Commit | Line | Data |
---|---|---|
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, | |
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 | ||
1da177e4 LT |
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> | |
1da177e4 | 23 | #include <linux/highmem.h> |
97d1f15b | 24 | #include <linux/mutex.h> |
22b8ce94 DH |
25 | #include <linux/slab.h> |
26 | #include <linux/vmalloc.h> | |
1da177e4 | 27 | #include <asm/sections.h> |
7d12e780 | 28 | #include <asm/irq_regs.h> |
e8edc6e0 | 29 | #include <asm/ptrace.h> |
1da177e4 LT |
30 | |
31 | struct profile_hit { | |
32 | u32 pc, hits; | |
33 | }; | |
34 | #define PROFILE_GRPSHIFT 3 | |
35 | #define PROFILE_GRPSZ (1 << PROFILE_GRPSHIFT) | |
36 | #define NR_PROFILE_HIT (PAGE_SIZE/sizeof(struct profile_hit)) | |
37 | #define NR_PROFILE_GRP (NR_PROFILE_HIT/PROFILE_GRPSZ) | |
38 | ||
39 | /* Oprofile timer tick hook */ | |
b012d346 | 40 | static int (*timer_hook)(struct pt_regs *) __read_mostly; |
1da177e4 LT |
41 | |
42 | static atomic_t *prof_buffer; | |
43 | static unsigned long prof_len, prof_shift; | |
07031e14 | 44 | |
ece8a684 | 45 | int prof_on __read_mostly; |
07031e14 IM |
46 | EXPORT_SYMBOL_GPL(prof_on); |
47 | ||
c309b917 | 48 | static cpumask_var_t prof_cpu_mask; |
1da177e4 LT |
49 | #ifdef CONFIG_SMP |
50 | static DEFINE_PER_CPU(struct profile_hit *[2], cpu_profile_hits); | |
51 | static DEFINE_PER_CPU(int, cpu_profile_flip); | |
97d1f15b | 52 | static DEFINE_MUTEX(profile_flip_mutex); |
1da177e4 LT |
53 | #endif /* CONFIG_SMP */ |
54 | ||
22b8ce94 | 55 | int profile_setup(char *str) |
1da177e4 | 56 | { |
22b8ce94 DH |
57 | static char schedstr[] = "schedule"; |
58 | static char sleepstr[] = "sleep"; | |
59 | static char kvmstr[] = "kvm"; | |
1da177e4 LT |
60 | int par; |
61 | ||
ece8a684 | 62 | if (!strncmp(str, sleepstr, strlen(sleepstr))) { |
b3da2a73 | 63 | #ifdef CONFIG_SCHEDSTATS |
ece8a684 IM |
64 | prof_on = SLEEP_PROFILING; |
65 | if (str[strlen(sleepstr)] == ',') | |
66 | str += strlen(sleepstr) + 1; | |
67 | if (get_option(&str, &par)) | |
68 | prof_shift = par; | |
69 | printk(KERN_INFO | |
70 | "kernel sleep profiling enabled (shift: %ld)\n", | |
71 | prof_shift); | |
b3da2a73 MG |
72 | #else |
73 | printk(KERN_WARNING | |
74 | "kernel sleep profiling requires CONFIG_SCHEDSTATS\n"); | |
75 | #endif /* CONFIG_SCHEDSTATS */ | |
a75acf85 | 76 | } else if (!strncmp(str, schedstr, strlen(schedstr))) { |
1da177e4 | 77 | prof_on = SCHED_PROFILING; |
dfaa9c94 WLII |
78 | if (str[strlen(schedstr)] == ',') |
79 | str += strlen(schedstr) + 1; | |
80 | if (get_option(&str, &par)) | |
81 | prof_shift = par; | |
82 | printk(KERN_INFO | |
83 | "kernel schedule profiling enabled (shift: %ld)\n", | |
84 | prof_shift); | |
07031e14 IM |
85 | } else if (!strncmp(str, kvmstr, strlen(kvmstr))) { |
86 | prof_on = KVM_PROFILING; | |
87 | if (str[strlen(kvmstr)] == ',') | |
88 | str += strlen(kvmstr) + 1; | |
89 | if (get_option(&str, &par)) | |
90 | prof_shift = par; | |
91 | printk(KERN_INFO | |
92 | "kernel KVM profiling enabled (shift: %ld)\n", | |
93 | prof_shift); | |
dfaa9c94 | 94 | } else if (get_option(&str, &par)) { |
1da177e4 LT |
95 | prof_shift = par; |
96 | prof_on = CPU_PROFILING; | |
97 | printk(KERN_INFO "kernel profiling enabled (shift: %ld)\n", | |
98 | prof_shift); | |
99 | } | |
100 | return 1; | |
101 | } | |
102 | __setup("profile=", profile_setup); | |
103 | ||
104 | ||
ce05fcc3 | 105 | int __ref profile_init(void) |
1da177e4 | 106 | { |
22b8ce94 | 107 | int buffer_bytes; |
1ad82fd5 | 108 | if (!prof_on) |
22b8ce94 | 109 | return 0; |
1ad82fd5 | 110 | |
1da177e4 LT |
111 | /* only text is profiled */ |
112 | prof_len = (_etext - _stext) >> prof_shift; | |
22b8ce94 | 113 | buffer_bytes = prof_len*sizeof(atomic_t); |
22b8ce94 | 114 | |
c309b917 RR |
115 | if (!alloc_cpumask_var(&prof_cpu_mask, GFP_KERNEL)) |
116 | return -ENOMEM; | |
117 | ||
acd89579 HD |
118 | cpumask_copy(prof_cpu_mask, cpu_possible_mask); |
119 | ||
22b8ce94 DH |
120 | prof_buffer = kzalloc(buffer_bytes, GFP_KERNEL); |
121 | if (prof_buffer) | |
122 | return 0; | |
123 | ||
124 | prof_buffer = alloc_pages_exact(buffer_bytes, GFP_KERNEL|__GFP_ZERO); | |
125 | if (prof_buffer) | |
126 | return 0; | |
127 | ||
128 | prof_buffer = vmalloc(buffer_bytes); | |
129 | if (prof_buffer) | |
130 | return 0; | |
131 | ||
c309b917 | 132 | free_cpumask_var(prof_cpu_mask); |
22b8ce94 | 133 | return -ENOMEM; |
1da177e4 LT |
134 | } |
135 | ||
136 | /* Profile event notifications */ | |
1ad82fd5 | 137 | |
e041c683 AS |
138 | static BLOCKING_NOTIFIER_HEAD(task_exit_notifier); |
139 | static ATOMIC_NOTIFIER_HEAD(task_free_notifier); | |
140 | static BLOCKING_NOTIFIER_HEAD(munmap_notifier); | |
1ad82fd5 PC |
141 | |
142 | void profile_task_exit(struct task_struct *task) | |
1da177e4 | 143 | { |
e041c683 | 144 | blocking_notifier_call_chain(&task_exit_notifier, 0, task); |
1da177e4 | 145 | } |
1ad82fd5 PC |
146 | |
147 | int profile_handoff_task(struct task_struct *task) | |
1da177e4 LT |
148 | { |
149 | int ret; | |
e041c683 | 150 | ret = atomic_notifier_call_chain(&task_free_notifier, 0, task); |
1da177e4 LT |
151 | return (ret == NOTIFY_OK) ? 1 : 0; |
152 | } | |
153 | ||
154 | void profile_munmap(unsigned long addr) | |
155 | { | |
e041c683 | 156 | blocking_notifier_call_chain(&munmap_notifier, 0, (void *)addr); |
1da177e4 LT |
157 | } |
158 | ||
1ad82fd5 | 159 | int task_handoff_register(struct notifier_block *n) |
1da177e4 | 160 | { |
e041c683 | 161 | return atomic_notifier_chain_register(&task_free_notifier, n); |
1da177e4 | 162 | } |
1ad82fd5 | 163 | EXPORT_SYMBOL_GPL(task_handoff_register); |
1da177e4 | 164 | |
1ad82fd5 | 165 | int task_handoff_unregister(struct notifier_block *n) |
1da177e4 | 166 | { |
e041c683 | 167 | return atomic_notifier_chain_unregister(&task_free_notifier, n); |
1da177e4 | 168 | } |
1ad82fd5 | 169 | EXPORT_SYMBOL_GPL(task_handoff_unregister); |
1da177e4 | 170 | |
1ad82fd5 | 171 | int profile_event_register(enum profile_type type, struct notifier_block *n) |
1da177e4 LT |
172 | { |
173 | int err = -EINVAL; | |
1ad82fd5 | 174 | |
1da177e4 | 175 | switch (type) { |
1ad82fd5 PC |
176 | case PROFILE_TASK_EXIT: |
177 | err = blocking_notifier_chain_register( | |
178 | &task_exit_notifier, n); | |
179 | break; | |
180 | case PROFILE_MUNMAP: | |
181 | err = blocking_notifier_chain_register( | |
182 | &munmap_notifier, n); | |
183 | break; | |
1da177e4 | 184 | } |
1ad82fd5 | 185 | |
1da177e4 LT |
186 | return err; |
187 | } | |
1ad82fd5 | 188 | EXPORT_SYMBOL_GPL(profile_event_register); |
1da177e4 | 189 | |
1ad82fd5 | 190 | int profile_event_unregister(enum profile_type type, struct notifier_block *n) |
1da177e4 LT |
191 | { |
192 | int err = -EINVAL; | |
1ad82fd5 | 193 | |
1da177e4 | 194 | switch (type) { |
1ad82fd5 PC |
195 | case PROFILE_TASK_EXIT: |
196 | err = blocking_notifier_chain_unregister( | |
197 | &task_exit_notifier, n); | |
198 | break; | |
199 | case PROFILE_MUNMAP: | |
200 | err = blocking_notifier_chain_unregister( | |
201 | &munmap_notifier, n); | |
202 | break; | |
1da177e4 LT |
203 | } |
204 | ||
1da177e4 LT |
205 | return err; |
206 | } | |
1ad82fd5 | 207 | EXPORT_SYMBOL_GPL(profile_event_unregister); |
1da177e4 LT |
208 | |
209 | int register_timer_hook(int (*hook)(struct pt_regs *)) | |
210 | { | |
211 | if (timer_hook) | |
212 | return -EBUSY; | |
213 | timer_hook = hook; | |
214 | return 0; | |
215 | } | |
1ad82fd5 | 216 | EXPORT_SYMBOL_GPL(register_timer_hook); |
1da177e4 LT |
217 | |
218 | void unregister_timer_hook(int (*hook)(struct pt_regs *)) | |
219 | { | |
220 | WARN_ON(hook != timer_hook); | |
221 | timer_hook = NULL; | |
222 | /* make sure all CPUs see the NULL hook */ | |
fbd568a3 | 223 | synchronize_sched(); /* Allow ongoing interrupts to complete. */ |
1da177e4 | 224 | } |
1da177e4 | 225 | EXPORT_SYMBOL_GPL(unregister_timer_hook); |
1da177e4 | 226 | |
1da177e4 LT |
227 | |
228 | #ifdef CONFIG_SMP | |
229 | /* | |
230 | * Each cpu has a pair of open-addressed hashtables for pending | |
231 | * profile hits. read_profile() IPI's all cpus to request them | |
232 | * to flip buffers and flushes their contents to prof_buffer itself. | |
233 | * Flip requests are serialized by the profile_flip_mutex. The sole | |
234 | * use of having a second hashtable is for avoiding cacheline | |
235 | * contention that would otherwise happen during flushes of pending | |
236 | * profile hits required for the accuracy of reported profile hits | |
237 | * and so resurrect the interrupt livelock issue. | |
238 | * | |
239 | * The open-addressed hashtables are indexed by profile buffer slot | |
240 | * and hold the number of pending hits to that profile buffer slot on | |
241 | * a cpu in an entry. When the hashtable overflows, all pending hits | |
242 | * are accounted to their corresponding profile buffer slots with | |
243 | * atomic_add() and the hashtable emptied. As numerous pending hits | |
244 | * may be accounted to a profile buffer slot in a hashtable entry, | |
245 | * this amortizes a number of atomic profile buffer increments likely | |
246 | * to be far larger than the number of entries in the hashtable, | |
247 | * particularly given that the number of distinct profile buffer | |
248 | * positions to which hits are accounted during short intervals (e.g. | |
249 | * several seconds) is usually very small. Exclusion from buffer | |
250 | * flipping is provided by interrupt disablement (note that for | |
ece8a684 IM |
251 | * SCHED_PROFILING or SLEEP_PROFILING profile_hit() may be called from |
252 | * process context). | |
1da177e4 LT |
253 | * The hash function is meant to be lightweight as opposed to strong, |
254 | * and was vaguely inspired by ppc64 firmware-supported inverted | |
255 | * pagetable hash functions, but uses a full hashtable full of finite | |
256 | * collision chains, not just pairs of them. | |
257 | * | |
258 | * -- wli | |
259 | */ | |
260 | static void __profile_flip_buffers(void *unused) | |
261 | { | |
262 | int cpu = smp_processor_id(); | |
263 | ||
264 | per_cpu(cpu_profile_flip, cpu) = !per_cpu(cpu_profile_flip, cpu); | |
265 | } | |
266 | ||
267 | static void profile_flip_buffers(void) | |
268 | { | |
269 | int i, j, cpu; | |
270 | ||
97d1f15b | 271 | mutex_lock(&profile_flip_mutex); |
1da177e4 LT |
272 | j = 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)[j]; | |
277 | for (i = 0; i < NR_PROFILE_HIT; ++i) { | |
278 | if (!hits[i].hits) { | |
279 | if (hits[i].pc) | |
280 | hits[i].pc = 0; | |
281 | continue; | |
282 | } | |
283 | atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]); | |
284 | hits[i].hits = hits[i].pc = 0; | |
285 | } | |
286 | } | |
97d1f15b | 287 | mutex_unlock(&profile_flip_mutex); |
1da177e4 LT |
288 | } |
289 | ||
290 | static void profile_discard_flip_buffers(void) | |
291 | { | |
292 | int i, cpu; | |
293 | ||
97d1f15b | 294 | mutex_lock(&profile_flip_mutex); |
1da177e4 LT |
295 | i = per_cpu(cpu_profile_flip, get_cpu()); |
296 | put_cpu(); | |
15c8b6c1 | 297 | on_each_cpu(__profile_flip_buffers, NULL, 1); |
1da177e4 LT |
298 | for_each_online_cpu(cpu) { |
299 | struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[i]; | |
300 | memset(hits, 0, NR_PROFILE_HIT*sizeof(struct profile_hit)); | |
301 | } | |
97d1f15b | 302 | mutex_unlock(&profile_flip_mutex); |
1da177e4 LT |
303 | } |
304 | ||
ece8a684 | 305 | void profile_hits(int type, void *__pc, unsigned int nr_hits) |
1da177e4 LT |
306 | { |
307 | unsigned long primary, secondary, flags, pc = (unsigned long)__pc; | |
308 | int i, j, cpu; | |
309 | struct profile_hit *hits; | |
310 | ||
311 | if (prof_on != type || !prof_buffer) | |
312 | return; | |
313 | pc = min((pc - (unsigned long)_stext) >> prof_shift, prof_len - 1); | |
314 | i = primary = (pc & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT; | |
315 | secondary = (~(pc << 1) & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT; | |
316 | cpu = get_cpu(); | |
317 | hits = per_cpu(cpu_profile_hits, cpu)[per_cpu(cpu_profile_flip, cpu)]; | |
318 | if (!hits) { | |
319 | put_cpu(); | |
320 | return; | |
321 | } | |
ece8a684 IM |
322 | /* |
323 | * We buffer the global profiler buffer into a per-CPU | |
324 | * queue and thus reduce the number of global (and possibly | |
325 | * NUMA-alien) accesses. The write-queue is self-coalescing: | |
326 | */ | |
1da177e4 LT |
327 | local_irq_save(flags); |
328 | do { | |
329 | for (j = 0; j < PROFILE_GRPSZ; ++j) { | |
330 | if (hits[i + j].pc == pc) { | |
ece8a684 | 331 | hits[i + j].hits += nr_hits; |
1da177e4 LT |
332 | goto out; |
333 | } else if (!hits[i + j].hits) { | |
334 | hits[i + j].pc = pc; | |
ece8a684 | 335 | hits[i + j].hits = nr_hits; |
1da177e4 LT |
336 | goto out; |
337 | } | |
338 | } | |
339 | i = (i + secondary) & (NR_PROFILE_HIT - 1); | |
340 | } while (i != primary); | |
ece8a684 IM |
341 | |
342 | /* | |
343 | * Add the current hit(s) and flush the write-queue out | |
344 | * to the global buffer: | |
345 | */ | |
346 | atomic_add(nr_hits, &prof_buffer[pc]); | |
1da177e4 LT |
347 | for (i = 0; i < NR_PROFILE_HIT; ++i) { |
348 | atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]); | |
349 | hits[i].pc = hits[i].hits = 0; | |
350 | } | |
351 | out: | |
352 | local_irq_restore(flags); | |
353 | put_cpu(); | |
354 | } | |
355 | ||
84196414 | 356 | static int __cpuinit profile_cpu_callback(struct notifier_block *info, |
1da177e4 LT |
357 | unsigned long action, void *__cpu) |
358 | { | |
359 | int node, cpu = (unsigned long)__cpu; | |
360 | struct page *page; | |
361 | ||
362 | switch (action) { | |
363 | case CPU_UP_PREPARE: | |
8bb78442 | 364 | case CPU_UP_PREPARE_FROZEN: |
1da177e4 LT |
365 | node = cpu_to_node(cpu); |
366 | per_cpu(cpu_profile_flip, cpu) = 0; | |
367 | if (!per_cpu(cpu_profile_hits, cpu)[1]) { | |
6484eb3e | 368 | page = alloc_pages_exact_node(node, |
4199cfa0 | 369 | GFP_KERNEL | __GFP_ZERO, |
fbd98167 | 370 | 0); |
1da177e4 LT |
371 | if (!page) |
372 | return NOTIFY_BAD; | |
373 | per_cpu(cpu_profile_hits, cpu)[1] = page_address(page); | |
374 | } | |
375 | if (!per_cpu(cpu_profile_hits, cpu)[0]) { | |
6484eb3e | 376 | page = alloc_pages_exact_node(node, |
4199cfa0 | 377 | GFP_KERNEL | __GFP_ZERO, |
fbd98167 | 378 | 0); |
1da177e4 LT |
379 | if (!page) |
380 | goto out_free; | |
381 | per_cpu(cpu_profile_hits, cpu)[0] = page_address(page); | |
382 | } | |
383 | break; | |
1ad82fd5 | 384 | out_free: |
1da177e4 LT |
385 | page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]); |
386 | per_cpu(cpu_profile_hits, cpu)[1] = NULL; | |
387 | __free_page(page); | |
388 | return NOTIFY_BAD; | |
389 | case CPU_ONLINE: | |
8bb78442 | 390 | case CPU_ONLINE_FROZEN: |
c309b917 RR |
391 | if (prof_cpu_mask != NULL) |
392 | cpumask_set_cpu(cpu, prof_cpu_mask); | |
1da177e4 LT |
393 | break; |
394 | case CPU_UP_CANCELED: | |
8bb78442 | 395 | case CPU_UP_CANCELED_FROZEN: |
1da177e4 | 396 | case CPU_DEAD: |
8bb78442 | 397 | case CPU_DEAD_FROZEN: |
c309b917 RR |
398 | if (prof_cpu_mask != NULL) |
399 | cpumask_clear_cpu(cpu, prof_cpu_mask); | |
1da177e4 LT |
400 | if (per_cpu(cpu_profile_hits, cpu)[0]) { |
401 | page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[0]); | |
402 | per_cpu(cpu_profile_hits, cpu)[0] = NULL; | |
403 | __free_page(page); | |
404 | } | |
405 | if (per_cpu(cpu_profile_hits, cpu)[1]) { | |
406 | page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]); | |
407 | per_cpu(cpu_profile_hits, cpu)[1] = NULL; | |
408 | __free_page(page); | |
409 | } | |
410 | break; | |
411 | } | |
412 | return NOTIFY_OK; | |
413 | } | |
1da177e4 LT |
414 | #else /* !CONFIG_SMP */ |
415 | #define profile_flip_buffers() do { } while (0) | |
416 | #define profile_discard_flip_buffers() do { } while (0) | |
02316067 | 417 | #define profile_cpu_callback NULL |
1da177e4 | 418 | |
ece8a684 | 419 | void profile_hits(int type, void *__pc, unsigned int nr_hits) |
1da177e4 LT |
420 | { |
421 | unsigned long pc; | |
422 | ||
423 | if (prof_on != type || !prof_buffer) | |
424 | return; | |
425 | pc = ((unsigned long)__pc - (unsigned long)_stext) >> prof_shift; | |
ece8a684 | 426 | atomic_add(nr_hits, &prof_buffer[min(pc, prof_len - 1)]); |
1da177e4 LT |
427 | } |
428 | #endif /* !CONFIG_SMP */ | |
bbe1a59b AM |
429 | EXPORT_SYMBOL_GPL(profile_hits); |
430 | ||
7d12e780 | 431 | void profile_tick(int type) |
1da177e4 | 432 | { |
7d12e780 DH |
433 | struct pt_regs *regs = get_irq_regs(); |
434 | ||
1da177e4 LT |
435 | if (type == CPU_PROFILING && timer_hook) |
436 | timer_hook(regs); | |
c309b917 RR |
437 | if (!user_mode(regs) && prof_cpu_mask != NULL && |
438 | cpumask_test_cpu(smp_processor_id(), prof_cpu_mask)) | |
1da177e4 LT |
439 | profile_hit(type, (void *)profile_pc(regs)); |
440 | } | |
441 | ||
442 | #ifdef CONFIG_PROC_FS | |
443 | #include <linux/proc_fs.h> | |
444 | #include <asm/uaccess.h> | |
1da177e4 | 445 | |
1ad82fd5 | 446 | static int prof_cpu_mask_read_proc(char *page, char **start, off_t off, |
1da177e4 LT |
447 | int count, int *eof, void *data) |
448 | { | |
c309b917 | 449 | int len = cpumask_scnprintf(page, count, data); |
1da177e4 LT |
450 | if (count - len < 2) |
451 | return -EINVAL; | |
452 | len += sprintf(page + len, "\n"); | |
453 | return len; | |
454 | } | |
455 | ||
1ad82fd5 PC |
456 | static int prof_cpu_mask_write_proc(struct file *file, |
457 | const char __user *buffer, unsigned long count, void *data) | |
1da177e4 | 458 | { |
c309b917 | 459 | struct cpumask *mask = data; |
1da177e4 | 460 | unsigned long full_count = count, err; |
c309b917 | 461 | cpumask_var_t new_value; |
1da177e4 | 462 | |
c309b917 RR |
463 | if (!alloc_cpumask_var(&new_value, GFP_KERNEL)) |
464 | return -ENOMEM; | |
1da177e4 | 465 | |
c309b917 RR |
466 | err = cpumask_parse_user(buffer, count, new_value); |
467 | if (!err) { | |
468 | cpumask_copy(mask, new_value); | |
469 | err = full_count; | |
470 | } | |
471 | free_cpumask_var(new_value); | |
472 | return err; | |
1da177e4 LT |
473 | } |
474 | ||
475 | void create_prof_cpu_mask(struct proc_dir_entry *root_irq_dir) | |
476 | { | |
477 | struct proc_dir_entry *entry; | |
478 | ||
479 | /* create /proc/irq/prof_cpu_mask */ | |
1ad82fd5 PC |
480 | entry = create_proc_entry("prof_cpu_mask", 0600, root_irq_dir); |
481 | if (!entry) | |
1da177e4 | 482 | return; |
c309b917 | 483 | entry->data = prof_cpu_mask; |
1da177e4 LT |
484 | entry->read_proc = prof_cpu_mask_read_proc; |
485 | entry->write_proc = prof_cpu_mask_write_proc; | |
486 | } | |
487 | ||
488 | /* | |
489 | * This function accesses profiling information. The returned data is | |
490 | * binary: the sampling step and the actual contents of the profile | |
491 | * buffer. Use of the program readprofile is recommended in order to | |
492 | * get meaningful info out of these data. | |
493 | */ | |
494 | static ssize_t | |
495 | read_profile(struct file *file, char __user *buf, size_t count, loff_t *ppos) | |
496 | { | |
497 | unsigned long p = *ppos; | |
498 | ssize_t read; | |
1ad82fd5 | 499 | char *pnt; |
1da177e4 LT |
500 | unsigned int sample_step = 1 << prof_shift; |
501 | ||
502 | profile_flip_buffers(); | |
503 | if (p >= (prof_len+1)*sizeof(unsigned int)) | |
504 | return 0; | |
505 | if (count > (prof_len+1)*sizeof(unsigned int) - p) | |
506 | count = (prof_len+1)*sizeof(unsigned int) - p; | |
507 | read = 0; | |
508 | ||
509 | while (p < sizeof(unsigned int) && count > 0) { | |
1ad82fd5 | 510 | if (put_user(*((char *)(&sample_step)+p), buf)) |
064b022c | 511 | return -EFAULT; |
1da177e4 LT |
512 | buf++; p++; count--; read++; |
513 | } | |
514 | pnt = (char *)prof_buffer + p - sizeof(atomic_t); | |
1ad82fd5 | 515 | if (copy_to_user(buf, (void *)pnt, count)) |
1da177e4 LT |
516 | return -EFAULT; |
517 | read += count; | |
518 | *ppos += read; | |
519 | return read; | |
520 | } | |
521 | ||
522 | /* | |
523 | * Writing to /proc/profile resets the counters | |
524 | * | |
525 | * Writing a 'profiling multiplier' value into it also re-sets the profiling | |
526 | * interrupt frequency, on architectures that support this. | |
527 | */ | |
528 | static ssize_t write_profile(struct file *file, const char __user *buf, | |
529 | size_t count, loff_t *ppos) | |
530 | { | |
531 | #ifdef CONFIG_SMP | |
1ad82fd5 | 532 | extern int setup_profiling_timer(unsigned int multiplier); |
1da177e4 LT |
533 | |
534 | if (count == sizeof(int)) { | |
535 | unsigned int multiplier; | |
536 | ||
537 | if (copy_from_user(&multiplier, buf, sizeof(int))) | |
538 | return -EFAULT; | |
539 | ||
540 | if (setup_profiling_timer(multiplier)) | |
541 | return -EINVAL; | |
542 | } | |
543 | #endif | |
544 | profile_discard_flip_buffers(); | |
545 | memset(prof_buffer, 0, prof_len * sizeof(atomic_t)); | |
546 | return count; | |
547 | } | |
548 | ||
15ad7cdc | 549 | static const struct file_operations proc_profile_operations = { |
1da177e4 LT |
550 | .read = read_profile, |
551 | .write = write_profile, | |
552 | }; | |
553 | ||
554 | #ifdef CONFIG_SMP | |
60a51513 | 555 | static void profile_nop(void *unused) |
1da177e4 LT |
556 | { |
557 | } | |
558 | ||
22b8ce94 | 559 | static int create_hash_tables(void) |
1da177e4 LT |
560 | { |
561 | int cpu; | |
562 | ||
563 | for_each_online_cpu(cpu) { | |
564 | int node = cpu_to_node(cpu); | |
565 | struct page *page; | |
566 | ||
6484eb3e | 567 | page = alloc_pages_exact_node(node, |
fbd98167 CL |
568 | GFP_KERNEL | __GFP_ZERO | GFP_THISNODE, |
569 | 0); | |
1da177e4 LT |
570 | if (!page) |
571 | goto out_cleanup; | |
572 | per_cpu(cpu_profile_hits, cpu)[1] | |
573 | = (struct profile_hit *)page_address(page); | |
6484eb3e | 574 | page = alloc_pages_exact_node(node, |
fbd98167 CL |
575 | GFP_KERNEL | __GFP_ZERO | GFP_THISNODE, |
576 | 0); | |
1da177e4 LT |
577 | if (!page) |
578 | goto out_cleanup; | |
579 | per_cpu(cpu_profile_hits, cpu)[0] | |
580 | = (struct profile_hit *)page_address(page); | |
581 | } | |
582 | return 0; | |
583 | out_cleanup: | |
584 | prof_on = 0; | |
d59dd462 | 585 | smp_mb(); |
15c8b6c1 | 586 | on_each_cpu(profile_nop, NULL, 1); |
1da177e4 LT |
587 | for_each_online_cpu(cpu) { |
588 | struct page *page; | |
589 | ||
590 | if (per_cpu(cpu_profile_hits, cpu)[0]) { | |
591 | page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[0]); | |
592 | per_cpu(cpu_profile_hits, cpu)[0] = NULL; | |
593 | __free_page(page); | |
594 | } | |
595 | if (per_cpu(cpu_profile_hits, cpu)[1]) { | |
596 | page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]); | |
597 | per_cpu(cpu_profile_hits, cpu)[1] = NULL; | |
598 | __free_page(page); | |
599 | } | |
600 | } | |
601 | return -1; | |
602 | } | |
603 | #else | |
604 | #define create_hash_tables() ({ 0; }) | |
605 | #endif | |
606 | ||
84196414 | 607 | int __ref create_proc_profile(void) /* false positive from hotcpu_notifier */ |
1da177e4 LT |
608 | { |
609 | struct proc_dir_entry *entry; | |
610 | ||
611 | if (!prof_on) | |
612 | return 0; | |
613 | if (create_hash_tables()) | |
22b8ce94 | 614 | return -ENOMEM; |
c33fff0a DL |
615 | entry = proc_create("profile", S_IWUSR | S_IRUGO, |
616 | NULL, &proc_profile_operations); | |
1ad82fd5 | 617 | if (!entry) |
1da177e4 | 618 | return 0; |
1da177e4 LT |
619 | entry->size = (1+prof_len) * sizeof(atomic_t); |
620 | hotcpu_notifier(profile_cpu_callback, 0); | |
621 | return 0; | |
622 | } | |
623 | module_init(create_proc_profile); | |
624 | #endif /* CONFIG_PROC_FS */ |