kthread: rename probe_kthread_data() to kthread_probe_data()
[linux-2.6-block.git] / kernel / profile.c
CommitLineData
1da177e4
LT
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,
6d49e352 11 * Nadia Yvette Chambers, Oracle, July 2004
1da177e4 12 * Amortized hit count accounting via per-cpu open-addressed hashtables
6d49e352
NYC
13 * to resolve timer interrupt livelocks, Nadia Yvette Chambers,
14 * Oracle, 2004
1da177e4
LT
15 */
16
9984de1a 17#include <linux/export.h>
1da177e4
LT
18#include <linux/profile.h>
19#include <linux/bootmem.h>
20#include <linux/notifier.h>
21#include <linux/mm.h>
22#include <linux/cpumask.h>
23#include <linux/cpu.h>
1da177e4 24#include <linux/highmem.h>
97d1f15b 25#include <linux/mutex.h>
22b8ce94
DH
26#include <linux/slab.h>
27#include <linux/vmalloc.h>
1da177e4 28#include <asm/sections.h>
7d12e780 29#include <asm/irq_regs.h>
e8edc6e0 30#include <asm/ptrace.h>
1da177e4
LT
31
32struct profile_hit {
33 u32 pc, hits;
34};
35#define PROFILE_GRPSHIFT 3
36#define PROFILE_GRPSZ (1 << PROFILE_GRPSHIFT)
37#define NR_PROFILE_HIT (PAGE_SIZE/sizeof(struct profile_hit))
38#define NR_PROFILE_GRP (NR_PROFILE_HIT/PROFILE_GRPSZ)
39
1da177e4
LT
40static atomic_t *prof_buffer;
41static unsigned long prof_len, prof_shift;
07031e14 42
ece8a684 43int prof_on __read_mostly;
07031e14
IM
44EXPORT_SYMBOL_GPL(prof_on);
45
c309b917 46static cpumask_var_t prof_cpu_mask;
ade356b9 47#if defined(CONFIG_SMP) && defined(CONFIG_PROC_FS)
1da177e4
LT
48static DEFINE_PER_CPU(struct profile_hit *[2], cpu_profile_hits);
49static DEFINE_PER_CPU(int, cpu_profile_flip);
97d1f15b 50static DEFINE_MUTEX(profile_flip_mutex);
1da177e4
LT
51#endif /* CONFIG_SMP */
52
22b8ce94 53int profile_setup(char *str)
1da177e4 54{
f3da64d1
FF
55 static const char schedstr[] = "schedule";
56 static const char sleepstr[] = "sleep";
57 static const char kvmstr[] = "kvm";
1da177e4
LT
58 int par;
59
ece8a684 60 if (!strncmp(str, sleepstr, strlen(sleepstr))) {
b3da2a73 61#ifdef CONFIG_SCHEDSTATS
cb251765 62 force_schedstat_enabled();
ece8a684
IM
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;
aba871f1 68 pr_info("kernel sleep profiling enabled (shift: %ld)\n",
ece8a684 69 prof_shift);
b3da2a73 70#else
aba871f1 71 pr_warn("kernel sleep profiling requires CONFIG_SCHEDSTATS\n");
b3da2a73 72#endif /* CONFIG_SCHEDSTATS */
a75acf85 73 } else if (!strncmp(str, schedstr, strlen(schedstr))) {
1da177e4 74 prof_on = SCHED_PROFILING;
dfaa9c94
WLII
75 if (str[strlen(schedstr)] == ',')
76 str += strlen(schedstr) + 1;
77 if (get_option(&str, &par))
78 prof_shift = par;
aba871f1 79 pr_info("kernel schedule profiling enabled (shift: %ld)\n",
dfaa9c94 80 prof_shift);
07031e14
IM
81 } else if (!strncmp(str, kvmstr, strlen(kvmstr))) {
82 prof_on = KVM_PROFILING;
83 if (str[strlen(kvmstr)] == ',')
84 str += strlen(kvmstr) + 1;
85 if (get_option(&str, &par))
86 prof_shift = par;
aba871f1 87 pr_info("kernel KVM profiling enabled (shift: %ld)\n",
07031e14 88 prof_shift);
dfaa9c94 89 } else if (get_option(&str, &par)) {
1da177e4
LT
90 prof_shift = par;
91 prof_on = CPU_PROFILING;
aba871f1 92 pr_info("kernel profiling enabled (shift: %ld)\n",
1da177e4
LT
93 prof_shift);
94 }
95 return 1;
96}
97__setup("profile=", profile_setup);
98
99
ce05fcc3 100int __ref profile_init(void)
1da177e4 101{
22b8ce94 102 int buffer_bytes;
1ad82fd5 103 if (!prof_on)
22b8ce94 104 return 0;
1ad82fd5 105
1da177e4
LT
106 /* only text is profiled */
107 prof_len = (_etext - _stext) >> prof_shift;
22b8ce94 108 buffer_bytes = prof_len*sizeof(atomic_t);
22b8ce94 109
c309b917
RR
110 if (!alloc_cpumask_var(&prof_cpu_mask, GFP_KERNEL))
111 return -ENOMEM;
112
acd89579
HD
113 cpumask_copy(prof_cpu_mask, cpu_possible_mask);
114
b62f495d 115 prof_buffer = kzalloc(buffer_bytes, GFP_KERNEL|__GFP_NOWARN);
22b8ce94
DH
116 if (prof_buffer)
117 return 0;
118
b62f495d
MG
119 prof_buffer = alloc_pages_exact(buffer_bytes,
120 GFP_KERNEL|__GFP_ZERO|__GFP_NOWARN);
22b8ce94
DH
121 if (prof_buffer)
122 return 0;
123
559fa6e7
JJ
124 prof_buffer = vzalloc(buffer_bytes);
125 if (prof_buffer)
22b8ce94
DH
126 return 0;
127
c309b917 128 free_cpumask_var(prof_cpu_mask);
22b8ce94 129 return -ENOMEM;
1da177e4
LT
130}
131
132/* Profile event notifications */
1ad82fd5 133
e041c683
AS
134static BLOCKING_NOTIFIER_HEAD(task_exit_notifier);
135static ATOMIC_NOTIFIER_HEAD(task_free_notifier);
136static BLOCKING_NOTIFIER_HEAD(munmap_notifier);
1ad82fd5
PC
137
138void profile_task_exit(struct task_struct *task)
1da177e4 139{
e041c683 140 blocking_notifier_call_chain(&task_exit_notifier, 0, task);
1da177e4 141}
1ad82fd5
PC
142
143int profile_handoff_task(struct task_struct *task)
1da177e4
LT
144{
145 int ret;
e041c683 146 ret = atomic_notifier_call_chain(&task_free_notifier, 0, task);
1da177e4
LT
147 return (ret == NOTIFY_OK) ? 1 : 0;
148}
149
150void profile_munmap(unsigned long addr)
151{
e041c683 152 blocking_notifier_call_chain(&munmap_notifier, 0, (void *)addr);
1da177e4
LT
153}
154
1ad82fd5 155int task_handoff_register(struct notifier_block *n)
1da177e4 156{
e041c683 157 return atomic_notifier_chain_register(&task_free_notifier, n);
1da177e4 158}
1ad82fd5 159EXPORT_SYMBOL_GPL(task_handoff_register);
1da177e4 160
1ad82fd5 161int task_handoff_unregister(struct notifier_block *n)
1da177e4 162{
e041c683 163 return atomic_notifier_chain_unregister(&task_free_notifier, n);
1da177e4 164}
1ad82fd5 165EXPORT_SYMBOL_GPL(task_handoff_unregister);
1da177e4 166
1ad82fd5 167int profile_event_register(enum profile_type type, struct notifier_block *n)
1da177e4
LT
168{
169 int err = -EINVAL;
1ad82fd5 170
1da177e4 171 switch (type) {
1ad82fd5
PC
172 case PROFILE_TASK_EXIT:
173 err = blocking_notifier_chain_register(
174 &task_exit_notifier, n);
175 break;
176 case PROFILE_MUNMAP:
177 err = blocking_notifier_chain_register(
178 &munmap_notifier, n);
179 break;
1da177e4 180 }
1ad82fd5 181
1da177e4
LT
182 return err;
183}
1ad82fd5 184EXPORT_SYMBOL_GPL(profile_event_register);
1da177e4 185
1ad82fd5 186int profile_event_unregister(enum profile_type type, struct notifier_block *n)
1da177e4
LT
187{
188 int err = -EINVAL;
1ad82fd5 189
1da177e4 190 switch (type) {
1ad82fd5
PC
191 case PROFILE_TASK_EXIT:
192 err = blocking_notifier_chain_unregister(
193 &task_exit_notifier, n);
194 break;
195 case PROFILE_MUNMAP:
196 err = blocking_notifier_chain_unregister(
197 &munmap_notifier, n);
198 break;
1da177e4
LT
199 }
200
1da177e4
LT
201 return err;
202}
1ad82fd5 203EXPORT_SYMBOL_GPL(profile_event_unregister);
1da177e4 204
ade356b9 205#if defined(CONFIG_SMP) && defined(CONFIG_PROC_FS)
1da177e4
LT
206/*
207 * Each cpu has a pair of open-addressed hashtables for pending
208 * profile hits. read_profile() IPI's all cpus to request them
209 * to flip buffers and flushes their contents to prof_buffer itself.
210 * Flip requests are serialized by the profile_flip_mutex. The sole
211 * use of having a second hashtable is for avoiding cacheline
212 * contention that would otherwise happen during flushes of pending
213 * profile hits required for the accuracy of reported profile hits
214 * and so resurrect the interrupt livelock issue.
215 *
216 * The open-addressed hashtables are indexed by profile buffer slot
217 * and hold the number of pending hits to that profile buffer slot on
218 * a cpu in an entry. When the hashtable overflows, all pending hits
219 * are accounted to their corresponding profile buffer slots with
220 * atomic_add() and the hashtable emptied. As numerous pending hits
221 * may be accounted to a profile buffer slot in a hashtable entry,
222 * this amortizes a number of atomic profile buffer increments likely
223 * to be far larger than the number of entries in the hashtable,
224 * particularly given that the number of distinct profile buffer
225 * positions to which hits are accounted during short intervals (e.g.
226 * several seconds) is usually very small. Exclusion from buffer
227 * flipping is provided by interrupt disablement (note that for
ece8a684
IM
228 * SCHED_PROFILING or SLEEP_PROFILING profile_hit() may be called from
229 * process context).
1da177e4
LT
230 * The hash function is meant to be lightweight as opposed to strong,
231 * and was vaguely inspired by ppc64 firmware-supported inverted
232 * pagetable hash functions, but uses a full hashtable full of finite
233 * collision chains, not just pairs of them.
234 *
6d49e352 235 * -- nyc
1da177e4
LT
236 */
237static void __profile_flip_buffers(void *unused)
238{
239 int cpu = smp_processor_id();
240
241 per_cpu(cpu_profile_flip, cpu) = !per_cpu(cpu_profile_flip, cpu);
242}
243
244static void profile_flip_buffers(void)
245{
246 int i, j, cpu;
247
97d1f15b 248 mutex_lock(&profile_flip_mutex);
1da177e4
LT
249 j = per_cpu(cpu_profile_flip, get_cpu());
250 put_cpu();
15c8b6c1 251 on_each_cpu(__profile_flip_buffers, NULL, 1);
1da177e4
LT
252 for_each_online_cpu(cpu) {
253 struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[j];
254 for (i = 0; i < NR_PROFILE_HIT; ++i) {
255 if (!hits[i].hits) {
256 if (hits[i].pc)
257 hits[i].pc = 0;
258 continue;
259 }
260 atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
261 hits[i].hits = hits[i].pc = 0;
262 }
263 }
97d1f15b 264 mutex_unlock(&profile_flip_mutex);
1da177e4
LT
265}
266
267static void profile_discard_flip_buffers(void)
268{
269 int i, cpu;
270
97d1f15b 271 mutex_lock(&profile_flip_mutex);
1da177e4
LT
272 i = per_cpu(cpu_profile_flip, get_cpu());
273 put_cpu();
15c8b6c1 274 on_each_cpu(__profile_flip_buffers, NULL, 1);
1da177e4
LT
275 for_each_online_cpu(cpu) {
276 struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[i];
277 memset(hits, 0, NR_PROFILE_HIT*sizeof(struct profile_hit));
278 }
97d1f15b 279 mutex_unlock(&profile_flip_mutex);
1da177e4
LT
280}
281
6f7bd76f 282static void do_profile_hits(int type, void *__pc, unsigned int nr_hits)
1da177e4
LT
283{
284 unsigned long primary, secondary, flags, pc = (unsigned long)__pc;
285 int i, j, cpu;
286 struct profile_hit *hits;
287
1da177e4
LT
288 pc = min((pc - (unsigned long)_stext) >> prof_shift, prof_len - 1);
289 i = primary = (pc & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
290 secondary = (~(pc << 1) & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
291 cpu = get_cpu();
292 hits = per_cpu(cpu_profile_hits, cpu)[per_cpu(cpu_profile_flip, cpu)];
293 if (!hits) {
294 put_cpu();
295 return;
296 }
ece8a684
IM
297 /*
298 * We buffer the global profiler buffer into a per-CPU
299 * queue and thus reduce the number of global (and possibly
300 * NUMA-alien) accesses. The write-queue is self-coalescing:
301 */
1da177e4
LT
302 local_irq_save(flags);
303 do {
304 for (j = 0; j < PROFILE_GRPSZ; ++j) {
305 if (hits[i + j].pc == pc) {
ece8a684 306 hits[i + j].hits += nr_hits;
1da177e4
LT
307 goto out;
308 } else if (!hits[i + j].hits) {
309 hits[i + j].pc = pc;
ece8a684 310 hits[i + j].hits = nr_hits;
1da177e4
LT
311 goto out;
312 }
313 }
314 i = (i + secondary) & (NR_PROFILE_HIT - 1);
315 } while (i != primary);
ece8a684
IM
316
317 /*
318 * Add the current hit(s) and flush the write-queue out
319 * to the global buffer:
320 */
321 atomic_add(nr_hits, &prof_buffer[pc]);
1da177e4
LT
322 for (i = 0; i < NR_PROFILE_HIT; ++i) {
323 atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
324 hits[i].pc = hits[i].hits = 0;
325 }
326out:
327 local_irq_restore(flags);
328 put_cpu();
329}
330
e722d8da 331static int profile_dead_cpu(unsigned int cpu)
1da177e4 332{
1da177e4 333 struct page *page;
e722d8da 334 int i;
1da177e4 335
e722d8da
SAS
336 if (prof_cpu_mask != NULL)
337 cpumask_clear_cpu(cpu, prof_cpu_mask);
338
339 for (i = 0; i < 2; i++) {
340 if (per_cpu(cpu_profile_hits, cpu)[i]) {
341 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[i]);
342 per_cpu(cpu_profile_hits, cpu)[i] = NULL;
1da177e4
LT
343 __free_page(page);
344 }
e722d8da
SAS
345 }
346 return 0;
347}
348
349static int profile_prepare_cpu(unsigned int cpu)
350{
351 int i, node = cpu_to_mem(cpu);
352 struct page *page;
353
354 per_cpu(cpu_profile_flip, cpu) = 0;
355
356 for (i = 0; i < 2; i++) {
357 if (per_cpu(cpu_profile_hits, cpu)[i])
358 continue;
359
360 page = __alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
361 if (!page) {
362 profile_dead_cpu(cpu);
363 return -ENOMEM;
1da177e4 364 }
e722d8da
SAS
365 per_cpu(cpu_profile_hits, cpu)[i] = page_address(page);
366
1da177e4 367 }
e722d8da
SAS
368 return 0;
369}
370
371static int profile_online_cpu(unsigned int cpu)
372{
373 if (prof_cpu_mask != NULL)
374 cpumask_set_cpu(cpu, prof_cpu_mask);
375
376 return 0;
1da177e4 377}
e722d8da 378
1da177e4
LT
379#else /* !CONFIG_SMP */
380#define profile_flip_buffers() do { } while (0)
381#define profile_discard_flip_buffers() do { } while (0)
382
6f7bd76f 383static void do_profile_hits(int type, void *__pc, unsigned int nr_hits)
1da177e4
LT
384{
385 unsigned long pc;
1da177e4 386 pc = ((unsigned long)__pc - (unsigned long)_stext) >> prof_shift;
ece8a684 387 atomic_add(nr_hits, &prof_buffer[min(pc, prof_len - 1)]);
1da177e4
LT
388}
389#endif /* !CONFIG_SMP */
6f7bd76f
RM
390
391void profile_hits(int type, void *__pc, unsigned int nr_hits)
392{
393 if (prof_on != type || !prof_buffer)
394 return;
395 do_profile_hits(type, __pc, nr_hits);
396}
bbe1a59b
AM
397EXPORT_SYMBOL_GPL(profile_hits);
398
7d12e780 399void profile_tick(int type)
1da177e4 400{
7d12e780
DH
401 struct pt_regs *regs = get_irq_regs();
402
c309b917
RR
403 if (!user_mode(regs) && prof_cpu_mask != NULL &&
404 cpumask_test_cpu(smp_processor_id(), prof_cpu_mask))
1da177e4
LT
405 profile_hit(type, (void *)profile_pc(regs));
406}
407
408#ifdef CONFIG_PROC_FS
409#include <linux/proc_fs.h>
583a22e7 410#include <linux/seq_file.h>
1da177e4 411#include <asm/uaccess.h>
1da177e4 412
583a22e7 413static int prof_cpu_mask_proc_show(struct seq_file *m, void *v)
1da177e4 414{
ccbd59c1 415 seq_printf(m, "%*pb\n", cpumask_pr_args(prof_cpu_mask));
583a22e7
AD
416 return 0;
417}
418
419static int prof_cpu_mask_proc_open(struct inode *inode, struct file *file)
420{
421 return single_open(file, prof_cpu_mask_proc_show, NULL);
1da177e4
LT
422}
423
583a22e7
AD
424static ssize_t prof_cpu_mask_proc_write(struct file *file,
425 const char __user *buffer, size_t count, loff_t *pos)
1da177e4 426{
c309b917 427 cpumask_var_t new_value;
583a22e7 428 int err;
1da177e4 429
c309b917
RR
430 if (!alloc_cpumask_var(&new_value, GFP_KERNEL))
431 return -ENOMEM;
1da177e4 432
c309b917
RR
433 err = cpumask_parse_user(buffer, count, new_value);
434 if (!err) {
583a22e7
AD
435 cpumask_copy(prof_cpu_mask, new_value);
436 err = count;
c309b917
RR
437 }
438 free_cpumask_var(new_value);
439 return err;
1da177e4
LT
440}
441
583a22e7
AD
442static const struct file_operations prof_cpu_mask_proc_fops = {
443 .open = prof_cpu_mask_proc_open,
444 .read = seq_read,
445 .llseek = seq_lseek,
446 .release = single_release,
447 .write = prof_cpu_mask_proc_write,
448};
449
fbd387ae 450void create_prof_cpu_mask(void)
1da177e4 451{
1da177e4 452 /* create /proc/irq/prof_cpu_mask */
fbd387ae 453 proc_create("irq/prof_cpu_mask", 0600, NULL, &prof_cpu_mask_proc_fops);
1da177e4
LT
454}
455
456/*
457 * This function accesses profiling information. The returned data is
458 * binary: the sampling step and the actual contents of the profile
459 * buffer. Use of the program readprofile is recommended in order to
460 * get meaningful info out of these data.
461 */
462static ssize_t
463read_profile(struct file *file, char __user *buf, size_t count, loff_t *ppos)
464{
465 unsigned long p = *ppos;
466 ssize_t read;
1ad82fd5 467 char *pnt;
1da177e4
LT
468 unsigned int sample_step = 1 << prof_shift;
469
470 profile_flip_buffers();
471 if (p >= (prof_len+1)*sizeof(unsigned int))
472 return 0;
473 if (count > (prof_len+1)*sizeof(unsigned int) - p)
474 count = (prof_len+1)*sizeof(unsigned int) - p;
475 read = 0;
476
477 while (p < sizeof(unsigned int) && count > 0) {
1ad82fd5 478 if (put_user(*((char *)(&sample_step)+p), buf))
064b022c 479 return -EFAULT;
1da177e4
LT
480 buf++; p++; count--; read++;
481 }
482 pnt = (char *)prof_buffer + p - sizeof(atomic_t);
1ad82fd5 483 if (copy_to_user(buf, (void *)pnt, count))
1da177e4
LT
484 return -EFAULT;
485 read += count;
486 *ppos += read;
487 return read;
488}
489
490/*
491 * Writing to /proc/profile resets the counters
492 *
493 * Writing a 'profiling multiplier' value into it also re-sets the profiling
494 * interrupt frequency, on architectures that support this.
495 */
496static ssize_t write_profile(struct file *file, const char __user *buf,
497 size_t count, loff_t *ppos)
498{
499#ifdef CONFIG_SMP
1ad82fd5 500 extern int setup_profiling_timer(unsigned int multiplier);
1da177e4
LT
501
502 if (count == sizeof(int)) {
503 unsigned int multiplier;
504
505 if (copy_from_user(&multiplier, buf, sizeof(int)))
506 return -EFAULT;
507
508 if (setup_profiling_timer(multiplier))
509 return -EINVAL;
510 }
511#endif
512 profile_discard_flip_buffers();
513 memset(prof_buffer, 0, prof_len * sizeof(atomic_t));
514 return count;
515}
516
15ad7cdc 517static const struct file_operations proc_profile_operations = {
1da177e4
LT
518 .read = read_profile,
519 .write = write_profile,
6038f373 520 .llseek = default_llseek,
1da177e4
LT
521};
522
e722d8da 523int __ref create_proc_profile(void)
1da177e4 524{
e722d8da
SAS
525 struct proc_dir_entry *entry;
526#ifdef CONFIG_SMP
527 enum cpuhp_state online_state;
1da177e4
LT
528#endif
529
c270a817 530 int err = 0;
1da177e4
LT
531
532 if (!prof_on)
533 return 0;
e722d8da
SAS
534#ifdef CONFIG_SMP
535 err = cpuhp_setup_state(CPUHP_PROFILE_PREPARE, "PROFILE_PREPARE",
536 profile_prepare_cpu, profile_dead_cpu);
537 if (err)
538 return err;
539
540 err = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "AP_PROFILE_ONLINE",
541 profile_online_cpu, NULL);
542 if (err < 0)
543 goto err_state_prep;
544 online_state = err;
545 err = 0;
546#endif
c33fff0a
DL
547 entry = proc_create("profile", S_IWUSR | S_IRUGO,
548 NULL, &proc_profile_operations);
1ad82fd5 549 if (!entry)
e722d8da 550 goto err_state_onl;
271a15ea 551 proc_set_size(entry, (1 + prof_len) * sizeof(atomic_t));
c270a817 552
e722d8da
SAS
553 return err;
554err_state_onl:
555#ifdef CONFIG_SMP
556 cpuhp_remove_state(online_state);
557err_state_prep:
558 cpuhp_remove_state(CPUHP_PROFILE_PREPARE);
559#endif
c270a817 560 return err;
1da177e4 561}
c96d6660 562subsys_initcall(create_proc_profile);
1da177e4 563#endif /* CONFIG_PROC_FS */