Merge branches 'tracing/ftrace', 'tracing/function-graph-tracer' and 'tracing/urgent...
[pandora-kernel.git] / arch / x86 / kernel / process.c
1 #include <linux/errno.h>
2 #include <linux/kernel.h>
3 #include <linux/mm.h>
4 #include <linux/smp.h>
5 #include <linux/slab.h>
6 #include <linux/sched.h>
7 #include <linux/module.h>
8 #include <linux/pm.h>
9 #include <linux/clockchips.h>
10 #include <linux/ftrace.h>
11 #include <asm/system.h>
12
13 unsigned long idle_halt;
14 EXPORT_SYMBOL(idle_halt);
15 unsigned long idle_nomwait;
16 EXPORT_SYMBOL(idle_nomwait);
17
18 struct kmem_cache *task_xstate_cachep;
19
20 int arch_dup_task_struct(struct task_struct *dst, struct task_struct *src)
21 {
22         *dst = *src;
23         if (src->thread.xstate) {
24                 dst->thread.xstate = kmem_cache_alloc(task_xstate_cachep,
25                                                       GFP_KERNEL);
26                 if (!dst->thread.xstate)
27                         return -ENOMEM;
28                 WARN_ON((unsigned long)dst->thread.xstate & 15);
29                 memcpy(dst->thread.xstate, src->thread.xstate, xstate_size);
30         }
31         return 0;
32 }
33
34 void free_thread_xstate(struct task_struct *tsk)
35 {
36         if (tsk->thread.xstate) {
37                 kmem_cache_free(task_xstate_cachep, tsk->thread.xstate);
38                 tsk->thread.xstate = NULL;
39         }
40 }
41
42 void free_thread_info(struct thread_info *ti)
43 {
44         free_thread_xstate(ti->task);
45         free_pages((unsigned long)ti, get_order(THREAD_SIZE));
46 }
47
48 void arch_task_cache_init(void)
49 {
50         task_xstate_cachep =
51                 kmem_cache_create("task_xstate", xstate_size,
52                                   __alignof__(union thread_xstate),
53                                   SLAB_PANIC, NULL);
54 }
55
56 /*
57  * Idle related variables and functions
58  */
59 unsigned long boot_option_idle_override = 0;
60 EXPORT_SYMBOL(boot_option_idle_override);
61
62 /*
63  * Powermanagement idle function, if any..
64  */
65 void (*pm_idle)(void);
66 EXPORT_SYMBOL(pm_idle);
67
68 #ifdef CONFIG_X86_32
69 /*
70  * This halt magic was a workaround for ancient floppy DMA
71  * wreckage. It should be safe to remove.
72  */
73 static int hlt_counter;
74 void disable_hlt(void)
75 {
76         hlt_counter++;
77 }
78 EXPORT_SYMBOL(disable_hlt);
79
80 void enable_hlt(void)
81 {
82         hlt_counter--;
83 }
84 EXPORT_SYMBOL(enable_hlt);
85
86 static inline int hlt_use_halt(void)
87 {
88         return (!hlt_counter && boot_cpu_data.hlt_works_ok);
89 }
90 #else
91 static inline int hlt_use_halt(void)
92 {
93         return 1;
94 }
95 #endif
96
97 /*
98  * We use this if we don't have any better
99  * idle routine..
100  */
101 void default_idle(void)
102 {
103         if (hlt_use_halt()) {
104                 struct power_trace it;
105
106                 trace_power_start(&it, POWER_CSTATE, 1);
107                 current_thread_info()->status &= ~TS_POLLING;
108                 /*
109                  * TS_POLLING-cleared state must be visible before we
110                  * test NEED_RESCHED:
111                  */
112                 smp_mb();
113
114                 if (!need_resched())
115                         safe_halt();    /* enables interrupts racelessly */
116                 else
117                         local_irq_enable();
118                 current_thread_info()->status |= TS_POLLING;
119                 trace_power_end(&it);
120         } else {
121                 local_irq_enable();
122                 /* loop is done by the caller */
123                 cpu_relax();
124         }
125 }
126 #ifdef CONFIG_APM_MODULE
127 EXPORT_SYMBOL(default_idle);
128 #endif
129
130 static void do_nothing(void *unused)
131 {
132 }
133
134 /*
135  * cpu_idle_wait - Used to ensure that all the CPUs discard old value of
136  * pm_idle and update to new pm_idle value. Required while changing pm_idle
137  * handler on SMP systems.
138  *
139  * Caller must have changed pm_idle to the new value before the call. Old
140  * pm_idle value will not be used by any CPU after the return of this function.
141  */
142 void cpu_idle_wait(void)
143 {
144         smp_mb();
145         /* kick all the CPUs so that they exit out of pm_idle */
146         smp_call_function(do_nothing, NULL, 1);
147 }
148 EXPORT_SYMBOL_GPL(cpu_idle_wait);
149
150 /*
151  * This uses new MONITOR/MWAIT instructions on P4 processors with PNI,
152  * which can obviate IPI to trigger checking of need_resched.
153  * We execute MONITOR against need_resched and enter optimized wait state
154  * through MWAIT. Whenever someone changes need_resched, we would be woken
155  * up from MWAIT (without an IPI).
156  *
157  * New with Core Duo processors, MWAIT can take some hints based on CPU
158  * capability.
159  */
160 void mwait_idle_with_hints(unsigned long ax, unsigned long cx)
161 {
162         struct power_trace it;
163
164         trace_power_start(&it, POWER_CSTATE, (ax>>4)+1);
165         if (!need_resched()) {
166                 __monitor((void *)&current_thread_info()->flags, 0, 0);
167                 smp_mb();
168                 if (!need_resched())
169                         __mwait(ax, cx);
170         }
171         trace_power_end(&it);
172 }
173
174 /* Default MONITOR/MWAIT with no hints, used for default C1 state */
175 static void mwait_idle(void)
176 {
177         struct power_trace it;
178         if (!need_resched()) {
179                 trace_power_start(&it, POWER_CSTATE, 1);
180                 __monitor((void *)&current_thread_info()->flags, 0, 0);
181                 smp_mb();
182                 if (!need_resched())
183                         __sti_mwait(0, 0);
184                 else
185                         local_irq_enable();
186                 trace_power_end(&it);
187         } else
188                 local_irq_enable();
189 }
190
191 /*
192  * On SMP it's slightly faster (but much more power-consuming!)
193  * to poll the ->work.need_resched flag instead of waiting for the
194  * cross-CPU IPI to arrive. Use this option with caution.
195  */
196 static void poll_idle(void)
197 {
198         struct power_trace it;
199
200         trace_power_start(&it, POWER_CSTATE, 0);
201         local_irq_enable();
202         while (!need_resched())
203                 cpu_relax();
204         trace_power_end(&it);
205 }
206
207 /*
208  * mwait selection logic:
209  *
210  * It depends on the CPU. For AMD CPUs that support MWAIT this is
211  * wrong. Family 0x10 and 0x11 CPUs will enter C1 on HLT. Powersavings
212  * then depend on a clock divisor and current Pstate of the core. If
213  * all cores of a processor are in halt state (C1) the processor can
214  * enter the C1E (C1 enhanced) state. If mwait is used this will never
215  * happen.
216  *
217  * idle=mwait overrides this decision and forces the usage of mwait.
218  */
219 static int __cpuinitdata force_mwait;
220
221 #define MWAIT_INFO                      0x05
222 #define MWAIT_ECX_EXTENDED_INFO         0x01
223 #define MWAIT_EDX_C1                    0xf0
224
225 static int __cpuinit mwait_usable(const struct cpuinfo_x86 *c)
226 {
227         u32 eax, ebx, ecx, edx;
228
229         if (force_mwait)
230                 return 1;
231
232         if (c->cpuid_level < MWAIT_INFO)
233                 return 0;
234
235         cpuid(MWAIT_INFO, &eax, &ebx, &ecx, &edx);
236         /* Check, whether EDX has extended info about MWAIT */
237         if (!(ecx & MWAIT_ECX_EXTENDED_INFO))
238                 return 1;
239
240         /*
241          * edx enumeratios MONITOR/MWAIT extensions. Check, whether
242          * C1  supports MWAIT
243          */
244         return (edx & MWAIT_EDX_C1);
245 }
246
247 /*
248  * Check for AMD CPUs, which have potentially C1E support
249  */
250 static int __cpuinit check_c1e_idle(const struct cpuinfo_x86 *c)
251 {
252         if (c->x86_vendor != X86_VENDOR_AMD)
253                 return 0;
254
255         if (c->x86 < 0x0F)
256                 return 0;
257
258         /* Family 0x0f models < rev F do not have C1E */
259         if (c->x86 == 0x0f && c->x86_model < 0x40)
260                 return 0;
261
262         return 1;
263 }
264
265 static cpumask_t c1e_mask = CPU_MASK_NONE;
266 static int c1e_detected;
267
268 void c1e_remove_cpu(int cpu)
269 {
270         cpu_clear(cpu, c1e_mask);
271 }
272
273 /*
274  * C1E aware idle routine. We check for C1E active in the interrupt
275  * pending message MSR. If we detect C1E, then we handle it the same
276  * way as C3 power states (local apic timer and TSC stop)
277  */
278 static void c1e_idle(void)
279 {
280         if (need_resched())
281                 return;
282
283         if (!c1e_detected) {
284                 u32 lo, hi;
285
286                 rdmsr(MSR_K8_INT_PENDING_MSG, lo, hi);
287                 if (lo & K8_INTP_C1E_ACTIVE_MASK) {
288                         c1e_detected = 1;
289                         if (!boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
290                                 mark_tsc_unstable("TSC halt in AMD C1E");
291                         printk(KERN_INFO "System has AMD C1E enabled\n");
292                         set_cpu_cap(&boot_cpu_data, X86_FEATURE_AMDC1E);
293                 }
294         }
295
296         if (c1e_detected) {
297                 int cpu = smp_processor_id();
298
299                 if (!cpu_isset(cpu, c1e_mask)) {
300                         cpu_set(cpu, c1e_mask);
301                         /*
302                          * Force broadcast so ACPI can not interfere. Needs
303                          * to run with interrupts enabled as it uses
304                          * smp_function_call.
305                          */
306                         local_irq_enable();
307                         clockevents_notify(CLOCK_EVT_NOTIFY_BROADCAST_FORCE,
308                                            &cpu);
309                         printk(KERN_INFO "Switch to broadcast mode on CPU%d\n",
310                                cpu);
311                         local_irq_disable();
312                 }
313                 clockevents_notify(CLOCK_EVT_NOTIFY_BROADCAST_ENTER, &cpu);
314
315                 default_idle();
316
317                 /*
318                  * The switch back from broadcast mode needs to be
319                  * called with interrupts disabled.
320                  */
321                  local_irq_disable();
322                  clockevents_notify(CLOCK_EVT_NOTIFY_BROADCAST_EXIT, &cpu);
323                  local_irq_enable();
324         } else
325                 default_idle();
326 }
327
328 void __cpuinit select_idle_routine(const struct cpuinfo_x86 *c)
329 {
330 #ifdef CONFIG_X86_SMP
331         if (pm_idle == poll_idle && smp_num_siblings > 1) {
332                 printk(KERN_WARNING "WARNING: polling idle and HT enabled,"
333                         " performance may degrade.\n");
334         }
335 #endif
336         if (pm_idle)
337                 return;
338
339         if (cpu_has(c, X86_FEATURE_MWAIT) && mwait_usable(c)) {
340                 /*
341                  * One CPU supports mwait => All CPUs supports mwait
342                  */
343                 printk(KERN_INFO "using mwait in idle threads.\n");
344                 pm_idle = mwait_idle;
345         } else if (check_c1e_idle(c)) {
346                 printk(KERN_INFO "using C1E aware idle routine\n");
347                 pm_idle = c1e_idle;
348         } else
349                 pm_idle = default_idle;
350 }
351
352 static int __init idle_setup(char *str)
353 {
354         if (!str)
355                 return -EINVAL;
356
357         if (!strcmp(str, "poll")) {
358                 printk("using polling idle threads.\n");
359                 pm_idle = poll_idle;
360         } else if (!strcmp(str, "mwait"))
361                 force_mwait = 1;
362         else if (!strcmp(str, "halt")) {
363                 /*
364                  * When the boot option of idle=halt is added, halt is
365                  * forced to be used for CPU idle. In such case CPU C2/C3
366                  * won't be used again.
367                  * To continue to load the CPU idle driver, don't touch
368                  * the boot_option_idle_override.
369                  */
370                 pm_idle = default_idle;
371                 idle_halt = 1;
372                 return 0;
373         } else if (!strcmp(str, "nomwait")) {
374                 /*
375                  * If the boot option of "idle=nomwait" is added,
376                  * it means that mwait will be disabled for CPU C2/C3
377                  * states. In such case it won't touch the variable
378                  * of boot_option_idle_override.
379                  */
380                 idle_nomwait = 1;
381                 return 0;
382         } else
383                 return -1;
384
385         boot_option_idle_override = 1;
386         return 0;
387 }
388 early_param("idle", idle_setup);
389