ACPI: create Processor Aggregator Device driver
Shaohua Li [Mon, 27 Jul 2009 22:11:02 +0000 (18:11 -0400)]
ACPI 4.0 created the logical "processor aggregator device" as
a mechinism for platforms to ask the OS to force otherwise busy
processors to enter (power saving) idle.

The intent is to lower power consumption to ride-out
transient electrical and thermal emergencies,
rather than powering off the server.

On platforms that can save more power/performance via P-states,
the platform will first exhaust P-states before forcing idle.
However, the relative benefit of P-states vs. idle states
is platform dependent, and thus this driver need not know
or care about it.

This driver does not use the kernel's CPU hot-plug mechanism
because after the transient emergency is over, the system must
be returned to its normal state, and hotplug would permanently
break both cpusets and binding.

So to force idle, the driver creates a power saving thread.
The scheduler will migrate the thread to the preferred CPU.
The thread has max priority and has SCHED_RR policy,
so it can occupy one CPU.  To save power, the thread will
invoke the deep C-state entry instructions.

To avoid starvation, the thread will sleep 5% of the time
time for every second (current RT scheduler has threshold
to avoid starvation, but if other CPUs are idle,
the CPU can borrow CPU timer from other,
which makes the mechanism not work here)

Vaidyanathan Srinivasan has proposed scheduler enhancements
to allow injecting idle time into the system.  This driver doesn't
depend on those enhancements, but could cut over to them
when they are available.

Peter Z. does not favor upstreaming this driver until
the those scheduler enhancements are in place.  However,
we favor upstreaming this driver now because it is useful
now, and can be enhanced over time.

Signed-off-by: Shaohua Li <shaohua.li@intel.com>
NACKed-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Vaidyanathan Srinivasan <svaidy@linux.vnet.ibm.com>
Signed-off-by: Len Brown <len.brown@intel.com>

MAINTAINERS
drivers/acpi/Kconfig
drivers/acpi/Makefile
drivers/acpi/acpi_pad.c [new file with mode: 0644]

index ebc2691..082df56 100644 (file)
@@ -288,6 +288,14 @@ L: linux-pci@vger.kernel.org
 S:     Supported
 F:     drivers/pci/hotplug/acpi*
 
+ACPI PROCESSOR AGGREGATOR DRIVER
+P:     Shaohua Li
+M:     shaohua.li@intel.com
+L:     linux-acpi@vger.kernel.org
+W:     http://www.lesswatts.org/projects/acpi/
+S:     Supported
+F:     drivers/acpi/acpi_pad.c
+
 ACPI THERMAL DRIVER
 P:     Zhang Rui
 M:     rui.zhang@intel.com
index 7ec7d88..13531ed 100644 (file)
@@ -196,6 +196,17 @@ config ACPI_HOTPLUG_CPU
        select ACPI_CONTAINER
        default y
 
+config ACPI_PROCESSOR_AGGREGATOR
+       tristate "Processor Aggregator"
+       depends on ACPI_PROCESSOR
+       depends on EXPERIMENTAL
+       help
+         ACPI 4.0 defines processor Aggregator, which enables OS to perform
+         specfic processor configuration and control that applies to all
+         processors in the platform. Currently only logical processor idling
+         is defined, which is to reduce power consumption. This driver
+         support the new device.
+
 config ACPI_THERMAL
        tristate "Thermal Zone"
        depends on ACPI_PROCESSOR
index 03a985b..8dae61c 100644 (file)
@@ -61,3 +61,5 @@ obj-$(CONFIG_ACPI_SBS)                += sbs.o
 processor-y                    := processor_core.o processor_throttling.o
 processor-y                    += processor_idle.o processor_thermal.o
 processor-$(CONFIG_CPU_FREQ)   += processor_perflib.o
+
+obj-$(CONFIG_ACPI_PROCESSOR_AGGREGATOR) += acpi_pad.o
diff --git a/drivers/acpi/acpi_pad.c b/drivers/acpi/acpi_pad.c
new file mode 100644 (file)
index 0000000..0d2cdb8
--- /dev/null
@@ -0,0 +1,514 @@
+/*
+ * acpi_pad.c ACPI Processor Aggregator Driver
+ *
+ * Copyright (c) 2009, Intel Corporation.
+ *
+ * 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.
+ *
+ */
+
+#include <linux/kernel.h>
+#include <linux/cpumask.h>
+#include <linux/module.h>
+#include <linux/init.h>
+#include <linux/types.h>
+#include <linux/kthread.h>
+#include <linux/freezer.h>
+#include <linux/cpu.h>
+#include <linux/clockchips.h>
+#include <acpi/acpi_bus.h>
+#include <acpi/acpi_drivers.h>
+
+#define ACPI_PROCESSOR_AGGREGATOR_CLASS        "processor_aggregator"
+#define ACPI_PROCESSOR_AGGREGATOR_DEVICE_NAME "Processor Aggregator"
+#define ACPI_PROCESSOR_AGGREGATOR_NOTIFY 0x80
+static DEFINE_MUTEX(isolated_cpus_lock);
+
+#define MWAIT_SUBSTATE_MASK    (0xf)
+#define MWAIT_CSTATE_MASK      (0xf)
+#define MWAIT_SUBSTATE_SIZE    (4)
+#define CPUID_MWAIT_LEAF (5)
+#define CPUID5_ECX_EXTENSIONS_SUPPORTED (0x1)
+#define CPUID5_ECX_INTERRUPT_BREAK     (0x2)
+static unsigned long power_saving_mwait_eax;
+static void power_saving_mwait_init(void)
+{
+       unsigned int eax, ebx, ecx, edx;
+       unsigned int highest_cstate = 0;
+       unsigned int highest_subcstate = 0;
+       int i;
+
+       if (!boot_cpu_has(X86_FEATURE_MWAIT))
+               return;
+       if (boot_cpu_data.cpuid_level < CPUID_MWAIT_LEAF)
+               return;
+
+       cpuid(CPUID_MWAIT_LEAF, &eax, &ebx, &ecx, &edx);
+
+       if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED) ||
+           !(ecx & CPUID5_ECX_INTERRUPT_BREAK))
+               return;
+
+       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;
+               }
+       }
+       power_saving_mwait_eax = (highest_cstate << MWAIT_SUBSTATE_SIZE) |
+               (highest_subcstate - 1);
+
+       for_each_online_cpu(i)
+               clockevents_notify(CLOCK_EVT_NOTIFY_BROADCAST_ON, &i);
+
+#if defined(CONFIG_GENERIC_TIME) && defined(CONFIG_X86)
+       switch (boot_cpu_data.x86_vendor) {
+       case X86_VENDOR_AMD:
+       case X86_VENDOR_INTEL:
+               /*
+                * AMD Fam10h TSC will tick in all
+                * C/P/S0/S1 states when this bit is set.
+                */
+               if (boot_cpu_has(X86_FEATURE_NONSTOP_TSC))
+                       return;
+
+               /*FALL THROUGH*/
+       default:
+               /* TSC could halt in idle, so notify users */
+               mark_tsc_unstable("TSC halts in idle");
+       }
+#endif
+}
+
+static unsigned long cpu_weight[NR_CPUS];
+static int tsk_in_cpu[NR_CPUS] = {[0 ... NR_CPUS-1] = -1};
+static DECLARE_BITMAP(pad_busy_cpus_bits, NR_CPUS);
+static void round_robin_cpu(unsigned int tsk_index)
+{
+       struct cpumask *pad_busy_cpus = to_cpumask(pad_busy_cpus_bits);
+       cpumask_var_t tmp;
+       int cpu;
+       unsigned long min_weight = -1, preferred_cpu;
+
+       if (!alloc_cpumask_var(&tmp, GFP_KERNEL))
+               return;
+
+       mutex_lock(&isolated_cpus_lock);
+       cpumask_clear(tmp);
+       for_each_cpu(cpu, pad_busy_cpus)
+               cpumask_or(tmp, tmp, topology_thread_cpumask(cpu));
+       cpumask_andnot(tmp, cpu_online_mask, tmp);
+       /* avoid HT sibilings if possible */
+       if (cpumask_empty(tmp))
+               cpumask_andnot(tmp, cpu_online_mask, pad_busy_cpus);
+       if (cpumask_empty(tmp)) {
+               mutex_unlock(&isolated_cpus_lock);
+               return;
+       }
+       for_each_cpu(cpu, tmp) {
+               if (cpu_weight[cpu] < min_weight) {
+                       min_weight = cpu_weight[cpu];
+                       preferred_cpu = cpu;
+               }
+       }
+
+       if (tsk_in_cpu[tsk_index] != -1)
+               cpumask_clear_cpu(tsk_in_cpu[tsk_index], pad_busy_cpus);
+       tsk_in_cpu[tsk_index] = preferred_cpu;
+       cpumask_set_cpu(preferred_cpu, pad_busy_cpus);
+       cpu_weight[preferred_cpu]++;
+       mutex_unlock(&isolated_cpus_lock);
+
+       set_cpus_allowed_ptr(current, cpumask_of(preferred_cpu));
+}
+
+static void exit_round_robin(unsigned int tsk_index)
+{
+       struct cpumask *pad_busy_cpus = to_cpumask(pad_busy_cpus_bits);
+       cpumask_clear_cpu(tsk_in_cpu[tsk_index], pad_busy_cpus);
+       tsk_in_cpu[tsk_index] = -1;
+}
+
+static unsigned int idle_pct = 5; /* percentage */
+static unsigned int round_robin_time = 10; /* second */
+static int power_saving_thread(void *data)
+{
+       struct sched_param param = {.sched_priority = 1};
+       int do_sleep;
+       unsigned int tsk_index = (unsigned long)data;
+       u64 last_jiffies = 0;
+
+       sched_setscheduler(current, SCHED_RR, &param);
+
+       while (!kthread_should_stop()) {
+               int cpu;
+               u64 expire_time;
+
+               try_to_freeze();
+
+               /* round robin to cpus */
+               if (last_jiffies + round_robin_time * HZ < jiffies) {
+                       last_jiffies = jiffies;
+                       round_robin_cpu(tsk_index);
+               }
+
+               do_sleep = 0;
+
+               current_thread_info()->status &= ~TS_POLLING;
+               /*
+                * TS_POLLING-cleared state must be visible before we test
+                * NEED_RESCHED:
+                */
+               smp_mb();
+
+               expire_time = jiffies + HZ * (100 - idle_pct) / 100;
+
+               while (!need_resched()) {
+                       local_irq_disable();
+                       cpu = smp_processor_id();
+                       clockevents_notify(CLOCK_EVT_NOTIFY_BROADCAST_ENTER,
+                               &cpu);
+                       stop_critical_timings();
+
+                       __monitor((void *)&current_thread_info()->flags, 0, 0);
+                       smp_mb();
+                       if (!need_resched())
+                               __mwait(power_saving_mwait_eax, 1);
+
+                       start_critical_timings();
+                       clockevents_notify(CLOCK_EVT_NOTIFY_BROADCAST_EXIT,
+                               &cpu);
+                       local_irq_enable();
+
+                       if (jiffies > expire_time) {
+                               do_sleep = 1;
+                               break;
+                       }
+               }
+
+               current_thread_info()->status |= TS_POLLING;
+
+               /*
+                * current sched_rt has threshold for rt task running time.
+                * When a rt task uses 95% CPU time, the rt thread will be
+                * scheduled out for 5% CPU time to not starve other tasks. But
+                * the mechanism only works when all CPUs have RT task running,
+                * as if one CPU hasn't RT task, RT task from other CPUs will
+                * borrow CPU time from this CPU and cause RT task use > 95%
+                * CPU time. To make 'avoid staration' work, takes a nap here.
+                */
+               if (do_sleep)
+                       schedule_timeout_killable(HZ * idle_pct / 100);
+       }
+
+       exit_round_robin(tsk_index);
+       return 0;
+}
+
+static struct task_struct *ps_tsks[NR_CPUS];
+static unsigned int ps_tsk_num;
+static int create_power_saving_task(void)
+{
+       ps_tsks[ps_tsk_num] = kthread_run(power_saving_thread,
+               (void *)(unsigned long)ps_tsk_num,
+               "power_saving/%d", ps_tsk_num);
+       if (ps_tsks[ps_tsk_num]) {
+               ps_tsk_num++;
+               return 0;
+       }
+       return -EINVAL;
+}
+
+static void destroy_power_saving_task(void)
+{
+       if (ps_tsk_num > 0) {
+               ps_tsk_num--;
+               kthread_stop(ps_tsks[ps_tsk_num]);
+       }
+}
+
+static void set_power_saving_task_num(unsigned int num)
+{
+       if (num > ps_tsk_num) {
+               while (ps_tsk_num < num) {
+                       if (create_power_saving_task())
+                               return;
+               }
+       } else if (num < ps_tsk_num) {
+               while (ps_tsk_num > num)
+                       destroy_power_saving_task();
+       }
+}
+
+static int acpi_pad_idle_cpus(unsigned int num_cpus)
+{
+       get_online_cpus();
+
+       num_cpus = min_t(unsigned int, num_cpus, num_online_cpus());
+       set_power_saving_task_num(num_cpus);
+
+       put_online_cpus();
+       return 0;
+}
+
+static uint32_t acpi_pad_idle_cpus_num(void)
+{
+       return ps_tsk_num;
+}
+
+static ssize_t acpi_pad_rrtime_store(struct device *dev,
+       struct device_attribute *attr, const char *buf, size_t count)
+{
+       unsigned long num;
+       if (strict_strtoul(buf, 0, &num))
+               return -EINVAL;
+       if (num < 1 || num >= 100)
+               return -EINVAL;
+       mutex_lock(&isolated_cpus_lock);
+       round_robin_time = num;
+       mutex_unlock(&isolated_cpus_lock);
+       return count;
+}
+
+static ssize_t acpi_pad_rrtime_show(struct device *dev,
+       struct device_attribute *attr, char *buf)
+{
+       return scnprintf(buf, PAGE_SIZE, "%d", round_robin_time);
+}
+static DEVICE_ATTR(rrtime, S_IRUGO|S_IWUSR,
+       acpi_pad_rrtime_show,
+       acpi_pad_rrtime_store);
+
+static ssize_t acpi_pad_idlepct_store(struct device *dev,
+       struct device_attribute *attr, const char *buf, size_t count)
+{
+       unsigned long num;
+       if (strict_strtoul(buf, 0, &num))
+               return -EINVAL;
+       if (num < 1 || num >= 100)
+               return -EINVAL;
+       mutex_lock(&isolated_cpus_lock);
+       idle_pct = num;
+       mutex_unlock(&isolated_cpus_lock);
+       return count;
+}
+
+static ssize_t acpi_pad_idlepct_show(struct device *dev,
+       struct device_attribute *attr, char *buf)
+{
+       return scnprintf(buf, PAGE_SIZE, "%d", idle_pct);
+}
+static DEVICE_ATTR(idlepct, S_IRUGO|S_IWUSR,
+       acpi_pad_idlepct_show,
+       acpi_pad_idlepct_store);
+
+static ssize_t acpi_pad_idlecpus_store(struct device *dev,
+       struct device_attribute *attr, const char *buf, size_t count)
+{
+       unsigned long num;
+       if (strict_strtoul(buf, 0, &num))
+               return -EINVAL;
+       mutex_lock(&isolated_cpus_lock);
+       acpi_pad_idle_cpus(num);
+       mutex_unlock(&isolated_cpus_lock);
+       return count;
+}
+
+static ssize_t acpi_pad_idlecpus_show(struct device *dev,
+       struct device_attribute *attr, char *buf)
+{
+       return cpumask_scnprintf(buf, PAGE_SIZE,
+               to_cpumask(pad_busy_cpus_bits));
+}
+static DEVICE_ATTR(idlecpus, S_IRUGO|S_IWUSR,
+       acpi_pad_idlecpus_show,
+       acpi_pad_idlecpus_store);
+
+static int acpi_pad_add_sysfs(struct acpi_device *device)
+{
+       int result;
+
+       result = device_create_file(&device->dev, &dev_attr_idlecpus);
+       if (result)
+               return -ENODEV;
+       result = device_create_file(&device->dev, &dev_attr_idlepct);
+       if (result) {
+               device_remove_file(&device->dev, &dev_attr_idlecpus);
+               return -ENODEV;
+       }
+       result = device_create_file(&device->dev, &dev_attr_rrtime);
+       if (result) {
+               device_remove_file(&device->dev, &dev_attr_idlecpus);
+               device_remove_file(&device->dev, &dev_attr_idlepct);
+               return -ENODEV;
+       }
+       return 0;
+}
+
+static void acpi_pad_remove_sysfs(struct acpi_device *device)
+{
+       device_remove_file(&device->dev, &dev_attr_idlecpus);
+       device_remove_file(&device->dev, &dev_attr_idlepct);
+       device_remove_file(&device->dev, &dev_attr_rrtime);
+}
+
+/* Query firmware how many CPUs should be idle */
+static int acpi_pad_pur(acpi_handle handle, int *num_cpus)
+{
+       struct acpi_buffer buffer = {ACPI_ALLOCATE_BUFFER, NULL};
+       acpi_status status;
+       union acpi_object *package;
+       int rev, num, ret = -EINVAL;
+
+       status = acpi_evaluate_object(handle, "_PUR", NULL, &buffer);
+       if (ACPI_FAILURE(status))
+               return -EINVAL;
+       package = buffer.pointer;
+       if (package->type != ACPI_TYPE_PACKAGE || package->package.count != 2)
+               goto out;
+       rev = package->package.elements[0].integer.value;
+       num = package->package.elements[1].integer.value;
+       if (rev != 1)
+               goto out;
+       *num_cpus = num;
+       ret = 0;
+out:
+       kfree(buffer.pointer);
+       return ret;
+}
+
+/* Notify firmware how many CPUs are idle */
+static void acpi_pad_ost(acpi_handle handle, int stat,
+       uint32_t idle_cpus)
+{
+       union acpi_object params[3] = {
+               {.type = ACPI_TYPE_INTEGER,},
+               {.type = ACPI_TYPE_INTEGER,},
+               {.type = ACPI_TYPE_BUFFER,},
+       };
+       struct acpi_object_list arg_list = {3, params};
+
+       params[0].integer.value = ACPI_PROCESSOR_AGGREGATOR_NOTIFY;
+       params[1].integer.value =  stat;
+       params[2].buffer.length = 4;
+       params[2].buffer.pointer = (void *)&idle_cpus;
+       acpi_evaluate_object(handle, "_OST", &arg_list, NULL);
+}
+
+static void acpi_pad_handle_notify(acpi_handle handle)
+{
+       int num_cpus, ret;
+       uint32_t idle_cpus;
+
+       mutex_lock(&isolated_cpus_lock);
+       if (acpi_pad_pur(handle, &num_cpus)) {
+               mutex_unlock(&isolated_cpus_lock);
+               return;
+       }
+       ret = acpi_pad_idle_cpus(num_cpus);
+       idle_cpus = acpi_pad_idle_cpus_num();
+       if (!ret)
+               acpi_pad_ost(handle, 0, idle_cpus);
+       else
+               acpi_pad_ost(handle, 1, 0);
+       mutex_unlock(&isolated_cpus_lock);
+}
+
+static void acpi_pad_notify(acpi_handle handle, u32 event,
+       void *data)
+{
+       struct acpi_device *device = data;
+
+       switch (event) {
+       case ACPI_PROCESSOR_AGGREGATOR_NOTIFY:
+               acpi_pad_handle_notify(handle);
+               acpi_bus_generate_proc_event(device, event, 0);
+               acpi_bus_generate_netlink_event(device->pnp.device_class,
+                       dev_name(&device->dev), event, 0);
+               break;
+       default:
+               printk(KERN_WARNING"Unsupported event [0x%x]\n", event);
+               break;
+       }
+}
+
+static int acpi_pad_add(struct acpi_device *device)
+{
+       acpi_status status;
+
+       strcpy(acpi_device_name(device), ACPI_PROCESSOR_AGGREGATOR_DEVICE_NAME);
+       strcpy(acpi_device_class(device), ACPI_PROCESSOR_AGGREGATOR_CLASS);
+
+       if (acpi_pad_add_sysfs(device))
+               return -ENODEV;
+
+       status = acpi_install_notify_handler(device->handle,
+               ACPI_DEVICE_NOTIFY, acpi_pad_notify, device);
+       if (ACPI_FAILURE(status)) {
+               acpi_pad_remove_sysfs(device);
+               return -ENODEV;
+       }
+
+       return 0;
+}
+
+static int acpi_pad_remove(struct acpi_device *device,
+       int type)
+{
+       mutex_lock(&isolated_cpus_lock);
+       acpi_pad_idle_cpus(0);
+       mutex_unlock(&isolated_cpus_lock);
+
+       acpi_remove_notify_handler(device->handle,
+               ACPI_DEVICE_NOTIFY, acpi_pad_notify);
+       acpi_pad_remove_sysfs(device);
+       return 0;
+}
+
+static const struct acpi_device_id pad_device_ids[] = {
+       {"ACPI000C", 0},
+       {"", 0},
+};
+MODULE_DEVICE_TABLE(acpi, pad_device_ids);
+
+static struct acpi_driver acpi_pad_driver = {
+       .name = "processor_aggregator",
+       .class = ACPI_PROCESSOR_AGGREGATOR_CLASS,
+       .ids = pad_device_ids,
+       .ops = {
+               .add = acpi_pad_add,
+               .remove = acpi_pad_remove,
+       },
+};
+
+static int __init acpi_pad_init(void)
+{
+       power_saving_mwait_init();
+       if (power_saving_mwait_eax == 0)
+               return -EINVAL;
+
+       return acpi_bus_register_driver(&acpi_pad_driver);
+}
+
+static void __exit acpi_pad_exit(void)
+{
+       acpi_bus_unregister_driver(&acpi_pad_driver);
+}
+
+module_init(acpi_pad_init);
+module_exit(acpi_pad_exit);
+MODULE_AUTHOR("Shaohua Li<shaohua.li@intel.com>");
+MODULE_DESCRIPTION("ACPI Processor Aggregator Driver");
+MODULE_LICENSE("GPL");