2 * latencytop.c: Latency display infrastructure
4 * (C) Copyright 2008 Intel Corporation
5 * Author: Arjan van de Ven <arjan@linux.intel.com>
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
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.
20 * This code tracks 2 levels of statistics:
21 * 1) System level latency
22 * 2) Per process latency
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.
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.
35 * The information is exported via /proc/latency_stats and /proc/<pid>/latency.
36 * These files look like this:
38 * Latency Top version : v0.1
39 * 70 59433 4897 i915_irq_wait drm_ioctl vfs_ioctl do_vfs_ioctl sys_ioctl
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
46 * (note: the average latency is the accumulated latency divided by the number
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>
55 #include <linux/latencytop.h>
56 #include <linux/export.h>
57 #include <linux/sched.h>
58 #include <linux/sched/debug.h>
59 #include <linux/sched/stat.h>
60 #include <linux/list.h>
61 #include <linux/stacktrace.h>
63 static DEFINE_RAW_SPINLOCK(latency_lock
);
66 static struct latency_record latency_record
[MAXLR
];
68 int latencytop_enabled
;
70 void clear_tsk_latency_tracing(struct task_struct
*p
)
74 raw_spin_lock_irqsave(&latency_lock
, flags
);
75 memset(&p
->latency_record
, 0, sizeof(p
->latency_record
));
76 p
->latency_record_count
= 0;
77 raw_spin_unlock_irqrestore(&latency_lock
, flags
);
80 static void clear_global_latency_tracing(void)
84 raw_spin_lock_irqsave(&latency_lock
, flags
);
85 memset(&latency_record
, 0, sizeof(latency_record
));
86 raw_spin_unlock_irqrestore(&latency_lock
, flags
);
90 account_global_scheduler_latency(struct task_struct
*tsk
,
91 struct latency_record
*lat
)
93 int firstnonnull
= MAXLR
+ 1;
96 /* skip kernel threads for now */
100 for (i
= 0; i
< MAXLR
; i
++) {
103 /* Nothing stored: */
104 if (!latency_record
[i
].backtrace
[0]) {
105 if (firstnonnull
> i
)
109 for (q
= 0; q
< LT_BACKTRACEDEPTH
; q
++) {
110 unsigned long record
= lat
->backtrace
[q
];
112 if (latency_record
[i
].backtrace
[q
] != record
) {
117 /* 0 entry marks end of backtrace: */
122 latency_record
[i
].count
++;
123 latency_record
[i
].time
+= lat
->time
;
124 if (lat
->time
> latency_record
[i
].max
)
125 latency_record
[i
].max
= lat
->time
;
134 /* Allocted a new one: */
135 memcpy(&latency_record
[i
], lat
, sizeof(struct latency_record
));
139 * __account_scheduler_latency - record an occurred latency
140 * @tsk - the task struct of the task hitting the latency
141 * @usecs - the duration of the latency in microseconds
142 * @inter - 1 if the sleep was interruptible, 0 if uninterruptible
144 * This function is the main entry point for recording latency entries
145 * as called by the scheduler.
147 * This function has a few special cases to deal with normal 'non-latency'
148 * sleeps: specifically, interruptible sleep longer than 5 msec is skipped
149 * since this usually is caused by waiting for events via select() and co.
151 * Negative latencies (caused by time going backwards) are also explicitly
155 __account_scheduler_latency(struct task_struct
*tsk
, int usecs
, int inter
)
159 struct latency_record lat
;
161 /* Long interruptible waits are generally user requested... */
162 if (inter
&& usecs
> 5000)
165 /* Negative sleeps are time going backwards */
166 /* Zero-time sleeps are non-interesting */
170 memset(&lat
, 0, sizeof(lat
));
175 stack_trace_save_tsk(tsk
, lat
.backtrace
, LT_BACKTRACEDEPTH
, 0);
177 raw_spin_lock_irqsave(&latency_lock
, flags
);
179 account_global_scheduler_latency(tsk
, &lat
);
181 for (i
= 0; i
< tsk
->latency_record_count
; i
++) {
182 struct latency_record
*mylat
;
185 mylat
= &tsk
->latency_record
[i
];
186 for (q
= 0; q
< LT_BACKTRACEDEPTH
; q
++) {
187 unsigned long record
= lat
.backtrace
[q
];
189 if (mylat
->backtrace
[q
] != record
) {
194 /* 0 entry is end of backtrace */
200 mylat
->time
+= lat
.time
;
201 if (lat
.time
> mylat
->max
)
202 mylat
->max
= lat
.time
;
208 * short term hack; if we're > 32 we stop; future we recycle:
210 if (tsk
->latency_record_count
>= LT_SAVECOUNT
)
213 /* Allocated a new one: */
214 i
= tsk
->latency_record_count
++;
215 memcpy(&tsk
->latency_record
[i
], &lat
, sizeof(struct latency_record
));
218 raw_spin_unlock_irqrestore(&latency_lock
, flags
);
221 static int lstats_show(struct seq_file
*m
, void *v
)
225 seq_puts(m
, "Latency Top version : v0.1\n");
227 for (i
= 0; i
< MAXLR
; i
++) {
228 struct latency_record
*lr
= &latency_record
[i
];
230 if (lr
->backtrace
[0]) {
232 seq_printf(m
, "%i %lu %lu",
233 lr
->count
, lr
->time
, lr
->max
);
234 for (q
= 0; q
< LT_BACKTRACEDEPTH
; q
++) {
235 unsigned long bt
= lr
->backtrace
[q
];
240 seq_printf(m
, " %ps", (void *)bt
);
249 lstats_write(struct file
*file
, const char __user
*buf
, size_t count
,
252 clear_global_latency_tracing();
257 static int lstats_open(struct inode
*inode
, struct file
*filp
)
259 return single_open(filp
, lstats_show
, NULL
);
262 static const struct file_operations lstats_fops
= {
265 .write
= lstats_write
,
267 .release
= single_release
,
270 static int __init
init_lstats_procfs(void)
272 proc_create("latency_stats", 0644, NULL
, &lstats_fops
);
276 int sysctl_latencytop(struct ctl_table
*table
, int write
,
277 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
281 err
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
282 if (latencytop_enabled
)
283 force_schedstat_enabled();
287 device_initcall(init_lstats_procfs
);