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/list.h>
59 #include <linux/stacktrace.h>
61 static DEFINE_RAW_SPINLOCK(latency_lock
);
64 static struct latency_record latency_record
[MAXLR
];
66 int latencytop_enabled
;
68 void clear_all_latency_tracing(struct task_struct
*p
)
72 if (!latencytop_enabled
)
75 raw_spin_lock_irqsave(&latency_lock
, flags
);
76 memset(&p
->latency_record
, 0, sizeof(p
->latency_record
));
77 p
->latency_record_count
= 0;
78 raw_spin_unlock_irqrestore(&latency_lock
, flags
);
81 static void clear_global_latency_tracing(void)
85 raw_spin_lock_irqsave(&latency_lock
, flags
);
86 memset(&latency_record
, 0, sizeof(latency_record
));
87 raw_spin_unlock_irqrestore(&latency_lock
, flags
);
91 account_global_scheduler_latency(struct task_struct
*tsk
,
92 struct latency_record
*lat
)
94 int firstnonnull
= MAXLR
+ 1;
97 if (!latencytop_enabled
)
100 /* skip kernel threads for now */
104 for (i
= 0; i
< MAXLR
; i
++) {
107 /* Nothing stored: */
108 if (!latency_record
[i
].backtrace
[0]) {
109 if (firstnonnull
> i
)
113 for (q
= 0; q
< LT_BACKTRACEDEPTH
; q
++) {
114 unsigned long record
= lat
->backtrace
[q
];
116 if (latency_record
[i
].backtrace
[q
] != record
) {
121 /* 0 and ULONG_MAX entries mean end of backtrace: */
122 if (record
== 0 || record
== ULONG_MAX
)
126 latency_record
[i
].count
++;
127 latency_record
[i
].time
+= lat
->time
;
128 if (lat
->time
> latency_record
[i
].max
)
129 latency_record
[i
].max
= lat
->time
;
138 /* Allocted a new one: */
139 memcpy(&latency_record
[i
], lat
, sizeof(struct latency_record
));
143 * Iterator to store a backtrace into a latency record entry
145 static inline void store_stacktrace(struct task_struct
*tsk
,
146 struct latency_record
*lat
)
148 struct stack_trace trace
;
150 memset(&trace
, 0, sizeof(trace
));
151 trace
.max_entries
= LT_BACKTRACEDEPTH
;
152 trace
.entries
= &lat
->backtrace
[0];
153 save_stack_trace_tsk(tsk
, &trace
);
157 * __account_scheduler_latency - record an occurred latency
158 * @tsk - the task struct of the task hitting the latency
159 * @usecs - the duration of the latency in microseconds
160 * @inter - 1 if the sleep was interruptible, 0 if uninterruptible
162 * This function is the main entry point for recording latency entries
163 * as called by the scheduler.
165 * This function has a few special cases to deal with normal 'non-latency'
166 * sleeps: specifically, interruptible sleep longer than 5 msec is skipped
167 * since this usually is caused by waiting for events via select() and co.
169 * Negative latencies (caused by time going backwards) are also explicitly
173 __account_scheduler_latency(struct task_struct
*tsk
, int usecs
, int inter
)
177 struct latency_record lat
;
179 /* Long interruptible waits are generally user requested... */
180 if (inter
&& usecs
> 5000)
183 /* Negative sleeps are time going backwards */
184 /* Zero-time sleeps are non-interesting */
188 memset(&lat
, 0, sizeof(lat
));
192 store_stacktrace(tsk
, &lat
);
194 raw_spin_lock_irqsave(&latency_lock
, flags
);
196 account_global_scheduler_latency(tsk
, &lat
);
198 for (i
= 0; i
< tsk
->latency_record_count
; i
++) {
199 struct latency_record
*mylat
;
202 mylat
= &tsk
->latency_record
[i
];
203 for (q
= 0; q
< LT_BACKTRACEDEPTH
; q
++) {
204 unsigned long record
= lat
.backtrace
[q
];
206 if (mylat
->backtrace
[q
] != record
) {
211 /* 0 and ULONG_MAX entries mean end of backtrace: */
212 if (record
== 0 || record
== ULONG_MAX
)
217 mylat
->time
+= lat
.time
;
218 if (lat
.time
> mylat
->max
)
219 mylat
->max
= lat
.time
;
225 * short term hack; if we're > 32 we stop; future we recycle:
227 if (tsk
->latency_record_count
>= LT_SAVECOUNT
)
230 /* Allocated a new one: */
231 i
= tsk
->latency_record_count
++;
232 memcpy(&tsk
->latency_record
[i
], &lat
, sizeof(struct latency_record
));
235 raw_spin_unlock_irqrestore(&latency_lock
, flags
);
238 static int lstats_show(struct seq_file
*m
, void *v
)
242 seq_puts(m
, "Latency Top version : v0.1\n");
244 for (i
= 0; i
< MAXLR
; i
++) {
245 struct latency_record
*lr
= &latency_record
[i
];
247 if (lr
->backtrace
[0]) {
249 seq_printf(m
, "%i %lu %lu",
250 lr
->count
, lr
->time
, lr
->max
);
251 for (q
= 0; q
< LT_BACKTRACEDEPTH
; q
++) {
252 unsigned long bt
= lr
->backtrace
[q
];
257 seq_printf(m
, " %ps", (void *)bt
);
266 lstats_write(struct file
*file
, const char __user
*buf
, size_t count
,
269 clear_global_latency_tracing();
274 static int lstats_open(struct inode
*inode
, struct file
*filp
)
276 return single_open(filp
, lstats_show
, NULL
);
279 static const struct file_operations lstats_fops
= {
282 .write
= lstats_write
,
284 .release
= single_release
,
287 static int __init
init_lstats_procfs(void)
289 proc_create("latency_stats", 0644, NULL
, &lstats_fops
);
293 int sysctl_latencytop(struct ctl_table
*table
, int write
,
294 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
298 err
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
299 if (latencytop_enabled
)
300 force_schedstat_enabled();
304 device_initcall(init_lstats_procfs
);