3b7a1b055122a698c0f97a05eae8cbe6a229bb6f
[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/highmem.h>
24 #include <linux/mutex.h>
25 #include <asm/sections.h>
26 #include <asm/semaphore.h>
27 #include <asm/irq_regs.h>
28 #include <asm/ptrace.h>
29
30 struct profile_hit {
31         u32 pc, hits;
32 };
33 #define PROFILE_GRPSHIFT        3
34 #define PROFILE_GRPSZ           (1 << PROFILE_GRPSHIFT)
35 #define NR_PROFILE_HIT          (PAGE_SIZE/sizeof(struct profile_hit))
36 #define NR_PROFILE_GRP          (NR_PROFILE_HIT/PROFILE_GRPSZ)
37
38 /* Oprofile timer tick hook */
39 static int (*timer_hook)(struct pt_regs *) __read_mostly;
40
41 static atomic_t *prof_buffer;
42 static unsigned long prof_len, prof_shift;
43
44 int prof_on __read_mostly;
45 EXPORT_SYMBOL_GPL(prof_on);
46
47 static cpumask_t prof_cpu_mask = CPU_MASK_ALL;
48 #ifdef CONFIG_SMP
49 static DEFINE_PER_CPU(struct profile_hit *[2], cpu_profile_hits);
50 static DEFINE_PER_CPU(int, cpu_profile_flip);
51 static DEFINE_MUTEX(profile_flip_mutex);
52 #endif /* CONFIG_SMP */
53
54 static int __init profile_setup(char *str)
55 {
56         static char __initdata schedstr[] = "schedule";
57         static char __initdata sleepstr[] = "sleep";
58         static char __initdata kvmstr[] = "kvm";
59         int par;
60
61         if (!strncmp(str, sleepstr, strlen(sleepstr))) {
62 #ifdef CONFIG_SCHEDSTATS
63                 prof_on = SLEEP_PROFILING;
64                 if (str[strlen(sleepstr)] == ',')
65                         str += strlen(sleepstr) + 1;
66                 if (get_option(&str, &par))
67                         prof_shift = par;
68                 printk(KERN_INFO
69                         "kernel sleep profiling enabled (shift: %ld)\n",
70                         prof_shift);
71 #else
72                 printk(KERN_WARNING
73                         "kernel sleep profiling requires CONFIG_SCHEDSTATS\n");
74 #endif /* CONFIG_SCHEDSTATS */
75         } else if (!strncmp(str, schedstr, strlen(schedstr))) {
76                 prof_on = SCHED_PROFILING;
77                 if (str[strlen(schedstr)] == ',')
78                         str += strlen(schedstr) + 1;
79                 if (get_option(&str, &par))
80                         prof_shift = par;
81                 printk(KERN_INFO
82                         "kernel schedule profiling enabled (shift: %ld)\n",
83                         prof_shift);
84         } else if (!strncmp(str, kvmstr, strlen(kvmstr))) {
85                 prof_on = KVM_PROFILING;
86                 if (str[strlen(kvmstr)] == ',')
87                         str += strlen(kvmstr) + 1;
88                 if (get_option(&str, &par))
89                         prof_shift = par;
90                 printk(KERN_INFO
91                         "kernel KVM profiling enabled (shift: %ld)\n",
92                         prof_shift);
93         } else if (get_option(&str, &par)) {
94                 prof_shift = par;
95                 prof_on = CPU_PROFILING;
96                 printk(KERN_INFO "kernel profiling enabled (shift: %ld)\n",
97                         prof_shift);
98         }
99         return 1;
100 }
101 __setup("profile=", profile_setup);
102
103
104 void __init profile_init(void)
105 {
106         if (!prof_on)
107                 return;
108
109         /* only text is profiled */
110         prof_len = (_etext - _stext) >> prof_shift;
111         prof_buffer = alloc_bootmem(prof_len*sizeof(atomic_t));
112 }
113
114 /* Profile event notifications */
115
116 #ifdef CONFIG_PROFILING
117
118 static BLOCKING_NOTIFIER_HEAD(task_exit_notifier);
119 static ATOMIC_NOTIFIER_HEAD(task_free_notifier);
120 static BLOCKING_NOTIFIER_HEAD(munmap_notifier);
121
122 void profile_task_exit(struct task_struct *task)
123 {
124         blocking_notifier_call_chain(&task_exit_notifier, 0, task);
125 }
126
127 int profile_handoff_task(struct task_struct *task)
128 {
129         int ret;
130         ret = atomic_notifier_call_chain(&task_free_notifier, 0, task);
131         return (ret == NOTIFY_OK) ? 1 : 0;
132 }
133
134 void profile_munmap(unsigned long addr)
135 {
136         blocking_notifier_call_chain(&munmap_notifier, 0, (void *)addr);
137 }
138
139 int task_handoff_register(struct notifier_block *n)
140 {
141         return atomic_notifier_chain_register(&task_free_notifier, n);
142 }
143 EXPORT_SYMBOL_GPL(task_handoff_register);
144
145 int task_handoff_unregister(struct notifier_block *n)
146 {
147         return atomic_notifier_chain_unregister(&task_free_notifier, n);
148 }
149 EXPORT_SYMBOL_GPL(task_handoff_unregister);
150
151 int profile_event_register(enum profile_type type, struct notifier_block *n)
152 {
153         int err = -EINVAL;
154
155         switch (type) {
156         case PROFILE_TASK_EXIT:
157                 err = blocking_notifier_chain_register(
158                                 &task_exit_notifier, n);
159                 break;
160         case PROFILE_MUNMAP:
161                 err = blocking_notifier_chain_register(
162                                 &munmap_notifier, n);
163                 break;
164         }
165
166         return err;
167 }
168 EXPORT_SYMBOL_GPL(profile_event_register);
169
170 int profile_event_unregister(enum profile_type type, struct notifier_block *n)
171 {
172         int err = -EINVAL;
173
174         switch (type) {
175         case PROFILE_TASK_EXIT:
176                 err = blocking_notifier_chain_unregister(
177                                 &task_exit_notifier, n);
178                 break;
179         case PROFILE_MUNMAP:
180                 err = blocking_notifier_chain_unregister(
181                                 &munmap_notifier, n);
182                 break;
183         }
184
185         return err;
186 }
187 EXPORT_SYMBOL_GPL(profile_event_unregister);
188
189 int register_timer_hook(int (*hook)(struct pt_regs *))
190 {
191         if (timer_hook)
192                 return -EBUSY;
193         timer_hook = hook;
194         return 0;
195 }
196 EXPORT_SYMBOL_GPL(register_timer_hook);
197
198 void unregister_timer_hook(int (*hook)(struct pt_regs *))
199 {
200         WARN_ON(hook != timer_hook);
201         timer_hook = NULL;
202         /* make sure all CPUs see the NULL hook */
203         synchronize_sched();  /* Allow ongoing interrupts to complete. */
204 }
205 EXPORT_SYMBOL_GPL(unregister_timer_hook);
206
207 #endif /* CONFIG_PROFILING */
208
209
210 #ifdef CONFIG_SMP
211 /*
212  * Each cpu has a pair of open-addressed hashtables for pending
213  * profile hits. read_profile() IPI's all cpus to request them
214  * to flip buffers and flushes their contents to prof_buffer itself.
215  * Flip requests are serialized by the profile_flip_mutex. The sole
216  * use of having a second hashtable is for avoiding cacheline
217  * contention that would otherwise happen during flushes of pending
218  * profile hits required for the accuracy of reported profile hits
219  * and so resurrect the interrupt livelock issue.
220  *
221  * The open-addressed hashtables are indexed by profile buffer slot
222  * and hold the number of pending hits to that profile buffer slot on
223  * a cpu in an entry. When the hashtable overflows, all pending hits
224  * are accounted to their corresponding profile buffer slots with
225  * atomic_add() and the hashtable emptied. As numerous pending hits
226  * may be accounted to a profile buffer slot in a hashtable entry,
227  * this amortizes a number of atomic profile buffer increments likely
228  * to be far larger than the number of entries in the hashtable,
229  * particularly given that the number of distinct profile buffer
230  * positions to which hits are accounted during short intervals (e.g.
231  * several seconds) is usually very small. Exclusion from buffer
232  * flipping is provided by interrupt disablement (note that for
233  * SCHED_PROFILING or SLEEP_PROFILING profile_hit() may be called from
234  * process context).
235  * The hash function is meant to be lightweight as opposed to strong,
236  * and was vaguely inspired by ppc64 firmware-supported inverted
237  * pagetable hash functions, but uses a full hashtable full of finite
238  * collision chains, not just pairs of them.
239  *
240  * -- wli
241  */
242 static void __profile_flip_buffers(void *unused)
243 {
244         int cpu = smp_processor_id();
245
246         per_cpu(cpu_profile_flip, cpu) = !per_cpu(cpu_profile_flip, cpu);
247 }
248
249 static void profile_flip_buffers(void)
250 {
251         int i, j, cpu;
252
253         mutex_lock(&profile_flip_mutex);
254         j = per_cpu(cpu_profile_flip, get_cpu());
255         put_cpu();
256         on_each_cpu(__profile_flip_buffers, NULL, 0, 1);
257         for_each_online_cpu(cpu) {
258                 struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[j];
259                 for (i = 0; i < NR_PROFILE_HIT; ++i) {
260                         if (!hits[i].hits) {
261                                 if (hits[i].pc)
262                                         hits[i].pc = 0;
263                                 continue;
264                         }
265                         atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
266                         hits[i].hits = hits[i].pc = 0;
267                 }
268         }
269         mutex_unlock(&profile_flip_mutex);
270 }
271
272 static void profile_discard_flip_buffers(void)
273 {
274         int i, cpu;
275
276         mutex_lock(&profile_flip_mutex);
277         i = per_cpu(cpu_profile_flip, get_cpu());
278         put_cpu();
279         on_each_cpu(__profile_flip_buffers, NULL, 0, 1);
280         for_each_online_cpu(cpu) {
281                 struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[i];
282                 memset(hits, 0, NR_PROFILE_HIT*sizeof(struct profile_hit));
283         }
284         mutex_unlock(&profile_flip_mutex);
285 }
286
287 void profile_hits(int type, void *__pc, unsigned int nr_hits)
288 {
289         unsigned long primary, secondary, flags, pc = (unsigned long)__pc;
290         int i, j, cpu;
291         struct profile_hit *hits;
292
293         if (prof_on != type || !prof_buffer)
294                 return;
295         pc = min((pc - (unsigned long)_stext) >> prof_shift, prof_len - 1);
296         i = primary = (pc & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
297         secondary = (~(pc << 1) & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
298         cpu = get_cpu();
299         hits = per_cpu(cpu_profile_hits, cpu)[per_cpu(cpu_profile_flip, cpu)];
300         if (!hits) {
301                 put_cpu();
302                 return;
303         }
304         /*
305          * We buffer the global profiler buffer into a per-CPU
306          * queue and thus reduce the number of global (and possibly
307          * NUMA-alien) accesses. The write-queue is self-coalescing:
308          */
309         local_irq_save(flags);
310         do {
311                 for (j = 0; j < PROFILE_GRPSZ; ++j) {
312                         if (hits[i + j].pc == pc) {
313                                 hits[i + j].hits += nr_hits;
314                                 goto out;
315                         } else if (!hits[i + j].hits) {
316                                 hits[i + j].pc = pc;
317                                 hits[i + j].hits = nr_hits;
318                                 goto out;
319                         }
320                 }
321                 i = (i + secondary) & (NR_PROFILE_HIT - 1);
322         } while (i != primary);
323
324         /*
325          * Add the current hit(s) and flush the write-queue out
326          * to the global buffer:
327          */
328         atomic_add(nr_hits, &prof_buffer[pc]);
329         for (i = 0; i < NR_PROFILE_HIT; ++i) {
330                 atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
331                 hits[i].pc = hits[i].hits = 0;
332         }
333 out:
334         local_irq_restore(flags);
335         put_cpu();
336 }
337
338 static int __devinit profile_cpu_callback(struct notifier_block *info,
339                                         unsigned long action, void *__cpu)
340 {
341         int node, cpu = (unsigned long)__cpu;
342         struct page *page;
343
344         switch (action) {
345         case CPU_UP_PREPARE:
346         case CPU_UP_PREPARE_FROZEN:
347                 node = cpu_to_node(cpu);
348                 per_cpu(cpu_profile_flip, cpu) = 0;
349                 if (!per_cpu(cpu_profile_hits, cpu)[1]) {
350                         page = alloc_pages_node(node,
351                                         GFP_KERNEL | __GFP_ZERO,
352                                         0);
353                         if (!page)
354                                 return NOTIFY_BAD;
355                         per_cpu(cpu_profile_hits, cpu)[1] = page_address(page);
356                 }
357                 if (!per_cpu(cpu_profile_hits, cpu)[0]) {
358                         page = alloc_pages_node(node,
359                                         GFP_KERNEL | __GFP_ZERO,
360                                         0);
361                         if (!page)
362                                 goto out_free;
363                         per_cpu(cpu_profile_hits, cpu)[0] = page_address(page);
364                 }
365                 break;
366 out_free:
367                 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
368                 per_cpu(cpu_profile_hits, cpu)[1] = NULL;
369                 __free_page(page);
370                 return NOTIFY_BAD;
371         case CPU_ONLINE:
372         case CPU_ONLINE_FROZEN:
373                 cpu_set(cpu, prof_cpu_mask);
374                 break;
375         case CPU_UP_CANCELED:
376         case CPU_UP_CANCELED_FROZEN:
377         case CPU_DEAD:
378         case CPU_DEAD_FROZEN:
379                 cpu_clear(cpu, prof_cpu_mask);
380                 if (per_cpu(cpu_profile_hits, cpu)[0]) {
381                         page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[0]);
382                         per_cpu(cpu_profile_hits, cpu)[0] = NULL;
383                         __free_page(page);
384                 }
385                 if (per_cpu(cpu_profile_hits, cpu)[1]) {
386                         page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
387                         per_cpu(cpu_profile_hits, cpu)[1] = NULL;
388                         __free_page(page);
389                 }
390                 break;
391         }
392         return NOTIFY_OK;
393 }
394 #else /* !CONFIG_SMP */
395 #define profile_flip_buffers()          do { } while (0)
396 #define profile_discard_flip_buffers()  do { } while (0)
397 #define profile_cpu_callback            NULL
398
399 void profile_hits(int type, void *__pc, unsigned int nr_hits)
400 {
401         unsigned long pc;
402
403         if (prof_on != type || !prof_buffer)
404                 return;
405         pc = ((unsigned long)__pc - (unsigned long)_stext) >> prof_shift;
406         atomic_add(nr_hits, &prof_buffer[min(pc, prof_len - 1)]);
407 }
408 #endif /* !CONFIG_SMP */
409 EXPORT_SYMBOL_GPL(profile_hits);
410
411 void profile_tick(int type)
412 {
413         struct pt_regs *regs = get_irq_regs();
414
415         if (type == CPU_PROFILING && timer_hook)
416                 timer_hook(regs);
417         if (!user_mode(regs) && cpu_isset(smp_processor_id(), prof_cpu_mask))
418                 profile_hit(type, (void *)profile_pc(regs));
419 }
420
421 #ifdef CONFIG_PROC_FS
422 #include <linux/proc_fs.h>
423 #include <asm/uaccess.h>
424 #include <asm/ptrace.h>
425
426 static int prof_cpu_mask_read_proc(char *page, char **start, off_t off,
427                         int count, int *eof, void *data)
428 {
429         int len = cpumask_scnprintf(page, count, *(cpumask_t *)data);
430         if (count - len < 2)
431                 return -EINVAL;
432         len += sprintf(page + len, "\n");
433         return len;
434 }
435
436 static int prof_cpu_mask_write_proc(struct file *file,
437         const char __user *buffer,  unsigned long count, void *data)
438 {
439         cpumask_t *mask = (cpumask_t *)data;
440         unsigned long full_count = count, err;
441         cpumask_t new_value;
442
443         err = cpumask_parse_user(buffer, count, new_value);
444         if (err)
445                 return err;
446
447         *mask = new_value;
448         return full_count;
449 }
450
451 void create_prof_cpu_mask(struct proc_dir_entry *root_irq_dir)
452 {
453         struct proc_dir_entry *entry;
454
455         /* create /proc/irq/prof_cpu_mask */
456         entry = create_proc_entry("prof_cpu_mask", 0600, root_irq_dir);
457         if (!entry)
458                 return;
459         entry->data = (void *)&prof_cpu_mask;
460         entry->read_proc = prof_cpu_mask_read_proc;
461         entry->write_proc = prof_cpu_mask_write_proc;
462 }
463
464 /*
465  * This function accesses profiling information. The returned data is
466  * binary: the sampling step and the actual contents of the profile
467  * buffer. Use of the program readprofile is recommended in order to
468  * get meaningful info out of these data.
469  */
470 static ssize_t
471 read_profile(struct file *file, char __user *buf, size_t count, loff_t *ppos)
472 {
473         unsigned long p = *ppos;
474         ssize_t read;
475         char *pnt;
476         unsigned int sample_step = 1 << prof_shift;
477
478         profile_flip_buffers();
479         if (p >= (prof_len+1)*sizeof(unsigned int))
480                 return 0;
481         if (count > (prof_len+1)*sizeof(unsigned int) - p)
482                 count = (prof_len+1)*sizeof(unsigned int) - p;
483         read = 0;
484
485         while (p < sizeof(unsigned int) && count > 0) {
486                 if (put_user(*((char *)(&sample_step)+p), buf))
487                         return -EFAULT;
488                 buf++; p++; count--; read++;
489         }
490         pnt = (char *)prof_buffer + p - sizeof(atomic_t);
491         if (copy_to_user(buf, (void *)pnt, count))
492                 return -EFAULT;
493         read += count;
494         *ppos += read;
495         return read;
496 }
497
498 /*
499  * Writing to /proc/profile resets the counters
500  *
501  * Writing a 'profiling multiplier' value into it also re-sets the profiling
502  * interrupt frequency, on architectures that support this.
503  */
504 static ssize_t write_profile(struct file *file, const char __user *buf,
505                              size_t count, loff_t *ppos)
506 {
507 #ifdef CONFIG_SMP
508         extern int setup_profiling_timer(unsigned int multiplier);
509
510         if (count == sizeof(int)) {
511                 unsigned int multiplier;
512
513                 if (copy_from_user(&multiplier, buf, sizeof(int)))
514                         return -EFAULT;
515
516                 if (setup_profiling_timer(multiplier))
517                         return -EINVAL;
518         }
519 #endif
520         profile_discard_flip_buffers();
521         memset(prof_buffer, 0, prof_len * sizeof(atomic_t));
522         return count;
523 }
524
525 static const struct file_operations proc_profile_operations = {
526         .read           = read_profile,
527         .write          = write_profile,
528 };
529
530 #ifdef CONFIG_SMP
531 static void __init profile_nop(void *unused)
532 {
533 }
534
535 static int __init create_hash_tables(void)
536 {
537         int cpu;
538
539         for_each_online_cpu(cpu) {
540                 int node = cpu_to_node(cpu);
541                 struct page *page;
542
543                 page = alloc_pages_node(node,
544                                 GFP_KERNEL | __GFP_ZERO | GFP_THISNODE,
545                                 0);
546                 if (!page)
547                         goto out_cleanup;
548                 per_cpu(cpu_profile_hits, cpu)[1]
549                                 = (struct profile_hit *)page_address(page);
550                 page = alloc_pages_node(node,
551                                 GFP_KERNEL | __GFP_ZERO | GFP_THISNODE,
552                                 0);
553                 if (!page)
554                         goto out_cleanup;
555                 per_cpu(cpu_profile_hits, cpu)[0]
556                                 = (struct profile_hit *)page_address(page);
557         }
558         return 0;
559 out_cleanup:
560         prof_on = 0;
561         smp_mb();
562         on_each_cpu(profile_nop, NULL, 0, 1);
563         for_each_online_cpu(cpu) {
564                 struct page *page;
565
566                 if (per_cpu(cpu_profile_hits, cpu)[0]) {
567                         page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[0]);
568                         per_cpu(cpu_profile_hits, cpu)[0] = NULL;
569                         __free_page(page);
570                 }
571                 if (per_cpu(cpu_profile_hits, cpu)[1]) {
572                         page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
573                         per_cpu(cpu_profile_hits, cpu)[1] = NULL;
574                         __free_page(page);
575                 }
576         }
577         return -1;
578 }
579 #else
580 #define create_hash_tables()                    ({ 0; })
581 #endif
582
583 static int __init create_proc_profile(void)
584 {
585         struct proc_dir_entry *entry;
586
587         if (!prof_on)
588                 return 0;
589         if (create_hash_tables())
590                 return -1;
591         entry = create_proc_entry("profile", S_IWUSR | S_IRUGO, NULL);
592         if (!entry)
593                 return 0;
594         entry->proc_fops = &proc_profile_operations;
595         entry->size = (1+prof_len) * sizeof(atomic_t);
596         hotcpu_notifier(profile_cpu_callback, 0);
597         return 0;
598 }
599 module_init(create_proc_profile);
600 #endif /* CONFIG_PROC_FS */