Merge git://git.kernel.org/pub/scm/linux/kernel/git/davem/sparc-2.6
[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/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 #ifdef CONFIG_PROFILING
116
117 static BLOCKING_NOTIFIER_HEAD(task_exit_notifier);
118 static ATOMIC_NOTIFIER_HEAD(task_free_notifier);
119 static BLOCKING_NOTIFIER_HEAD(munmap_notifier);
120
121 void profile_task_exit(struct task_struct *task)
122 {
123         blocking_notifier_call_chain(&task_exit_notifier, 0, task);
124 }
125
126 int profile_handoff_task(struct task_struct *task)
127 {
128         int ret;
129         ret = atomic_notifier_call_chain(&task_free_notifier, 0, task);
130         return (ret == NOTIFY_OK) ? 1 : 0;
131 }
132
133 void profile_munmap(unsigned long addr)
134 {
135         blocking_notifier_call_chain(&munmap_notifier, 0, (void *)addr);
136 }
137
138 int task_handoff_register(struct notifier_block *n)
139 {
140         return atomic_notifier_chain_register(&task_free_notifier, n);
141 }
142 EXPORT_SYMBOL_GPL(task_handoff_register);
143
144 int task_handoff_unregister(struct notifier_block *n)
145 {
146         return atomic_notifier_chain_unregister(&task_free_notifier, n);
147 }
148 EXPORT_SYMBOL_GPL(task_handoff_unregister);
149
150 int profile_event_register(enum profile_type type, struct notifier_block *n)
151 {
152         int err = -EINVAL;
153
154         switch (type) {
155         case PROFILE_TASK_EXIT:
156                 err = blocking_notifier_chain_register(
157                                 &task_exit_notifier, n);
158                 break;
159         case PROFILE_MUNMAP:
160                 err = blocking_notifier_chain_register(
161                                 &munmap_notifier, n);
162                 break;
163         }
164
165         return err;
166 }
167 EXPORT_SYMBOL_GPL(profile_event_register);
168
169 int profile_event_unregister(enum profile_type type, struct notifier_block *n)
170 {
171         int err = -EINVAL;
172
173         switch (type) {
174         case PROFILE_TASK_EXIT:
175                 err = blocking_notifier_chain_unregister(
176                                 &task_exit_notifier, n);
177                 break;
178         case PROFILE_MUNMAP:
179                 err = blocking_notifier_chain_unregister(
180                                 &munmap_notifier, n);
181                 break;
182         }
183
184         return err;
185 }
186 EXPORT_SYMBOL_GPL(profile_event_unregister);
187
188 int register_timer_hook(int (*hook)(struct pt_regs *))
189 {
190         if (timer_hook)
191                 return -EBUSY;
192         timer_hook = hook;
193         return 0;
194 }
195 EXPORT_SYMBOL_GPL(register_timer_hook);
196
197 void unregister_timer_hook(int (*hook)(struct pt_regs *))
198 {
199         WARN_ON(hook != timer_hook);
200         timer_hook = NULL;
201         /* make sure all CPUs see the NULL hook */
202         synchronize_sched();  /* Allow ongoing interrupts to complete. */
203 }
204 EXPORT_SYMBOL_GPL(unregister_timer_hook);
205
206 #endif /* CONFIG_PROFILING */
207
208
209 #ifdef CONFIG_SMP
210 /*
211  * Each cpu has a pair of open-addressed hashtables for pending
212  * profile hits. read_profile() IPI's all cpus to request them
213  * to flip buffers and flushes their contents to prof_buffer itself.
214  * Flip requests are serialized by the profile_flip_mutex. The sole
215  * use of having a second hashtable is for avoiding cacheline
216  * contention that would otherwise happen during flushes of pending
217  * profile hits required for the accuracy of reported profile hits
218  * and so resurrect the interrupt livelock issue.
219  *
220  * The open-addressed hashtables are indexed by profile buffer slot
221  * and hold the number of pending hits to that profile buffer slot on
222  * a cpu in an entry. When the hashtable overflows, all pending hits
223  * are accounted to their corresponding profile buffer slots with
224  * atomic_add() and the hashtable emptied. As numerous pending hits
225  * may be accounted to a profile buffer slot in a hashtable entry,
226  * this amortizes a number of atomic profile buffer increments likely
227  * to be far larger than the number of entries in the hashtable,
228  * particularly given that the number of distinct profile buffer
229  * positions to which hits are accounted during short intervals (e.g.
230  * several seconds) is usually very small. Exclusion from buffer
231  * flipping is provided by interrupt disablement (note that for
232  * SCHED_PROFILING or SLEEP_PROFILING profile_hit() may be called from
233  * process context).
234  * The hash function is meant to be lightweight as opposed to strong,
235  * and was vaguely inspired by ppc64 firmware-supported inverted
236  * pagetable hash functions, but uses a full hashtable full of finite
237  * collision chains, not just pairs of them.
238  *
239  * -- wli
240  */
241 static void __profile_flip_buffers(void *unused)
242 {
243         int cpu = smp_processor_id();
244
245         per_cpu(cpu_profile_flip, cpu) = !per_cpu(cpu_profile_flip, cpu);
246 }
247
248 static void profile_flip_buffers(void)
249 {
250         int i, j, cpu;
251
252         mutex_lock(&profile_flip_mutex);
253         j = per_cpu(cpu_profile_flip, get_cpu());
254         put_cpu();
255         on_each_cpu(__profile_flip_buffers, NULL, 0, 1);
256         for_each_online_cpu(cpu) {
257                 struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[j];
258                 for (i = 0; i < NR_PROFILE_HIT; ++i) {
259                         if (!hits[i].hits) {
260                                 if (hits[i].pc)
261                                         hits[i].pc = 0;
262                                 continue;
263                         }
264                         atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
265                         hits[i].hits = hits[i].pc = 0;
266                 }
267         }
268         mutex_unlock(&profile_flip_mutex);
269 }
270
271 static void profile_discard_flip_buffers(void)
272 {
273         int i, cpu;
274
275         mutex_lock(&profile_flip_mutex);
276         i = per_cpu(cpu_profile_flip, get_cpu());
277         put_cpu();
278         on_each_cpu(__profile_flip_buffers, NULL, 0, 1);
279         for_each_online_cpu(cpu) {
280                 struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[i];
281                 memset(hits, 0, NR_PROFILE_HIT*sizeof(struct profile_hit));
282         }
283         mutex_unlock(&profile_flip_mutex);
284 }
285
286 void profile_hits(int type, void *__pc, unsigned int nr_hits)
287 {
288         unsigned long primary, secondary, flags, pc = (unsigned long)__pc;
289         int i, j, cpu;
290         struct profile_hit *hits;
291
292         if (prof_on != type || !prof_buffer)
293                 return;
294         pc = min((pc - (unsigned long)_stext) >> prof_shift, prof_len - 1);
295         i = primary = (pc & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
296         secondary = (~(pc << 1) & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
297         cpu = get_cpu();
298         hits = per_cpu(cpu_profile_hits, cpu)[per_cpu(cpu_profile_flip, cpu)];
299         if (!hits) {
300                 put_cpu();
301                 return;
302         }
303         /*
304          * We buffer the global profiler buffer into a per-CPU
305          * queue and thus reduce the number of global (and possibly
306          * NUMA-alien) accesses. The write-queue is self-coalescing:
307          */
308         local_irq_save(flags);
309         do {
310                 for (j = 0; j < PROFILE_GRPSZ; ++j) {
311                         if (hits[i + j].pc == pc) {
312                                 hits[i + j].hits += nr_hits;
313                                 goto out;
314                         } else if (!hits[i + j].hits) {
315                                 hits[i + j].pc = pc;
316                                 hits[i + j].hits = nr_hits;
317                                 goto out;
318                         }
319                 }
320                 i = (i + secondary) & (NR_PROFILE_HIT - 1);
321         } while (i != primary);
322
323         /*
324          * Add the current hit(s) and flush the write-queue out
325          * to the global buffer:
326          */
327         atomic_add(nr_hits, &prof_buffer[pc]);
328         for (i = 0; i < NR_PROFILE_HIT; ++i) {
329                 atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
330                 hits[i].pc = hits[i].hits = 0;
331         }
332 out:
333         local_irq_restore(flags);
334         put_cpu();
335 }
336
337 static int __devinit profile_cpu_callback(struct notifier_block *info,
338                                         unsigned long action, void *__cpu)
339 {
340         int node, cpu = (unsigned long)__cpu;
341         struct page *page;
342
343         switch (action) {
344         case CPU_UP_PREPARE:
345         case CPU_UP_PREPARE_FROZEN:
346                 node = cpu_to_node(cpu);
347                 per_cpu(cpu_profile_flip, cpu) = 0;
348                 if (!per_cpu(cpu_profile_hits, cpu)[1]) {
349                         page = alloc_pages_node(node,
350                                         GFP_KERNEL | __GFP_ZERO,
351                                         0);
352                         if (!page)
353                                 return NOTIFY_BAD;
354                         per_cpu(cpu_profile_hits, cpu)[1] = page_address(page);
355                 }
356                 if (!per_cpu(cpu_profile_hits, cpu)[0]) {
357                         page = alloc_pages_node(node,
358                                         GFP_KERNEL | __GFP_ZERO,
359                                         0);
360                         if (!page)
361                                 goto out_free;
362                         per_cpu(cpu_profile_hits, cpu)[0] = page_address(page);
363                 }
364                 break;
365 out_free:
366                 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
367                 per_cpu(cpu_profile_hits, cpu)[1] = NULL;
368                 __free_page(page);
369                 return NOTIFY_BAD;
370         case CPU_ONLINE:
371         case CPU_ONLINE_FROZEN:
372                 cpu_set(cpu, prof_cpu_mask);
373                 break;
374         case CPU_UP_CANCELED:
375         case CPU_UP_CANCELED_FROZEN:
376         case CPU_DEAD:
377         case CPU_DEAD_FROZEN:
378                 cpu_clear(cpu, prof_cpu_mask);
379                 if (per_cpu(cpu_profile_hits, cpu)[0]) {
380                         page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[0]);
381                         per_cpu(cpu_profile_hits, cpu)[0] = NULL;
382                         __free_page(page);
383                 }
384                 if (per_cpu(cpu_profile_hits, cpu)[1]) {
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                 }
389                 break;
390         }
391         return NOTIFY_OK;
392 }
393 #else /* !CONFIG_SMP */
394 #define profile_flip_buffers()          do { } while (0)
395 #define profile_discard_flip_buffers()  do { } while (0)
396 #define profile_cpu_callback            NULL
397
398 void profile_hits(int type, void *__pc, unsigned int nr_hits)
399 {
400         unsigned long pc;
401
402         if (prof_on != type || !prof_buffer)
403                 return;
404         pc = ((unsigned long)__pc - (unsigned long)_stext) >> prof_shift;
405         atomic_add(nr_hits, &prof_buffer[min(pc, prof_len - 1)]);
406 }
407 #endif /* !CONFIG_SMP */
408 EXPORT_SYMBOL_GPL(profile_hits);
409
410 void profile_tick(int type)
411 {
412         struct pt_regs *regs = get_irq_regs();
413
414         if (type == CPU_PROFILING && timer_hook)
415                 timer_hook(regs);
416         if (!user_mode(regs) && cpu_isset(smp_processor_id(), prof_cpu_mask))
417                 profile_hit(type, (void *)profile_pc(regs));
418 }
419
420 #ifdef CONFIG_PROC_FS
421 #include <linux/proc_fs.h>
422 #include <asm/uaccess.h>
423 #include <asm/ptrace.h>
424
425 static int prof_cpu_mask_read_proc(char *page, char **start, off_t off,
426                         int count, int *eof, void *data)
427 {
428         int len = cpumask_scnprintf(page, count, *(cpumask_t *)data);
429         if (count - len < 2)
430                 return -EINVAL;
431         len += sprintf(page + len, "\n");
432         return len;
433 }
434
435 static int prof_cpu_mask_write_proc(struct file *file,
436         const char __user *buffer,  unsigned long count, void *data)
437 {
438         cpumask_t *mask = (cpumask_t *)data;
439         unsigned long full_count = count, err;
440         cpumask_t new_value;
441
442         err = cpumask_parse_user(buffer, count, new_value);
443         if (err)
444                 return err;
445
446         *mask = new_value;
447         return full_count;
448 }
449
450 void create_prof_cpu_mask(struct proc_dir_entry *root_irq_dir)
451 {
452         struct proc_dir_entry *entry;
453
454         /* create /proc/irq/prof_cpu_mask */
455         entry = create_proc_entry("prof_cpu_mask", 0600, root_irq_dir);
456         if (!entry)
457                 return;
458         entry->data = (void *)&prof_cpu_mask;
459         entry->read_proc = prof_cpu_mask_read_proc;
460         entry->write_proc = prof_cpu_mask_write_proc;
461 }
462
463 /*
464  * This function accesses profiling information. The returned data is
465  * binary: the sampling step and the actual contents of the profile
466  * buffer. Use of the program readprofile is recommended in order to
467  * get meaningful info out of these data.
468  */
469 static ssize_t
470 read_profile(struct file *file, char __user *buf, size_t count, loff_t *ppos)
471 {
472         unsigned long p = *ppos;
473         ssize_t read;
474         char *pnt;
475         unsigned int sample_step = 1 << prof_shift;
476
477         profile_flip_buffers();
478         if (p >= (prof_len+1)*sizeof(unsigned int))
479                 return 0;
480         if (count > (prof_len+1)*sizeof(unsigned int) - p)
481                 count = (prof_len+1)*sizeof(unsigned int) - p;
482         read = 0;
483
484         while (p < sizeof(unsigned int) && count > 0) {
485                 if (put_user(*((char *)(&sample_step)+p), buf))
486                         return -EFAULT;
487                 buf++; p++; count--; read++;
488         }
489         pnt = (char *)prof_buffer + p - sizeof(atomic_t);
490         if (copy_to_user(buf, (void *)pnt, count))
491                 return -EFAULT;
492         read += count;
493         *ppos += read;
494         return read;
495 }
496
497 /*
498  * Writing to /proc/profile resets the counters
499  *
500  * Writing a 'profiling multiplier' value into it also re-sets the profiling
501  * interrupt frequency, on architectures that support this.
502  */
503 static ssize_t write_profile(struct file *file, const char __user *buf,
504                              size_t count, loff_t *ppos)
505 {
506 #ifdef CONFIG_SMP
507         extern int setup_profiling_timer(unsigned int multiplier);
508
509         if (count == sizeof(int)) {
510                 unsigned int multiplier;
511
512                 if (copy_from_user(&multiplier, buf, sizeof(int)))
513                         return -EFAULT;
514
515                 if (setup_profiling_timer(multiplier))
516                         return -EINVAL;
517         }
518 #endif
519         profile_discard_flip_buffers();
520         memset(prof_buffer, 0, prof_len * sizeof(atomic_t));
521         return count;
522 }
523
524 static const struct file_operations proc_profile_operations = {
525         .read           = read_profile,
526         .write          = write_profile,
527 };
528
529 #ifdef CONFIG_SMP
530 static void __init profile_nop(void *unused)
531 {
532 }
533
534 static int __init create_hash_tables(void)
535 {
536         int cpu;
537
538         for_each_online_cpu(cpu) {
539                 int node = cpu_to_node(cpu);
540                 struct page *page;
541
542                 page = alloc_pages_node(node,
543                                 GFP_KERNEL | __GFP_ZERO | GFP_THISNODE,
544                                 0);
545                 if (!page)
546                         goto out_cleanup;
547                 per_cpu(cpu_profile_hits, cpu)[1]
548                                 = (struct profile_hit *)page_address(page);
549                 page = alloc_pages_node(node,
550                                 GFP_KERNEL | __GFP_ZERO | GFP_THISNODE,
551                                 0);
552                 if (!page)
553                         goto out_cleanup;
554                 per_cpu(cpu_profile_hits, cpu)[0]
555                                 = (struct profile_hit *)page_address(page);
556         }
557         return 0;
558 out_cleanup:
559         prof_on = 0;
560         smp_mb();
561         on_each_cpu(profile_nop, NULL, 0, 1);
562         for_each_online_cpu(cpu) {
563                 struct page *page;
564
565                 if (per_cpu(cpu_profile_hits, cpu)[0]) {
566                         page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[0]);
567                         per_cpu(cpu_profile_hits, cpu)[0] = NULL;
568                         __free_page(page);
569                 }
570                 if (per_cpu(cpu_profile_hits, cpu)[1]) {
571                         page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
572                         per_cpu(cpu_profile_hits, cpu)[1] = NULL;
573                         __free_page(page);
574                 }
575         }
576         return -1;
577 }
578 #else
579 #define create_hash_tables()                    ({ 0; })
580 #endif
581
582 static int __init create_proc_profile(void)
583 {
584         struct proc_dir_entry *entry;
585
586         if (!prof_on)
587                 return 0;
588         if (create_hash_tables())
589                 return -1;
590         entry = proc_create("profile", S_IWUSR | S_IRUGO,
591                             NULL, &proc_profile_operations);
592         if (!entry)
593                 return 0;
594         entry->size = (1+prof_len) * sizeof(atomic_t);
595         hotcpu_notifier(profile_cpu_callback, 0);
596         return 0;
597 }
598 module_init(create_proc_profile);
599 #endif /* CONFIG_PROC_FS */