xref: /linux/kernel/profile.c (revision 447e140e66fd226350b3ce86cffc965eaae4c856)
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, &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, &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(&prof_cpu_mask, GFP_KERNEL))
118 		return -ENOMEM;
119 
120 	cpumask_copy(prof_cpu_mask, cpu_possible_mask);
121 
122 	prof_buffer = kzalloc(buffer_bytes, GFP_KERNEL|__GFP_NOWARN);
123 	if (prof_buffer)
124 		return 0;
125 
126 	prof_buffer = alloc_pages_exact(buffer_bytes,
127 					GFP_KERNEL|__GFP_ZERO|__GFP_NOWARN);
128 	if (prof_buffer)
129 		return 0;
130 
131 	prof_buffer = vzalloc(buffer_bytes);
132 	if (prof_buffer)
133 		return 0;
134 
135 	free_cpumask_var(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(__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) {
189 			if (!hits[i].hits) {
190 				if (hits[i].pc)
191 					hits[i].pc = 0;
192 				continue;
193 			}
194 			atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
195 			hits[i].hits = hits[i].pc = 0;
196 		}
197 	}
198 	mutex_unlock(&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(__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));
212 	}
213 	mutex_unlock(&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(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;
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(prof_cpu_mask))
271 		cpumask_clear_cpu(cpu, 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(node, GFP_KERNEL | __GFP_ZERO, 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(prof_cpu_mask))
308 		cpumask_set_cpu(cpu, 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(prof_cpu_mask) &&
338 	    cpumask_test_cpu(smp_processor_id(), prof_cpu_mask))
339 		profile_hit(type, (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 /*
348  * This function accesses profiling information. The returned data is
349  * binary: the sampling step and the actual contents of the profile
350  * buffer. Use of the program readprofile is recommended in order to
351  * get meaningful info out of these data.
352  */
353 static ssize_t
354 read_profile(struct file *file, char __user *buf, size_t count, loff_t *ppos)
355 {
356 	unsigned long p = *ppos;
357 	ssize_t read;
358 	char *pnt;
359 	unsigned long sample_step = 1UL << prof_shift;
360 
361 	profile_flip_buffers();
362 	if (p >= (prof_len+1)*sizeof(unsigned int))
363 		return 0;
364 	if (count > (prof_len+1)*sizeof(unsigned int) - p)
365 		count = (prof_len+1)*sizeof(unsigned int) - p;
366 	read = 0;
367 
368 	while (p < sizeof(unsigned int) && count > 0) {
369 		if (put_user(*((char *)(&sample_step)+p), buf))
370 			return -EFAULT;
371 		buf++; p++; count--; read++;
372 	}
373 	pnt = (char *)prof_buffer + p - sizeof(atomic_t);
374 	if (copy_to_user(buf, (void *)pnt, count))
375 		return -EFAULT;
376 	read += count;
377 	*ppos += read;
378 	return read;
379 }
380 
381 /* default is to not implement this call */
382 int __weak setup_profiling_timer(unsigned mult)
383 {
384 	return -EINVAL;
385 }
386 
387 /*
388  * Writing to /proc/profile resets the counters
389  *
390  * Writing a 'profiling multiplier' value into it also re-sets the profiling
391  * interrupt frequency, on architectures that support this.
392  */
393 static ssize_t write_profile(struct file *file, const char __user *buf,
394 			     size_t count, loff_t *ppos)
395 {
396 #ifdef CONFIG_SMP
397 	if (count == sizeof(int)) {
398 		unsigned int multiplier;
399 
400 		if (copy_from_user(&multiplier, buf, sizeof(int)))
401 			return -EFAULT;
402 
403 		if (setup_profiling_timer(multiplier))
404 			return -EINVAL;
405 	}
406 #endif
407 	profile_discard_flip_buffers();
408 	memset(prof_buffer, 0, prof_len * sizeof(atomic_t));
409 	return count;
410 }
411 
412 static const struct proc_ops profile_proc_ops = {
413 	.proc_read	= read_profile,
414 	.proc_write	= write_profile,
415 	.proc_lseek	= default_llseek,
416 };
417 
418 int __ref create_proc_profile(void)
419 {
420 	struct proc_dir_entry *entry;
421 #ifdef CONFIG_SMP
422 	enum cpuhp_state online_state;
423 #endif
424 
425 	int err = 0;
426 
427 	if (!prof_on)
428 		return 0;
429 #ifdef CONFIG_SMP
430 	err = cpuhp_setup_state(CPUHP_PROFILE_PREPARE, "PROFILE_PREPARE",
431 				profile_prepare_cpu, profile_dead_cpu);
432 	if (err)
433 		return err;
434 
435 	err = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "AP_PROFILE_ONLINE",
436 				profile_online_cpu, NULL);
437 	if (err < 0)
438 		goto err_state_prep;
439 	online_state = err;
440 	err = 0;
441 #endif
442 	entry = proc_create("profile", S_IWUSR | S_IRUGO,
443 			    NULL, &profile_proc_ops);
444 	if (!entry)
445 		goto err_state_onl;
446 	proc_set_size(entry, (1 + prof_len) * sizeof(atomic_t));
447 
448 	return err;
449 err_state_onl:
450 #ifdef CONFIG_SMP
451 	cpuhp_remove_state(online_state);
452 err_state_prep:
453 	cpuhp_remove_state(CPUHP_PROFILE_PREPARE);
454 #endif
455 	return err;
456 }
457 subsys_initcall(create_proc_profile);
458 #endif /* CONFIG_PROC_FS */
459