intel_idle: native hardware cpuidle driver for latest Intel processors
authorLen Brown <len.brown@intel.com>
Mon, 8 Mar 2010 19:07:30 +0000 (14:07 -0500)
committerLen Brown <len.brown@intel.com>
Fri, 28 May 2010 18:26:20 +0000 (14:26 -0400)
This EXPERIMENTAL driver supersedes acpi_idle on
Intel Atom Processors, Intel Core i3/i5/i7 Processors
and associated Intel Xeon processors.

It does not support the Intel Core2 processor or earlier.

For kernels configured with ACPI, CONFIG_INTEL_IDLE=y
allows intel_idle to probe before the ACPI processor driver.
Booting with "intel_idle.max_cstate=0" disables intel_idle
and the system will fall back on ACPI's "acpi_idle".

Typical Linux distributions load ACPI processor module early,
making CONFIG_INTEL_IDLE=m not easily useful on ACPI platforms.

intel_idle probes all processors at module_init time.
Processors that are hot-added later will be limited
to using C1 in idle.

Signed-off-by: Len Brown <len.brown@intel.com>
MAINTAINERS
drivers/Makefile
drivers/acpi/processor_driver.c
drivers/idle/Kconfig
drivers/idle/Makefile
drivers/idle/intel_idle.c [new file with mode: 0755]

index d329b053a7185b2a31c3e92b95168b94f90b9618..276e79bb462cba6cff8795d8c09fcd7e0e4b5db2 100644 (file)
@@ -2850,6 +2850,13 @@ T:       git git://git.kernel.org/pub/scm/linux/kernel/git/dtor/input.git
 S:     Maintained
 F:     drivers/input/
 
+INTEL IDLE DRIVER
+M:     Len Brown <lenb@kernel.org>
+L:     linux-pm@lists.linux-foundation.org
+T:     git git://git.kernel.org/pub/scm/linux/kernel/git/lenb/linux-idle-2.6.git
+S:     Supported
+F:     drivers/idle/intel_idle.c
+
 INTEL FRAMEBUFFER DRIVER (excluding 810 and 815)
 M:     Maik Broemme <mbroemme@plusserver.de>
 L:     linux-fbdev@vger.kernel.org
index f42a03029b7c01d58193ee4b4b63d2660852a46c..91874e048552364280d7905aa817e957be2f9c48 100644 (file)
@@ -10,6 +10,7 @@ obj-$(CONFIG_PCI)             += pci/
 obj-$(CONFIG_PARISC)           += parisc/
 obj-$(CONFIG_RAPIDIO)          += rapidio/
 obj-y                          += video/
+obj-y                          += idle/
 obj-$(CONFIG_ACPI)             += acpi/
 obj-$(CONFIG_SFI)              += sfi/
 # PnP must come after ACPI since it will eventually need to check if acpi
@@ -91,7 +92,6 @@ obj-$(CONFIG_EISA)            += eisa/
 obj-y                          += lguest/
 obj-$(CONFIG_CPU_FREQ)         += cpufreq/
 obj-$(CONFIG_CPU_IDLE)         += cpuidle/
-obj-y                          += idle/
 obj-$(CONFIG_MMC)              += mmc/
 obj-$(CONFIG_MEMSTICK)         += memstick/
 obj-$(CONFIG_NEW_LEDS)         += leds/
index deefa8591746c569539478e3af094e3c4a388a88..b1034a9ada4e74da88970cbdc65dd9c1959d9988 100644 (file)
@@ -922,9 +922,13 @@ static int __init acpi_processor_init(void)
                return -ENOMEM;
 #endif
 
-       if (!cpuidle_register_driver(&acpi_idle_driver))
+       if (!cpuidle_register_driver(&acpi_idle_driver)) {
                printk(KERN_DEBUG "ACPI: %s registered with cpuidle\n",
                        acpi_idle_driver.name);
+       } else {
+               printk(KERN_DEBUG "ACPI: acpi_idle yielding to %s",
+                       cpuidle_get_driver()->name);
+       }
 
        result = acpi_bus_register_driver(&acpi_processor_driver);
        if (result < 0)
index f15e90a453d1c416fb007b046f41ee6bc0c4d866..fb5c5186d4aae6b838732cea9e370409a1198310 100644 (file)
@@ -1,3 +1,14 @@
+config INTEL_IDLE
+       tristate "Cpuidle Driver for Intel Processors"
+       depends on CPU_IDLE
+       depends on X86
+       depends on CPU_SUP_INTEL
+       depends on EXPERIMENTAL
+       help
+         Enable intel_idle, a cpuidle driver that includes knowledge of
+         native Intel hardware idle features.  The acpi_idle driver
+         can be configured at the same time, in order to handle
+         processors intel_idle does not support.
 
 menu "Memory power savings"
 depends on X86_64
index 5f68fc377e21d89679bd4867aa03f583deda2896..23d295cf10f2a785eefb17224bf13aa154970e1d 100644 (file)
@@ -1,2 +1,3 @@
 obj-$(CONFIG_I7300_IDLE)                       += i7300_idle.o
+obj-$(CONFIG_INTEL_IDLE)                       += intel_idle.o
 
diff --git a/drivers/idle/intel_idle.c b/drivers/idle/intel_idle.c
new file mode 100755 (executable)
index 0000000..54f0fb4
--- /dev/null
@@ -0,0 +1,461 @@
+/*
+ * intel_idle.c - native hardware idle loop for modern Intel processors
+ *
+ * Copyright (c) 2010, Intel Corporation.
+ * Len Brown <len.brown@intel.com>
+ *
+ * This program is free software; you can redistribute it and/or modify it
+ * under the terms and conditions of the GNU General Public License,
+ * version 2, as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
+ * more details.
+ *
+ * You should have received a copy of the GNU General Public License along with
+ * this program; if not, write to the Free Software Foundation, Inc.,
+ * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
+ */
+
+/*
+ * intel_idle is a cpuidle driver that loads on specific Intel processors
+ * in lieu of the legacy ACPI processor_idle driver.  The intent is to
+ * make Linux more efficient on these processors, as intel_idle knows
+ * more than ACPI, as well as make Linux more immune to ACPI BIOS bugs.
+ */
+
+/*
+ * Design Assumptions
+ *
+ * All CPUs have same idle states as boot CPU
+ *
+ * Chipset BM_STS (bus master status) bit is a NOP
+ *     for preventing entry into deep C-stats
+ */
+
+/*
+ * Known limitations
+ *
+ * The driver currently initializes for_each_online_cpu() upon modprobe.
+ * It it unaware of subsequent processors hot-added to the system.
+ * This means that if you boot with maxcpus=n and later online
+ * processors above n, those processors will use C1 only.
+ *
+ * ACPI has a .suspend hack to turn off deep c-statees during suspend
+ * to avoid complications with the lapic timer workaround.
+ * Have not seen issues with suspend, but may need same workaround here.
+ *
+ * There is currently no kernel-based automatic probing/loading mechanism
+ * if the driver is built as a module.
+ */
+
+/* un-comment DEBUG to enable pr_debug() statements */
+#define DEBUG
+
+#include <linux/kernel.h>
+#include <linux/cpuidle.h>
+#include <linux/clockchips.h>
+#include <linux/hrtimer.h>     /* ktime_get_real() */
+#include <trace/events/power.h>
+#include <linux/sched.h>
+
+#define INTEL_IDLE_VERSION "0.4"
+#define PREFIX "intel_idle: "
+
+#define MWAIT_SUBSTATE_MASK    (0xf)
+#define MWAIT_CSTATE_MASK      (0xf)
+#define MWAIT_SUBSTATE_SIZE    (4)
+#define MWAIT_MAX_NUM_CSTATES  8
+#define CPUID_MWAIT_LEAF (5)
+#define CPUID5_ECX_EXTENSIONS_SUPPORTED (0x1)
+#define CPUID5_ECX_INTERRUPT_BREAK     (0x2)
+
+static struct cpuidle_driver intel_idle_driver = {
+       .name = "intel_idle",
+       .owner = THIS_MODULE,
+};
+/* intel_idle.max_cstate=0 disables driver */
+static int max_cstate = MWAIT_MAX_NUM_CSTATES - 1;
+static int power_policy = 7; /* 0 = max perf; 15 = max powersave */
+
+static unsigned int substates;
+static int (*choose_substate)(int);
+
+/* Reliable LAPIC Timer States, bit 1 for C1 etc.  */
+static unsigned int lapic_timer_reliable_states;
+
+static struct cpuidle_device *intel_idle_cpuidle_devices;
+static int intel_idle(struct cpuidle_device *dev, struct cpuidle_state *state);
+
+static struct cpuidle_state *cpuidle_state_table;
+
+/*
+ * States are indexed by the cstate number,
+ * which is also the index into the MWAIT hint array.
+ * Thus C0 is a dummy.
+ */
+static struct cpuidle_state nehalem_cstates[MWAIT_MAX_NUM_CSTATES] = {
+       { /* MWAIT C0 */ },
+       { /* MWAIT C1 */
+               .name = "NHM-C1",
+               .desc = "MWAIT 0x00",
+               .driver_data = (void *) 0x00,
+               .flags = CPUIDLE_FLAG_TIME_VALID,
+               .exit_latency = 3,
+               .power_usage = 1000,
+               .target_residency = 6,
+               .enter = &intel_idle },
+       { /* MWAIT C2 */
+               .name = "NHM-C3",
+               .desc = "MWAIT 0x10",
+               .driver_data = (void *) 0x10,
+               .flags = CPUIDLE_FLAG_TIME_VALID,
+               .exit_latency = 20,
+               .power_usage = 500,
+               .target_residency = 80,
+               .enter = &intel_idle },
+       { /* MWAIT C3 */
+               .name = "NHM-C6",
+               .desc = "MWAIT 0x20",
+               .driver_data = (void *) 0x20,
+               .flags = CPUIDLE_FLAG_TIME_VALID,
+               .exit_latency = 200,
+               .power_usage = 350,
+               .target_residency = 800,
+               .enter = &intel_idle },
+};
+
+static struct cpuidle_state atom_cstates[MWAIT_MAX_NUM_CSTATES] = {
+       { /* MWAIT C0 */ },
+       { /* MWAIT C1 */
+               .name = "ATM-C1",
+               .desc = "MWAIT 0x00",
+               .driver_data = (void *) 0x00,
+               .flags = CPUIDLE_FLAG_TIME_VALID,
+               .exit_latency = 1,
+               .power_usage = 1000,
+               .target_residency = 4,
+               .enter = &intel_idle },
+       { /* MWAIT C2 */
+               .name = "ATM-C2",
+               .desc = "MWAIT 0x10",
+               .driver_data = (void *) 0x10,
+               .flags = CPUIDLE_FLAG_TIME_VALID,
+               .exit_latency = 20,
+               .power_usage = 500,
+               .target_residency = 80,
+               .enter = &intel_idle },
+       { /* MWAIT C3 */ },
+       { /* MWAIT C4 */
+               .name = "ATM-C4",
+               .desc = "MWAIT 0x30",
+               .driver_data = (void *) 0x30,
+               .flags = CPUIDLE_FLAG_TIME_VALID,
+               .exit_latency = 100,
+               .power_usage = 250,
+               .target_residency = 400,
+               .enter = &intel_idle },
+       { /* MWAIT C5 */ },
+       { /* MWAIT C6 */
+               .name = "ATM-C6",
+               .desc = "MWAIT 0x40",
+               .driver_data = (void *) 0x40,
+               .flags = CPUIDLE_FLAG_TIME_VALID,
+               .exit_latency = 200,
+               .power_usage = 150,
+               .target_residency = 800,
+               .enter = NULL },        /* disabled */
+};
+
+/*
+ * choose_tunable_substate()
+ *
+ * Run-time decision on which C-state substate to invoke
+ * If power_policy = 0, choose shallowest substate (0)
+ * If power_policy = 15, choose deepest substate
+ * If power_policy = middle, choose middle substate etc.
+ */
+static int choose_tunable_substate(int cstate)
+{
+       unsigned int num_substates;
+       unsigned int substate_choice;
+
+       power_policy &= 0xF;    /* valid range: 0-15 */
+       cstate &= 7;    /* valid range: 0-7 */
+
+       num_substates = (substates >> ((cstate) * 4)) & MWAIT_SUBSTATE_MASK;
+
+       if (num_substates <= 1)
+               return 0;
+
+       substate_choice = ((power_policy + (power_policy + 1) *
+                               (num_substates - 1)) / 16);
+
+       return substate_choice;
+}
+
+/*
+ * choose_zero_substate()
+ */
+static int choose_zero_substate(int cstate)
+{
+       return 0;
+}
+
+/**
+ * intel_idle
+ * @dev: cpuidle_device
+ * @state: cpuidle state
+ *
+ */
+static int intel_idle(struct cpuidle_device *dev, struct cpuidle_state *state)
+{
+       unsigned long ecx = 1; /* break on interrupt flag */
+       unsigned long eax = (unsigned long)cpuidle_get_statedata(state);
+       unsigned int cstate;
+       ktime_t kt_before, kt_after;
+       s64 usec_delta;
+       int cpu = smp_processor_id();
+
+       cstate = (((eax) >> MWAIT_SUBSTATE_SIZE) & MWAIT_CSTATE_MASK) + 1;
+
+       eax = eax + (choose_substate)(cstate);
+
+       local_irq_disable();
+
+       if (!(lapic_timer_reliable_states & (1 << (cstate))))
+               clockevents_notify(CLOCK_EVT_NOTIFY_BROADCAST_ENTER, &cpu);
+
+       kt_before = ktime_get_real();
+
+       stop_critical_timings();
+#ifndef MODULE
+       trace_power_start(POWER_CSTATE, (eax >> 4) + 1);
+#endif
+       if (!need_resched()) {
+
+               __monitor((void *)&current_thread_info()->flags, 0, 0);
+               smp_mb();
+               if (!need_resched())
+                       __mwait(eax, ecx);
+       }
+
+       start_critical_timings();
+
+       kt_after = ktime_get_real();
+       usec_delta = ktime_to_us(ktime_sub(kt_after, kt_before));
+
+       local_irq_enable();
+
+       if (!(lapic_timer_reliable_states & (1 << (cstate))))
+               clockevents_notify(CLOCK_EVT_NOTIFY_BROADCAST_EXIT, &cpu);
+
+       return usec_delta;
+}
+
+/*
+ * intel_idle_probe()
+ */
+static int intel_idle_probe(void)
+{
+       unsigned int eax, ebx, ecx, edx;
+
+       if (max_cstate == 0) {
+               pr_debug(PREFIX "disabled\n");
+               return -EPERM;
+       }
+
+       if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL)
+               return -ENODEV;
+
+       if (!boot_cpu_has(X86_FEATURE_MWAIT))
+               return -ENODEV;
+
+       if (boot_cpu_data.cpuid_level < CPUID_MWAIT_LEAF)
+               return -ENODEV;
+
+       cpuid(CPUID_MWAIT_LEAF, &eax, &ebx, &ecx, &edx);
+
+       if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED) ||
+               !(ecx & CPUID5_ECX_INTERRUPT_BREAK))
+                       return -ENODEV;
+#ifdef DEBUG
+       if (substates == 0)     /* can over-ride via modparam */
+#endif
+               substates = edx;
+
+       pr_debug(PREFIX "MWAIT substates: 0x%x\n", substates);
+
+       if (boot_cpu_has(X86_FEATURE_ARAT))     /* Always Reliable APIC Timer */
+               lapic_timer_reliable_states = 0xFFFFFFFF;
+
+       if (boot_cpu_data.x86 != 6)     /* family 6 */
+               return -ENODEV;
+
+       switch (boot_cpu_data.x86_model) {
+
+       case 0x1A:      /* Core i7, Xeon 5500 series */
+       case 0x1E:      /* Core i7 and i5 Processor - Lynnfield Jasper Forest */
+       case 0x1F:      /* Core i7 and i5 Processor - Nehalem */
+       case 0x2E:      /* Nehalem-EX Xeon */
+               lapic_timer_reliable_states = (1 << 1);  /* C1 */
+
+       case 0x25:      /* Westmere */
+       case 0x2C:      /* Westmere */
+               cpuidle_state_table = nehalem_cstates;
+               choose_substate = choose_tunable_substate;
+               break;
+
+       case 0x1C:      /* 28 - Atom Processor */
+               lapic_timer_reliable_states = (1 << 2) | (1 << 1); /* C2, C1 */
+               cpuidle_state_table = atom_cstates;
+               choose_substate = choose_zero_substate;
+               break;
+#ifdef FUTURE_USE
+       case 0x17:      /* 23 - Core 2 Duo */
+               lapic_timer_reliable_states = (1 << 2) | (1 << 1); /* C2, C1 */
+#endif
+
+       default:
+               pr_debug(PREFIX "does not run on family %d model %d\n",
+                       boot_cpu_data.x86, boot_cpu_data.x86_model);
+               return -ENODEV;
+       }
+
+       pr_debug(PREFIX "v" INTEL_IDLE_VERSION
+               " model 0x%X\n", boot_cpu_data.x86_model);
+
+       pr_debug(PREFIX "lapic_timer_reliable_states 0x%x\n",
+               lapic_timer_reliable_states);
+       return 0;
+}
+
+/*
+ * intel_idle_cpuidle_devices_uninit()
+ * unregister, free cpuidle_devices
+ */
+static void intel_idle_cpuidle_devices_uninit(void)
+{
+       int i;
+       struct cpuidle_device *dev;
+
+       for_each_online_cpu(i) {
+               dev = per_cpu_ptr(intel_idle_cpuidle_devices, i);
+               cpuidle_unregister_device(dev);
+       }
+
+       free_percpu(intel_idle_cpuidle_devices);
+       return;
+}
+/*
+ * intel_idle_cpuidle_devices_init()
+ * allocate, initialize, register cpuidle_devices
+ */
+static int intel_idle_cpuidle_devices_init(void)
+{
+       int i, cstate;
+       struct cpuidle_device *dev;
+
+       intel_idle_cpuidle_devices = alloc_percpu(struct cpuidle_device);
+       if (intel_idle_cpuidle_devices == NULL)
+               return -ENOMEM;
+
+       for_each_online_cpu(i) {
+               dev = per_cpu_ptr(intel_idle_cpuidle_devices, i);
+
+               dev->state_count = 1;
+
+               for (cstate = 1; cstate < MWAIT_MAX_NUM_CSTATES; ++cstate) {
+                       int num_substates;
+
+                       if (cstate > max_cstate) {
+                               printk(PREFIX "max_cstate %d reached\n",
+                                       max_cstate);
+                               break;
+                       }
+
+                       /* does the state exist in CPUID.MWAIT? */
+                       num_substates = (substates >> ((cstate) * 4))
+                                               & MWAIT_SUBSTATE_MASK;
+                       if (num_substates == 0)
+                               continue;
+                       /* is the state not enabled? */
+                       if (cpuidle_state_table[cstate].enter == NULL) {
+                               /* does the driver not know about the state? */
+                               if (*cpuidle_state_table[cstate].name == '\0')
+                                       pr_debug(PREFIX "unaware of model 0x%x"
+                                               " MWAIT %d please"
+                                               " contact lenb@kernel.org",
+                                       boot_cpu_data.x86_model, cstate);
+                               continue;
+                       }
+
+                       if ((cstate > 2) &&
+                               !boot_cpu_has(X86_FEATURE_NONSTOP_TSC))
+                               mark_tsc_unstable("TSC halts in idle"
+                                       " states deeper than C2");
+
+                       dev->states[dev->state_count] = /* structure copy */
+                               cpuidle_state_table[cstate];
+
+                       dev->state_count += 1;
+               }
+
+               dev->cpu = i;
+               if (cpuidle_register_device(dev)) {
+                       pr_debug(PREFIX "cpuidle_register_device %d failed!\n",
+                                i);
+                       intel_idle_cpuidle_devices_uninit();
+                       return -EIO;
+               }
+       }
+
+       return 0;
+}
+
+
+static int __init intel_idle_init(void)
+{
+       int retval;
+
+       retval = intel_idle_probe();
+       if (retval)
+               return retval;
+
+       retval = cpuidle_register_driver(&intel_idle_driver);
+       if (retval) {
+               printk(KERN_DEBUG PREFIX "intel_idle yielding to %s",
+                       cpuidle_get_driver()->name);
+               return retval;
+       }
+
+       retval = intel_idle_cpuidle_devices_init();
+       if (retval) {
+               cpuidle_unregister_driver(&intel_idle_driver);
+               return retval;
+       }
+
+       return 0;
+}
+
+static void __exit intel_idle_exit(void)
+{
+       intel_idle_cpuidle_devices_uninit();
+       cpuidle_unregister_driver(&intel_idle_driver);
+
+       return;
+}
+
+module_init(intel_idle_init);
+module_exit(intel_idle_exit);
+
+module_param(power_policy, int, 0644);
+module_param(max_cstate, int, 0444);
+#ifdef DEBUG
+module_param(substates, int, 0444);
+#endif
+
+MODULE_AUTHOR("Len Brown <len.brown@intel.com>");
+MODULE_DESCRIPTION("Cpuidle driver for Intel Hardware v" INTEL_IDLE_VERSION);
+MODULE_LICENSE("GPL");