acpi_pad.c revision 3b8cb427e9281790f36e847e46cb1d005a50cec0 
1/*
2 * acpi_pad.c ACPI Processor Aggregator Driver
3 *
4 * Copyright (c) 2009, Intel Corporation.
5 *
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.
9 *
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
13 * more details.
14 *
15 * You should have received a copy of the GNU General Public License along with
16 * this program; if not, write to the Free Software Foundation, Inc.,
17 * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 */
20
21#include <linux/kernel.h>
22#include <linux/cpumask.h>
23#include <linux/module.h>
24#include <linux/init.h>
25#include <linux/types.h>
26#include <linux/kthread.h>
27#include <linux/freezer.h>
28#include <linux/cpu.h>
29#include <linux/clockchips.h>
30#include <acpi/acpi_bus.h>
31#include <acpi/acpi_drivers.h>
32
33#define ACPI_PROCESSOR_AGGREGATOR_CLASS	"processor_aggregator"
34#define ACPI_PROCESSOR_AGGREGATOR_DEVICE_NAME "Processor Aggregator"
35#define ACPI_PROCESSOR_AGGREGATOR_NOTIFY 0x80
36static DEFINE_MUTEX(isolated_cpus_lock);
37
38#define MWAIT_SUBSTATE_MASK	(0xf)
39#define MWAIT_CSTATE_MASK	(0xf)
40#define MWAIT_SUBSTATE_SIZE	(4)
41#define CPUID_MWAIT_LEAF (5)
42#define CPUID5_ECX_EXTENSIONS_SUPPORTED (0x1)
43#define CPUID5_ECX_INTERRUPT_BREAK	(0x2)
44static unsigned long power_saving_mwait_eax;
45static void power_saving_mwait_init(void)
46{
47	unsigned int eax, ebx, ecx, edx;
48	unsigned int highest_cstate = 0;
49	unsigned int highest_subcstate = 0;
50	int i;
51
52	if (!boot_cpu_has(X86_FEATURE_MWAIT))
53		return;
54	if (boot_cpu_data.cpuid_level < CPUID_MWAIT_LEAF)
55		return;
56
57	cpuid(CPUID_MWAIT_LEAF, &eax, &ebx, &ecx, &edx);
58
59	if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED) ||
60	    !(ecx & CPUID5_ECX_INTERRUPT_BREAK))
61		return;
62
63	edx >>= MWAIT_SUBSTATE_SIZE;
64	for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) {
65		if (edx & MWAIT_SUBSTATE_MASK) {
66			highest_cstate = i;
67			highest_subcstate = edx & MWAIT_SUBSTATE_MASK;
68		}
69	}
70	power_saving_mwait_eax = (highest_cstate << MWAIT_SUBSTATE_SIZE) |
71		(highest_subcstate - 1);
72
73	for_each_online_cpu(i)
74		clockevents_notify(CLOCK_EVT_NOTIFY_BROADCAST_ON, &i);
75
76#if defined(CONFIG_GENERIC_TIME) && defined(CONFIG_X86)
77	switch (boot_cpu_data.x86_vendor) {
78	case X86_VENDOR_AMD:
79	case X86_VENDOR_INTEL:
80		/*
81		 * AMD Fam10h TSC will tick in all
82		 * C/P/S0/S1 states when this bit is set.
83		 */
84		if (boot_cpu_has(X86_FEATURE_NONSTOP_TSC))
85			return;
86
87		/*FALL THROUGH*/
88	default:
89		/* TSC could halt in idle, so notify users */
90		mark_tsc_unstable("TSC halts in idle");
91	}
92#endif
93}
94
95static unsigned long cpu_weight[NR_CPUS];
96static int tsk_in_cpu[NR_CPUS] = {[0 ... NR_CPUS-1] = -1};
97static DECLARE_BITMAP(pad_busy_cpus_bits, NR_CPUS);
98static void round_robin_cpu(unsigned int tsk_index)
99{
100	struct cpumask *pad_busy_cpus = to_cpumask(pad_busy_cpus_bits);
101	cpumask_var_t tmp;
102	int cpu;
103	unsigned long min_weight = -1, preferred_cpu;
104
105	if (!alloc_cpumask_var(&tmp, GFP_KERNEL))
106		return;
107
108	mutex_lock(&isolated_cpus_lock);
109	cpumask_clear(tmp);
110	for_each_cpu(cpu, pad_busy_cpus)
111		cpumask_or(tmp, tmp, topology_thread_cpumask(cpu));
112	cpumask_andnot(tmp, cpu_online_mask, tmp);
113	/* avoid HT sibilings if possible */
114	if (cpumask_empty(tmp))
115		cpumask_andnot(tmp, cpu_online_mask, pad_busy_cpus);
116	if (cpumask_empty(tmp)) {
117		mutex_unlock(&isolated_cpus_lock);
118		return;
119	}
120	for_each_cpu(cpu, tmp) {
121		if (cpu_weight[cpu] < min_weight) {
122			min_weight = cpu_weight[cpu];
123			preferred_cpu = cpu;
124		}
125	}
126
127	if (tsk_in_cpu[tsk_index] != -1)
128		cpumask_clear_cpu(tsk_in_cpu[tsk_index], pad_busy_cpus);
129	tsk_in_cpu[tsk_index] = preferred_cpu;
130	cpumask_set_cpu(preferred_cpu, pad_busy_cpus);
131	cpu_weight[preferred_cpu]++;
132	mutex_unlock(&isolated_cpus_lock);
133
134	set_cpus_allowed_ptr(current, cpumask_of(preferred_cpu));
135}
136
137static void exit_round_robin(unsigned int tsk_index)
138{
139	struct cpumask *pad_busy_cpus = to_cpumask(pad_busy_cpus_bits);
140	cpumask_clear_cpu(tsk_in_cpu[tsk_index], pad_busy_cpus);
141	tsk_in_cpu[tsk_index] = -1;
142}
143
144static unsigned int idle_pct = 5; /* percentage */
145static unsigned int round_robin_time = 10; /* second */
146static int power_saving_thread(void *data)
147{
148	struct sched_param param = {.sched_priority = 1};
149	int do_sleep;
150	unsigned int tsk_index = (unsigned long)data;
151	u64 last_jiffies = 0;
152
153	sched_setscheduler(current, SCHED_RR, &param);
154
155	while (!kthread_should_stop()) {
156		int cpu;
157		u64 expire_time;
158
159		try_to_freeze();
160
161		/* round robin to cpus */
162		if (last_jiffies + round_robin_time * HZ < jiffies) {
163			last_jiffies = jiffies;
164			round_robin_cpu(tsk_index);
165		}
166
167		do_sleep = 0;
168
169		current_thread_info()->status &= ~TS_POLLING;
170		/*
171		 * TS_POLLING-cleared state must be visible before we test
172		 * NEED_RESCHED:
173		 */
174		smp_mb();
175
176		expire_time = jiffies + HZ * (100 - idle_pct) / 100;
177
178		while (!need_resched()) {
179			local_irq_disable();
180			cpu = smp_processor_id();
181			clockevents_notify(CLOCK_EVT_NOTIFY_BROADCAST_ENTER,
182				&cpu);
183			stop_critical_timings();
184
185			__monitor((void *)&current_thread_info()->flags, 0, 0);
186			smp_mb();
187			if (!need_resched())
188				__mwait(power_saving_mwait_eax, 1);
189
190			start_critical_timings();
191			clockevents_notify(CLOCK_EVT_NOTIFY_BROADCAST_EXIT,
192				&cpu);
193			local_irq_enable();
194
195			if (jiffies > expire_time) {
196				do_sleep = 1;
197				break;
198			}
199		}
200
201		current_thread_info()->status |= TS_POLLING;
202
203		/*
204		 * current sched_rt has threshold for rt task running time.
205		 * When a rt task uses 95% CPU time, the rt thread will be
206		 * scheduled out for 5% CPU time to not starve other tasks. But
207		 * the mechanism only works when all CPUs have RT task running,
208		 * as if one CPU hasn't RT task, RT task from other CPUs will
209		 * borrow CPU time from this CPU and cause RT task use > 95%
210		 * CPU time. To make 'avoid starvation' work, takes a nap here.
211		 */
212		if (do_sleep)
213			schedule_timeout_killable(HZ * idle_pct / 100);
214	}
215
216	exit_round_robin(tsk_index);
217	return 0;
218}
219
220static struct task_struct *ps_tsks[NR_CPUS];
221static unsigned int ps_tsk_num;
222static int create_power_saving_task(void)
223{
224	int rc = -ENOMEM;
225
226	ps_tsks[ps_tsk_num] = kthread_run(power_saving_thread,
227		(void *)(unsigned long)ps_tsk_num,
228		"power_saving/%d", ps_tsk_num);
229	rc = IS_ERR(ps_tsks[ps_tsk_num]) ? PTR_ERR(ps_tsks[ps_tsk_num]) : 0;
230	if (!rc)
231		ps_tsk_num++;
232	else
233		ps_tsks[ps_tsk_num] = NULL;
234
235	return rc;
236}
237
238static void destroy_power_saving_task(void)
239{
240	if (ps_tsk_num > 0) {
241		ps_tsk_num--;
242		kthread_stop(ps_tsks[ps_tsk_num]);
243		ps_tsks[ps_tsk_num] = NULL;
244	}
245}
246
247static void set_power_saving_task_num(unsigned int num)
248{
249	if (num > ps_tsk_num) {
250		while (ps_tsk_num < num) {
251			if (create_power_saving_task())
252				return;
253		}
254	} else if (num < ps_tsk_num) {
255		while (ps_tsk_num > num)
256			destroy_power_saving_task();
257	}
258}
259
260static void acpi_pad_idle_cpus(unsigned int num_cpus)
261{
262	get_online_cpus();
263
264	num_cpus = min_t(unsigned int, num_cpus, num_online_cpus());
265	set_power_saving_task_num(num_cpus);
266
267	put_online_cpus();
268}
269
270static uint32_t acpi_pad_idle_cpus_num(void)
271{
272	return ps_tsk_num;
273}
274
275static ssize_t acpi_pad_rrtime_store(struct device *dev,
276	struct device_attribute *attr, const char *buf, size_t count)
277{
278	unsigned long num;
279	if (strict_strtoul(buf, 0, &num))
280		return -EINVAL;
281	if (num < 1 || num >= 100)
282		return -EINVAL;
283	mutex_lock(&isolated_cpus_lock);
284	round_robin_time = num;
285	mutex_unlock(&isolated_cpus_lock);
286	return count;
287}
288
289static ssize_t acpi_pad_rrtime_show(struct device *dev,
290	struct device_attribute *attr, char *buf)
291{
292	return scnprintf(buf, PAGE_SIZE, "%d", round_robin_time);
293}
294static DEVICE_ATTR(rrtime, S_IRUGO|S_IWUSR,
295	acpi_pad_rrtime_show,
296	acpi_pad_rrtime_store);
297
298static ssize_t acpi_pad_idlepct_store(struct device *dev,
299	struct device_attribute *attr, const char *buf, size_t count)
300{
301	unsigned long num;
302	if (strict_strtoul(buf, 0, &num))
303		return -EINVAL;
304	if (num < 1 || num >= 100)
305		return -EINVAL;
306	mutex_lock(&isolated_cpus_lock);
307	idle_pct = num;
308	mutex_unlock(&isolated_cpus_lock);
309	return count;
310}
311
312static ssize_t acpi_pad_idlepct_show(struct device *dev,
313	struct device_attribute *attr, char *buf)
314{
315	return scnprintf(buf, PAGE_SIZE, "%d", idle_pct);
316}
317static DEVICE_ATTR(idlepct, S_IRUGO|S_IWUSR,
318	acpi_pad_idlepct_show,
319	acpi_pad_idlepct_store);
320
321static ssize_t acpi_pad_idlecpus_store(struct device *dev,
322	struct device_attribute *attr, const char *buf, size_t count)
323{
324	unsigned long num;
325	if (strict_strtoul(buf, 0, &num))
326		return -EINVAL;
327	mutex_lock(&isolated_cpus_lock);
328	acpi_pad_idle_cpus(num);
329	mutex_unlock(&isolated_cpus_lock);
330	return count;
331}
332
333static ssize_t acpi_pad_idlecpus_show(struct device *dev,
334	struct device_attribute *attr, char *buf)
335{
336	return cpumask_scnprintf(buf, PAGE_SIZE,
337		to_cpumask(pad_busy_cpus_bits));
338}
339static DEVICE_ATTR(idlecpus, S_IRUGO|S_IWUSR,
340	acpi_pad_idlecpus_show,
341	acpi_pad_idlecpus_store);
342
343static int acpi_pad_add_sysfs(struct acpi_device *device)
344{
345	int result;
346
347	result = device_create_file(&device->dev, &dev_attr_idlecpus);
348	if (result)
349		return -ENODEV;
350	result = device_create_file(&device->dev, &dev_attr_idlepct);
351	if (result) {
352		device_remove_file(&device->dev, &dev_attr_idlecpus);
353		return -ENODEV;
354	}
355	result = device_create_file(&device->dev, &dev_attr_rrtime);
356	if (result) {
357		device_remove_file(&device->dev, &dev_attr_idlecpus);
358		device_remove_file(&device->dev, &dev_attr_idlepct);
359		return -ENODEV;
360	}
361	return 0;
362}
363
364static void acpi_pad_remove_sysfs(struct acpi_device *device)
365{
366	device_remove_file(&device->dev, &dev_attr_idlecpus);
367	device_remove_file(&device->dev, &dev_attr_idlepct);
368	device_remove_file(&device->dev, &dev_attr_rrtime);
369}
370
371/* Query firmware how many CPUs should be idle */
372static int acpi_pad_pur(acpi_handle handle, int *num_cpus)
373{
374	struct acpi_buffer buffer = {ACPI_ALLOCATE_BUFFER, NULL};
375	union acpi_object *package;
376	int rev, num, ret = -EINVAL;
377
378	if (ACPI_FAILURE(acpi_evaluate_object(handle, "_PUR", NULL, &buffer)))
379		return -EINVAL;
380
381	if (!buffer.length || !buffer.pointer)
382		return -EINVAL;
383
384	package = buffer.pointer;
385	if (package->type != ACPI_TYPE_PACKAGE || package->package.count != 2)
386		goto out;
387	rev = package->package.elements[0].integer.value;
388	num = package->package.elements[1].integer.value;
389	if (rev != 1 || num < 0)
390		goto out;
391	*num_cpus = num;
392	ret = 0;
393out:
394	kfree(buffer.pointer);
395	return ret;
396}
397
398/* Notify firmware how many CPUs are idle */
399static void acpi_pad_ost(acpi_handle handle, int stat,
400	uint32_t idle_cpus)
401{
402	union acpi_object params[3] = {
403		{.type = ACPI_TYPE_INTEGER,},
404		{.type = ACPI_TYPE_INTEGER,},
405		{.type = ACPI_TYPE_BUFFER,},
406	};
407	struct acpi_object_list arg_list = {3, params};
408
409	params[0].integer.value = ACPI_PROCESSOR_AGGREGATOR_NOTIFY;
410	params[1].integer.value =  stat;
411	params[2].buffer.length = 4;
412	params[2].buffer.pointer = (void *)&idle_cpus;
413	acpi_evaluate_object(handle, "_OST", &arg_list, NULL);
414}
415
416static void acpi_pad_handle_notify(acpi_handle handle)
417{
418	int num_cpus;
419	uint32_t idle_cpus;
420
421	mutex_lock(&isolated_cpus_lock);
422	if (acpi_pad_pur(handle, &num_cpus)) {
423		mutex_unlock(&isolated_cpus_lock);
424		return;
425	}
426	acpi_pad_idle_cpus(num_cpus);
427	idle_cpus = acpi_pad_idle_cpus_num();
428	acpi_pad_ost(handle, 0, idle_cpus);
429	mutex_unlock(&isolated_cpus_lock);
430}
431
432static void acpi_pad_notify(acpi_handle handle, u32 event,
433	void *data)
434{
435	struct acpi_device *device = data;
436
437	switch (event) {
438	case ACPI_PROCESSOR_AGGREGATOR_NOTIFY:
439		acpi_pad_handle_notify(handle);
440		acpi_bus_generate_proc_event(device, event, 0);
441		acpi_bus_generate_netlink_event(device->pnp.device_class,
442			dev_name(&device->dev), event, 0);
443		break;
444	default:
445		printk(KERN_WARNING"Unsupported event [0x%x]\n", event);
446		break;
447	}
448}
449
450static int acpi_pad_add(struct acpi_device *device)
451{
452	acpi_status status;
453
454	strcpy(acpi_device_name(device), ACPI_PROCESSOR_AGGREGATOR_DEVICE_NAME);
455	strcpy(acpi_device_class(device), ACPI_PROCESSOR_AGGREGATOR_CLASS);
456
457	if (acpi_pad_add_sysfs(device))
458		return -ENODEV;
459
460	status = acpi_install_notify_handler(device->handle,
461		ACPI_DEVICE_NOTIFY, acpi_pad_notify, device);
462	if (ACPI_FAILURE(status)) {
463		acpi_pad_remove_sysfs(device);
464		return -ENODEV;
465	}
466
467	return 0;
468}
469
470static int acpi_pad_remove(struct acpi_device *device,
471	int type)
472{
473	mutex_lock(&isolated_cpus_lock);
474	acpi_pad_idle_cpus(0);
475	mutex_unlock(&isolated_cpus_lock);
476
477	acpi_remove_notify_handler(device->handle,
478		ACPI_DEVICE_NOTIFY, acpi_pad_notify);
479	acpi_pad_remove_sysfs(device);
480	return 0;
481}
482
483static const struct acpi_device_id pad_device_ids[] = {
484	{"ACPI000C", 0},
485	{"", 0},
486};
487MODULE_DEVICE_TABLE(acpi, pad_device_ids);
488
489static struct acpi_driver acpi_pad_driver = {
490	.name = "processor_aggregator",
491	.class = ACPI_PROCESSOR_AGGREGATOR_CLASS,
492	.ids = pad_device_ids,
493	.ops = {
494		.add = acpi_pad_add,
495		.remove = acpi_pad_remove,
496	},
497};
498
499static int __init acpi_pad_init(void)
500{
501	power_saving_mwait_init();
502	if (power_saving_mwait_eax == 0)
503		return -EINVAL;
504
505	return acpi_bus_register_driver(&acpi_pad_driver);
506}
507
508static void __exit acpi_pad_exit(void)
509{
510	acpi_bus_unregister_driver(&acpi_pad_driver);
511}
512
513module_init(acpi_pad_init);
514module_exit(acpi_pad_exit);
515MODULE_AUTHOR("Shaohua Li<shaohua.li@intel.com>");
516MODULE_DESCRIPTION("ACPI Processor Aggregator Driver");
517MODULE_LICENSE("GPL");
518