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;
114 pr_warn("profiling shift: %u too large\n", prof_shift);
119 buffer_bytes = prof_len*sizeof(atomic_t);
121 if (!alloc_cpumask_var(&prof_cpu_mask, GFP_KERNEL))
124 cpumask_copy(prof_cpu_mask, cpu_possible_mask);
126 prof_buffer = kzalloc(buffer_bytes, GFP_KERNEL|__GFP_NOWARN);
130 prof_buffer = alloc_pages_exact(buffer_bytes,
131 GFP_KERNEL|__GFP_ZERO|__GFP_NOWARN);
135 prof_buffer = vzalloc(buffer_bytes);
139 free_cpumask_var(prof_cpu_mask);
143 #if defined(CONFIG_SMP) && defined(CONFIG_PROC_FS)
145 * Each cpu has a pair of open-addressed hashtables for pending
146 * profile hits. read_profile() IPI's all cpus to request them
147 * to flip buffers and flushes their contents to prof_buffer itself.
148 * Flip requests are serialized by the profile_flip_mutex. The sole
149 * use of having a second hashtable is for avoiding cacheline
150 * contention that would otherwise happen during flushes of pending
151 * profile hits required for the accuracy of reported profile hits
152 * and so resurrect the interrupt livelock issue.
154 * The open-addressed hashtables are indexed by profile buffer slot
155 * and hold the number of pending hits to that profile buffer slot on
156 * a cpu in an entry. When the hashtable overflows, all pending hits
157 * are accounted to their corresponding profile buffer slots with
158 * atomic_add() and the hashtable emptied. As numerous pending hits
159 * may be accounted to a profile buffer slot in a hashtable entry,
160 * this amortizes a number of atomic profile buffer increments likely
161 * to be far larger than the number of entries in the hashtable,
162 * particularly given that the number of distinct profile buffer
163 * positions to which hits are accounted during short intervals (e.g.
164 * several seconds) is usually very small. Exclusion from buffer
165 * flipping is provided by interrupt disablement (note that for
166 * SCHED_PROFILING or SLEEP_PROFILING profile_hit() may be called from
168 * The hash function is meant to be lightweight as opposed to strong,
169 * and was vaguely inspired by ppc64 firmware-supported inverted
170 * pagetable hash functions, but uses a full hashtable full of finite
171 * collision chains, not just pairs of them.
175 static void __profile_flip_buffers(void *unused)
177 int cpu = smp_processor_id();
179 per_cpu(cpu_profile_flip, cpu) = !per_cpu(cpu_profile_flip, cpu);
182 static void profile_flip_buffers(void)
186 mutex_lock(&profile_flip_mutex);
187 j = per_cpu(cpu_profile_flip, get_cpu());
189 on_each_cpu(__profile_flip_buffers, NULL, 1);
190 for_each_online_cpu(cpu) {
191 struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[j];
192 for (i = 0; i < NR_PROFILE_HIT; ++i) {
198 atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
199 hits[i].hits = hits[i].pc = 0;
202 mutex_unlock(&profile_flip_mutex);
205 static void profile_discard_flip_buffers(void)
209 mutex_lock(&profile_flip_mutex);
210 i = per_cpu(cpu_profile_flip, get_cpu());
212 on_each_cpu(__profile_flip_buffers, NULL, 1);
213 for_each_online_cpu(cpu) {
214 struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[i];
215 memset(hits, 0, NR_PROFILE_HIT*sizeof(struct profile_hit));
217 mutex_unlock(&profile_flip_mutex);
220 static void do_profile_hits(int type, void *__pc, unsigned int nr_hits)
222 unsigned long primary, secondary, flags, pc = (unsigned long)__pc;
224 struct profile_hit *hits;
226 pc = min((pc - (unsigned long)_stext) >> prof_shift, prof_len - 1);
227 i = primary = (pc & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
228 secondary = (~(pc << 1) & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
230 hits = per_cpu(cpu_profile_hits, cpu)[per_cpu(cpu_profile_flip, cpu)];
236 * We buffer the global profiler buffer into a per-CPU
237 * queue and thus reduce the number of global (and possibly
238 * NUMA-alien) accesses. The write-queue is self-coalescing:
240 local_irq_save(flags);
242 for (j = 0; j < PROFILE_GRPSZ; ++j) {
243 if (hits[i + j].pc == pc) {
244 hits[i + j].hits += nr_hits;
246 } else if (!hits[i + j].hits) {
248 hits[i + j].hits = nr_hits;
252 i = (i + secondary) & (NR_PROFILE_HIT - 1);
253 } while (i != primary);
256 * Add the current hit(s) and flush the write-queue out
257 * to the global buffer:
259 atomic_add(nr_hits, &prof_buffer[pc]);
260 for (i = 0; i < NR_PROFILE_HIT; ++i) {
261 atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
262 hits[i].pc = hits[i].hits = 0;
265 local_irq_restore(flags);
269 static int profile_dead_cpu(unsigned int cpu)
274 if (cpumask_available(prof_cpu_mask))
275 cpumask_clear_cpu(cpu, prof_cpu_mask);
277 for (i = 0; i < 2; i++) {
278 if (per_cpu(cpu_profile_hits, cpu)[i]) {
279 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[i]);
280 per_cpu(cpu_profile_hits, cpu)[i] = NULL;
287 static int profile_prepare_cpu(unsigned int cpu)
289 int i, node = cpu_to_mem(cpu);
292 per_cpu(cpu_profile_flip, cpu) = 0;
294 for (i = 0; i < 2; i++) {
295 if (per_cpu(cpu_profile_hits, cpu)[i])
298 page = __alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
300 profile_dead_cpu(cpu);
303 per_cpu(cpu_profile_hits, cpu)[i] = page_address(page);
309 static int profile_online_cpu(unsigned int cpu)
311 if (cpumask_available(prof_cpu_mask))
312 cpumask_set_cpu(cpu, prof_cpu_mask);
317 #else /* !CONFIG_SMP */
318 #define profile_flip_buffers() do { } while (0)
319 #define profile_discard_flip_buffers() do { } while (0)
321 static void do_profile_hits(int type, void *__pc, unsigned int nr_hits)
324 pc = ((unsigned long)__pc - (unsigned long)_stext) >> prof_shift;
325 atomic_add(nr_hits, &prof_buffer[min(pc, prof_len - 1)]);
327 #endif /* !CONFIG_SMP */
329 void profile_hits(int type, void *__pc, unsigned int nr_hits)
331 if (prof_on != type || !prof_buffer)
333 do_profile_hits(type, __pc, nr_hits);
335 EXPORT_SYMBOL_GPL(profile_hits);
337 void profile_tick(int type)
339 struct pt_regs *regs = get_irq_regs();
341 if (!user_mode(regs) && cpumask_available(prof_cpu_mask) &&
342 cpumask_test_cpu(smp_processor_id(), prof_cpu_mask))
343 profile_hit(type, (void *)profile_pc(regs));
346 #ifdef CONFIG_PROC_FS
347 #include <linux/proc_fs.h>
348 #include <linux/seq_file.h>
349 #include <linux/uaccess.h>
351 static int prof_cpu_mask_proc_show(struct seq_file *m, void *v)
353 seq_printf(m, "%*pb\n", cpumask_pr_args(prof_cpu_mask));
357 static int prof_cpu_mask_proc_open(struct inode *inode, struct file *file)
359 return single_open(file, prof_cpu_mask_proc_show, NULL);
362 static ssize_t prof_cpu_mask_proc_write(struct file *file,
363 const char __user *buffer, size_t count, loff_t *pos)
365 cpumask_var_t new_value;
368 if (!zalloc_cpumask_var(&new_value, GFP_KERNEL))
371 err = cpumask_parse_user(buffer, count, new_value);
373 cpumask_copy(prof_cpu_mask, new_value);
376 free_cpumask_var(new_value);
380 static const struct proc_ops prof_cpu_mask_proc_ops = {
381 .proc_open = prof_cpu_mask_proc_open,
382 .proc_read = seq_read,
383 .proc_lseek = seq_lseek,
384 .proc_release = single_release,
385 .proc_write = prof_cpu_mask_proc_write,
388 void create_prof_cpu_mask(void)
390 /* create /proc/irq/prof_cpu_mask */
391 proc_create("irq/prof_cpu_mask", 0600, NULL, &prof_cpu_mask_proc_ops);
395 * This function accesses profiling information. The returned data is
396 * binary: the sampling step and the actual contents of the profile
397 * buffer. Use of the program readprofile is recommended in order to
398 * get meaningful info out of these data.
401 read_profile(struct file *file, char __user *buf, size_t count, loff_t *ppos)
403 unsigned long p = *ppos;
406 unsigned long sample_step = 1UL << prof_shift;
408 profile_flip_buffers();
409 if (p >= (prof_len+1)*sizeof(unsigned int))
411 if (count > (prof_len+1)*sizeof(unsigned int) - p)
412 count = (prof_len+1)*sizeof(unsigned int) - p;
415 while (p < sizeof(unsigned int) && count > 0) {
416 if (put_user(*((char *)(&sample_step)+p), buf))
418 buf++; p++; count--; read++;
420 pnt = (char *)prof_buffer + p - sizeof(atomic_t);
421 if (copy_to_user(buf, (void *)pnt, count))
429 * Writing to /proc/profile resets the counters
431 * Writing a 'profiling multiplier' value into it also re-sets the profiling
432 * interrupt frequency, on architectures that support this.
434 static ssize_t write_profile(struct file *file, const char __user *buf,
435 size_t count, loff_t *ppos)
438 extern int setup_profiling_timer(unsigned int multiplier);
440 if (count == sizeof(int)) {
441 unsigned int multiplier;
443 if (copy_from_user(&multiplier, buf, sizeof(int)))
446 if (setup_profiling_timer(multiplier))
450 profile_discard_flip_buffers();
451 memset(prof_buffer, 0, prof_len * sizeof(atomic_t));
455 static const struct proc_ops profile_proc_ops = {
456 .proc_read = read_profile,
457 .proc_write = write_profile,
458 .proc_lseek = default_llseek,
461 int __ref create_proc_profile(void)
463 struct proc_dir_entry *entry;
465 enum cpuhp_state online_state;
473 err = cpuhp_setup_state(CPUHP_PROFILE_PREPARE, "PROFILE_PREPARE",
474 profile_prepare_cpu, profile_dead_cpu);
478 err = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "AP_PROFILE_ONLINE",
479 profile_online_cpu, NULL);
485 entry = proc_create("profile", S_IWUSR | S_IRUGO,
486 NULL, &profile_proc_ops);
489 proc_set_size(entry, (1 + prof_len) * sizeof(atomic_t));
494 cpuhp_remove_state(online_state);
496 cpuhp_remove_state(CPUHP_PROFILE_PREPARE);
500 subsys_initcall(create_proc_profile);
501 #endif /* CONFIG_PROC_FS */