[PATCH] use smp_mb/wmb/rmb where possible
[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 DEFINE_RWLOCK(handoff_lock);
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 | __GFP_ZERO, 0);
330                         if (!page)
331                                 return NOTIFY_BAD;
332                         per_cpu(cpu_profile_hits, cpu)[1] = page_address(page);
333                 }
334                 if (!per_cpu(cpu_profile_hits, cpu)[0]) {
335                         page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
336                         if (!page)
337                                 goto out_free;
338                         per_cpu(cpu_profile_hits, cpu)[0] = page_address(page);
339                 }
340                 break;
341         out_free:
342                 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
343                 per_cpu(cpu_profile_hits, cpu)[1] = NULL;
344                 __free_page(page);
345                 return NOTIFY_BAD;
346         case CPU_ONLINE:
347                 cpu_set(cpu, prof_cpu_mask);
348                 break;
349         case CPU_UP_CANCELED:
350         case CPU_DEAD:
351                 cpu_clear(cpu, prof_cpu_mask);
352                 if (per_cpu(cpu_profile_hits, cpu)[0]) {
353                         page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[0]);
354                         per_cpu(cpu_profile_hits, cpu)[0] = NULL;
355                         __free_page(page);
356                 }
357                 if (per_cpu(cpu_profile_hits, cpu)[1]) {
358                         page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
359                         per_cpu(cpu_profile_hits, cpu)[1] = NULL;
360                         __free_page(page);
361                 }
362                 break;
363         }
364         return NOTIFY_OK;
365 }
366 #endif /* CONFIG_HOTPLUG_CPU */
367 #else /* !CONFIG_SMP */
368 #define profile_flip_buffers()          do { } while (0)
369 #define profile_discard_flip_buffers()  do { } while (0)
370
371 void profile_hit(int type, void *__pc)
372 {
373         unsigned long pc;
374
375         if (prof_on != type || !prof_buffer)
376                 return;
377         pc = ((unsigned long)__pc - (unsigned long)_stext) >> prof_shift;
378         atomic_inc(&prof_buffer[min(pc, prof_len - 1)]);
379 }
380 #endif /* !CONFIG_SMP */
381
382 void profile_tick(int type, struct pt_regs *regs)
383 {
384         if (type == CPU_PROFILING && timer_hook)
385                 timer_hook(regs);
386         if (!user_mode(regs) && cpu_isset(smp_processor_id(), prof_cpu_mask))
387                 profile_hit(type, (void *)profile_pc(regs));
388 }
389
390 #ifdef CONFIG_PROC_FS
391 #include <linux/proc_fs.h>
392 #include <asm/uaccess.h>
393 #include <asm/ptrace.h>
394
395 static int prof_cpu_mask_read_proc (char *page, char **start, off_t off,
396                         int count, int *eof, void *data)
397 {
398         int len = cpumask_scnprintf(page, count, *(cpumask_t *)data);
399         if (count - len < 2)
400                 return -EINVAL;
401         len += sprintf(page + len, "\n");
402         return len;
403 }
404
405 static int prof_cpu_mask_write_proc (struct file *file, const char __user *buffer,
406                                         unsigned long count, void *data)
407 {
408         cpumask_t *mask = (cpumask_t *)data;
409         unsigned long full_count = count, err;
410         cpumask_t new_value;
411
412         err = cpumask_parse(buffer, count, new_value);
413         if (err)
414                 return err;
415
416         *mask = new_value;
417         return full_count;
418 }
419
420 void create_prof_cpu_mask(struct proc_dir_entry *root_irq_dir)
421 {
422         struct proc_dir_entry *entry;
423
424         /* create /proc/irq/prof_cpu_mask */
425         if (!(entry = create_proc_entry("prof_cpu_mask", 0600, root_irq_dir)))
426                 return;
427         entry->nlink = 1;
428         entry->data = (void *)&prof_cpu_mask;
429         entry->read_proc = prof_cpu_mask_read_proc;
430         entry->write_proc = prof_cpu_mask_write_proc;
431 }
432
433 /*
434  * This function accesses profiling information. The returned data is
435  * binary: the sampling step and the actual contents of the profile
436  * buffer. Use of the program readprofile is recommended in order to
437  * get meaningful info out of these data.
438  */
439 static ssize_t
440 read_profile(struct file *file, char __user *buf, size_t count, loff_t *ppos)
441 {
442         unsigned long p = *ppos;
443         ssize_t read;
444         char * pnt;
445         unsigned int sample_step = 1 << prof_shift;
446
447         profile_flip_buffers();
448         if (p >= (prof_len+1)*sizeof(unsigned int))
449                 return 0;
450         if (count > (prof_len+1)*sizeof(unsigned int) - p)
451                 count = (prof_len+1)*sizeof(unsigned int) - p;
452         read = 0;
453
454         while (p < sizeof(unsigned int) && count > 0) {
455                 put_user(*((char *)(&sample_step)+p),buf);
456                 buf++; p++; count--; read++;
457         }
458         pnt = (char *)prof_buffer + p - sizeof(atomic_t);
459         if (copy_to_user(buf,(void *)pnt,count))
460                 return -EFAULT;
461         read += count;
462         *ppos += read;
463         return read;
464 }
465
466 /*
467  * Writing to /proc/profile resets the counters
468  *
469  * Writing a 'profiling multiplier' value into it also re-sets the profiling
470  * interrupt frequency, on architectures that support this.
471  */
472 static ssize_t write_profile(struct file *file, const char __user *buf,
473                              size_t count, loff_t *ppos)
474 {
475 #ifdef CONFIG_SMP
476         extern int setup_profiling_timer (unsigned int multiplier);
477
478         if (count == sizeof(int)) {
479                 unsigned int multiplier;
480
481                 if (copy_from_user(&multiplier, buf, sizeof(int)))
482                         return -EFAULT;
483
484                 if (setup_profiling_timer(multiplier))
485                         return -EINVAL;
486         }
487 #endif
488         profile_discard_flip_buffers();
489         memset(prof_buffer, 0, prof_len * sizeof(atomic_t));
490         return count;
491 }
492
493 static struct file_operations proc_profile_operations = {
494         .read           = read_profile,
495         .write          = write_profile,
496 };
497
498 #ifdef CONFIG_SMP
499 static void __init profile_nop(void *unused)
500 {
501 }
502
503 static int __init create_hash_tables(void)
504 {
505         int cpu;
506
507         for_each_online_cpu(cpu) {
508                 int node = cpu_to_node(cpu);
509                 struct page *page;
510
511                 page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
512                 if (!page)
513                         goto out_cleanup;
514                 per_cpu(cpu_profile_hits, cpu)[1]
515                                 = (struct profile_hit *)page_address(page);
516                 page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
517                 if (!page)
518                         goto out_cleanup;
519                 per_cpu(cpu_profile_hits, cpu)[0]
520                                 = (struct profile_hit *)page_address(page);
521         }
522         return 0;
523 out_cleanup:
524         prof_on = 0;
525         smp_mb();
526         on_each_cpu(profile_nop, NULL, 0, 1);
527         for_each_online_cpu(cpu) {
528                 struct page *page;
529
530                 if (per_cpu(cpu_profile_hits, cpu)[0]) {
531                         page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[0]);
532                         per_cpu(cpu_profile_hits, cpu)[0] = NULL;
533                         __free_page(page);
534                 }
535                 if (per_cpu(cpu_profile_hits, cpu)[1]) {
536                         page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
537                         per_cpu(cpu_profile_hits, cpu)[1] = NULL;
538                         __free_page(page);
539                 }
540         }
541         return -1;
542 }
543 #else
544 #define create_hash_tables()                    ({ 0; })
545 #endif
546
547 static int __init create_proc_profile(void)
548 {
549         struct proc_dir_entry *entry;
550
551         if (!prof_on)
552                 return 0;
553         if (create_hash_tables())
554                 return -1;
555         if (!(entry = create_proc_entry("profile", S_IWUSR | S_IRUGO, NULL)))
556                 return 0;
557         entry->proc_fops = &proc_profile_operations;
558         entry->size = (1+prof_len) * sizeof(atomic_t);
559         hotcpu_notifier(profile_cpu_callback, 0);
560         return 0;
561 }
562 module_init(create_proc_profile);
563 #endif /* CONFIG_PROC_FS */