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;
45 static unsigned short int prof_shift;
47 int prof_on __read_mostly;
48 EXPORT_SYMBOL_GPL(prof_on);
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 */
57 int profile_setup(char *str)
59 static const char schedstr[] = "schedule";
60 static const char sleepstr[] = "sleep";
61 static const char kvmstr[] = "kvm";
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",
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",
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",
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",
101 __setup("profile=", profile_setup);
104 int __ref profile_init(void)
110 /* only text is profiled */
111 prof_len = (_etext - _stext) >> prof_shift;
112 buffer_bytes = prof_len*sizeof(atomic_t);
114 if (!alloc_cpumask_var(&prof_cpu_mask, GFP_KERNEL))
117 cpumask_copy(prof_cpu_mask, cpu_possible_mask);
119 prof_buffer = kzalloc(buffer_bytes, GFP_KERNEL|__GFP_NOWARN);
123 prof_buffer = alloc_pages_exact(buffer_bytes,
124 GFP_KERNEL|__GFP_ZERO|__GFP_NOWARN);
128 prof_buffer = vzalloc(buffer_bytes);
132 free_cpumask_var(prof_cpu_mask);
136 #if defined(CONFIG_SMP) && defined(CONFIG_PROC_FS)
138 * Each cpu has a pair of open-addressed hashtables for pending
139 * profile hits. read_profile() IPI's all cpus to request them
140 * to flip buffers and flushes their contents to prof_buffer itself.
141 * Flip requests are serialized by the profile_flip_mutex. The sole
142 * use of having a second hashtable is for avoiding cacheline
143 * contention that would otherwise happen during flushes of pending
144 * profile hits required for the accuracy of reported profile hits
145 * and so resurrect the interrupt livelock issue.
147 * The open-addressed hashtables are indexed by profile buffer slot
148 * and hold the number of pending hits to that profile buffer slot on
149 * a cpu in an entry. When the hashtable overflows, all pending hits
150 * are accounted to their corresponding profile buffer slots with
151 * atomic_add() and the hashtable emptied. As numerous pending hits
152 * may be accounted to a profile buffer slot in a hashtable entry,
153 * this amortizes a number of atomic profile buffer increments likely
154 * to be far larger than the number of entries in the hashtable,
155 * particularly given that the number of distinct profile buffer
156 * positions to which hits are accounted during short intervals (e.g.
157 * several seconds) is usually very small. Exclusion from buffer
158 * flipping is provided by interrupt disablement (note that for
159 * SCHED_PROFILING or SLEEP_PROFILING profile_hit() may be called from
161 * The hash function is meant to be lightweight as opposed to strong,
162 * and was vaguely inspired by ppc64 firmware-supported inverted
163 * pagetable hash functions, but uses a full hashtable full of finite
164 * collision chains, not just pairs of them.
168 static void __profile_flip_buffers(void *unused)
170 int cpu = smp_processor_id();
172 per_cpu(cpu_profile_flip, cpu) = !per_cpu(cpu_profile_flip, cpu);
175 static void profile_flip_buffers(void)
179 mutex_lock(&profile_flip_mutex);
180 j = per_cpu(cpu_profile_flip, get_cpu());
182 on_each_cpu(__profile_flip_buffers, NULL, 1);
183 for_each_online_cpu(cpu) {
184 struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[j];
185 for (i = 0; i < NR_PROFILE_HIT; ++i) {
191 atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
192 hits[i].hits = hits[i].pc = 0;
195 mutex_unlock(&profile_flip_mutex);
198 static void profile_discard_flip_buffers(void)
202 mutex_lock(&profile_flip_mutex);
203 i = per_cpu(cpu_profile_flip, get_cpu());
205 on_each_cpu(__profile_flip_buffers, NULL, 1);
206 for_each_online_cpu(cpu) {
207 struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[i];
208 memset(hits, 0, NR_PROFILE_HIT*sizeof(struct profile_hit));
210 mutex_unlock(&profile_flip_mutex);
213 static void do_profile_hits(int type, void *__pc, unsigned int nr_hits)
215 unsigned long primary, secondary, flags, pc = (unsigned long)__pc;
217 struct profile_hit *hits;
219 pc = min((pc - (unsigned long)_stext) >> prof_shift, prof_len - 1);
220 i = primary = (pc & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
221 secondary = (~(pc << 1) & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
223 hits = per_cpu(cpu_profile_hits, cpu)[per_cpu(cpu_profile_flip, cpu)];
229 * We buffer the global profiler buffer into a per-CPU
230 * queue and thus reduce the number of global (and possibly
231 * NUMA-alien) accesses. The write-queue is self-coalescing:
233 local_irq_save(flags);
235 for (j = 0; j < PROFILE_GRPSZ; ++j) {
236 if (hits[i + j].pc == pc) {
237 hits[i + j].hits += nr_hits;
239 } else if (!hits[i + j].hits) {
241 hits[i + j].hits = nr_hits;
245 i = (i + secondary) & (NR_PROFILE_HIT - 1);
246 } while (i != primary);
249 * Add the current hit(s) and flush the write-queue out
250 * to the global buffer:
252 atomic_add(nr_hits, &prof_buffer[pc]);
253 for (i = 0; i < NR_PROFILE_HIT; ++i) {
254 atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
255 hits[i].pc = hits[i].hits = 0;
258 local_irq_restore(flags);
262 static int profile_dead_cpu(unsigned int cpu)
267 if (cpumask_available(prof_cpu_mask))
268 cpumask_clear_cpu(cpu, prof_cpu_mask);
270 for (i = 0; i < 2; i++) {
271 if (per_cpu(cpu_profile_hits, cpu)[i]) {
272 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[i]);
273 per_cpu(cpu_profile_hits, cpu)[i] = NULL;
280 static int profile_prepare_cpu(unsigned int cpu)
282 int i, node = cpu_to_mem(cpu);
285 per_cpu(cpu_profile_flip, cpu) = 0;
287 for (i = 0; i < 2; i++) {
288 if (per_cpu(cpu_profile_hits, cpu)[i])
291 page = __alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
293 profile_dead_cpu(cpu);
296 per_cpu(cpu_profile_hits, cpu)[i] = page_address(page);
302 static int profile_online_cpu(unsigned int cpu)
304 if (cpumask_available(prof_cpu_mask))
305 cpumask_set_cpu(cpu, prof_cpu_mask);
310 #else /* !CONFIG_SMP */
311 #define profile_flip_buffers() do { } while (0)
312 #define profile_discard_flip_buffers() do { } while (0)
314 static void do_profile_hits(int type, void *__pc, unsigned int nr_hits)
317 pc = ((unsigned long)__pc - (unsigned long)_stext) >> prof_shift;
318 atomic_add(nr_hits, &prof_buffer[min(pc, prof_len - 1)]);
320 #endif /* !CONFIG_SMP */
322 void profile_hits(int type, void *__pc, unsigned int nr_hits)
324 if (prof_on != type || !prof_buffer)
326 do_profile_hits(type, __pc, nr_hits);
328 EXPORT_SYMBOL_GPL(profile_hits);
330 void profile_tick(int type)
332 struct pt_regs *regs = get_irq_regs();
334 if (!user_mode(regs) && cpumask_available(prof_cpu_mask) &&
335 cpumask_test_cpu(smp_processor_id(), prof_cpu_mask))
336 profile_hit(type, (void *)profile_pc(regs));
339 #ifdef CONFIG_PROC_FS
340 #include <linux/proc_fs.h>
341 #include <linux/seq_file.h>
342 #include <linux/uaccess.h>
344 static int prof_cpu_mask_proc_show(struct seq_file *m, void *v)
346 seq_printf(m, "%*pb\n", cpumask_pr_args(prof_cpu_mask));
350 static int prof_cpu_mask_proc_open(struct inode *inode, struct file *file)
352 return single_open(file, prof_cpu_mask_proc_show, NULL);
355 static ssize_t prof_cpu_mask_proc_write(struct file *file,
356 const char __user *buffer, size_t count, loff_t *pos)
358 cpumask_var_t new_value;
361 if (!zalloc_cpumask_var(&new_value, GFP_KERNEL))
364 err = cpumask_parse_user(buffer, count, new_value);
366 cpumask_copy(prof_cpu_mask, new_value);
369 free_cpumask_var(new_value);
373 static const struct proc_ops prof_cpu_mask_proc_ops = {
374 .proc_open = prof_cpu_mask_proc_open,
375 .proc_read = seq_read,
376 .proc_lseek = seq_lseek,
377 .proc_release = single_release,
378 .proc_write = prof_cpu_mask_proc_write,
381 void create_prof_cpu_mask(void)
383 /* create /proc/irq/prof_cpu_mask */
384 proc_create("irq/prof_cpu_mask", 0600, NULL, &prof_cpu_mask_proc_ops);
388 * This function accesses profiling information. The returned data is
389 * binary: the sampling step and the actual contents of the profile
390 * buffer. Use of the program readprofile is recommended in order to
391 * get meaningful info out of these data.
394 read_profile(struct file *file, char __user *buf, size_t count, loff_t *ppos)
396 unsigned long p = *ppos;
399 unsigned long sample_step = 1UL << prof_shift;
401 profile_flip_buffers();
402 if (p >= (prof_len+1)*sizeof(unsigned int))
404 if (count > (prof_len+1)*sizeof(unsigned int) - p)
405 count = (prof_len+1)*sizeof(unsigned int) - p;
408 while (p < sizeof(unsigned int) && count > 0) {
409 if (put_user(*((char *)(&sample_step)+p), buf))
411 buf++; p++; count--; read++;
413 pnt = (char *)prof_buffer + p - sizeof(atomic_t);
414 if (copy_to_user(buf, (void *)pnt, count))
422 * Writing to /proc/profile resets the counters
424 * Writing a 'profiling multiplier' value into it also re-sets the profiling
425 * interrupt frequency, on architectures that support this.
427 static ssize_t write_profile(struct file *file, const char __user *buf,
428 size_t count, loff_t *ppos)
431 extern int setup_profiling_timer(unsigned int multiplier);
433 if (count == sizeof(int)) {
434 unsigned int multiplier;
436 if (copy_from_user(&multiplier, buf, sizeof(int)))
439 if (setup_profiling_timer(multiplier))
443 profile_discard_flip_buffers();
444 memset(prof_buffer, 0, prof_len * sizeof(atomic_t));
448 static const struct proc_ops profile_proc_ops = {
449 .proc_read = read_profile,
450 .proc_write = write_profile,
451 .proc_lseek = default_llseek,
454 int __ref create_proc_profile(void)
456 struct proc_dir_entry *entry;
458 enum cpuhp_state online_state;
466 err = cpuhp_setup_state(CPUHP_PROFILE_PREPARE, "PROFILE_PREPARE",
467 profile_prepare_cpu, profile_dead_cpu);
471 err = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "AP_PROFILE_ONLINE",
472 profile_online_cpu, NULL);
478 entry = proc_create("profile", S_IWUSR | S_IRUGO,
479 NULL, &profile_proc_ops);
482 proc_set_size(entry, (1 + prof_len) * sizeof(atomic_t));
487 cpuhp_remove_state(online_state);
489 cpuhp_remove_state(CPUHP_PROFILE_PREPARE);
493 subsys_initcall(create_proc_profile);
494 #endif /* CONFIG_PROC_FS */