Loading...
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 */
16
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>
29
30#define ACPI_PROCESSOR_AGGREGATOR_CLASS "acpi_pad"
31#define ACPI_PROCESSOR_AGGREGATOR_DEVICE_NAME "Processor Aggregator"
32#define ACPI_PROCESSOR_AGGREGATOR_NOTIFY 0x80
33static DEFINE_MUTEX(isolated_cpus_lock);
34static DEFINE_MUTEX(round_robin_lock);
35
36static unsigned long power_saving_mwait_eax;
37
38static unsigned char tsc_detected_unstable;
39static unsigned char tsc_marked_unstable;
40
41static void power_saving_mwait_init(void)
42{
43 unsigned int eax, ebx, ecx, edx;
44 unsigned int highest_cstate = 0;
45 unsigned int highest_subcstate = 0;
46 int i;
47
48 if (!boot_cpu_has(X86_FEATURE_MWAIT))
49 return;
50 if (boot_cpu_data.cpuid_level < CPUID_MWAIT_LEAF)
51 return;
52
53 cpuid(CPUID_MWAIT_LEAF, &eax, &ebx, &ecx, &edx);
54
55 if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED) ||
56 !(ecx & CPUID5_ECX_INTERRUPT_BREAK))
57 return;
58
59 edx >>= MWAIT_SUBSTATE_SIZE;
60 for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) {
61 if (edx & MWAIT_SUBSTATE_MASK) {
62 highest_cstate = i;
63 highest_subcstate = edx & MWAIT_SUBSTATE_MASK;
64 }
65 }
66 power_saving_mwait_eax = (highest_cstate << MWAIT_SUBSTATE_SIZE) |
67 (highest_subcstate - 1);
68
69#if defined(CONFIG_X86)
70 switch (boot_cpu_data.x86_vendor) {
71 case X86_VENDOR_AMD:
72 case X86_VENDOR_INTEL:
73 /*
74 * AMD Fam10h TSC will tick in all
75 * C/P/S0/S1 states when this bit is set.
76 */
77 if (!boot_cpu_has(X86_FEATURE_NONSTOP_TSC))
78 tsc_detected_unstable = 1;
79 break;
80 default:
81 /* TSC could halt in idle */
82 tsc_detected_unstable = 1;
83 }
84#endif
85}
86
87static unsigned long cpu_weight[NR_CPUS];
88static int tsk_in_cpu[NR_CPUS] = {[0 ... NR_CPUS-1] = -1};
89static DECLARE_BITMAP(pad_busy_cpus_bits, NR_CPUS);
90static void round_robin_cpu(unsigned int tsk_index)
91{
92 struct cpumask *pad_busy_cpus = to_cpumask(pad_busy_cpus_bits);
93 cpumask_var_t tmp;
94 int cpu;
95 unsigned long min_weight = -1;
96 unsigned long uninitialized_var(preferred_cpu);
97
98 if (!alloc_cpumask_var(&tmp, GFP_KERNEL))
99 return;
100
101 mutex_lock(&round_robin_lock);
102 cpumask_clear(tmp);
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);
111 return;
112 }
113 for_each_cpu(cpu, tmp) {
114 if (cpu_weight[cpu] < min_weight) {
115 min_weight = cpu_weight[cpu];
116 preferred_cpu = cpu;
117 }
118 }
119
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);
126
127 set_cpus_allowed_ptr(current, cpumask_of(preferred_cpu));
128}
129
130static void exit_round_robin(unsigned int tsk_index)
131{
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;
135}
136
137static unsigned int idle_pct = 5; /* percentage */
138static unsigned int round_robin_time = 1; /* second */
139static int power_saving_thread(void *data)
140{
141 struct sched_param param = {.sched_priority = 1};
142 int do_sleep;
143 unsigned int tsk_index = (unsigned long)data;
144 u64 last_jiffies = 0;
145
146 sched_setscheduler(current, SCHED_RR, ¶m);
147
148 while (!kthread_should_stop()) {
149 unsigned long expire_time;
150
151 /* round robin to cpus */
152 expire_time = last_jiffies + round_robin_time * HZ;
153 if (time_before(expire_time, jiffies)) {
154 last_jiffies = jiffies;
155 round_robin_cpu(tsk_index);
156 }
157
158 do_sleep = 0;
159
160 expire_time = jiffies + HZ * (100 - idle_pct) / 100;
161
162 while (!need_resched()) {
163 if (tsc_detected_unstable && !tsc_marked_unstable) {
164 /* TSC could halt in idle, so notify users */
165 mark_tsc_unstable("TSC halts in idle");
166 tsc_marked_unstable = 1;
167 }
168 local_irq_disable();
169 tick_broadcast_enable();
170 tick_broadcast_enter();
171 stop_critical_timings();
172
173 mwait_idle_with_hints(power_saving_mwait_eax, 1);
174
175 start_critical_timings();
176 tick_broadcast_exit();
177 local_irq_enable();
178
179 if (time_before(expire_time, jiffies)) {
180 do_sleep = 1;
181 break;
182 }
183 }
184
185 /*
186 * current sched_rt has threshold for rt task running time.
187 * When a rt task uses 95% CPU time, the rt thread will be
188 * scheduled out for 5% CPU time to not starve other tasks. But
189 * the mechanism only works when all CPUs have RT task running,
190 * as if one CPU hasn't RT task, RT task from other CPUs will
191 * borrow CPU time from this CPU and cause RT task use > 95%
192 * CPU time. To make 'avoid starvation' work, takes a nap here.
193 */
194 if (unlikely(do_sleep))
195 schedule_timeout_killable(HZ * idle_pct / 100);
196
197 /* If an external event has set the need_resched flag, then
198 * we need to deal with it, or this loop will continue to
199 * spin without calling __mwait().
200 */
201 if (unlikely(need_resched()))
202 schedule();
203 }
204
205 exit_round_robin(tsk_index);
206 return 0;
207}
208
209static struct task_struct *ps_tsks[NR_CPUS];
210static unsigned int ps_tsk_num;
211static int create_power_saving_task(void)
212{
213 int rc;
214
215 ps_tsks[ps_tsk_num] = kthread_run(power_saving_thread,
216 (void *)(unsigned long)ps_tsk_num,
217 "acpi_pad/%d", ps_tsk_num);
218
219 if (IS_ERR(ps_tsks[ps_tsk_num])) {
220 rc = PTR_ERR(ps_tsks[ps_tsk_num]);
221 ps_tsks[ps_tsk_num] = NULL;
222 } else {
223 rc = 0;
224 ps_tsk_num++;
225 }
226
227 return rc;
228}
229
230static void destroy_power_saving_task(void)
231{
232 if (ps_tsk_num > 0) {
233 ps_tsk_num--;
234 kthread_stop(ps_tsks[ps_tsk_num]);
235 ps_tsks[ps_tsk_num] = NULL;
236 }
237}
238
239static void set_power_saving_task_num(unsigned int num)
240{
241 if (num > ps_tsk_num) {
242 while (ps_tsk_num < num) {
243 if (create_power_saving_task())
244 return;
245 }
246 } else if (num < ps_tsk_num) {
247 while (ps_tsk_num > num)
248 destroy_power_saving_task();
249 }
250}
251
252static void acpi_pad_idle_cpus(unsigned int num_cpus)
253{
254 get_online_cpus();
255
256 num_cpus = min_t(unsigned int, num_cpus, num_online_cpus());
257 set_power_saving_task_num(num_cpus);
258
259 put_online_cpus();
260}
261
262static uint32_t acpi_pad_idle_cpus_num(void)
263{
264 return ps_tsk_num;
265}
266
267static ssize_t acpi_pad_rrtime_store(struct device *dev,
268 struct device_attribute *attr, const char *buf, size_t count)
269{
270 unsigned long num;
271 if (kstrtoul(buf, 0, &num))
272 return -EINVAL;
273 if (num < 1 || num >= 100)
274 return -EINVAL;
275 mutex_lock(&isolated_cpus_lock);
276 round_robin_time = num;
277 mutex_unlock(&isolated_cpus_lock);
278 return count;
279}
280
281static ssize_t acpi_pad_rrtime_show(struct device *dev,
282 struct device_attribute *attr, char *buf)
283{
284 return scnprintf(buf, PAGE_SIZE, "%d\n", round_robin_time);
285}
286static DEVICE_ATTR(rrtime, S_IRUGO|S_IWUSR,
287 acpi_pad_rrtime_show,
288 acpi_pad_rrtime_store);
289
290static ssize_t acpi_pad_idlepct_store(struct device *dev,
291 struct device_attribute *attr, const char *buf, size_t count)
292{
293 unsigned long num;
294 if (kstrtoul(buf, 0, &num))
295 return -EINVAL;
296 if (num < 1 || num >= 100)
297 return -EINVAL;
298 mutex_lock(&isolated_cpus_lock);
299 idle_pct = num;
300 mutex_unlock(&isolated_cpus_lock);
301 return count;
302}
303
304static ssize_t acpi_pad_idlepct_show(struct device *dev,
305 struct device_attribute *attr, char *buf)
306{
307 return scnprintf(buf, PAGE_SIZE, "%d\n", idle_pct);
308}
309static DEVICE_ATTR(idlepct, S_IRUGO|S_IWUSR,
310 acpi_pad_idlepct_show,
311 acpi_pad_idlepct_store);
312
313static ssize_t acpi_pad_idlecpus_store(struct device *dev,
314 struct device_attribute *attr, const char *buf, size_t count)
315{
316 unsigned long num;
317 if (kstrtoul(buf, 0, &num))
318 return -EINVAL;
319 mutex_lock(&isolated_cpus_lock);
320 acpi_pad_idle_cpus(num);
321 mutex_unlock(&isolated_cpus_lock);
322 return count;
323}
324
325static ssize_t acpi_pad_idlecpus_show(struct device *dev,
326 struct device_attribute *attr, char *buf)
327{
328 return cpumap_print_to_pagebuf(false, buf,
329 to_cpumask(pad_busy_cpus_bits));
330}
331
332static DEVICE_ATTR(idlecpus, S_IRUGO|S_IWUSR,
333 acpi_pad_idlecpus_show,
334 acpi_pad_idlecpus_store);
335
336static int acpi_pad_add_sysfs(struct acpi_device *device)
337{
338 int result;
339
340 result = device_create_file(&device->dev, &dev_attr_idlecpus);
341 if (result)
342 return -ENODEV;
343 result = device_create_file(&device->dev, &dev_attr_idlepct);
344 if (result) {
345 device_remove_file(&device->dev, &dev_attr_idlecpus);
346 return -ENODEV;
347 }
348 result = device_create_file(&device->dev, &dev_attr_rrtime);
349 if (result) {
350 device_remove_file(&device->dev, &dev_attr_idlecpus);
351 device_remove_file(&device->dev, &dev_attr_idlepct);
352 return -ENODEV;
353 }
354 return 0;
355}
356
357static void acpi_pad_remove_sysfs(struct acpi_device *device)
358{
359 device_remove_file(&device->dev, &dev_attr_idlecpus);
360 device_remove_file(&device->dev, &dev_attr_idlepct);
361 device_remove_file(&device->dev, &dev_attr_rrtime);
362}
363
364/*
365 * Query firmware how many CPUs should be idle
366 * return -1 on failure
367 */
368static int acpi_pad_pur(acpi_handle handle)
369{
370 struct acpi_buffer buffer = {ACPI_ALLOCATE_BUFFER, NULL};
371 union acpi_object *package;
372 int num = -1;
373
374 if (ACPI_FAILURE(acpi_evaluate_object(handle, "_PUR", NULL, &buffer)))
375 return num;
376
377 if (!buffer.length || !buffer.pointer)
378 return num;
379
380 package = buffer.pointer;
381
382 if (package->type == ACPI_TYPE_PACKAGE &&
383 package->package.count == 2 &&
384 package->package.elements[0].integer.value == 1) /* rev 1 */
385
386 num = package->package.elements[1].integer.value;
387
388 kfree(buffer.pointer);
389 return num;
390}
391
392static void acpi_pad_handle_notify(acpi_handle handle)
393{
394 int num_cpus;
395 uint32_t idle_cpus;
396 struct acpi_buffer param = {
397 .length = 4,
398 .pointer = (void *)&idle_cpus,
399 };
400
401 mutex_lock(&isolated_cpus_lock);
402 num_cpus = acpi_pad_pur(handle);
403 if (num_cpus < 0) {
404 mutex_unlock(&isolated_cpus_lock);
405 return;
406 }
407 acpi_pad_idle_cpus(num_cpus);
408 idle_cpus = acpi_pad_idle_cpus_num();
409 acpi_evaluate_ost(handle, ACPI_PROCESSOR_AGGREGATOR_NOTIFY, 0, ¶m);
410 mutex_unlock(&isolated_cpus_lock);
411}
412
413static void acpi_pad_notify(acpi_handle handle, u32 event,
414 void *data)
415{
416 struct acpi_device *device = data;
417
418 switch (event) {
419 case ACPI_PROCESSOR_AGGREGATOR_NOTIFY:
420 acpi_pad_handle_notify(handle);
421 acpi_bus_generate_netlink_event(device->pnp.device_class,
422 dev_name(&device->dev), event, 0);
423 break;
424 default:
425 pr_warn("Unsupported event [0x%x]\n", event);
426 break;
427 }
428}
429
430static int acpi_pad_add(struct acpi_device *device)
431{
432 acpi_status status;
433
434 strcpy(acpi_device_name(device), ACPI_PROCESSOR_AGGREGATOR_DEVICE_NAME);
435 strcpy(acpi_device_class(device), ACPI_PROCESSOR_AGGREGATOR_CLASS);
436
437 if (acpi_pad_add_sysfs(device))
438 return -ENODEV;
439
440 status = acpi_install_notify_handler(device->handle,
441 ACPI_DEVICE_NOTIFY, acpi_pad_notify, device);
442 if (ACPI_FAILURE(status)) {
443 acpi_pad_remove_sysfs(device);
444 return -ENODEV;
445 }
446
447 return 0;
448}
449
450static int acpi_pad_remove(struct acpi_device *device)
451{
452 mutex_lock(&isolated_cpus_lock);
453 acpi_pad_idle_cpus(0);
454 mutex_unlock(&isolated_cpus_lock);
455
456 acpi_remove_notify_handler(device->handle,
457 ACPI_DEVICE_NOTIFY, acpi_pad_notify);
458 acpi_pad_remove_sysfs(device);
459 return 0;
460}
461
462static const struct acpi_device_id pad_device_ids[] = {
463 {"ACPI000C", 0},
464 {"", 0},
465};
466MODULE_DEVICE_TABLE(acpi, pad_device_ids);
467
468static struct acpi_driver acpi_pad_driver = {
469 .name = "processor_aggregator",
470 .class = ACPI_PROCESSOR_AGGREGATOR_CLASS,
471 .ids = pad_device_ids,
472 .ops = {
473 .add = acpi_pad_add,
474 .remove = acpi_pad_remove,
475 },
476};
477
478static int __init acpi_pad_init(void)
479{
480 power_saving_mwait_init();
481 if (power_saving_mwait_eax == 0)
482 return -EINVAL;
483
484 return acpi_bus_register_driver(&acpi_pad_driver);
485}
486
487static void __exit acpi_pad_exit(void)
488{
489 acpi_bus_unregister_driver(&acpi_pad_driver);
490}
491
492module_init(acpi_pad_init);
493module_exit(acpi_pad_exit);
494MODULE_AUTHOR("Shaohua Li<shaohua.li@intel.com>");
495MODULE_DESCRIPTION("ACPI Processor Aggregator Driver");
496MODULE_LICENSE("GPL");
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * acpi_pad.c ACPI Processor Aggregator Driver
4 *
5 * Copyright (c) 2009, Intel Corporation.
6 */
7
8#include <linux/kernel.h>
9#include <linux/cpumask.h>
10#include <linux/module.h>
11#include <linux/init.h>
12#include <linux/types.h>
13#include <linux/kthread.h>
14#include <uapi/linux/sched/types.h>
15#include <linux/freezer.h>
16#include <linux/cpu.h>
17#include <linux/tick.h>
18#include <linux/slab.h>
19#include <linux/acpi.h>
20#include <linux/perf_event.h>
21#include <asm/mwait.h>
22#include <xen/xen.h>
23
24#define ACPI_PROCESSOR_AGGREGATOR_CLASS "acpi_pad"
25#define ACPI_PROCESSOR_AGGREGATOR_DEVICE_NAME "Processor Aggregator"
26#define ACPI_PROCESSOR_AGGREGATOR_NOTIFY 0x80
27static DEFINE_MUTEX(isolated_cpus_lock);
28static DEFINE_MUTEX(round_robin_lock);
29
30static unsigned long power_saving_mwait_eax;
31
32static unsigned char tsc_detected_unstable;
33static unsigned char tsc_marked_unstable;
34
35static void power_saving_mwait_init(void)
36{
37 unsigned int eax, ebx, ecx, edx;
38 unsigned int highest_cstate = 0;
39 unsigned int highest_subcstate = 0;
40 int i;
41
42 if (!boot_cpu_has(X86_FEATURE_MWAIT))
43 return;
44 if (boot_cpu_data.cpuid_level < CPUID_MWAIT_LEAF)
45 return;
46
47 cpuid(CPUID_MWAIT_LEAF, &eax, &ebx, &ecx, &edx);
48
49 if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED) ||
50 !(ecx & CPUID5_ECX_INTERRUPT_BREAK))
51 return;
52
53 edx >>= MWAIT_SUBSTATE_SIZE;
54 for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) {
55 if (edx & MWAIT_SUBSTATE_MASK) {
56 highest_cstate = i;
57 highest_subcstate = edx & MWAIT_SUBSTATE_MASK;
58 }
59 }
60 power_saving_mwait_eax = (highest_cstate << MWAIT_SUBSTATE_SIZE) |
61 (highest_subcstate - 1);
62
63#if defined(CONFIG_X86)
64 switch (boot_cpu_data.x86_vendor) {
65 case X86_VENDOR_HYGON:
66 case X86_VENDOR_AMD:
67 case X86_VENDOR_INTEL:
68 case X86_VENDOR_ZHAOXIN:
69 /*
70 * AMD Fam10h TSC will tick in all
71 * C/P/S0/S1 states when this bit is set.
72 */
73 if (!boot_cpu_has(X86_FEATURE_NONSTOP_TSC))
74 tsc_detected_unstable = 1;
75 break;
76 default:
77 /* TSC could halt in idle */
78 tsc_detected_unstable = 1;
79 }
80#endif
81}
82
83static unsigned long cpu_weight[NR_CPUS];
84static int tsk_in_cpu[NR_CPUS] = {[0 ... NR_CPUS-1] = -1};
85static DECLARE_BITMAP(pad_busy_cpus_bits, NR_CPUS);
86static void round_robin_cpu(unsigned int tsk_index)
87{
88 struct cpumask *pad_busy_cpus = to_cpumask(pad_busy_cpus_bits);
89 cpumask_var_t tmp;
90 int cpu;
91 unsigned long min_weight = -1;
92 unsigned long preferred_cpu;
93
94 if (!alloc_cpumask_var(&tmp, GFP_KERNEL))
95 return;
96
97 mutex_lock(&round_robin_lock);
98 cpumask_clear(tmp);
99 for_each_cpu(cpu, pad_busy_cpus)
100 cpumask_or(tmp, tmp, topology_sibling_cpumask(cpu));
101 cpumask_andnot(tmp, cpu_online_mask, tmp);
102 /* avoid HT sibilings if possible */
103 if (cpumask_empty(tmp))
104 cpumask_andnot(tmp, cpu_online_mask, pad_busy_cpus);
105 if (cpumask_empty(tmp)) {
106 mutex_unlock(&round_robin_lock);
107 free_cpumask_var(tmp);
108 return;
109 }
110 for_each_cpu(cpu, tmp) {
111 if (cpu_weight[cpu] < min_weight) {
112 min_weight = cpu_weight[cpu];
113 preferred_cpu = cpu;
114 }
115 }
116
117 if (tsk_in_cpu[tsk_index] != -1)
118 cpumask_clear_cpu(tsk_in_cpu[tsk_index], pad_busy_cpus);
119 tsk_in_cpu[tsk_index] = preferred_cpu;
120 cpumask_set_cpu(preferred_cpu, pad_busy_cpus);
121 cpu_weight[preferred_cpu]++;
122 mutex_unlock(&round_robin_lock);
123
124 set_cpus_allowed_ptr(current, cpumask_of(preferred_cpu));
125
126 free_cpumask_var(tmp);
127}
128
129static void exit_round_robin(unsigned int tsk_index)
130{
131 struct cpumask *pad_busy_cpus = to_cpumask(pad_busy_cpus_bits);
132
133 cpumask_clear_cpu(tsk_in_cpu[tsk_index], pad_busy_cpus);
134 tsk_in_cpu[tsk_index] = -1;
135}
136
137static unsigned int idle_pct = 5; /* percentage */
138static unsigned int round_robin_time = 1; /* second */
139static int power_saving_thread(void *data)
140{
141 int do_sleep;
142 unsigned int tsk_index = (unsigned long)data;
143 u64 last_jiffies = 0;
144
145 sched_set_fifo_low(current);
146
147 while (!kthread_should_stop()) {
148 unsigned long expire_time;
149
150 /* round robin to cpus */
151 expire_time = last_jiffies + round_robin_time * HZ;
152 if (time_before(expire_time, jiffies)) {
153 last_jiffies = jiffies;
154 round_robin_cpu(tsk_index);
155 }
156
157 do_sleep = 0;
158
159 expire_time = jiffies + HZ * (100 - idle_pct) / 100;
160
161 while (!need_resched()) {
162 if (tsc_detected_unstable && !tsc_marked_unstable) {
163 /* TSC could halt in idle, so notify users */
164 mark_tsc_unstable("TSC halts in idle");
165 tsc_marked_unstable = 1;
166 }
167 local_irq_disable();
168
169 perf_lopwr_cb(true);
170
171 tick_broadcast_enable();
172 tick_broadcast_enter();
173 stop_critical_timings();
174
175 mwait_idle_with_hints(power_saving_mwait_eax, 1);
176
177 start_critical_timings();
178 tick_broadcast_exit();
179
180 perf_lopwr_cb(false);
181
182 local_irq_enable();
183
184 if (time_before(expire_time, jiffies)) {
185 do_sleep = 1;
186 break;
187 }
188 }
189
190 /*
191 * current sched_rt has threshold for rt task running time.
192 * When a rt task uses 95% CPU time, the rt thread will be
193 * scheduled out for 5% CPU time to not starve other tasks. But
194 * the mechanism only works when all CPUs have RT task running,
195 * as if one CPU hasn't RT task, RT task from other CPUs will
196 * borrow CPU time from this CPU and cause RT task use > 95%
197 * CPU time. To make 'avoid starvation' work, takes a nap here.
198 */
199 if (unlikely(do_sleep))
200 schedule_timeout_killable(HZ * idle_pct / 100);
201
202 /* If an external event has set the need_resched flag, then
203 * we need to deal with it, or this loop will continue to
204 * spin without calling __mwait().
205 */
206 if (unlikely(need_resched()))
207 schedule();
208 }
209
210 exit_round_robin(tsk_index);
211 return 0;
212}
213
214static struct task_struct *ps_tsks[NR_CPUS];
215static unsigned int ps_tsk_num;
216static int create_power_saving_task(void)
217{
218 int rc;
219
220 ps_tsks[ps_tsk_num] = kthread_run(power_saving_thread,
221 (void *)(unsigned long)ps_tsk_num,
222 "acpi_pad/%d", ps_tsk_num);
223
224 if (IS_ERR(ps_tsks[ps_tsk_num])) {
225 rc = PTR_ERR(ps_tsks[ps_tsk_num]);
226 ps_tsks[ps_tsk_num] = NULL;
227 } else {
228 rc = 0;
229 ps_tsk_num++;
230 }
231
232 return rc;
233}
234
235static void destroy_power_saving_task(void)
236{
237 if (ps_tsk_num > 0) {
238 ps_tsk_num--;
239 kthread_stop(ps_tsks[ps_tsk_num]);
240 ps_tsks[ps_tsk_num] = NULL;
241 }
242}
243
244static void set_power_saving_task_num(unsigned int num)
245{
246 if (num > ps_tsk_num) {
247 while (ps_tsk_num < num) {
248 if (create_power_saving_task())
249 return;
250 }
251 } else if (num < ps_tsk_num) {
252 while (ps_tsk_num > num)
253 destroy_power_saving_task();
254 }
255}
256
257static void acpi_pad_idle_cpus(unsigned int num_cpus)
258{
259 cpus_read_lock();
260
261 num_cpus = min_t(unsigned int, num_cpus, num_online_cpus());
262 set_power_saving_task_num(num_cpus);
263
264 cpus_read_unlock();
265}
266
267static uint32_t acpi_pad_idle_cpus_num(void)
268{
269 return ps_tsk_num;
270}
271
272static ssize_t rrtime_store(struct device *dev,
273 struct device_attribute *attr, const char *buf, size_t count)
274{
275 unsigned long num;
276
277 if (kstrtoul(buf, 0, &num))
278 return -EINVAL;
279 if (num < 1 || num >= 100)
280 return -EINVAL;
281 mutex_lock(&isolated_cpus_lock);
282 round_robin_time = num;
283 mutex_unlock(&isolated_cpus_lock);
284 return count;
285}
286
287static ssize_t rrtime_show(struct device *dev,
288 struct device_attribute *attr, char *buf)
289{
290 return sysfs_emit(buf, "%d\n", round_robin_time);
291}
292static DEVICE_ATTR_RW(rrtime);
293
294static ssize_t idlepct_store(struct device *dev,
295 struct device_attribute *attr, const char *buf, size_t count)
296{
297 unsigned long num;
298
299 if (kstrtoul(buf, 0, &num))
300 return -EINVAL;
301 if (num < 1 || num >= 100)
302 return -EINVAL;
303 mutex_lock(&isolated_cpus_lock);
304 idle_pct = num;
305 mutex_unlock(&isolated_cpus_lock);
306 return count;
307}
308
309static ssize_t idlepct_show(struct device *dev,
310 struct device_attribute *attr, char *buf)
311{
312 return sysfs_emit(buf, "%d\n", idle_pct);
313}
314static DEVICE_ATTR_RW(idlepct);
315
316static ssize_t idlecpus_store(struct device *dev,
317 struct device_attribute *attr, const char *buf, size_t count)
318{
319 unsigned long num;
320
321 if (kstrtoul(buf, 0, &num))
322 return -EINVAL;
323 mutex_lock(&isolated_cpus_lock);
324 acpi_pad_idle_cpus(num);
325 mutex_unlock(&isolated_cpus_lock);
326 return count;
327}
328
329static ssize_t idlecpus_show(struct device *dev,
330 struct device_attribute *attr, char *buf)
331{
332 return cpumap_print_to_pagebuf(false, buf,
333 to_cpumask(pad_busy_cpus_bits));
334}
335
336static DEVICE_ATTR_RW(idlecpus);
337
338static int acpi_pad_add_sysfs(struct acpi_device *device)
339{
340 int result;
341
342 result = device_create_file(&device->dev, &dev_attr_idlecpus);
343 if (result)
344 return -ENODEV;
345 result = device_create_file(&device->dev, &dev_attr_idlepct);
346 if (result) {
347 device_remove_file(&device->dev, &dev_attr_idlecpus);
348 return -ENODEV;
349 }
350 result = device_create_file(&device->dev, &dev_attr_rrtime);
351 if (result) {
352 device_remove_file(&device->dev, &dev_attr_idlecpus);
353 device_remove_file(&device->dev, &dev_attr_idlepct);
354 return -ENODEV;
355 }
356 return 0;
357}
358
359static void acpi_pad_remove_sysfs(struct acpi_device *device)
360{
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);
364}
365
366/*
367 * Query firmware how many CPUs should be idle
368 * return -1 on failure
369 */
370static int acpi_pad_pur(acpi_handle handle)
371{
372 struct acpi_buffer buffer = {ACPI_ALLOCATE_BUFFER, NULL};
373 union acpi_object *package;
374 int num = -1;
375
376 if (ACPI_FAILURE(acpi_evaluate_object(handle, "_PUR", NULL, &buffer)))
377 return num;
378
379 if (!buffer.length || !buffer.pointer)
380 return num;
381
382 package = buffer.pointer;
383
384 if (package->type == ACPI_TYPE_PACKAGE &&
385 package->package.count == 2 &&
386 package->package.elements[0].integer.value == 1) /* rev 1 */
387
388 num = package->package.elements[1].integer.value;
389
390 kfree(buffer.pointer);
391 return num;
392}
393
394static void acpi_pad_handle_notify(acpi_handle handle)
395{
396 int num_cpus;
397 uint32_t idle_cpus;
398 struct acpi_buffer param = {
399 .length = 4,
400 .pointer = (void *)&idle_cpus,
401 };
402
403 mutex_lock(&isolated_cpus_lock);
404 num_cpus = acpi_pad_pur(handle);
405 if (num_cpus < 0) {
406 mutex_unlock(&isolated_cpus_lock);
407 return;
408 }
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);
413}
414
415static void acpi_pad_notify(acpi_handle handle, u32 event,
416 void *data)
417{
418 struct acpi_device *device = data;
419
420 switch (event) {
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);
425 break;
426 default:
427 pr_warn("Unsupported event [0x%x]\n", event);
428 break;
429 }
430}
431
432static int acpi_pad_add(struct acpi_device *device)
433{
434 acpi_status status;
435
436 strcpy(acpi_device_name(device), ACPI_PROCESSOR_AGGREGATOR_DEVICE_NAME);
437 strcpy(acpi_device_class(device), ACPI_PROCESSOR_AGGREGATOR_CLASS);
438
439 if (acpi_pad_add_sysfs(device))
440 return -ENODEV;
441
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);
446 return -ENODEV;
447 }
448
449 return 0;
450}
451
452static void acpi_pad_remove(struct acpi_device *device)
453{
454 mutex_lock(&isolated_cpus_lock);
455 acpi_pad_idle_cpus(0);
456 mutex_unlock(&isolated_cpus_lock);
457
458 acpi_remove_notify_handler(device->handle,
459 ACPI_DEVICE_NOTIFY, acpi_pad_notify);
460 acpi_pad_remove_sysfs(device);
461}
462
463static const struct acpi_device_id pad_device_ids[] = {
464 {"ACPI000C", 0},
465 {"", 0},
466};
467MODULE_DEVICE_TABLE(acpi, pad_device_ids);
468
469static struct acpi_driver acpi_pad_driver = {
470 .name = "processor_aggregator",
471 .class = ACPI_PROCESSOR_AGGREGATOR_CLASS,
472 .ids = pad_device_ids,
473 .ops = {
474 .add = acpi_pad_add,
475 .remove = acpi_pad_remove,
476 },
477};
478
479static int __init acpi_pad_init(void)
480{
481 /* Xen ACPI PAD is used when running as Xen Dom0. */
482 if (xen_initial_domain())
483 return -ENODEV;
484
485 power_saving_mwait_init();
486 if (power_saving_mwait_eax == 0)
487 return -EINVAL;
488
489 return acpi_bus_register_driver(&acpi_pad_driver);
490}
491
492static void __exit acpi_pad_exit(void)
493{
494 acpi_bus_unregister_driver(&acpi_pad_driver);
495}
496
497module_init(acpi_pad_init);
498module_exit(acpi_pad_exit);
499MODULE_AUTHOR("Shaohua Li<shaohua.li@intel.com>");
500MODULE_DESCRIPTION("ACPI Processor Aggregator Driver");
501MODULE_LICENSE("GPL");