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
27 static DEFINE_MUTEX(isolated_cpus_lock);
28 static DEFINE_MUTEX(round_robin_lock);
29
30 static unsigned long power_saving_mwait_eax;
31
32 static unsigned char tsc_detected_unstable;
33 static unsigned char tsc_marked_unstable;
34
power_saving_mwait_init(void)35 static 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 case X86_VENDOR_CENTAUR:
70 /*
71 * AMD Fam10h TSC will tick in all
72 * C/P/S0/S1 states when this bit is set.
73 */
74 if (!boot_cpu_has(X86_FEATURE_NONSTOP_TSC))
75 tsc_detected_unstable = 1;
76 break;
77 default:
78 /* TSC could halt in idle */
79 tsc_detected_unstable = 1;
80 }
81 #endif
82 }
83
84 static unsigned long cpu_weight[NR_CPUS];
85 static int tsk_in_cpu[NR_CPUS] = {[0 ... NR_CPUS-1] = -1};
86 static DECLARE_BITMAP(pad_busy_cpus_bits, NR_CPUS);
round_robin_cpu(unsigned int tsk_index)87 static void round_robin_cpu(unsigned int tsk_index)
88 {
89 struct cpumask *pad_busy_cpus = to_cpumask(pad_busy_cpus_bits);
90 cpumask_var_t tmp;
91 int cpu;
92 unsigned long min_weight = -1;
93 unsigned long preferred_cpu;
94
95 if (!alloc_cpumask_var(&tmp, GFP_KERNEL))
96 return;
97
98 mutex_lock(&round_robin_lock);
99 cpumask_clear(tmp);
100 for_each_cpu(cpu, pad_busy_cpus)
101 cpumask_or(tmp, tmp, topology_sibling_cpumask(cpu));
102 cpumask_andnot(tmp, cpu_online_mask, tmp);
103 /* avoid HT sibilings if possible */
104 if (cpumask_empty(tmp))
105 cpumask_andnot(tmp, cpu_online_mask, pad_busy_cpus);
106 if (cpumask_empty(tmp)) {
107 mutex_unlock(&round_robin_lock);
108 free_cpumask_var(tmp);
109 return;
110 }
111 for_each_cpu(cpu, tmp) {
112 if (cpu_weight[cpu] < min_weight) {
113 min_weight = cpu_weight[cpu];
114 preferred_cpu = cpu;
115 }
116 }
117
118 if (tsk_in_cpu[tsk_index] != -1)
119 cpumask_clear_cpu(tsk_in_cpu[tsk_index], pad_busy_cpus);
120 tsk_in_cpu[tsk_index] = preferred_cpu;
121 cpumask_set_cpu(preferred_cpu, pad_busy_cpus);
122 cpu_weight[preferred_cpu]++;
123 mutex_unlock(&round_robin_lock);
124
125 set_cpus_allowed_ptr(current, cpumask_of(preferred_cpu));
126
127 free_cpumask_var(tmp);
128 }
129
exit_round_robin(unsigned int tsk_index)130 static void exit_round_robin(unsigned int tsk_index)
131 {
132 struct cpumask *pad_busy_cpus = to_cpumask(pad_busy_cpus_bits);
133
134 if (tsk_in_cpu[tsk_index] != -1) {
135 cpumask_clear_cpu(tsk_in_cpu[tsk_index], pad_busy_cpus);
136 tsk_in_cpu[tsk_index] = -1;
137 }
138 }
139
140 static unsigned int idle_pct = 5; /* percentage */
141 static unsigned int round_robin_time = 1; /* second */
power_saving_thread(void * data)142 static int power_saving_thread(void *data)
143 {
144 int do_sleep;
145 unsigned int tsk_index = (unsigned long)data;
146 u64 last_jiffies = 0;
147
148 sched_set_fifo_low(current);
149
150 while (!kthread_should_stop()) {
151 unsigned long expire_time;
152
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);
158 }
159
160 do_sleep = 0;
161
162 expire_time = jiffies + HZ * (100 - idle_pct) / 100;
163
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;
169 }
170 local_irq_disable();
171
172 perf_lopwr_cb(true);
173
174 tick_broadcast_enable();
175 tick_broadcast_enter();
176 stop_critical_timings();
177
178 mwait_idle_with_hints(power_saving_mwait_eax, 1);
179
180 start_critical_timings();
181 tick_broadcast_exit();
182
183 perf_lopwr_cb(false);
184
185 local_irq_enable();
186
187 if (time_before(expire_time, jiffies)) {
188 do_sleep = 1;
189 break;
190 }
191 }
192
193 /*
194 * current sched_rt has threshold for rt task running time.
195 * When a rt task uses 95% CPU time, the rt thread will be
196 * scheduled out for 5% CPU time to not starve other tasks. But
197 * the mechanism only works when all CPUs have RT task running,
198 * as if one CPU hasn't RT task, RT task from other CPUs will
199 * borrow CPU time from this CPU and cause RT task use > 95%
200 * CPU time. To make 'avoid starvation' work, takes a nap here.
201 */
202 if (unlikely(do_sleep))
203 schedule_timeout_killable(HZ * idle_pct / 100);
204
205 /* If an external event has set the need_resched flag, then
206 * we need to deal with it, or this loop will continue to
207 * spin without calling __mwait().
208 */
209 if (unlikely(need_resched()))
210 schedule();
211 }
212
213 exit_round_robin(tsk_index);
214 return 0;
215 }
216
217 static struct task_struct *ps_tsks[NR_CPUS];
218 static unsigned int ps_tsk_num;
create_power_saving_task(void)219 static int create_power_saving_task(void)
220 {
221 int rc;
222
223 ps_tsks[ps_tsk_num] = kthread_run(power_saving_thread,
224 (void *)(unsigned long)ps_tsk_num,
225 "acpi_pad/%d", ps_tsk_num);
226
227 if (IS_ERR(ps_tsks[ps_tsk_num])) {
228 rc = PTR_ERR(ps_tsks[ps_tsk_num]);
229 ps_tsks[ps_tsk_num] = NULL;
230 } else {
231 rc = 0;
232 ps_tsk_num++;
233 }
234
235 return rc;
236 }
237
destroy_power_saving_task(void)238 static 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
set_power_saving_task_num(unsigned int num)247 static 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
acpi_pad_idle_cpus(unsigned int num_cpus)260 static void acpi_pad_idle_cpus(unsigned int num_cpus)
261 {
262 cpus_read_lock();
263
264 num_cpus = min_t(unsigned int, num_cpus, num_online_cpus());
265 set_power_saving_task_num(num_cpus);
266
267 cpus_read_unlock();
268 }
269
acpi_pad_idle_cpus_num(void)270 static uint32_t acpi_pad_idle_cpus_num(void)
271 {
272 return ps_tsk_num;
273 }
274
rrtime_store(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)275 static ssize_t rrtime_store(struct device *dev,
276 struct device_attribute *attr, const char *buf, size_t count)
277 {
278 unsigned long num;
279
280 if (kstrtoul(buf, 0, &num))
281 return -EINVAL;
282 if (num < 1 || num >= 100)
283 return -EINVAL;
284 mutex_lock(&isolated_cpus_lock);
285 round_robin_time = num;
286 mutex_unlock(&isolated_cpus_lock);
287 return count;
288 }
289
rrtime_show(struct device * dev,struct device_attribute * attr,char * buf)290 static ssize_t rrtime_show(struct device *dev,
291 struct device_attribute *attr, char *buf)
292 {
293 return sysfs_emit(buf, "%d\n", round_robin_time);
294 }
295 static DEVICE_ATTR_RW(rrtime);
296
idlepct_store(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)297 static ssize_t idlepct_store(struct device *dev,
298 struct device_attribute *attr, const char *buf, size_t count)
299 {
300 unsigned long num;
301
302 if (kstrtoul(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
idlepct_show(struct device * dev,struct device_attribute * attr,char * buf)312 static ssize_t idlepct_show(struct device *dev,
313 struct device_attribute *attr, char *buf)
314 {
315 return sysfs_emit(buf, "%d\n", idle_pct);
316 }
317 static DEVICE_ATTR_RW(idlepct);
318
idlecpus_store(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)319 static ssize_t idlecpus_store(struct device *dev,
320 struct device_attribute *attr, const char *buf, size_t count)
321 {
322 unsigned long num;
323
324 if (kstrtoul(buf, 0, &num))
325 return -EINVAL;
326 mutex_lock(&isolated_cpus_lock);
327 acpi_pad_idle_cpus(num);
328 mutex_unlock(&isolated_cpus_lock);
329 return count;
330 }
331
idlecpus_show(struct device * dev,struct device_attribute * attr,char * buf)332 static ssize_t idlecpus_show(struct device *dev,
333 struct device_attribute *attr, char *buf)
334 {
335 return cpumap_print_to_pagebuf(false, buf,
336 to_cpumask(pad_busy_cpus_bits));
337 }
338
339 static DEVICE_ATTR_RW(idlecpus);
340
acpi_pad_add_sysfs(struct acpi_device * device)341 static int acpi_pad_add_sysfs(struct acpi_device *device)
342 {
343 int result;
344
345 result = device_create_file(&device->dev, &dev_attr_idlecpus);
346 if (result)
347 return -ENODEV;
348 result = device_create_file(&device->dev, &dev_attr_idlepct);
349 if (result) {
350 device_remove_file(&device->dev, &dev_attr_idlecpus);
351 return -ENODEV;
352 }
353 result = device_create_file(&device->dev, &dev_attr_rrtime);
354 if (result) {
355 device_remove_file(&device->dev, &dev_attr_idlecpus);
356 device_remove_file(&device->dev, &dev_attr_idlepct);
357 return -ENODEV;
358 }
359 return 0;
360 }
361
acpi_pad_remove_sysfs(struct acpi_device * device)362 static void acpi_pad_remove_sysfs(struct acpi_device *device)
363 {
364 device_remove_file(&device->dev, &dev_attr_idlecpus);
365 device_remove_file(&device->dev, &dev_attr_idlepct);
366 device_remove_file(&device->dev, &dev_attr_rrtime);
367 }
368
369 /*
370 * Query firmware how many CPUs should be idle
371 * return -1 on failure
372 */
acpi_pad_pur(acpi_handle handle)373 static int acpi_pad_pur(acpi_handle handle)
374 {
375 struct acpi_buffer buffer = {ACPI_ALLOCATE_BUFFER, NULL};
376 union acpi_object *package;
377 int num = -1;
378
379 if (ACPI_FAILURE(acpi_evaluate_object(handle, "_PUR", NULL, &buffer)))
380 return num;
381
382 if (!buffer.length || !buffer.pointer)
383 return num;
384
385 package = buffer.pointer;
386
387 if (package->type == ACPI_TYPE_PACKAGE &&
388 package->package.count == 2 &&
389 package->package.elements[0].integer.value == 1) /* rev 1 */
390
391 num = package->package.elements[1].integer.value;
392
393 kfree(buffer.pointer);
394 return num;
395 }
396
acpi_pad_handle_notify(acpi_handle handle)397 static void acpi_pad_handle_notify(acpi_handle handle)
398 {
399 int num_cpus;
400 uint32_t idle_cpus;
401 struct acpi_buffer param = {
402 .length = 4,
403 .pointer = (void *)&idle_cpus,
404 };
405
406 mutex_lock(&isolated_cpus_lock);
407 num_cpus = acpi_pad_pur(handle);
408 if (num_cpus < 0) {
409 mutex_unlock(&isolated_cpus_lock);
410 return;
411 }
412 acpi_pad_idle_cpus(num_cpus);
413 idle_cpus = acpi_pad_idle_cpus_num();
414 acpi_evaluate_ost(handle, ACPI_PROCESSOR_AGGREGATOR_NOTIFY, 0, ¶m);
415 mutex_unlock(&isolated_cpus_lock);
416 }
417
acpi_pad_notify(acpi_handle handle,u32 event,void * data)418 static void acpi_pad_notify(acpi_handle handle, u32 event,
419 void *data)
420 {
421 struct acpi_device *device = data;
422
423 switch (event) {
424 case ACPI_PROCESSOR_AGGREGATOR_NOTIFY:
425 acpi_pad_handle_notify(handle);
426 acpi_bus_generate_netlink_event(device->pnp.device_class,
427 dev_name(&device->dev), event, 0);
428 break;
429 default:
430 pr_warn("Unsupported event [0x%x]\n", event);
431 break;
432 }
433 }
434
acpi_pad_add(struct acpi_device * device)435 static int acpi_pad_add(struct acpi_device *device)
436 {
437 acpi_status status;
438
439 strcpy(acpi_device_name(device), ACPI_PROCESSOR_AGGREGATOR_DEVICE_NAME);
440 strcpy(acpi_device_class(device), ACPI_PROCESSOR_AGGREGATOR_CLASS);
441
442 if (acpi_pad_add_sysfs(device))
443 return -ENODEV;
444
445 status = acpi_install_notify_handler(device->handle,
446 ACPI_DEVICE_NOTIFY, acpi_pad_notify, device);
447 if (ACPI_FAILURE(status)) {
448 acpi_pad_remove_sysfs(device);
449 return -ENODEV;
450 }
451
452 return 0;
453 }
454
acpi_pad_remove(struct acpi_device * device)455 static void acpi_pad_remove(struct acpi_device *device)
456 {
457 mutex_lock(&isolated_cpus_lock);
458 acpi_pad_idle_cpus(0);
459 mutex_unlock(&isolated_cpus_lock);
460
461 acpi_remove_notify_handler(device->handle,
462 ACPI_DEVICE_NOTIFY, acpi_pad_notify);
463 acpi_pad_remove_sysfs(device);
464 }
465
466 static const struct acpi_device_id pad_device_ids[] = {
467 {"ACPI000C", 0},
468 {"", 0},
469 };
470 MODULE_DEVICE_TABLE(acpi, pad_device_ids);
471
472 static struct acpi_driver acpi_pad_driver = {
473 .name = "processor_aggregator",
474 .class = ACPI_PROCESSOR_AGGREGATOR_CLASS,
475 .ids = pad_device_ids,
476 .ops = {
477 .add = acpi_pad_add,
478 .remove = acpi_pad_remove,
479 },
480 };
481
acpi_pad_init(void)482 static int __init acpi_pad_init(void)
483 {
484 /* Xen ACPI PAD is used when running as Xen Dom0. */
485 if (xen_initial_domain())
486 return -ENODEV;
487
488 power_saving_mwait_init();
489 if (power_saving_mwait_eax == 0)
490 return -EINVAL;
491
492 return acpi_bus_register_driver(&acpi_pad_driver);
493 }
494
acpi_pad_exit(void)495 static void __exit acpi_pad_exit(void)
496 {
497 acpi_bus_unregister_driver(&acpi_pad_driver);
498 }
499
500 module_init(acpi_pad_init);
501 module_exit(acpi_pad_exit);
502 MODULE_AUTHOR("Shaohua Li<shaohua.li@intel.com>");
503 MODULE_DESCRIPTION("ACPI Processor Aggregator Driver");
504 MODULE_LICENSE("GPL");
505