1
2
3
4
5
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
70
71
72 if (!boot_cpu_has(X86_FEATURE_NONSTOP_TSC))
73 tsc_detected_unstable = 1;
74 break;
75 default:
76
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 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
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;
136static unsigned int round_robin_time = 1;
137static int power_saving_thread(void *data)
138{
139 int do_sleep;
140 unsigned int tsk_index = (unsigned long)data;
141 u64 last_jiffies = 0;
142
143 sched_set_fifo_low(current);
144
145 while (!kthread_should_stop()) {
146 unsigned long expire_time;
147
148
149 expire_time = last_jiffies + round_robin_time * HZ;
150 if (time_before(expire_time, jiffies)) {
151 last_jiffies = jiffies;
152 round_robin_cpu(tsk_index);
153 }
154
155 do_sleep = 0;
156
157 expire_time = jiffies + HZ * (100 - idle_pct) / 100;
158
159 while (!need_resched()) {
160 if (tsc_detected_unstable && !tsc_marked_unstable) {
161
162 mark_tsc_unstable("TSC halts in idle");
163 tsc_marked_unstable = 1;
164 }
165 local_irq_disable();
166 tick_broadcast_enable();
167 tick_broadcast_enter();
168 stop_critical_timings();
169
170 mwait_idle_with_hints(power_saving_mwait_eax, 1);
171
172 start_critical_timings();
173 tick_broadcast_exit();
174 local_irq_enable();
175
176 if (time_before(expire_time, jiffies)) {
177 do_sleep = 1;
178 break;
179 }
180 }
181
182
183
184
185
186
187
188
189
190
191 if (unlikely(do_sleep))
192 schedule_timeout_killable(HZ * idle_pct / 100);
193
194
195
196
197
198 if (unlikely(need_resched()))
199 schedule();
200 }
201
202 exit_round_robin(tsk_index);
203 return 0;
204}
205
206static struct task_struct *ps_tsks[NR_CPUS];
207static unsigned int ps_tsk_num;
208static int create_power_saving_task(void)
209{
210 int rc;
211
212 ps_tsks[ps_tsk_num] = kthread_run(power_saving_thread,
213 (void *)(unsigned long)ps_tsk_num,
214 "acpi_pad/%d", ps_tsk_num);
215
216 if (IS_ERR(ps_tsks[ps_tsk_num])) {
217 rc = PTR_ERR(ps_tsks[ps_tsk_num]);
218 ps_tsks[ps_tsk_num] = NULL;
219 } else {
220 rc = 0;
221 ps_tsk_num++;
222 }
223
224 return rc;
225}
226
227static void destroy_power_saving_task(void)
228{
229 if (ps_tsk_num > 0) {
230 ps_tsk_num--;
231 kthread_stop(ps_tsks[ps_tsk_num]);
232 ps_tsks[ps_tsk_num] = NULL;
233 }
234}
235
236static void set_power_saving_task_num(unsigned int num)
237{
238 if (num > ps_tsk_num) {
239 while (ps_tsk_num < num) {
240 if (create_power_saving_task())
241 return;
242 }
243 } else if (num < ps_tsk_num) {
244 while (ps_tsk_num > num)
245 destroy_power_saving_task();
246 }
247}
248
249static void acpi_pad_idle_cpus(unsigned int num_cpus)
250{
251 get_online_cpus();
252
253 num_cpus = min_t(unsigned int, num_cpus, num_online_cpus());
254 set_power_saving_task_num(num_cpus);
255
256 put_online_cpus();
257}
258
259static uint32_t acpi_pad_idle_cpus_num(void)
260{
261 return ps_tsk_num;
262}
263
264static ssize_t rrtime_store(struct device *dev,
265 struct device_attribute *attr, const char *buf, size_t count)
266{
267 unsigned long num;
268 if (kstrtoul(buf, 0, &num))
269 return -EINVAL;
270 if (num < 1 || num >= 100)
271 return -EINVAL;
272 mutex_lock(&isolated_cpus_lock);
273 round_robin_time = num;
274 mutex_unlock(&isolated_cpus_lock);
275 return count;
276}
277
278static ssize_t rrtime_show(struct device *dev,
279 struct device_attribute *attr, char *buf)
280{
281 return scnprintf(buf, PAGE_SIZE, "%d\n", round_robin_time);
282}
283static DEVICE_ATTR_RW(rrtime);
284
285static ssize_t idlepct_store(struct device *dev,
286 struct device_attribute *attr, const char *buf, size_t count)
287{
288 unsigned long num;
289 if (kstrtoul(buf, 0, &num))
290 return -EINVAL;
291 if (num < 1 || num >= 100)
292 return -EINVAL;
293 mutex_lock(&isolated_cpus_lock);
294 idle_pct = num;
295 mutex_unlock(&isolated_cpus_lock);
296 return count;
297}
298
299static ssize_t idlepct_show(struct device *dev,
300 struct device_attribute *attr, char *buf)
301{
302 return scnprintf(buf, PAGE_SIZE, "%d\n", idle_pct);
303}
304static DEVICE_ATTR_RW(idlepct);
305
306static ssize_t idlecpus_store(struct device *dev,
307 struct device_attribute *attr, const char *buf, size_t count)
308{
309 unsigned long num;
310 if (kstrtoul(buf, 0, &num))
311 return -EINVAL;
312 mutex_lock(&isolated_cpus_lock);
313 acpi_pad_idle_cpus(num);
314 mutex_unlock(&isolated_cpus_lock);
315 return count;
316}
317
318static ssize_t idlecpus_show(struct device *dev,
319 struct device_attribute *attr, char *buf)
320{
321 return cpumap_print_to_pagebuf(false, buf,
322 to_cpumask(pad_busy_cpus_bits));
323}
324
325static DEVICE_ATTR_RW(idlecpus);
326
327static int acpi_pad_add_sysfs(struct acpi_device *device)
328{
329 int result;
330
331 result = device_create_file(&device->dev, &dev_attr_idlecpus);
332 if (result)
333 return -ENODEV;
334 result = device_create_file(&device->dev, &dev_attr_idlepct);
335 if (result) {
336 device_remove_file(&device->dev, &dev_attr_idlecpus);
337 return -ENODEV;
338 }
339 result = device_create_file(&device->dev, &dev_attr_rrtime);
340 if (result) {
341 device_remove_file(&device->dev, &dev_attr_idlecpus);
342 device_remove_file(&device->dev, &dev_attr_idlepct);
343 return -ENODEV;
344 }
345 return 0;
346}
347
348static void acpi_pad_remove_sysfs(struct acpi_device *device)
349{
350 device_remove_file(&device->dev, &dev_attr_idlecpus);
351 device_remove_file(&device->dev, &dev_attr_idlepct);
352 device_remove_file(&device->dev, &dev_attr_rrtime);
353}
354
355
356
357
358
359static 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)
376
377 num = package->package.elements[1].integer.value;
378
379 kfree(buffer.pointer);
380 return num;
381}
382
383static 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
392 mutex_lock(&isolated_cpus_lock);
393 num_cpus = acpi_pad_pur(handle);
394 if (num_cpus < 0) {
395 mutex_unlock(&isolated_cpus_lock);
396 return;
397 }
398 acpi_pad_idle_cpus(num_cpus);
399 idle_cpus = acpi_pad_idle_cpus_num();
400 acpi_evaluate_ost(handle, ACPI_PROCESSOR_AGGREGATOR_NOTIFY, 0, ¶m);
401 mutex_unlock(&isolated_cpus_lock);
402}
403
404static void acpi_pad_notify(acpi_handle handle, u32 event,
405 void *data)
406{
407 struct acpi_device *device = data;
408
409 switch (event) {
410 case ACPI_PROCESSOR_AGGREGATOR_NOTIFY:
411 acpi_pad_handle_notify(handle);
412 acpi_bus_generate_netlink_event(device->pnp.device_class,
413 dev_name(&device->dev), event, 0);
414 break;
415 default:
416 pr_warn("Unsupported event [0x%x]\n", event);
417 break;
418 }
419}
420
421static int acpi_pad_add(struct acpi_device *device)
422{
423 acpi_status status;
424
425 strcpy(acpi_device_name(device), ACPI_PROCESSOR_AGGREGATOR_DEVICE_NAME);
426 strcpy(acpi_device_class(device), ACPI_PROCESSOR_AGGREGATOR_CLASS);
427
428 if (acpi_pad_add_sysfs(device))
429 return -ENODEV;
430
431 status = acpi_install_notify_handler(device->handle,
432 ACPI_DEVICE_NOTIFY, acpi_pad_notify, device);
433 if (ACPI_FAILURE(status)) {
434 acpi_pad_remove_sysfs(device);
435 return -ENODEV;
436 }
437
438 return 0;
439}
440
441static int acpi_pad_remove(struct acpi_device *device)
442{
443 mutex_lock(&isolated_cpus_lock);
444 acpi_pad_idle_cpus(0);
445 mutex_unlock(&isolated_cpus_lock);
446
447 acpi_remove_notify_handler(device->handle,
448 ACPI_DEVICE_NOTIFY, acpi_pad_notify);
449 acpi_pad_remove_sysfs(device);
450 return 0;
451}
452
453static const struct acpi_device_id pad_device_ids[] = {
454 {"ACPI000C", 0},
455 {"", 0},
456};
457MODULE_DEVICE_TABLE(acpi, pad_device_ids);
458
459static struct acpi_driver acpi_pad_driver = {
460 .name = "processor_aggregator",
461 .class = ACPI_PROCESSOR_AGGREGATOR_CLASS,
462 .ids = pad_device_ids,
463 .ops = {
464 .add = acpi_pad_add,
465 .remove = acpi_pad_remove,
466 },
467};
468
469static int __init acpi_pad_init(void)
470{
471
472 if (xen_initial_domain())
473 return -ENODEV;
474
475 power_saving_mwait_init();
476 if (power_saving_mwait_eax == 0)
477 return -EINVAL;
478
479 return acpi_bus_register_driver(&acpi_pad_driver);
480}
481
482static void __exit acpi_pad_exit(void)
483{
484 acpi_bus_unregister_driver(&acpi_pad_driver);
485}
486
487module_init(acpi_pad_init);
488module_exit(acpi_pad_exit);
489MODULE_AUTHOR("Shaohua Li<shaohua.li@intel.com>");
490MODULE_DESCRIPTION("ACPI Processor Aggregator Driver");
491MODULE_LICENSE("GPL");
492
