Linux Audio

Check our new training course

Loading...
v3.1
  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
 70static struct cpuidle_driver intel_idle_driver = {
 71	.name = "intel_idle",
 72	.owner = THIS_MODULE,
 73};
 74/* intel_idle.max_cstate=0 disables driver */
 75static int max_cstate = MWAIT_MAX_NUM_CSTATES - 1;
 76
 77static unsigned int mwait_substates;
 78
 79#define LAPIC_TIMER_ALWAYS_RELIABLE 0xFFFFFFFF
 80/* Reliable LAPIC Timer States, bit 1 for C1 etc.  */
 81static unsigned int lapic_timer_reliable_states = (1 << 1);	 /* Default to only C1 */
 82
 83static struct cpuidle_device __percpu *intel_idle_cpuidle_devices;
 84static int intel_idle(struct cpuidle_device *dev, struct cpuidle_state *state);
 85
 86static 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 */
 92static 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 */
107static 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
135static 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
171static 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 * @state: cpuidle state
 
213 *
 
214 */
215static int intel_idle(struct cpuidle_device *dev, struct cpuidle_state *state)
 
216{
217	unsigned long ecx = 1; /* break on interrupt flag */
218	unsigned long eax = (unsigned long)cpuidle_get_statedata(state);
 
219	unsigned int cstate;
220	ktime_t kt_before, kt_after;
221	s64 usec_delta;
222	int cpu = smp_processor_id();
223
224	cstate = (((eax) >> MWAIT_SUBSTATE_SIZE) & MWAIT_CSTATE_MASK) + 1;
225
226	local_irq_disable();
227
228	/*
229	 * leave_mm() to avoid costly and often unnecessary wakeups
230	 * for flushing the user TLB's associated with the active mm.
231	 */
232	if (state->flags & CPUIDLE_FLAG_TLB_FLUSHED)
233		leave_mm(cpu);
234
235	if (!(lapic_timer_reliable_states & (1 << (cstate))))
236		clockevents_notify(CLOCK_EVT_NOTIFY_BROADCAST_ENTER, &cpu);
237
238	kt_before = ktime_get_real();
239
240	stop_critical_timings();
241	if (!need_resched()) {
242
243		__monitor((void *)&current_thread_info()->flags, 0, 0);
244		smp_mb();
245		if (!need_resched())
246			__mwait(eax, ecx);
247	}
248
249	start_critical_timings();
250
251	kt_after = ktime_get_real();
252	usec_delta = ktime_to_us(ktime_sub(kt_after, kt_before));
253
254	local_irq_enable();
255
256	if (!(lapic_timer_reliable_states & (1 << (cstate))))
257		clockevents_notify(CLOCK_EVT_NOTIFY_BROADCAST_EXIT, &cpu);
258
259	return usec_delta;
260}
261
262static void __setup_broadcast_timer(void *arg)
 
 
 
 
 
 
 
263{
264	unsigned long reason = (unsigned long)arg;
265	int cpu = smp_processor_id();
266
267	reason = reason ?
268		CLOCK_EVT_NOTIFY_BROADCAST_ON : CLOCK_EVT_NOTIFY_BROADCAST_OFF;
269
270	clockevents_notify(reason, &cpu);
271}
272
273static int setup_broadcast_cpuhp_notify(struct notifier_block *n,
274		unsigned long action, void *hcpu)
275{
276	int hotcpu = (unsigned long)hcpu;
277
278	switch (action & 0xf) {
279	case CPU_ONLINE:
280		smp_call_function_single(hotcpu, __setup_broadcast_timer,
281			(void *)true, 1);
282		break;
283	}
284	return NOTIFY_OK;
285}
286
287static struct notifier_block setup_broadcast_notifier = {
288	.notifier_call = setup_broadcast_cpuhp_notify,
289};
290
291static void auto_demotion_disable(void *dummy)
292{
293	unsigned long long msr_bits;
294
295	rdmsrl(MSR_NHM_SNB_PKG_CST_CFG_CTL, msr_bits);
296	msr_bits &= ~auto_demotion_disable_flags;
297	wrmsrl(MSR_NHM_SNB_PKG_CST_CFG_CTL, msr_bits);
298}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
299
300/*
301 * intel_idle_probe()
302 */
303static int intel_idle_probe(void)
304{
305	unsigned int eax, ebx, ecx;
 
306
307	if (max_cstate == 0) {
308		pr_debug(PREFIX "disabled\n");
309		return -EPERM;
310	}
311
312	if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL)
313		return -ENODEV;
314
315	if (!boot_cpu_has(X86_FEATURE_MWAIT))
 
 
316		return -ENODEV;
 
317
318	if (boot_cpu_data.cpuid_level < CPUID_MWAIT_LEAF)
319		return -ENODEV;
320
321	cpuid(CPUID_MWAIT_LEAF, &eax, &ebx, &ecx, &mwait_substates);
322
323	if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED) ||
324		!(ecx & CPUID5_ECX_INTERRUPT_BREAK))
 
325			return -ENODEV;
326
327	pr_debug(PREFIX "MWAIT substates: 0x%x\n", mwait_substates);
328
329
330	if (boot_cpu_data.x86 != 6)	/* family 6 */
331		return -ENODEV;
332
333	switch (boot_cpu_data.x86_model) {
334
335	case 0x1A:	/* Core i7, Xeon 5500 series */
336	case 0x1E:	/* Core i7 and i5 Processor - Lynnfield Jasper Forest */
337	case 0x1F:	/* Core i7 and i5 Processor - Nehalem */
338	case 0x2E:	/* Nehalem-EX Xeon */
339	case 0x2F:	/* Westmere-EX Xeon */
340	case 0x25:	/* Westmere */
341	case 0x2C:	/* Westmere */
342		cpuidle_state_table = nehalem_cstates;
343		auto_demotion_disable_flags =
344			(NHM_C1_AUTO_DEMOTE | NHM_C3_AUTO_DEMOTE);
345		break;
346
347	case 0x1C:	/* 28 - Atom Processor */
348		cpuidle_state_table = atom_cstates;
349		break;
350
351	case 0x26:	/* 38 - Lincroft Atom Processor */
352		cpuidle_state_table = atom_cstates;
353		auto_demotion_disable_flags = ATM_LNC_C6_AUTO_DEMOTE;
354		break;
355
356	case 0x2A:	/* SNB */
357	case 0x2D:	/* SNB Xeon */
358		cpuidle_state_table = snb_cstates;
359		break;
360
361	default:
362		pr_debug(PREFIX "does not run on family %d model %d\n",
363			boot_cpu_data.x86, boot_cpu_data.x86_model);
364		return -ENODEV;
365	}
366
367	if (boot_cpu_has(X86_FEATURE_ARAT))	/* Always Reliable APIC Timer */
368		lapic_timer_reliable_states = LAPIC_TIMER_ALWAYS_RELIABLE;
369	else {
370		smp_call_function(__setup_broadcast_timer, (void *)true, 1);
371		register_cpu_notifier(&setup_broadcast_notifier);
372	}
373
374	pr_debug(PREFIX "v" INTEL_IDLE_VERSION
375		" model 0x%X\n", boot_cpu_data.x86_model);
376
377	pr_debug(PREFIX "lapic_timer_reliable_states 0x%x\n",
378		lapic_timer_reliable_states);
379	return 0;
380}
381
382/*
383 * intel_idle_cpuidle_devices_uninit()
384 * unregister, free cpuidle_devices
385 */
386static void intel_idle_cpuidle_devices_uninit(void)
387{
388	int i;
389	struct cpuidle_device *dev;
390
391	for_each_online_cpu(i) {
392		dev = per_cpu_ptr(intel_idle_cpuidle_devices, i);
393		cpuidle_unregister_device(dev);
394	}
 
395
396	free_percpu(intel_idle_cpuidle_devices);
397	return;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
398}
 
399/*
400 * intel_idle_cpuidle_devices_init()
401 * allocate, initialize, register cpuidle_devices
402 */
403static int intel_idle_cpuidle_devices_init(void)
 
 
 
 
404{
405	int i, cstate;
406	struct cpuidle_device *dev;
407
408	intel_idle_cpuidle_devices = alloc_percpu(struct cpuidle_device);
409	if (intel_idle_cpuidle_devices == NULL)
410		return -ENOMEM;
411
412	for_each_online_cpu(i) {
413		dev = per_cpu_ptr(intel_idle_cpuidle_devices, i);
 
 
 
 
 
 
 
 
 
 
 
 
414
415		dev->state_count = 1;
 
 
 
 
 
416
417		for (cstate = 1; cstate < MWAIT_MAX_NUM_CSTATES; ++cstate) {
418			int num_substates;
 
 
 
 
419
420			if (cstate > max_cstate) {
421				printk(PREFIX "max_cstate %d reached\n",
422					max_cstate);
423				break;
424			}
 
425
426			/* does the state exist in CPUID.MWAIT? */
427			num_substates = (mwait_substates >> ((cstate) * 4))
428						& MWAIT_SUBSTATE_MASK;
429			if (num_substates == 0)
430				continue;
431			/* is the state not enabled? */
432			if (cpuidle_state_table[cstate].enter == NULL) {
433				/* does the driver not know about the state? */
434				if (*cpuidle_state_table[cstate].name == '\0')
435					pr_debug(PREFIX "unaware of model 0x%x"
436						" MWAIT %d please"
437						" contact lenb@kernel.org",
438					boot_cpu_data.x86_model, cstate);
439				continue;
440			}
441
442			if ((cstate > 2) &&
443				!boot_cpu_has(X86_FEATURE_NONSTOP_TSC))
444				mark_tsc_unstable("TSC halts in idle"
445					" states deeper than C2");
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
446
447			dev->states[dev->state_count] =	/* structure copy */
448				cpuidle_state_table[cstate];
449
450			dev->state_count += 1;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
451		}
452
453		dev->cpu = i;
454		if (cpuidle_register_device(dev)) {
455			pr_debug(PREFIX "cpuidle_register_device %d failed!\n",
456				 i);
457			intel_idle_cpuidle_devices_uninit();
458			return -EIO;
 
 
 
 
 
 
 
 
 
 
459		}
 
 
 
 
 
 
 
 
 
 
 
460	}
461	if (auto_demotion_disable_flags)
462		smp_call_function(auto_demotion_disable, NULL, 1);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
463
464	return 0;
465}
466
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
467
468static int __init intel_idle_init(void)
469{
470	int retval;
471
472	/* Do not load intel_idle at all for now if idle= is passed */
473	if (boot_option_idle_override != IDLE_NO_OVERRIDE)
474		return -ENODEV;
475
476	retval = intel_idle_probe();
477	if (retval)
478		return retval;
479
 
 
 
 
 
480	retval = cpuidle_register_driver(&intel_idle_driver);
481	if (retval) {
 
482		printk(KERN_DEBUG PREFIX "intel_idle yielding to %s",
483			cpuidle_get_driver()->name);
484		return retval;
485	}
486
487	retval = intel_idle_cpuidle_devices_init();
488	if (retval) {
489		cpuidle_unregister_driver(&intel_idle_driver);
490		return retval;
491	}
 
 
 
 
 
492
493	return 0;
494}
495
496static void __exit intel_idle_exit(void)
497{
498	intel_idle_cpuidle_devices_uninit();
499	cpuidle_unregister_driver(&intel_idle_driver);
 
 
 
500
501	if (lapic_timer_reliable_states != LAPIC_TIMER_ALWAYS_RELIABLE) {
502		smp_call_function(__setup_broadcast_timer, (void *)false, 1);
503		unregister_cpu_notifier(&setup_broadcast_notifier);
504	}
505
506	return;
507}
 
508
509module_init(intel_idle_init);
510module_exit(intel_idle_exit);
511
 
 
 
512module_param(max_cstate, int, 0444);
513
514MODULE_AUTHOR("Len Brown <len.brown@intel.com>");
515MODULE_DESCRIPTION("Cpuidle driver for Intel Hardware v" INTEL_IDLE_VERSION);
516MODULE_LICENSE("GPL");
v4.10.11
   1/*
   2 * intel_idle.c - native hardware idle loop for modern Intel processors
   3 *
   4 * Copyright (c) 2013, 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 */
  50
  51/* un-comment DEBUG to enable pr_debug() statements */
  52#define DEBUG
  53
  54#include <linux/kernel.h>
  55#include <linux/cpuidle.h>
  56#include <linux/tick.h>
 
  57#include <trace/events/power.h>
  58#include <linux/sched.h>
  59#include <linux/notifier.h>
  60#include <linux/cpu.h>
  61#include <linux/moduleparam.h>
  62#include <asm/cpu_device_id.h>
  63#include <asm/intel-family.h>
  64#include <asm/mwait.h>
  65#include <asm/msr.h>
  66
  67#define INTEL_IDLE_VERSION "0.4.1"
  68#define PREFIX "intel_idle: "
  69
  70static struct cpuidle_driver intel_idle_driver = {
  71	.name = "intel_idle",
  72	.owner = THIS_MODULE,
  73};
  74/* intel_idle.max_cstate=0 disables driver */
  75static int max_cstate = CPUIDLE_STATE_MAX - 1;
  76
  77static unsigned int mwait_substates;
  78
  79#define LAPIC_TIMER_ALWAYS_RELIABLE 0xFFFFFFFF
  80/* Reliable LAPIC Timer States, bit 1 for C1 etc.  */
  81static unsigned int lapic_timer_reliable_states = (1 << 1);	 /* Default to only C1 */
  82
  83struct idle_cpu {
  84	struct cpuidle_state *state_table;
  85
  86	/*
  87	 * Hardware C-state auto-demotion may not always be optimal.
  88	 * Indicate which enable bits to clear here.
  89	 */
  90	unsigned long auto_demotion_disable_flags;
  91	bool byt_auto_demotion_disable_flag;
  92	bool disable_promotion_to_c1e;
  93};
  94
  95static const struct idle_cpu *icpu;
  96static struct cpuidle_device __percpu *intel_idle_cpuidle_devices;
  97static int intel_idle(struct cpuidle_device *dev,
  98			struct cpuidle_driver *drv, int index);
  99static void intel_idle_freeze(struct cpuidle_device *dev,
 100			      struct cpuidle_driver *drv, int index);
 101static struct cpuidle_state *cpuidle_state_table;
 102
 103/*
 104 * Set this flag for states where the HW flushes the TLB for us
 105 * and so we don't need cross-calls to keep it consistent.
 106 * If this flag is set, SW flushes the TLB, so even if the
 107 * HW doesn't do the flushing, this flag is safe to use.
 108 */
 109#define CPUIDLE_FLAG_TLB_FLUSHED	0x10000
 110
 111/*
 112 * MWAIT takes an 8-bit "hint" in EAX "suggesting"
 113 * the C-state (top nibble) and sub-state (bottom nibble)
 114 * 0x00 means "MWAIT(C1)", 0x10 means "MWAIT(C2)" etc.
 115 *
 116 * We store the hint at the top of our "flags" for each state.
 117 */
 118#define flg2MWAIT(flags) (((flags) >> 24) & 0xFF)
 119#define MWAIT2flg(eax) ((eax & 0xFF) << 24)
 120
 121/*
 122 * States are indexed by the cstate number,
 123 * which is also the index into the MWAIT hint array.
 124 * Thus C0 is a dummy.
 125 */
 126static struct cpuidle_state nehalem_cstates[] = {
 127	{
 
 128		.name = "C1-NHM",
 129		.desc = "MWAIT 0x00",
 130		.flags = MWAIT2flg(0x00),
 
 131		.exit_latency = 3,
 132		.target_residency = 6,
 133		.enter = &intel_idle,
 134		.enter_freeze = intel_idle_freeze, },
 135	{
 136		.name = "C1E-NHM",
 137		.desc = "MWAIT 0x01",
 138		.flags = MWAIT2flg(0x01),
 139		.exit_latency = 10,
 140		.target_residency = 20,
 141		.enter = &intel_idle,
 142		.enter_freeze = intel_idle_freeze, },
 143	{
 144		.name = "C3-NHM",
 145		.desc = "MWAIT 0x10",
 146		.flags = MWAIT2flg(0x10) | CPUIDLE_FLAG_TLB_FLUSHED,
 
 147		.exit_latency = 20,
 148		.target_residency = 80,
 149		.enter = &intel_idle,
 150		.enter_freeze = intel_idle_freeze, },
 151	{
 152		.name = "C6-NHM",
 153		.desc = "MWAIT 0x20",
 154		.flags = MWAIT2flg(0x20) | CPUIDLE_FLAG_TLB_FLUSHED,
 
 155		.exit_latency = 200,
 156		.target_residency = 800,
 157		.enter = &intel_idle,
 158		.enter_freeze = intel_idle_freeze, },
 159	{
 160		.enter = NULL }
 161};
 162
 163static struct cpuidle_state snb_cstates[] = {
 164	{
 
 165		.name = "C1-SNB",
 166		.desc = "MWAIT 0x00",
 167		.flags = MWAIT2flg(0x00),
 168		.exit_latency = 2,
 169		.target_residency = 2,
 170		.enter = &intel_idle,
 171		.enter_freeze = intel_idle_freeze, },
 172	{
 173		.name = "C1E-SNB",
 174		.desc = "MWAIT 0x01",
 175		.flags = MWAIT2flg(0x01),
 176		.exit_latency = 10,
 177		.target_residency = 20,
 178		.enter = &intel_idle,
 179		.enter_freeze = intel_idle_freeze, },
 180	{
 181		.name = "C3-SNB",
 182		.desc = "MWAIT 0x10",
 183		.flags = MWAIT2flg(0x10) | CPUIDLE_FLAG_TLB_FLUSHED,
 
 184		.exit_latency = 80,
 185		.target_residency = 211,
 186		.enter = &intel_idle,
 187		.enter_freeze = intel_idle_freeze, },
 188	{
 189		.name = "C6-SNB",
 190		.desc = "MWAIT 0x20",
 191		.flags = MWAIT2flg(0x20) | CPUIDLE_FLAG_TLB_FLUSHED,
 
 192		.exit_latency = 104,
 193		.target_residency = 345,
 194		.enter = &intel_idle,
 195		.enter_freeze = intel_idle_freeze, },
 196	{
 197		.name = "C7-SNB",
 198		.desc = "MWAIT 0x30",
 199		.flags = MWAIT2flg(0x30) | CPUIDLE_FLAG_TLB_FLUSHED,
 
 200		.exit_latency = 109,
 201		.target_residency = 345,
 202		.enter = &intel_idle,
 203		.enter_freeze = intel_idle_freeze, },
 204	{
 205		.enter = NULL }
 206};
 207
 208static struct cpuidle_state byt_cstates[] = {
 209	{
 210		.name = "C1-BYT",
 
 211		.desc = "MWAIT 0x00",
 212		.flags = MWAIT2flg(0x00),
 
 213		.exit_latency = 1,
 214		.target_residency = 1,
 215		.enter = &intel_idle,
 216		.enter_freeze = intel_idle_freeze, },
 217	{
 218		.name = "C6N-BYT",
 219		.desc = "MWAIT 0x58",
 220		.flags = MWAIT2flg(0x58) | CPUIDLE_FLAG_TLB_FLUSHED,
 221		.exit_latency = 300,
 222		.target_residency = 275,
 223		.enter = &intel_idle,
 224		.enter_freeze = intel_idle_freeze, },
 225	{
 226		.name = "C6S-BYT",
 227		.desc = "MWAIT 0x52",
 228		.flags = MWAIT2flg(0x52) | CPUIDLE_FLAG_TLB_FLUSHED,
 229		.exit_latency = 500,
 230		.target_residency = 560,
 231		.enter = &intel_idle,
 232		.enter_freeze = intel_idle_freeze, },
 233	{
 234		.name = "C7-BYT",
 235		.desc = "MWAIT 0x60",
 236		.flags = MWAIT2flg(0x60) | CPUIDLE_FLAG_TLB_FLUSHED,
 237		.exit_latency = 1200,
 238		.target_residency = 4000,
 239		.enter = &intel_idle,
 240		.enter_freeze = intel_idle_freeze, },
 241	{
 242		.name = "C7S-BYT",
 243		.desc = "MWAIT 0x64",
 244		.flags = MWAIT2flg(0x64) | CPUIDLE_FLAG_TLB_FLUSHED,
 245		.exit_latency = 10000,
 246		.target_residency = 20000,
 247		.enter = &intel_idle,
 248		.enter_freeze = intel_idle_freeze, },
 249	{
 250		.enter = NULL }
 251};
 252
 253static struct cpuidle_state cht_cstates[] = {
 254	{
 255		.name = "C1-CHT",
 256		.desc = "MWAIT 0x00",
 257		.flags = MWAIT2flg(0x00),
 258		.exit_latency = 1,
 259		.target_residency = 1,
 260		.enter = &intel_idle,
 261		.enter_freeze = intel_idle_freeze, },
 262	{
 263		.name = "C6N-CHT",
 264		.desc = "MWAIT 0x58",
 265		.flags = MWAIT2flg(0x58) | CPUIDLE_FLAG_TLB_FLUSHED,
 266		.exit_latency = 80,
 267		.target_residency = 275,
 268		.enter = &intel_idle,
 269		.enter_freeze = intel_idle_freeze, },
 270	{
 271		.name = "C6S-CHT",
 272		.desc = "MWAIT 0x52",
 273		.flags = MWAIT2flg(0x52) | CPUIDLE_FLAG_TLB_FLUSHED,
 274		.exit_latency = 200,
 275		.target_residency = 560,
 276		.enter = &intel_idle,
 277		.enter_freeze = intel_idle_freeze, },
 278	{
 279		.name = "C7-CHT",
 280		.desc = "MWAIT 0x60",
 281		.flags = MWAIT2flg(0x60) | CPUIDLE_FLAG_TLB_FLUSHED,
 282		.exit_latency = 1200,
 283		.target_residency = 4000,
 284		.enter = &intel_idle,
 285		.enter_freeze = intel_idle_freeze, },
 286	{
 287		.name = "C7S-CHT",
 288		.desc = "MWAIT 0x64",
 289		.flags = MWAIT2flg(0x64) | CPUIDLE_FLAG_TLB_FLUSHED,
 290		.exit_latency = 10000,
 291		.target_residency = 20000,
 292		.enter = &intel_idle,
 293		.enter_freeze = intel_idle_freeze, },
 294	{
 295		.enter = NULL }
 296};
 297
 298static struct cpuidle_state ivb_cstates[] = {
 299	{
 300		.name = "C1-IVB",
 301		.desc = "MWAIT 0x00",
 302		.flags = MWAIT2flg(0x00),
 303		.exit_latency = 1,
 304		.target_residency = 1,
 305		.enter = &intel_idle,
 306		.enter_freeze = intel_idle_freeze, },
 307	{
 308		.name = "C1E-IVB",
 309		.desc = "MWAIT 0x01",
 310		.flags = MWAIT2flg(0x01),
 311		.exit_latency = 10,
 312		.target_residency = 20,
 313		.enter = &intel_idle,
 314		.enter_freeze = intel_idle_freeze, },
 315	{
 316		.name = "C3-IVB",
 317		.desc = "MWAIT 0x10",
 318		.flags = MWAIT2flg(0x10) | CPUIDLE_FLAG_TLB_FLUSHED,
 319		.exit_latency = 59,
 320		.target_residency = 156,
 321		.enter = &intel_idle,
 322		.enter_freeze = intel_idle_freeze, },
 323	{
 324		.name = "C6-IVB",
 325		.desc = "MWAIT 0x20",
 326		.flags = MWAIT2flg(0x20) | CPUIDLE_FLAG_TLB_FLUSHED,
 327		.exit_latency = 80,
 328		.target_residency = 300,
 329		.enter = &intel_idle,
 330		.enter_freeze = intel_idle_freeze, },
 331	{
 332		.name = "C7-IVB",
 333		.desc = "MWAIT 0x30",
 334		.flags = MWAIT2flg(0x30) | CPUIDLE_FLAG_TLB_FLUSHED,
 335		.exit_latency = 87,
 336		.target_residency = 300,
 337		.enter = &intel_idle,
 338		.enter_freeze = intel_idle_freeze, },
 339	{
 340		.enter = NULL }
 341};
 342
 343static struct cpuidle_state ivt_cstates[] = {
 344	{
 345		.name = "C1-IVT",
 346		.desc = "MWAIT 0x00",
 347		.flags = MWAIT2flg(0x00),
 348		.exit_latency = 1,
 349		.target_residency = 1,
 350		.enter = &intel_idle,
 351		.enter_freeze = intel_idle_freeze, },
 352	{
 353		.name = "C1E-IVT",
 354		.desc = "MWAIT 0x01",
 355		.flags = MWAIT2flg(0x01),
 356		.exit_latency = 10,
 357		.target_residency = 80,
 358		.enter = &intel_idle,
 359		.enter_freeze = intel_idle_freeze, },
 360	{
 361		.name = "C3-IVT",
 362		.desc = "MWAIT 0x10",
 363		.flags = MWAIT2flg(0x10) | CPUIDLE_FLAG_TLB_FLUSHED,
 364		.exit_latency = 59,
 365		.target_residency = 156,
 366		.enter = &intel_idle,
 367		.enter_freeze = intel_idle_freeze, },
 368	{
 369		.name = "C6-IVT",
 370		.desc = "MWAIT 0x20",
 371		.flags = MWAIT2flg(0x20) | CPUIDLE_FLAG_TLB_FLUSHED,
 372		.exit_latency = 82,
 373		.target_residency = 300,
 374		.enter = &intel_idle,
 375		.enter_freeze = intel_idle_freeze, },
 376	{
 377		.enter = NULL }
 378};
 379
 380static struct cpuidle_state ivt_cstates_4s[] = {
 381	{
 382		.name = "C1-IVT-4S",
 383		.desc = "MWAIT 0x00",
 384		.flags = MWAIT2flg(0x00),
 385		.exit_latency = 1,
 386		.target_residency = 1,
 387		.enter = &intel_idle,
 388		.enter_freeze = intel_idle_freeze, },
 389	{
 390		.name = "C1E-IVT-4S",
 391		.desc = "MWAIT 0x01",
 392		.flags = MWAIT2flg(0x01),
 393		.exit_latency = 10,
 394		.target_residency = 250,
 395		.enter = &intel_idle,
 396		.enter_freeze = intel_idle_freeze, },
 397	{
 398		.name = "C3-IVT-4S",
 399		.desc = "MWAIT 0x10",
 400		.flags = MWAIT2flg(0x10) | CPUIDLE_FLAG_TLB_FLUSHED,
 401		.exit_latency = 59,
 402		.target_residency = 300,
 403		.enter = &intel_idle,
 404		.enter_freeze = intel_idle_freeze, },
 405	{
 406		.name = "C6-IVT-4S",
 407		.desc = "MWAIT 0x20",
 408		.flags = MWAIT2flg(0x20) | CPUIDLE_FLAG_TLB_FLUSHED,
 409		.exit_latency = 84,
 410		.target_residency = 400,
 411		.enter = &intel_idle,
 412		.enter_freeze = intel_idle_freeze, },
 413	{
 414		.enter = NULL }
 415};
 416
 417static struct cpuidle_state ivt_cstates_8s[] = {
 418	{
 419		.name = "C1-IVT-8S",
 420		.desc = "MWAIT 0x00",
 421		.flags = MWAIT2flg(0x00),
 422		.exit_latency = 1,
 423		.target_residency = 1,
 424		.enter = &intel_idle,
 425		.enter_freeze = intel_idle_freeze, },
 426	{
 427		.name = "C1E-IVT-8S",
 428		.desc = "MWAIT 0x01",
 429		.flags = MWAIT2flg(0x01),
 430		.exit_latency = 10,
 431		.target_residency = 500,
 432		.enter = &intel_idle,
 433		.enter_freeze = intel_idle_freeze, },
 434	{
 435		.name = "C3-IVT-8S",
 436		.desc = "MWAIT 0x10",
 437		.flags = MWAIT2flg(0x10) | CPUIDLE_FLAG_TLB_FLUSHED,
 438		.exit_latency = 59,
 439		.target_residency = 600,
 440		.enter = &intel_idle,
 441		.enter_freeze = intel_idle_freeze, },
 442	{
 443		.name = "C6-IVT-8S",
 444		.desc = "MWAIT 0x20",
 445		.flags = MWAIT2flg(0x20) | CPUIDLE_FLAG_TLB_FLUSHED,
 446		.exit_latency = 88,
 447		.target_residency = 700,
 448		.enter = &intel_idle,
 449		.enter_freeze = intel_idle_freeze, },
 450	{
 451		.enter = NULL }
 452};
 453
 454static struct cpuidle_state hsw_cstates[] = {
 455	{
 456		.name = "C1-HSW",
 457		.desc = "MWAIT 0x00",
 458		.flags = MWAIT2flg(0x00),
 459		.exit_latency = 2,
 460		.target_residency = 2,
 461		.enter = &intel_idle,
 462		.enter_freeze = intel_idle_freeze, },
 463	{
 464		.name = "C1E-HSW",
 465		.desc = "MWAIT 0x01",
 466		.flags = MWAIT2flg(0x01),
 467		.exit_latency = 10,
 468		.target_residency = 20,
 469		.enter = &intel_idle,
 470		.enter_freeze = intel_idle_freeze, },
 471	{
 472		.name = "C3-HSW",
 473		.desc = "MWAIT 0x10",
 474		.flags = MWAIT2flg(0x10) | CPUIDLE_FLAG_TLB_FLUSHED,
 475		.exit_latency = 33,
 476		.target_residency = 100,
 477		.enter = &intel_idle,
 478		.enter_freeze = intel_idle_freeze, },
 479	{
 480		.name = "C6-HSW",
 481		.desc = "MWAIT 0x20",
 482		.flags = MWAIT2flg(0x20) | CPUIDLE_FLAG_TLB_FLUSHED,
 483		.exit_latency = 133,
 484		.target_residency = 400,
 485		.enter = &intel_idle,
 486		.enter_freeze = intel_idle_freeze, },
 487	{
 488		.name = "C7s-HSW",
 489		.desc = "MWAIT 0x32",
 490		.flags = MWAIT2flg(0x32) | CPUIDLE_FLAG_TLB_FLUSHED,
 491		.exit_latency = 166,
 492		.target_residency = 500,
 493		.enter = &intel_idle,
 494		.enter_freeze = intel_idle_freeze, },
 495	{
 496		.name = "C8-HSW",
 497		.desc = "MWAIT 0x40",
 498		.flags = MWAIT2flg(0x40) | CPUIDLE_FLAG_TLB_FLUSHED,
 499		.exit_latency = 300,
 500		.target_residency = 900,
 501		.enter = &intel_idle,
 502		.enter_freeze = intel_idle_freeze, },
 503	{
 504		.name = "C9-HSW",
 505		.desc = "MWAIT 0x50",
 506		.flags = MWAIT2flg(0x50) | CPUIDLE_FLAG_TLB_FLUSHED,
 507		.exit_latency = 600,
 508		.target_residency = 1800,
 509		.enter = &intel_idle,
 510		.enter_freeze = intel_idle_freeze, },
 511	{
 512		.name = "C10-HSW",
 513		.desc = "MWAIT 0x60",
 514		.flags = MWAIT2flg(0x60) | CPUIDLE_FLAG_TLB_FLUSHED,
 515		.exit_latency = 2600,
 516		.target_residency = 7700,
 517		.enter = &intel_idle,
 518		.enter_freeze = intel_idle_freeze, },
 519	{
 520		.enter = NULL }
 521};
 522static struct cpuidle_state bdw_cstates[] = {
 523	{
 524		.name = "C1-BDW",
 525		.desc = "MWAIT 0x00",
 526		.flags = MWAIT2flg(0x00),
 527		.exit_latency = 2,
 528		.target_residency = 2,
 529		.enter = &intel_idle,
 530		.enter_freeze = intel_idle_freeze, },
 531	{
 532		.name = "C1E-BDW",
 533		.desc = "MWAIT 0x01",
 534		.flags = MWAIT2flg(0x01),
 535		.exit_latency = 10,
 536		.target_residency = 20,
 537		.enter = &intel_idle,
 538		.enter_freeze = intel_idle_freeze, },
 539	{
 540		.name = "C3-BDW",
 541		.desc = "MWAIT 0x10",
 542		.flags = MWAIT2flg(0x10) | CPUIDLE_FLAG_TLB_FLUSHED,
 543		.exit_latency = 40,
 544		.target_residency = 100,
 545		.enter = &intel_idle,
 546		.enter_freeze = intel_idle_freeze, },
 547	{
 548		.name = "C6-BDW",
 549		.desc = "MWAIT 0x20",
 550		.flags = MWAIT2flg(0x20) | CPUIDLE_FLAG_TLB_FLUSHED,
 551		.exit_latency = 133,
 552		.target_residency = 400,
 553		.enter = &intel_idle,
 554		.enter_freeze = intel_idle_freeze, },
 555	{
 556		.name = "C7s-BDW",
 557		.desc = "MWAIT 0x32",
 558		.flags = MWAIT2flg(0x32) | CPUIDLE_FLAG_TLB_FLUSHED,
 559		.exit_latency = 166,
 560		.target_residency = 500,
 561		.enter = &intel_idle,
 562		.enter_freeze = intel_idle_freeze, },
 563	{
 564		.name = "C8-BDW",
 565		.desc = "MWAIT 0x40",
 566		.flags = MWAIT2flg(0x40) | CPUIDLE_FLAG_TLB_FLUSHED,
 567		.exit_latency = 300,
 568		.target_residency = 900,
 569		.enter = &intel_idle,
 570		.enter_freeze = intel_idle_freeze, },
 571	{
 572		.name = "C9-BDW",
 573		.desc = "MWAIT 0x50",
 574		.flags = MWAIT2flg(0x50) | CPUIDLE_FLAG_TLB_FLUSHED,
 575		.exit_latency = 600,
 576		.target_residency = 1800,
 577		.enter = &intel_idle,
 578		.enter_freeze = intel_idle_freeze, },
 579	{
 580		.name = "C10-BDW",
 581		.desc = "MWAIT 0x60",
 582		.flags = MWAIT2flg(0x60) | CPUIDLE_FLAG_TLB_FLUSHED,
 583		.exit_latency = 2600,
 584		.target_residency = 7700,
 585		.enter = &intel_idle,
 586		.enter_freeze = intel_idle_freeze, },
 587	{
 588		.enter = NULL }
 589};
 590
 591static struct cpuidle_state skl_cstates[] = {
 592	{
 593		.name = "C1-SKL",
 594		.desc = "MWAIT 0x00",
 595		.flags = MWAIT2flg(0x00),
 596		.exit_latency = 2,
 597		.target_residency = 2,
 598		.enter = &intel_idle,
 599		.enter_freeze = intel_idle_freeze, },
 600	{
 601		.name = "C1E-SKL",
 602		.desc = "MWAIT 0x01",
 603		.flags = MWAIT2flg(0x01),
 604		.exit_latency = 10,
 605		.target_residency = 20,
 606		.enter = &intel_idle,
 607		.enter_freeze = intel_idle_freeze, },
 608	{
 609		.name = "C3-SKL",
 610		.desc = "MWAIT 0x10",
 611		.flags = MWAIT2flg(0x10) | CPUIDLE_FLAG_TLB_FLUSHED,
 612		.exit_latency = 70,
 613		.target_residency = 100,
 614		.enter = &intel_idle,
 615		.enter_freeze = intel_idle_freeze, },
 616	{
 617		.name = "C6-SKL",
 618		.desc = "MWAIT 0x20",
 619		.flags = MWAIT2flg(0x20) | CPUIDLE_FLAG_TLB_FLUSHED,
 620		.exit_latency = 85,
 621		.target_residency = 200,
 622		.enter = &intel_idle,
 623		.enter_freeze = intel_idle_freeze, },
 624	{
 625		.name = "C7s-SKL",
 626		.desc = "MWAIT 0x33",
 627		.flags = MWAIT2flg(0x33) | CPUIDLE_FLAG_TLB_FLUSHED,
 628		.exit_latency = 124,
 629		.target_residency = 800,
 630		.enter = &intel_idle,
 631		.enter_freeze = intel_idle_freeze, },
 632	{
 633		.name = "C8-SKL",
 634		.desc = "MWAIT 0x40",
 635		.flags = MWAIT2flg(0x40) | CPUIDLE_FLAG_TLB_FLUSHED,
 636		.exit_latency = 200,
 637		.target_residency = 800,
 638		.enter = &intel_idle,
 639		.enter_freeze = intel_idle_freeze, },
 640	{
 641		.name = "C9-SKL",
 642		.desc = "MWAIT 0x50",
 643		.flags = MWAIT2flg(0x50) | CPUIDLE_FLAG_TLB_FLUSHED,
 644		.exit_latency = 480,
 645		.target_residency = 5000,
 646		.enter = &intel_idle,
 647		.enter_freeze = intel_idle_freeze, },
 648	{
 649		.name = "C10-SKL",
 650		.desc = "MWAIT 0x60",
 651		.flags = MWAIT2flg(0x60) | CPUIDLE_FLAG_TLB_FLUSHED,
 652		.exit_latency = 890,
 653		.target_residency = 5000,
 654		.enter = &intel_idle,
 655		.enter_freeze = intel_idle_freeze, },
 656	{
 657		.enter = NULL }
 658};
 659
 660static struct cpuidle_state skx_cstates[] = {
 661	{
 662		.name = "C1-SKX",
 663		.desc = "MWAIT 0x00",
 664		.flags = MWAIT2flg(0x00),
 665		.exit_latency = 2,
 666		.target_residency = 2,
 667		.enter = &intel_idle,
 668		.enter_freeze = intel_idle_freeze, },
 669	{
 670		.name = "C1E-SKX",
 671		.desc = "MWAIT 0x01",
 672		.flags = MWAIT2flg(0x01),
 673		.exit_latency = 10,
 674		.target_residency = 20,
 675		.enter = &intel_idle,
 676		.enter_freeze = intel_idle_freeze, },
 677	{
 678		.name = "C6-SKX",
 679		.desc = "MWAIT 0x20",
 680		.flags = MWAIT2flg(0x20) | CPUIDLE_FLAG_TLB_FLUSHED,
 681		.exit_latency = 133,
 682		.target_residency = 600,
 683		.enter = &intel_idle,
 684		.enter_freeze = intel_idle_freeze, },
 685	{
 686		.enter = NULL }
 687};
 688
 689static struct cpuidle_state atom_cstates[] = {
 690	{
 691		.name = "C1E-ATM",
 692		.desc = "MWAIT 0x00",
 693		.flags = MWAIT2flg(0x00),
 694		.exit_latency = 10,
 695		.target_residency = 20,
 696		.enter = &intel_idle,
 697		.enter_freeze = intel_idle_freeze, },
 698	{
 699		.name = "C2-ATM",
 700		.desc = "MWAIT 0x10",
 701		.flags = MWAIT2flg(0x10),
 
 702		.exit_latency = 20,
 703		.target_residency = 80,
 704		.enter = &intel_idle,
 705		.enter_freeze = intel_idle_freeze, },
 706	{
 707		.name = "C4-ATM",
 708		.desc = "MWAIT 0x30",
 709		.flags = MWAIT2flg(0x30) | CPUIDLE_FLAG_TLB_FLUSHED,
 
 710		.exit_latency = 100,
 711		.target_residency = 400,
 712		.enter = &intel_idle,
 713		.enter_freeze = intel_idle_freeze, },
 714	{
 715		.name = "C6-ATM",
 716		.desc = "MWAIT 0x52",
 717		.flags = MWAIT2flg(0x52) | CPUIDLE_FLAG_TLB_FLUSHED,
 718		.exit_latency = 140,
 719		.target_residency = 560,
 720		.enter = &intel_idle,
 721		.enter_freeze = intel_idle_freeze, },
 722	{
 723		.enter = NULL }
 724};
 725static struct cpuidle_state tangier_cstates[] = {
 726	{
 727		.name = "C1-TNG",
 728		.desc = "MWAIT 0x00",
 729		.flags = MWAIT2flg(0x00),
 730		.exit_latency = 1,
 731		.target_residency = 4,
 732		.enter = &intel_idle,
 733		.enter_freeze = intel_idle_freeze, },
 734	{
 735		.name = "C4-TNG",
 736		.desc = "MWAIT 0x30",
 737		.flags = MWAIT2flg(0x30) | CPUIDLE_FLAG_TLB_FLUSHED,
 738		.exit_latency = 100,
 739		.target_residency = 400,
 740		.enter = &intel_idle,
 741		.enter_freeze = intel_idle_freeze, },
 742	{
 743		.name = "C6-TNG",
 744		.desc = "MWAIT 0x52",
 745		.flags = MWAIT2flg(0x52) | CPUIDLE_FLAG_TLB_FLUSHED,
 746		.exit_latency = 140,
 747		.target_residency = 560,
 748		.enter = &intel_idle,
 749		.enter_freeze = intel_idle_freeze, },
 750	{
 751		.name = "C7-TNG",
 752		.desc = "MWAIT 0x60",
 753		.flags = MWAIT2flg(0x60) | CPUIDLE_FLAG_TLB_FLUSHED,
 754		.exit_latency = 1200,
 755		.target_residency = 4000,
 756		.enter = &intel_idle,
 757		.enter_freeze = intel_idle_freeze, },
 758	{
 759		.name = "C9-TNG",
 760		.desc = "MWAIT 0x64",
 761		.flags = MWAIT2flg(0x64) | CPUIDLE_FLAG_TLB_FLUSHED,
 762		.exit_latency = 10000,
 763		.target_residency = 20000,
 764		.enter = &intel_idle,
 765		.enter_freeze = intel_idle_freeze, },
 766	{
 767		.enter = NULL }
 768};
 769static struct cpuidle_state avn_cstates[] = {
 770	{
 771		.name = "C1-AVN",
 772		.desc = "MWAIT 0x00",
 773		.flags = MWAIT2flg(0x00),
 774		.exit_latency = 2,
 775		.target_residency = 2,
 776		.enter = &intel_idle,
 777		.enter_freeze = intel_idle_freeze, },
 778	{
 779		.name = "C6-AVN",
 780		.desc = "MWAIT 0x51",
 781		.flags = MWAIT2flg(0x51) | CPUIDLE_FLAG_TLB_FLUSHED,
 782		.exit_latency = 15,
 783		.target_residency = 45,
 784		.enter = &intel_idle,
 785		.enter_freeze = intel_idle_freeze, },
 786	{
 787		.enter = NULL }
 788};
 789static struct cpuidle_state knl_cstates[] = {
 790	{
 791		.name = "C1-KNL",
 792		.desc = "MWAIT 0x00",
 793		.flags = MWAIT2flg(0x00),
 794		.exit_latency = 1,
 795		.target_residency = 2,
 796		.enter = &intel_idle,
 797		.enter_freeze = intel_idle_freeze },
 798	{
 799		.name = "C6-KNL",
 800		.desc = "MWAIT 0x10",
 801		.flags = MWAIT2flg(0x10) | CPUIDLE_FLAG_TLB_FLUSHED,
 802		.exit_latency = 120,
 803		.target_residency = 500,
 804		.enter = &intel_idle,
 805		.enter_freeze = intel_idle_freeze },
 806	{
 807		.enter = NULL }
 808};
 809
 810static struct cpuidle_state bxt_cstates[] = {
 811	{
 812		.name = "C1-BXT",
 813		.desc = "MWAIT 0x00",
 814		.flags = MWAIT2flg(0x00),
 815		.exit_latency = 2,
 816		.target_residency = 2,
 817		.enter = &intel_idle,
 818		.enter_freeze = intel_idle_freeze, },
 819	{
 820		.name = "C1E-BXT",
 821		.desc = "MWAIT 0x01",
 822		.flags = MWAIT2flg(0x01),
 823		.exit_latency = 10,
 824		.target_residency = 20,
 825		.enter = &intel_idle,
 826		.enter_freeze = intel_idle_freeze, },
 827	{
 828		.name = "C6-BXT",
 829		.desc = "MWAIT 0x20",
 830		.flags = MWAIT2flg(0x20) | CPUIDLE_FLAG_TLB_FLUSHED,
 831		.exit_latency = 133,
 832		.target_residency = 133,
 833		.enter = &intel_idle,
 834		.enter_freeze = intel_idle_freeze, },
 835	{
 836		.name = "C7s-BXT",
 837		.desc = "MWAIT 0x31",
 838		.flags = MWAIT2flg(0x31) | CPUIDLE_FLAG_TLB_FLUSHED,
 839		.exit_latency = 155,
 840		.target_residency = 155,
 841		.enter = &intel_idle,
 842		.enter_freeze = intel_idle_freeze, },
 843	{
 844		.name = "C8-BXT",
 845		.desc = "MWAIT 0x40",
 846		.flags = MWAIT2flg(0x40) | CPUIDLE_FLAG_TLB_FLUSHED,
 847		.exit_latency = 1000,
 848		.target_residency = 1000,
 849		.enter = &intel_idle,
 850		.enter_freeze = intel_idle_freeze, },
 851	{
 852		.name = "C9-BXT",
 853		.desc = "MWAIT 0x50",
 854		.flags = MWAIT2flg(0x50) | CPUIDLE_FLAG_TLB_FLUSHED,
 855		.exit_latency = 2000,
 856		.target_residency = 2000,
 857		.enter = &intel_idle,
 858		.enter_freeze = intel_idle_freeze, },
 859	{
 860		.name = "C10-BXT",
 861		.desc = "MWAIT 0x60",
 862		.flags = MWAIT2flg(0x60) | CPUIDLE_FLAG_TLB_FLUSHED,
 863		.exit_latency = 10000,
 864		.target_residency = 10000,
 865		.enter = &intel_idle,
 866		.enter_freeze = intel_idle_freeze, },
 867	{
 868		.enter = NULL }
 869};
 870
 871static struct cpuidle_state dnv_cstates[] = {
 872	{
 873		.name = "C1-DNV",
 874		.desc = "MWAIT 0x00",
 875		.flags = MWAIT2flg(0x00),
 876		.exit_latency = 2,
 877		.target_residency = 2,
 878		.enter = &intel_idle,
 879		.enter_freeze = intel_idle_freeze, },
 880	{
 881		.name = "C1E-DNV",
 882		.desc = "MWAIT 0x01",
 883		.flags = MWAIT2flg(0x01),
 884		.exit_latency = 10,
 885		.target_residency = 20,
 886		.enter = &intel_idle,
 887		.enter_freeze = intel_idle_freeze, },
 888	{
 889		.name = "C6-DNV",
 890		.desc = "MWAIT 0x20",
 891		.flags = MWAIT2flg(0x20) | CPUIDLE_FLAG_TLB_FLUSHED,
 892		.exit_latency = 50,
 893		.target_residency = 500,
 894		.enter = &intel_idle,
 895		.enter_freeze = intel_idle_freeze, },
 896	{
 897		.enter = NULL }
 898};
 899
 900/**
 901 * intel_idle
 902 * @dev: cpuidle_device
 903 * @drv: cpuidle driver
 904 * @index: index of cpuidle state
 905 *
 906 * Must be called under local_irq_disable().
 907 */
 908static __cpuidle int intel_idle(struct cpuidle_device *dev,
 909				struct cpuidle_driver *drv, int index)
 910{
 911	unsigned long ecx = 1; /* break on interrupt flag */
 912	struct cpuidle_state *state = &drv->states[index];
 913	unsigned long eax = flg2MWAIT(state->flags);
 914	unsigned int cstate;
 
 
 915	int cpu = smp_processor_id();
 916
 917	cstate = (((eax) >> MWAIT_SUBSTATE_SIZE) & MWAIT_CSTATE_MASK) + 1;
 918
 
 
 919	/*
 920	 * leave_mm() to avoid costly and often unnecessary wakeups
 921	 * for flushing the user TLB's associated with the active mm.
 922	 */
 923	if (state->flags & CPUIDLE_FLAG_TLB_FLUSHED)
 924		leave_mm(cpu);
 925
 926	if (!(lapic_timer_reliable_states & (1 << (cstate))))
 927		tick_broadcast_enter();
 
 
 
 
 
 
 
 
 
 
 
 928
 929	mwait_idle_with_hints(eax, ecx);
 
 
 
 
 
 930
 931	if (!(lapic_timer_reliable_states & (1 << (cstate))))
 932		tick_broadcast_exit();
 933
 934	return index;
 935}
 936
 937/**
 938 * intel_idle_freeze - simplified "enter" callback routine for suspend-to-idle
 939 * @dev: cpuidle_device
 940 * @drv: cpuidle driver
 941 * @index: state index
 942 */
 943static void intel_idle_freeze(struct cpuidle_device *dev,
 944			     struct cpuidle_driver *drv, int index)
 945{
 946	unsigned long ecx = 1; /* break on interrupt flag */
 947	unsigned long eax = flg2MWAIT(drv->states[index].flags);
 
 
 
 948
 949	mwait_idle_with_hints(eax, ecx);
 950}
 951
 952static void __setup_broadcast_timer(bool on)
 
 953{
 954	if (on)
 955		tick_broadcast_enable();
 956	else
 957		tick_broadcast_disable();
 
 
 
 
 
 958}
 959
 960static void auto_demotion_disable(void)
 
 
 
 
 961{
 962	unsigned long long msr_bits;
 963
 964	rdmsrl(MSR_NHM_SNB_PKG_CST_CFG_CTL, msr_bits);
 965	msr_bits &= ~(icpu->auto_demotion_disable_flags);
 966	wrmsrl(MSR_NHM_SNB_PKG_CST_CFG_CTL, msr_bits);
 967}
 968static void c1e_promotion_disable(void)
 969{
 970	unsigned long long msr_bits;
 971
 972	rdmsrl(MSR_IA32_POWER_CTL, msr_bits);
 973	msr_bits &= ~0x2;
 974	wrmsrl(MSR_IA32_POWER_CTL, msr_bits);
 975}
 976
 977static const struct idle_cpu idle_cpu_nehalem = {
 978	.state_table = nehalem_cstates,
 979	.auto_demotion_disable_flags = NHM_C1_AUTO_DEMOTE | NHM_C3_AUTO_DEMOTE,
 980	.disable_promotion_to_c1e = true,
 981};
 982
 983static const struct idle_cpu idle_cpu_atom = {
 984	.state_table = atom_cstates,
 985};
 986
 987static const struct idle_cpu idle_cpu_tangier = {
 988	.state_table = tangier_cstates,
 989};
 990
 991static const struct idle_cpu idle_cpu_lincroft = {
 992	.state_table = atom_cstates,
 993	.auto_demotion_disable_flags = ATM_LNC_C6_AUTO_DEMOTE,
 994};
 995
 996static const struct idle_cpu idle_cpu_snb = {
 997	.state_table = snb_cstates,
 998	.disable_promotion_to_c1e = true,
 999};
1000
1001static const struct idle_cpu idle_cpu_byt = {
1002	.state_table = byt_cstates,
1003	.disable_promotion_to_c1e = true,
1004	.byt_auto_demotion_disable_flag = true,
1005};
1006
1007static const struct idle_cpu idle_cpu_cht = {
1008	.state_table = cht_cstates,
1009	.disable_promotion_to_c1e = true,
1010	.byt_auto_demotion_disable_flag = true,
1011};
1012
1013static const struct idle_cpu idle_cpu_ivb = {
1014	.state_table = ivb_cstates,
1015	.disable_promotion_to_c1e = true,
1016};
1017
1018static const struct idle_cpu idle_cpu_ivt = {
1019	.state_table = ivt_cstates,
1020	.disable_promotion_to_c1e = true,
1021};
1022
1023static const struct idle_cpu idle_cpu_hsw = {
1024	.state_table = hsw_cstates,
1025	.disable_promotion_to_c1e = true,
1026};
1027
1028static const struct idle_cpu idle_cpu_bdw = {
1029	.state_table = bdw_cstates,
1030	.disable_promotion_to_c1e = true,
1031};
1032
1033static const struct idle_cpu idle_cpu_skl = {
1034	.state_table = skl_cstates,
1035	.disable_promotion_to_c1e = true,
1036};
1037
1038static const struct idle_cpu idle_cpu_skx = {
1039	.state_table = skx_cstates,
1040	.disable_promotion_to_c1e = true,
1041};
1042
1043static const struct idle_cpu idle_cpu_avn = {
1044	.state_table = avn_cstates,
1045	.disable_promotion_to_c1e = true,
1046};
1047
1048static const struct idle_cpu idle_cpu_knl = {
1049	.state_table = knl_cstates,
1050};
1051
1052static const struct idle_cpu idle_cpu_bxt = {
1053	.state_table = bxt_cstates,
1054	.disable_promotion_to_c1e = true,
1055};
1056
1057static const struct idle_cpu idle_cpu_dnv = {
1058	.state_table = dnv_cstates,
1059	.disable_promotion_to_c1e = true,
1060};
1061
1062#define ICPU(model, cpu) \
1063	{ X86_VENDOR_INTEL, 6, model, X86_FEATURE_MWAIT, (unsigned long)&cpu }
1064
1065static const struct x86_cpu_id intel_idle_ids[] __initconst = {
1066	ICPU(INTEL_FAM6_NEHALEM_EP,		idle_cpu_nehalem),
1067	ICPU(INTEL_FAM6_NEHALEM,		idle_cpu_nehalem),
1068	ICPU(INTEL_FAM6_NEHALEM_G,		idle_cpu_nehalem),
1069	ICPU(INTEL_FAM6_WESTMERE,		idle_cpu_nehalem),
1070	ICPU(INTEL_FAM6_WESTMERE_EP,		idle_cpu_nehalem),
1071	ICPU(INTEL_FAM6_NEHALEM_EX,		idle_cpu_nehalem),
1072	ICPU(INTEL_FAM6_ATOM_PINEVIEW,		idle_cpu_atom),
1073	ICPU(INTEL_FAM6_ATOM_LINCROFT,		idle_cpu_lincroft),
1074	ICPU(INTEL_FAM6_WESTMERE_EX,		idle_cpu_nehalem),
1075	ICPU(INTEL_FAM6_SANDYBRIDGE,		idle_cpu_snb),
1076	ICPU(INTEL_FAM6_SANDYBRIDGE_X,		idle_cpu_snb),
1077	ICPU(INTEL_FAM6_ATOM_CEDARVIEW,		idle_cpu_atom),
1078	ICPU(INTEL_FAM6_ATOM_SILVERMONT1,	idle_cpu_byt),
1079	ICPU(INTEL_FAM6_ATOM_MERRIFIELD,	idle_cpu_tangier),
1080	ICPU(INTEL_FAM6_ATOM_AIRMONT,		idle_cpu_cht),
1081	ICPU(INTEL_FAM6_IVYBRIDGE,		idle_cpu_ivb),
1082	ICPU(INTEL_FAM6_IVYBRIDGE_X,		idle_cpu_ivt),
1083	ICPU(INTEL_FAM6_HASWELL_CORE,		idle_cpu_hsw),
1084	ICPU(INTEL_FAM6_HASWELL_X,		idle_cpu_hsw),
1085	ICPU(INTEL_FAM6_HASWELL_ULT,		idle_cpu_hsw),
1086	ICPU(INTEL_FAM6_HASWELL_GT3E,		idle_cpu_hsw),
1087	ICPU(INTEL_FAM6_ATOM_SILVERMONT2,	idle_cpu_avn),
1088	ICPU(INTEL_FAM6_BROADWELL_CORE,		idle_cpu_bdw),
1089	ICPU(INTEL_FAM6_BROADWELL_GT3E,		idle_cpu_bdw),
1090	ICPU(INTEL_FAM6_BROADWELL_X,		idle_cpu_bdw),
1091	ICPU(INTEL_FAM6_BROADWELL_XEON_D,	idle_cpu_bdw),
1092	ICPU(INTEL_FAM6_SKYLAKE_MOBILE,		idle_cpu_skl),
1093	ICPU(INTEL_FAM6_SKYLAKE_DESKTOP,	idle_cpu_skl),
1094	ICPU(INTEL_FAM6_KABYLAKE_MOBILE,	idle_cpu_skl),
1095	ICPU(INTEL_FAM6_KABYLAKE_DESKTOP,	idle_cpu_skl),
1096	ICPU(INTEL_FAM6_SKYLAKE_X,		idle_cpu_skx),
1097	ICPU(INTEL_FAM6_XEON_PHI_KNL,		idle_cpu_knl),
1098	ICPU(INTEL_FAM6_XEON_PHI_KNM,		idle_cpu_knl),
1099	ICPU(INTEL_FAM6_ATOM_GOLDMONT,		idle_cpu_bxt),
1100	ICPU(INTEL_FAM6_ATOM_DENVERTON,		idle_cpu_dnv),
1101	{}
1102};
1103
1104/*
1105 * intel_idle_probe()
1106 */
1107static int __init intel_idle_probe(void)
1108{
1109	unsigned int eax, ebx, ecx;
1110	const struct x86_cpu_id *id;
1111
1112	if (max_cstate == 0) {
1113		pr_debug(PREFIX "disabled\n");
1114		return -EPERM;
1115	}
1116
1117	id = x86_match_cpu(intel_idle_ids);
1118	if (!id) {
1119		if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL &&
1120		    boot_cpu_data.x86 == 6)
1121			pr_debug(PREFIX "does not run on family %d model %d\n",
1122				boot_cpu_data.x86, boot_cpu_data.x86_model);
1123		return -ENODEV;
1124	}
1125
1126	if (boot_cpu_data.cpuid_level < CPUID_MWAIT_LEAF)
1127		return -ENODEV;
1128
1129	cpuid(CPUID_MWAIT_LEAF, &eax, &ebx, &ecx, &mwait_substates);
1130
1131	if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED) ||
1132	    !(ecx & CPUID5_ECX_INTERRUPT_BREAK) ||
1133	    !mwait_substates)
1134			return -ENODEV;
1135
1136	pr_debug(PREFIX "MWAIT substates: 0x%x\n", mwait_substates);
1137
1138	icpu = (const struct idle_cpu *)id->driver_data;
1139	cpuidle_state_table = icpu->state_table;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1140
1141	pr_debug(PREFIX "v" INTEL_IDLE_VERSION
1142		" model 0x%X\n", boot_cpu_data.x86_model);
1143
 
 
1144	return 0;
1145}
1146
1147/*
1148 * intel_idle_cpuidle_devices_uninit()
1149 * Unregisters the cpuidle devices.
1150 */
1151static void intel_idle_cpuidle_devices_uninit(void)
1152{
1153	int i;
1154	struct cpuidle_device *dev;
1155
1156	for_each_online_cpu(i) {
1157		dev = per_cpu_ptr(intel_idle_cpuidle_devices, i);
1158		cpuidle_unregister_device(dev);
1159	}
1160}
1161
1162/*
1163 * ivt_idle_state_table_update(void)
1164 *
1165 * Tune IVT multi-socket targets
1166 * Assumption: num_sockets == (max_package_num + 1)
1167 */
1168static void ivt_idle_state_table_update(void)
1169{
1170	/* IVT uses a different table for 1-2, 3-4, and > 4 sockets */
1171	int cpu, package_num, num_sockets = 1;
1172
1173	for_each_online_cpu(cpu) {
1174		package_num = topology_physical_package_id(cpu);
1175		if (package_num + 1 > num_sockets) {
1176			num_sockets = package_num + 1;
1177
1178			if (num_sockets > 4) {
1179				cpuidle_state_table = ivt_cstates_8s;
1180				return;
1181			}
1182		}
1183	}
1184
1185	if (num_sockets > 2)
1186		cpuidle_state_table = ivt_cstates_4s;
1187
1188	/* else, 1 and 2 socket systems use default ivt_cstates */
1189}
1190
1191/*
1192 * Translate IRTL (Interrupt Response Time Limit) MSR to usec
 
1193 */
1194
1195static unsigned int irtl_ns_units[] = {
1196	1, 32, 1024, 32768, 1048576, 33554432, 0, 0 };
1197
1198static unsigned long long irtl_2_usec(unsigned long long irtl)
1199{
1200	unsigned long long ns;
 
1201
1202	if (!irtl)
1203		return 0;
 
1204
1205	ns = irtl_ns_units[(irtl >> 10) & 0x7];
1206
1207	return div64_u64((irtl & 0x3FF) * ns, 1000);
1208}
1209/*
1210 * bxt_idle_state_table_update(void)
1211 *
1212 * On BXT, we trust the IRTL to show the definitive maximum latency
1213 * We use the same value for target_residency.
1214 */
1215static void bxt_idle_state_table_update(void)
1216{
1217	unsigned long long msr;
1218	unsigned int usec;
1219
1220	rdmsrl(MSR_PKGC6_IRTL, msr);
1221	usec = irtl_2_usec(msr);
1222	if (usec) {
1223		bxt_cstates[2].exit_latency = usec;
1224		bxt_cstates[2].target_residency = usec;
1225	}
1226
1227	rdmsrl(MSR_PKGC7_IRTL, msr);
1228	usec = irtl_2_usec(msr);
1229	if (usec) {
1230		bxt_cstates[3].exit_latency = usec;
1231		bxt_cstates[3].target_residency = usec;
1232	}
1233
1234	rdmsrl(MSR_PKGC8_IRTL, msr);
1235	usec = irtl_2_usec(msr);
1236	if (usec) {
1237		bxt_cstates[4].exit_latency = usec;
1238		bxt_cstates[4].target_residency = usec;
1239	}
1240
1241	rdmsrl(MSR_PKGC9_IRTL, msr);
1242	usec = irtl_2_usec(msr);
1243	if (usec) {
1244		bxt_cstates[5].exit_latency = usec;
1245		bxt_cstates[5].target_residency = usec;
1246	}
 
 
 
 
 
 
 
 
 
1247
1248	rdmsrl(MSR_PKGC10_IRTL, msr);
1249	usec = irtl_2_usec(msr);
1250	if (usec) {
1251		bxt_cstates[6].exit_latency = usec;
1252		bxt_cstates[6].target_residency = usec;
1253	}
1254
1255}
1256/*
1257 * sklh_idle_state_table_update(void)
1258 *
1259 * On SKL-H (model 0x5e) disable C8 and C9 if:
1260 * C10 is enabled and SGX disabled
1261 */
1262static void sklh_idle_state_table_update(void)
1263{
1264	unsigned long long msr;
1265	unsigned int eax, ebx, ecx, edx;
1266
1267
1268	/* if PC10 disabled via cmdline intel_idle.max_cstate=7 or shallower */
1269	if (max_cstate <= 7)
1270		return;
1271
1272	/* if PC10 not present in CPUID.MWAIT.EDX */
1273	if ((mwait_substates & (0xF << 28)) == 0)
1274		return;
1275
1276	rdmsrl(MSR_NHM_SNB_PKG_CST_CFG_CTL, msr);
1277
1278	/* PC10 is not enabled in PKG C-state limit */
1279	if ((msr & 0xF) != 8)
1280		return;
1281
1282	ecx = 0;
1283	cpuid(7, &eax, &ebx, &ecx, &edx);
1284
1285	/* if SGX is present */
1286	if (ebx & (1 << 2)) {
1287
1288		rdmsrl(MSR_IA32_FEATURE_CONTROL, msr);
1289
1290		/* if SGX is enabled */
1291		if (msr & (1 << 18))
1292			return;
1293	}
1294
1295	skl_cstates[5].disabled = 1;	/* C8-SKL */
1296	skl_cstates[6].disabled = 1;	/* C9-SKL */
1297}
1298/*
1299 * intel_idle_state_table_update()
1300 *
1301 * Update the default state_table for this CPU-id
1302 */
1303
1304static void intel_idle_state_table_update(void)
1305{
1306	switch (boot_cpu_data.x86_model) {
1307
1308	case INTEL_FAM6_IVYBRIDGE_X:
1309		ivt_idle_state_table_update();
1310		break;
1311	case INTEL_FAM6_ATOM_GOLDMONT:
1312		bxt_idle_state_table_update();
1313		break;
1314	case INTEL_FAM6_SKYLAKE_DESKTOP:
1315		sklh_idle_state_table_update();
1316		break;
1317	}
1318}
1319
1320/*
1321 * intel_idle_cpuidle_driver_init()
1322 * allocate, initialize cpuidle_states
1323 */
1324static void __init intel_idle_cpuidle_driver_init(void)
1325{
1326	int cstate;
1327	struct cpuidle_driver *drv = &intel_idle_driver;
1328
1329	intel_idle_state_table_update();
1330
1331	drv->state_count = 1;
1332
1333	for (cstate = 0; cstate < CPUIDLE_STATE_MAX; ++cstate) {
1334		int num_substates, mwait_hint, mwait_cstate;
1335
1336		if ((cpuidle_state_table[cstate].enter == NULL) &&
1337		    (cpuidle_state_table[cstate].enter_freeze == NULL))
1338			break;
1339
1340		if (cstate + 1 > max_cstate) {
1341			printk(PREFIX "max_cstate %d reached\n",
1342				max_cstate);
1343			break;
1344		}
1345
1346		mwait_hint = flg2MWAIT(cpuidle_state_table[cstate].flags);
1347		mwait_cstate = MWAIT_HINT2CSTATE(mwait_hint);
1348
1349		/* number of sub-states for this state in CPUID.MWAIT */
1350		num_substates = (mwait_substates >> ((mwait_cstate + 1) * 4))
1351					& MWAIT_SUBSTATE_MASK;
1352
1353		/* if NO sub-states for this state in CPUID, skip it */
1354		if (num_substates == 0)
1355			continue;
1356
1357		/* if state marked as disabled, skip it */
1358		if (cpuidle_state_table[cstate].disabled != 0) {
1359			pr_debug(PREFIX "state %s is disabled",
1360				cpuidle_state_table[cstate].name);
1361			continue;
1362		}
1363
1364
1365		if (((mwait_cstate + 1) > 2) &&
1366			!boot_cpu_has(X86_FEATURE_NONSTOP_TSC))
1367			mark_tsc_unstable("TSC halts in idle"
1368					" states deeper than C2");
1369
1370		drv->states[drv->state_count] =	/* structure copy */
1371			cpuidle_state_table[cstate];
1372
1373		drv->state_count += 1;
1374	}
1375
1376	if (icpu->byt_auto_demotion_disable_flag) {
1377		wrmsrl(MSR_CC6_DEMOTION_POLICY_CONFIG, 0);
1378		wrmsrl(MSR_MC6_DEMOTION_POLICY_CONFIG, 0);
1379	}
1380}
1381
1382
1383/*
1384 * intel_idle_cpu_init()
1385 * allocate, initialize, register cpuidle_devices
1386 * @cpu: cpu/core to initialize
1387 */
1388static int intel_idle_cpu_init(unsigned int cpu)
1389{
1390	struct cpuidle_device *dev;
1391
1392	dev = per_cpu_ptr(intel_idle_cpuidle_devices, cpu);
1393	dev->cpu = cpu;
1394
1395	if (cpuidle_register_device(dev)) {
1396		pr_debug(PREFIX "cpuidle_register_device %d failed!\n", cpu);
1397		return -EIO;
1398	}
1399
1400	if (icpu->auto_demotion_disable_flags)
1401		auto_demotion_disable();
1402
1403	if (icpu->disable_promotion_to_c1e)
1404		c1e_promotion_disable();
1405
1406	return 0;
1407}
1408
1409static int intel_idle_cpu_online(unsigned int cpu)
1410{
1411	struct cpuidle_device *dev;
1412
1413	if (lapic_timer_reliable_states != LAPIC_TIMER_ALWAYS_RELIABLE)
1414		__setup_broadcast_timer(true);
1415
1416	/*
1417	 * Some systems can hotplug a cpu at runtime after
1418	 * the kernel has booted, we have to initialize the
1419	 * driver in this case
1420	 */
1421	dev = per_cpu_ptr(intel_idle_cpuidle_devices, cpu);
1422	if (!dev->registered)
1423		return intel_idle_cpu_init(cpu);
1424
1425	return 0;
1426}
1427
1428static int __init intel_idle_init(void)
1429{
1430	int retval;
1431
1432	/* Do not load intel_idle at all for now if idle= is passed */
1433	if (boot_option_idle_override != IDLE_NO_OVERRIDE)
1434		return -ENODEV;
1435
1436	retval = intel_idle_probe();
1437	if (retval)
1438		return retval;
1439
1440	intel_idle_cpuidle_devices = alloc_percpu(struct cpuidle_device);
1441	if (intel_idle_cpuidle_devices == NULL)
1442		return -ENOMEM;
1443
1444	intel_idle_cpuidle_driver_init();
1445	retval = cpuidle_register_driver(&intel_idle_driver);
1446	if (retval) {
1447		struct cpuidle_driver *drv = cpuidle_get_driver();
1448		printk(KERN_DEBUG PREFIX "intel_idle yielding to %s",
1449			drv ? drv->name : "none");
1450		goto init_driver_fail;
1451	}
1452
1453	if (boot_cpu_has(X86_FEATURE_ARAT))	/* Always Reliable APIC Timer */
1454		lapic_timer_reliable_states = LAPIC_TIMER_ALWAYS_RELIABLE;
1455
1456	retval = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "idle/intel:online",
1457				   intel_idle_cpu_online, NULL);
1458	if (retval < 0)
1459		goto hp_setup_fail;
1460
1461	pr_debug(PREFIX "lapic_timer_reliable_states 0x%x\n",
1462		lapic_timer_reliable_states);
1463
1464	return 0;
 
1465
1466hp_setup_fail:
 
1467	intel_idle_cpuidle_devices_uninit();
1468	cpuidle_unregister_driver(&intel_idle_driver);
1469init_driver_fail:
1470	free_percpu(intel_idle_cpuidle_devices);
1471	return retval;
1472
 
 
 
 
 
 
1473}
1474device_initcall(intel_idle_init);
1475
1476/*
1477 * We are not really modular, but we used to support that.  Meaning we also
1478 * support "intel_idle.max_cstate=..." at boot and also a read-only export of
1479 * it at /sys/module/intel_idle/parameters/max_cstate -- so using module_param
1480 * is the easiest way (currently) to continue doing that.
1481 */
1482module_param(max_cstate, int, 0444);