1 // SPDX-License-Identifier: GPL-2.0-only
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.
9 * Scheduler profiling support, Arjan van de Ven and Ingo Molnar,
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,
18 #include <linux/export.h>
19 #include <linux/profile.h>
20 #include <linux/memblock.h>
21 #include <linux/notifier.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>
31 #include <asm/sections.h>
32 #include <asm/irq_regs.h>
33 #include <asm/ptrace.h>
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)
43 static atomic_t *prof_buffer;
44 static unsigned long prof_len, prof_shift;
46 int prof_on __read_mostly;
47 EXPORT_SYMBOL_GPL(prof_on);
49 static cpumask_var_t prof_cpu_mask;
50 #if defined(CONFIG_SMP) && defined(CONFIG_PROC_FS)
51 static DEFINE_PER_CPU(struct profile_hit *[2], cpu_profile_hits);
52 static DEFINE_PER_CPU(int, cpu_profile_flip);
53 static DEFINE_MUTEX(profile_flip_mutex);
54 #endif /* CONFIG_SMP */
56 int profile_setup(char *str)
58 static const char schedstr[] = "schedule";
59 static const char sleepstr[] = "sleep";
60 static const char kvmstr[] = "kvm";
63 if (!strncmp(str, sleepstr, strlen(sleepstr))) {
64 #ifdef CONFIG_SCHEDSTATS
65 force_schedstat_enabled();
66 prof_on = SLEEP_PROFILING;
67 if (str[strlen(sleepstr)] == ',')
68 str += strlen(sleepstr) + 1;
69 if (get_option(&str, &par))
71 pr_info("kernel sleep profiling enabled (shift: %ld)\n",
74 pr_warn("kernel sleep profiling requires CONFIG_SCHEDSTATS\n");
75 #endif /* CONFIG_SCHEDSTATS */
76 } else if (!strncmp(str, schedstr, strlen(schedstr))) {
77 prof_on = SCHED_PROFILING;
78 if (str[strlen(schedstr)] == ',')
79 str += strlen(schedstr) + 1;
80 if (get_option(&str, &par))
82 pr_info("kernel schedule profiling enabled (shift: %ld)\n",
84 } else if (!strncmp(str, kvmstr, strlen(kvmstr))) {
85 prof_on = KVM_PROFILING;
86 if (str[strlen(kvmstr)] == ',')
87 str += strlen(kvmstr) + 1;
88 if (get_option(&str, &par))
90 pr_info("kernel KVM profiling enabled (shift: %ld)\n",
92 } else if (get_option(&str, &par)) {
94 prof_on = CPU_PROFILING;
95 pr_info("kernel profiling enabled (shift: %ld)\n",
100 __setup("profile=", profile_setup);
103 int __ref profile_init(void)
109 /* only text is profiled */
110 prof_len = (_etext - _stext) >> prof_shift;
111 buffer_bytes = prof_len*sizeof(atomic_t);
113 if (!alloc_cpumask_var(&prof_cpu_mask, GFP_KERNEL))
116 cpumask_copy(prof_cpu_mask, cpu_possible_mask);
118 prof_buffer = kzalloc(buffer_bytes, GFP_KERNEL|__GFP_NOWARN);
122 prof_buffer = alloc_pages_exact(buffer_bytes,
123 GFP_KERNEL|__GFP_ZERO|__GFP_NOWARN);
127 prof_buffer = vzalloc(buffer_bytes);
131 free_cpumask_var(prof_cpu_mask);
135 /* Profile event notifications */
137 static BLOCKING_NOTIFIER_HEAD(task_exit_notifier);
138 static ATOMIC_NOTIFIER_HEAD(task_free_notifier);
139 static BLOCKING_NOTIFIER_HEAD(munmap_notifier);
141 void profile_task_exit(struct task_struct *task)
143 blocking_notifier_call_chain(&task_exit_notifier, 0, task);
146 int profile_handoff_task(struct task_struct *task)
149 ret = atomic_notifier_call_chain(&task_free_notifier, 0, task);
150 return (ret == NOTIFY_OK) ? 1 : 0;
153 void profile_munmap(unsigned long addr)
155 blocking_notifier_call_chain(&munmap_notifier, 0, (void *)addr);
158 int task_handoff_register(struct notifier_block *n)
160 return atomic_notifier_chain_register(&task_free_notifier, n);
162 EXPORT_SYMBOL_GPL(task_handoff_register);
164 int task_handoff_unregister(struct notifier_block *n)
166 return atomic_notifier_chain_unregister(&task_free_notifier, n);
168 EXPORT_SYMBOL_GPL(task_handoff_unregister);
170 int profile_event_register(enum profile_type type, struct notifier_block *n)
175 case PROFILE_TASK_EXIT:
176 err = blocking_notifier_chain_register(
177 &task_exit_notifier, n);
180 err = blocking_notifier_chain_register(
181 &munmap_notifier, n);
187 EXPORT_SYMBOL_GPL(profile_event_register);
189 int profile_event_unregister(enum profile_type type, struct notifier_block *n)
194 case PROFILE_TASK_EXIT:
195 err = blocking_notifier_chain_unregister(
196 &task_exit_notifier, n);
199 err = blocking_notifier_chain_unregister(
200 &munmap_notifier, n);
206 EXPORT_SYMBOL_GPL(profile_event_unregister);
208 #if defined(CONFIG_SMP) && defined(CONFIG_PROC_FS)
210 * Each cpu has a pair of open-addressed hashtables for pending
211 * profile hits. read_profile() IPI's all cpus to request them
212 * to flip buffers and flushes their contents to prof_buffer itself.
213 * Flip requests are serialized by the profile_flip_mutex. The sole
214 * use of having a second hashtable is for avoiding cacheline
215 * contention that would otherwise happen during flushes of pending
216 * profile hits required for the accuracy of reported profile hits
217 * and so resurrect the interrupt livelock issue.
219 * The open-addressed hashtables are indexed by profile buffer slot
220 * and hold the number of pending hits to that profile buffer slot on
221 * a cpu in an entry. When the hashtable overflows, all pending hits
222 * are accounted to their corresponding profile buffer slots with
223 * atomic_add() and the hashtable emptied. As numerous pending hits
224 * may be accounted to a profile buffer slot in a hashtable entry,
225 * this amortizes a number of atomic profile buffer increments likely
226 * to be far larger than the number of entries in the hashtable,
227 * particularly given that the number of distinct profile buffer
228 * positions to which hits are accounted during short intervals (e.g.
229 * several seconds) is usually very small. Exclusion from buffer
230 * flipping is provided by interrupt disablement (note that for
231 * SCHED_PROFILING or SLEEP_PROFILING profile_hit() may be called from
233 * The hash function is meant to be lightweight as opposed to strong,
234 * and was vaguely inspired by ppc64 firmware-supported inverted
235 * pagetable hash functions, but uses a full hashtable full of finite
236 * collision chains, not just pairs of them.
240 static void __profile_flip_buffers(void *unused)
242 int cpu = smp_processor_id();
244 per_cpu(cpu_profile_flip, cpu) = !per_cpu(cpu_profile_flip, cpu);
247 static void profile_flip_buffers(void)
251 mutex_lock(&profile_flip_mutex);
252 j = per_cpu(cpu_profile_flip, get_cpu());
254 on_each_cpu(__profile_flip_buffers, NULL, 1);
255 for_each_online_cpu(cpu) {
256 struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[j];
257 for (i = 0; i < NR_PROFILE_HIT; ++i) {
263 atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
264 hits[i].hits = hits[i].pc = 0;
267 mutex_unlock(&profile_flip_mutex);
270 static void profile_discard_flip_buffers(void)
274 mutex_lock(&profile_flip_mutex);
275 i = per_cpu(cpu_profile_flip, get_cpu());
277 on_each_cpu(__profile_flip_buffers, NULL, 1);
278 for_each_online_cpu(cpu) {
279 struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[i];
280 memset(hits, 0, NR_PROFILE_HIT*sizeof(struct profile_hit));
282 mutex_unlock(&profile_flip_mutex);
285 static void do_profile_hits(int type, void *__pc, unsigned int nr_hits)
287 unsigned long primary, secondary, flags, pc = (unsigned long)__pc;
289 struct profile_hit *hits;
291 pc = min((pc - (unsigned long)_stext) >> prof_shift, prof_len - 1);
292 i = primary = (pc & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
293 secondary = (~(pc << 1) & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
295 hits = per_cpu(cpu_profile_hits, cpu)[per_cpu(cpu_profile_flip, cpu)];
301 * We buffer the global profiler buffer into a per-CPU
302 * queue and thus reduce the number of global (and possibly
303 * NUMA-alien) accesses. The write-queue is self-coalescing:
305 local_irq_save(flags);
307 for (j = 0; j < PROFILE_GRPSZ; ++j) {
308 if (hits[i + j].pc == pc) {
309 hits[i + j].hits += nr_hits;
311 } else if (!hits[i + j].hits) {
313 hits[i + j].hits = nr_hits;
317 i = (i + secondary) & (NR_PROFILE_HIT - 1);
318 } while (i != primary);
321 * Add the current hit(s) and flush the write-queue out
322 * to the global buffer:
324 atomic_add(nr_hits, &prof_buffer[pc]);
325 for (i = 0; i < NR_PROFILE_HIT; ++i) {
326 atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
327 hits[i].pc = hits[i].hits = 0;
330 local_irq_restore(flags);
334 static int profile_dead_cpu(unsigned int cpu)
339 if (cpumask_available(prof_cpu_mask))
340 cpumask_clear_cpu(cpu, prof_cpu_mask);
342 for (i = 0; i < 2; i++) {
343 if (per_cpu(cpu_profile_hits, cpu)[i]) {
344 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[i]);
345 per_cpu(cpu_profile_hits, cpu)[i] = NULL;
352 static int profile_prepare_cpu(unsigned int cpu)
354 int i, node = cpu_to_mem(cpu);
357 per_cpu(cpu_profile_flip, cpu) = 0;
359 for (i = 0; i < 2; i++) {
360 if (per_cpu(cpu_profile_hits, cpu)[i])
363 page = __alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
365 profile_dead_cpu(cpu);
368 per_cpu(cpu_profile_hits, cpu)[i] = page_address(page);
374 static int profile_online_cpu(unsigned int cpu)
376 if (cpumask_available(prof_cpu_mask))
377 cpumask_set_cpu(cpu, prof_cpu_mask);
382 #else /* !CONFIG_SMP */
383 #define profile_flip_buffers() do { } while (0)
384 #define profile_discard_flip_buffers() do { } while (0)
386 static void do_profile_hits(int type, void *__pc, unsigned int nr_hits)
389 pc = ((unsigned long)__pc - (unsigned long)_stext) >> prof_shift;
390 atomic_add(nr_hits, &prof_buffer[min(pc, prof_len - 1)]);
392 #endif /* !CONFIG_SMP */
394 void profile_hits(int type, void *__pc, unsigned int nr_hits)
396 if (prof_on != type || !prof_buffer)
398 do_profile_hits(type, __pc, nr_hits);
400 EXPORT_SYMBOL_GPL(profile_hits);
402 void profile_tick(int type)
404 struct pt_regs *regs = get_irq_regs();
406 if (!user_mode(regs) && cpumask_available(prof_cpu_mask) &&
407 cpumask_test_cpu(smp_processor_id(), prof_cpu_mask))
408 profile_hit(type, (void *)profile_pc(regs));
411 #ifdef CONFIG_PROC_FS
412 #include <linux/proc_fs.h>
413 #include <linux/seq_file.h>
414 #include <linux/uaccess.h>
416 static int prof_cpu_mask_proc_show(struct seq_file *m, void *v)
418 seq_printf(m, "%*pb\n", cpumask_pr_args(prof_cpu_mask));
422 static int prof_cpu_mask_proc_open(struct inode *inode, struct file *file)
424 return single_open(file, prof_cpu_mask_proc_show, NULL);
427 static ssize_t prof_cpu_mask_proc_write(struct file *file,
428 const char __user *buffer, size_t count, loff_t *pos)
430 cpumask_var_t new_value;
433 if (!alloc_cpumask_var(&new_value, GFP_KERNEL))
436 err = cpumask_parse_user(buffer, count, new_value);
438 cpumask_copy(prof_cpu_mask, new_value);
441 free_cpumask_var(new_value);
445 static const struct proc_ops prof_cpu_mask_proc_ops = {
446 .proc_open = prof_cpu_mask_proc_open,
447 .proc_read = seq_read,
448 .proc_lseek = seq_lseek,
449 .proc_release = single_release,
450 .proc_write = prof_cpu_mask_proc_write,
453 void create_prof_cpu_mask(void)
455 /* create /proc/irq/prof_cpu_mask */
456 proc_create("irq/prof_cpu_mask", 0600, NULL, &prof_cpu_mask_proc_ops);
460 * This function accesses profiling information. The returned data is
461 * binary: the sampling step and the actual contents of the profile
462 * buffer. Use of the program readprofile is recommended in order to
463 * get meaningful info out of these data.
466 read_profile(struct file *file, char __user *buf, size_t count, loff_t *ppos)
468 unsigned long p = *ppos;
471 unsigned int sample_step = 1 << prof_shift;
473 profile_flip_buffers();
474 if (p >= (prof_len+1)*sizeof(unsigned int))
476 if (count > (prof_len+1)*sizeof(unsigned int) - p)
477 count = (prof_len+1)*sizeof(unsigned int) - p;
480 while (p < sizeof(unsigned int) && count > 0) {
481 if (put_user(*((char *)(&sample_step)+p), buf))
483 buf++; p++; count--; read++;
485 pnt = (char *)prof_buffer + p - sizeof(atomic_t);
486 if (copy_to_user(buf, (void *)pnt, count))
494 * Writing to /proc/profile resets the counters
496 * Writing a 'profiling multiplier' value into it also re-sets the profiling
497 * interrupt frequency, on architectures that support this.
499 static ssize_t write_profile(struct file *file, const char __user *buf,
500 size_t count, loff_t *ppos)
503 extern int setup_profiling_timer(unsigned int multiplier);
505 if (count == sizeof(int)) {
506 unsigned int multiplier;
508 if (copy_from_user(&multiplier, buf, sizeof(int)))
511 if (setup_profiling_timer(multiplier))
515 profile_discard_flip_buffers();
516 memset(prof_buffer, 0, prof_len * sizeof(atomic_t));
520 static const struct proc_ops profile_proc_ops = {
521 .proc_read = read_profile,
522 .proc_write = write_profile,
523 .proc_lseek = default_llseek,
526 int __ref create_proc_profile(void)
528 struct proc_dir_entry *entry;
530 enum cpuhp_state online_state;
538 err = cpuhp_setup_state(CPUHP_PROFILE_PREPARE, "PROFILE_PREPARE",
539 profile_prepare_cpu, profile_dead_cpu);
543 err = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "AP_PROFILE_ONLINE",
544 profile_online_cpu, NULL);
550 entry = proc_create("profile", S_IWUSR | S_IRUGO,
551 NULL, &profile_proc_ops);
554 proc_set_size(entry, (1 + prof_len) * sizeof(atomic_t));
559 cpuhp_remove_state(online_state);
561 cpuhp_remove_state(CPUHP_PROFILE_PREPARE);
565 subsys_initcall(create_proc_profile);
566 #endif /* CONFIG_PROC_FS */