Memoryless nodes: Allow profiling data to fall back to other nodes
[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/profile.h>
24 #include <linux/highmem.h>
25 #include <linux/mutex.h>
26 #include <asm/sections.h>
27 #include <asm/semaphore.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 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_t prof_cpu_mask = CPU_MASK_ALL;
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 static int __init profile_setup(char * str)
56 {
57         static char __initdata schedstr[] = "schedule";
58         static char __initdata sleepstr[] = "sleep";
59         static char __initdata kvmstr[] = "kvm";
60         int par;
61
62         if (!strncmp(str, sleepstr, strlen(sleepstr))) {
63                 prof_on = SLEEP_PROFILING;
64                 if (str[strlen(sleepstr)] == ',')
65                         str += strlen(sleepstr) + 1;
66                 if (get_option(&str, &par))
67                         prof_shift = par;
68                 printk(KERN_INFO
69                         "kernel sleep profiling enabled (shift: %ld)\n",
70                         prof_shift);
71         } else if (!strncmp(str, schedstr, strlen(schedstr))) {
72                 prof_on = SCHED_PROFILING;
73                 if (str[strlen(schedstr)] == ',')
74                         str += strlen(schedstr) + 1;
75                 if (get_option(&str, &par))
76                         prof_shift = par;
77                 printk(KERN_INFO
78                         "kernel schedule profiling enabled (shift: %ld)\n",
79                         prof_shift);
80         } else if (!strncmp(str, kvmstr, strlen(kvmstr))) {
81                 prof_on = KVM_PROFILING;
82                 if (str[strlen(kvmstr)] == ',')
83                         str += strlen(kvmstr) + 1;
84                 if (get_option(&str, &par))
85                         prof_shift = par;
86                 printk(KERN_INFO
87                         "kernel KVM profiling enabled (shift: %ld)\n",
88                         prof_shift);
89         } else if (get_option(&str, &par)) {
90                 prof_shift = par;
91                 prof_on = CPU_PROFILING;
92                 printk(KERN_INFO "kernel profiling enabled (shift: %ld)\n",
93                         prof_shift);
94         }
95         return 1;
96 }
97 __setup("profile=", profile_setup);
98
99
100 void __init profile_init(void)
101 {
102         if (!prof_on) 
103                 return;
104  
105         /* only text is profiled */
106         prof_len = (_etext - _stext) >> prof_shift;
107         prof_buffer = alloc_bootmem(prof_len*sizeof(atomic_t));
108 }
109
110 /* Profile event notifications */
111  
112 #ifdef CONFIG_PROFILING
113  
114 static BLOCKING_NOTIFIER_HEAD(task_exit_notifier);
115 static ATOMIC_NOTIFIER_HEAD(task_free_notifier);
116 static BLOCKING_NOTIFIER_HEAD(munmap_notifier);
117  
118 void profile_task_exit(struct task_struct * task)
119 {
120         blocking_notifier_call_chain(&task_exit_notifier, 0, task);
121 }
122  
123 int profile_handoff_task(struct task_struct * task)
124 {
125         int ret;
126         ret = atomic_notifier_call_chain(&task_free_notifier, 0, task);
127         return (ret == NOTIFY_OK) ? 1 : 0;
128 }
129
130 void profile_munmap(unsigned long addr)
131 {
132         blocking_notifier_call_chain(&munmap_notifier, 0, (void *)addr);
133 }
134
135 int task_handoff_register(struct notifier_block * n)
136 {
137         return atomic_notifier_chain_register(&task_free_notifier, n);
138 }
139
140 int task_handoff_unregister(struct notifier_block * n)
141 {
142         return atomic_notifier_chain_unregister(&task_free_notifier, n);
143 }
144
145 int profile_event_register(enum profile_type type, struct notifier_block * n)
146 {
147         int err = -EINVAL;
148  
149         switch (type) {
150                 case PROFILE_TASK_EXIT:
151                         err = blocking_notifier_chain_register(
152                                         &task_exit_notifier, n);
153                         break;
154                 case PROFILE_MUNMAP:
155                         err = blocking_notifier_chain_register(
156                                         &munmap_notifier, n);
157                         break;
158         }
159  
160         return err;
161 }
162
163  
164 int profile_event_unregister(enum profile_type type, struct notifier_block * n)
165 {
166         int err = -EINVAL;
167  
168         switch (type) {
169                 case PROFILE_TASK_EXIT:
170                         err = blocking_notifier_chain_unregister(
171                                         &task_exit_notifier, n);
172                         break;
173                 case PROFILE_MUNMAP:
174                         err = blocking_notifier_chain_unregister(
175                                         &munmap_notifier, n);
176                         break;
177         }
178
179         return err;
180 }
181
182 int register_timer_hook(int (*hook)(struct pt_regs *))
183 {
184         if (timer_hook)
185                 return -EBUSY;
186         timer_hook = hook;
187         return 0;
188 }
189
190 void unregister_timer_hook(int (*hook)(struct pt_regs *))
191 {
192         WARN_ON(hook != timer_hook);
193         timer_hook = NULL;
194         /* make sure all CPUs see the NULL hook */
195         synchronize_sched();  /* Allow ongoing interrupts to complete. */
196 }
197
198 EXPORT_SYMBOL_GPL(register_timer_hook);
199 EXPORT_SYMBOL_GPL(unregister_timer_hook);
200 EXPORT_SYMBOL_GPL(task_handoff_register);
201 EXPORT_SYMBOL_GPL(task_handoff_unregister);
202 EXPORT_SYMBOL_GPL(profile_event_register);
203 EXPORT_SYMBOL_GPL(profile_event_unregister);
204
205 #endif /* CONFIG_PROFILING */
206
207
208 #ifdef CONFIG_SMP
209 /*
210  * Each cpu has a pair of open-addressed hashtables for pending
211  * profile hits. read_profile() IPI's all cpus to request them
212  * to flip buffers and flushes their contents to prof_buffer itself.
213  * Flip requests are serialized by the profile_flip_mutex. The sole
214  * use of having a second hashtable is for avoiding cacheline
215  * contention that would otherwise happen during flushes of pending
216  * profile hits required for the accuracy of reported profile hits
217  * and so resurrect the interrupt livelock issue.
218  *
219  * The open-addressed hashtables are indexed by profile buffer slot
220  * and hold the number of pending hits to that profile buffer slot on
221  * a cpu in an entry. When the hashtable overflows, all pending hits
222  * are accounted to their corresponding profile buffer slots with
223  * atomic_add() and the hashtable emptied. As numerous pending hits
224  * may be accounted to a profile buffer slot in a hashtable entry,
225  * this amortizes a number of atomic profile buffer increments likely
226  * to be far larger than the number of entries in the hashtable,
227  * particularly given that the number of distinct profile buffer
228  * positions to which hits are accounted during short intervals (e.g.
229  * several seconds) is usually very small. Exclusion from buffer
230  * flipping is provided by interrupt disablement (note that for
231  * SCHED_PROFILING or SLEEP_PROFILING profile_hit() may be called from
232  * process context).
233  * The hash function is meant to be lightweight as opposed to strong,
234  * and was vaguely inspired by ppc64 firmware-supported inverted
235  * pagetable hash functions, but uses a full hashtable full of finite
236  * collision chains, not just pairs of them.
237  *
238  * -- wli
239  */
240 static void __profile_flip_buffers(void *unused)
241 {
242         int cpu = smp_processor_id();
243
244         per_cpu(cpu_profile_flip, cpu) = !per_cpu(cpu_profile_flip, cpu);
245 }
246
247 static void profile_flip_buffers(void)
248 {
249         int i, j, cpu;
250
251         mutex_lock(&profile_flip_mutex);
252         j = per_cpu(cpu_profile_flip, get_cpu());
253         put_cpu();
254         on_each_cpu(__profile_flip_buffers, NULL, 0, 1);
255         for_each_online_cpu(cpu) {
256                 struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[j];
257                 for (i = 0; i < NR_PROFILE_HIT; ++i) {
258                         if (!hits[i].hits) {
259                                 if (hits[i].pc)
260                                         hits[i].pc = 0;
261                                 continue;
262                         }
263                         atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
264                         hits[i].hits = hits[i].pc = 0;
265                 }
266         }
267         mutex_unlock(&profile_flip_mutex);
268 }
269
270 static void profile_discard_flip_buffers(void)
271 {
272         int i, cpu;
273
274         mutex_lock(&profile_flip_mutex);
275         i = per_cpu(cpu_profile_flip, get_cpu());
276         put_cpu();
277         on_each_cpu(__profile_flip_buffers, NULL, 0, 1);
278         for_each_online_cpu(cpu) {
279                 struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[i];
280                 memset(hits, 0, NR_PROFILE_HIT*sizeof(struct profile_hit));
281         }
282         mutex_unlock(&profile_flip_mutex);
283 }
284
285 void profile_hits(int type, void *__pc, unsigned int nr_hits)
286 {
287         unsigned long primary, secondary, flags, pc = (unsigned long)__pc;
288         int i, j, cpu;
289         struct profile_hit *hits;
290
291         if (prof_on != type || !prof_buffer)
292                 return;
293         pc = min((pc - (unsigned long)_stext) >> prof_shift, prof_len - 1);
294         i = primary = (pc & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
295         secondary = (~(pc << 1) & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
296         cpu = get_cpu();
297         hits = per_cpu(cpu_profile_hits, cpu)[per_cpu(cpu_profile_flip, cpu)];
298         if (!hits) {
299                 put_cpu();
300                 return;
301         }
302         /*
303          * We buffer the global profiler buffer into a per-CPU
304          * queue and thus reduce the number of global (and possibly
305          * NUMA-alien) accesses. The write-queue is self-coalescing:
306          */
307         local_irq_save(flags);
308         do {
309                 for (j = 0; j < PROFILE_GRPSZ; ++j) {
310                         if (hits[i + j].pc == pc) {
311                                 hits[i + j].hits += nr_hits;
312                                 goto out;
313                         } else if (!hits[i + j].hits) {
314                                 hits[i + j].pc = pc;
315                                 hits[i + j].hits = nr_hits;
316                                 goto out;
317                         }
318                 }
319                 i = (i + secondary) & (NR_PROFILE_HIT - 1);
320         } while (i != primary);
321
322         /*
323          * Add the current hit(s) and flush the write-queue out
324          * to the global buffer:
325          */
326         atomic_add(nr_hits, &prof_buffer[pc]);
327         for (i = 0; i < NR_PROFILE_HIT; ++i) {
328                 atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
329                 hits[i].pc = hits[i].hits = 0;
330         }
331 out:
332         local_irq_restore(flags);
333         put_cpu();
334 }
335
336 static int __devinit profile_cpu_callback(struct notifier_block *info,
337                                         unsigned long action, void *__cpu)
338 {
339         int node, cpu = (unsigned long)__cpu;
340         struct page *page;
341
342         switch (action) {
343         case CPU_UP_PREPARE:
344         case CPU_UP_PREPARE_FROZEN:
345                 node = cpu_to_node(cpu);
346                 per_cpu(cpu_profile_flip, cpu) = 0;
347                 if (!per_cpu(cpu_profile_hits, cpu)[1]) {
348                         page = alloc_pages_node(node,
349                                         GFP_KERNEL | __GFP_ZERO,
350                                         0);
351                         if (!page)
352                                 return NOTIFY_BAD;
353                         per_cpu(cpu_profile_hits, cpu)[1] = page_address(page);
354                 }
355                 if (!per_cpu(cpu_profile_hits, cpu)[0]) {
356                         page = alloc_pages_node(node,
357                                         GFP_KERNEL | __GFP_ZERO,
358                                         0);
359                         if (!page)
360                                 goto out_free;
361                         per_cpu(cpu_profile_hits, cpu)[0] = page_address(page);
362                 }
363                 break;
364         out_free:
365                 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
366                 per_cpu(cpu_profile_hits, cpu)[1] = NULL;
367                 __free_page(page);
368                 return NOTIFY_BAD;
369         case CPU_ONLINE:
370         case CPU_ONLINE_FROZEN:
371                 cpu_set(cpu, prof_cpu_mask);
372                 break;
373         case CPU_UP_CANCELED:
374         case CPU_UP_CANCELED_FROZEN:
375         case CPU_DEAD:
376         case CPU_DEAD_FROZEN:
377                 cpu_clear(cpu, prof_cpu_mask);
378                 if (per_cpu(cpu_profile_hits, cpu)[0]) {
379                         page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[0]);
380                         per_cpu(cpu_profile_hits, cpu)[0] = NULL;
381                         __free_page(page);
382                 }
383                 if (per_cpu(cpu_profile_hits, cpu)[1]) {
384                         page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
385                         per_cpu(cpu_profile_hits, cpu)[1] = NULL;
386                         __free_page(page);
387                 }
388                 break;
389         }
390         return NOTIFY_OK;
391 }
392 #else /* !CONFIG_SMP */
393 #define profile_flip_buffers()          do { } while (0)
394 #define profile_discard_flip_buffers()  do { } while (0)
395 #define profile_cpu_callback            NULL
396
397 void profile_hits(int type, void *__pc, unsigned int nr_hits)
398 {
399         unsigned long pc;
400
401         if (prof_on != type || !prof_buffer)
402                 return;
403         pc = ((unsigned long)__pc - (unsigned long)_stext) >> prof_shift;
404         atomic_add(nr_hits, &prof_buffer[min(pc, prof_len - 1)]);
405 }
406 #endif /* !CONFIG_SMP */
407
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, const char __user *buffer,
436                                         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         if (!(entry = create_proc_entry("prof_cpu_mask", 0600, root_irq_dir)))
456                 return;
457         entry->data = (void *)&prof_cpu_mask;
458         entry->read_proc = prof_cpu_mask_read_proc;
459         entry->write_proc = prof_cpu_mask_write_proc;
460 }
461
462 /*
463  * This function accesses profiling information. The returned data is
464  * binary: the sampling step and the actual contents of the profile
465  * buffer. Use of the program readprofile is recommended in order to
466  * get meaningful info out of these data.
467  */
468 static ssize_t
469 read_profile(struct file *file, char __user *buf, size_t count, loff_t *ppos)
470 {
471         unsigned long p = *ppos;
472         ssize_t read;
473         char * pnt;
474         unsigned int sample_step = 1 << prof_shift;
475
476         profile_flip_buffers();
477         if (p >= (prof_len+1)*sizeof(unsigned int))
478                 return 0;
479         if (count > (prof_len+1)*sizeof(unsigned int) - p)
480                 count = (prof_len+1)*sizeof(unsigned int) - p;
481         read = 0;
482
483         while (p < sizeof(unsigned int) && count > 0) {
484                 if (put_user(*((char *)(&sample_step)+p),buf))
485                         return -EFAULT;
486                 buf++; p++; count--; read++;
487         }
488         pnt = (char *)prof_buffer + p - sizeof(atomic_t);
489         if (copy_to_user(buf,(void *)pnt,count))
490                 return -EFAULT;
491         read += count;
492         *ppos += read;
493         return read;
494 }
495
496 /*
497  * Writing to /proc/profile resets the counters
498  *
499  * Writing a 'profiling multiplier' value into it also re-sets the profiling
500  * interrupt frequency, on architectures that support this.
501  */
502 static ssize_t write_profile(struct file *file, const char __user *buf,
503                              size_t count, loff_t *ppos)
504 {
505 #ifdef CONFIG_SMP
506         extern int setup_profiling_timer (unsigned int multiplier);
507
508         if (count == sizeof(int)) {
509                 unsigned int multiplier;
510
511                 if (copy_from_user(&multiplier, buf, sizeof(int)))
512                         return -EFAULT;
513
514                 if (setup_profiling_timer(multiplier))
515                         return -EINVAL;
516         }
517 #endif
518         profile_discard_flip_buffers();
519         memset(prof_buffer, 0, prof_len * sizeof(atomic_t));
520         return count;
521 }
522
523 static const struct file_operations proc_profile_operations = {
524         .read           = read_profile,
525         .write          = write_profile,
526 };
527
528 #ifdef CONFIG_SMP
529 static void __init profile_nop(void *unused)
530 {
531 }
532
533 static int __init create_hash_tables(void)
534 {
535         int cpu;
536
537         for_each_online_cpu(cpu) {
538                 int node = cpu_to_node(cpu);
539                 struct page *page;
540
541                 page = alloc_pages_node(node,
542                                 GFP_KERNEL | __GFP_ZERO | GFP_THISNODE,
543                                 0);
544                 if (!page)
545                         goto out_cleanup;
546                 per_cpu(cpu_profile_hits, cpu)[1]
547                                 = (struct profile_hit *)page_address(page);
548                 page = alloc_pages_node(node,
549                                 GFP_KERNEL | __GFP_ZERO | GFP_THISNODE,
550                                 0);
551                 if (!page)
552                         goto out_cleanup;
553                 per_cpu(cpu_profile_hits, cpu)[0]
554                                 = (struct profile_hit *)page_address(page);
555         }
556         return 0;
557 out_cleanup:
558         prof_on = 0;
559         smp_mb();
560         on_each_cpu(profile_nop, NULL, 0, 1);
561         for_each_online_cpu(cpu) {
562                 struct page *page;
563
564                 if (per_cpu(cpu_profile_hits, cpu)[0]) {
565                         page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[0]);
566                         per_cpu(cpu_profile_hits, cpu)[0] = NULL;
567                         __free_page(page);
568                 }
569                 if (per_cpu(cpu_profile_hits, cpu)[1]) {
570                         page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
571                         per_cpu(cpu_profile_hits, cpu)[1] = NULL;
572                         __free_page(page);
573                 }
574         }
575         return -1;
576 }
577 #else
578 #define create_hash_tables()                    ({ 0; })
579 #endif
580
581 static int __init create_proc_profile(void)
582 {
583         struct proc_dir_entry *entry;
584
585         if (!prof_on)
586                 return 0;
587         if (create_hash_tables())
588                 return -1;
589         if (!(entry = create_proc_entry("profile", S_IWUSR | S_IRUGO, NULL)))
590                 return 0;
591         entry->proc_fops = &proc_profile_operations;
592         entry->size = (1+prof_len) * sizeof(atomic_t);
593         hotcpu_notifier(profile_cpu_callback, 0);
594         return 0;
595 }
596 module_init(create_proc_profile);
597 #endif /* CONFIG_PROC_FS */