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";
62 const char *select = NULL;
65 if (!strncmp(str, sleepstr, strlen(sleepstr))) {
66 #ifdef CONFIG_SCHEDSTATS
67 force_schedstat_enabled();
68 prof_on = SLEEP_PROFILING;
71 pr_warn("kernel sleep profiling requires CONFIG_SCHEDSTATS\n");
72 #endif /* CONFIG_SCHEDSTATS */
73 } else if (!strncmp(str, schedstr, strlen(schedstr))) {
74 prof_on = SCHED_PROFILING;
76 } else if (!strncmp(str, kvmstr, strlen(kvmstr))) {
77 prof_on = KVM_PROFILING;
79 } else if (get_option(&str, &par)) {
80 prof_shift = clamp(par, 0, BITS_PER_LONG - 1);
81 prof_on = CPU_PROFILING;
82 pr_info("kernel profiling enabled (shift: %u)\n",
87 if (str[strlen(select)] == ',')
88 str += strlen(select) + 1;
89 if (get_option(&str, &par))
90 prof_shift = clamp(par, 0, BITS_PER_LONG - 1);
91 pr_info("kernel %s profiling enabled (shift: %u)\n",
97 __setup("profile=", profile_setup);
100 int __ref profile_init(void)
106 /* only text is profiled */
107 prof_len = (_etext - _stext) >> prof_shift;
110 pr_warn("profiling shift: %u too large\n", prof_shift);
115 buffer_bytes = prof_len*sizeof(atomic_t);
117 if (!alloc_cpumask_var(&prof_cpu_mask, GFP_KERNEL))
120 cpumask_copy(prof_cpu_mask, cpu_possible_mask);
122 prof_buffer = kzalloc(buffer_bytes, GFP_KERNEL|__GFP_NOWARN);
126 prof_buffer = alloc_pages_exact(buffer_bytes,
127 GFP_KERNEL|__GFP_ZERO|__GFP_NOWARN);
131 prof_buffer = vzalloc(buffer_bytes);
135 free_cpumask_var(prof_cpu_mask);
139 #if defined(CONFIG_SMP) && defined(CONFIG_PROC_FS)
141 * Each cpu has a pair of open-addressed hashtables for pending
142 * profile hits. read_profile() IPI's all cpus to request them
143 * to flip buffers and flushes their contents to prof_buffer itself.
144 * Flip requests are serialized by the profile_flip_mutex. The sole
145 * use of having a second hashtable is for avoiding cacheline
146 * contention that would otherwise happen during flushes of pending
147 * profile hits required for the accuracy of reported profile hits
148 * and so resurrect the interrupt livelock issue.
150 * The open-addressed hashtables are indexed by profile buffer slot
151 * and hold the number of pending hits to that profile buffer slot on
152 * a cpu in an entry. When the hashtable overflows, all pending hits
153 * are accounted to their corresponding profile buffer slots with
154 * atomic_add() and the hashtable emptied. As numerous pending hits
155 * may be accounted to a profile buffer slot in a hashtable entry,
156 * this amortizes a number of atomic profile buffer increments likely
157 * to be far larger than the number of entries in the hashtable,
158 * particularly given that the number of distinct profile buffer
159 * positions to which hits are accounted during short intervals (e.g.
160 * several seconds) is usually very small. Exclusion from buffer
161 * flipping is provided by interrupt disablement (note that for
162 * SCHED_PROFILING or SLEEP_PROFILING profile_hit() may be called from
164 * The hash function is meant to be lightweight as opposed to strong,
165 * and was vaguely inspired by ppc64 firmware-supported inverted
166 * pagetable hash functions, but uses a full hashtable full of finite
167 * collision chains, not just pairs of them.
171 static void __profile_flip_buffers(void *unused)
173 int cpu = smp_processor_id();
175 per_cpu(cpu_profile_flip, cpu) = !per_cpu(cpu_profile_flip, cpu);
178 static void profile_flip_buffers(void)
182 mutex_lock(&profile_flip_mutex);
183 j = per_cpu(cpu_profile_flip, get_cpu());
185 on_each_cpu(__profile_flip_buffers, NULL, 1);
186 for_each_online_cpu(cpu) {
187 struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[j];
188 for (i = 0; i < NR_PROFILE_HIT; ++i) {
194 atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
195 hits[i].hits = hits[i].pc = 0;
198 mutex_unlock(&profile_flip_mutex);
201 static void profile_discard_flip_buffers(void)
205 mutex_lock(&profile_flip_mutex);
206 i = per_cpu(cpu_profile_flip, get_cpu());
208 on_each_cpu(__profile_flip_buffers, NULL, 1);
209 for_each_online_cpu(cpu) {
210 struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[i];
211 memset(hits, 0, NR_PROFILE_HIT*sizeof(struct profile_hit));
213 mutex_unlock(&profile_flip_mutex);
216 static void do_profile_hits(int type, void *__pc, unsigned int nr_hits)
218 unsigned long primary, secondary, flags, pc = (unsigned long)__pc;
220 struct profile_hit *hits;
222 pc = min((pc - (unsigned long)_stext) >> prof_shift, prof_len - 1);
223 i = primary = (pc & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
224 secondary = (~(pc << 1) & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
226 hits = per_cpu(cpu_profile_hits, cpu)[per_cpu(cpu_profile_flip, cpu)];
232 * We buffer the global profiler buffer into a per-CPU
233 * queue and thus reduce the number of global (and possibly
234 * NUMA-alien) accesses. The write-queue is self-coalescing:
236 local_irq_save(flags);
238 for (j = 0; j < PROFILE_GRPSZ; ++j) {
239 if (hits[i + j].pc == pc) {
240 hits[i + j].hits += nr_hits;
242 } else if (!hits[i + j].hits) {
244 hits[i + j].hits = nr_hits;
248 i = (i + secondary) & (NR_PROFILE_HIT - 1);
249 } while (i != primary);
252 * Add the current hit(s) and flush the write-queue out
253 * to the global buffer:
255 atomic_add(nr_hits, &prof_buffer[pc]);
256 for (i = 0; i < NR_PROFILE_HIT; ++i) {
257 atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
258 hits[i].pc = hits[i].hits = 0;
261 local_irq_restore(flags);
265 static int profile_dead_cpu(unsigned int cpu)
270 if (cpumask_available(prof_cpu_mask))
271 cpumask_clear_cpu(cpu, prof_cpu_mask);
273 for (i = 0; i < 2; i++) {
274 if (per_cpu(cpu_profile_hits, cpu)[i]) {
275 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[i]);
276 per_cpu(cpu_profile_hits, cpu)[i] = NULL;
283 static int profile_prepare_cpu(unsigned int cpu)
285 int i, node = cpu_to_mem(cpu);
288 per_cpu(cpu_profile_flip, cpu) = 0;
290 for (i = 0; i < 2; i++) {
291 if (per_cpu(cpu_profile_hits, cpu)[i])
294 page = __alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
296 profile_dead_cpu(cpu);
299 per_cpu(cpu_profile_hits, cpu)[i] = page_address(page);
305 static int profile_online_cpu(unsigned int cpu)
307 if (cpumask_available(prof_cpu_mask))
308 cpumask_set_cpu(cpu, prof_cpu_mask);
313 #else /* !CONFIG_SMP */
314 #define profile_flip_buffers() do { } while (0)
315 #define profile_discard_flip_buffers() do { } while (0)
317 static void do_profile_hits(int type, void *__pc, unsigned int nr_hits)
320 pc = ((unsigned long)__pc - (unsigned long)_stext) >> prof_shift;
321 atomic_add(nr_hits, &prof_buffer[min(pc, prof_len - 1)]);
323 #endif /* !CONFIG_SMP */
325 void profile_hits(int type, void *__pc, unsigned int nr_hits)
327 if (prof_on != type || !prof_buffer)
329 do_profile_hits(type, __pc, nr_hits);
331 EXPORT_SYMBOL_GPL(profile_hits);
333 void profile_tick(int type)
335 struct pt_regs *regs = get_irq_regs();
337 if (!user_mode(regs) && cpumask_available(prof_cpu_mask) &&
338 cpumask_test_cpu(smp_processor_id(), prof_cpu_mask))
339 profile_hit(type, (void *)profile_pc(regs));
342 #ifdef CONFIG_PROC_FS
343 #include <linux/proc_fs.h>
344 #include <linux/seq_file.h>
345 #include <linux/uaccess.h>
347 static int prof_cpu_mask_proc_show(struct seq_file *m, void *v)
349 seq_printf(m, "%*pb\n", cpumask_pr_args(prof_cpu_mask));
353 static int prof_cpu_mask_proc_open(struct inode *inode, struct file *file)
355 return single_open(file, prof_cpu_mask_proc_show, NULL);
358 static ssize_t prof_cpu_mask_proc_write(struct file *file,
359 const char __user *buffer, size_t count, loff_t *pos)
361 cpumask_var_t new_value;
364 if (!zalloc_cpumask_var(&new_value, GFP_KERNEL))
367 err = cpumask_parse_user(buffer, count, new_value);
369 cpumask_copy(prof_cpu_mask, new_value);
372 free_cpumask_var(new_value);
376 static const struct proc_ops prof_cpu_mask_proc_ops = {
377 .proc_open = prof_cpu_mask_proc_open,
378 .proc_read = seq_read,
379 .proc_lseek = seq_lseek,
380 .proc_release = single_release,
381 .proc_write = prof_cpu_mask_proc_write,
384 void create_prof_cpu_mask(void)
386 /* create /proc/irq/prof_cpu_mask */
387 proc_create("irq/prof_cpu_mask", 0600, NULL, &prof_cpu_mask_proc_ops);
391 * This function accesses profiling information. The returned data is
392 * binary: the sampling step and the actual contents of the profile
393 * buffer. Use of the program readprofile is recommended in order to
394 * get meaningful info out of these data.
397 read_profile(struct file *file, char __user *buf, size_t count, loff_t *ppos)
399 unsigned long p = *ppos;
402 unsigned long sample_step = 1UL << prof_shift;
404 profile_flip_buffers();
405 if (p >= (prof_len+1)*sizeof(unsigned int))
407 if (count > (prof_len+1)*sizeof(unsigned int) - p)
408 count = (prof_len+1)*sizeof(unsigned int) - p;
411 while (p < sizeof(unsigned int) && count > 0) {
412 if (put_user(*((char *)(&sample_step)+p), buf))
414 buf++; p++; count--; read++;
416 pnt = (char *)prof_buffer + p - sizeof(atomic_t);
417 if (copy_to_user(buf, (void *)pnt, count))
424 /* default is to not implement this call */
425 int __weak setup_profiling_timer(unsigned mult)
431 * Writing to /proc/profile resets the counters
433 * Writing a 'profiling multiplier' value into it also re-sets the profiling
434 * interrupt frequency, on architectures that support this.
436 static ssize_t write_profile(struct file *file, const char __user *buf,
437 size_t count, loff_t *ppos)
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 */