xref: /linux/drivers/idle/intel_idle.c (revision e978aa7d7d57d04eb5f88a7507c4fb98577def77)
1 /*
2  * intel_idle.c - native hardware idle loop for modern Intel processors
3  *
4  * Copyright (c) 2010, Intel Corporation.
5  * Len Brown <len.brown@intel.com>
6  *
7  * This program is free software; you can redistribute it and/or modify it
8  * under the terms and conditions of the GNU General Public License,
9  * version 2, as published by the Free Software Foundation.
10  *
11  * This program is distributed in the hope it will be useful, but WITHOUT
12  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
13  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
14  * more details.
15  *
16  * You should have received a copy of the GNU General Public License along with
17  * this program; if not, write to the Free Software Foundation, Inc.,
18  * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
19  */
20 
21 /*
22  * intel_idle is a cpuidle driver that loads on specific Intel processors
23  * in lieu of the legacy ACPI processor_idle driver.  The intent is to
24  * make Linux more efficient on these processors, as intel_idle knows
25  * more than ACPI, as well as make Linux more immune to ACPI BIOS bugs.
26  */
27 
28 /*
29  * Design Assumptions
30  *
31  * All CPUs have same idle states as boot CPU
32  *
33  * Chipset BM_STS (bus master status) bit is a NOP
34  *	for preventing entry into deep C-stats
35  */
36 
37 /*
38  * Known limitations
39  *
40  * The driver currently initializes for_each_online_cpu() upon modprobe.
41  * It it unaware of subsequent processors hot-added to the system.
42  * This means that if you boot with maxcpus=n and later online
43  * processors above n, those processors will use C1 only.
44  *
45  * ACPI has a .suspend hack to turn off deep c-statees during suspend
46  * to avoid complications with the lapic timer workaround.
47  * Have not seen issues with suspend, but may need same workaround here.
48  *
49  * There is currently no kernel-based automatic probing/loading mechanism
50  * if the driver is built as a module.
51  */
52 
53 /* un-comment DEBUG to enable pr_debug() statements */
54 #define DEBUG
55 
56 #include <linux/kernel.h>
57 #include <linux/cpuidle.h>
58 #include <linux/clockchips.h>
59 #include <linux/hrtimer.h>	/* ktime_get_real() */
60 #include <trace/events/power.h>
61 #include <linux/sched.h>
62 #include <linux/notifier.h>
63 #include <linux/cpu.h>
64 #include <asm/mwait.h>
65 #include <asm/msr.h>
66 
67 #define INTEL_IDLE_VERSION "0.4"
68 #define PREFIX "intel_idle: "
69 
70 static struct cpuidle_driver intel_idle_driver = {
71 	.name = "intel_idle",
72 	.owner = THIS_MODULE,
73 };
74 /* intel_idle.max_cstate=0 disables driver */
75 static int max_cstate = MWAIT_MAX_NUM_CSTATES - 1;
76 
77 static unsigned int mwait_substates;
78 
79 #define LAPIC_TIMER_ALWAYS_RELIABLE 0xFFFFFFFF
80 /* Reliable LAPIC Timer States, bit 1 for C1 etc.  */
81 static unsigned int lapic_timer_reliable_states = (1 << 1);	 /* Default to only C1 */
82 
83 static struct cpuidle_device __percpu *intel_idle_cpuidle_devices;
84 static int intel_idle(struct cpuidle_device *dev, int index);
85 
86 static struct cpuidle_state *cpuidle_state_table;
87 
88 /*
89  * Hardware C-state auto-demotion may not always be optimal.
90  * Indicate which enable bits to clear here.
91  */
92 static unsigned long long auto_demotion_disable_flags;
93 
94 /*
95  * Set this flag for states where the HW flushes the TLB for us
96  * and so we don't need cross-calls to keep it consistent.
97  * If this flag is set, SW flushes the TLB, so even if the
98  * HW doesn't do the flushing, this flag is safe to use.
99  */
100 #define CPUIDLE_FLAG_TLB_FLUSHED	0x10000
101 
102 /*
103  * States are indexed by the cstate number,
104  * which is also the index into the MWAIT hint array.
105  * Thus C0 is a dummy.
106  */
107 static struct cpuidle_state nehalem_cstates[MWAIT_MAX_NUM_CSTATES] = {
108 	{ /* MWAIT C0 */ },
109 	{ /* MWAIT C1 */
110 		.name = "C1-NHM",
111 		.desc = "MWAIT 0x00",
112 		.driver_data = (void *) 0x00,
113 		.flags = CPUIDLE_FLAG_TIME_VALID,
114 		.exit_latency = 3,
115 		.target_residency = 6,
116 		.enter = &intel_idle },
117 	{ /* MWAIT C2 */
118 		.name = "C3-NHM",
119 		.desc = "MWAIT 0x10",
120 		.driver_data = (void *) 0x10,
121 		.flags = CPUIDLE_FLAG_TIME_VALID | CPUIDLE_FLAG_TLB_FLUSHED,
122 		.exit_latency = 20,
123 		.target_residency = 80,
124 		.enter = &intel_idle },
125 	{ /* MWAIT C3 */
126 		.name = "C6-NHM",
127 		.desc = "MWAIT 0x20",
128 		.driver_data = (void *) 0x20,
129 		.flags = CPUIDLE_FLAG_TIME_VALID | CPUIDLE_FLAG_TLB_FLUSHED,
130 		.exit_latency = 200,
131 		.target_residency = 800,
132 		.enter = &intel_idle },
133 };
134 
135 static struct cpuidle_state snb_cstates[MWAIT_MAX_NUM_CSTATES] = {
136 	{ /* MWAIT C0 */ },
137 	{ /* MWAIT C1 */
138 		.name = "C1-SNB",
139 		.desc = "MWAIT 0x00",
140 		.driver_data = (void *) 0x00,
141 		.flags = CPUIDLE_FLAG_TIME_VALID,
142 		.exit_latency = 1,
143 		.target_residency = 1,
144 		.enter = &intel_idle },
145 	{ /* MWAIT C2 */
146 		.name = "C3-SNB",
147 		.desc = "MWAIT 0x10",
148 		.driver_data = (void *) 0x10,
149 		.flags = CPUIDLE_FLAG_TIME_VALID | CPUIDLE_FLAG_TLB_FLUSHED,
150 		.exit_latency = 80,
151 		.target_residency = 211,
152 		.enter = &intel_idle },
153 	{ /* MWAIT C3 */
154 		.name = "C6-SNB",
155 		.desc = "MWAIT 0x20",
156 		.driver_data = (void *) 0x20,
157 		.flags = CPUIDLE_FLAG_TIME_VALID | CPUIDLE_FLAG_TLB_FLUSHED,
158 		.exit_latency = 104,
159 		.target_residency = 345,
160 		.enter = &intel_idle },
161 	{ /* MWAIT C4 */
162 		.name = "C7-SNB",
163 		.desc = "MWAIT 0x30",
164 		.driver_data = (void *) 0x30,
165 		.flags = CPUIDLE_FLAG_TIME_VALID | CPUIDLE_FLAG_TLB_FLUSHED,
166 		.exit_latency = 109,
167 		.target_residency = 345,
168 		.enter = &intel_idle },
169 };
170 
171 static struct cpuidle_state atom_cstates[MWAIT_MAX_NUM_CSTATES] = {
172 	{ /* MWAIT C0 */ },
173 	{ /* MWAIT C1 */
174 		.name = "C1-ATM",
175 		.desc = "MWAIT 0x00",
176 		.driver_data = (void *) 0x00,
177 		.flags = CPUIDLE_FLAG_TIME_VALID,
178 		.exit_latency = 1,
179 		.target_residency = 4,
180 		.enter = &intel_idle },
181 	{ /* MWAIT C2 */
182 		.name = "C2-ATM",
183 		.desc = "MWAIT 0x10",
184 		.driver_data = (void *) 0x10,
185 		.flags = CPUIDLE_FLAG_TIME_VALID,
186 		.exit_latency = 20,
187 		.target_residency = 80,
188 		.enter = &intel_idle },
189 	{ /* MWAIT C3 */ },
190 	{ /* MWAIT C4 */
191 		.name = "C4-ATM",
192 		.desc = "MWAIT 0x30",
193 		.driver_data = (void *) 0x30,
194 		.flags = CPUIDLE_FLAG_TIME_VALID | CPUIDLE_FLAG_TLB_FLUSHED,
195 		.exit_latency = 100,
196 		.target_residency = 400,
197 		.enter = &intel_idle },
198 	{ /* MWAIT C5 */ },
199 	{ /* MWAIT C6 */
200 		.name = "C6-ATM",
201 		.desc = "MWAIT 0x52",
202 		.driver_data = (void *) 0x52,
203 		.flags = CPUIDLE_FLAG_TIME_VALID | CPUIDLE_FLAG_TLB_FLUSHED,
204 		.exit_latency = 140,
205 		.target_residency = 560,
206 		.enter = &intel_idle },
207 };
208 
209 /**
210  * intel_idle
211  * @dev: cpuidle_device
212  * @index: index of cpuidle state
213  *
214  */
215 static int intel_idle(struct cpuidle_device *dev, int index)
216 {
217 	unsigned long ecx = 1; /* break on interrupt flag */
218 	struct cpuidle_state *state = &dev->states[index];
219 	unsigned long eax = (unsigned long)cpuidle_get_statedata(state);
220 	unsigned int cstate;
221 	ktime_t kt_before, kt_after;
222 	s64 usec_delta;
223 	int cpu = smp_processor_id();
224 
225 	cstate = (((eax) >> MWAIT_SUBSTATE_SIZE) & MWAIT_CSTATE_MASK) + 1;
226 
227 	local_irq_disable();
228 
229 	/*
230 	 * leave_mm() to avoid costly and often unnecessary wakeups
231 	 * for flushing the user TLB's associated with the active mm.
232 	 */
233 	if (state->flags & CPUIDLE_FLAG_TLB_FLUSHED)
234 		leave_mm(cpu);
235 
236 	if (!(lapic_timer_reliable_states & (1 << (cstate))))
237 		clockevents_notify(CLOCK_EVT_NOTIFY_BROADCAST_ENTER, &cpu);
238 
239 	kt_before = ktime_get_real();
240 
241 	stop_critical_timings();
242 	if (!need_resched()) {
243 
244 		__monitor((void *)&current_thread_info()->flags, 0, 0);
245 		smp_mb();
246 		if (!need_resched())
247 			__mwait(eax, ecx);
248 	}
249 
250 	start_critical_timings();
251 
252 	kt_after = ktime_get_real();
253 	usec_delta = ktime_to_us(ktime_sub(kt_after, kt_before));
254 
255 	local_irq_enable();
256 
257 	if (!(lapic_timer_reliable_states & (1 << (cstate))))
258 		clockevents_notify(CLOCK_EVT_NOTIFY_BROADCAST_EXIT, &cpu);
259 
260 	/* Update cpuidle counters */
261 	dev->last_residency = (int)usec_delta;
262 
263 	return index;
264 }
265 
266 static void __setup_broadcast_timer(void *arg)
267 {
268 	unsigned long reason = (unsigned long)arg;
269 	int cpu = smp_processor_id();
270 
271 	reason = reason ?
272 		CLOCK_EVT_NOTIFY_BROADCAST_ON : CLOCK_EVT_NOTIFY_BROADCAST_OFF;
273 
274 	clockevents_notify(reason, &cpu);
275 }
276 
277 static int setup_broadcast_cpuhp_notify(struct notifier_block *n,
278 		unsigned long action, void *hcpu)
279 {
280 	int hotcpu = (unsigned long)hcpu;
281 
282 	switch (action & 0xf) {
283 	case CPU_ONLINE:
284 		smp_call_function_single(hotcpu, __setup_broadcast_timer,
285 			(void *)true, 1);
286 		break;
287 	}
288 	return NOTIFY_OK;
289 }
290 
291 static struct notifier_block setup_broadcast_notifier = {
292 	.notifier_call = setup_broadcast_cpuhp_notify,
293 };
294 
295 static void auto_demotion_disable(void *dummy)
296 {
297 	unsigned long long msr_bits;
298 
299 	rdmsrl(MSR_NHM_SNB_PKG_CST_CFG_CTL, msr_bits);
300 	msr_bits &= ~auto_demotion_disable_flags;
301 	wrmsrl(MSR_NHM_SNB_PKG_CST_CFG_CTL, msr_bits);
302 }
303 
304 /*
305  * intel_idle_probe()
306  */
307 static int intel_idle_probe(void)
308 {
309 	unsigned int eax, ebx, ecx;
310 
311 	if (max_cstate == 0) {
312 		pr_debug(PREFIX "disabled\n");
313 		return -EPERM;
314 	}
315 
316 	if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL)
317 		return -ENODEV;
318 
319 	if (!boot_cpu_has(X86_FEATURE_MWAIT))
320 		return -ENODEV;
321 
322 	if (boot_cpu_data.cpuid_level < CPUID_MWAIT_LEAF)
323 		return -ENODEV;
324 
325 	cpuid(CPUID_MWAIT_LEAF, &eax, &ebx, &ecx, &mwait_substates);
326 
327 	if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED) ||
328 		!(ecx & CPUID5_ECX_INTERRUPT_BREAK))
329 			return -ENODEV;
330 
331 	pr_debug(PREFIX "MWAIT substates: 0x%x\n", mwait_substates);
332 
333 
334 	if (boot_cpu_data.x86 != 6)	/* family 6 */
335 		return -ENODEV;
336 
337 	switch (boot_cpu_data.x86_model) {
338 
339 	case 0x1A:	/* Core i7, Xeon 5500 series */
340 	case 0x1E:	/* Core i7 and i5 Processor - Lynnfield Jasper Forest */
341 	case 0x1F:	/* Core i7 and i5 Processor - Nehalem */
342 	case 0x2E:	/* Nehalem-EX Xeon */
343 	case 0x2F:	/* Westmere-EX Xeon */
344 	case 0x25:	/* Westmere */
345 	case 0x2C:	/* Westmere */
346 		cpuidle_state_table = nehalem_cstates;
347 		auto_demotion_disable_flags =
348 			(NHM_C1_AUTO_DEMOTE | NHM_C3_AUTO_DEMOTE);
349 		break;
350 
351 	case 0x1C:	/* 28 - Atom Processor */
352 		cpuidle_state_table = atom_cstates;
353 		break;
354 
355 	case 0x26:	/* 38 - Lincroft Atom Processor */
356 		cpuidle_state_table = atom_cstates;
357 		auto_demotion_disable_flags = ATM_LNC_C6_AUTO_DEMOTE;
358 		break;
359 
360 	case 0x2A:	/* SNB */
361 	case 0x2D:	/* SNB Xeon */
362 		cpuidle_state_table = snb_cstates;
363 		break;
364 
365 	default:
366 		pr_debug(PREFIX "does not run on family %d model %d\n",
367 			boot_cpu_data.x86, boot_cpu_data.x86_model);
368 		return -ENODEV;
369 	}
370 
371 	if (boot_cpu_has(X86_FEATURE_ARAT))	/* Always Reliable APIC Timer */
372 		lapic_timer_reliable_states = LAPIC_TIMER_ALWAYS_RELIABLE;
373 	else {
374 		smp_call_function(__setup_broadcast_timer, (void *)true, 1);
375 		register_cpu_notifier(&setup_broadcast_notifier);
376 	}
377 
378 	pr_debug(PREFIX "v" INTEL_IDLE_VERSION
379 		" model 0x%X\n", boot_cpu_data.x86_model);
380 
381 	pr_debug(PREFIX "lapic_timer_reliable_states 0x%x\n",
382 		lapic_timer_reliable_states);
383 	return 0;
384 }
385 
386 /*
387  * intel_idle_cpuidle_devices_uninit()
388  * unregister, free cpuidle_devices
389  */
390 static void intel_idle_cpuidle_devices_uninit(void)
391 {
392 	int i;
393 	struct cpuidle_device *dev;
394 
395 	for_each_online_cpu(i) {
396 		dev = per_cpu_ptr(intel_idle_cpuidle_devices, i);
397 		cpuidle_unregister_device(dev);
398 	}
399 
400 	free_percpu(intel_idle_cpuidle_devices);
401 	return;
402 }
403 /*
404  * intel_idle_cpuidle_devices_init()
405  * allocate, initialize, register cpuidle_devices
406  */
407 static int intel_idle_cpuidle_devices_init(void)
408 {
409 	int i, cstate;
410 	struct cpuidle_device *dev;
411 
412 	intel_idle_cpuidle_devices = alloc_percpu(struct cpuidle_device);
413 	if (intel_idle_cpuidle_devices == NULL)
414 		return -ENOMEM;
415 
416 	for_each_online_cpu(i) {
417 		dev = per_cpu_ptr(intel_idle_cpuidle_devices, i);
418 
419 		dev->state_count = 1;
420 
421 		for (cstate = 1; cstate < MWAIT_MAX_NUM_CSTATES; ++cstate) {
422 			int num_substates;
423 
424 			if (cstate > max_cstate) {
425 				printk(PREFIX "max_cstate %d reached\n",
426 					max_cstate);
427 				break;
428 			}
429 
430 			/* does the state exist in CPUID.MWAIT? */
431 			num_substates = (mwait_substates >> ((cstate) * 4))
432 						& MWAIT_SUBSTATE_MASK;
433 			if (num_substates == 0)
434 				continue;
435 			/* is the state not enabled? */
436 			if (cpuidle_state_table[cstate].enter == NULL) {
437 				/* does the driver not know about the state? */
438 				if (*cpuidle_state_table[cstate].name == '\0')
439 					pr_debug(PREFIX "unaware of model 0x%x"
440 						" MWAIT %d please"
441 						" contact lenb@kernel.org",
442 					boot_cpu_data.x86_model, cstate);
443 				continue;
444 			}
445 
446 			if ((cstate > 2) &&
447 				!boot_cpu_has(X86_FEATURE_NONSTOP_TSC))
448 				mark_tsc_unstable("TSC halts in idle"
449 					" states deeper than C2");
450 
451 			dev->states[dev->state_count] =	/* structure copy */
452 				cpuidle_state_table[cstate];
453 
454 			dev->state_count += 1;
455 		}
456 
457 		dev->cpu = i;
458 		if (cpuidle_register_device(dev)) {
459 			pr_debug(PREFIX "cpuidle_register_device %d failed!\n",
460 				 i);
461 			intel_idle_cpuidle_devices_uninit();
462 			return -EIO;
463 		}
464 	}
465 	if (auto_demotion_disable_flags)
466 		smp_call_function(auto_demotion_disable, NULL, 1);
467 
468 	return 0;
469 }
470 
471 
472 static int __init intel_idle_init(void)
473 {
474 	int retval;
475 
476 	/* Do not load intel_idle at all for now if idle= is passed */
477 	if (boot_option_idle_override != IDLE_NO_OVERRIDE)
478 		return -ENODEV;
479 
480 	retval = intel_idle_probe();
481 	if (retval)
482 		return retval;
483 
484 	retval = cpuidle_register_driver(&intel_idle_driver);
485 	if (retval) {
486 		printk(KERN_DEBUG PREFIX "intel_idle yielding to %s",
487 			cpuidle_get_driver()->name);
488 		return retval;
489 	}
490 
491 	retval = intel_idle_cpuidle_devices_init();
492 	if (retval) {
493 		cpuidle_unregister_driver(&intel_idle_driver);
494 		return retval;
495 	}
496 
497 	return 0;
498 }
499 
500 static void __exit intel_idle_exit(void)
501 {
502 	intel_idle_cpuidle_devices_uninit();
503 	cpuidle_unregister_driver(&intel_idle_driver);
504 
505 	if (lapic_timer_reliable_states != LAPIC_TIMER_ALWAYS_RELIABLE) {
506 		smp_call_function(__setup_broadcast_timer, (void *)false, 1);
507 		unregister_cpu_notifier(&setup_broadcast_notifier);
508 	}
509 
510 	return;
511 }
512 
513 module_init(intel_idle_init);
514 module_exit(intel_idle_exit);
515 
516 module_param(max_cstate, int, 0444);
517 
518 MODULE_AUTHOR("Len Brown <len.brown@intel.com>");
519 MODULE_DESCRIPTION("Cpuidle driver for Intel Hardware v" INTEL_IDLE_VERSION);
520 MODULE_LICENSE("GPL");
521