ACPI: Processor native C-states using MWAIT
authorVenkatesh Pallipadi <venkatesh.pallipadi@intel.com>
Mon, 25 Sep 2006 23:28:13 +0000 (16:28 -0700)
committerLen Brown <len.brown@intel.com>
Sat, 14 Oct 2006 04:35:39 +0000 (00:35 -0400)
Intel processors starting with the Core Duo support
support processor native C-state using the MWAIT instruction.
Refer: Intel Architecture Software Developer's Manual
http://www.intel.com/design/Pentium4/manuals/253668.htm

Platform firmware exports the support for Native C-state to OS using
ACPI _PDC and _CST methods.
Refer: Intel Processor Vendor-Specific ACPI: Interface Specification
http://www.intel.com/technology/iapc/acpi/downloads/302223.htm

With Processor Native C-state, we use 'MWAIT' instruction on the processor
to enter different C-states (C1, C2, C3).  We won't use the special IO
ports to enter C-state and no SMM mode etc required to enter C-state.
Overall this will mean better C-state support.

One major advantage of using MWAIT for all C-states is, with this and
"treat interrupt as break event" feature of MWAIT, we can now get accurate
timing for the time spent in C1, C2, ..  states.

Signed-off-by: Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Len Brown <len.brown@intel.com>
arch/i386/kernel/acpi/cstate.c
arch/i386/kernel/process.c
arch/x86_64/kernel/process.c
drivers/acpi/processor_idle.c
include/acpi/pdc_intel.h
include/acpi/processor.h
include/asm-i386/processor.h
include/asm-x86_64/processor.h

index 25db49ef1770aa11ce148f1d35789e85bc125151..20563e52c62248853e9d4fbc9af48461c27696bd 100644 (file)
@@ -10,6 +10,7 @@
 #include <linux/module.h>
 #include <linux/init.h>
 #include <linux/acpi.h>
+#include <linux/cpu.h>
 
 #include <acpi/processor.h>
 #include <asm/acpi.h>
@@ -41,5 +42,124 @@ void acpi_processor_power_init_bm_check(struct acpi_processor_flags *flags,
                flags->bm_check = 1;
        }
 }
-
 EXPORT_SYMBOL(acpi_processor_power_init_bm_check);
+
+/* The code below handles cstate entry with monitor-mwait pair on Intel*/
+
+struct cstate_entry_s {
+       struct {
+               unsigned int eax;
+               unsigned int ecx;
+       } states[ACPI_PROCESSOR_MAX_POWER];
+};
+static struct cstate_entry_s *cpu_cstate_entry;        /* per CPU ptr */
+
+static short mwait_supported[ACPI_PROCESSOR_MAX_POWER];
+
+#define MWAIT_SUBSTATE_MASK    (0xf)
+#define MWAIT_SUBSTATE_SIZE    (4)
+
+#define CPUID_MWAIT_LEAF (5)
+#define CPUID5_ECX_EXTENSIONS_SUPPORTED (0x1)
+#define CPUID5_ECX_INTERRUPT_BREAK     (0x2)
+
+#define MWAIT_ECX_INTERRUPT_BREAK      (0x1)
+
+#define NATIVE_CSTATE_BEYOND_HALT      (2)
+
+int acpi_processor_ffh_cstate_probe(unsigned int cpu,
+               struct acpi_processor_cx *cx, struct acpi_power_register *reg)
+{
+       struct cstate_entry_s *percpu_entry;
+       struct cpuinfo_x86 *c = cpu_data + cpu;
+
+       cpumask_t saved_mask;
+       int retval;
+       unsigned int eax, ebx, ecx, edx;
+       unsigned int edx_part;
+       unsigned int cstate_type; /* C-state type and not ACPI C-state type */
+       unsigned int num_cstate_subtype;
+
+       if (!cpu_cstate_entry || c->cpuid_level < CPUID_MWAIT_LEAF )
+               return -1;
+
+       if (reg->bit_offset != NATIVE_CSTATE_BEYOND_HALT)
+               return -1;
+
+       percpu_entry = per_cpu_ptr(cpu_cstate_entry, cpu);
+       percpu_entry->states[cx->index].eax = 0;
+       percpu_entry->states[cx->index].ecx = 0;
+
+       /* Make sure we are running on right CPU */
+       saved_mask = current->cpus_allowed;
+       retval = set_cpus_allowed(current, cpumask_of_cpu(cpu));
+       if (retval)
+               return -1;
+
+       cpuid(CPUID_MWAIT_LEAF, &eax, &ebx, &ecx, &edx);
+
+       /* Check whether this particular cx_type (in CST) is supported or not */
+       cstate_type = (cx->address >> MWAIT_SUBSTATE_SIZE) + 1;
+       edx_part = edx >> (cstate_type * MWAIT_SUBSTATE_SIZE);
+       num_cstate_subtype = edx_part & MWAIT_SUBSTATE_MASK;
+
+       retval = 0;
+       if (num_cstate_subtype < (cx->address & MWAIT_SUBSTATE_MASK)) {
+               retval = -1;
+               goto out;
+       }
+
+       /* mwait ecx extensions INTERRUPT_BREAK should be supported for C2/C3 */
+       if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED) ||
+           !(ecx & CPUID5_ECX_INTERRUPT_BREAK)) {
+               retval = -1;
+               goto out;
+       }
+       percpu_entry->states[cx->index].ecx = MWAIT_ECX_INTERRUPT_BREAK;
+
+       /* Use the hint in CST */
+       percpu_entry->states[cx->index].eax = cx->address;
+
+       if (!mwait_supported[cstate_type]) {
+               mwait_supported[cstate_type] = 1;
+               printk(KERN_DEBUG "Monitor-Mwait will be used to enter C-%d "
+                      "state\n", cx->type);
+       }
+
+out:
+       set_cpus_allowed(current, saved_mask);
+       return retval;
+}
+EXPORT_SYMBOL_GPL(acpi_processor_ffh_cstate_probe);
+
+void acpi_processor_ffh_cstate_enter(struct acpi_processor_cx *cx)
+{
+       unsigned int cpu = smp_processor_id();
+       struct cstate_entry_s *percpu_entry;
+
+       percpu_entry = per_cpu_ptr(cpu_cstate_entry, cpu);
+       mwait_idle_with_hints(percpu_entry->states[cx->index].eax,
+                             percpu_entry->states[cx->index].ecx);
+}
+EXPORT_SYMBOL_GPL(acpi_processor_ffh_cstate_enter);
+
+static int __init ffh_cstate_init(void)
+{
+       struct cpuinfo_x86 *c = &boot_cpu_data;
+       if (c->x86_vendor != X86_VENDOR_INTEL)
+               return -1;
+
+       cpu_cstate_entry = alloc_percpu(struct cstate_entry_s);
+       return 0;
+}
+
+static void __exit ffh_cstate_exit(void)
+{
+       if (cpu_cstate_entry) {
+               free_percpu(cpu_cstate_entry);
+               cpu_cstate_entry = NULL;
+       }
+}
+
+arch_initcall(ffh_cstate_init);
+__exitcall(ffh_cstate_exit);
index b0a07801d9df90abf71f923fa27e55269cacf319..57d375900afb0643cf64f10b52ed49462131d001 100644 (file)
@@ -236,20 +236,28 @@ EXPORT_SYMBOL_GPL(cpu_idle_wait);
  * We execute MONITOR against need_resched and enter optimized wait state
  * through MWAIT. Whenever someone changes need_resched, we would be woken
  * up from MWAIT (without an IPI).
+ *
+ * New with Core Duo processors, MWAIT can take some hints based on CPU
+ * capability.
  */
-static void mwait_idle(void)
+void mwait_idle_with_hints(unsigned long eax, unsigned long ecx)
 {
-       local_irq_enable();
-
-       while (!need_resched()) {
+       if (!need_resched()) {
                __monitor((void *)&current_thread_info()->flags, 0, 0);
                smp_mb();
-               if (need_resched())
-                       break;
-               __mwait(0, 0);
+               if (!need_resched())
+                       __mwait(eax, ecx);
        }
 }
 
+/* Default MONITOR/MWAIT with no hints, used for default C1 state */
+static void mwait_idle(void)
+{
+       local_irq_enable();
+       while (!need_resched())
+               mwait_idle_with_hints(0, 0);
+}
+
 void __devinit select_idle_routine(const struct cpuinfo_x86 *c)
 {
        if (cpu_has(c, X86_FEATURE_MWAIT)) {
index 5e95b257ee26208612b8f28229612e8ce6796660..49f7fac6229e5f4e920d58fcdec1fd5e49e8667b 100644 (file)
@@ -238,20 +238,28 @@ void cpu_idle (void)
  * We execute MONITOR against need_resched and enter optimized wait state
  * through MWAIT. Whenever someone changes need_resched, we would be woken
  * up from MWAIT (without an IPI).
+ *
+ * New with Core Duo processors, MWAIT can take some hints based on CPU
+ * capability.
  */
-static void mwait_idle(void)
+void mwait_idle_with_hints(unsigned long eax, unsigned long ecx)
 {
-       local_irq_enable();
-
-       while (!need_resched()) {
+       if (!need_resched()) {
                __monitor((void *)&current_thread_info()->flags, 0, 0);
                smp_mb();
-               if (need_resched())
-                       break;
-               __mwait(0, 0);
+               if (!need_resched())
+                       __mwait(eax, ecx);
        }
 }
 
+/* Default MONITOR/MWAIT with no hints, used for default C1 state */
+static void mwait_idle(void)
+{
+       local_irq_enable();
+       while (!need_resched())
+               mwait_idle_with_hints(0,0);
+}
+
 void __cpuinit select_idle_routine(const struct cpuinfo_x86 *c)
 {
        static int printed;
index 0a395fca843b46686b748ffe1583515cc863e9af..429a39dbd75d19475583da403bc18450b29b5ea7 100644 (file)
@@ -219,6 +219,23 @@ static void acpi_safe_halt(void)
 
 static atomic_t c3_cpu_count;
 
+/* Common C-state entry for C2, C3, .. */
+static void acpi_cstate_enter(struct acpi_processor_cx *cstate)
+{
+       if (cstate->space_id == ACPI_CSTATE_FFH) {
+               /* Call into architectural FFH based C-state */
+               acpi_processor_ffh_cstate_enter(cstate);
+       } else {
+               int unused;
+               /* IO port based C-state */
+               inb(cstate->address);
+               /* Dummy wait op - must do something useless after P_LVL2 read
+                  because chipsets cannot guarantee that STPCLK# signal
+                  gets asserted in time to freeze execution properly. */
+               unused = inl(acpi_fadt.xpm_tmr_blk.address);
+       }
+}
+
 static void acpi_processor_idle(void)
 {
        struct acpi_processor *pr = NULL;
@@ -361,11 +378,7 @@ static void acpi_processor_idle(void)
                /* Get start time (ticks) */
                t1 = inl(acpi_fadt.xpm_tmr_blk.address);
                /* Invoke C2 */
-               inb(cx->address);
-               /* Dummy wait op - must do something useless after P_LVL2 read
-                  because chipsets cannot guarantee that STPCLK# signal
-                  gets asserted in time to freeze execution properly. */
-               t2 = inl(acpi_fadt.xpm_tmr_blk.address);
+               acpi_cstate_enter(cx);
                /* Get end time (ticks) */
                t2 = inl(acpi_fadt.xpm_tmr_blk.address);
 
@@ -401,9 +414,7 @@ static void acpi_processor_idle(void)
                /* Get start time (ticks) */
                t1 = inl(acpi_fadt.xpm_tmr_blk.address);
                /* Invoke C3 */
-               inb(cx->address);
-               /* Dummy wait op (see above) */
-               t2 = inl(acpi_fadt.xpm_tmr_blk.address);
+               acpi_cstate_enter(cx);
                /* Get end time (ticks) */
                t2 = inl(acpi_fadt.xpm_tmr_blk.address);
                if (pr->flags.bm_check) {
@@ -628,20 +639,16 @@ static int acpi_processor_get_power_info_fadt(struct acpi_processor *pr)
        return 0;
 }
 
-static int acpi_processor_get_power_info_default_c1(struct acpi_processor *pr)
+static int acpi_processor_get_power_info_default(struct acpi_processor *pr)
 {
-
-       /* Zero initialize all the C-states info. */
-       memset(pr->power.states, 0, sizeof(pr->power.states));
-
-       /* set the first C-State to C1 */
-       pr->power.states[ACPI_STATE_C1].type = ACPI_STATE_C1;
-
-       /* the C0 state only exists as a filler in our array,
-        * and all processors need to support C1 */
+       if (!pr->power.states[ACPI_STATE_C1].valid) {
+               /* set the first C-State to C1 */
+               /* all processors need to support C1 */
+               pr->power.states[ACPI_STATE_C1].type = ACPI_STATE_C1;
+               pr->power.states[ACPI_STATE_C1].valid = 1;
+       }
+       /* the C0 state only exists as a filler in our array */
        pr->power.states[ACPI_STATE_C0].valid = 1;
-       pr->power.states[ACPI_STATE_C1].valid = 1;
-
        return 0;
 }
 
@@ -658,12 +665,7 @@ static int acpi_processor_get_power_info_cst(struct acpi_processor *pr)
        if (nocst)
                return -ENODEV;
 
-       current_count = 1;
-
-       /* Zero initialize C2 onwards and prepare for fresh CST lookup */
-       for (i = 2; i < ACPI_PROCESSOR_MAX_POWER; i++)
-               memset(&(pr->power.states[i]), 0, 
-                               sizeof(struct acpi_processor_cx));
+       current_count = 0;
 
        status = acpi_evaluate_object(pr->handle, "_CST", NULL, &buffer);
        if (ACPI_FAILURE(status)) {
@@ -718,22 +720,39 @@ static int acpi_processor_get_power_info_cst(struct acpi_processor *pr)
                    (reg->space_id != ACPI_ADR_SPACE_FIXED_HARDWARE))
                        continue;
 
-               cx.address = (reg->space_id == ACPI_ADR_SPACE_FIXED_HARDWARE) ?
-                   0 : reg->address;
-
                /* There should be an easy way to extract an integer... */
                obj = (union acpi_object *)&(element->package.elements[1]);
                if (obj->type != ACPI_TYPE_INTEGER)
                        continue;
 
                cx.type = obj->integer.value;
-
-               if ((cx.type != ACPI_STATE_C1) &&
-                   (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO))
-                       continue;
-
-               if ((cx.type < ACPI_STATE_C2) || (cx.type > ACPI_STATE_C3))
-                       continue;
+               /*
+                * Some buggy BIOSes won't list C1 in _CST -
+                * Let acpi_processor_get_power_info_default() handle them later
+                */
+               if (i == 1 && cx.type != ACPI_STATE_C1)
+                       current_count++;
+
+               cx.address = reg->address;
+               cx.index = current_count + 1;
+
+               cx.space_id = ACPI_CSTATE_SYSTEMIO;
+               if (reg->space_id == ACPI_ADR_SPACE_FIXED_HARDWARE) {
+                       if (acpi_processor_ffh_cstate_probe
+                                       (pr->id, &cx, reg) == 0) {
+                               cx.space_id = ACPI_CSTATE_FFH;
+                       } else if (cx.type != ACPI_STATE_C1) {
+                               /*
+                                * C1 is a special case where FIXED_HARDWARE
+                                * can be handled in non-MWAIT way as well.
+                                * In that case, save this _CST entry info.
+                                * That is, we retain space_id of SYSTEM_IO for
+                                * halt based C1.
+                                * Otherwise, ignore this info and continue.
+                                */
+                               continue;
+                       }
+               }
 
                obj = (union acpi_object *)&(element->package.elements[2]);
                if (obj->type != ACPI_TYPE_INTEGER)
@@ -938,12 +957,18 @@ static int acpi_processor_get_power_info(struct acpi_processor *pr)
        /* NOTE: the idle thread may not be running while calling
         * this function */
 
-       /* Adding C1 state */
-       acpi_processor_get_power_info_default_c1(pr);
+       /* Zero initialize all the C-states info. */
+       memset(pr->power.states, 0, sizeof(pr->power.states));
+
        result = acpi_processor_get_power_info_cst(pr);
        if (result == -ENODEV)
                acpi_processor_get_power_info_fadt(pr);
 
+       if (result)
+               return result;
+
+       acpi_processor_get_power_info_default(pr);
+
        pr->power.count = acpi_processor_power_verify(pr);
 
        /*
index c5472be6f3a2c41744471e61dafa0386bf038397..e72bfdd887f9b285ccbb3530be8096d61aeb74a8 100644 (file)
@@ -13,6 +13,7 @@
 #define ACPI_PDC_SMP_C_SWCOORD         (0x0040)
 #define ACPI_PDC_SMP_T_SWCOORD         (0x0080)
 #define ACPI_PDC_C_C1_FFH              (0x0100)
+#define ACPI_PDC_C_C2C3_FFH            (0x0200)
 
 #define ACPI_PDC_EST_CAPABILITY_SMP    (ACPI_PDC_SMP_C1PT | \
                                         ACPI_PDC_C_C1_HALT | \
                                         ACPI_PDC_SMP_P_SWCOORD | \
                                         ACPI_PDC_P_FFH)
 
-#define ACPI_PDC_C_CAPABILITY_SMP      (ACPI_PDC_SMP_C2C3 | \
-                                        ACPI_PDC_SMP_C1PT | \
-                                        ACPI_PDC_C_C1_HALT)
+#define ACPI_PDC_C_CAPABILITY_SMP      (ACPI_PDC_SMP_C2C3  | \
+                                        ACPI_PDC_SMP_C1PT  | \
+                                        ACPI_PDC_C_C1_HALT | \
+                                        ACPI_PDC_C_C1_FFH  | \
+                                        ACPI_PDC_C_C2C3_FFH)
 
 #endif                         /* __PDC_INTEL_H__ */
index 9dd5b75961f84c19ba9a731eb071f15950577d75..7798d2a9f793aaab8005c2a45d3362a35029ea5b 100644 (file)
@@ -29,6 +29,9 @@
 #define DOMAIN_COORD_TYPE_SW_ANY       0xfd
 #define DOMAIN_COORD_TYPE_HW_ALL       0xfe
 
+#define ACPI_CSTATE_SYSTEMIO   (0)
+#define ACPI_CSTATE_FFH                (1)
+
 /* Power Management */
 
 struct acpi_processor_cx;
@@ -58,6 +61,8 @@ struct acpi_processor_cx {
        u8 valid;
        u8 type;
        u32 address;
+       u8 space_id;
+       u8 index;
        u32 latency;
        u32 latency_ticks;
        u32 power;
@@ -206,6 +211,9 @@ void arch_acpi_processor_init_pdc(struct acpi_processor *pr);
 #ifdef ARCH_HAS_POWER_INIT
 void acpi_processor_power_init_bm_check(struct acpi_processor_flags *flags,
                                        unsigned int cpu);
+int acpi_processor_ffh_cstate_probe(unsigned int cpu,
+               struct acpi_processor_cx *cx, struct acpi_power_register *reg);
+void acpi_processor_ffh_cstate_enter(struct acpi_processor_cx *cstate);
 #else
 static inline void acpi_processor_power_init_bm_check(struct
                                                      acpi_processor_flags
@@ -214,6 +222,16 @@ static inline void acpi_processor_power_init_bm_check(struct
        flags->bm_check = 1;
        return;
 }
+static inline int acpi_processor_ffh_cstate_probe(unsigned int cpu,
+               struct acpi_processor_cx *cx, struct acpi_power_register *reg)
+{
+       return -1;
+}
+static inline void acpi_processor_ffh_cstate_enter(
+               struct acpi_processor_cx *cstate)
+{
+       return;
+}
 #endif
 
 /* in processor_perflib.c */
index 2277127696d264579687a2e693768a6e4d8ba206..e0ddca94d50c8f42c71e05e1796086fdf84ad0ef 100644 (file)
@@ -306,6 +306,8 @@ static inline void __mwait(unsigned long eax, unsigned long ecx)
                : :"a" (eax), "c" (ecx));
 }
 
+extern void mwait_idle_with_hints(unsigned long eax, unsigned long ecx);
+
 /* from system description table in BIOS.  Mostly for MCA use, but
 others may find it useful. */
 extern unsigned int machine_id;
index de9c3147ee4c3dddb639210dc17a9b74b406a040..cef17e0f828cc5d0080cf967d11e80a228f4d18b 100644 (file)
@@ -475,6 +475,8 @@ static inline void __mwait(unsigned long eax, unsigned long ecx)
                : :"a" (eax), "c" (ecx));
 }
 
+extern void mwait_idle_with_hints(unsigned long eax, unsigned long ecx);
+
 #define stack_current() \
 ({                                                             \
        struct thread_info *ti;                                 \