PCI: Tolerate hierarchies with no Root Port
[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;
1da177e4
LT
47#ifdef CONFIG_SMP
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
ece8a684
IM
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;
aba871f1 67 pr_info("kernel sleep profiling enabled (shift: %ld)\n",
ece8a684 68 prof_shift);
b3da2a73 69#else
aba871f1 70 pr_warn("kernel sleep profiling requires CONFIG_SCHEDSTATS\n");
b3da2a73 71#endif /* CONFIG_SCHEDSTATS */
a75acf85 72 } else if (!strncmp(str, schedstr, strlen(schedstr))) {
1da177e4 73 prof_on = SCHED_PROFILING;
dfaa9c94
WLII
74 if (str[strlen(schedstr)] == ',')
75 str += strlen(schedstr) + 1;
76 if (get_option(&str, &par))
77 prof_shift = par;
aba871f1 78 pr_info("kernel schedule profiling enabled (shift: %ld)\n",
dfaa9c94 79 prof_shift);
07031e14
IM
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;
aba871f1 86 pr_info("kernel KVM profiling enabled (shift: %ld)\n",
07031e14 87 prof_shift);
dfaa9c94 88 } else if (get_option(&str, &par)) {
1da177e4
LT
89 prof_shift = par;
90 prof_on = CPU_PROFILING;
aba871f1 91 pr_info("kernel profiling enabled (shift: %ld)\n",
1da177e4
LT
92 prof_shift);
93 }
94 return 1;
95}
96__setup("profile=", profile_setup);
97
98
ce05fcc3 99int __ref profile_init(void)
1da177e4 100{
22b8ce94 101 int buffer_bytes;
1ad82fd5 102 if (!prof_on)
22b8ce94 103 return 0;
1ad82fd5 104
1da177e4
LT
105 /* only text is profiled */
106 prof_len = (_etext - _stext) >> prof_shift;
22b8ce94 107 buffer_bytes = prof_len*sizeof(atomic_t);
22b8ce94 108
c309b917
RR
109 if (!alloc_cpumask_var(&prof_cpu_mask, GFP_KERNEL))
110 return -ENOMEM;
111
acd89579
HD
112 cpumask_copy(prof_cpu_mask, cpu_possible_mask);
113
b62f495d 114 prof_buffer = kzalloc(buffer_bytes, GFP_KERNEL|__GFP_NOWARN);
22b8ce94
DH
115 if (prof_buffer)
116 return 0;
117
b62f495d
MG
118 prof_buffer = alloc_pages_exact(buffer_bytes,
119 GFP_KERNEL|__GFP_ZERO|__GFP_NOWARN);
22b8ce94
DH
120 if (prof_buffer)
121 return 0;
122
559fa6e7
JJ
123 prof_buffer = vzalloc(buffer_bytes);
124 if (prof_buffer)
22b8ce94
DH
125 return 0;
126
c309b917 127 free_cpumask_var(prof_cpu_mask);
22b8ce94 128 return -ENOMEM;
1da177e4
LT
129}
130
131/* Profile event notifications */
1ad82fd5 132
e041c683
AS
133static BLOCKING_NOTIFIER_HEAD(task_exit_notifier);
134static ATOMIC_NOTIFIER_HEAD(task_free_notifier);
135static BLOCKING_NOTIFIER_HEAD(munmap_notifier);
1ad82fd5
PC
136
137void profile_task_exit(struct task_struct *task)
1da177e4 138{
e041c683 139 blocking_notifier_call_chain(&task_exit_notifier, 0, task);
1da177e4 140}
1ad82fd5
PC
141
142int profile_handoff_task(struct task_struct *task)
1da177e4
LT
143{
144 int ret;
e041c683 145 ret = atomic_notifier_call_chain(&task_free_notifier, 0, task);
1da177e4
LT
146 return (ret == NOTIFY_OK) ? 1 : 0;
147}
148
149void profile_munmap(unsigned long addr)
150{
e041c683 151 blocking_notifier_call_chain(&munmap_notifier, 0, (void *)addr);
1da177e4
LT
152}
153
1ad82fd5 154int task_handoff_register(struct notifier_block *n)
1da177e4 155{
e041c683 156 return atomic_notifier_chain_register(&task_free_notifier, n);
1da177e4 157}
1ad82fd5 158EXPORT_SYMBOL_GPL(task_handoff_register);
1da177e4 159
1ad82fd5 160int task_handoff_unregister(struct notifier_block *n)
1da177e4 161{
e041c683 162 return atomic_notifier_chain_unregister(&task_free_notifier, n);
1da177e4 163}
1ad82fd5 164EXPORT_SYMBOL_GPL(task_handoff_unregister);
1da177e4 165
1ad82fd5 166int profile_event_register(enum profile_type type, struct notifier_block *n)
1da177e4
LT
167{
168 int err = -EINVAL;
1ad82fd5 169
1da177e4 170 switch (type) {
1ad82fd5
PC
171 case PROFILE_TASK_EXIT:
172 err = blocking_notifier_chain_register(
173 &task_exit_notifier, n);
174 break;
175 case PROFILE_MUNMAP:
176 err = blocking_notifier_chain_register(
177 &munmap_notifier, n);
178 break;
1da177e4 179 }
1ad82fd5 180
1da177e4
LT
181 return err;
182}
1ad82fd5 183EXPORT_SYMBOL_GPL(profile_event_register);
1da177e4 184
1ad82fd5 185int profile_event_unregister(enum profile_type type, struct notifier_block *n)
1da177e4
LT
186{
187 int err = -EINVAL;
1ad82fd5 188
1da177e4 189 switch (type) {
1ad82fd5
PC
190 case PROFILE_TASK_EXIT:
191 err = blocking_notifier_chain_unregister(
192 &task_exit_notifier, n);
193 break;
194 case PROFILE_MUNMAP:
195 err = blocking_notifier_chain_unregister(
196 &munmap_notifier, n);
197 break;
1da177e4
LT
198 }
199
1da177e4
LT
200 return err;
201}
1ad82fd5 202EXPORT_SYMBOL_GPL(profile_event_unregister);
1da177e4 203
1da177e4
LT
204#ifdef CONFIG_SMP
205/*
206 * Each cpu has a pair of open-addressed hashtables for pending
207 * profile hits. read_profile() IPI's all cpus to request them
208 * to flip buffers and flushes their contents to prof_buffer itself.
209 * Flip requests are serialized by the profile_flip_mutex. The sole
210 * use of having a second hashtable is for avoiding cacheline
211 * contention that would otherwise happen during flushes of pending
212 * profile hits required for the accuracy of reported profile hits
213 * and so resurrect the interrupt livelock issue.
214 *
215 * The open-addressed hashtables are indexed by profile buffer slot
216 * and hold the number of pending hits to that profile buffer slot on
217 * a cpu in an entry. When the hashtable overflows, all pending hits
218 * are accounted to their corresponding profile buffer slots with
219 * atomic_add() and the hashtable emptied. As numerous pending hits
220 * may be accounted to a profile buffer slot in a hashtable entry,
221 * this amortizes a number of atomic profile buffer increments likely
222 * to be far larger than the number of entries in the hashtable,
223 * particularly given that the number of distinct profile buffer
224 * positions to which hits are accounted during short intervals (e.g.
225 * several seconds) is usually very small. Exclusion from buffer
226 * flipping is provided by interrupt disablement (note that for
ece8a684
IM
227 * SCHED_PROFILING or SLEEP_PROFILING profile_hit() may be called from
228 * process context).
1da177e4
LT
229 * The hash function is meant to be lightweight as opposed to strong,
230 * and was vaguely inspired by ppc64 firmware-supported inverted
231 * pagetable hash functions, but uses a full hashtable full of finite
232 * collision chains, not just pairs of them.
233 *
6d49e352 234 * -- nyc
1da177e4
LT
235 */
236static void __profile_flip_buffers(void *unused)
237{
238 int cpu = smp_processor_id();
239
240 per_cpu(cpu_profile_flip, cpu) = !per_cpu(cpu_profile_flip, cpu);
241}
242
243static void profile_flip_buffers(void)
244{
245 int i, j, cpu;
246
97d1f15b 247 mutex_lock(&profile_flip_mutex);
1da177e4
LT
248 j = per_cpu(cpu_profile_flip, get_cpu());
249 put_cpu();
15c8b6c1 250 on_each_cpu(__profile_flip_buffers, NULL, 1);
1da177e4
LT
251 for_each_online_cpu(cpu) {
252 struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[j];
253 for (i = 0; i < NR_PROFILE_HIT; ++i) {
254 if (!hits[i].hits) {
255 if (hits[i].pc)
256 hits[i].pc = 0;
257 continue;
258 }
259 atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
260 hits[i].hits = hits[i].pc = 0;
261 }
262 }
97d1f15b 263 mutex_unlock(&profile_flip_mutex);
1da177e4
LT
264}
265
266static void profile_discard_flip_buffers(void)
267{
268 int i, cpu;
269
97d1f15b 270 mutex_lock(&profile_flip_mutex);
1da177e4
LT
271 i = per_cpu(cpu_profile_flip, get_cpu());
272 put_cpu();
15c8b6c1 273 on_each_cpu(__profile_flip_buffers, NULL, 1);
1da177e4
LT
274 for_each_online_cpu(cpu) {
275 struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[i];
276 memset(hits, 0, NR_PROFILE_HIT*sizeof(struct profile_hit));
277 }
97d1f15b 278 mutex_unlock(&profile_flip_mutex);
1da177e4
LT
279}
280
6f7bd76f 281static void do_profile_hits(int type, void *__pc, unsigned int nr_hits)
1da177e4
LT
282{
283 unsigned long primary, secondary, flags, pc = (unsigned long)__pc;
284 int i, j, cpu;
285 struct profile_hit *hits;
286
1da177e4
LT
287 pc = min((pc - (unsigned long)_stext) >> prof_shift, prof_len - 1);
288 i = primary = (pc & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
289 secondary = (~(pc << 1) & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
290 cpu = get_cpu();
291 hits = per_cpu(cpu_profile_hits, cpu)[per_cpu(cpu_profile_flip, cpu)];
292 if (!hits) {
293 put_cpu();
294 return;
295 }
ece8a684
IM
296 /*
297 * We buffer the global profiler buffer into a per-CPU
298 * queue and thus reduce the number of global (and possibly
299 * NUMA-alien) accesses. The write-queue is self-coalescing:
300 */
1da177e4
LT
301 local_irq_save(flags);
302 do {
303 for (j = 0; j < PROFILE_GRPSZ; ++j) {
304 if (hits[i + j].pc == pc) {
ece8a684 305 hits[i + j].hits += nr_hits;
1da177e4
LT
306 goto out;
307 } else if (!hits[i + j].hits) {
308 hits[i + j].pc = pc;
ece8a684 309 hits[i + j].hits = nr_hits;
1da177e4
LT
310 goto out;
311 }
312 }
313 i = (i + secondary) & (NR_PROFILE_HIT - 1);
314 } while (i != primary);
ece8a684
IM
315
316 /*
317 * Add the current hit(s) and flush the write-queue out
318 * to the global buffer:
319 */
320 atomic_add(nr_hits, &prof_buffer[pc]);
1da177e4
LT
321 for (i = 0; i < NR_PROFILE_HIT; ++i) {
322 atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
323 hits[i].pc = hits[i].hits = 0;
324 }
325out:
326 local_irq_restore(flags);
327 put_cpu();
328}
329
0db0628d 330static int profile_cpu_callback(struct notifier_block *info,
1da177e4
LT
331 unsigned long action, void *__cpu)
332{
333 int node, cpu = (unsigned long)__cpu;
334 struct page *page;
335
336 switch (action) {
337 case CPU_UP_PREPARE:
8bb78442 338 case CPU_UP_PREPARE_FROZEN:
3dd6b5fb 339 node = cpu_to_mem(cpu);
1da177e4
LT
340 per_cpu(cpu_profile_flip, cpu) = 0;
341 if (!per_cpu(cpu_profile_hits, cpu)[1]) {
6484eb3e 342 page = alloc_pages_exact_node(node,
4199cfa0 343 GFP_KERNEL | __GFP_ZERO,
fbd98167 344 0);
1da177e4 345 if (!page)
80b5184c 346 return notifier_from_errno(-ENOMEM);
1da177e4
LT
347 per_cpu(cpu_profile_hits, cpu)[1] = page_address(page);
348 }
349 if (!per_cpu(cpu_profile_hits, cpu)[0]) {
6484eb3e 350 page = alloc_pages_exact_node(node,
4199cfa0 351 GFP_KERNEL | __GFP_ZERO,
fbd98167 352 0);
1da177e4
LT
353 if (!page)
354 goto out_free;
355 per_cpu(cpu_profile_hits, cpu)[0] = page_address(page);
356 }
357 break;
1ad82fd5 358out_free:
1da177e4
LT
359 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
360 per_cpu(cpu_profile_hits, cpu)[1] = NULL;
361 __free_page(page);
80b5184c 362 return notifier_from_errno(-ENOMEM);
1da177e4 363 case CPU_ONLINE:
8bb78442 364 case CPU_ONLINE_FROZEN:
c309b917
RR
365 if (prof_cpu_mask != NULL)
366 cpumask_set_cpu(cpu, prof_cpu_mask);
1da177e4
LT
367 break;
368 case CPU_UP_CANCELED:
8bb78442 369 case CPU_UP_CANCELED_FROZEN:
1da177e4 370 case CPU_DEAD:
8bb78442 371 case CPU_DEAD_FROZEN:
c309b917
RR
372 if (prof_cpu_mask != NULL)
373 cpumask_clear_cpu(cpu, prof_cpu_mask);
1da177e4
LT
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}
1da177e4
LT
388#else /* !CONFIG_SMP */
389#define profile_flip_buffers() do { } while (0)
390#define profile_discard_flip_buffers() do { } while (0)
02316067 391#define profile_cpu_callback NULL
1da177e4 392
6f7bd76f 393static void do_profile_hits(int type, void *__pc, unsigned int nr_hits)
1da177e4
LT
394{
395 unsigned long pc;
1da177e4 396 pc = ((unsigned long)__pc - (unsigned long)_stext) >> prof_shift;
ece8a684 397 atomic_add(nr_hits, &prof_buffer[min(pc, prof_len - 1)]);
1da177e4
LT
398}
399#endif /* !CONFIG_SMP */
6f7bd76f
RM
400
401void profile_hits(int type, void *__pc, unsigned int nr_hits)
402{
403 if (prof_on != type || !prof_buffer)
404 return;
405 do_profile_hits(type, __pc, nr_hits);
406}
bbe1a59b
AM
407EXPORT_SYMBOL_GPL(profile_hits);
408
7d12e780 409void profile_tick(int type)
1da177e4 410{
7d12e780
DH
411 struct pt_regs *regs = get_irq_regs();
412
c309b917
RR
413 if (!user_mode(regs) && prof_cpu_mask != NULL &&
414 cpumask_test_cpu(smp_processor_id(), prof_cpu_mask))
1da177e4
LT
415 profile_hit(type, (void *)profile_pc(regs));
416}
417
418#ifdef CONFIG_PROC_FS
419#include <linux/proc_fs.h>
583a22e7 420#include <linux/seq_file.h>
1da177e4 421#include <asm/uaccess.h>
1da177e4 422
583a22e7 423static int prof_cpu_mask_proc_show(struct seq_file *m, void *v)
1da177e4 424{
ccbd59c1 425 seq_printf(m, "%*pb\n", cpumask_pr_args(prof_cpu_mask));
583a22e7
AD
426 return 0;
427}
428
429static int prof_cpu_mask_proc_open(struct inode *inode, struct file *file)
430{
431 return single_open(file, prof_cpu_mask_proc_show, NULL);
1da177e4
LT
432}
433
583a22e7
AD
434static ssize_t prof_cpu_mask_proc_write(struct file *file,
435 const char __user *buffer, size_t count, loff_t *pos)
1da177e4 436{
c309b917 437 cpumask_var_t new_value;
583a22e7 438 int err;
1da177e4 439
c309b917
RR
440 if (!alloc_cpumask_var(&new_value, GFP_KERNEL))
441 return -ENOMEM;
1da177e4 442
c309b917
RR
443 err = cpumask_parse_user(buffer, count, new_value);
444 if (!err) {
583a22e7
AD
445 cpumask_copy(prof_cpu_mask, new_value);
446 err = count;
c309b917
RR
447 }
448 free_cpumask_var(new_value);
449 return err;
1da177e4
LT
450}
451
583a22e7
AD
452static const struct file_operations prof_cpu_mask_proc_fops = {
453 .open = prof_cpu_mask_proc_open,
454 .read = seq_read,
455 .llseek = seq_lseek,
456 .release = single_release,
457 .write = prof_cpu_mask_proc_write,
458};
459
fbd387ae 460void create_prof_cpu_mask(void)
1da177e4 461{
1da177e4 462 /* create /proc/irq/prof_cpu_mask */
fbd387ae 463 proc_create("irq/prof_cpu_mask", 0600, NULL, &prof_cpu_mask_proc_fops);
1da177e4
LT
464}
465
466/*
467 * This function accesses profiling information. The returned data is
468 * binary: the sampling step and the actual contents of the profile
469 * buffer. Use of the program readprofile is recommended in order to
470 * get meaningful info out of these data.
471 */
472static ssize_t
473read_profile(struct file *file, char __user *buf, size_t count, loff_t *ppos)
474{
475 unsigned long p = *ppos;
476 ssize_t read;
1ad82fd5 477 char *pnt;
1da177e4
LT
478 unsigned int sample_step = 1 << prof_shift;
479
480 profile_flip_buffers();
481 if (p >= (prof_len+1)*sizeof(unsigned int))
482 return 0;
483 if (count > (prof_len+1)*sizeof(unsigned int) - p)
484 count = (prof_len+1)*sizeof(unsigned int) - p;
485 read = 0;
486
487 while (p < sizeof(unsigned int) && count > 0) {
1ad82fd5 488 if (put_user(*((char *)(&sample_step)+p), buf))
064b022c 489 return -EFAULT;
1da177e4
LT
490 buf++; p++; count--; read++;
491 }
492 pnt = (char *)prof_buffer + p - sizeof(atomic_t);
1ad82fd5 493 if (copy_to_user(buf, (void *)pnt, count))
1da177e4
LT
494 return -EFAULT;
495 read += count;
496 *ppos += read;
497 return read;
498}
499
500/*
501 * Writing to /proc/profile resets the counters
502 *
503 * Writing a 'profiling multiplier' value into it also re-sets the profiling
504 * interrupt frequency, on architectures that support this.
505 */
506static ssize_t write_profile(struct file *file, const char __user *buf,
507 size_t count, loff_t *ppos)
508{
509#ifdef CONFIG_SMP
1ad82fd5 510 extern int setup_profiling_timer(unsigned int multiplier);
1da177e4
LT
511
512 if (count == sizeof(int)) {
513 unsigned int multiplier;
514
515 if (copy_from_user(&multiplier, buf, sizeof(int)))
516 return -EFAULT;
517
518 if (setup_profiling_timer(multiplier))
519 return -EINVAL;
520 }
521#endif
522 profile_discard_flip_buffers();
523 memset(prof_buffer, 0, prof_len * sizeof(atomic_t));
524 return count;
525}
526
15ad7cdc 527static const struct file_operations proc_profile_operations = {
1da177e4
LT
528 .read = read_profile,
529 .write = write_profile,
6038f373 530 .llseek = default_llseek,
1da177e4
LT
531};
532
533#ifdef CONFIG_SMP
60a51513 534static void profile_nop(void *unused)
1da177e4
LT
535{
536}
537
22b8ce94 538static int create_hash_tables(void)
1da177e4
LT
539{
540 int cpu;
541
542 for_each_online_cpu(cpu) {
3dd6b5fb 543 int node = cpu_to_mem(cpu);
1da177e4
LT
544 struct page *page;
545
6484eb3e 546 page = alloc_pages_exact_node(node,
e97ca8e5 547 GFP_KERNEL | __GFP_ZERO | __GFP_THISNODE,
fbd98167 548 0);
1da177e4
LT
549 if (!page)
550 goto out_cleanup;
551 per_cpu(cpu_profile_hits, cpu)[1]
552 = (struct profile_hit *)page_address(page);
6484eb3e 553 page = alloc_pages_exact_node(node,
e97ca8e5 554 GFP_KERNEL | __GFP_ZERO | __GFP_THISNODE,
fbd98167 555 0);
1da177e4
LT
556 if (!page)
557 goto out_cleanup;
558 per_cpu(cpu_profile_hits, cpu)[0]
559 = (struct profile_hit *)page_address(page);
560 }
561 return 0;
562out_cleanup:
563 prof_on = 0;
d59dd462 564 smp_mb();
15c8b6c1 565 on_each_cpu(profile_nop, NULL, 1);
1da177e4
LT
566 for_each_online_cpu(cpu) {
567 struct page *page;
568
569 if (per_cpu(cpu_profile_hits, cpu)[0]) {
570 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[0]);
571 per_cpu(cpu_profile_hits, cpu)[0] = NULL;
572 __free_page(page);
573 }
574 if (per_cpu(cpu_profile_hits, cpu)[1]) {
575 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
576 per_cpu(cpu_profile_hits, cpu)[1] = NULL;
577 __free_page(page);
578 }
579 }
580 return -1;
581}
582#else
583#define create_hash_tables() ({ 0; })
584#endif
585
84196414 586int __ref create_proc_profile(void) /* false positive from hotcpu_notifier */
1da177e4
LT
587{
588 struct proc_dir_entry *entry;
c270a817 589 int err = 0;
1da177e4
LT
590
591 if (!prof_on)
592 return 0;
c270a817
SB
593
594 cpu_notifier_register_begin();
595
596 if (create_hash_tables()) {
597 err = -ENOMEM;
598 goto out;
599 }
600
c33fff0a
DL
601 entry = proc_create("profile", S_IWUSR | S_IRUGO,
602 NULL, &proc_profile_operations);
1ad82fd5 603 if (!entry)
c270a817 604 goto out;
271a15ea 605 proc_set_size(entry, (1 + prof_len) * sizeof(atomic_t));
c270a817
SB
606 __hotcpu_notifier(profile_cpu_callback, 0);
607
608out:
609 cpu_notifier_register_done();
610 return err;
1da177e4 611}
c96d6660 612subsys_initcall(create_proc_profile);
1da177e4 613#endif /* CONFIG_PROC_FS */