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