1delays - Information on the various kernel delay / sleep mechanisms
2-------------------------------------------------------------------
3
4This document seeks to answer the common question: "What is the
5RightWay (TM) to insert a delay?"
6
7This question is most often faced by driver writers who have to
8deal with hardware delays and who may not be the most intimately
9familiar with the inner workings of the Linux Kernel.
10
11
12Inserting Delays
13----------------
14
15The first, and most important, question you need to ask is "Is my
16code in an atomic context?"  This should be followed closely by "Does
17it really need to delay in atomic context?" If so...
18
19ATOMIC CONTEXT:
20	You must use the *delay family of functions. These
21	functions use the jiffie estimation of clock speed
22	and will busy wait for enough loop cycles to achieve
23	the desired delay:
24
25	ndelay(unsigned long nsecs)
26	udelay(unsigned long usecs)
27	mdelay(unsigned long msecs)
28
29	udelay is the generally preferred API; ndelay-level
30	precision may not actually exist on many non-PC devices.
31
32	mdelay is macro wrapper around udelay, to account for
33	possible overflow when passing large arguments to udelay.
34	In general, use of mdelay is discouraged and code should
35	be refactored to allow for the use of msleep.
36
37NON-ATOMIC CONTEXT:
38	You should use the *sleep[_range] family of functions.
39	There are a few more options here, while any of them may
40	work correctly, using the "right" sleep function will
41	help the scheduler, power management, and just make your
42	driver better :)
43
44	-- Backed by busy-wait loop:
45		udelay(unsigned long usecs)
46	-- Backed by hrtimers:
47		usleep_range(unsigned long min, unsigned long max)
48	-- Backed by jiffies / legacy_timers
49		msleep(unsigned long msecs)
50		msleep_interruptible(unsigned long msecs)
51
52	Unlike the *delay family, the underlying mechanism
53	driving each of these calls varies, thus there are
54	quirks you should be aware of.
55
56
57	SLEEPING FOR "A FEW" USECS ( < ~10us? ):
58		* Use udelay
59
60		- Why not usleep?
61			On slower systems, (embedded, OR perhaps a speed-
62			stepped PC!) the overhead of setting up the hrtimers
63			for usleep *may* not be worth it. Such an evaluation
64			will obviously depend on your specific situation, but
65			it is something to be aware of.
66
67	SLEEPING FOR ~USECS OR SMALL MSECS ( 10us - 20ms):
68		* Use usleep_range
69
70		- Why not msleep for (1ms - 20ms)?
71			Explained originally here:
72				http://lkml.org/lkml/2007/8/3/250
73			msleep(1~20) may not do what the caller intends, and
74			will often sleep longer (~20 ms actual sleep for any
75			value given in the 1~20ms range). In many cases this
76			is not the desired behavior.
77
78		- Why is there no "usleep" / What is a good range?
79			Since usleep_range is built on top of hrtimers, the
80			wakeup will be very precise (ish), thus a simple
81			usleep function would likely introduce a large number
82			of undesired interrupts.
83
84			With the introduction of a range, the scheduler is
85			free to coalesce your wakeup with any other wakeup
86			that may have happened for other reasons, or at the
87			worst case, fire an interrupt for your upper bound.
88
89			The larger a range you supply, the greater a chance
90			that you will not trigger an interrupt; this should
91			be balanced with what is an acceptable upper bound on
92			delay / performance for your specific code path. Exact
93			tolerances here are very situation specific, thus it
94			is left to the caller to determine a reasonable range.
95
96	SLEEPING FOR LARGER MSECS ( 10ms+ )
97		* Use msleep or possibly msleep_interruptible
98
99		- What's the difference?
100			msleep sets the current task to TASK_UNINTERRUPTIBLE
101			whereas msleep_interruptible sets the current task to
102			TASK_INTERRUPTIBLE before scheduling the sleep. In
103			short, the difference is whether the sleep can be ended
104			early by a signal. In general, just use msleep unless
105			you know you have a need for the interruptible variant.
106