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