[linux.git] / kernel / latencytop.c

/*
 * latencytop.c: Latency display infrastructure
 *
 * (C) Copyright 2008 Intel Corporation
 * Author: Arjan van de Ven <[email protected]>
 *
 * This program is free software; you can redistribute it and/or
 * modify it under the terms of the GNU General Public License
 * as published by the Free Software Foundation; version 2
 * of the License.
 */

/*
 * CONFIG_LATENCYTOP enables a kernel latency tracking infrastructure that is
 * used by the "latencytop" userspace tool. The latency that is tracked is not
 * the 'traditional' interrupt latency (which is primarily caused by something
 * else consuming CPU), but instead, it is the latency an application encounters
 * because the kernel sleeps on its behalf for various reasons.
 *
 * This code tracks 2 levels of statistics:
 * 1) System level latency
 * 2) Per process latency
 *
 * The latency is stored in fixed sized data structures in an accumulated form;
 * if the "same" latency cause is hit twice, this will be tracked as one entry
 * in the data structure. Both the count, total accumulated latency and maximum
 * latency are tracked in this data structure. When the fixed size structure is
 * full, no new causes are tracked until the buffer is flushed by writing to
 * the /proc file; the userspace tool does this on a regular basis.
 *
 * A latency cause is identified by a stringified backtrace at the point that
 * the scheduler gets invoked. The userland tool will use this string to
 * identify the cause of the latency in human readable form.
 *
 * The information is exported via /proc/latency_stats and /proc/<pid>/latency.
 * These files look like this:
 *
 * Latency Top version : v0.1
 * 70 59433 4897 i915_irq_wait drm_ioctl vfs_ioctl do_vfs_ioctl sys_ioctl
 * |    |    |    |
 * |    |    |    +----> the stringified backtrace
 * |    |    +---------> The maximum latency for this entry in microseconds
 * |    +--------------> The accumulated latency for this entry (microseconds)
 * +-------------------> The number of times this entry is hit
 *
 * (note: the average latency is the accumulated latency divided by the number
 * of times)
 */

#include <linux/kallsyms.h>
#include <linux/seq_file.h>
#include <linux/notifier.h>
#include <linux/spinlock.h>
#include <linux/proc_fs.h>
#include <linux/latencytop.h>
#include <linux/export.h>
#include <linux/sched.h>
#include <linux/sched/debug.h>
#include <linux/list.h>
#include <linux/stacktrace.h>

static DEFINE_RAW_SPINLOCK(latency_lock);

#define MAXLR 128
static struct latency_record latency_record[MAXLR];

int latencytop_enabled;

void clear_all_latency_tracing(struct task_struct *p)
{
	unsigned long flags;

	if (!latencytop_enabled)
		return;

	raw_spin_lock_irqsave(&latency_lock, flags);
	memset(&p->latency_record, 0, sizeof(p->latency_record));
	p->latency_record_count = 0;
	raw_spin_unlock_irqrestore(&latency_lock, flags);
}

static void clear_global_latency_tracing(void)
{
	unsigned long flags;

	raw_spin_lock_irqsave(&latency_lock, flags);
	memset(&latency_record, 0, sizeof(latency_record));
	raw_spin_unlock_irqrestore(&latency_lock, flags);
}

static void __sched
account_global_scheduler_latency(struct task_struct *tsk,
				 struct latency_record *lat)
{
	int firstnonnull = MAXLR + 1;
	int i;

	if (!latencytop_enabled)
		return;

	/* skip kernel threads for now */
	if (!tsk->mm)
		return;

	for (i = 0; i < MAXLR; i++) {
		int q, same = 1;

		/* Nothing stored: */
		if (!latency_record[i].backtrace[0]) {
			if (firstnonnull > i)
				firstnonnull = i;
			continue;
		}
		for (q = 0; q < LT_BACKTRACEDEPTH; q++) {
			unsigned long record = lat->backtrace[q];

			if (latency_record[i].backtrace[q] != record) {
				same = 0;
				break;
			}

			/* 0 and ULONG_MAX entries mean end of backtrace: */
			if (record == 0 || record == ULONG_MAX)
				break;
		}
		if (same) {
			latency_record[i].count++;
			latency_record[i].time += lat->time;
			if (lat->time > latency_record[i].max)
				latency_record[i].max = lat->time;
			return;
		}
	}

	i = firstnonnull;
	if (i >= MAXLR - 1)
		return;

	/* Allocted a new one: */
	memcpy(&latency_record[i], lat, sizeof(struct latency_record));
}

/*
 * Iterator to store a backtrace into a latency record entry
 */
static inline void store_stacktrace(struct task_struct *tsk,
					struct latency_record *lat)
{
	struct stack_trace trace;

	memset(&trace, 0, sizeof(trace));
	trace.max_entries = LT_BACKTRACEDEPTH;
	trace.entries = &lat->backtrace[0];
	save_stack_trace_tsk(tsk, &trace);
}

/**
 * __account_scheduler_latency - record an occurred latency
 * @tsk - the task struct of the task hitting the latency
 * @usecs - the duration of the latency in microseconds
 * @inter - 1 if the sleep was interruptible, 0 if uninterruptible
 *
 * This function is the main entry point for recording latency entries
 * as called by the scheduler.
 *
 * This function has a few special cases to deal with normal 'non-latency'
 * sleeps: specifically, interruptible sleep longer than 5 msec is skipped
 * since this usually is caused by waiting for events via select() and co.
 *
 * Negative latencies (caused by time going backwards) are also explicitly
 * skipped.
 */
void __sched
__account_scheduler_latency(struct task_struct *tsk, int usecs, int inter)
{
	unsigned long flags;
	int i, q;
	struct latency_record lat;

	/* Long interruptible waits are generally user requested... */
	if (inter && usecs > 5000)
		return;

	/* Negative sleeps are time going backwards */
	/* Zero-time sleeps are non-interesting */
	if (usecs <= 0)
		return;

	memset(&lat, 0, sizeof(lat));
	lat.count = 1;
	lat.time = usecs;
	lat.max = usecs;
	store_stacktrace(tsk, &lat);

	raw_spin_lock_irqsave(&latency_lock, flags);

	account_global_scheduler_latency(tsk, &lat);

	for (i = 0; i < tsk->latency_record_count; i++) {
		struct latency_record *mylat;
		int same = 1;

		mylat = &tsk->latency_record[i];
		for (q = 0; q < LT_BACKTRACEDEPTH; q++) {
			unsigned long record = lat.backtrace[q];

			if (mylat->backtrace[q] != record) {
				same = 0;
				break;
			}

			/* 0 and ULONG_MAX entries mean end of backtrace: */
			if (record == 0 || record == ULONG_MAX)
				break;
		}
		if (same) {
			mylat->count++;
			mylat->time += lat.time;
			if (lat.time > mylat->max)
				mylat->max = lat.time;
			goto out_unlock;
		}
	}

	/*
	 * short term hack; if we're > 32 we stop; future we recycle:
	 */
	if (tsk->latency_record_count >= LT_SAVECOUNT)
		goto out_unlock;

	/* Allocated a new one: */
	i = tsk->latency_record_count++;
	memcpy(&tsk->latency_record[i], &lat, sizeof(struct latency_record));

out_unlock:
	raw_spin_unlock_irqrestore(&latency_lock, flags);
}

static int lstats_show(struct seq_file *m, void *v)
{
	int i;

	seq_puts(m, "Latency Top version : v0.1\n");

	for (i = 0; i < MAXLR; i++) {
		struct latency_record *lr = &latency_record[i];

		if (lr->backtrace[0]) {
			int q;
			seq_printf(m, "%i %lu %lu",
				   lr->count, lr->time, lr->max);
			for (q = 0; q < LT_BACKTRACEDEPTH; q++) {
				unsigned long bt = lr->backtrace[q];
				if (!bt)
					break;
				if (bt == ULONG_MAX)
					break;
				seq_printf(m, " %ps", (void *)bt);
			}
			seq_puts(m, "\n");
		}
	}
	return 0;
}

static ssize_t
lstats_write(struct file *file, const char __user *buf, size_t count,
	     loff_t *offs)
{
	clear_global_latency_tracing();

	return count;
}

static int lstats_open(struct inode *inode, struct file *filp)
{
	return single_open(filp, lstats_show, NULL);
}

static const struct file_operations lstats_fops = {
	.open		= lstats_open,
	.read		= seq_read,
	.write		= lstats_write,
	.llseek		= seq_lseek,
	.release	= single_release,
};

static int __init init_lstats_procfs(void)
{
	proc_create("latency_stats", 0644, NULL, &lstats_fops);
	return 0;
}

int sysctl_latencytop(struct ctl_table *table, int write,
			void __user *buffer, size_t *lenp, loff_t *ppos)
{
	int err;

	err = proc_dointvec(table, write, buffer, lenp, ppos);
	if (latencytop_enabled)
		force_schedstat_enabled();

	return err;
}
device_initcall(init_lstats_procfs);
Commit	Line	Data
9745512c AV	1	/*
	2	* latencytop.c: Latency display infrastructure
	3	*
	4	* (C) Copyright 2008 Intel Corporation
	5	* Author: Arjan van de Ven <[email protected]>
	6	*
	7	* This program is free software; you can redistribute it and/or
	8	* modify it under the terms of the GNU General Public License
	9	* as published by the Free Software Foundation; version 2
	10	* of the License.
	11	*/
ad0b0fd5 AV	12
	13	/*
	14	* CONFIG_LATENCYTOP enables a kernel latency tracking infrastructure that is
	15	* used by the "latencytop" userspace tool. The latency that is tracked is not
	16	* the 'traditional' interrupt latency (which is primarily caused by something
	17	* else consuming CPU), but instead, it is the latency an application encounters
	18	* because the kernel sleeps on its behalf for various reasons.
	19	*
	20	* This code tracks 2 levels of statistics:
	21	* 1) System level latency
	22	* 2) Per process latency
	23	*
	24	* The latency is stored in fixed sized data structures in an accumulated form;
	25	* if the "same" latency cause is hit twice, this will be tracked as one entry
	26	* in the data structure. Both the count, total accumulated latency and maximum
	27	* latency are tracked in this data structure. When the fixed size structure is
	28	* full, no new causes are tracked until the buffer is flushed by writing to
	29	* the /proc file; the userspace tool does this on a regular basis.
	30	*
	31	* A latency cause is identified by a stringified backtrace at the point that
	32	* the scheduler gets invoked. The userland tool will use this string to
	33	* identify the cause of the latency in human readable form.
	34	*
	35	* The information is exported via /proc/latency_stats and /proc/<pid>/latency.
	36	* These files look like this:
	37	*
	38	* Latency Top version : v0.1
	39	* 70 59433 4897 i915_irq_wait drm_ioctl vfs_ioctl do_vfs_ioctl sys_ioctl
	40	* \| \| \| \|
	41	* \| \| \| +----> the stringified backtrace
	42	* \| \| +---------> The maximum latency for this entry in microseconds
	43	* \| +--------------> The accumulated latency for this entry (microseconds)
	44	* +-------------------> The number of times this entry is hit
	45	*
	46	* (note: the average latency is the accumulated latency divided by the number
	47	* of times)
	48	*/
	49
9745512c AV	50	#include <linux/kallsyms.h>
	51	#include <linux/seq_file.h>
	52	#include <linux/notifier.h>
	53	#include <linux/spinlock.h>
	54	#include <linux/proc_fs.h>
cb251765	55	#include <linux/latencytop.h>
9984de1a	56	#include <linux/export.h>
9745512c	57	#include <linux/sched.h>
b17b0153	58	#include <linux/sched/debug.h>
9745512c	59	#include <linux/list.h>
9745512c AV	60	#include <linux/stacktrace.h>
9745512c AV	61
757455d4	62	static DEFINE_RAW_SPINLOCK(latency_lock);
9745512c AV	63
	64	#define MAXLR 128
	65	static struct latency_record latency_record[MAXLR];
	66
	67	int latencytop_enabled;
	68
	69	void clear_all_latency_tracing(struct task_struct *p)
	70	{
	71	unsigned long flags;
	72
	73	if (!latencytop_enabled)
	74	return;
	75
757455d4	76	raw_spin_lock_irqsave(&latency_lock, flags);
9745512c AV	77	memset(&p->latency_record, 0, sizeof(p->latency_record));
9745512c AV	78	p->latency_record_count = 0;
757455d4	79	raw_spin_unlock_irqrestore(&latency_lock, flags);
9745512c AV	80	}
	81
	82	static void clear_global_latency_tracing(void)
	83	{
	84	unsigned long flags;
	85
757455d4	86	raw_spin_lock_irqsave(&latency_lock, flags);
9745512c	87	memset(&latency_record, 0, sizeof(latency_record));
757455d4	88	raw_spin_unlock_irqrestore(&latency_lock, flags);
9745512c AV	89	}
	90
	91	static void __sched
eaa1809b FF	92	account_global_scheduler_latency(struct task_struct *tsk,
eaa1809b FF	93	struct latency_record *lat)
9745512c AV	94	{
	95	int firstnonnull = MAXLR + 1;
	96	int i;
	97
	98	if (!latencytop_enabled)
	99	return;
	100
	101	/* skip kernel threads for now */
	102	if (!tsk->mm)
	103	return;
	104
	105	for (i = 0; i < MAXLR; i++) {
19fb518c DA	106	int q, same = 1;
19fb518c DA	107
9745512c AV	108	/* Nothing stored: */
	109	if (!latency_record[i].backtrace[0]) {
	110	if (firstnonnull > i)
	111	firstnonnull = i;
	112	continue;
	113	}
ad0b0fd5	114	for (q = 0; q < LT_BACKTRACEDEPTH; q++) {
19fb518c DA	115	unsigned long record = lat->backtrace[q];
	116
	117	if (latency_record[i].backtrace[q] != record) {
9745512c	118	same = 0;
9745512c	119	break;
19fb518c DA	120	}
	121
	122	/* 0 and ULONG_MAX entries mean end of backtrace: */
	123	if (record == 0 \|\| record == ULONG_MAX)
9745512c AV	124	break;
	125	}
	126	if (same) {
	127	latency_record[i].count++;
	128	latency_record[i].time += lat->time;
	129	if (lat->time > latency_record[i].max)
	130	latency_record[i].max = lat->time;
	131	return;
	132	}
	133	}
	134
	135	i = firstnonnull;
	136	if (i >= MAXLR - 1)
	137	return;
	138
	139	/* Allocted a new one: */
	140	memcpy(&latency_record[i], lat, sizeof(struct latency_record));
	141	}
	142
ad0b0fd5 AV	143	/*
	144	* Iterator to store a backtrace into a latency record entry
	145	*/
	146	static inline void store_stacktrace(struct task_struct *tsk,
	147	struct latency_record *lat)
9745512c AV	148	{
	149	struct stack_trace trace;
	150
	151	memset(&trace, 0, sizeof(trace));
	152	trace.max_entries = LT_BACKTRACEDEPTH;
	153	trace.entries = &lat->backtrace[0];
9745512c AV	154	save_stack_trace_tsk(tsk, &trace);
	155	}
	156
ad0b0fd5	157	/**
25985edc	158	* __account_scheduler_latency - record an occurred latency
ad0b0fd5 AV	159	* @tsk - the task struct of the task hitting the latency
	160	* @usecs - the duration of the latency in microseconds
	161	* @inter - 1 if the sleep was interruptible, 0 if uninterruptible
	162	*
	163	* This function is the main entry point for recording latency entries
	164	* as called by the scheduler.
	165	*
	166	* This function has a few special cases to deal with normal 'non-latency'
	167	* sleeps: specifically, interruptible sleep longer than 5 msec is skipped
	168	* since this usually is caused by waiting for events via select() and co.
	169	*
	170	* Negative latencies (caused by time going backwards) are also explicitly
	171	* skipped.
	172	*/
9745512c	173	void __sched
ad0b0fd5	174	__account_scheduler_latency(struct task_struct *tsk, int usecs, int inter)
9745512c AV	175	{
	176	unsigned long flags;
	177	int i, q;
	178	struct latency_record lat;
	179
9745512c AV	180	/* Long interruptible waits are generally user requested... */
	181	if (inter && usecs > 5000)
	182	return;
	183
ad0b0fd5 AV	184	/* Negative sleeps are time going backwards */
	185	/* Zero-time sleeps are non-interesting */
	186	if (usecs <= 0)
	187	return;
	188
9745512c AV	189	memset(&lat, 0, sizeof(lat));
	190	lat.count = 1;
	191	lat.time = usecs;
	192	lat.max = usecs;
	193	store_stacktrace(tsk, &lat);
	194
757455d4	195	raw_spin_lock_irqsave(&latency_lock, flags);
9745512c AV	196
	197	account_global_scheduler_latency(tsk, &lat);
	198
38715258	199	for (i = 0; i < tsk->latency_record_count; i++) {
9745512c AV	200	struct latency_record *mylat;
9745512c AV	201	int same = 1;
19fb518c	202
9745512c	203	mylat = &tsk->latency_record[i];
ad0b0fd5	204	for (q = 0; q < LT_BACKTRACEDEPTH; q++) {
19fb518c DA	205	unsigned long record = lat.backtrace[q];
	206
	207	if (mylat->backtrace[q] != record) {
9745512c	208	same = 0;
9745512c	209	break;
19fb518c DA	210	}
	211
	212	/* 0 and ULONG_MAX entries mean end of backtrace: */
	213	if (record == 0 \|\| record == ULONG_MAX)
9745512c AV	214	break;
	215	}
	216	if (same) {
	217	mylat->count++;
	218	mylat->time += lat.time;
	219	if (lat.time > mylat->max)
	220	mylat->max = lat.time;
	221	goto out_unlock;
	222	}
	223	}
	224
38715258 KC	225	/*
	226	* short term hack; if we're > 32 we stop; future we recycle:
	227	*/
	228	if (tsk->latency_record_count >= LT_SAVECOUNT)
	229	goto out_unlock;
	230
9745512c	231	/* Allocated a new one: */
38715258	232	i = tsk->latency_record_count++;
9745512c AV	233	memcpy(&tsk->latency_record[i], &lat, sizeof(struct latency_record));
	234
	235	out_unlock:
757455d4	236	raw_spin_unlock_irqrestore(&latency_lock, flags);
9745512c AV	237	}
	238
	239	static int lstats_show(struct seq_file m, void v)
	240	{
	241	int i;
	242
	243	seq_puts(m, "Latency Top version : v0.1\n");
	244
	245	for (i = 0; i < MAXLR; i++) {
34e49d4f JP	246	struct latency_record *lr = &latency_record[i];
	247
	248	if (lr->backtrace[0]) {
9745512c	249	int q;
34e49d4f JP	250	seq_printf(m, "%i %lu %lu",
34e49d4f JP	251	lr->count, lr->time, lr->max);
9745512c	252	for (q = 0; q < LT_BACKTRACEDEPTH; q++) {
34e49d4f JP	253	unsigned long bt = lr->backtrace[q];
34e49d4f JP	254	if (!bt)
9745512c	255	break;
34e49d4f	256	if (bt == ULONG_MAX)
9745512c	257	break;
34e49d4f	258	seq_printf(m, " %ps", (void *)bt);
9745512c	259	}
eaa1809b	260	seq_puts(m, "\n");
9745512c AV	261	}
	262	}
	263	return 0;
	264	}
	265
	266	static ssize_t
	267	lstats_write(struct file file, const char __user buf, size_t count,
	268	loff_t *offs)
	269	{
	270	clear_global_latency_tracing();
	271
	272	return count;
	273	}
	274
	275	static int lstats_open(struct inode inode, struct file filp)
	276	{
	277	return single_open(filp, lstats_show, NULL);
	278	}
	279
ad0b0fd5	280	static const struct file_operations lstats_fops = {
9745512c AV	281	.open = lstats_open,
	282	.read = seq_read,
	283	.write = lstats_write,
	284	.llseek = seq_lseek,
	285	.release = single_release,
	286	};
	287
	288	static int __init init_lstats_procfs(void)
	289	{
c33fff0a	290	proc_create("latency_stats", 0644, NULL, &lstats_fops);
9745512c AV	291	return 0;
9745512c AV	292	}
cb251765 MG	293
	294	int sysctl_latencytop(struct ctl_table *table, int write,
	295	void __user buffer, size_t lenp, loff_t *ppos)
	296	{
	297	int err;
	298
	299	err = proc_dointvec(table, write, buffer, lenp, ppos);
	300	if (latencytop_enabled)
	301	force_schedstat_enabled();
	302
	303	return err;
	304	}
ad0b0fd5	305	device_initcall(init_lstats_procfs);