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 <uapi/linux/sched/types.h>
24#include <linux/freezer.h>
25#include <linux/cpu.h>
26#include <linux/tick.h>
27#include <linux/slab.h>
28#include <linux/acpi.h>
29#include <asm/mwait.h>
30#include <xen/xen.h>
31
32#define ACPI_PROCESSOR_AGGREGATOR_CLASS "acpi_pad"
33#define ACPI_PROCESSOR_AGGREGATOR_DEVICE_NAME "Processor Aggregator"
34#define ACPI_PROCESSOR_AGGREGATOR_NOTIFY 0x80
35static DEFINE_MUTEX(isolated_cpus_lock);
36static DEFINE_MUTEX(round_robin_lock);
37
38static unsigned long power_saving_mwait_eax;
39
40static unsigned char tsc_detected_unstable;
41static unsigned char tsc_marked_unstable;
42
43static void power_saving_mwait_init(void)
44{
45 unsigned int eax, ebx, ecx, edx;
46 unsigned int highest_cstate = 0;
47 unsigned int highest_subcstate = 0;
48 int i;
49
50 if (!boot_cpu_has(X86_FEATURE_MWAIT))
51 return;
52 if (boot_cpu_data.cpuid_level < CPUID_MWAIT_LEAF)
53 return;
54
55 cpuid(CPUID_MWAIT_LEAF, &eax, &ebx, &ecx, &edx);
56
57 if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED) ||
58 !(ecx & CPUID5_ECX_INTERRUPT_BREAK))
59 return;
60
61 edx >>= MWAIT_SUBSTATE_SIZE;
62 for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) {
63 if (edx & MWAIT_SUBSTATE_MASK) {
64 highest_cstate = i;
65 highest_subcstate = edx & MWAIT_SUBSTATE_MASK;
66 }
67 }
68 power_saving_mwait_eax = (highest_cstate << MWAIT_SUBSTATE_SIZE) |
69 (highest_subcstate - 1);
70
71#if defined(CONFIG_X86)
72 switch (boot_cpu_data.x86_vendor) {
73 case X86_VENDOR_HYGON:
74 case X86_VENDOR_AMD:
75 case X86_VENDOR_INTEL:
76 /*
77 * AMD Fam10h TSC will tick in all
78 * C/P/S0/S1 states when this bit is set.
79 */
80 if (!boot_cpu_has(X86_FEATURE_NONSTOP_TSC))
81 tsc_detected_unstable = 1;
82 break;
83 default:
84 /* TSC could halt in idle */
85 tsc_detected_unstable = 1;
86 }
87#endif
88}
89
90static unsigned long cpu_weight[NR_CPUS];
91static int tsk_in_cpu[NR_CPUS] = {[0 ... NR_CPUS-1] = -1};
92static DECLARE_BITMAP(pad_busy_cpus_bits, NR_CPUS);
93static void round_robin_cpu(unsigned int tsk_index)
94{
95 struct cpumask *pad_busy_cpus = to_cpumask(pad_busy_cpus_bits);
96 cpumask_var_t tmp;
97 int cpu;
98 unsigned long min_weight = -1;
99 unsigned long uninitialized_var(preferred_cpu);
100
101 if (!alloc_cpumask_var(&tmp, GFP_KERNEL))
102 return;
103
104 mutex_lock(&round_robin_lock);
105 cpumask_clear(tmp);
106 for_each_cpu(cpu, pad_busy_cpus)
107 cpumask_or(tmp, tmp, topology_sibling_cpumask(cpu));
108 cpumask_andnot(tmp, cpu_online_mask, tmp);
109 /* avoid HT sibilings if possible */
110 if (cpumask_empty(tmp))
111 cpumask_andnot(tmp, cpu_online_mask, pad_busy_cpus);
112 if (cpumask_empty(tmp)) {
113 mutex_unlock(&round_robin_lock);
114 free_cpumask_var(tmp);
115 return;
116 }
117 for_each_cpu(cpu, tmp) {
118 if (cpu_weight[cpu] < min_weight) {
119 min_weight = cpu_weight[cpu];
120 preferred_cpu = cpu;
121 }
122 }
123
124 if (tsk_in_cpu[tsk_index] != -1)
125 cpumask_clear_cpu(tsk_in_cpu[tsk_index], pad_busy_cpus);
126 tsk_in_cpu[tsk_index] = preferred_cpu;
127 cpumask_set_cpu(preferred_cpu, pad_busy_cpus);
128 cpu_weight[preferred_cpu]++;
129 mutex_unlock(&round_robin_lock);
130
131 set_cpus_allowed_ptr(current, cpumask_of(preferred_cpu));
132
133 free_cpumask_var(tmp);
134}
135
136static void exit_round_robin(unsigned int tsk_index)
137{
138 struct cpumask *pad_busy_cpus = to_cpumask(pad_busy_cpus_bits);
139 cpumask_clear_cpu(tsk_in_cpu[tsk_index], pad_busy_cpus);
140 tsk_in_cpu[tsk_index] = -1;
141}
142
143static unsigned int idle_pct = 5; /* percentage */
144static unsigned int round_robin_time = 1; /* second */
145static int power_saving_thread(void *data)
146{
147 struct sched_param param = {.sched_priority = 1};
148 int do_sleep;
149 unsigned int tsk_index = (unsigned long)data;
150 u64 last_jiffies = 0;
151
152 sched_setscheduler(current, SCHED_RR, &param);
153
154 while (!kthread_should_stop()) {
155 unsigned long expire_time;
156
157 /* round robin to cpus */
158 expire_time = last_jiffies + round_robin_time * HZ;
159 if (time_before(expire_time, jiffies)) {
160 last_jiffies = jiffies;
161 round_robin_cpu(tsk_index);
162 }
163
164 do_sleep = 0;
165
166 expire_time = jiffies + HZ * (100 - idle_pct) / 100;
167
168 while (!need_resched()) {
169 if (tsc_detected_unstable && !tsc_marked_unstable) {
170 /* TSC could halt in idle, so notify users */
171 mark_tsc_unstable("TSC halts in idle");
172 tsc_marked_unstable = 1;
173 }
174 local_irq_disable();
175 tick_broadcast_enable();
176 tick_broadcast_enter();
177 stop_critical_timings();
178
179 mwait_idle_with_hints(power_saving_mwait_eax, 1);
180
181 start_critical_timings();
182 tick_broadcast_exit();
183 local_irq_enable();
184
185 if (time_before(expire_time, jiffies)) {
186 do_sleep = 1;
187 break;
188 }
189 }
190
191 /*
192 * current sched_rt has threshold for rt task running time.
193 * When a rt task uses 95% CPU time, the rt thread will be
194 * scheduled out for 5% CPU time to not starve other tasks. But
195 * the mechanism only works when all CPUs have RT task running,
196 * as if one CPU hasn't RT task, RT task from other CPUs will
197 * borrow CPU time from this CPU and cause RT task use > 95%
198 * CPU time. To make 'avoid starvation' work, takes a nap here.
199 */
200 if (unlikely(do_sleep))
201 schedule_timeout_killable(HZ * idle_pct / 100);
202
203 /* If an external event has set the need_resched flag, then
204 * we need to deal with it, or this loop will continue to
205 * spin without calling __mwait().
206 */
207 if (unlikely(need_resched()))
208 schedule();
209 }
210
211 exit_round_robin(tsk_index);
212 return 0;
213}
214
215static struct task_struct *ps_tsks[NR_CPUS];
216static unsigned int ps_tsk_num;
217static int create_power_saving_task(void)
218{
219 int rc;
220
221 ps_tsks[ps_tsk_num] = kthread_run(power_saving_thread,
222 (void *)(unsigned long)ps_tsk_num,
223 "acpi_pad/%d", ps_tsk_num);
224
225 if (IS_ERR(ps_tsks[ps_tsk_num])) {
226 rc = PTR_ERR(ps_tsks[ps_tsk_num]);
227 ps_tsks[ps_tsk_num] = NULL;
228 } else {
229 rc = 0;
230 ps_tsk_num++;
231 }
232
233 return rc;
234}
235
236static void destroy_power_saving_task(void)
237{
238 if (ps_tsk_num > 0) {
239 ps_tsk_num--;
240 kthread_stop(ps_tsks[ps_tsk_num]);
241 ps_tsks[ps_tsk_num] = NULL;
242 }
243}
244
245static void set_power_saving_task_num(unsigned int num)
246{
247 if (num > ps_tsk_num) {
248 while (ps_tsk_num < num) {
249 if (create_power_saving_task())
250 return;
251 }
252 } else if (num < ps_tsk_num) {
253 while (ps_tsk_num > num)
254 destroy_power_saving_task();
255 }
256}
257
258static void acpi_pad_idle_cpus(unsigned int num_cpus)
259{
260 get_online_cpus();
261
262 num_cpus = min_t(unsigned int, num_cpus, num_online_cpus());
263 set_power_saving_task_num(num_cpus);
264
265 put_online_cpus();
266}
267
268static uint32_t acpi_pad_idle_cpus_num(void)
269{
270 return ps_tsk_num;
271}
272
273static ssize_t acpi_pad_rrtime_store(struct device *dev,
274 struct device_attribute *attr, const char *buf, size_t count)
275{
276 unsigned long num;
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 acpi_pad_rrtime_show(struct device *dev,
288 struct device_attribute *attr, char *buf)
289{
290 return scnprintf(buf, PAGE_SIZE, "%d\n", round_robin_time);
291}
292static DEVICE_ATTR(rrtime, S_IRUGO|S_IWUSR,
293 acpi_pad_rrtime_show,
294 acpi_pad_rrtime_store);
295
296static ssize_t acpi_pad_idlepct_store(struct device *dev,
297 struct device_attribute *attr, const char *buf, size_t count)
298{
299 unsigned long num;
300 if (kstrtoul(buf, 0, &num))
301 return -EINVAL;
302 if (num < 1 || num >= 100)
303 return -EINVAL;
304 mutex_lock(&isolated_cpus_lock);
305 idle_pct = num;
306 mutex_unlock(&isolated_cpus_lock);
307 return count;
308}
309
310static ssize_t acpi_pad_idlepct_show(struct device *dev,
311 struct device_attribute *attr, char *buf)
312{
313 return scnprintf(buf, PAGE_SIZE, "%d\n", idle_pct);
314}
315static DEVICE_ATTR(idlepct, S_IRUGO|S_IWUSR,
316 acpi_pad_idlepct_show,
317 acpi_pad_idlepct_store);
318
319static ssize_t acpi_pad_idlecpus_store(struct device *dev,
320 struct device_attribute *attr, const char *buf, size_t count)
321{
322 unsigned long num;
323 if (kstrtoul(buf, 0, &num))
324 return -EINVAL;
325 mutex_lock(&isolated_cpus_lock);
326 acpi_pad_idle_cpus(num);
327 mutex_unlock(&isolated_cpus_lock);
328 return count;
329}
330
331static ssize_t acpi_pad_idlecpus_show(struct device *dev,
332 struct device_attribute *attr, char *buf)
333{
334 return cpumap_print_to_pagebuf(false, buf,
335 to_cpumask(pad_busy_cpus_bits));
336}
337
338static DEVICE_ATTR(idlecpus, S_IRUGO|S_IWUSR,
339 acpi_pad_idlecpus_show,
340 acpi_pad_idlecpus_store);
341
342static int acpi_pad_add_sysfs(struct acpi_device *device)
343{
344 int result;
345
346 result = device_create_file(&device->dev, &dev_attr_idlecpus);
347 if (result)
348 return -ENODEV;
349 result = device_create_file(&device->dev, &dev_attr_idlepct);
350 if (result) {
351 device_remove_file(&device->dev, &dev_attr_idlecpus);
352 return -ENODEV;
353 }
354 result = device_create_file(&device->dev, &dev_attr_rrtime);
355 if (result) {
356 device_remove_file(&device->dev, &dev_attr_idlecpus);
357 device_remove_file(&device->dev, &dev_attr_idlepct);
358 return -ENODEV;
359 }
360 return 0;
361}
362
363static void acpi_pad_remove_sysfs(struct acpi_device *device)
364{
365 device_remove_file(&device->dev, &dev_attr_idlecpus);
366 device_remove_file(&device->dev, &dev_attr_idlepct);
367 device_remove_file(&device->dev, &dev_attr_rrtime);
368}
369
370/*
371 * Query firmware how many CPUs should be idle
372 * return -1 on failure
373 */
374static int acpi_pad_pur(acpi_handle handle)
375{
376 struct acpi_buffer buffer = {ACPI_ALLOCATE_BUFFER, NULL};
377 union acpi_object *package;
378 int num = -1;
379
380 if (ACPI_FAILURE(acpi_evaluate_object(handle, "_PUR", NULL, &buffer)))
381 return num;
382
383 if (!buffer.length || !buffer.pointer)
384 return num;
385
386 package = buffer.pointer;
387
388 if (package->type == ACPI_TYPE_PACKAGE &&
389 package->package.count == 2 &&
390 package->package.elements[0].integer.value == 1) /* rev 1 */
391
392 num = package->package.elements[1].integer.value;
393
394 kfree(buffer.pointer);
395 return num;
396}
397
398static void acpi_pad_handle_notify(acpi_handle handle)
399{
400 int num_cpus;
401 uint32_t idle_cpus;
402 struct acpi_buffer param = {
403 .length = 4,
404 .pointer = (void *)&idle_cpus,
405 };
406
407 mutex_lock(&isolated_cpus_lock);
408 num_cpus = acpi_pad_pur(handle);
409 if (num_cpus < 0) {
410 mutex_unlock(&isolated_cpus_lock);
411 return;
412 }
413 acpi_pad_idle_cpus(num_cpus);
414 idle_cpus = acpi_pad_idle_cpus_num();
415 acpi_evaluate_ost(handle, ACPI_PROCESSOR_AGGREGATOR_NOTIFY, 0, &param);
416 mutex_unlock(&isolated_cpus_lock);
417}
418
419static void acpi_pad_notify(acpi_handle handle, u32 event,
420 void *data)
421{
422 struct acpi_device *device = data;
423
424 switch (event) {
425 case ACPI_PROCESSOR_AGGREGATOR_NOTIFY:
426 acpi_pad_handle_notify(handle);
427 acpi_bus_generate_netlink_event(device->pnp.device_class,
428 dev_name(&device->dev), event, 0);
429 break;
430 default:
431 pr_warn("Unsupported event [0x%x]\n", event);
432 break;
433 }
434}
435
436static int acpi_pad_add(struct acpi_device *device)
437{
438 acpi_status status;
439
440 strcpy(acpi_device_name(device), ACPI_PROCESSOR_AGGREGATOR_DEVICE_NAME);
441 strcpy(acpi_device_class(device), ACPI_PROCESSOR_AGGREGATOR_CLASS);
442
443 if (acpi_pad_add_sysfs(device))
444 return -ENODEV;
445
446 status = acpi_install_notify_handler(device->handle,
447 ACPI_DEVICE_NOTIFY, acpi_pad_notify, device);
448 if (ACPI_FAILURE(status)) {
449 acpi_pad_remove_sysfs(device);
450 return -ENODEV;
451 }
452
453 return 0;
454}
455
456static int acpi_pad_remove(struct acpi_device *device)
457{
458 mutex_lock(&isolated_cpus_lock);
459 acpi_pad_idle_cpus(0);
460 mutex_unlock(&isolated_cpus_lock);
461
462 acpi_remove_notify_handler(device->handle,
463 ACPI_DEVICE_NOTIFY, acpi_pad_notify);
464 acpi_pad_remove_sysfs(device);
465 return 0;
466}
467
468static const struct acpi_device_id pad_device_ids[] = {
469 {"ACPI000C", 0},
470 {"", 0},
471};
472MODULE_DEVICE_TABLE(acpi, pad_device_ids);
473
474static struct acpi_driver acpi_pad_driver = {
475 .name = "processor_aggregator",
476 .class = ACPI_PROCESSOR_AGGREGATOR_CLASS,
477 .ids = pad_device_ids,
478 .ops = {
479 .add = acpi_pad_add,
480 .remove = acpi_pad_remove,
481 },
482};
483
484static int __init acpi_pad_init(void)
485{
486 /* Xen ACPI PAD is used when running as Xen Dom0. */
487 if (xen_initial_domain())
488 return -ENODEV;
489
490 power_saving_mwait_init();
491 if (power_saving_mwait_eax == 0)
492 return -EINVAL;
493
494 return acpi_bus_register_driver(&acpi_pad_driver);
495}
496
497static void __exit acpi_pad_exit(void)
498{
499 acpi_bus_unregister_driver(&acpi_pad_driver);
500}
501
502module_init(acpi_pad_init);
503module_exit(acpi_pad_exit);
504MODULE_AUTHOR("Shaohua Li<shaohua.li@intel.com>");
505MODULE_DESCRIPTION("ACPI Processor Aggregator Driver");
506MODULE_LICENSE("GPL");
507