Linux Kernel  3.7.1
 All Data Structures Namespaces Files Functions Variables Typedefs Enumerations Enumerator Macros Groups Pages
intel_idle.c
Go to the documentation of this file.
1 /*
2  * intel_idle.c - native hardware idle loop for modern Intel processors
3  *
4  * Copyright (c) 2010, Intel Corporation.
5  * Len Brown <[email protected]>
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/cpu_device_id.h>
66 #include <asm/mwait.h>
67 #include <asm/msr.h>
68 
69 #define INTEL_IDLE_VERSION "0.4"
70 #define PREFIX "intel_idle: "
71 
72 static struct cpuidle_driver intel_idle_driver = {
73  .name = "intel_idle",
74  .owner = THIS_MODULE,
75 };
76 /* intel_idle.max_cstate=0 disables driver */
77 static int max_cstate = MWAIT_MAX_NUM_CSTATES - 1;
78 
79 static unsigned int mwait_substates;
80 
81 #define LAPIC_TIMER_ALWAYS_RELIABLE 0xFFFFFFFF
82 /* Reliable LAPIC Timer States, bit 1 for C1 etc. */
83 static unsigned int lapic_timer_reliable_states = (1 << 1); /* Default to only C1 */
84 
85 struct idle_cpu {
87 
88  /*
89  * Hardware C-state auto-demotion may not always be optimal.
90  * Indicate which enable bits to clear here.
91  */
93 };
94 
95 static const struct idle_cpu *icpu;
96 static struct cpuidle_device __percpu *intel_idle_cpuidle_devices;
97 static int intel_idle(struct cpuidle_device *dev,
98  struct cpuidle_driver *drv, int index);
99 static int intel_idle_cpu_init(int cpu);
100 
101 static 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  * States are indexed by the cstate number,
113  * which is also the index into the MWAIT hint array.
114  * Thus C0 is a dummy.
115  */
116 static struct cpuidle_state nehalem_cstates[MWAIT_MAX_NUM_CSTATES] = {
117  { /* MWAIT C0 */ },
118  { /* MWAIT C1 */
119  .name = "C1-NHM",
120  .desc = "MWAIT 0x00",
121  .flags = CPUIDLE_FLAG_TIME_VALID,
122  .exit_latency = 3,
123  .target_residency = 6,
124  .enter = &intel_idle },
125  { /* MWAIT C2 */
126  .name = "C3-NHM",
127  .desc = "MWAIT 0x10",
129  .exit_latency = 20,
130  .target_residency = 80,
131  .enter = &intel_idle },
132  { /* MWAIT C3 */
133  .name = "C6-NHM",
134  .desc = "MWAIT 0x20",
136  .exit_latency = 200,
137  .target_residency = 800,
138  .enter = &intel_idle },
139 };
140 
141 static struct cpuidle_state snb_cstates[MWAIT_MAX_NUM_CSTATES] = {
142  { /* MWAIT C0 */ },
143  { /* MWAIT C1 */
144  .name = "C1-SNB",
145  .desc = "MWAIT 0x00",
146  .flags = CPUIDLE_FLAG_TIME_VALID,
147  .exit_latency = 1,
148  .target_residency = 1,
149  .enter = &intel_idle },
150  { /* MWAIT C2 */
151  .name = "C3-SNB",
152  .desc = "MWAIT 0x10",
154  .exit_latency = 80,
155  .target_residency = 211,
156  .enter = &intel_idle },
157  { /* MWAIT C3 */
158  .name = "C6-SNB",
159  .desc = "MWAIT 0x20",
161  .exit_latency = 104,
162  .target_residency = 345,
163  .enter = &intel_idle },
164  { /* MWAIT C4 */
165  .name = "C7-SNB",
166  .desc = "MWAIT 0x30",
168  .exit_latency = 109,
169  .target_residency = 345,
170  .enter = &intel_idle },
171 };
172 
173 static struct cpuidle_state ivb_cstates[MWAIT_MAX_NUM_CSTATES] = {
174  { /* MWAIT C0 */ },
175  { /* MWAIT C1 */
176  .name = "C1-IVB",
177  .desc = "MWAIT 0x00",
178  .flags = CPUIDLE_FLAG_TIME_VALID,
179  .exit_latency = 1,
180  .target_residency = 1,
181  .enter = &intel_idle },
182  { /* MWAIT C2 */
183  .name = "C3-IVB",
184  .desc = "MWAIT 0x10",
186  .exit_latency = 59,
187  .target_residency = 156,
188  .enter = &intel_idle },
189  { /* MWAIT C3 */
190  .name = "C6-IVB",
191  .desc = "MWAIT 0x20",
193  .exit_latency = 80,
194  .target_residency = 300,
195  .enter = &intel_idle },
196  { /* MWAIT C4 */
197  .name = "C7-IVB",
198  .desc = "MWAIT 0x30",
200  .exit_latency = 87,
201  .target_residency = 300,
202  .enter = &intel_idle },
203 };
204 
205 static struct cpuidle_state atom_cstates[MWAIT_MAX_NUM_CSTATES] = {
206  { /* MWAIT C0 */ },
207  { /* MWAIT C1 */
208  .name = "C1-ATM",
209  .desc = "MWAIT 0x00",
210  .flags = CPUIDLE_FLAG_TIME_VALID,
211  .exit_latency = 1,
212  .target_residency = 4,
213  .enter = &intel_idle },
214  { /* MWAIT C2 */
215  .name = "C2-ATM",
216  .desc = "MWAIT 0x10",
217  .flags = CPUIDLE_FLAG_TIME_VALID,
218  .exit_latency = 20,
219  .target_residency = 80,
220  .enter = &intel_idle },
221  { /* MWAIT C3 */ },
222  { /* MWAIT C4 */
223  .name = "C4-ATM",
224  .desc = "MWAIT 0x30",
226  .exit_latency = 100,
227  .target_residency = 400,
228  .enter = &intel_idle },
229  { /* MWAIT C5 */ },
230  { /* MWAIT C6 */
231  .name = "C6-ATM",
232  .desc = "MWAIT 0x52",
234  .exit_latency = 140,
235  .target_residency = 560,
236  .enter = &intel_idle },
237 };
238 
239 static long get_driver_data(int cstate)
240 {
241  int driver_data;
242  switch (cstate) {
243 
244  case 1: /* MWAIT C1 */
245  driver_data = 0x00;
246  break;
247  case 2: /* MWAIT C2 */
248  driver_data = 0x10;
249  break;
250  case 3: /* MWAIT C3 */
251  driver_data = 0x20;
252  break;
253  case 4: /* MWAIT C4 */
254  driver_data = 0x30;
255  break;
256  case 5: /* MWAIT C5 */
257  driver_data = 0x40;
258  break;
259  case 6: /* MWAIT C6 */
260  driver_data = 0x52;
261  break;
262  default:
263  driver_data = 0x00;
264  }
265  return driver_data;
266 }
267 
276 static int intel_idle(struct cpuidle_device *dev,
277  struct cpuidle_driver *drv, int index)
278 {
279  unsigned long ecx = 1; /* break on interrupt flag */
280  struct cpuidle_state *state = &drv->states[index];
281  struct cpuidle_state_usage *state_usage = &dev->states_usage[index];
282  unsigned long eax = (unsigned long)cpuidle_get_statedata(state_usage);
283  unsigned int cstate;
284  ktime_t kt_before, kt_after;
285  s64 usec_delta;
286  int cpu = smp_processor_id();
287 
288  cstate = (((eax) >> MWAIT_SUBSTATE_SIZE) & MWAIT_CSTATE_MASK) + 1;
289 
290  /*
291  * leave_mm() to avoid costly and often unnecessary wakeups
292  * for flushing the user TLB's associated with the active mm.
293  */
294  if (state->flags & CPUIDLE_FLAG_TLB_FLUSHED)
295  leave_mm(cpu);
296 
297  if (!(lapic_timer_reliable_states & (1 << (cstate))))
298  clockevents_notify(CLOCK_EVT_NOTIFY_BROADCAST_ENTER, &cpu);
299 
300  kt_before = ktime_get_real();
301 
303  if (!need_resched()) {
304 
305  __monitor((void *)&current_thread_info()->flags, 0, 0);
306  smp_mb();
307  if (!need_resched())
308  __mwait(eax, ecx);
309  }
310 
312 
313  kt_after = ktime_get_real();
314  usec_delta = ktime_to_us(ktime_sub(kt_after, kt_before));
315 
317 
318  if (!(lapic_timer_reliable_states & (1 << (cstate))))
319  clockevents_notify(CLOCK_EVT_NOTIFY_BROADCAST_EXIT, &cpu);
320 
321  /* Update cpuidle counters */
322  dev->last_residency = (int)usec_delta;
323 
324  return index;
325 }
326 
327 static void __setup_broadcast_timer(void *arg)
328 {
329  unsigned long reason = (unsigned long)arg;
330  int cpu = smp_processor_id();
331 
332  reason = reason ?
333  CLOCK_EVT_NOTIFY_BROADCAST_ON : CLOCK_EVT_NOTIFY_BROADCAST_OFF;
334 
335  clockevents_notify(reason, &cpu);
336 }
337 
338 static int cpu_hotplug_notify(struct notifier_block *n,
339  unsigned long action, void *hcpu)
340 {
341  int hotcpu = (unsigned long)hcpu;
342  struct cpuidle_device *dev;
343 
344  switch (action & 0xf) {
345  case CPU_ONLINE:
346 
347  if (lapic_timer_reliable_states != LAPIC_TIMER_ALWAYS_RELIABLE)
348  smp_call_function_single(hotcpu, __setup_broadcast_timer,
349  (void *)true, 1);
350 
351  /*
352  * Some systems can hotplug a cpu at runtime after
353  * the kernel has booted, we have to initialize the
354  * driver in this case
355  */
356  dev = per_cpu_ptr(intel_idle_cpuidle_devices, hotcpu);
357  if (!dev->registered)
358  intel_idle_cpu_init(hotcpu);
359 
360  break;
361  }
362  return NOTIFY_OK;
363 }
364 
365 static struct notifier_block cpu_hotplug_notifier = {
366  .notifier_call = cpu_hotplug_notify,
367 };
368 
369 static void auto_demotion_disable(void *dummy)
370 {
371  unsigned long long msr_bits;
372 
373  rdmsrl(MSR_NHM_SNB_PKG_CST_CFG_CTL, msr_bits);
374  msr_bits &= ~(icpu->auto_demotion_disable_flags);
375  wrmsrl(MSR_NHM_SNB_PKG_CST_CFG_CTL, msr_bits);
376 }
377 
378 static const struct idle_cpu idle_cpu_nehalem = {
379  .state_table = nehalem_cstates,
380  .auto_demotion_disable_flags = NHM_C1_AUTO_DEMOTE | NHM_C3_AUTO_DEMOTE,
381 };
382 
383 static const struct idle_cpu idle_cpu_atom = {
384  .state_table = atom_cstates,
385 };
386 
387 static const struct idle_cpu idle_cpu_lincroft = {
388  .state_table = atom_cstates,
389  .auto_demotion_disable_flags = ATM_LNC_C6_AUTO_DEMOTE,
390 };
391 
392 static const struct idle_cpu idle_cpu_snb = {
393  .state_table = snb_cstates,
394 };
395 
396 static const struct idle_cpu idle_cpu_ivb = {
397  .state_table = ivb_cstates,
398 };
399 
400 #define ICPU(model, cpu) \
401  { X86_VENDOR_INTEL, 6, model, X86_FEATURE_MWAIT, (unsigned long)&cpu }
402 
403 static const struct x86_cpu_id intel_idle_ids[] = {
404  ICPU(0x1a, idle_cpu_nehalem),
405  ICPU(0x1e, idle_cpu_nehalem),
406  ICPU(0x1f, idle_cpu_nehalem),
407  ICPU(0x25, idle_cpu_nehalem),
408  ICPU(0x2c, idle_cpu_nehalem),
409  ICPU(0x2e, idle_cpu_nehalem),
410  ICPU(0x1c, idle_cpu_atom),
411  ICPU(0x26, idle_cpu_lincroft),
412  ICPU(0x2f, idle_cpu_nehalem),
413  ICPU(0x2a, idle_cpu_snb),
414  ICPU(0x2d, idle_cpu_snb),
415  ICPU(0x3a, idle_cpu_ivb),
416  ICPU(0x3e, idle_cpu_ivb),
417  {}
418 };
419 MODULE_DEVICE_TABLE(x86cpu, intel_idle_ids);
420 
421 /*
422  * intel_idle_probe()
423  */
424 static int intel_idle_probe(void)
425 {
426  unsigned int eax, ebx, ecx;
427  const struct x86_cpu_id *id;
428 
429  if (max_cstate == 0) {
430  pr_debug(PREFIX "disabled\n");
431  return -EPERM;
432  }
433 
434  id = x86_match_cpu(intel_idle_ids);
435  if (!id) {
436  if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL &&
437  boot_cpu_data.x86 == 6)
438  pr_debug(PREFIX "does not run on family %d model %d\n",
439  boot_cpu_data.x86, boot_cpu_data.x86_model);
440  return -ENODEV;
441  }
442 
443  if (boot_cpu_data.cpuid_level < CPUID_MWAIT_LEAF)
444  return -ENODEV;
445 
446  cpuid(CPUID_MWAIT_LEAF, &eax, &ebx, &ecx, &mwait_substates);
447 
448  if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED) ||
449  !(ecx & CPUID5_ECX_INTERRUPT_BREAK) ||
450  !mwait_substates)
451  return -ENODEV;
452 
453  pr_debug(PREFIX "MWAIT substates: 0x%x\n", mwait_substates);
454 
455  icpu = (const struct idle_cpu *)id->driver_data;
456  cpuidle_state_table = icpu->state_table;
457 
458  if (boot_cpu_has(X86_FEATURE_ARAT)) /* Always Reliable APIC Timer */
459  lapic_timer_reliable_states = LAPIC_TIMER_ALWAYS_RELIABLE;
460  else
461  on_each_cpu(__setup_broadcast_timer, (void *)true, 1);
462 
463  register_cpu_notifier(&cpu_hotplug_notifier);
464 
466  " model 0x%X\n", boot_cpu_data.x86_model);
467 
468  pr_debug(PREFIX "lapic_timer_reliable_states 0x%x\n",
469  lapic_timer_reliable_states);
470  return 0;
471 }
472 
473 /*
474  * intel_idle_cpuidle_devices_uninit()
475  * unregister, free cpuidle_devices
476  */
477 static void intel_idle_cpuidle_devices_uninit(void)
478 {
479  int i;
480  struct cpuidle_device *dev;
481 
483  dev = per_cpu_ptr(intel_idle_cpuidle_devices, i);
485  }
486 
487  free_percpu(intel_idle_cpuidle_devices);
488  return;
489 }
490 /*
491  * intel_idle_cpuidle_driver_init()
492  * allocate, initialize cpuidle_states
493  */
494 static int intel_idle_cpuidle_driver_init(void)
495 {
496  int cstate;
497  struct cpuidle_driver *drv = &intel_idle_driver;
498 
499  drv->state_count = 1;
500 
501  for (cstate = 1; cstate < MWAIT_MAX_NUM_CSTATES; ++cstate) {
502  int num_substates;
503 
504  if (cstate > max_cstate) {
505  printk(PREFIX "max_cstate %d reached\n",
506  max_cstate);
507  break;
508  }
509 
510  /* does the state exist in CPUID.MWAIT? */
511  num_substates = (mwait_substates >> ((cstate) * 4))
513  if (num_substates == 0)
514  continue;
515  /* is the state not enabled? */
516  if (cpuidle_state_table[cstate].enter == NULL) {
517  /* does the driver not know about the state? */
518  if (*cpuidle_state_table[cstate].name == '\0')
519  pr_debug(PREFIX "unaware of model 0x%x"
520  " MWAIT %d please"
521  " contact [email protected]",
522  boot_cpu_data.x86_model, cstate);
523  continue;
524  }
525 
526  if ((cstate > 2) &&
527  !boot_cpu_has(X86_FEATURE_NONSTOP_TSC))
528  mark_tsc_unstable("TSC halts in idle"
529  " states deeper than C2");
530 
531  drv->states[drv->state_count] = /* structure copy */
532  cpuidle_state_table[cstate];
533 
534  drv->state_count += 1;
535  }
536 
537  if (icpu->auto_demotion_disable_flags)
538  on_each_cpu(auto_demotion_disable, NULL, 1);
539 
540  return 0;
541 }
542 
543 
544 /*
545  * intel_idle_cpu_init()
546  * allocate, initialize, register cpuidle_devices
547  * @cpu: cpu/core to initialize
548  */
549 static int intel_idle_cpu_init(int cpu)
550 {
551  int cstate;
552  struct cpuidle_device *dev;
553 
554  dev = per_cpu_ptr(intel_idle_cpuidle_devices, cpu);
555 
556  dev->state_count = 1;
557 
558  for (cstate = 1; cstate < MWAIT_MAX_NUM_CSTATES; ++cstate) {
559  int num_substates;
560 
561  if (cstate > max_cstate) {
562  printk(PREFIX "max_cstate %d reached\n", max_cstate);
563  break;
564  }
565 
566  /* does the state exist in CPUID.MWAIT? */
567  num_substates = (mwait_substates >> ((cstate) * 4))
569  if (num_substates == 0)
570  continue;
571  /* is the state not enabled? */
572  if (cpuidle_state_table[cstate].enter == NULL)
573  continue;
574 
575  dev->states_usage[dev->state_count].driver_data =
576  (void *)get_driver_data(cstate);
577 
578  dev->state_count += 1;
579  }
580 
581  dev->cpu = cpu;
582 
583  if (cpuidle_register_device(dev)) {
584  pr_debug(PREFIX "cpuidle_register_device %d failed!\n", cpu);
585  intel_idle_cpuidle_devices_uninit();
586  return -EIO;
587  }
588 
589  if (icpu->auto_demotion_disable_flags)
590  smp_call_function_single(cpu, auto_demotion_disable, NULL, 1);
591 
592  return 0;
593 }
594 
595 static int __init intel_idle_init(void)
596 {
597  int retval, i;
598 
599  /* Do not load intel_idle at all for now if idle= is passed */
601  return -ENODEV;
602 
603  retval = intel_idle_probe();
604  if (retval)
605  return retval;
606 
607  intel_idle_cpuidle_driver_init();
608  retval = cpuidle_register_driver(&intel_idle_driver);
609  if (retval) {
610  struct cpuidle_driver *drv = cpuidle_get_driver();
611  printk(KERN_DEBUG PREFIX "intel_idle yielding to %s",
612  drv ? drv->name : "none");
613  return retval;
614  }
615 
616  intel_idle_cpuidle_devices = alloc_percpu(struct cpuidle_device);
617  if (intel_idle_cpuidle_devices == NULL)
618  return -ENOMEM;
619 
621  retval = intel_idle_cpu_init(i);
622  if (retval) {
623  cpuidle_unregister_driver(&intel_idle_driver);
624  return retval;
625  }
626  }
627 
628  return 0;
629 }
630 
631 static void __exit intel_idle_exit(void)
632 {
633  intel_idle_cpuidle_devices_uninit();
634  cpuidle_unregister_driver(&intel_idle_driver);
635 
636 
637  if (lapic_timer_reliable_states != LAPIC_TIMER_ALWAYS_RELIABLE)
638  on_each_cpu(__setup_broadcast_timer, (void *)false, 1);
639  unregister_cpu_notifier(&cpu_hotplug_notifier);
640 
641  return;
642 }
643 
644 module_init(intel_idle_init);
645 module_exit(intel_idle_exit);
646 
647 module_param(max_cstate, int, 0444);
648 
649 MODULE_AUTHOR("Len Brown <[email protected]>");
650 MODULE_DESCRIPTION("Cpuidle driver for Intel Hardware v" INTEL_IDLE_VERSION);
651 MODULE_LICENSE("GPL");