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