page allocator: do not check NUMA node ID when the caller knows the node is valid
[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 <linux/slab.h>
26 #include <linux/vmalloc.h>
27 #include <asm/sections.h>
28 #include <asm/irq_regs.h>
29 #include <asm/ptrace.h>
30
31 struct profile_hit {
32         u32 pc, hits;
33 };
34 #define PROFILE_GRPSHIFT        3
35 #define PROFILE_GRPSZ           (1 << PROFILE_GRPSHIFT)
36 #define NR_PROFILE_HIT          (PAGE_SIZE/sizeof(struct profile_hit))
37 #define NR_PROFILE_GRP          (NR_PROFILE_HIT/PROFILE_GRPSZ)
38
39 /* Oprofile timer tick hook */
40 static int (*timer_hook)(struct pt_regs *) __read_mostly;
41
42 static atomic_t *prof_buffer;
43 static unsigned long prof_len, prof_shift;
44
45 int prof_on __read_mostly;
46 EXPORT_SYMBOL_GPL(prof_on);
47
48 static cpumask_var_t prof_cpu_mask;
49 #ifdef CONFIG_SMP
50 static DEFINE_PER_CPU(struct profile_hit *[2], cpu_profile_hits);
51 static DEFINE_PER_CPU(int, cpu_profile_flip);
52 static DEFINE_MUTEX(profile_flip_mutex);
53 #endif /* CONFIG_SMP */
54
55 int profile_setup(char *str)
56 {
57         static char schedstr[] = "schedule";
58         static char sleepstr[] = "sleep";
59         static char kvmstr[] = "kvm";
60         int par;
61
62         if (!strncmp(str, sleepstr, strlen(sleepstr))) {
63 #ifdef CONFIG_SCHEDSTATS
64                 prof_on = SLEEP_PROFILING;
65                 if (str[strlen(sleepstr)] == ',')
66                         str += strlen(sleepstr) + 1;
67                 if (get_option(&str, &par))
68                         prof_shift = par;
69                 printk(KERN_INFO
70                         "kernel sleep profiling enabled (shift: %ld)\n",
71                         prof_shift);
72 #else
73                 printk(KERN_WARNING
74                         "kernel sleep profiling requires CONFIG_SCHEDSTATS\n");
75 #endif /* CONFIG_SCHEDSTATS */
76         } else if (!strncmp(str, schedstr, strlen(schedstr))) {
77                 prof_on = SCHED_PROFILING;
78                 if (str[strlen(schedstr)] == ',')
79                         str += strlen(schedstr) + 1;
80                 if (get_option(&str, &par))
81                         prof_shift = par;
82                 printk(KERN_INFO
83                         "kernel schedule profiling enabled (shift: %ld)\n",
84                         prof_shift);
85         } else if (!strncmp(str, kvmstr, strlen(kvmstr))) {
86                 prof_on = KVM_PROFILING;
87                 if (str[strlen(kvmstr)] == ',')
88                         str += strlen(kvmstr) + 1;
89                 if (get_option(&str, &par))
90                         prof_shift = par;
91                 printk(KERN_INFO
92                         "kernel KVM profiling enabled (shift: %ld)\n",
93                         prof_shift);
94         } else if (get_option(&str, &par)) {
95                 prof_shift = par;
96                 prof_on = CPU_PROFILING;
97                 printk(KERN_INFO "kernel profiling enabled (shift: %ld)\n",
98                         prof_shift);
99         }
100         return 1;
101 }
102 __setup("profile=", profile_setup);
103
104
105 int __ref profile_init(void)
106 {
107         int buffer_bytes;
108         if (!prof_on)
109                 return 0;
110
111         /* only text is profiled */
112         prof_len = (_etext - _stext) >> prof_shift;
113         buffer_bytes = prof_len*sizeof(atomic_t);
114
115         if (!alloc_cpumask_var(&prof_cpu_mask, GFP_KERNEL))
116                 return -ENOMEM;
117
118         cpumask_copy(prof_cpu_mask, cpu_possible_mask);
119
120         prof_buffer = kzalloc(buffer_bytes, GFP_KERNEL);
121         if (prof_buffer)
122                 return 0;
123
124         prof_buffer = alloc_pages_exact(buffer_bytes, GFP_KERNEL|__GFP_ZERO);
125         if (prof_buffer)
126                 return 0;
127
128         prof_buffer = vmalloc(buffer_bytes);
129         if (prof_buffer)
130                 return 0;
131
132         free_cpumask_var(prof_cpu_mask);
133         return -ENOMEM;
134 }
135
136 /* Profile event notifications */
137
138 static BLOCKING_NOTIFIER_HEAD(task_exit_notifier);
139 static ATOMIC_NOTIFIER_HEAD(task_free_notifier);
140 static BLOCKING_NOTIFIER_HEAD(munmap_notifier);
141
142 void profile_task_exit(struct task_struct *task)
143 {
144         blocking_notifier_call_chain(&task_exit_notifier, 0, task);
145 }
146
147 int profile_handoff_task(struct task_struct *task)
148 {
149         int ret;
150         ret = atomic_notifier_call_chain(&task_free_notifier, 0, task);
151         return (ret == NOTIFY_OK) ? 1 : 0;
152 }
153
154 void profile_munmap(unsigned long addr)
155 {
156         blocking_notifier_call_chain(&munmap_notifier, 0, (void *)addr);
157 }
158
159 int task_handoff_register(struct notifier_block *n)
160 {
161         return atomic_notifier_chain_register(&task_free_notifier, n);
162 }
163 EXPORT_SYMBOL_GPL(task_handoff_register);
164
165 int task_handoff_unregister(struct notifier_block *n)
166 {
167         return atomic_notifier_chain_unregister(&task_free_notifier, n);
168 }
169 EXPORT_SYMBOL_GPL(task_handoff_unregister);
170
171 int profile_event_register(enum profile_type type, struct notifier_block *n)
172 {
173         int err = -EINVAL;
174
175         switch (type) {
176         case PROFILE_TASK_EXIT:
177                 err = blocking_notifier_chain_register(
178                                 &task_exit_notifier, n);
179                 break;
180         case PROFILE_MUNMAP:
181                 err = blocking_notifier_chain_register(
182                                 &munmap_notifier, n);
183                 break;
184         }
185
186         return err;
187 }
188 EXPORT_SYMBOL_GPL(profile_event_register);
189
190 int profile_event_unregister(enum profile_type type, struct notifier_block *n)
191 {
192         int err = -EINVAL;
193
194         switch (type) {
195         case PROFILE_TASK_EXIT:
196                 err = blocking_notifier_chain_unregister(
197                                 &task_exit_notifier, n);
198                 break;
199         case PROFILE_MUNMAP:
200                 err = blocking_notifier_chain_unregister(
201                                 &munmap_notifier, n);
202                 break;
203         }
204
205         return err;
206 }
207 EXPORT_SYMBOL_GPL(profile_event_unregister);
208
209 int register_timer_hook(int (*hook)(struct pt_regs *))
210 {
211         if (timer_hook)
212                 return -EBUSY;
213         timer_hook = hook;
214         return 0;
215 }
216 EXPORT_SYMBOL_GPL(register_timer_hook);
217
218 void unregister_timer_hook(int (*hook)(struct pt_regs *))
219 {
220         WARN_ON(hook != timer_hook);
221         timer_hook = NULL;
222         /* make sure all CPUs see the NULL hook */
223         synchronize_sched();  /* Allow ongoing interrupts to complete. */
224 }
225 EXPORT_SYMBOL_GPL(unregister_timer_hook);
226
227
228 #ifdef CONFIG_SMP
229 /*
230  * Each cpu has a pair of open-addressed hashtables for pending
231  * profile hits. read_profile() IPI's all cpus to request them
232  * to flip buffers and flushes their contents to prof_buffer itself.
233  * Flip requests are serialized by the profile_flip_mutex. The sole
234  * use of having a second hashtable is for avoiding cacheline
235  * contention that would otherwise happen during flushes of pending
236  * profile hits required for the accuracy of reported profile hits
237  * and so resurrect the interrupt livelock issue.
238  *
239  * The open-addressed hashtables are indexed by profile buffer slot
240  * and hold the number of pending hits to that profile buffer slot on
241  * a cpu in an entry. When the hashtable overflows, all pending hits
242  * are accounted to their corresponding profile buffer slots with
243  * atomic_add() and the hashtable emptied. As numerous pending hits
244  * may be accounted to a profile buffer slot in a hashtable entry,
245  * this amortizes a number of atomic profile buffer increments likely
246  * to be far larger than the number of entries in the hashtable,
247  * particularly given that the number of distinct profile buffer
248  * positions to which hits are accounted during short intervals (e.g.
249  * several seconds) is usually very small. Exclusion from buffer
250  * flipping is provided by interrupt disablement (note that for
251  * SCHED_PROFILING or SLEEP_PROFILING profile_hit() may be called from
252  * process context).
253  * The hash function is meant to be lightweight as opposed to strong,
254  * and was vaguely inspired by ppc64 firmware-supported inverted
255  * pagetable hash functions, but uses a full hashtable full of finite
256  * collision chains, not just pairs of them.
257  *
258  * -- wli
259  */
260 static void __profile_flip_buffers(void *unused)
261 {
262         int cpu = smp_processor_id();
263
264         per_cpu(cpu_profile_flip, cpu) = !per_cpu(cpu_profile_flip, cpu);
265 }
266
267 static void profile_flip_buffers(void)
268 {
269         int i, j, cpu;
270
271         mutex_lock(&profile_flip_mutex);
272         j = 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)[j];
277                 for (i = 0; i < NR_PROFILE_HIT; ++i) {
278                         if (!hits[i].hits) {
279                                 if (hits[i].pc)
280                                         hits[i].pc = 0;
281                                 continue;
282                         }
283                         atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
284                         hits[i].hits = hits[i].pc = 0;
285                 }
286         }
287         mutex_unlock(&profile_flip_mutex);
288 }
289
290 static void profile_discard_flip_buffers(void)
291 {
292         int i, cpu;
293
294         mutex_lock(&profile_flip_mutex);
295         i = per_cpu(cpu_profile_flip, get_cpu());
296         put_cpu();
297         on_each_cpu(__profile_flip_buffers, NULL, 1);
298         for_each_online_cpu(cpu) {
299                 struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[i];
300                 memset(hits, 0, NR_PROFILE_HIT*sizeof(struct profile_hit));
301         }
302         mutex_unlock(&profile_flip_mutex);
303 }
304
305 void profile_hits(int type, void *__pc, unsigned int nr_hits)
306 {
307         unsigned long primary, secondary, flags, pc = (unsigned long)__pc;
308         int i, j, cpu;
309         struct profile_hit *hits;
310
311         if (prof_on != type || !prof_buffer)
312                 return;
313         pc = min((pc - (unsigned long)_stext) >> prof_shift, prof_len - 1);
314         i = primary = (pc & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
315         secondary = (~(pc << 1) & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
316         cpu = get_cpu();
317         hits = per_cpu(cpu_profile_hits, cpu)[per_cpu(cpu_profile_flip, cpu)];
318         if (!hits) {
319                 put_cpu();
320                 return;
321         }
322         /*
323          * We buffer the global profiler buffer into a per-CPU
324          * queue and thus reduce the number of global (and possibly
325          * NUMA-alien) accesses. The write-queue is self-coalescing:
326          */
327         local_irq_save(flags);
328         do {
329                 for (j = 0; j < PROFILE_GRPSZ; ++j) {
330                         if (hits[i + j].pc == pc) {
331                                 hits[i + j].hits += nr_hits;
332                                 goto out;
333                         } else if (!hits[i + j].hits) {
334                                 hits[i + j].pc = pc;
335                                 hits[i + j].hits = nr_hits;
336                                 goto out;
337                         }
338                 }
339                 i = (i + secondary) & (NR_PROFILE_HIT - 1);
340         } while (i != primary);
341
342         /*
343          * Add the current hit(s) and flush the write-queue out
344          * to the global buffer:
345          */
346         atomic_add(nr_hits, &prof_buffer[pc]);
347         for (i = 0; i < NR_PROFILE_HIT; ++i) {
348                 atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
349                 hits[i].pc = hits[i].hits = 0;
350         }
351 out:
352         local_irq_restore(flags);
353         put_cpu();
354 }
355
356 static int __cpuinit profile_cpu_callback(struct notifier_block *info,
357                                         unsigned long action, void *__cpu)
358 {
359         int node, cpu = (unsigned long)__cpu;
360         struct page *page;
361
362         switch (action) {
363         case CPU_UP_PREPARE:
364         case CPU_UP_PREPARE_FROZEN:
365                 node = cpu_to_node(cpu);
366                 per_cpu(cpu_profile_flip, cpu) = 0;
367                 if (!per_cpu(cpu_profile_hits, cpu)[1]) {
368                         page = alloc_pages_exact_node(node,
369                                         GFP_KERNEL | __GFP_ZERO,
370                                         0);
371                         if (!page)
372                                 return NOTIFY_BAD;
373                         per_cpu(cpu_profile_hits, cpu)[1] = page_address(page);
374                 }
375                 if (!per_cpu(cpu_profile_hits, cpu)[0]) {
376                         page = alloc_pages_exact_node(node,
377                                         GFP_KERNEL | __GFP_ZERO,
378                                         0);
379                         if (!page)
380                                 goto out_free;
381                         per_cpu(cpu_profile_hits, cpu)[0] = page_address(page);
382                 }
383                 break;
384 out_free:
385                 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
386                 per_cpu(cpu_profile_hits, cpu)[1] = NULL;
387                 __free_page(page);
388                 return NOTIFY_BAD;
389         case CPU_ONLINE:
390         case CPU_ONLINE_FROZEN:
391                 if (prof_cpu_mask != NULL)
392                         cpumask_set_cpu(cpu, prof_cpu_mask);
393                 break;
394         case CPU_UP_CANCELED:
395         case CPU_UP_CANCELED_FROZEN:
396         case CPU_DEAD:
397         case CPU_DEAD_FROZEN:
398                 if (prof_cpu_mask != NULL)
399                         cpumask_clear_cpu(cpu, prof_cpu_mask);
400                 if (per_cpu(cpu_profile_hits, cpu)[0]) {
401                         page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[0]);
402                         per_cpu(cpu_profile_hits, cpu)[0] = NULL;
403                         __free_page(page);
404                 }
405                 if (per_cpu(cpu_profile_hits, cpu)[1]) {
406                         page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
407                         per_cpu(cpu_profile_hits, cpu)[1] = NULL;
408                         __free_page(page);
409                 }
410                 break;
411         }
412         return NOTIFY_OK;
413 }
414 #else /* !CONFIG_SMP */
415 #define profile_flip_buffers()          do { } while (0)
416 #define profile_discard_flip_buffers()  do { } while (0)
417 #define profile_cpu_callback            NULL
418
419 void profile_hits(int type, void *__pc, unsigned int nr_hits)
420 {
421         unsigned long pc;
422
423         if (prof_on != type || !prof_buffer)
424                 return;
425         pc = ((unsigned long)__pc - (unsigned long)_stext) >> prof_shift;
426         atomic_add(nr_hits, &prof_buffer[min(pc, prof_len - 1)]);
427 }
428 #endif /* !CONFIG_SMP */
429 EXPORT_SYMBOL_GPL(profile_hits);
430
431 void profile_tick(int type)
432 {
433         struct pt_regs *regs = get_irq_regs();
434
435         if (type == CPU_PROFILING && timer_hook)
436                 timer_hook(regs);
437         if (!user_mode(regs) && prof_cpu_mask != NULL &&
438             cpumask_test_cpu(smp_processor_id(), prof_cpu_mask))
439                 profile_hit(type, (void *)profile_pc(regs));
440 }
441
442 #ifdef CONFIG_PROC_FS
443 #include <linux/proc_fs.h>
444 #include <asm/uaccess.h>
445
446 static int prof_cpu_mask_read_proc(char *page, char **start, off_t off,
447                         int count, int *eof, void *data)
448 {
449         int len = cpumask_scnprintf(page, count, data);
450         if (count - len < 2)
451                 return -EINVAL;
452         len += sprintf(page + len, "\n");
453         return len;
454 }
455
456 static int prof_cpu_mask_write_proc(struct file *file,
457         const char __user *buffer,  unsigned long count, void *data)
458 {
459         struct cpumask *mask = data;
460         unsigned long full_count = count, err;
461         cpumask_var_t new_value;
462
463         if (!alloc_cpumask_var(&new_value, GFP_KERNEL))
464                 return -ENOMEM;
465
466         err = cpumask_parse_user(buffer, count, new_value);
467         if (!err) {
468                 cpumask_copy(mask, new_value);
469                 err = full_count;
470         }
471         free_cpumask_var(new_value);
472         return err;
473 }
474
475 void create_prof_cpu_mask(struct proc_dir_entry *root_irq_dir)
476 {
477         struct proc_dir_entry *entry;
478
479         /* create /proc/irq/prof_cpu_mask */
480         entry = create_proc_entry("prof_cpu_mask", 0600, root_irq_dir);
481         if (!entry)
482                 return;
483         entry->data = prof_cpu_mask;
484         entry->read_proc = prof_cpu_mask_read_proc;
485         entry->write_proc = prof_cpu_mask_write_proc;
486 }
487
488 /*
489  * This function accesses profiling information. The returned data is
490  * binary: the sampling step and the actual contents of the profile
491  * buffer. Use of the program readprofile is recommended in order to
492  * get meaningful info out of these data.
493  */
494 static ssize_t
495 read_profile(struct file *file, char __user *buf, size_t count, loff_t *ppos)
496 {
497         unsigned long p = *ppos;
498         ssize_t read;
499         char *pnt;
500         unsigned int sample_step = 1 << prof_shift;
501
502         profile_flip_buffers();
503         if (p >= (prof_len+1)*sizeof(unsigned int))
504                 return 0;
505         if (count > (prof_len+1)*sizeof(unsigned int) - p)
506                 count = (prof_len+1)*sizeof(unsigned int) - p;
507         read = 0;
508
509         while (p < sizeof(unsigned int) && count > 0) {
510                 if (put_user(*((char *)(&sample_step)+p), buf))
511                         return -EFAULT;
512                 buf++; p++; count--; read++;
513         }
514         pnt = (char *)prof_buffer + p - sizeof(atomic_t);
515         if (copy_to_user(buf, (void *)pnt, count))
516                 return -EFAULT;
517         read += count;
518         *ppos += read;
519         return read;
520 }
521
522 /*
523  * Writing to /proc/profile resets the counters
524  *
525  * Writing a 'profiling multiplier' value into it also re-sets the profiling
526  * interrupt frequency, on architectures that support this.
527  */
528 static ssize_t write_profile(struct file *file, const char __user *buf,
529                              size_t count, loff_t *ppos)
530 {
531 #ifdef CONFIG_SMP
532         extern int setup_profiling_timer(unsigned int multiplier);
533
534         if (count == sizeof(int)) {
535                 unsigned int multiplier;
536
537                 if (copy_from_user(&multiplier, buf, sizeof(int)))
538                         return -EFAULT;
539
540                 if (setup_profiling_timer(multiplier))
541                         return -EINVAL;
542         }
543 #endif
544         profile_discard_flip_buffers();
545         memset(prof_buffer, 0, prof_len * sizeof(atomic_t));
546         return count;
547 }
548
549 static const struct file_operations proc_profile_operations = {
550         .read           = read_profile,
551         .write          = write_profile,
552 };
553
554 #ifdef CONFIG_SMP
555 static void profile_nop(void *unused)
556 {
557 }
558
559 static int create_hash_tables(void)
560 {
561         int cpu;
562
563         for_each_online_cpu(cpu) {
564                 int node = cpu_to_node(cpu);
565                 struct page *page;
566
567                 page = alloc_pages_exact_node(node,
568                                 GFP_KERNEL | __GFP_ZERO | GFP_THISNODE,
569                                 0);
570                 if (!page)
571                         goto out_cleanup;
572                 per_cpu(cpu_profile_hits, cpu)[1]
573                                 = (struct profile_hit *)page_address(page);
574                 page = alloc_pages_exact_node(node,
575                                 GFP_KERNEL | __GFP_ZERO | GFP_THISNODE,
576                                 0);
577                 if (!page)
578                         goto out_cleanup;
579                 per_cpu(cpu_profile_hits, cpu)[0]
580                                 = (struct profile_hit *)page_address(page);
581         }
582         return 0;
583 out_cleanup:
584         prof_on = 0;
585         smp_mb();
586         on_each_cpu(profile_nop, NULL, 1);
587         for_each_online_cpu(cpu) {
588                 struct page *page;
589
590                 if (per_cpu(cpu_profile_hits, cpu)[0]) {
591                         page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[0]);
592                         per_cpu(cpu_profile_hits, cpu)[0] = NULL;
593                         __free_page(page);
594                 }
595                 if (per_cpu(cpu_profile_hits, cpu)[1]) {
596                         page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
597                         per_cpu(cpu_profile_hits, cpu)[1] = NULL;
598                         __free_page(page);
599                 }
600         }
601         return -1;
602 }
603 #else
604 #define create_hash_tables()                    ({ 0; })
605 #endif
606
607 int __ref create_proc_profile(void) /* false positive from hotcpu_notifier */
608 {
609         struct proc_dir_entry *entry;
610
611         if (!prof_on)
612                 return 0;
613         if (create_hash_tables())
614                 return -ENOMEM;
615         entry = proc_create("profile", S_IWUSR | S_IRUGO,
616                             NULL, &proc_profile_operations);
617         if (!entry)
618                 return 0;
619         entry->size = (1+prof_len) * sizeof(atomic_t);
620         hotcpu_notifier(profile_cpu_callback, 0);
621         return 0;
622 }
623 module_init(create_proc_profile);
624 #endif /* CONFIG_PROC_FS */