xref: /linux/drivers/idle/intel_idle.c (revision d229807f669ba3dea9f64467ee965051c4366aed)
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 <linux/module.h>
65 #include <asm/mwait.h>
66 #include <asm/msr.h>
67 
68 #define INTEL_IDLE_VERSION "0.4"
69 #define PREFIX "intel_idle: "
70 
71 static struct cpuidle_driver intel_idle_driver = {
72 	.name = "intel_idle",
73 	.owner = THIS_MODULE,
74 };
75 /* intel_idle.max_cstate=0 disables driver */
76 static int max_cstate = MWAIT_MAX_NUM_CSTATES - 1;
77 
78 static unsigned int mwait_substates;
79 
80 #define LAPIC_TIMER_ALWAYS_RELIABLE 0xFFFFFFFF
81 /* Reliable LAPIC Timer States, bit 1 for C1 etc.  */
82 static unsigned int lapic_timer_reliable_states = (1 << 1);	 /* Default to only C1 */
83 
84 static struct cpuidle_device __percpu *intel_idle_cpuidle_devices;
85 static int intel_idle(struct cpuidle_device *dev, struct cpuidle_state *state);
86 
87 static struct cpuidle_state *cpuidle_state_table;
88 
89 /*
90  * Hardware C-state auto-demotion may not always be optimal.
91  * Indicate which enable bits to clear here.
92  */
93 static unsigned long long auto_demotion_disable_flags;
94 
95 /*
96  * Set this flag for states where the HW flushes the TLB for us
97  * and so we don't need cross-calls to keep it consistent.
98  * If this flag is set, SW flushes the TLB, so even if the
99  * HW doesn't do the flushing, this flag is safe to use.
100  */
101 #define CPUIDLE_FLAG_TLB_FLUSHED	0x10000
102 
103 /*
104  * States are indexed by the cstate number,
105  * which is also the index into the MWAIT hint array.
106  * Thus C0 is a dummy.
107  */
108 static struct cpuidle_state nehalem_cstates[MWAIT_MAX_NUM_CSTATES] = {
109 	{ /* MWAIT C0 */ },
110 	{ /* MWAIT C1 */
111 		.name = "C1-NHM",
112 		.desc = "MWAIT 0x00",
113 		.driver_data = (void *) 0x00,
114 		.flags = CPUIDLE_FLAG_TIME_VALID,
115 		.exit_latency = 3,
116 		.target_residency = 6,
117 		.enter = &intel_idle },
118 	{ /* MWAIT C2 */
119 		.name = "C3-NHM",
120 		.desc = "MWAIT 0x10",
121 		.driver_data = (void *) 0x10,
122 		.flags = CPUIDLE_FLAG_TIME_VALID | CPUIDLE_FLAG_TLB_FLUSHED,
123 		.exit_latency = 20,
124 		.target_residency = 80,
125 		.enter = &intel_idle },
126 	{ /* MWAIT C3 */
127 		.name = "C6-NHM",
128 		.desc = "MWAIT 0x20",
129 		.driver_data = (void *) 0x20,
130 		.flags = CPUIDLE_FLAG_TIME_VALID | CPUIDLE_FLAG_TLB_FLUSHED,
131 		.exit_latency = 200,
132 		.target_residency = 800,
133 		.enter = &intel_idle },
134 };
135 
136 static struct cpuidle_state snb_cstates[MWAIT_MAX_NUM_CSTATES] = {
137 	{ /* MWAIT C0 */ },
138 	{ /* MWAIT C1 */
139 		.name = "C1-SNB",
140 		.desc = "MWAIT 0x00",
141 		.driver_data = (void *) 0x00,
142 		.flags = CPUIDLE_FLAG_TIME_VALID,
143 		.exit_latency = 1,
144 		.target_residency = 1,
145 		.enter = &intel_idle },
146 	{ /* MWAIT C2 */
147 		.name = "C3-SNB",
148 		.desc = "MWAIT 0x10",
149 		.driver_data = (void *) 0x10,
150 		.flags = CPUIDLE_FLAG_TIME_VALID | CPUIDLE_FLAG_TLB_FLUSHED,
151 		.exit_latency = 80,
152 		.target_residency = 211,
153 		.enter = &intel_idle },
154 	{ /* MWAIT C3 */
155 		.name = "C6-SNB",
156 		.desc = "MWAIT 0x20",
157 		.driver_data = (void *) 0x20,
158 		.flags = CPUIDLE_FLAG_TIME_VALID | CPUIDLE_FLAG_TLB_FLUSHED,
159 		.exit_latency = 104,
160 		.target_residency = 345,
161 		.enter = &intel_idle },
162 	{ /* MWAIT C4 */
163 		.name = "C7-SNB",
164 		.desc = "MWAIT 0x30",
165 		.driver_data = (void *) 0x30,
166 		.flags = CPUIDLE_FLAG_TIME_VALID | CPUIDLE_FLAG_TLB_FLUSHED,
167 		.exit_latency = 109,
168 		.target_residency = 345,
169 		.enter = &intel_idle },
170 };
171 
172 static struct cpuidle_state atom_cstates[MWAIT_MAX_NUM_CSTATES] = {
173 	{ /* MWAIT C0 */ },
174 	{ /* MWAIT C1 */
175 		.name = "C1-ATM",
176 		.desc = "MWAIT 0x00",
177 		.driver_data = (void *) 0x00,
178 		.flags = CPUIDLE_FLAG_TIME_VALID,
179 		.exit_latency = 1,
180 		.target_residency = 4,
181 		.enter = &intel_idle },
182 	{ /* MWAIT C2 */
183 		.name = "C2-ATM",
184 		.desc = "MWAIT 0x10",
185 		.driver_data = (void *) 0x10,
186 		.flags = CPUIDLE_FLAG_TIME_VALID,
187 		.exit_latency = 20,
188 		.target_residency = 80,
189 		.enter = &intel_idle },
190 	{ /* MWAIT C3 */ },
191 	{ /* MWAIT C4 */
192 		.name = "C4-ATM",
193 		.desc = "MWAIT 0x30",
194 		.driver_data = (void *) 0x30,
195 		.flags = CPUIDLE_FLAG_TIME_VALID | CPUIDLE_FLAG_TLB_FLUSHED,
196 		.exit_latency = 100,
197 		.target_residency = 400,
198 		.enter = &intel_idle },
199 	{ /* MWAIT C5 */ },
200 	{ /* MWAIT C6 */
201 		.name = "C6-ATM",
202 		.desc = "MWAIT 0x52",
203 		.driver_data = (void *) 0x52,
204 		.flags = CPUIDLE_FLAG_TIME_VALID | CPUIDLE_FLAG_TLB_FLUSHED,
205 		.exit_latency = 140,
206 		.target_residency = 560,
207 		.enter = &intel_idle },
208 };
209 
210 /**
211  * intel_idle
212  * @dev: cpuidle_device
213  * @state: cpuidle state
214  *
215  */
216 static int intel_idle(struct cpuidle_device *dev, struct cpuidle_state *state)
217 {
218 	unsigned long ecx = 1; /* break on interrupt flag */
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 	return usec_delta;
261 }
262 
263 static void __setup_broadcast_timer(void *arg)
264 {
265 	unsigned long reason = (unsigned long)arg;
266 	int cpu = smp_processor_id();
267 
268 	reason = reason ?
269 		CLOCK_EVT_NOTIFY_BROADCAST_ON : CLOCK_EVT_NOTIFY_BROADCAST_OFF;
270 
271 	clockevents_notify(reason, &cpu);
272 }
273 
274 static int setup_broadcast_cpuhp_notify(struct notifier_block *n,
275 		unsigned long action, void *hcpu)
276 {
277 	int hotcpu = (unsigned long)hcpu;
278 
279 	switch (action & 0xf) {
280 	case CPU_ONLINE:
281 		smp_call_function_single(hotcpu, __setup_broadcast_timer,
282 			(void *)true, 1);
283 		break;
284 	}
285 	return NOTIFY_OK;
286 }
287 
288 static struct notifier_block setup_broadcast_notifier = {
289 	.notifier_call = setup_broadcast_cpuhp_notify,
290 };
291 
292 static void auto_demotion_disable(void *dummy)
293 {
294 	unsigned long long msr_bits;
295 
296 	rdmsrl(MSR_NHM_SNB_PKG_CST_CFG_CTL, msr_bits);
297 	msr_bits &= ~auto_demotion_disable_flags;
298 	wrmsrl(MSR_NHM_SNB_PKG_CST_CFG_CTL, msr_bits);
299 }
300 
301 /*
302  * intel_idle_probe()
303  */
304 static int intel_idle_probe(void)
305 {
306 	unsigned int eax, ebx, ecx;
307 
308 	if (max_cstate == 0) {
309 		pr_debug(PREFIX "disabled\n");
310 		return -EPERM;
311 	}
312 
313 	if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL)
314 		return -ENODEV;
315 
316 	if (!boot_cpu_has(X86_FEATURE_MWAIT))
317 		return -ENODEV;
318 
319 	if (boot_cpu_data.cpuid_level < CPUID_MWAIT_LEAF)
320 		return -ENODEV;
321 
322 	cpuid(CPUID_MWAIT_LEAF, &eax, &ebx, &ecx, &mwait_substates);
323 
324 	if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED) ||
325 		!(ecx & CPUID5_ECX_INTERRUPT_BREAK))
326 			return -ENODEV;
327 
328 	pr_debug(PREFIX "MWAIT substates: 0x%x\n", mwait_substates);
329 
330 
331 	if (boot_cpu_data.x86 != 6)	/* family 6 */
332 		return -ENODEV;
333 
334 	switch (boot_cpu_data.x86_model) {
335 
336 	case 0x1A:	/* Core i7, Xeon 5500 series */
337 	case 0x1E:	/* Core i7 and i5 Processor - Lynnfield Jasper Forest */
338 	case 0x1F:	/* Core i7 and i5 Processor - Nehalem */
339 	case 0x2E:	/* Nehalem-EX Xeon */
340 	case 0x2F:	/* Westmere-EX Xeon */
341 	case 0x25:	/* Westmere */
342 	case 0x2C:	/* Westmere */
343 		cpuidle_state_table = nehalem_cstates;
344 		auto_demotion_disable_flags =
345 			(NHM_C1_AUTO_DEMOTE | NHM_C3_AUTO_DEMOTE);
346 		break;
347 
348 	case 0x1C:	/* 28 - Atom Processor */
349 		cpuidle_state_table = atom_cstates;
350 		break;
351 
352 	case 0x26:	/* 38 - Lincroft Atom Processor */
353 		cpuidle_state_table = atom_cstates;
354 		auto_demotion_disable_flags = ATM_LNC_C6_AUTO_DEMOTE;
355 		break;
356 
357 	case 0x2A:	/* SNB */
358 	case 0x2D:	/* SNB Xeon */
359 		cpuidle_state_table = snb_cstates;
360 		break;
361 
362 	default:
363 		pr_debug(PREFIX "does not run on family %d model %d\n",
364 			boot_cpu_data.x86, boot_cpu_data.x86_model);
365 		return -ENODEV;
366 	}
367 
368 	if (boot_cpu_has(X86_FEATURE_ARAT))	/* Always Reliable APIC Timer */
369 		lapic_timer_reliable_states = LAPIC_TIMER_ALWAYS_RELIABLE;
370 	else {
371 		smp_call_function(__setup_broadcast_timer, (void *)true, 1);
372 		register_cpu_notifier(&setup_broadcast_notifier);
373 	}
374 
375 	pr_debug(PREFIX "v" INTEL_IDLE_VERSION
376 		" model 0x%X\n", boot_cpu_data.x86_model);
377 
378 	pr_debug(PREFIX "lapic_timer_reliable_states 0x%x\n",
379 		lapic_timer_reliable_states);
380 	return 0;
381 }
382 
383 /*
384  * intel_idle_cpuidle_devices_uninit()
385  * unregister, free cpuidle_devices
386  */
387 static void intel_idle_cpuidle_devices_uninit(void)
388 {
389 	int i;
390 	struct cpuidle_device *dev;
391 
392 	for_each_online_cpu(i) {
393 		dev = per_cpu_ptr(intel_idle_cpuidle_devices, i);
394 		cpuidle_unregister_device(dev);
395 	}
396 
397 	free_percpu(intel_idle_cpuidle_devices);
398 	return;
399 }
400 /*
401  * intel_idle_cpuidle_devices_init()
402  * allocate, initialize, register cpuidle_devices
403  */
404 static int intel_idle_cpuidle_devices_init(void)
405 {
406 	int i, cstate;
407 	struct cpuidle_device *dev;
408 
409 	intel_idle_cpuidle_devices = alloc_percpu(struct cpuidle_device);
410 	if (intel_idle_cpuidle_devices == NULL)
411 		return -ENOMEM;
412 
413 	for_each_online_cpu(i) {
414 		dev = per_cpu_ptr(intel_idle_cpuidle_devices, i);
415 
416 		dev->state_count = 1;
417 
418 		for (cstate = 1; cstate < MWAIT_MAX_NUM_CSTATES; ++cstate) {
419 			int num_substates;
420 
421 			if (cstate > max_cstate) {
422 				printk(PREFIX "max_cstate %d reached\n",
423 					max_cstate);
424 				break;
425 			}
426 
427 			/* does the state exist in CPUID.MWAIT? */
428 			num_substates = (mwait_substates >> ((cstate) * 4))
429 						& MWAIT_SUBSTATE_MASK;
430 			if (num_substates == 0)
431 				continue;
432 			/* is the state not enabled? */
433 			if (cpuidle_state_table[cstate].enter == NULL) {
434 				/* does the driver not know about the state? */
435 				if (*cpuidle_state_table[cstate].name == '\0')
436 					pr_debug(PREFIX "unaware of model 0x%x"
437 						" MWAIT %d please"
438 						" contact lenb@kernel.org",
439 					boot_cpu_data.x86_model, cstate);
440 				continue;
441 			}
442 
443 			if ((cstate > 2) &&
444 				!boot_cpu_has(X86_FEATURE_NONSTOP_TSC))
445 				mark_tsc_unstable("TSC halts in idle"
446 					" states deeper than C2");
447 
448 			dev->states[dev->state_count] =	/* structure copy */
449 				cpuidle_state_table[cstate];
450 
451 			dev->state_count += 1;
452 		}
453 
454 		dev->cpu = i;
455 		if (cpuidle_register_device(dev)) {
456 			pr_debug(PREFIX "cpuidle_register_device %d failed!\n",
457 				 i);
458 			intel_idle_cpuidle_devices_uninit();
459 			return -EIO;
460 		}
461 	}
462 	if (auto_demotion_disable_flags)
463 		smp_call_function(auto_demotion_disable, NULL, 1);
464 
465 	return 0;
466 }
467 
468 
469 static int __init intel_idle_init(void)
470 {
471 	int retval;
472 
473 	/* Do not load intel_idle at all for now if idle= is passed */
474 	if (boot_option_idle_override != IDLE_NO_OVERRIDE)
475 		return -ENODEV;
476 
477 	retval = intel_idle_probe();
478 	if (retval)
479 		return retval;
480 
481 	retval = cpuidle_register_driver(&intel_idle_driver);
482 	if (retval) {
483 		printk(KERN_DEBUG PREFIX "intel_idle yielding to %s",
484 			cpuidle_get_driver()->name);
485 		return retval;
486 	}
487 
488 	retval = intel_idle_cpuidle_devices_init();
489 	if (retval) {
490 		cpuidle_unregister_driver(&intel_idle_driver);
491 		return retval;
492 	}
493 
494 	return 0;
495 }
496 
497 static void __exit intel_idle_exit(void)
498 {
499 	intel_idle_cpuidle_devices_uninit();
500 	cpuidle_unregister_driver(&intel_idle_driver);
501 
502 	if (lapic_timer_reliable_states != LAPIC_TIMER_ALWAYS_RELIABLE) {
503 		smp_call_function(__setup_broadcast_timer, (void *)false, 1);
504 		unregister_cpu_notifier(&setup_broadcast_notifier);
505 	}
506 
507 	return;
508 }
509 
510 module_init(intel_idle_init);
511 module_exit(intel_idle_exit);
512 
513 module_param(max_cstate, int, 0444);
514 
515 MODULE_AUTHOR("Len Brown <len.brown@intel.com>");
516 MODULE_DESCRIPTION("Cpuidle driver for Intel Hardware v" INTEL_IDLE_VERSION);
517 MODULE_LICENSE("GPL");
518