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 | |
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 | const char *select = NULL; |
63 | int par; |
64 | |
65 | if (!strncmp(str, sleepstr, strlen(sleepstr))) { |
66 | #ifdef CONFIG_SCHEDSTATS |
67 | force_schedstat_enabled(); |
68 | prof_on = SLEEP_PROFILING; |
69 | select = sleepstr; |
70 | #else |
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; |
75 | select = schedstr; |
76 | } else if (!strncmp(str, kvmstr, strlen(kvmstr))) { |
77 | prof_on = KVM_PROFILING; |
78 | select = kvmstr; |
79 | } else if (get_option(str: &str, pint: &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" , |
83 | prof_shift); |
84 | } |
85 | |
86 | if (select) { |
87 | if (str[strlen(select)] == ',') |
88 | str += strlen(select) + 1; |
89 | if (get_option(str: &str, pint: &par)) |
90 | prof_shift = clamp(par, 0, BITS_PER_LONG - 1); |
91 | pr_info("kernel %s profiling enabled (shift: %u)\n" , |
92 | select, prof_shift); |
93 | } |
94 | |
95 | return 1; |
96 | } |
97 | __setup("profile=" , profile_setup); |
98 | |
99 | |
100 | int __ref profile_init(void) |
101 | { |
102 | int buffer_bytes; |
103 | if (!prof_on) |
104 | return 0; |
105 | |
106 | /* only text is profiled */ |
107 | prof_len = (_etext - _stext) >> prof_shift; |
108 | |
109 | if (!prof_len) { |
110 | pr_warn("profiling shift: %u too large\n" , prof_shift); |
111 | prof_on = 0; |
112 | return -EINVAL; |
113 | } |
114 | |
115 | buffer_bytes = prof_len*sizeof(atomic_t); |
116 | |
117 | if (!alloc_cpumask_var(mask: &prof_cpu_mask, GFP_KERNEL)) |
118 | return -ENOMEM; |
119 | |
120 | cpumask_copy(dstp: prof_cpu_mask, cpu_possible_mask); |
121 | |
122 | prof_buffer = kzalloc(size: buffer_bytes, GFP_KERNEL|__GFP_NOWARN); |
123 | if (prof_buffer) |
124 | return 0; |
125 | |
126 | prof_buffer = alloc_pages_exact(size: buffer_bytes, |
127 | GFP_KERNEL|__GFP_ZERO|__GFP_NOWARN); |
128 | if (prof_buffer) |
129 | return 0; |
130 | |
131 | prof_buffer = vzalloc(size: buffer_bytes); |
132 | if (prof_buffer) |
133 | return 0; |
134 | |
135 | free_cpumask_var(mask: prof_cpu_mask); |
136 | return -ENOMEM; |
137 | } |
138 | |
139 | #if defined(CONFIG_SMP) && defined(CONFIG_PROC_FS) |
140 | /* |
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. |
149 | * |
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 |
163 | * process context). |
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. |
168 | * |
169 | * -- nyc |
170 | */ |
171 | static void __profile_flip_buffers(void *unused) |
172 | { |
173 | int cpu = smp_processor_id(); |
174 | |
175 | per_cpu(cpu_profile_flip, cpu) = !per_cpu(cpu_profile_flip, cpu); |
176 | } |
177 | |
178 | static void profile_flip_buffers(void) |
179 | { |
180 | int i, j, cpu; |
181 | |
182 | mutex_lock(&profile_flip_mutex); |
183 | j = per_cpu(cpu_profile_flip, get_cpu()); |
184 | put_cpu(); |
185 | on_each_cpu(func: __profile_flip_buffers, NULL, wait: 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) { |
189 | if (!hits[i].hits) { |
190 | if (hits[i].pc) |
191 | hits[i].pc = 0; |
192 | continue; |
193 | } |
194 | atomic_add(i: hits[i].hits, v: &prof_buffer[hits[i].pc]); |
195 | hits[i].hits = hits[i].pc = 0; |
196 | } |
197 | } |
198 | mutex_unlock(lock: &profile_flip_mutex); |
199 | } |
200 | |
201 | static void profile_discard_flip_buffers(void) |
202 | { |
203 | int i, cpu; |
204 | |
205 | mutex_lock(&profile_flip_mutex); |
206 | i = per_cpu(cpu_profile_flip, get_cpu()); |
207 | put_cpu(); |
208 | on_each_cpu(func: __profile_flip_buffers, NULL, wait: 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)); |
212 | } |
213 | mutex_unlock(lock: &profile_flip_mutex); |
214 | } |
215 | |
216 | static void do_profile_hits(int type, void *__pc, unsigned int nr_hits) |
217 | { |
218 | unsigned long primary, secondary, flags, pc = (unsigned long)__pc; |
219 | int i, j, cpu; |
220 | struct profile_hit *hits; |
221 | |
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; |
225 | cpu = get_cpu(); |
226 | hits = per_cpu(cpu_profile_hits, cpu)[per_cpu(cpu_profile_flip, cpu)]; |
227 | if (!hits) { |
228 | put_cpu(); |
229 | return; |
230 | } |
231 | /* |
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: |
235 | */ |
236 | local_irq_save(flags); |
237 | do { |
238 | for (j = 0; j < PROFILE_GRPSZ; ++j) { |
239 | if (hits[i + j].pc == pc) { |
240 | hits[i + j].hits += nr_hits; |
241 | goto out; |
242 | } else if (!hits[i + j].hits) { |
243 | hits[i + j].pc = pc; |
244 | hits[i + j].hits = nr_hits; |
245 | goto out; |
246 | } |
247 | } |
248 | i = (i + secondary) & (NR_PROFILE_HIT - 1); |
249 | } while (i != primary); |
250 | |
251 | /* |
252 | * Add the current hit(s) and flush the write-queue out |
253 | * to the global buffer: |
254 | */ |
255 | atomic_add(i: nr_hits, v: &prof_buffer[pc]); |
256 | for (i = 0; i < NR_PROFILE_HIT; ++i) { |
257 | atomic_add(i: hits[i].hits, v: &prof_buffer[hits[i].pc]); |
258 | hits[i].pc = hits[i].hits = 0; |
259 | } |
260 | out: |
261 | local_irq_restore(flags); |
262 | put_cpu(); |
263 | } |
264 | |
265 | static int profile_dead_cpu(unsigned int cpu) |
266 | { |
267 | struct page *page; |
268 | int i; |
269 | |
270 | if (cpumask_available(mask: prof_cpu_mask)) |
271 | cpumask_clear_cpu(cpu, dstp: prof_cpu_mask); |
272 | |
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; |
277 | __free_page(page); |
278 | } |
279 | } |
280 | return 0; |
281 | } |
282 | |
283 | static int profile_prepare_cpu(unsigned int cpu) |
284 | { |
285 | int i, node = cpu_to_mem(cpu); |
286 | struct page *page; |
287 | |
288 | per_cpu(cpu_profile_flip, cpu) = 0; |
289 | |
290 | for (i = 0; i < 2; i++) { |
291 | if (per_cpu(cpu_profile_hits, cpu)[i]) |
292 | continue; |
293 | |
294 | page = __alloc_pages_node(nid: node, GFP_KERNEL | __GFP_ZERO, order: 0); |
295 | if (!page) { |
296 | profile_dead_cpu(cpu); |
297 | return -ENOMEM; |
298 | } |
299 | per_cpu(cpu_profile_hits, cpu)[i] = page_address(page); |
300 | |
301 | } |
302 | return 0; |
303 | } |
304 | |
305 | static int profile_online_cpu(unsigned int cpu) |
306 | { |
307 | if (cpumask_available(mask: prof_cpu_mask)) |
308 | cpumask_set_cpu(cpu, dstp: prof_cpu_mask); |
309 | |
310 | return 0; |
311 | } |
312 | |
313 | #else /* !CONFIG_SMP */ |
314 | #define profile_flip_buffers() do { } while (0) |
315 | #define profile_discard_flip_buffers() do { } while (0) |
316 | |
317 | static void do_profile_hits(int type, void *__pc, unsigned int nr_hits) |
318 | { |
319 | unsigned long pc; |
320 | pc = ((unsigned long)__pc - (unsigned long)_stext) >> prof_shift; |
321 | atomic_add(nr_hits, &prof_buffer[min(pc, prof_len - 1)]); |
322 | } |
323 | #endif /* !CONFIG_SMP */ |
324 | |
325 | void profile_hits(int type, void *__pc, unsigned int nr_hits) |
326 | { |
327 | if (prof_on != type || !prof_buffer) |
328 | return; |
329 | do_profile_hits(type, __pc, nr_hits); |
330 | } |
331 | EXPORT_SYMBOL_GPL(profile_hits); |
332 | |
333 | void profile_tick(int type) |
334 | { |
335 | struct pt_regs *regs = get_irq_regs(); |
336 | |
337 | if (!user_mode(regs) && cpumask_available(mask: prof_cpu_mask) && |
338 | cpumask_test_cpu(smp_processor_id(), cpumask: prof_cpu_mask)) |
339 | profile_hit(type, ip: (void *)profile_pc(regs)); |
340 | } |
341 | |
342 | #ifdef CONFIG_PROC_FS |
343 | #include <linux/proc_fs.h> |
344 | #include <linux/seq_file.h> |
345 | #include <linux/uaccess.h> |
346 | |
347 | static int prof_cpu_mask_proc_show(struct seq_file *m, void *v) |
348 | { |
349 | seq_printf(m, fmt: "%*pb\n" , cpumask_pr_args(prof_cpu_mask)); |
350 | return 0; |
351 | } |
352 | |
353 | static int prof_cpu_mask_proc_open(struct inode *inode, struct file *file) |
354 | { |
355 | return single_open(file, prof_cpu_mask_proc_show, NULL); |
356 | } |
357 | |
358 | static ssize_t prof_cpu_mask_proc_write(struct file *file, |
359 | const char __user *buffer, size_t count, loff_t *pos) |
360 | { |
361 | cpumask_var_t new_value; |
362 | int err; |
363 | |
364 | if (!zalloc_cpumask_var(mask: &new_value, GFP_KERNEL)) |
365 | return -ENOMEM; |
366 | |
367 | err = cpumask_parse_user(buf: buffer, len: count, dstp: new_value); |
368 | if (!err) { |
369 | cpumask_copy(dstp: prof_cpu_mask, srcp: new_value); |
370 | err = count; |
371 | } |
372 | free_cpumask_var(mask: new_value); |
373 | return err; |
374 | } |
375 | |
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, |
382 | }; |
383 | |
384 | void create_prof_cpu_mask(void) |
385 | { |
386 | /* create /proc/irq/prof_cpu_mask */ |
387 | proc_create(name: "irq/prof_cpu_mask" , mode: 0600, NULL, proc_ops: &prof_cpu_mask_proc_ops); |
388 | } |
389 | |
390 | /* |
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. |
395 | */ |
396 | static ssize_t |
397 | read_profile(struct file *file, char __user *buf, size_t count, loff_t *ppos) |
398 | { |
399 | unsigned long p = *ppos; |
400 | ssize_t read; |
401 | char *pnt; |
402 | unsigned long sample_step = 1UL << prof_shift; |
403 | |
404 | profile_flip_buffers(); |
405 | if (p >= (prof_len+1)*sizeof(unsigned int)) |
406 | return 0; |
407 | if (count > (prof_len+1)*sizeof(unsigned int) - p) |
408 | count = (prof_len+1)*sizeof(unsigned int) - p; |
409 | read = 0; |
410 | |
411 | while (p < sizeof(unsigned int) && count > 0) { |
412 | if (put_user(*((char *)(&sample_step)+p), buf)) |
413 | return -EFAULT; |
414 | buf++; p++; count--; read++; |
415 | } |
416 | pnt = (char *)prof_buffer + p - sizeof(atomic_t); |
417 | if (copy_to_user(to: buf, from: (void *)pnt, n: count)) |
418 | return -EFAULT; |
419 | read += count; |
420 | *ppos += read; |
421 | return read; |
422 | } |
423 | |
424 | /* default is to not implement this call */ |
425 | int __weak setup_profiling_timer(unsigned mult) |
426 | { |
427 | return -EINVAL; |
428 | } |
429 | |
430 | /* |
431 | * Writing to /proc/profile resets the counters |
432 | * |
433 | * Writing a 'profiling multiplier' value into it also re-sets the profiling |
434 | * interrupt frequency, on architectures that support this. |
435 | */ |
436 | static ssize_t write_profile(struct file *file, const char __user *buf, |
437 | size_t count, loff_t *ppos) |
438 | { |
439 | #ifdef CONFIG_SMP |
440 | if (count == sizeof(int)) { |
441 | unsigned int multiplier; |
442 | |
443 | if (copy_from_user(to: &multiplier, from: buf, n: sizeof(int))) |
444 | return -EFAULT; |
445 | |
446 | if (setup_profiling_timer(multiplier)) |
447 | return -EINVAL; |
448 | } |
449 | #endif |
450 | profile_discard_flip_buffers(); |
451 | memset(prof_buffer, 0, prof_len * sizeof(atomic_t)); |
452 | return count; |
453 | } |
454 | |
455 | static const struct proc_ops profile_proc_ops = { |
456 | .proc_read = read_profile, |
457 | .proc_write = write_profile, |
458 | .proc_lseek = default_llseek, |
459 | }; |
460 | |
461 | int __ref create_proc_profile(void) |
462 | { |
463 | struct proc_dir_entry *entry; |
464 | #ifdef CONFIG_SMP |
465 | enum cpuhp_state online_state; |
466 | #endif |
467 | |
468 | int err = 0; |
469 | |
470 | if (!prof_on) |
471 | return 0; |
472 | #ifdef CONFIG_SMP |
473 | err = cpuhp_setup_state(state: CPUHP_PROFILE_PREPARE, name: "PROFILE_PREPARE" , |
474 | startup: profile_prepare_cpu, teardown: profile_dead_cpu); |
475 | if (err) |
476 | return err; |
477 | |
478 | err = cpuhp_setup_state(state: CPUHP_AP_ONLINE_DYN, name: "AP_PROFILE_ONLINE" , |
479 | startup: profile_online_cpu, NULL); |
480 | if (err < 0) |
481 | goto err_state_prep; |
482 | online_state = err; |
483 | err = 0; |
484 | #endif |
485 | entry = proc_create(name: "profile" , S_IWUSR | S_IRUGO, |
486 | NULL, proc_ops: &profile_proc_ops); |
487 | if (!entry) |
488 | goto err_state_onl; |
489 | proc_set_size(entry, (1 + prof_len) * sizeof(atomic_t)); |
490 | |
491 | return err; |
492 | err_state_onl: |
493 | #ifdef CONFIG_SMP |
494 | cpuhp_remove_state(state: online_state); |
495 | err_state_prep: |
496 | cpuhp_remove_state(state: CPUHP_PROFILE_PREPARE); |
497 | #endif |
498 | return err; |
499 | } |
500 | subsys_initcall(create_proc_profile); |
501 | #endif /* CONFIG_PROC_FS */ |
502 | |