kernel.org linux-2.6.10
[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/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 *);
39
40 static atomic_t *prof_buffer;
41 static unsigned long prof_len, prof_shift;
42 static int prof_on;
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         int par;
53
54         if (!strncmp(str, "schedule", 8)) {
55                 prof_on = SCHED_PROFILING;
56                 printk(KERN_INFO "kernel schedule profiling enabled\n");
57                 if (str[7] == ',')
58                         str += 8;
59         }
60         if (get_option(&str,&par)) {
61                 prof_shift = par;
62                 prof_on = CPU_PROFILING;
63                 printk(KERN_INFO "kernel profiling enabled (shift: %ld)\n",
64                         prof_shift);
65         }
66         return 1;
67 }
68 __setup("profile=", profile_setup);
69
70
71 void __init profile_init(void)
72 {
73         if (!prof_on) 
74                 return;
75  
76         /* only text is profiled */
77         prof_len = (_etext - _stext) >> prof_shift;
78         prof_buffer = alloc_bootmem(prof_len*sizeof(atomic_t));
79 }
80
81 /* Profile event notifications */
82  
83 #ifdef CONFIG_PROFILING
84  
85 static DECLARE_RWSEM(profile_rwsem);
86 static rwlock_t handoff_lock = RW_LOCK_UNLOCKED;
87 static struct notifier_block * task_exit_notifier;
88 static struct notifier_block * task_free_notifier;
89 static struct notifier_block * munmap_notifier;
90  
91 void profile_task_exit(struct task_struct * task)
92 {
93         down_read(&profile_rwsem);
94         notifier_call_chain(&task_exit_notifier, 0, task);
95         up_read(&profile_rwsem);
96 }
97  
98 int profile_handoff_task(struct task_struct * task)
99 {
100         int ret;
101         read_lock(&handoff_lock);
102         ret = notifier_call_chain(&task_free_notifier, 0, task);
103         read_unlock(&handoff_lock);
104         return (ret == NOTIFY_OK) ? 1 : 0;
105 }
106
107 void profile_munmap(unsigned long addr)
108 {
109         down_read(&profile_rwsem);
110         notifier_call_chain(&munmap_notifier, 0, (void *)addr);
111         up_read(&profile_rwsem);
112 }
113
114 int task_handoff_register(struct notifier_block * n)
115 {
116         int err = -EINVAL;
117
118         write_lock(&handoff_lock);
119         err = notifier_chain_register(&task_free_notifier, n);
120         write_unlock(&handoff_lock);
121         return err;
122 }
123
124 int task_handoff_unregister(struct notifier_block * n)
125 {
126         int err = -EINVAL;
127
128         write_lock(&handoff_lock);
129         err = notifier_chain_unregister(&task_free_notifier, n);
130         write_unlock(&handoff_lock);
131         return err;
132 }
133
134 int profile_event_register(enum profile_type type, struct notifier_block * n)
135 {
136         int err = -EINVAL;
137  
138         down_write(&profile_rwsem);
139  
140         switch (type) {
141                 case PROFILE_TASK_EXIT:
142                         err = notifier_chain_register(&task_exit_notifier, n);
143                         break;
144                 case PROFILE_MUNMAP:
145                         err = notifier_chain_register(&munmap_notifier, n);
146                         break;
147         }
148  
149         up_write(&profile_rwsem);
150  
151         return err;
152 }
153
154  
155 int profile_event_unregister(enum profile_type type, struct notifier_block * n)
156 {
157         int err = -EINVAL;
158  
159         down_write(&profile_rwsem);
160  
161         switch (type) {
162                 case PROFILE_TASK_EXIT:
163                         err = notifier_chain_unregister(&task_exit_notifier, n);
164                         break;
165                 case PROFILE_MUNMAP:
166                         err = notifier_chain_unregister(&munmap_notifier, n);
167                         break;
168         }
169
170         up_write(&profile_rwsem);
171         return err;
172 }
173
174 int register_timer_hook(int (*hook)(struct pt_regs *))
175 {
176         if (timer_hook)
177                 return -EBUSY;
178         timer_hook = hook;
179         return 0;
180 }
181
182 void unregister_timer_hook(int (*hook)(struct pt_regs *))
183 {
184         WARN_ON(hook != timer_hook);
185         timer_hook = NULL;
186         /* make sure all CPUs see the NULL hook */
187         synchronize_kernel();
188 }
189
190 EXPORT_SYMBOL_GPL(register_timer_hook);
191 EXPORT_SYMBOL_GPL(unregister_timer_hook);
192 EXPORT_SYMBOL_GPL(task_handoff_register);
193 EXPORT_SYMBOL_GPL(task_handoff_unregister);
194
195 #endif /* CONFIG_PROFILING */
196
197 EXPORT_SYMBOL_GPL(profile_event_register);
198 EXPORT_SYMBOL_GPL(profile_event_unregister);
199
200 #ifdef CONFIG_SMP
201 /*
202  * Each cpu has a pair of open-addressed hashtables for pending
203  * profile hits. read_profile() IPI's all cpus to request them
204  * to flip buffers and flushes their contents to prof_buffer itself.
205  * Flip requests are serialized by the profile_flip_mutex. The sole
206  * use of having a second hashtable is for avoiding cacheline
207  * contention that would otherwise happen during flushes of pending
208  * profile hits required for the accuracy of reported profile hits
209  * and so resurrect the interrupt livelock issue.
210  *
211  * The open-addressed hashtables are indexed by profile buffer slot
212  * and hold the number of pending hits to that profile buffer slot on
213  * a cpu in an entry. When the hashtable overflows, all pending hits
214  * are accounted to their corresponding profile buffer slots with
215  * atomic_add() and the hashtable emptied. As numerous pending hits
216  * may be accounted to a profile buffer slot in a hashtable entry,
217  * this amortizes a number of atomic profile buffer increments likely
218  * to be far larger than the number of entries in the hashtable,
219  * particularly given that the number of distinct profile buffer
220  * positions to which hits are accounted during short intervals (e.g.
221  * several seconds) is usually very small. Exclusion from buffer
222  * flipping is provided by interrupt disablement (note that for
223  * SCHED_PROFILING profile_hit() may be called from process context).
224  * The hash function is meant to be lightweight as opposed to strong,
225  * and was vaguely inspired by ppc64 firmware-supported inverted
226  * pagetable hash functions, but uses a full hashtable full of finite
227  * collision chains, not just pairs of them.
228  *
229  * -- wli
230  */
231 static void __profile_flip_buffers(void *unused)
232 {
233         int cpu = smp_processor_id();
234
235         per_cpu(cpu_profile_flip, cpu) = !per_cpu(cpu_profile_flip, cpu);
236 }
237
238 static void profile_flip_buffers(void)
239 {
240         int i, j, cpu;
241
242         down(&profile_flip_mutex);
243         j = per_cpu(cpu_profile_flip, get_cpu());
244         put_cpu();
245         on_each_cpu(__profile_flip_buffers, NULL, 0, 1);
246         for_each_online_cpu(cpu) {
247                 struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[j];
248                 for (i = 0; i < NR_PROFILE_HIT; ++i) {
249                         if (!hits[i].hits) {
250                                 if (hits[i].pc)
251                                         hits[i].pc = 0;
252                                 continue;
253                         }
254                         atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
255                         hits[i].hits = hits[i].pc = 0;
256                 }
257         }
258         up(&profile_flip_mutex);
259 }
260
261 static void profile_discard_flip_buffers(void)
262 {
263         int i, cpu;
264
265         down(&profile_flip_mutex);
266         i = per_cpu(cpu_profile_flip, get_cpu());
267         put_cpu();
268         on_each_cpu(__profile_flip_buffers, NULL, 0, 1);
269         for_each_online_cpu(cpu) {
270                 struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[i];
271                 memset(hits, 0, NR_PROFILE_HIT*sizeof(struct profile_hit));
272         }
273         up(&profile_flip_mutex);
274 }
275
276 void profile_hit(int type, void *__pc)
277 {
278         unsigned long primary, secondary, flags, pc = (unsigned long)__pc;
279         int i, j, cpu;
280         struct profile_hit *hits;
281
282         if (prof_on != type || !prof_buffer)
283                 return;
284         pc = min((pc - (unsigned long)_stext) >> prof_shift, prof_len - 1);
285         i = primary = (pc & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
286         secondary = (~(pc << 1) & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
287         cpu = get_cpu();
288         hits = per_cpu(cpu_profile_hits, cpu)[per_cpu(cpu_profile_flip, cpu)];
289         if (!hits) {
290                 put_cpu();
291                 return;
292         }
293         local_irq_save(flags);
294         do {
295                 for (j = 0; j < PROFILE_GRPSZ; ++j) {
296                         if (hits[i + j].pc == pc) {
297                                 hits[i + j].hits++;
298                                 goto out;
299                         } else if (!hits[i + j].hits) {
300                                 hits[i + j].pc = pc;
301                                 hits[i + j].hits = 1;
302                                 goto out;
303                         }
304                 }
305                 i = (i + secondary) & (NR_PROFILE_HIT - 1);
306         } while (i != primary);
307         atomic_inc(&prof_buffer[pc]);
308         for (i = 0; i < NR_PROFILE_HIT; ++i) {
309                 atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
310                 hits[i].pc = hits[i].hits = 0;
311         }
312 out:
313         local_irq_restore(flags);
314         put_cpu();
315 }
316
317 #ifdef CONFIG_HOTPLUG_CPU
318 static int __devinit profile_cpu_callback(struct notifier_block *info,
319                                         unsigned long action, void *__cpu)
320 {
321         int node, cpu = (unsigned long)__cpu;
322         struct page *page;
323
324         switch (action) {
325         case CPU_UP_PREPARE:
326                 node = cpu_to_node(cpu);
327                 per_cpu(cpu_profile_flip, cpu) = 0;
328                 if (!per_cpu(cpu_profile_hits, cpu)[1]) {
329                         page = alloc_pages_node(node, GFP_KERNEL, 0);
330                         if (!page)
331                                 return NOTIFY_BAD;
332                         clear_highpage(page);
333                         per_cpu(cpu_profile_hits, cpu)[1] = page_address(page);
334                 }
335                 if (!per_cpu(cpu_profile_hits, cpu)[0]) {
336                         page = alloc_pages_node(node, GFP_KERNEL, 0);
337                         if (!page)
338                                 goto out_free;
339                         clear_highpage(page);
340                         per_cpu(cpu_profile_hits, cpu)[0] = page_address(page);
341                 }
342                 break;
343         out_free:
344                 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
345                 per_cpu(cpu_profile_hits, cpu)[1] = NULL;
346                 __free_page(page);
347                 return NOTIFY_BAD;
348         case CPU_ONLINE:
349                 cpu_set(cpu, prof_cpu_mask);
350                 break;
351         case CPU_UP_CANCELED:
352         case CPU_DEAD:
353                 cpu_clear(cpu, prof_cpu_mask);
354                 if (per_cpu(cpu_profile_hits, cpu)[0]) {
355                         page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[0]);
356                         per_cpu(cpu_profile_hits, cpu)[0] = NULL;
357                         __free_page(page);
358                 }
359                 if (per_cpu(cpu_profile_hits, cpu)[1]) {
360                         page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
361                         per_cpu(cpu_profile_hits, cpu)[1] = NULL;
362                         __free_page(page);
363                 }
364                 break;
365         }
366         return NOTIFY_OK;
367 }
368 #endif /* CONFIG_HOTPLUG_CPU */
369 #else /* !CONFIG_SMP */
370 #define profile_flip_buffers()          do { } while (0)
371 #define profile_discard_flip_buffers()  do { } while (0)
372
373 void profile_hit(int type, void *__pc)
374 {
375         unsigned long pc;
376
377         if (prof_on != type || !prof_buffer)
378                 return;
379         pc = ((unsigned long)__pc - (unsigned long)_stext) >> prof_shift;
380         atomic_inc(&prof_buffer[min(pc, prof_len - 1)]);
381 }
382 #endif /* !CONFIG_SMP */
383
384 void profile_tick(int type, struct pt_regs *regs)
385 {
386         if (type == CPU_PROFILING && timer_hook)
387                 timer_hook(regs);
388         if (!user_mode(regs) && cpu_isset(smp_processor_id(), prof_cpu_mask))
389                 profile_hit(type, (void *)profile_pc(regs));
390 }
391
392 #ifdef CONFIG_PROC_FS
393 #include <linux/proc_fs.h>
394 #include <asm/uaccess.h>
395 #include <asm/ptrace.h>
396
397 static int prof_cpu_mask_read_proc (char *page, char **start, off_t off,
398                         int count, int *eof, void *data)
399 {
400         int len = cpumask_scnprintf(page, count, *(cpumask_t *)data);
401         if (count - len < 2)
402                 return -EINVAL;
403         len += sprintf(page + len, "\n");
404         return len;
405 }
406
407 static int prof_cpu_mask_write_proc (struct file *file, const char __user *buffer,
408                                         unsigned long count, void *data)
409 {
410         cpumask_t *mask = (cpumask_t *)data;
411         unsigned long full_count = count, err;
412         cpumask_t new_value;
413
414         err = cpumask_parse(buffer, count, new_value);
415         if (err)
416                 return err;
417
418         *mask = new_value;
419         return full_count;
420 }
421
422 void create_prof_cpu_mask(struct proc_dir_entry *root_irq_dir)
423 {
424         struct proc_dir_entry *entry;
425
426         /* create /proc/irq/prof_cpu_mask */
427         if (!(entry = create_proc_entry("prof_cpu_mask", 0600, root_irq_dir)))
428                 return;
429         entry->nlink = 1;
430         entry->data = (void *)&prof_cpu_mask;
431         entry->read_proc = prof_cpu_mask_read_proc;
432         entry->write_proc = prof_cpu_mask_write_proc;
433 }
434
435 /*
436  * This function accesses profiling information. The returned data is
437  * binary: the sampling step and the actual contents of the profile
438  * buffer. Use of the program readprofile is recommended in order to
439  * get meaningful info out of these data.
440  */
441 static ssize_t
442 read_profile(struct file *file, char __user *buf, size_t count, loff_t *ppos)
443 {
444         unsigned long p = *ppos;
445         ssize_t read;
446         char * pnt;
447         unsigned int sample_step = 1 << prof_shift;
448
449         profile_flip_buffers();
450         if (p >= (prof_len+1)*sizeof(unsigned int))
451                 return 0;
452         if (count > (prof_len+1)*sizeof(unsigned int) - p)
453                 count = (prof_len+1)*sizeof(unsigned int) - p;
454         read = 0;
455
456         while (p < sizeof(unsigned int) && count > 0) {
457                 put_user(*((char *)(&sample_step)+p),buf);
458                 buf++; p++; count--; read++;
459         }
460         pnt = (char *)prof_buffer + p - sizeof(atomic_t);
461         if (copy_to_user(buf,(void *)pnt,count))
462                 return -EFAULT;
463         read += count;
464         *ppos += read;
465         return read;
466 }
467
468 /*
469  * Writing to /proc/profile resets the counters
470  *
471  * Writing a 'profiling multiplier' value into it also re-sets the profiling
472  * interrupt frequency, on architectures that support this.
473  */
474 static ssize_t write_profile(struct file *file, const char __user *buf,
475                              size_t count, loff_t *ppos)
476 {
477 #ifdef CONFIG_SMP
478         extern int setup_profiling_timer (unsigned int multiplier);
479
480         if (count == sizeof(int)) {
481                 unsigned int multiplier;
482
483                 if (copy_from_user(&multiplier, buf, sizeof(int)))
484                         return -EFAULT;
485
486                 if (setup_profiling_timer(multiplier))
487                         return -EINVAL;
488         }
489 #endif
490         profile_discard_flip_buffers();
491         memset(prof_buffer, 0, prof_len * sizeof(atomic_t));
492         return count;
493 }
494
495 static struct file_operations proc_profile_operations = {
496         .read           = read_profile,
497         .write          = write_profile,
498 };
499
500 #ifdef CONFIG_SMP
501 static void __init profile_nop(void *unused)
502 {
503 }
504
505 static int __init create_hash_tables(void)
506 {
507         int cpu;
508
509         for_each_online_cpu(cpu) {
510                 int node = cpu_to_node(cpu);
511                 struct page *page;
512
513                 page = alloc_pages_node(node, GFP_KERNEL, 0);
514                 if (!page)
515                         goto out_cleanup;
516                 clear_highpage(page);
517                 per_cpu(cpu_profile_hits, cpu)[1]
518                                 = (struct profile_hit *)page_address(page);
519                 page = alloc_pages_node(node, GFP_KERNEL, 0);
520                 if (!page)
521                         goto out_cleanup;
522                 clear_highpage(page);
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         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 */