sched: document profile=sleep requiring CONFIG_SCHEDSTATS
[linux-2.6.git] / kernel / profile.c
1 /*
2  *  linux/kernel/profile.c
3  *  Simple profiling. Manages a direct-mapped profile hit count buffer,
4  *  with configurable resolution, support for restricting the cpus on
5  *  which profiling is done, and switching between cpu time and
6  *  schedule() calls via kernel command line parameters passed at boot.
7  *
8  *  Scheduler profiling support, Arjan van de Ven and Ingo Molnar,
9  *      Red Hat, July 2004
10  *  Consolidation of architecture support code for profiling,
11  *      William Irwin, Oracle, July 2004
12  *  Amortized hit count accounting via per-cpu open-addressed hashtables
13  *      to resolve timer interrupt livelocks, William Irwin, Oracle, 2004
14  */
15
16 #include <linux/module.h>
17 #include <linux/profile.h>
18 #include <linux/bootmem.h>
19 #include <linux/notifier.h>
20 #include <linux/mm.h>
21 #include <linux/cpumask.h>
22 #include <linux/cpu.h>
23 #include <linux/profile.h>
24 #include <linux/highmem.h>
25 #include <linux/mutex.h>
26 #include <asm/sections.h>
27 #include <asm/semaphore.h>
28 #include <asm/irq_regs.h>
29 #include <asm/ptrace.h>
30
31 struct profile_hit {
32         u32 pc, hits;
33 };
34 #define PROFILE_GRPSHIFT        3
35 #define PROFILE_GRPSZ           (1 << PROFILE_GRPSHIFT)
36 #define NR_PROFILE_HIT          (PAGE_SIZE/sizeof(struct profile_hit))
37 #define NR_PROFILE_GRP          (NR_PROFILE_HIT/PROFILE_GRPSZ)
38
39 /* Oprofile timer tick hook */
40 static int (*timer_hook)(struct pt_regs *) __read_mostly;
41
42 static atomic_t *prof_buffer;
43 static unsigned long prof_len, prof_shift;
44
45 int prof_on __read_mostly;
46 EXPORT_SYMBOL_GPL(prof_on);
47
48 static cpumask_t prof_cpu_mask = CPU_MASK_ALL;
49 #ifdef CONFIG_SMP
50 static DEFINE_PER_CPU(struct profile_hit *[2], cpu_profile_hits);
51 static DEFINE_PER_CPU(int, cpu_profile_flip);
52 static DEFINE_MUTEX(profile_flip_mutex);
53 #endif /* CONFIG_SMP */
54
55 static int __init profile_setup(char * str)
56 {
57         static char __initdata schedstr[] = "schedule";
58         static char __initdata sleepstr[] = "sleep";
59         static char __initdata kvmstr[] = "kvm";
60         int par;
61
62         if (!strncmp(str, sleepstr, strlen(sleepstr))) {
63 #ifdef CONFIG_SCHEDSTATS
64                 prof_on = SLEEP_PROFILING;
65                 if (str[strlen(sleepstr)] == ',')
66                         str += strlen(sleepstr) + 1;
67                 if (get_option(&str, &par))
68                         prof_shift = par;
69                 printk(KERN_INFO
70                         "kernel sleep profiling enabled (shift: %ld)\n",
71                         prof_shift);
72 #else
73                 printk(KERN_WARNING
74                         "kernel sleep profiling requires CONFIG_SCHEDSTATS\n");
75 #endif /* CONFIG_SCHEDSTATS */
76         } else if (!strncmp(str, schedstr, strlen(schedstr))) {
77                 prof_on = SCHED_PROFILING;
78                 if (str[strlen(schedstr)] == ',')
79                         str += strlen(schedstr) + 1;
80                 if (get_option(&str, &par))
81                         prof_shift = par;
82                 printk(KERN_INFO
83                         "kernel schedule profiling enabled (shift: %ld)\n",
84                         prof_shift);
85         } else if (!strncmp(str, kvmstr, strlen(kvmstr))) {
86                 prof_on = KVM_PROFILING;
87                 if (str[strlen(kvmstr)] == ',')
88                         str += strlen(kvmstr) + 1;
89                 if (get_option(&str, &par))
90                         prof_shift = par;
91                 printk(KERN_INFO
92                         "kernel KVM profiling enabled (shift: %ld)\n",
93                         prof_shift);
94         } else if (get_option(&str, &par)) {
95                 prof_shift = par;
96                 prof_on = CPU_PROFILING;
97                 printk(KERN_INFO "kernel profiling enabled (shift: %ld)\n",
98                         prof_shift);
99         }
100         return 1;
101 }
102 __setup("profile=", profile_setup);
103
104
105 void __init profile_init(void)
106 {
107         if (!prof_on) 
108                 return;
109  
110         /* only text is profiled */
111         prof_len = (_etext - _stext) >> prof_shift;
112         prof_buffer = alloc_bootmem(prof_len*sizeof(atomic_t));
113 }
114
115 /* Profile event notifications */
116  
117 #ifdef CONFIG_PROFILING
118  
119 static BLOCKING_NOTIFIER_HEAD(task_exit_notifier);
120 static ATOMIC_NOTIFIER_HEAD(task_free_notifier);
121 static BLOCKING_NOTIFIER_HEAD(munmap_notifier);
122  
123 void profile_task_exit(struct task_struct * task)
124 {
125         blocking_notifier_call_chain(&task_exit_notifier, 0, task);
126 }
127  
128 int profile_handoff_task(struct task_struct * task)
129 {
130         int ret;
131         ret = atomic_notifier_call_chain(&task_free_notifier, 0, task);
132         return (ret == NOTIFY_OK) ? 1 : 0;
133 }
134
135 void profile_munmap(unsigned long addr)
136 {
137         blocking_notifier_call_chain(&munmap_notifier, 0, (void *)addr);
138 }
139
140 int task_handoff_register(struct notifier_block * n)
141 {
142         return atomic_notifier_chain_register(&task_free_notifier, n);
143 }
144
145 int task_handoff_unregister(struct notifier_block * n)
146 {
147         return atomic_notifier_chain_unregister(&task_free_notifier, n);
148 }
149
150 int profile_event_register(enum profile_type type, struct notifier_block * n)
151 {
152         int err = -EINVAL;
153  
154         switch (type) {
155                 case PROFILE_TASK_EXIT:
156                         err = blocking_notifier_chain_register(
157                                         &task_exit_notifier, n);
158                         break;
159                 case PROFILE_MUNMAP:
160                         err = blocking_notifier_chain_register(
161                                         &munmap_notifier, n);
162                         break;
163         }
164  
165         return err;
166 }
167
168  
169 int profile_event_unregister(enum profile_type type, struct notifier_block * n)
170 {
171         int err = -EINVAL;
172  
173         switch (type) {
174                 case PROFILE_TASK_EXIT:
175                         err = blocking_notifier_chain_unregister(
176                                         &task_exit_notifier, n);
177                         break;
178                 case PROFILE_MUNMAP:
179                         err = blocking_notifier_chain_unregister(
180                                         &munmap_notifier, n);
181                         break;
182         }
183
184         return err;
185 }
186
187 int register_timer_hook(int (*hook)(struct pt_regs *))
188 {
189         if (timer_hook)
190                 return -EBUSY;
191         timer_hook = hook;
192         return 0;
193 }
194
195 void unregister_timer_hook(int (*hook)(struct pt_regs *))
196 {
197         WARN_ON(hook != timer_hook);
198         timer_hook = NULL;
199         /* make sure all CPUs see the NULL hook */
200         synchronize_sched();  /* Allow ongoing interrupts to complete. */
201 }
202
203 EXPORT_SYMBOL_GPL(register_timer_hook);
204 EXPORT_SYMBOL_GPL(unregister_timer_hook);
205 EXPORT_SYMBOL_GPL(task_handoff_register);
206 EXPORT_SYMBOL_GPL(task_handoff_unregister);
207 EXPORT_SYMBOL_GPL(profile_event_register);
208 EXPORT_SYMBOL_GPL(profile_event_unregister);
209
210 #endif /* CONFIG_PROFILING */
211
212
213 #ifdef CONFIG_SMP
214 /*
215  * Each cpu has a pair of open-addressed hashtables for pending
216  * profile hits. read_profile() IPI's all cpus to request them
217  * to flip buffers and flushes their contents to prof_buffer itself.
218  * Flip requests are serialized by the profile_flip_mutex. The sole
219  * use of having a second hashtable is for avoiding cacheline
220  * contention that would otherwise happen during flushes of pending
221  * profile hits required for the accuracy of reported profile hits
222  * and so resurrect the interrupt livelock issue.
223  *
224  * The open-addressed hashtables are indexed by profile buffer slot
225  * and hold the number of pending hits to that profile buffer slot on
226  * a cpu in an entry. When the hashtable overflows, all pending hits
227  * are accounted to their corresponding profile buffer slots with
228  * atomic_add() and the hashtable emptied. As numerous pending hits
229  * may be accounted to a profile buffer slot in a hashtable entry,
230  * this amortizes a number of atomic profile buffer increments likely
231  * to be far larger than the number of entries in the hashtable,
232  * particularly given that the number of distinct profile buffer
233  * positions to which hits are accounted during short intervals (e.g.
234  * several seconds) is usually very small. Exclusion from buffer
235  * flipping is provided by interrupt disablement (note that for
236  * SCHED_PROFILING or SLEEP_PROFILING profile_hit() may be called from
237  * process context).
238  * The hash function is meant to be lightweight as opposed to strong,
239  * and was vaguely inspired by ppc64 firmware-supported inverted
240  * pagetable hash functions, but uses a full hashtable full of finite
241  * collision chains, not just pairs of them.
242  *
243  * -- wli
244  */
245 static void __profile_flip_buffers(void *unused)
246 {
247         int cpu = smp_processor_id();
248
249         per_cpu(cpu_profile_flip, cpu) = !per_cpu(cpu_profile_flip, cpu);
250 }
251
252 static void profile_flip_buffers(void)
253 {
254         int i, j, cpu;
255
256         mutex_lock(&profile_flip_mutex);
257         j = per_cpu(cpu_profile_flip, get_cpu());
258         put_cpu();
259         on_each_cpu(__profile_flip_buffers, NULL, 0, 1);
260         for_each_online_cpu(cpu) {
261                 struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[j];
262                 for (i = 0; i < NR_PROFILE_HIT; ++i) {
263                         if (!hits[i].hits) {
264                                 if (hits[i].pc)
265                                         hits[i].pc = 0;
266                                 continue;
267                         }
268                         atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
269                         hits[i].hits = hits[i].pc = 0;
270                 }
271         }
272         mutex_unlock(&profile_flip_mutex);
273 }
274
275 static void profile_discard_flip_buffers(void)
276 {
277         int i, cpu;
278
279         mutex_lock(&profile_flip_mutex);
280         i = per_cpu(cpu_profile_flip, get_cpu());
281         put_cpu();
282         on_each_cpu(__profile_flip_buffers, NULL, 0, 1);
283         for_each_online_cpu(cpu) {
284                 struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[i];
285                 memset(hits, 0, NR_PROFILE_HIT*sizeof(struct profile_hit));
286         }
287         mutex_unlock(&profile_flip_mutex);
288 }
289
290 void profile_hits(int type, void *__pc, unsigned int nr_hits)
291 {
292         unsigned long primary, secondary, flags, pc = (unsigned long)__pc;
293         int i, j, cpu;
294         struct profile_hit *hits;
295
296         if (prof_on != type || !prof_buffer)
297                 return;
298         pc = min((pc - (unsigned long)_stext) >> prof_shift, prof_len - 1);
299         i = primary = (pc & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
300         secondary = (~(pc << 1) & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
301         cpu = get_cpu();
302         hits = per_cpu(cpu_profile_hits, cpu)[per_cpu(cpu_profile_flip, cpu)];
303         if (!hits) {
304                 put_cpu();
305                 return;
306         }
307         /*
308          * We buffer the global profiler buffer into a per-CPU
309          * queue and thus reduce the number of global (and possibly
310          * NUMA-alien) accesses. The write-queue is self-coalescing:
311          */
312         local_irq_save(flags);
313         do {
314                 for (j = 0; j < PROFILE_GRPSZ; ++j) {
315                         if (hits[i + j].pc == pc) {
316                                 hits[i + j].hits += nr_hits;
317                                 goto out;
318                         } else if (!hits[i + j].hits) {
319                                 hits[i + j].pc = pc;
320                                 hits[i + j].hits = nr_hits;
321                                 goto out;
322                         }
323                 }
324                 i = (i + secondary) & (NR_PROFILE_HIT - 1);
325         } while (i != primary);
326
327         /*
328          * Add the current hit(s) and flush the write-queue out
329          * to the global buffer:
330          */
331         atomic_add(nr_hits, &prof_buffer[pc]);
332         for (i = 0; i < NR_PROFILE_HIT; ++i) {
333                 atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
334                 hits[i].pc = hits[i].hits = 0;
335         }
336 out:
337         local_irq_restore(flags);
338         put_cpu();
339 }
340
341 static int __devinit profile_cpu_callback(struct notifier_block *info,
342                                         unsigned long action, void *__cpu)
343 {
344         int node, cpu = (unsigned long)__cpu;
345         struct page *page;
346
347         switch (action) {
348         case CPU_UP_PREPARE:
349         case CPU_UP_PREPARE_FROZEN:
350                 node = cpu_to_node(cpu);
351                 per_cpu(cpu_profile_flip, cpu) = 0;
352                 if (!per_cpu(cpu_profile_hits, cpu)[1]) {
353                         page = alloc_pages_node(node,
354                                         GFP_KERNEL | __GFP_ZERO,
355                                         0);
356                         if (!page)
357                                 return NOTIFY_BAD;
358                         per_cpu(cpu_profile_hits, cpu)[1] = page_address(page);
359                 }
360                 if (!per_cpu(cpu_profile_hits, cpu)[0]) {
361                         page = alloc_pages_node(node,
362                                         GFP_KERNEL | __GFP_ZERO,
363                                         0);
364                         if (!page)
365                                 goto out_free;
366                         per_cpu(cpu_profile_hits, cpu)[0] = page_address(page);
367                 }
368                 break;
369         out_free:
370                 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
371                 per_cpu(cpu_profile_hits, cpu)[1] = NULL;
372                 __free_page(page);
373                 return NOTIFY_BAD;
374         case CPU_ONLINE:
375         case CPU_ONLINE_FROZEN:
376                 cpu_set(cpu, prof_cpu_mask);
377                 break;
378         case CPU_UP_CANCELED:
379         case CPU_UP_CANCELED_FROZEN:
380         case CPU_DEAD:
381         case CPU_DEAD_FROZEN:
382                 cpu_clear(cpu, prof_cpu_mask);
383                 if (per_cpu(cpu_profile_hits, cpu)[0]) {
384                         page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[0]);
385                         per_cpu(cpu_profile_hits, cpu)[0] = NULL;
386                         __free_page(page);
387                 }
388                 if (per_cpu(cpu_profile_hits, cpu)[1]) {
389                         page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
390                         per_cpu(cpu_profile_hits, cpu)[1] = NULL;
391                         __free_page(page);
392                 }
393                 break;
394         }
395         return NOTIFY_OK;
396 }
397 #else /* !CONFIG_SMP */
398 #define profile_flip_buffers()          do { } while (0)
399 #define profile_discard_flip_buffers()  do { } while (0)
400 #define profile_cpu_callback            NULL
401
402 void profile_hits(int type, void *__pc, unsigned int nr_hits)
403 {
404         unsigned long pc;
405
406         if (prof_on != type || !prof_buffer)
407                 return;
408         pc = ((unsigned long)__pc - (unsigned long)_stext) >> prof_shift;
409         atomic_add(nr_hits, &prof_buffer[min(pc, prof_len - 1)]);
410 }
411 #endif /* !CONFIG_SMP */
412
413 EXPORT_SYMBOL_GPL(profile_hits);
414
415 void profile_tick(int type)
416 {
417         struct pt_regs *regs = get_irq_regs();
418
419         if (type == CPU_PROFILING && timer_hook)
420                 timer_hook(regs);
421         if (!user_mode(regs) && cpu_isset(smp_processor_id(), prof_cpu_mask))
422                 profile_hit(type, (void *)profile_pc(regs));
423 }
424
425 #ifdef CONFIG_PROC_FS
426 #include <linux/proc_fs.h>
427 #include <asm/uaccess.h>
428 #include <asm/ptrace.h>
429
430 static int prof_cpu_mask_read_proc (char *page, char **start, off_t off,
431                         int count, int *eof, void *data)
432 {
433         int len = cpumask_scnprintf(page, count, *(cpumask_t *)data);
434         if (count - len < 2)
435                 return -EINVAL;
436         len += sprintf(page + len, "\n");
437         return len;
438 }
439
440 static int prof_cpu_mask_write_proc (struct file *file, const char __user *buffer,
441                                         unsigned long count, void *data)
442 {
443         cpumask_t *mask = (cpumask_t *)data;
444         unsigned long full_count = count, err;
445         cpumask_t new_value;
446
447         err = cpumask_parse_user(buffer, count, new_value);
448         if (err)
449                 return err;
450
451         *mask = new_value;
452         return full_count;
453 }
454
455 void create_prof_cpu_mask(struct proc_dir_entry *root_irq_dir)
456 {
457         struct proc_dir_entry *entry;
458
459         /* create /proc/irq/prof_cpu_mask */
460         if (!(entry = create_proc_entry("prof_cpu_mask", 0600, root_irq_dir)))
461                 return;
462         entry->data = (void *)&prof_cpu_mask;
463         entry->read_proc = prof_cpu_mask_read_proc;
464         entry->write_proc = prof_cpu_mask_write_proc;
465 }
466
467 /*
468  * This function accesses profiling information. The returned data is
469  * binary: the sampling step and the actual contents of the profile
470  * buffer. Use of the program readprofile is recommended in order to
471  * get meaningful info out of these data.
472  */
473 static ssize_t
474 read_profile(struct file *file, char __user *buf, size_t count, loff_t *ppos)
475 {
476         unsigned long p = *ppos;
477         ssize_t read;
478         char * pnt;
479         unsigned int sample_step = 1 << prof_shift;
480
481         profile_flip_buffers();
482         if (p >= (prof_len+1)*sizeof(unsigned int))
483                 return 0;
484         if (count > (prof_len+1)*sizeof(unsigned int) - p)
485                 count = (prof_len+1)*sizeof(unsigned int) - p;
486         read = 0;
487
488         while (p < sizeof(unsigned int) && count > 0) {
489                 if (put_user(*((char *)(&sample_step)+p),buf))
490                         return -EFAULT;
491                 buf++; p++; count--; read++;
492         }
493         pnt = (char *)prof_buffer + p - sizeof(atomic_t);
494         if (copy_to_user(buf,(void *)pnt,count))
495                 return -EFAULT;
496         read += count;
497         *ppos += read;
498         return read;
499 }
500
501 /*
502  * Writing to /proc/profile resets the counters
503  *
504  * Writing a 'profiling multiplier' value into it also re-sets the profiling
505  * interrupt frequency, on architectures that support this.
506  */
507 static ssize_t write_profile(struct file *file, const char __user *buf,
508                              size_t count, loff_t *ppos)
509 {
510 #ifdef CONFIG_SMP
511         extern int setup_profiling_timer (unsigned int multiplier);
512
513         if (count == sizeof(int)) {
514                 unsigned int multiplier;
515
516                 if (copy_from_user(&multiplier, buf, sizeof(int)))
517                         return -EFAULT;
518
519                 if (setup_profiling_timer(multiplier))
520                         return -EINVAL;
521         }
522 #endif
523         profile_discard_flip_buffers();
524         memset(prof_buffer, 0, prof_len * sizeof(atomic_t));
525         return count;
526 }
527
528 static const struct file_operations proc_profile_operations = {
529         .read           = read_profile,
530         .write          = write_profile,
531 };
532
533 #ifdef CONFIG_SMP
534 static void __init profile_nop(void *unused)
535 {
536 }
537
538 static int __init create_hash_tables(void)
539 {
540         int cpu;
541
542         for_each_online_cpu(cpu) {
543                 int node = cpu_to_node(cpu);
544                 struct page *page;
545
546                 page = alloc_pages_node(node,
547                                 GFP_KERNEL | __GFP_ZERO | GFP_THISNODE,
548                                 0);
549                 if (!page)
550                         goto out_cleanup;
551                 per_cpu(cpu_profile_hits, cpu)[1]
552                                 = (struct profile_hit *)page_address(page);
553                 page = alloc_pages_node(node,
554                                 GFP_KERNEL | __GFP_ZERO | GFP_THISNODE,
555                                 0);
556                 if (!page)
557                         goto out_cleanup;
558                 per_cpu(cpu_profile_hits, cpu)[0]
559                                 = (struct profile_hit *)page_address(page);
560         }
561         return 0;
562 out_cleanup:
563         prof_on = 0;
564         smp_mb();
565         on_each_cpu(profile_nop, NULL, 0, 1);
566         for_each_online_cpu(cpu) {
567                 struct page *page;
568
569                 if (per_cpu(cpu_profile_hits, cpu)[0]) {
570                         page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[0]);
571                         per_cpu(cpu_profile_hits, cpu)[0] = NULL;
572                         __free_page(page);
573                 }
574                 if (per_cpu(cpu_profile_hits, cpu)[1]) {
575                         page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
576                         per_cpu(cpu_profile_hits, cpu)[1] = NULL;
577                         __free_page(page);
578                 }
579         }
580         return -1;
581 }
582 #else
583 #define create_hash_tables()                    ({ 0; })
584 #endif
585
586 static int __init create_proc_profile(void)
587 {
588         struct proc_dir_entry *entry;
589
590         if (!prof_on)
591                 return 0;
592         if (create_hash_tables())
593                 return -1;
594         if (!(entry = create_proc_entry("profile", S_IWUSR | S_IRUGO, NULL)))
595                 return 0;
596         entry->proc_fops = &proc_profile_operations;
597         entry->size = (1+prof_len) * sizeof(atomic_t);
598         hotcpu_notifier(profile_cpu_callback, 0);
599         return 0;
600 }
601 module_init(create_proc_profile);
602 #endif /* CONFIG_PROC_FS */