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