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