[PATCH] Profiling: require buffer allocation on the correct node
[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
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 *) __read_mostly;
39
40 static atomic_t *prof_buffer;
41 static unsigned long prof_len, prof_shift;
42 static int prof_on __read_mostly;
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 DEFINE_MUTEX(profile_flip_mutex);
48 #endif /* CONFIG_SMP */
49
50 static int __init profile_setup(char * str)
51 {
52         static char __initdata schedstr[] = "schedule";
53         int par;
54
55         if (!strncmp(str, schedstr, strlen(schedstr))) {
56                 prof_on = SCHED_PROFILING;
57                 if (str[strlen(schedstr)] == ',')
58                         str += strlen(schedstr) + 1;
59                 if (get_option(&str, &par))
60                         prof_shift = par;
61                 printk(KERN_INFO
62                         "kernel schedule profiling enabled (shift: %ld)\n",
63                         prof_shift);
64         } else if (get_option(&str, &par)) {
65                 prof_shift = par;
66                 prof_on = CPU_PROFILING;
67                 printk(KERN_INFO "kernel profiling enabled (shift: %ld)\n",
68                         prof_shift);
69         }
70         return 1;
71 }
72 __setup("profile=", profile_setup);
73
74
75 void __init profile_init(void)
76 {
77         if (!prof_on) 
78                 return;
79  
80         /* only text is profiled */
81         prof_len = (_etext - _stext) >> prof_shift;
82         prof_buffer = alloc_bootmem(prof_len*sizeof(atomic_t));
83 }
84
85 /* Profile event notifications */
86  
87 #ifdef CONFIG_PROFILING
88  
89 static BLOCKING_NOTIFIER_HEAD(task_exit_notifier);
90 static ATOMIC_NOTIFIER_HEAD(task_free_notifier);
91 static BLOCKING_NOTIFIER_HEAD(munmap_notifier);
92  
93 void profile_task_exit(struct task_struct * task)
94 {
95         blocking_notifier_call_chain(&task_exit_notifier, 0, task);
96 }
97  
98 int profile_handoff_task(struct task_struct * task)
99 {
100         int ret;
101         ret = atomic_notifier_call_chain(&task_free_notifier, 0, task);
102         return (ret == NOTIFY_OK) ? 1 : 0;
103 }
104
105 void profile_munmap(unsigned long addr)
106 {
107         blocking_notifier_call_chain(&munmap_notifier, 0, (void *)addr);
108 }
109
110 int task_handoff_register(struct notifier_block * n)
111 {
112         return atomic_notifier_chain_register(&task_free_notifier, n);
113 }
114
115 int task_handoff_unregister(struct notifier_block * n)
116 {
117         return atomic_notifier_chain_unregister(&task_free_notifier, n);
118 }
119
120 int profile_event_register(enum profile_type type, struct notifier_block * n)
121 {
122         int err = -EINVAL;
123  
124         switch (type) {
125                 case PROFILE_TASK_EXIT:
126                         err = blocking_notifier_chain_register(
127                                         &task_exit_notifier, n);
128                         break;
129                 case PROFILE_MUNMAP:
130                         err = blocking_notifier_chain_register(
131                                         &munmap_notifier, n);
132                         break;
133         }
134  
135         return err;
136 }
137
138  
139 int profile_event_unregister(enum profile_type type, struct notifier_block * n)
140 {
141         int err = -EINVAL;
142  
143         switch (type) {
144                 case PROFILE_TASK_EXIT:
145                         err = blocking_notifier_chain_unregister(
146                                         &task_exit_notifier, n);
147                         break;
148                 case PROFILE_MUNMAP:
149                         err = blocking_notifier_chain_unregister(
150                                         &munmap_notifier, n);
151                         break;
152         }
153
154         return err;
155 }
156
157 int register_timer_hook(int (*hook)(struct pt_regs *))
158 {
159         if (timer_hook)
160                 return -EBUSY;
161         timer_hook = hook;
162         return 0;
163 }
164
165 void unregister_timer_hook(int (*hook)(struct pt_regs *))
166 {
167         WARN_ON(hook != timer_hook);
168         timer_hook = NULL;
169         /* make sure all CPUs see the NULL hook */
170         synchronize_sched();  /* Allow ongoing interrupts to complete. */
171 }
172
173 EXPORT_SYMBOL_GPL(register_timer_hook);
174 EXPORT_SYMBOL_GPL(unregister_timer_hook);
175 EXPORT_SYMBOL_GPL(task_handoff_register);
176 EXPORT_SYMBOL_GPL(task_handoff_unregister);
177
178 #endif /* CONFIG_PROFILING */
179
180 EXPORT_SYMBOL_GPL(profile_event_register);
181 EXPORT_SYMBOL_GPL(profile_event_unregister);
182
183 #ifdef CONFIG_SMP
184 /*
185  * Each cpu has a pair of open-addressed hashtables for pending
186  * profile hits. read_profile() IPI's all cpus to request them
187  * to flip buffers and flushes their contents to prof_buffer itself.
188  * Flip requests are serialized by the profile_flip_mutex. The sole
189  * use of having a second hashtable is for avoiding cacheline
190  * contention that would otherwise happen during flushes of pending
191  * profile hits required for the accuracy of reported profile hits
192  * and so resurrect the interrupt livelock issue.
193  *
194  * The open-addressed hashtables are indexed by profile buffer slot
195  * and hold the number of pending hits to that profile buffer slot on
196  * a cpu in an entry. When the hashtable overflows, all pending hits
197  * are accounted to their corresponding profile buffer slots with
198  * atomic_add() and the hashtable emptied. As numerous pending hits
199  * may be accounted to a profile buffer slot in a hashtable entry,
200  * this amortizes a number of atomic profile buffer increments likely
201  * to be far larger than the number of entries in the hashtable,
202  * particularly given that the number of distinct profile buffer
203  * positions to which hits are accounted during short intervals (e.g.
204  * several seconds) is usually very small. Exclusion from buffer
205  * flipping is provided by interrupt disablement (note that for
206  * SCHED_PROFILING profile_hit() may be called from process context).
207  * The hash function is meant to be lightweight as opposed to strong,
208  * and was vaguely inspired by ppc64 firmware-supported inverted
209  * pagetable hash functions, but uses a full hashtable full of finite
210  * collision chains, not just pairs of them.
211  *
212  * -- wli
213  */
214 static void __profile_flip_buffers(void *unused)
215 {
216         int cpu = smp_processor_id();
217
218         per_cpu(cpu_profile_flip, cpu) = !per_cpu(cpu_profile_flip, cpu);
219 }
220
221 static void profile_flip_buffers(void)
222 {
223         int i, j, cpu;
224
225         mutex_lock(&profile_flip_mutex);
226         j = per_cpu(cpu_profile_flip, get_cpu());
227         put_cpu();
228         on_each_cpu(__profile_flip_buffers, NULL, 0, 1);
229         for_each_online_cpu(cpu) {
230                 struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[j];
231                 for (i = 0; i < NR_PROFILE_HIT; ++i) {
232                         if (!hits[i].hits) {
233                                 if (hits[i].pc)
234                                         hits[i].pc = 0;
235                                 continue;
236                         }
237                         atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
238                         hits[i].hits = hits[i].pc = 0;
239                 }
240         }
241         mutex_unlock(&profile_flip_mutex);
242 }
243
244 static void profile_discard_flip_buffers(void)
245 {
246         int i, cpu;
247
248         mutex_lock(&profile_flip_mutex);
249         i = per_cpu(cpu_profile_flip, get_cpu());
250         put_cpu();
251         on_each_cpu(__profile_flip_buffers, NULL, 0, 1);
252         for_each_online_cpu(cpu) {
253                 struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[i];
254                 memset(hits, 0, NR_PROFILE_HIT*sizeof(struct profile_hit));
255         }
256         mutex_unlock(&profile_flip_mutex);
257 }
258
259 void profile_hit(int type, void *__pc)
260 {
261         unsigned long primary, secondary, flags, pc = (unsigned long)__pc;
262         int i, j, cpu;
263         struct profile_hit *hits;
264
265         if (prof_on != type || !prof_buffer)
266                 return;
267         pc = min((pc - (unsigned long)_stext) >> prof_shift, prof_len - 1);
268         i = primary = (pc & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
269         secondary = (~(pc << 1) & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
270         cpu = get_cpu();
271         hits = per_cpu(cpu_profile_hits, cpu)[per_cpu(cpu_profile_flip, cpu)];
272         if (!hits) {
273                 put_cpu();
274                 return;
275         }
276         local_irq_save(flags);
277         do {
278                 for (j = 0; j < PROFILE_GRPSZ; ++j) {
279                         if (hits[i + j].pc == pc) {
280                                 hits[i + j].hits++;
281                                 goto out;
282                         } else if (!hits[i + j].hits) {
283                                 hits[i + j].pc = pc;
284                                 hits[i + j].hits = 1;
285                                 goto out;
286                         }
287                 }
288                 i = (i + secondary) & (NR_PROFILE_HIT - 1);
289         } while (i != primary);
290         atomic_inc(&prof_buffer[pc]);
291         for (i = 0; i < NR_PROFILE_HIT; ++i) {
292                 atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
293                 hits[i].pc = hits[i].hits = 0;
294         }
295 out:
296         local_irq_restore(flags);
297         put_cpu();
298 }
299
300 #ifdef CONFIG_HOTPLUG_CPU
301 static int __devinit profile_cpu_callback(struct notifier_block *info,
302                                         unsigned long action, void *__cpu)
303 {
304         int node, cpu = (unsigned long)__cpu;
305         struct page *page;
306
307         switch (action) {
308         case CPU_UP_PREPARE:
309                 node = cpu_to_node(cpu);
310                 per_cpu(cpu_profile_flip, cpu) = 0;
311                 if (!per_cpu(cpu_profile_hits, cpu)[1]) {
312                         page = alloc_pages_node(node,
313                                         GFP_KERNEL | __GFP_ZERO | GFP_THISNODE,
314                                         0);
315                         if (!page)
316                                 return NOTIFY_BAD;
317                         per_cpu(cpu_profile_hits, cpu)[1] = page_address(page);
318                 }
319                 if (!per_cpu(cpu_profile_hits, cpu)[0]) {
320                         page = alloc_pages_node(node,
321                                         GFP_KERNEL | __GFP_ZERO | GFP_THISNODE,
322                                         0);
323                         if (!page)
324                                 goto out_free;
325                         per_cpu(cpu_profile_hits, cpu)[0] = page_address(page);
326                 }
327                 break;
328         out_free:
329                 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
330                 per_cpu(cpu_profile_hits, cpu)[1] = NULL;
331                 __free_page(page);
332                 return NOTIFY_BAD;
333         case CPU_ONLINE:
334                 cpu_set(cpu, prof_cpu_mask);
335                 break;
336         case CPU_UP_CANCELED:
337         case CPU_DEAD:
338                 cpu_clear(cpu, prof_cpu_mask);
339                 if (per_cpu(cpu_profile_hits, cpu)[0]) {
340                         page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[0]);
341                         per_cpu(cpu_profile_hits, cpu)[0] = NULL;
342                         __free_page(page);
343                 }
344                 if (per_cpu(cpu_profile_hits, cpu)[1]) {
345                         page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
346                         per_cpu(cpu_profile_hits, cpu)[1] = NULL;
347                         __free_page(page);
348                 }
349                 break;
350         }
351         return NOTIFY_OK;
352 }
353 #endif /* CONFIG_HOTPLUG_CPU */
354 #else /* !CONFIG_SMP */
355 #define profile_flip_buffers()          do { } while (0)
356 #define profile_discard_flip_buffers()  do { } while (0)
357
358 void profile_hit(int type, void *__pc)
359 {
360         unsigned long pc;
361
362         if (prof_on != type || !prof_buffer)
363                 return;
364         pc = ((unsigned long)__pc - (unsigned long)_stext) >> prof_shift;
365         atomic_inc(&prof_buffer[min(pc, prof_len - 1)]);
366 }
367 #endif /* !CONFIG_SMP */
368
369 void profile_tick(int type, struct pt_regs *regs)
370 {
371         if (type == CPU_PROFILING && timer_hook)
372                 timer_hook(regs);
373         if (!user_mode(regs) && cpu_isset(smp_processor_id(), prof_cpu_mask))
374                 profile_hit(type, (void *)profile_pc(regs));
375 }
376
377 #ifdef CONFIG_PROC_FS
378 #include <linux/proc_fs.h>
379 #include <asm/uaccess.h>
380 #include <asm/ptrace.h>
381
382 static int prof_cpu_mask_read_proc (char *page, char **start, off_t off,
383                         int count, int *eof, void *data)
384 {
385         int len = cpumask_scnprintf(page, count, *(cpumask_t *)data);
386         if (count - len < 2)
387                 return -EINVAL;
388         len += sprintf(page + len, "\n");
389         return len;
390 }
391
392 static int prof_cpu_mask_write_proc (struct file *file, const char __user *buffer,
393                                         unsigned long count, void *data)
394 {
395         cpumask_t *mask = (cpumask_t *)data;
396         unsigned long full_count = count, err;
397         cpumask_t new_value;
398
399         err = cpumask_parse(buffer, count, new_value);
400         if (err)
401                 return err;
402
403         *mask = new_value;
404         return full_count;
405 }
406
407 void create_prof_cpu_mask(struct proc_dir_entry *root_irq_dir)
408 {
409         struct proc_dir_entry *entry;
410
411         /* create /proc/irq/prof_cpu_mask */
412         if (!(entry = create_proc_entry("prof_cpu_mask", 0600, root_irq_dir)))
413                 return;
414         entry->nlink = 1;
415         entry->data = (void *)&prof_cpu_mask;
416         entry->read_proc = prof_cpu_mask_read_proc;
417         entry->write_proc = prof_cpu_mask_write_proc;
418 }
419
420 /*
421  * This function accesses profiling information. The returned data is
422  * binary: the sampling step and the actual contents of the profile
423  * buffer. Use of the program readprofile is recommended in order to
424  * get meaningful info out of these data.
425  */
426 static ssize_t
427 read_profile(struct file *file, char __user *buf, size_t count, loff_t *ppos)
428 {
429         unsigned long p = *ppos;
430         ssize_t read;
431         char * pnt;
432         unsigned int sample_step = 1 << prof_shift;
433
434         profile_flip_buffers();
435         if (p >= (prof_len+1)*sizeof(unsigned int))
436                 return 0;
437         if (count > (prof_len+1)*sizeof(unsigned int) - p)
438                 count = (prof_len+1)*sizeof(unsigned int) - p;
439         read = 0;
440
441         while (p < sizeof(unsigned int) && count > 0) {
442                 put_user(*((char *)(&sample_step)+p),buf);
443                 buf++; p++; count--; read++;
444         }
445         pnt = (char *)prof_buffer + p - sizeof(atomic_t);
446         if (copy_to_user(buf,(void *)pnt,count))
447                 return -EFAULT;
448         read += count;
449         *ppos += read;
450         return read;
451 }
452
453 /*
454  * Writing to /proc/profile resets the counters
455  *
456  * Writing a 'profiling multiplier' value into it also re-sets the profiling
457  * interrupt frequency, on architectures that support this.
458  */
459 static ssize_t write_profile(struct file *file, const char __user *buf,
460                              size_t count, loff_t *ppos)
461 {
462 #ifdef CONFIG_SMP
463         extern int setup_profiling_timer (unsigned int multiplier);
464
465         if (count == sizeof(int)) {
466                 unsigned int multiplier;
467
468                 if (copy_from_user(&multiplier, buf, sizeof(int)))
469                         return -EFAULT;
470
471                 if (setup_profiling_timer(multiplier))
472                         return -EINVAL;
473         }
474 #endif
475         profile_discard_flip_buffers();
476         memset(prof_buffer, 0, prof_len * sizeof(atomic_t));
477         return count;
478 }
479
480 static struct file_operations proc_profile_operations = {
481         .read           = read_profile,
482         .write          = write_profile,
483 };
484
485 #ifdef CONFIG_SMP
486 static void __init profile_nop(void *unused)
487 {
488 }
489
490 static int __init create_hash_tables(void)
491 {
492         int cpu;
493
494         for_each_online_cpu(cpu) {
495                 int node = cpu_to_node(cpu);
496                 struct page *page;
497
498                 page = alloc_pages_node(node,
499                                 GFP_KERNEL | __GFP_ZERO | GFP_THISNODE,
500                                 0);
501                 if (!page)
502                         goto out_cleanup;
503                 per_cpu(cpu_profile_hits, cpu)[1]
504                                 = (struct profile_hit *)page_address(page);
505                 page = alloc_pages_node(node,
506                                 GFP_KERNEL | __GFP_ZERO | GFP_THISNODE,
507                                 0);
508                 if (!page)
509                         goto out_cleanup;
510                 per_cpu(cpu_profile_hits, cpu)[0]
511                                 = (struct profile_hit *)page_address(page);
512         }
513         return 0;
514 out_cleanup:
515         prof_on = 0;
516         smp_mb();
517         on_each_cpu(profile_nop, NULL, 0, 1);
518         for_each_online_cpu(cpu) {
519                 struct page *page;
520
521                 if (per_cpu(cpu_profile_hits, cpu)[0]) {
522                         page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[0]);
523                         per_cpu(cpu_profile_hits, cpu)[0] = NULL;
524                         __free_page(page);
525                 }
526                 if (per_cpu(cpu_profile_hits, cpu)[1]) {
527                         page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
528                         per_cpu(cpu_profile_hits, cpu)[1] = NULL;
529                         __free_page(page);
530                 }
531         }
532         return -1;
533 }
534 #else
535 #define create_hash_tables()                    ({ 0; })
536 #endif
537
538 static int __init create_proc_profile(void)
539 {
540         struct proc_dir_entry *entry;
541
542         if (!prof_on)
543                 return 0;
544         if (create_hash_tables())
545                 return -1;
546         if (!(entry = create_proc_entry("profile", S_IWUSR | S_IRUGO, NULL)))
547                 return 0;
548         entry->proc_fops = &proc_profile_operations;
549         entry->size = (1+prof_len) * sizeof(atomic_t);
550         hotcpu_notifier(profile_cpu_callback, 0);
551         return 0;
552 }
553 module_init(create_proc_profile);
554 #endif /* CONFIG_PROC_FS */