1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * linux/kernel/profile.c
4 * Simple profiling. Manages a direct-mapped profile hit count buffer,
5 * with configurable resolution, support for restricting the cpus on
6 * which profiling is done, and switching between cpu time and
7 * schedule() calls via kernel command line parameters passed at boot.
8 *
9 * Scheduler profiling support, Arjan van de Ven and Ingo Molnar,
10 * Red Hat, July 2004
11 * Consolidation of architecture support code for profiling,
12 * Nadia Yvette Chambers, Oracle, July 2004
13 * Amortized hit count accounting via per-cpu open-addressed hashtables
14 * to resolve timer interrupt livelocks, Nadia Yvette Chambers,
15 * Oracle, 2004
16 */
17
18 #include <linux/export.h>
19 #include <linux/profile.h>
20 #include <linux/memblock.h>
21 #include <linux/notifier.h>
22 #include <linux/mm.h>
23 #include <linux/cpumask.h>
24 #include <linux/cpu.h>
25 #include <linux/highmem.h>
26 #include <linux/mutex.h>
27 #include <linux/slab.h>
28 #include <linux/vmalloc.h>
29 #include <linux/sched/stat.h>
30
31 #include <asm/sections.h>
32 #include <asm/irq_regs.h>
33 #include <asm/ptrace.h>
34
35 struct profile_hit {
36 u32 pc, hits;
37 };
38 #define PROFILE_GRPSHIFT 3
39 #define PROFILE_GRPSZ (1 << PROFILE_GRPSHIFT)
40 #define NR_PROFILE_HIT (PAGE_SIZE/sizeof(struct profile_hit))
41 #define NR_PROFILE_GRP (NR_PROFILE_HIT/PROFILE_GRPSZ)
42
43 static atomic_t *prof_buffer;
44 static unsigned long prof_len;
45 static unsigned short int prof_shift;
46
47 int prof_on __read_mostly;
48 EXPORT_SYMBOL_GPL(prof_on);
49
50 static cpumask_var_t prof_cpu_mask;
51 #if defined(CONFIG_SMP) && defined(CONFIG_PROC_FS)
52 static DEFINE_PER_CPU(struct profile_hit *[2], cpu_profile_hits);
53 static DEFINE_PER_CPU(int, cpu_profile_flip);
54 static DEFINE_MUTEX(profile_flip_mutex);
55 #endif /* CONFIG_SMP */
56
profile_setup(char * str)57 int profile_setup(char *str)
58 {
59 static const char schedstr[] = "schedule";
60 static const char sleepstr[] = "sleep";
61 static const char kvmstr[] = "kvm";
62 int par;
63
64 if (!strncmp(str, sleepstr, strlen(sleepstr))) {
65 #ifdef CONFIG_SCHEDSTATS
66 force_schedstat_enabled();
67 prof_on = SLEEP_PROFILING;
68 if (str[strlen(sleepstr)] == ',')
69 str += strlen(sleepstr) + 1;
70 if (get_option(&str, &par))
71 prof_shift = clamp(par, 0, BITS_PER_LONG - 1);
72 pr_info("kernel sleep profiling enabled (shift: %u)\n",
73 prof_shift);
74 #else
75 pr_warn("kernel sleep profiling requires CONFIG_SCHEDSTATS\n");
76 #endif /* CONFIG_SCHEDSTATS */
77 } else if (!strncmp(str, schedstr, strlen(schedstr))) {
78 prof_on = SCHED_PROFILING;
79 if (str[strlen(schedstr)] == ',')
80 str += strlen(schedstr) + 1;
81 if (get_option(&str, &par))
82 prof_shift = clamp(par, 0, BITS_PER_LONG - 1);
83 pr_info("kernel schedule profiling enabled (shift: %u)\n",
84 prof_shift);
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 = clamp(par, 0, BITS_PER_LONG - 1);
91 pr_info("kernel KVM profiling enabled (shift: %u)\n",
92 prof_shift);
93 } else if (get_option(&str, &par)) {
94 prof_shift = clamp(par, 0, BITS_PER_LONG - 1);
95 prof_on = CPU_PROFILING;
96 pr_info("kernel profiling enabled (shift: %u)\n",
97 prof_shift);
98 }
99 return 1;
100 }
101 __setup("profile=", profile_setup);
102
103
profile_init(void)104 int __ref profile_init(void)
105 {
106 int buffer_bytes;
107 if (!prof_on)
108 return 0;
109
110 /* only text is profiled */
111 prof_len = (_etext - _stext) >> prof_shift;
112
113 if (!prof_len) {
114 pr_warn("profiling shift: %u too large\n", prof_shift);
115 prof_on = 0;
116 return -EINVAL;
117 }
118
119 buffer_bytes = prof_len*sizeof(atomic_t);
120
121 if (!alloc_cpumask_var(&prof_cpu_mask, GFP_KERNEL))
122 return -ENOMEM;
123
124 cpumask_copy(prof_cpu_mask, cpu_possible_mask);
125
126 prof_buffer = kzalloc(buffer_bytes, GFP_KERNEL|__GFP_NOWARN);
127 if (prof_buffer)
128 return 0;
129
130 prof_buffer = alloc_pages_exact(buffer_bytes,
131 GFP_KERNEL|__GFP_ZERO|__GFP_NOWARN);
132 if (prof_buffer)
133 return 0;
134
135 prof_buffer = vzalloc(buffer_bytes);
136 if (prof_buffer)
137 return 0;
138
139 free_cpumask_var(prof_cpu_mask);
140 return -ENOMEM;
141 }
142
143 /* Profile event notifications */
144
145 static BLOCKING_NOTIFIER_HEAD(task_exit_notifier);
146 static ATOMIC_NOTIFIER_HEAD(task_free_notifier);
147 static BLOCKING_NOTIFIER_HEAD(munmap_notifier);
148
profile_task_exit(struct task_struct * task)149 void profile_task_exit(struct task_struct *task)
150 {
151 blocking_notifier_call_chain(&task_exit_notifier, 0, task);
152 }
153
profile_handoff_task(struct task_struct * task)154 int profile_handoff_task(struct task_struct *task)
155 {
156 int ret;
157 ret = atomic_notifier_call_chain(&task_free_notifier, 0, task);
158 return (ret == NOTIFY_OK) ? 1 : 0;
159 }
160
profile_munmap(unsigned long addr)161 void profile_munmap(unsigned long addr)
162 {
163 blocking_notifier_call_chain(&munmap_notifier, 0, (void *)addr);
164 }
165
task_handoff_register(struct notifier_block * n)166 int task_handoff_register(struct notifier_block *n)
167 {
168 return atomic_notifier_chain_register(&task_free_notifier, n);
169 }
170 EXPORT_SYMBOL_GPL(task_handoff_register);
171
task_handoff_unregister(struct notifier_block * n)172 int task_handoff_unregister(struct notifier_block *n)
173 {
174 return atomic_notifier_chain_unregister(&task_free_notifier, n);
175 }
176 EXPORT_SYMBOL_GPL(task_handoff_unregister);
177
profile_event_register(enum profile_type type,struct notifier_block * n)178 int profile_event_register(enum profile_type type, struct notifier_block *n)
179 {
180 int err = -EINVAL;
181
182 switch (type) {
183 case PROFILE_TASK_EXIT:
184 err = blocking_notifier_chain_register(
185 &task_exit_notifier, n);
186 break;
187 case PROFILE_MUNMAP:
188 err = blocking_notifier_chain_register(
189 &munmap_notifier, n);
190 break;
191 }
192
193 return err;
194 }
195 EXPORT_SYMBOL_GPL(profile_event_register);
196
profile_event_unregister(enum profile_type type,struct notifier_block * n)197 int profile_event_unregister(enum profile_type type, struct notifier_block *n)
198 {
199 int err = -EINVAL;
200
201 switch (type) {
202 case PROFILE_TASK_EXIT:
203 err = blocking_notifier_chain_unregister(
204 &task_exit_notifier, n);
205 break;
206 case PROFILE_MUNMAP:
207 err = blocking_notifier_chain_unregister(
208 &munmap_notifier, n);
209 break;
210 }
211
212 return err;
213 }
214 EXPORT_SYMBOL_GPL(profile_event_unregister);
215
216 #if defined(CONFIG_SMP) && defined(CONFIG_PROC_FS)
217 /*
218 * Each cpu has a pair of open-addressed hashtables for pending
219 * profile hits. read_profile() IPI's all cpus to request them
220 * to flip buffers and flushes their contents to prof_buffer itself.
221 * Flip requests are serialized by the profile_flip_mutex. The sole
222 * use of having a second hashtable is for avoiding cacheline
223 * contention that would otherwise happen during flushes of pending
224 * profile hits required for the accuracy of reported profile hits
225 * and so resurrect the interrupt livelock issue.
226 *
227 * The open-addressed hashtables are indexed by profile buffer slot
228 * and hold the number of pending hits to that profile buffer slot on
229 * a cpu in an entry. When the hashtable overflows, all pending hits
230 * are accounted to their corresponding profile buffer slots with
231 * atomic_add() and the hashtable emptied. As numerous pending hits
232 * may be accounted to a profile buffer slot in a hashtable entry,
233 * this amortizes a number of atomic profile buffer increments likely
234 * to be far larger than the number of entries in the hashtable,
235 * particularly given that the number of distinct profile buffer
236 * positions to which hits are accounted during short intervals (e.g.
237 * several seconds) is usually very small. Exclusion from buffer
238 * flipping is provided by interrupt disablement (note that for
239 * SCHED_PROFILING or SLEEP_PROFILING profile_hit() may be called from
240 * process context).
241 * The hash function is meant to be lightweight as opposed to strong,
242 * and was vaguely inspired by ppc64 firmware-supported inverted
243 * pagetable hash functions, but uses a full hashtable full of finite
244 * collision chains, not just pairs of them.
245 *
246 * -- nyc
247 */
__profile_flip_buffers(void * unused)248 static void __profile_flip_buffers(void *unused)
249 {
250 int cpu = smp_processor_id();
251
252 per_cpu(cpu_profile_flip, cpu) = !per_cpu(cpu_profile_flip, cpu);
253 }
254
profile_flip_buffers(void)255 static void profile_flip_buffers(void)
256 {
257 int i, j, cpu;
258
259 mutex_lock(&profile_flip_mutex);
260 j = per_cpu(cpu_profile_flip, get_cpu());
261 put_cpu();
262 on_each_cpu(__profile_flip_buffers, NULL, 1);
263 for_each_online_cpu(cpu) {
264 struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[j];
265 for (i = 0; i < NR_PROFILE_HIT; ++i) {
266 if (!hits[i].hits) {
267 if (hits[i].pc)
268 hits[i].pc = 0;
269 continue;
270 }
271 atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
272 hits[i].hits = hits[i].pc = 0;
273 }
274 }
275 mutex_unlock(&profile_flip_mutex);
276 }
277
profile_discard_flip_buffers(void)278 static void profile_discard_flip_buffers(void)
279 {
280 int i, cpu;
281
282 mutex_lock(&profile_flip_mutex);
283 i = per_cpu(cpu_profile_flip, get_cpu());
284 put_cpu();
285 on_each_cpu(__profile_flip_buffers, NULL, 1);
286 for_each_online_cpu(cpu) {
287 struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[i];
288 memset(hits, 0, NR_PROFILE_HIT*sizeof(struct profile_hit));
289 }
290 mutex_unlock(&profile_flip_mutex);
291 }
292
do_profile_hits(int type,void * __pc,unsigned int nr_hits)293 static void do_profile_hits(int type, void *__pc, unsigned int nr_hits)
294 {
295 unsigned long primary, secondary, flags, pc = (unsigned long)__pc;
296 int i, j, cpu;
297 struct profile_hit *hits;
298
299 pc = min((pc - (unsigned long)_stext) >> prof_shift, prof_len - 1);
300 i = primary = (pc & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
301 secondary = (~(pc << 1) & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
302 cpu = get_cpu();
303 hits = per_cpu(cpu_profile_hits, cpu)[per_cpu(cpu_profile_flip, cpu)];
304 if (!hits) {
305 put_cpu();
306 return;
307 }
308 /*
309 * We buffer the global profiler buffer into a per-CPU
310 * queue and thus reduce the number of global (and possibly
311 * NUMA-alien) accesses. The write-queue is self-coalescing:
312 */
313 local_irq_save(flags);
314 do {
315 for (j = 0; j < PROFILE_GRPSZ; ++j) {
316 if (hits[i + j].pc == pc) {
317 hits[i + j].hits += nr_hits;
318 goto out;
319 } else if (!hits[i + j].hits) {
320 hits[i + j].pc = pc;
321 hits[i + j].hits = nr_hits;
322 goto out;
323 }
324 }
325 i = (i + secondary) & (NR_PROFILE_HIT - 1);
326 } while (i != primary);
327
328 /*
329 * Add the current hit(s) and flush the write-queue out
330 * to the global buffer:
331 */
332 atomic_add(nr_hits, &prof_buffer[pc]);
333 for (i = 0; i < NR_PROFILE_HIT; ++i) {
334 atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
335 hits[i].pc = hits[i].hits = 0;
336 }
337 out:
338 local_irq_restore(flags);
339 put_cpu();
340 }
341
profile_dead_cpu(unsigned int cpu)342 static int profile_dead_cpu(unsigned int cpu)
343 {
344 struct page *page;
345 int i;
346
347 if (cpumask_available(prof_cpu_mask))
348 cpumask_clear_cpu(cpu, prof_cpu_mask);
349
350 for (i = 0; i < 2; i++) {
351 if (per_cpu(cpu_profile_hits, cpu)[i]) {
352 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[i]);
353 per_cpu(cpu_profile_hits, cpu)[i] = NULL;
354 __free_page(page);
355 }
356 }
357 return 0;
358 }
359
profile_prepare_cpu(unsigned int cpu)360 static int profile_prepare_cpu(unsigned int cpu)
361 {
362 int i, node = cpu_to_mem(cpu);
363 struct page *page;
364
365 per_cpu(cpu_profile_flip, cpu) = 0;
366
367 for (i = 0; i < 2; i++) {
368 if (per_cpu(cpu_profile_hits, cpu)[i])
369 continue;
370
371 page = __alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
372 if (!page) {
373 profile_dead_cpu(cpu);
374 return -ENOMEM;
375 }
376 per_cpu(cpu_profile_hits, cpu)[i] = page_address(page);
377
378 }
379 return 0;
380 }
381
profile_online_cpu(unsigned int cpu)382 static int profile_online_cpu(unsigned int cpu)
383 {
384 if (cpumask_available(prof_cpu_mask))
385 cpumask_set_cpu(cpu, prof_cpu_mask);
386
387 return 0;
388 }
389
390 #else /* !CONFIG_SMP */
391 #define profile_flip_buffers() do { } while (0)
392 #define profile_discard_flip_buffers() do { } while (0)
393
do_profile_hits(int type,void * __pc,unsigned int nr_hits)394 static void do_profile_hits(int type, void *__pc, unsigned int nr_hits)
395 {
396 unsigned long pc;
397 pc = ((unsigned long)__pc - (unsigned long)_stext) >> prof_shift;
398 atomic_add(nr_hits, &prof_buffer[min(pc, prof_len - 1)]);
399 }
400 #endif /* !CONFIG_SMP */
401
profile_hits(int type,void * __pc,unsigned int nr_hits)402 void profile_hits(int type, void *__pc, unsigned int nr_hits)
403 {
404 if (prof_on != type || !prof_buffer)
405 return;
406 do_profile_hits(type, __pc, nr_hits);
407 }
408 EXPORT_SYMBOL_GPL(profile_hits);
409
profile_tick(int type)410 void profile_tick(int type)
411 {
412 struct pt_regs *regs = get_irq_regs();
413
414 if (!user_mode(regs) && cpumask_available(prof_cpu_mask) &&
415 cpumask_test_cpu(smp_processor_id(), prof_cpu_mask))
416 profile_hit(type, (void *)profile_pc(regs));
417 }
418
419 #ifdef CONFIG_PROC_FS
420 #include <linux/proc_fs.h>
421 #include <linux/seq_file.h>
422 #include <linux/uaccess.h>
423
prof_cpu_mask_proc_show(struct seq_file * m,void * v)424 static int prof_cpu_mask_proc_show(struct seq_file *m, void *v)
425 {
426 seq_printf(m, "%*pb\n", cpumask_pr_args(prof_cpu_mask));
427 return 0;
428 }
429
prof_cpu_mask_proc_open(struct inode * inode,struct file * file)430 static int prof_cpu_mask_proc_open(struct inode *inode, struct file *file)
431 {
432 return single_open(file, prof_cpu_mask_proc_show, NULL);
433 }
434
prof_cpu_mask_proc_write(struct file * file,const char __user * buffer,size_t count,loff_t * pos)435 static ssize_t prof_cpu_mask_proc_write(struct file *file,
436 const char __user *buffer, size_t count, loff_t *pos)
437 {
438 cpumask_var_t new_value;
439 int err;
440
441 if (!zalloc_cpumask_var(&new_value, GFP_KERNEL))
442 return -ENOMEM;
443
444 err = cpumask_parse_user(buffer, count, new_value);
445 if (!err) {
446 cpumask_copy(prof_cpu_mask, new_value);
447 err = count;
448 }
449 free_cpumask_var(new_value);
450 return err;
451 }
452
453 static const struct proc_ops prof_cpu_mask_proc_ops = {
454 .proc_open = prof_cpu_mask_proc_open,
455 .proc_read = seq_read,
456 .proc_lseek = seq_lseek,
457 .proc_release = single_release,
458 .proc_write = prof_cpu_mask_proc_write,
459 };
460
create_prof_cpu_mask(void)461 void create_prof_cpu_mask(void)
462 {
463 /* create /proc/irq/prof_cpu_mask */
464 proc_create("irq/prof_cpu_mask", 0600, NULL, &prof_cpu_mask_proc_ops);
465 }
466
467 /*
468 * This function accesses profiling information. The returned data is
469 * binary: the sampling step and the actual contents of the profile
470 * buffer. Use of the program readprofile is recommended in order to
471 * get meaningful info out of these data.
472 */
473 static ssize_t
read_profile(struct file * file,char __user * buf,size_t count,loff_t * ppos)474 read_profile(struct file *file, char __user *buf, size_t count, loff_t *ppos)
475 {
476 unsigned long p = *ppos;
477 ssize_t read;
478 char *pnt;
479 unsigned long sample_step = 1UL << prof_shift;
480
481 profile_flip_buffers();
482 if (p >= (prof_len+1)*sizeof(unsigned int))
483 return 0;
484 if (count > (prof_len+1)*sizeof(unsigned int) - p)
485 count = (prof_len+1)*sizeof(unsigned int) - p;
486 read = 0;
487
488 while (p < sizeof(unsigned int) && count > 0) {
489 if (put_user(*((char *)(&sample_step)+p), buf))
490 return -EFAULT;
491 buf++; p++; count--; read++;
492 }
493 pnt = (char *)prof_buffer + p - sizeof(atomic_t);
494 if (copy_to_user(buf, (void *)pnt, count))
495 return -EFAULT;
496 read += count;
497 *ppos += read;
498 return read;
499 }
500
501 /*
502 * Writing to /proc/profile resets the counters
503 *
504 * Writing a 'profiling multiplier' value into it also re-sets the profiling
505 * interrupt frequency, on architectures that support this.
506 */
write_profile(struct file * file,const char __user * buf,size_t count,loff_t * ppos)507 static ssize_t write_profile(struct file *file, const char __user *buf,
508 size_t count, loff_t *ppos)
509 {
510 #ifdef CONFIG_SMP
511 extern int setup_profiling_timer(unsigned int multiplier);
512
513 if (count == sizeof(int)) {
514 unsigned int multiplier;
515
516 if (copy_from_user(&multiplier, buf, sizeof(int)))
517 return -EFAULT;
518
519 if (setup_profiling_timer(multiplier))
520 return -EINVAL;
521 }
522 #endif
523 profile_discard_flip_buffers();
524 memset(prof_buffer, 0, prof_len * sizeof(atomic_t));
525 return count;
526 }
527
528 static const struct proc_ops profile_proc_ops = {
529 .proc_read = read_profile,
530 .proc_write = write_profile,
531 .proc_lseek = default_llseek,
532 };
533
create_proc_profile(void)534 int __ref create_proc_profile(void)
535 {
536 struct proc_dir_entry *entry;
537 #ifdef CONFIG_SMP
538 enum cpuhp_state online_state;
539 #endif
540
541 int err = 0;
542
543 if (!prof_on)
544 return 0;
545 #ifdef CONFIG_SMP
546 err = cpuhp_setup_state(CPUHP_PROFILE_PREPARE, "PROFILE_PREPARE",
547 profile_prepare_cpu, profile_dead_cpu);
548 if (err)
549 return err;
550
551 err = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "AP_PROFILE_ONLINE",
552 profile_online_cpu, NULL);
553 if (err < 0)
554 goto err_state_prep;
555 online_state = err;
556 err = 0;
557 #endif
558 entry = proc_create("profile", S_IWUSR | S_IRUGO,
559 NULL, &profile_proc_ops);
560 if (!entry)
561 goto err_state_onl;
562 proc_set_size(entry, (1 + prof_len) * sizeof(atomic_t));
563
564 return err;
565 err_state_onl:
566 #ifdef CONFIG_SMP
567 cpuhp_remove_state(online_state);
568 err_state_prep:
569 cpuhp_remove_state(CPUHP_PROFILE_PREPARE);
570 #endif
571 return err;
572 }
573 subsys_initcall(create_proc_profile);
574 #endif /* CONFIG_PROC_FS */
575