/acpi/actypes.h        |  5 +++--
  2 files changed, 10 insertions(+), 5 deletions(-)
drivers/acpi/processor_idle.c +7 -3
index 5dacf41d7cc0a..537b0119535ea 100644
@@ -444,6 +444,10 @@ static int acpi_processor_power_verify(struct acpi_processor *pr)
                 case ACPI_STATE_C3:
                         acpi_processor_power_verify_c3(pr, cx);
                         break;
+               case ACPI_STATE_C4:
+                       if (!cx->latency || !cx->address)
+                               break;
+                       cx->valid = 1;
                 }
                 if (!cx->valid)
                         continue;
@@ -685,7 +689,7 @@ static int __cpuidle acpi_idle_enter(struct cpuidle_device *dev,
                 return -EINVAL;
 
         if (cx->type != ACPI_STATE_C1) {
-               if (cx->type == ACPI_STATE_C3 && pr->flags.bm_check)
+               if (cx->type >= ACPI_STATE_C3 && pr->flags.bm_check)
                         return acpi_idle_enter_bm(drv, pr, cx, index);
 
                 /* C2 to C1 demotion. */
@@ -708,7 +712,7 @@ static int __cpuidle acpi_idle_enter_s2idle(struct cpuidle_device *dev,
  {
         struct acpi_processor_cx *cx = per_cpu(acpi_cstate[index], dev->cpu);
 
-       if (cx->type == ACPI_STATE_C3) {
+       if (cx->index == drv->state_count - 1) {
                 struct acpi_processor *pr = __this_cpu_read(processors);
 
                 if (unlikely(!pr))
@@ -754,7 +758,7 @@ static int acpi_processor_setup_cpuidle_cx(struct acpi_processor *pr,
                 if (lapic_timer_needs_broadcast(pr, cx))
                         state->flags |= CPUIDLE_FLAG_TIMER_STOP;
 
-               if (cx->type == ACPI_STATE_C3) {
+               if (cx->type >= ACPI_STATE_C3) {
                         state->flags |= CPUIDLE_FLAG_TLB_FLUSHED;
                         if (pr->flags.bm_check)
                                 state->flags |= CPUIDLE_FLAG_RCU_IDLE;
include/acpi/actypes.h +0 -0
index 8fe893d776dde..6c9f472af482c 100644
@@ -600,8 +600,9 @@ typedef u64 acpi_integer;
  #define ACPI_STATE_C1                   (u8) 1
  #define ACPI_STATE_C2                   (u8) 2
Quoted 8 lines

include/acpi/actypes.h typedef u64 acpi_integer; · line 605

Quoted 6 lines
include/acpi/actypes.h +3 -2
  #define ACPI_STATE_C3                   (u8) 3
-#define ACPI_C_STATES_MAX               ACPI_STATE_C3
-#define ACPI_C_STATE_COUNT              4
+#define ACPI_STATE_C4                   (u8) 4
+#define ACPI_C_STATES_MAX               ACPI_STATE_C4
+#define ACPI_C_STATE_COUNT              5

Don't you need to update ACPICA accordingly?

Thanks, I've got a PR opened.

Quoted 8 lines

include/acpi/actypes.h typedef u64 acpi_integer;

Quoted 5 lines
include/acpi/actypes.h +0 -1
  /*
   * Sleep type invalid value
--

The current point in the cycle is kind of unsuitable for changes like this. I'd prefer to revisit it after 6.18-rc1.

OK! I wanted this out a few weeks ago but was waiting for positive Q/A results.

I'll rebase/resend after 6.18-rc1.