2 * acpi_pad.c ACPI Processor Aggregator Driver
4 * Copyright (c) 2009, Intel Corporation.
6 * This program is free software; you can redistribute it and/or modify it
7 * under the terms and conditions of the GNU General Public License,
8 * version 2, as published by the Free Software Foundation.
10 * This program is distributed in the hope it will be useful, but WITHOUT
11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
17 #include <linux/kernel.h>
18 #include <linux/cpumask.h>
19 #include <linux/module.h>
20 #include <linux/init.h>
21 #include <linux/types.h>
22 #include <linux/kthread.h>
23 #include <linux/freezer.h>
24 #include <linux/cpu.h>
25 #include <linux/tick.h>
26 #include <linux/slab.h>
27 #include <linux/acpi.h>
28 #include <asm/mwait.h>
30 #define ACPI_PROCESSOR_AGGREGATOR_CLASS "acpi_pad"
31 #define ACPI_PROCESSOR_AGGREGATOR_DEVICE_NAME "Processor Aggregator"
32 #define ACPI_PROCESSOR_AGGREGATOR_NOTIFY 0x80
33 static DEFINE_MUTEX(isolated_cpus_lock);
34 static DEFINE_MUTEX(round_robin_lock);
36 static unsigned long power_saving_mwait_eax;
38 static unsigned char tsc_detected_unstable;
39 static unsigned char tsc_marked_unstable;
41 static void power_saving_mwait_init(void)
43 unsigned int eax, ebx, ecx, edx;
44 unsigned int highest_cstate = 0;
45 unsigned int highest_subcstate = 0;
48 if (!boot_cpu_has(X86_FEATURE_MWAIT))
50 if (boot_cpu_data.cpuid_level < CPUID_MWAIT_LEAF)
53 cpuid(CPUID_MWAIT_LEAF, &eax, &ebx, &ecx, &edx);
55 if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED) ||
56 !(ecx & CPUID5_ECX_INTERRUPT_BREAK))
59 edx >>= MWAIT_SUBSTATE_SIZE;
60 for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) {
61 if (edx & MWAIT_SUBSTATE_MASK) {
63 highest_subcstate = edx & MWAIT_SUBSTATE_MASK;
66 power_saving_mwait_eax = (highest_cstate << MWAIT_SUBSTATE_SIZE) |
67 (highest_subcstate - 1);
69 #if defined(CONFIG_X86)
70 switch (boot_cpu_data.x86_vendor) {
72 case X86_VENDOR_INTEL:
74 * AMD Fam10h TSC will tick in all
75 * C/P/S0/S1 states when this bit is set.
77 if (!boot_cpu_has(X86_FEATURE_NONSTOP_TSC))
78 tsc_detected_unstable = 1;
81 /* TSC could halt in idle */
82 tsc_detected_unstable = 1;
87 static unsigned long cpu_weight[NR_CPUS];
88 static int tsk_in_cpu[NR_CPUS] = {[0 ... NR_CPUS-1] = -1};
89 static DECLARE_BITMAP(pad_busy_cpus_bits, NR_CPUS);
90 static void round_robin_cpu(unsigned int tsk_index)
92 struct cpumask *pad_busy_cpus = to_cpumask(pad_busy_cpus_bits);
95 unsigned long min_weight = -1;
96 unsigned long uninitialized_var(preferred_cpu);
98 if (!alloc_cpumask_var(&tmp, GFP_KERNEL))
101 mutex_lock(&round_robin_lock);
103 for_each_cpu(cpu, pad_busy_cpus)
104 cpumask_or(tmp, tmp, topology_sibling_cpumask(cpu));
105 cpumask_andnot(tmp, cpu_online_mask, tmp);
106 /* avoid HT sibilings if possible */
107 if (cpumask_empty(tmp))
108 cpumask_andnot(tmp, cpu_online_mask, pad_busy_cpus);
109 if (cpumask_empty(tmp)) {
110 mutex_unlock(&round_robin_lock);
113 for_each_cpu(cpu, tmp) {
114 if (cpu_weight[cpu] < min_weight) {
115 min_weight = cpu_weight[cpu];
120 if (tsk_in_cpu[tsk_index] != -1)
121 cpumask_clear_cpu(tsk_in_cpu[tsk_index], pad_busy_cpus);
122 tsk_in_cpu[tsk_index] = preferred_cpu;
123 cpumask_set_cpu(preferred_cpu, pad_busy_cpus);
124 cpu_weight[preferred_cpu]++;
125 mutex_unlock(&round_robin_lock);
127 set_cpus_allowed_ptr(current, cpumask_of(preferred_cpu));
130 static void exit_round_robin(unsigned int tsk_index)
132 struct cpumask *pad_busy_cpus = to_cpumask(pad_busy_cpus_bits);
133 cpumask_clear_cpu(tsk_in_cpu[tsk_index], pad_busy_cpus);
134 tsk_in_cpu[tsk_index] = -1;
137 static unsigned int idle_pct = 5; /* percentage */
138 static unsigned int round_robin_time = 1; /* second */
139 static int power_saving_thread(void *data)
141 struct sched_param param = {.sched_priority = 1};
143 unsigned int tsk_index = (unsigned long)data;
144 u64 last_jiffies = 0;
146 sched_setscheduler(current, SCHED_RR, ¶m);
148 while (!kthread_should_stop()) {
149 unsigned long expire_time;
153 /* round robin to cpus */
154 expire_time = last_jiffies + round_robin_time * HZ;
155 if (time_before(expire_time, jiffies)) {
156 last_jiffies = jiffies;
157 round_robin_cpu(tsk_index);
162 expire_time = jiffies + HZ * (100 - idle_pct) / 100;
164 while (!need_resched()) {
165 if (tsc_detected_unstable && !tsc_marked_unstable) {
166 /* TSC could halt in idle, so notify users */
167 mark_tsc_unstable("TSC halts in idle");
168 tsc_marked_unstable = 1;
171 tick_broadcast_enable();
172 tick_broadcast_enter();
173 stop_critical_timings();
175 mwait_idle_with_hints(power_saving_mwait_eax, 1);
177 start_critical_timings();
178 tick_broadcast_exit();
181 if (time_before(expire_time, jiffies)) {
188 * current sched_rt has threshold for rt task running time.
189 * When a rt task uses 95% CPU time, the rt thread will be
190 * scheduled out for 5% CPU time to not starve other tasks. But
191 * the mechanism only works when all CPUs have RT task running,
192 * as if one CPU hasn't RT task, RT task from other CPUs will
193 * borrow CPU time from this CPU and cause RT task use > 95%
194 * CPU time. To make 'avoid starvation' work, takes a nap here.
196 if (unlikely(do_sleep))
197 schedule_timeout_killable(HZ * idle_pct / 100);
199 /* If an external event has set the need_resched flag, then
200 * we need to deal with it, or this loop will continue to
201 * spin without calling __mwait().
203 if (unlikely(need_resched()))
207 exit_round_robin(tsk_index);
211 static struct task_struct *ps_tsks[NR_CPUS];
212 static unsigned int ps_tsk_num;
213 static int create_power_saving_task(void)
217 ps_tsks[ps_tsk_num] = kthread_run(power_saving_thread,
218 (void *)(unsigned long)ps_tsk_num,
219 "acpi_pad/%d", ps_tsk_num);
221 if (IS_ERR(ps_tsks[ps_tsk_num])) {
222 rc = PTR_ERR(ps_tsks[ps_tsk_num]);
223 ps_tsks[ps_tsk_num] = NULL;
232 static void destroy_power_saving_task(void)
234 if (ps_tsk_num > 0) {
236 kthread_stop(ps_tsks[ps_tsk_num]);
237 ps_tsks[ps_tsk_num] = NULL;
241 static void set_power_saving_task_num(unsigned int num)
243 if (num > ps_tsk_num) {
244 while (ps_tsk_num < num) {
245 if (create_power_saving_task())
248 } else if (num < ps_tsk_num) {
249 while (ps_tsk_num > num)
250 destroy_power_saving_task();
254 static void acpi_pad_idle_cpus(unsigned int num_cpus)
258 num_cpus = min_t(unsigned int, num_cpus, num_online_cpus());
259 set_power_saving_task_num(num_cpus);
264 static uint32_t acpi_pad_idle_cpus_num(void)
269 static ssize_t acpi_pad_rrtime_store(struct device *dev,
270 struct device_attribute *attr, const char *buf, size_t count)
273 if (kstrtoul(buf, 0, &num))
275 if (num < 1 || num >= 100)
277 mutex_lock(&isolated_cpus_lock);
278 round_robin_time = num;
279 mutex_unlock(&isolated_cpus_lock);
283 static ssize_t acpi_pad_rrtime_show(struct device *dev,
284 struct device_attribute *attr, char *buf)
286 return scnprintf(buf, PAGE_SIZE, "%d\n", round_robin_time);
288 static DEVICE_ATTR(rrtime, S_IRUGO|S_IWUSR,
289 acpi_pad_rrtime_show,
290 acpi_pad_rrtime_store);
292 static ssize_t acpi_pad_idlepct_store(struct device *dev,
293 struct device_attribute *attr, const char *buf, size_t count)
296 if (kstrtoul(buf, 0, &num))
298 if (num < 1 || num >= 100)
300 mutex_lock(&isolated_cpus_lock);
302 mutex_unlock(&isolated_cpus_lock);
306 static ssize_t acpi_pad_idlepct_show(struct device *dev,
307 struct device_attribute *attr, char *buf)
309 return scnprintf(buf, PAGE_SIZE, "%d\n", idle_pct);
311 static DEVICE_ATTR(idlepct, S_IRUGO|S_IWUSR,
312 acpi_pad_idlepct_show,
313 acpi_pad_idlepct_store);
315 static ssize_t acpi_pad_idlecpus_store(struct device *dev,
316 struct device_attribute *attr, const char *buf, size_t count)
319 if (kstrtoul(buf, 0, &num))
321 mutex_lock(&isolated_cpus_lock);
322 acpi_pad_idle_cpus(num);
323 mutex_unlock(&isolated_cpus_lock);
327 static ssize_t acpi_pad_idlecpus_show(struct device *dev,
328 struct device_attribute *attr, char *buf)
330 return cpumap_print_to_pagebuf(false, buf,
331 to_cpumask(pad_busy_cpus_bits));
334 static DEVICE_ATTR(idlecpus, S_IRUGO|S_IWUSR,
335 acpi_pad_idlecpus_show,
336 acpi_pad_idlecpus_store);
338 static int acpi_pad_add_sysfs(struct acpi_device *device)
342 result = device_create_file(&device->dev, &dev_attr_idlecpus);
345 result = device_create_file(&device->dev, &dev_attr_idlepct);
347 device_remove_file(&device->dev, &dev_attr_idlecpus);
350 result = device_create_file(&device->dev, &dev_attr_rrtime);
352 device_remove_file(&device->dev, &dev_attr_idlecpus);
353 device_remove_file(&device->dev, &dev_attr_idlepct);
359 static void acpi_pad_remove_sysfs(struct acpi_device *device)
361 device_remove_file(&device->dev, &dev_attr_idlecpus);
362 device_remove_file(&device->dev, &dev_attr_idlepct);
363 device_remove_file(&device->dev, &dev_attr_rrtime);
367 * Query firmware how many CPUs should be idle
368 * return -1 on failure
370 static int acpi_pad_pur(acpi_handle handle)
372 struct acpi_buffer buffer = {ACPI_ALLOCATE_BUFFER, NULL};
373 union acpi_object *package;
376 if (ACPI_FAILURE(acpi_evaluate_object(handle, "_PUR", NULL, &buffer)))
379 if (!buffer.length || !buffer.pointer)
382 package = buffer.pointer;
384 if (package->type == ACPI_TYPE_PACKAGE &&
385 package->package.count == 2 &&
386 package->package.elements[0].integer.value == 1) /* rev 1 */
388 num = package->package.elements[1].integer.value;
390 kfree(buffer.pointer);
394 static void acpi_pad_handle_notify(acpi_handle handle)
398 struct acpi_buffer param = {
400 .pointer = (void *)&idle_cpus,
403 mutex_lock(&isolated_cpus_lock);
404 num_cpus = acpi_pad_pur(handle);
406 mutex_unlock(&isolated_cpus_lock);
409 acpi_pad_idle_cpus(num_cpus);
410 idle_cpus = acpi_pad_idle_cpus_num();
411 acpi_evaluate_ost(handle, ACPI_PROCESSOR_AGGREGATOR_NOTIFY, 0, ¶m);
412 mutex_unlock(&isolated_cpus_lock);
415 static void acpi_pad_notify(acpi_handle handle, u32 event,
418 struct acpi_device *device = data;
421 case ACPI_PROCESSOR_AGGREGATOR_NOTIFY:
422 acpi_pad_handle_notify(handle);
423 acpi_bus_generate_netlink_event(device->pnp.device_class,
424 dev_name(&device->dev), event, 0);
427 pr_warn("Unsupported event [0x%x]\n", event);
432 static int acpi_pad_add(struct acpi_device *device)
436 strcpy(acpi_device_name(device), ACPI_PROCESSOR_AGGREGATOR_DEVICE_NAME);
437 strcpy(acpi_device_class(device), ACPI_PROCESSOR_AGGREGATOR_CLASS);
439 if (acpi_pad_add_sysfs(device))
442 status = acpi_install_notify_handler(device->handle,
443 ACPI_DEVICE_NOTIFY, acpi_pad_notify, device);
444 if (ACPI_FAILURE(status)) {
445 acpi_pad_remove_sysfs(device);
452 static int acpi_pad_remove(struct acpi_device *device)
454 mutex_lock(&isolated_cpus_lock);
455 acpi_pad_idle_cpus(0);
456 mutex_unlock(&isolated_cpus_lock);
458 acpi_remove_notify_handler(device->handle,
459 ACPI_DEVICE_NOTIFY, acpi_pad_notify);
460 acpi_pad_remove_sysfs(device);
464 static const struct acpi_device_id pad_device_ids[] = {
468 MODULE_DEVICE_TABLE(acpi, pad_device_ids);
470 static struct acpi_driver acpi_pad_driver = {
471 .name = "processor_aggregator",
472 .class = ACPI_PROCESSOR_AGGREGATOR_CLASS,
473 .ids = pad_device_ids,
476 .remove = acpi_pad_remove,
480 static int __init acpi_pad_init(void)
482 power_saving_mwait_init();
483 if (power_saving_mwait_eax == 0)
486 return acpi_bus_register_driver(&acpi_pad_driver);
489 static void __exit acpi_pad_exit(void)
491 acpi_bus_unregister_driver(&acpi_pad_driver);
494 module_init(acpi_pad_init);
495 module_exit(acpi_pad_exit);
496 MODULE_AUTHOR("Shaohua Li<shaohua.li@intel.com>");
497 MODULE_DESCRIPTION("ACPI Processor Aggregator Driver");
498 MODULE_LICENSE("GPL");