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 <linux/platform_device.h>
22 #include <asm/mwait.h>
23 #include <xen/xen.h>
24
25 #define ACPI_PROCESSOR_AGGREGATOR_CLASS "acpi_pad"
26 #define ACPI_PROCESSOR_AGGREGATOR_DEVICE_NAME "Processor Aggregator"
27 #define ACPI_PROCESSOR_AGGREGATOR_NOTIFY 0x80
28
29 #define ACPI_PROCESSOR_AGGREGATOR_STATUS_SUCCESS 0
30 #define ACPI_PROCESSOR_AGGREGATOR_STATUS_NO_ACTION 1
31
32 static DEFINE_MUTEX(isolated_cpus_lock);
33 static DEFINE_MUTEX(round_robin_lock);
34
35 static unsigned long power_saving_mwait_eax;
36
37 static unsigned char tsc_detected_unstable;
38 static unsigned char tsc_marked_unstable;
39
power_saving_mwait_init(void)40 static void power_saving_mwait_init(void)
41 {
42 unsigned int eax, ebx, ecx, edx;
43 unsigned int highest_cstate = 0;
44 unsigned int highest_subcstate = 0;
45 int i;
46
47 if (!boot_cpu_has(X86_FEATURE_MWAIT))
48 return;
49 if (boot_cpu_data.cpuid_level < CPUID_MWAIT_LEAF)
50 return;
51
52 cpuid(CPUID_MWAIT_LEAF, &eax, &ebx, &ecx, &edx);
53
54 if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED) ||
55 !(ecx & CPUID5_ECX_INTERRUPT_BREAK))
56 return;
57
58 edx >>= MWAIT_SUBSTATE_SIZE;
59 for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) {
60 if (edx & MWAIT_SUBSTATE_MASK) {
61 highest_cstate = i;
62 highest_subcstate = edx & MWAIT_SUBSTATE_MASK;
63 }
64 }
65 power_saving_mwait_eax = (highest_cstate << MWAIT_SUBSTATE_SIZE) |
66 (highest_subcstate - 1);
67
68 #if defined(CONFIG_X86)
69 switch (boot_cpu_data.x86_vendor) {
70 case X86_VENDOR_HYGON:
71 case X86_VENDOR_AMD:
72 case X86_VENDOR_INTEL:
73 case X86_VENDOR_ZHAOXIN:
74 case X86_VENDOR_CENTAUR:
75 /*
76 * AMD Fam10h TSC will tick in all
77 * C/P/S0/S1 states when this bit is set.
78 */
79 if (!boot_cpu_has(X86_FEATURE_NONSTOP_TSC))
80 tsc_detected_unstable = 1;
81 break;
82 default:
83 /* TSC could halt in idle */
84 tsc_detected_unstable = 1;
85 }
86 #endif
87 }
88
89 static unsigned long cpu_weight[NR_CPUS];
90 static int tsk_in_cpu[NR_CPUS] = {[0 ... NR_CPUS-1] = -1};
91 static DECLARE_BITMAP(pad_busy_cpus_bits, NR_CPUS);
round_robin_cpu(unsigned int tsk_index)92 static void round_robin_cpu(unsigned int tsk_index)
93 {
94 struct cpumask *pad_busy_cpus = to_cpumask(pad_busy_cpus_bits);
95 cpumask_var_t tmp;
96 int cpu;
97 unsigned long min_weight = -1;
98 unsigned long preferred_cpu;
99
100 if (!alloc_cpumask_var(&tmp, GFP_KERNEL))
101 return;
102
103 mutex_lock(&round_robin_lock);
104 cpumask_clear(tmp);
105 for_each_cpu(cpu, pad_busy_cpus)
106 cpumask_or(tmp, tmp, topology_sibling_cpumask(cpu));
107 cpumask_andnot(tmp, cpu_online_mask, tmp);
108 /* avoid HT siblings if possible */
109 if (cpumask_empty(tmp))
110 cpumask_andnot(tmp, cpu_online_mask, pad_busy_cpus);
111 if (cpumask_empty(tmp)) {
112 mutex_unlock(&round_robin_lock);
113 free_cpumask_var(tmp);
114 return;
115 }
116 for_each_cpu(cpu, tmp) {
117 if (cpu_weight[cpu] < min_weight) {
118 min_weight = cpu_weight[cpu];
119 preferred_cpu = cpu;
120 }
121 }
122
123 if (tsk_in_cpu[tsk_index] != -1)
124 cpumask_clear_cpu(tsk_in_cpu[tsk_index], pad_busy_cpus);
125 tsk_in_cpu[tsk_index] = preferred_cpu;
126 cpumask_set_cpu(preferred_cpu, pad_busy_cpus);
127 cpu_weight[preferred_cpu]++;
128 mutex_unlock(&round_robin_lock);
129
130 set_cpus_allowed_ptr(current, cpumask_of(preferred_cpu));
131
132 free_cpumask_var(tmp);
133 }
134
exit_round_robin(unsigned int tsk_index)135 static void exit_round_robin(unsigned int tsk_index)
136 {
137 struct cpumask *pad_busy_cpus = to_cpumask(pad_busy_cpus_bits);
138
139 if (tsk_in_cpu[tsk_index] != -1) {
140 cpumask_clear_cpu(tsk_in_cpu[tsk_index], pad_busy_cpus);
141 tsk_in_cpu[tsk_index] = -1;
142 }
143 }
144
145 static unsigned int idle_pct = 5; /* percentage */
146 static unsigned int round_robin_time = 1; /* second */
power_saving_thread(void * data)147 static int power_saving_thread(void *data)
148 {
149 int do_sleep;
150 unsigned int tsk_index = (unsigned long)data;
151 u64 last_jiffies = 0;
152
153 sched_set_fifo_low(current);
154
155 while (!kthread_should_stop()) {
156 unsigned long expire_time;
157
158 /* round robin to cpus */
159 expire_time = last_jiffies + round_robin_time * HZ;
160 if (time_before(expire_time, jiffies)) {
161 last_jiffies = jiffies;
162 round_robin_cpu(tsk_index);
163 }
164
165 do_sleep = 0;
166
167 expire_time = jiffies + HZ * (100 - idle_pct) / 100;
168
169 while (!need_resched()) {
170 if (tsc_detected_unstable && !tsc_marked_unstable) {
171 /* TSC could halt in idle, so notify users */
172 mark_tsc_unstable("TSC halts in idle");
173 tsc_marked_unstable = 1;
174 }
175 local_irq_disable();
176
177 perf_lopwr_cb(true);
178
179 tick_broadcast_enable();
180 tick_broadcast_enter();
181 stop_critical_timings();
182
183 mwait_idle_with_hints(power_saving_mwait_eax, 1);
184
185 start_critical_timings();
186 tick_broadcast_exit();
187
188 perf_lopwr_cb(false);
189
190 local_irq_enable();
191
192 if (time_before(expire_time, jiffies)) {
193 do_sleep = 1;
194 break;
195 }
196 }
197
198 /*
199 * current sched_rt has threshold for rt task running time.
200 * When a rt task uses 95% CPU time, the rt thread will be
201 * scheduled out for 5% CPU time to not starve other tasks. But
202 * the mechanism only works when all CPUs have RT task running,
203 * as if one CPU hasn't RT task, RT task from other CPUs will
204 * borrow CPU time from this CPU and cause RT task use > 95%
205 * CPU time. To make 'avoid starvation' work, takes a nap here.
206 */
207 if (unlikely(do_sleep))
208 schedule_timeout_killable(HZ * idle_pct / 100);
209
210 /* If an external event has set the need_resched flag, then
211 * we need to deal with it, or this loop will continue to
212 * spin without calling __mwait().
213 */
214 if (unlikely(need_resched()))
215 schedule();
216 }
217
218 exit_round_robin(tsk_index);
219 return 0;
220 }
221
222 static struct task_struct *ps_tsks[NR_CPUS];
223 static unsigned int ps_tsk_num;
create_power_saving_task(void)224 static int create_power_saving_task(void)
225 {
226 int rc;
227
228 ps_tsks[ps_tsk_num] = kthread_run(power_saving_thread,
229 (void *)(unsigned long)ps_tsk_num,
230 "acpi_pad/%d", ps_tsk_num);
231
232 if (IS_ERR(ps_tsks[ps_tsk_num])) {
233 rc = PTR_ERR(ps_tsks[ps_tsk_num]);
234 ps_tsks[ps_tsk_num] = NULL;
235 } else {
236 rc = 0;
237 ps_tsk_num++;
238 }
239
240 return rc;
241 }
242
destroy_power_saving_task(void)243 static void destroy_power_saving_task(void)
244 {
245 if (ps_tsk_num > 0) {
246 ps_tsk_num--;
247 kthread_stop(ps_tsks[ps_tsk_num]);
248 ps_tsks[ps_tsk_num] = NULL;
249 }
250 }
251
set_power_saving_task_num(unsigned int num)252 static void set_power_saving_task_num(unsigned int num)
253 {
254 if (num > ps_tsk_num) {
255 while (ps_tsk_num < num) {
256 if (create_power_saving_task())
257 return;
258 }
259 } else if (num < ps_tsk_num) {
260 while (ps_tsk_num > num)
261 destroy_power_saving_task();
262 }
263 }
264
acpi_pad_idle_cpus(unsigned int num_cpus)265 static void acpi_pad_idle_cpus(unsigned int num_cpus)
266 {
267 cpus_read_lock();
268
269 num_cpus = min_t(unsigned int, num_cpus, num_online_cpus());
270 set_power_saving_task_num(num_cpus);
271
272 cpus_read_unlock();
273 }
274
acpi_pad_idle_cpus_num(void)275 static uint32_t acpi_pad_idle_cpus_num(void)
276 {
277 return ps_tsk_num;
278 }
279
rrtime_store(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)280 static ssize_t rrtime_store(struct device *dev,
281 struct device_attribute *attr, const char *buf, size_t count)
282 {
283 unsigned long num;
284
285 if (kstrtoul(buf, 0, &num))
286 return -EINVAL;
287 if (num < 1 || num >= 100)
288 return -EINVAL;
289 mutex_lock(&isolated_cpus_lock);
290 round_robin_time = num;
291 mutex_unlock(&isolated_cpus_lock);
292 return count;
293 }
294
rrtime_show(struct device * dev,struct device_attribute * attr,char * buf)295 static ssize_t rrtime_show(struct device *dev,
296 struct device_attribute *attr, char *buf)
297 {
298 return sysfs_emit(buf, "%d\n", round_robin_time);
299 }
300 static DEVICE_ATTR_RW(rrtime);
301
idlepct_store(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)302 static ssize_t idlepct_store(struct device *dev,
303 struct device_attribute *attr, const char *buf, size_t count)
304 {
305 unsigned long num;
306
307 if (kstrtoul(buf, 0, &num))
308 return -EINVAL;
309 if (num < 1 || num >= 100)
310 return -EINVAL;
311 mutex_lock(&isolated_cpus_lock);
312 idle_pct = num;
313 mutex_unlock(&isolated_cpus_lock);
314 return count;
315 }
316
idlepct_show(struct device * dev,struct device_attribute * attr,char * buf)317 static ssize_t idlepct_show(struct device *dev,
318 struct device_attribute *attr, char *buf)
319 {
320 return sysfs_emit(buf, "%d\n", idle_pct);
321 }
322 static DEVICE_ATTR_RW(idlepct);
323
idlecpus_store(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)324 static ssize_t idlecpus_store(struct device *dev,
325 struct device_attribute *attr, const char *buf, size_t count)
326 {
327 unsigned long num;
328
329 if (kstrtoul(buf, 0, &num))
330 return -EINVAL;
331 mutex_lock(&isolated_cpus_lock);
332 acpi_pad_idle_cpus(num);
333 mutex_unlock(&isolated_cpus_lock);
334 return count;
335 }
336
idlecpus_show(struct device * dev,struct device_attribute * attr,char * buf)337 static ssize_t idlecpus_show(struct device *dev,
338 struct device_attribute *attr, char *buf)
339 {
340 return cpumap_print_to_pagebuf(false, buf,
341 to_cpumask(pad_busy_cpus_bits));
342 }
343
344 static DEVICE_ATTR_RW(idlecpus);
345
346 static struct attribute *acpi_pad_attrs[] = {
347 &dev_attr_idlecpus.attr,
348 &dev_attr_idlepct.attr,
349 &dev_attr_rrtime.attr,
350 NULL
351 };
352
353 ATTRIBUTE_GROUPS(acpi_pad);
354
355 /*
356 * Query firmware how many CPUs should be idle
357 * return -1 on failure
358 */
acpi_pad_pur(acpi_handle handle)359 static int acpi_pad_pur(acpi_handle handle)
360 {
361 struct acpi_buffer buffer = {ACPI_ALLOCATE_BUFFER, NULL};
362 union acpi_object *package;
363 int num = -1;
364
365 if (ACPI_FAILURE(acpi_evaluate_object(handle, "_PUR", NULL, &buffer)))
366 return num;
367
368 if (!buffer.length || !buffer.pointer)
369 return num;
370
371 package = buffer.pointer;
372
373 if (package->type == ACPI_TYPE_PACKAGE &&
374 package->package.count == 2 &&
375 package->package.elements[0].integer.value == 1) /* rev 1 */
376
377 num = package->package.elements[1].integer.value;
378
379 kfree(buffer.pointer);
380 return num;
381 }
382
acpi_pad_handle_notify(acpi_handle handle)383 static void acpi_pad_handle_notify(acpi_handle handle)
384 {
385 int num_cpus;
386 uint32_t idle_cpus;
387 struct acpi_buffer param = {
388 .length = 4,
389 .pointer = (void *)&idle_cpus,
390 };
391 u32 status;
392
393 mutex_lock(&isolated_cpus_lock);
394 num_cpus = acpi_pad_pur(handle);
395 if (num_cpus < 0) {
396 /* The ACPI specification says that if no action was performed when
397 * processing the _PUR object, _OST should still be evaluated, albeit
398 * with a different status code.
399 */
400 status = ACPI_PROCESSOR_AGGREGATOR_STATUS_NO_ACTION;
401 } else {
402 status = ACPI_PROCESSOR_AGGREGATOR_STATUS_SUCCESS;
403 acpi_pad_idle_cpus(num_cpus);
404 }
405
406 idle_cpus = acpi_pad_idle_cpus_num();
407 acpi_evaluate_ost(handle, ACPI_PROCESSOR_AGGREGATOR_NOTIFY, status, ¶m);
408 mutex_unlock(&isolated_cpus_lock);
409 }
410
acpi_pad_notify(acpi_handle handle,u32 event,void * data)411 static void acpi_pad_notify(acpi_handle handle, u32 event,
412 void *data)
413 {
414 struct acpi_device *adev = data;
415
416 switch (event) {
417 case ACPI_PROCESSOR_AGGREGATOR_NOTIFY:
418 acpi_pad_handle_notify(handle);
419 acpi_bus_generate_netlink_event(adev->pnp.device_class,
420 dev_name(&adev->dev), event, 0);
421 break;
422 default:
423 pr_warn("Unsupported event [0x%x]\n", event);
424 break;
425 }
426 }
427
acpi_pad_probe(struct platform_device * pdev)428 static int acpi_pad_probe(struct platform_device *pdev)
429 {
430 struct acpi_device *adev = ACPI_COMPANION(&pdev->dev);
431 acpi_status status;
432
433 strscpy(acpi_device_name(adev), ACPI_PROCESSOR_AGGREGATOR_DEVICE_NAME);
434 strscpy(acpi_device_class(adev), ACPI_PROCESSOR_AGGREGATOR_CLASS);
435
436 status = acpi_install_notify_handler(adev->handle,
437 ACPI_DEVICE_NOTIFY, acpi_pad_notify, adev);
438
439 if (ACPI_FAILURE(status))
440 return -ENODEV;
441
442 return 0;
443 }
444
acpi_pad_remove(struct platform_device * pdev)445 static void acpi_pad_remove(struct platform_device *pdev)
446 {
447 struct acpi_device *adev = ACPI_COMPANION(&pdev->dev);
448
449 mutex_lock(&isolated_cpus_lock);
450 acpi_pad_idle_cpus(0);
451 mutex_unlock(&isolated_cpus_lock);
452
453 acpi_remove_notify_handler(adev->handle,
454 ACPI_DEVICE_NOTIFY, acpi_pad_notify);
455 }
456
457 static const struct acpi_device_id pad_device_ids[] = {
458 {"ACPI000C", 0},
459 {"", 0},
460 };
461 MODULE_DEVICE_TABLE(acpi, pad_device_ids);
462
463 static struct platform_driver acpi_pad_driver = {
464 .probe = acpi_pad_probe,
465 .remove_new = acpi_pad_remove,
466 .driver = {
467 .dev_groups = acpi_pad_groups,
468 .name = "processor_aggregator",
469 .acpi_match_table = pad_device_ids,
470 },
471 };
472
acpi_pad_init(void)473 static int __init acpi_pad_init(void)
474 {
475 /* Xen ACPI PAD is used when running as Xen Dom0. */
476 if (xen_initial_domain())
477 return -ENODEV;
478
479 power_saving_mwait_init();
480 if (power_saving_mwait_eax == 0)
481 return -EINVAL;
482
483 return platform_driver_register(&acpi_pad_driver);
484 }
485
acpi_pad_exit(void)486 static void __exit acpi_pad_exit(void)
487 {
488 platform_driver_unregister(&acpi_pad_driver);
489 }
490
491 module_init(acpi_pad_init);
492 module_exit(acpi_pad_exit);
493 MODULE_AUTHOR("Shaohua Li<shaohua.li@intel.com>");
494 MODULE_DESCRIPTION("ACPI Processor Aggregator Driver");
495 MODULE_LICENSE("GPL");
496