[PATCH] Add lockless helpers for remove_suid()
[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/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 static int prof_on __read_mostly;
44 static cpumask_t prof_cpu_mask = CPU_MASK_ALL;
45 #ifdef CONFIG_SMP
46 static DEFINE_PER_CPU(struct profile_hit *[2], cpu_profile_hits);
47 static DEFINE_PER_CPU(int, cpu_profile_flip);
48 static DEFINE_MUTEX(profile_flip_mutex);
49 #endif /* CONFIG_SMP */
50
51 static int __init profile_setup(char * str)
52 {
53         static char __initdata schedstr[] = "schedule";
54         int par;
55
56         if (!strncmp(str, schedstr, strlen(schedstr))) {
57                 prof_on = SCHED_PROFILING;
58                 if (str[strlen(schedstr)] == ',')
59                         str += strlen(schedstr) + 1;
60                 if (get_option(&str, &par))
61                         prof_shift = par;
62                 printk(KERN_INFO
63                         "kernel schedule profiling enabled (shift: %ld)\n",
64                         prof_shift);
65         } else if (get_option(&str, &par)) {
66                 prof_shift = par;
67                 prof_on = CPU_PROFILING;
68                 printk(KERN_INFO "kernel profiling enabled (shift: %ld)\n",
69                         prof_shift);
70         }
71         return 1;
72 }
73 __setup("profile=", profile_setup);
74
75
76 void __init profile_init(void)
77 {
78         if (!prof_on) 
79                 return;
80  
81         /* only text is profiled */
82         prof_len = (_etext - _stext) >> prof_shift;
83         prof_buffer = alloc_bootmem(prof_len*sizeof(atomic_t));
84 }
85
86 /* Profile event notifications */
87  
88 #ifdef CONFIG_PROFILING
89  
90 static BLOCKING_NOTIFIER_HEAD(task_exit_notifier);
91 static ATOMIC_NOTIFIER_HEAD(task_free_notifier);
92 static BLOCKING_NOTIFIER_HEAD(munmap_notifier);
93  
94 void profile_task_exit(struct task_struct * task)
95 {
96         blocking_notifier_call_chain(&task_exit_notifier, 0, task);
97 }
98  
99 int profile_handoff_task(struct task_struct * task)
100 {
101         int ret;
102         ret = atomic_notifier_call_chain(&task_free_notifier, 0, task);
103         return (ret == NOTIFY_OK) ? 1 : 0;
104 }
105
106 void profile_munmap(unsigned long addr)
107 {
108         blocking_notifier_call_chain(&munmap_notifier, 0, (void *)addr);
109 }
110
111 int task_handoff_register(struct notifier_block * n)
112 {
113         return atomic_notifier_chain_register(&task_free_notifier, n);
114 }
115
116 int task_handoff_unregister(struct notifier_block * n)
117 {
118         return atomic_notifier_chain_unregister(&task_free_notifier, n);
119 }
120
121 int profile_event_register(enum profile_type type, struct notifier_block * n)
122 {
123         int err = -EINVAL;
124  
125         switch (type) {
126                 case PROFILE_TASK_EXIT:
127                         err = blocking_notifier_chain_register(
128                                         &task_exit_notifier, n);
129                         break;
130                 case PROFILE_MUNMAP:
131                         err = blocking_notifier_chain_register(
132                                         &munmap_notifier, n);
133                         break;
134         }
135  
136         return err;
137 }
138
139  
140 int profile_event_unregister(enum profile_type type, struct notifier_block * n)
141 {
142         int err = -EINVAL;
143  
144         switch (type) {
145                 case PROFILE_TASK_EXIT:
146                         err = blocking_notifier_chain_unregister(
147                                         &task_exit_notifier, n);
148                         break;
149                 case PROFILE_MUNMAP:
150                         err = blocking_notifier_chain_unregister(
151                                         &munmap_notifier, n);
152                         break;
153         }
154
155         return err;
156 }
157
158 int register_timer_hook(int (*hook)(struct pt_regs *))
159 {
160         if (timer_hook)
161                 return -EBUSY;
162         timer_hook = hook;
163         return 0;
164 }
165
166 void unregister_timer_hook(int (*hook)(struct pt_regs *))
167 {
168         WARN_ON(hook != timer_hook);
169         timer_hook = NULL;
170         /* make sure all CPUs see the NULL hook */
171         synchronize_sched();  /* Allow ongoing interrupts to complete. */
172 }
173
174 EXPORT_SYMBOL_GPL(register_timer_hook);
175 EXPORT_SYMBOL_GPL(unregister_timer_hook);
176 EXPORT_SYMBOL_GPL(task_handoff_register);
177 EXPORT_SYMBOL_GPL(task_handoff_unregister);
178
179 #endif /* CONFIG_PROFILING */
180
181 EXPORT_SYMBOL_GPL(profile_event_register);
182 EXPORT_SYMBOL_GPL(profile_event_unregister);
183
184 #ifdef CONFIG_SMP
185 /*
186  * Each cpu has a pair of open-addressed hashtables for pending
187  * profile hits. read_profile() IPI's all cpus to request them
188  * to flip buffers and flushes their contents to prof_buffer itself.
189  * Flip requests are serialized by the profile_flip_mutex. The sole
190  * use of having a second hashtable is for avoiding cacheline
191  * contention that would otherwise happen during flushes of pending
192  * profile hits required for the accuracy of reported profile hits
193  * and so resurrect the interrupt livelock issue.
194  *
195  * The open-addressed hashtables are indexed by profile buffer slot
196  * and hold the number of pending hits to that profile buffer slot on
197  * a cpu in an entry. When the hashtable overflows, all pending hits
198  * are accounted to their corresponding profile buffer slots with
199  * atomic_add() and the hashtable emptied. As numerous pending hits
200  * may be accounted to a profile buffer slot in a hashtable entry,
201  * this amortizes a number of atomic profile buffer increments likely
202  * to be far larger than the number of entries in the hashtable,
203  * particularly given that the number of distinct profile buffer
204  * positions to which hits are accounted during short intervals (e.g.
205  * several seconds) is usually very small. Exclusion from buffer
206  * flipping is provided by interrupt disablement (note that for
207  * SCHED_PROFILING profile_hit() may be called from process context).
208  * The hash function is meant to be lightweight as opposed to strong,
209  * and was vaguely inspired by ppc64 firmware-supported inverted
210  * pagetable hash functions, but uses a full hashtable full of finite
211  * collision chains, not just pairs of them.
212  *
213  * -- wli
214  */
215 static void __profile_flip_buffers(void *unused)
216 {
217         int cpu = smp_processor_id();
218
219         per_cpu(cpu_profile_flip, cpu) = !per_cpu(cpu_profile_flip, cpu);
220 }
221
222 static void profile_flip_buffers(void)
223 {
224         int i, j, cpu;
225
226         mutex_lock(&profile_flip_mutex);
227         j = per_cpu(cpu_profile_flip, get_cpu());
228         put_cpu();
229         on_each_cpu(__profile_flip_buffers, NULL, 0, 1);
230         for_each_online_cpu(cpu) {
231                 struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[j];
232                 for (i = 0; i < NR_PROFILE_HIT; ++i) {
233                         if (!hits[i].hits) {
234                                 if (hits[i].pc)
235                                         hits[i].pc = 0;
236                                 continue;
237                         }
238                         atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
239                         hits[i].hits = hits[i].pc = 0;
240                 }
241         }
242         mutex_unlock(&profile_flip_mutex);
243 }
244
245 static void profile_discard_flip_buffers(void)
246 {
247         int i, cpu;
248
249         mutex_lock(&profile_flip_mutex);
250         i = per_cpu(cpu_profile_flip, get_cpu());
251         put_cpu();
252         on_each_cpu(__profile_flip_buffers, NULL, 0, 1);
253         for_each_online_cpu(cpu) {
254                 struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[i];
255                 memset(hits, 0, NR_PROFILE_HIT*sizeof(struct profile_hit));
256         }
257         mutex_unlock(&profile_flip_mutex);
258 }
259
260 void profile_hit(int type, void *__pc)
261 {
262         unsigned long primary, secondary, flags, pc = (unsigned long)__pc;
263         int i, j, cpu;
264         struct profile_hit *hits;
265
266         if (prof_on != type || !prof_buffer)
267                 return;
268         pc = min((pc - (unsigned long)_stext) >> prof_shift, prof_len - 1);
269         i = primary = (pc & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
270         secondary = (~(pc << 1) & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
271         cpu = get_cpu();
272         hits = per_cpu(cpu_profile_hits, cpu)[per_cpu(cpu_profile_flip, cpu)];
273         if (!hits) {
274                 put_cpu();
275                 return;
276         }
277         local_irq_save(flags);
278         do {
279                 for (j = 0; j < PROFILE_GRPSZ; ++j) {
280                         if (hits[i + j].pc == pc) {
281                                 hits[i + j].hits++;
282                                 goto out;
283                         } else if (!hits[i + j].hits) {
284                                 hits[i + j].pc = pc;
285                                 hits[i + j].hits = 1;
286                                 goto out;
287                         }
288                 }
289                 i = (i + secondary) & (NR_PROFILE_HIT - 1);
290         } while (i != primary);
291         atomic_inc(&prof_buffer[pc]);
292         for (i = 0; i < NR_PROFILE_HIT; ++i) {
293                 atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
294                 hits[i].pc = hits[i].hits = 0;
295         }
296 out:
297         local_irq_restore(flags);
298         put_cpu();
299 }
300
301 #ifdef CONFIG_HOTPLUG_CPU
302 static int __devinit profile_cpu_callback(struct notifier_block *info,
303                                         unsigned long action, void *__cpu)
304 {
305         int node, cpu = (unsigned long)__cpu;
306         struct page *page;
307
308         switch (action) {
309         case CPU_UP_PREPARE:
310                 node = cpu_to_node(cpu);
311                 per_cpu(cpu_profile_flip, cpu) = 0;
312                 if (!per_cpu(cpu_profile_hits, cpu)[1]) {
313                         page = alloc_pages_node(node,
314                                         GFP_KERNEL | __GFP_ZERO | GFP_THISNODE,
315                                         0);
316                         if (!page)
317                                 return NOTIFY_BAD;
318                         per_cpu(cpu_profile_hits, cpu)[1] = page_address(page);
319                 }
320                 if (!per_cpu(cpu_profile_hits, cpu)[0]) {
321                         page = alloc_pages_node(node,
322                                         GFP_KERNEL | __GFP_ZERO | GFP_THISNODE,
323                                         0);
324                         if (!page)
325                                 goto out_free;
326                         per_cpu(cpu_profile_hits, cpu)[0] = page_address(page);
327                 }
328                 break;
329         out_free:
330                 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
331                 per_cpu(cpu_profile_hits, cpu)[1] = NULL;
332                 __free_page(page);
333                 return NOTIFY_BAD;
334         case CPU_ONLINE:
335                 cpu_set(cpu, prof_cpu_mask);
336                 break;
337         case CPU_UP_CANCELED:
338         case CPU_DEAD:
339                 cpu_clear(cpu, prof_cpu_mask);
340                 if (per_cpu(cpu_profile_hits, cpu)[0]) {
341                         page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[0]);
342                         per_cpu(cpu_profile_hits, cpu)[0] = NULL;
343                         __free_page(page);
344                 }
345                 if (per_cpu(cpu_profile_hits, cpu)[1]) {
346                         page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
347                         per_cpu(cpu_profile_hits, cpu)[1] = NULL;
348                         __free_page(page);
349                 }
350                 break;
351         }
352         return NOTIFY_OK;
353 }
354 #endif /* CONFIG_HOTPLUG_CPU */
355 #else /* !CONFIG_SMP */
356 #define profile_flip_buffers()          do { } while (0)
357 #define profile_discard_flip_buffers()  do { } while (0)
358
359 void profile_hit(int type, void *__pc)
360 {
361         unsigned long pc;
362
363         if (prof_on != type || !prof_buffer)
364                 return;
365         pc = ((unsigned long)__pc - (unsigned long)_stext) >> prof_shift;
366         atomic_inc(&prof_buffer[min(pc, prof_len - 1)]);
367 }
368 #endif /* !CONFIG_SMP */
369
370 void profile_tick(int type)
371 {
372         struct pt_regs *regs = get_irq_regs();
373
374         if (type == CPU_PROFILING && timer_hook)
375                 timer_hook(regs);
376         if (!user_mode(regs) && cpu_isset(smp_processor_id(), prof_cpu_mask))
377                 profile_hit(type, (void *)profile_pc(regs));
378 }
379
380 #ifdef CONFIG_PROC_FS
381 #include <linux/proc_fs.h>
382 #include <asm/uaccess.h>
383 #include <asm/ptrace.h>
384
385 static int prof_cpu_mask_read_proc (char *page, char **start, off_t off,
386                         int count, int *eof, void *data)
387 {
388         int len = cpumask_scnprintf(page, count, *(cpumask_t *)data);
389         if (count - len < 2)
390                 return -EINVAL;
391         len += sprintf(page + len, "\n");
392         return len;
393 }
394
395 static int prof_cpu_mask_write_proc (struct file *file, const char __user *buffer,
396                                         unsigned long count, void *data)
397 {
398         cpumask_t *mask = (cpumask_t *)data;
399         unsigned long full_count = count, err;
400         cpumask_t new_value;
401
402         err = cpumask_parse_user(buffer, count, new_value);
403         if (err)
404                 return err;
405
406         *mask = new_value;
407         return full_count;
408 }
409
410 void create_prof_cpu_mask(struct proc_dir_entry *root_irq_dir)
411 {
412         struct proc_dir_entry *entry;
413
414         /* create /proc/irq/prof_cpu_mask */
415         if (!(entry = create_proc_entry("prof_cpu_mask", 0600, root_irq_dir)))
416                 return;
417         entry->nlink = 1;
418         entry->data = (void *)&prof_cpu_mask;
419         entry->read_proc = prof_cpu_mask_read_proc;
420         entry->write_proc = prof_cpu_mask_write_proc;
421 }
422
423 /*
424  * This function accesses profiling information. The returned data is
425  * binary: the sampling step and the actual contents of the profile
426  * buffer. Use of the program readprofile is recommended in order to
427  * get meaningful info out of these data.
428  */
429 static ssize_t
430 read_profile(struct file *file, char __user *buf, size_t count, loff_t *ppos)
431 {
432         unsigned long p = *ppos;
433         ssize_t read;
434         char * pnt;
435         unsigned int sample_step = 1 << prof_shift;
436
437         profile_flip_buffers();
438         if (p >= (prof_len+1)*sizeof(unsigned int))
439                 return 0;
440         if (count > (prof_len+1)*sizeof(unsigned int) - p)
441                 count = (prof_len+1)*sizeof(unsigned int) - p;
442         read = 0;
443
444         while (p < sizeof(unsigned int) && count > 0) {
445                 put_user(*((char *)(&sample_step)+p),buf);
446                 buf++; p++; count--; read++;
447         }
448         pnt = (char *)prof_buffer + p - sizeof(atomic_t);
449         if (copy_to_user(buf,(void *)pnt,count))
450                 return -EFAULT;
451         read += count;
452         *ppos += read;
453         return read;
454 }
455
456 /*
457  * Writing to /proc/profile resets the counters
458  *
459  * Writing a 'profiling multiplier' value into it also re-sets the profiling
460  * interrupt frequency, on architectures that support this.
461  */
462 static ssize_t write_profile(struct file *file, const char __user *buf,
463                              size_t count, loff_t *ppos)
464 {
465 #ifdef CONFIG_SMP
466         extern int setup_profiling_timer (unsigned int multiplier);
467
468         if (count == sizeof(int)) {
469                 unsigned int multiplier;
470
471                 if (copy_from_user(&multiplier, buf, sizeof(int)))
472                         return -EFAULT;
473
474                 if (setup_profiling_timer(multiplier))
475                         return -EINVAL;
476         }
477 #endif
478         profile_discard_flip_buffers();
479         memset(prof_buffer, 0, prof_len * sizeof(atomic_t));
480         return count;
481 }
482
483 static struct file_operations proc_profile_operations = {
484         .read           = read_profile,
485         .write          = write_profile,
486 };
487
488 #ifdef CONFIG_SMP
489 static void __init profile_nop(void *unused)
490 {
491 }
492
493 static int __init create_hash_tables(void)
494 {
495         int cpu;
496
497         for_each_online_cpu(cpu) {
498                 int node = cpu_to_node(cpu);
499                 struct page *page;
500
501                 page = alloc_pages_node(node,
502                                 GFP_KERNEL | __GFP_ZERO | GFP_THISNODE,
503                                 0);
504                 if (!page)
505                         goto out_cleanup;
506                 per_cpu(cpu_profile_hits, cpu)[1]
507                                 = (struct profile_hit *)page_address(page);
508                 page = alloc_pages_node(node,
509                                 GFP_KERNEL | __GFP_ZERO | GFP_THISNODE,
510                                 0);
511                 if (!page)
512                         goto out_cleanup;
513                 per_cpu(cpu_profile_hits, cpu)[0]
514                                 = (struct profile_hit *)page_address(page);
515         }
516         return 0;
517 out_cleanup:
518         prof_on = 0;
519         smp_mb();
520         on_each_cpu(profile_nop, NULL, 0, 1);
521         for_each_online_cpu(cpu) {
522                 struct page *page;
523
524                 if (per_cpu(cpu_profile_hits, cpu)[0]) {
525                         page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[0]);
526                         per_cpu(cpu_profile_hits, cpu)[0] = NULL;
527                         __free_page(page);
528                 }
529                 if (per_cpu(cpu_profile_hits, cpu)[1]) {
530                         page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
531                         per_cpu(cpu_profile_hits, cpu)[1] = NULL;
532                         __free_page(page);
533                 }
534         }
535         return -1;
536 }
537 #else
538 #define create_hash_tables()                    ({ 0; })
539 #endif
540
541 static int __init create_proc_profile(void)
542 {
543         struct proc_dir_entry *entry;
544
545         if (!prof_on)
546                 return 0;
547         if (create_hash_tables())
548                 return -1;
549         if (!(entry = create_proc_entry("profile", S_IWUSR | S_IRUGO, NULL)))
550                 return 0;
551         entry->proc_fops = &proc_profile_operations;
552         entry->size = (1+prof_len) * sizeof(atomic_t);
553         hotcpu_notifier(profile_cpu_callback, 0);
554         return 0;
555 }
556 module_init(create_proc_profile);
557 #endif /* CONFIG_PROC_FS */