1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
|
hrtimers - subsystem for high-resolution kernel timers
----------------------------------------------------
This patch introduces a new subsystem for high-resolution kernel timers.
One might ask the question: we already have a timer subsystem
(kernel/timers.c), why do we need two timer subsystems? After a lot of
back and forth trying to integrate high-resolution and high-precision
features into the existing timer framework, and after testing various
such high-resolution timer implementations in practice, we came to the
conclusion that the timer wheel code is fundamentally not suitable for
such an approach. We initially didn't believe this ('there must be a way
to solve this'), and spent a considerable effort trying to integrate
things into the timer wheel, but we failed. In hindsight, there are
several reasons why such integration is hard/impossible:
- the forced handling of low-resolution and high-resolution timers in
the same way leads to a lot of compromises, macro magic and #ifdef
mess. The timers.c code is very "tightly coded" around jiffies and
32-bitness assumptions, and has been honed and micro-optimized for a
relatively narrow use case (jiffies in a relatively narrow HZ range)
for many years - and thus even small extensions to it easily break
the wheel concept, leading to even worse compromises. The timer wheel
code is very good and tight code, there's zero problems with it in its
current usage - but it is simply not suitable to be extended for
high-res timers.
- the unpredictable [O(N)] overhead of cascading leads to delays which
necessitate a more complex handling of high resolution timers, which
in turn decreases robustness. Such a design still leads to rather large
timing inaccuracies. Cascading is a fundamental property of the timer
wheel concept, it cannot be 'designed out' without inevitably
degrading other portions of the timers.c code in an unacceptable way.
- the implementation of the current posix-timer subsystem on top of
the timer wheel has already introduced a quite complex handling of
the required readjusting of absolute CLOCK_REALTIME timers at
settimeofday or NTP time - further underlying our experience by
example: that the timer wheel data structure is too rigid for high-res
timers.
- the timer wheel code is most optimal for use cases which can be
identified as "timeouts". Such timeouts are usually set up to cover
error conditions in various I/O paths, such as networking and block
I/O. The vast majority of those timers never expire and are rarely
recascaded because the expected correct event arrives in time so they
can be removed from the timer wheel before any further processing of
them becomes necessary. Thus the users of these timeouts can accept
the granularity and precision tradeoffs of the timer wheel, and
largely expect the timer subsystem to have near-zero overhead.
Accurate timing for them is not a core purpose - in fact most of the
timeout values used are ad-hoc. For them it is at most a necessary
evil to guarantee the processing of actual timeout completions
(because most of the timeouts are deleted before completion), which
should thus be as cheap and unintrusive as possible.
The primary users of precision timers are user-space applications that
utilize nanosleep, posix-timers and itimer interfaces. Also, in-kernel
users like drivers and subsystems which require precise timed events
(e.g. multimedia) can benefit from the availability of a separate
high-resolution timer subsystem as well.
While this subsystem does not offer high-resolution clock sources just
yet, the hrtimer subsystem can be easily extended with high-resolution
clock capabilities, and patches for that exist and are maturing quickly.
The increasing demand for realtime and multimedia applications along
with other potential users for precise timers gives another reason to
separate the "timeout" and "precise timer" subsystems.
Another potential benefit is that such a separation allows even more
special-purpose optimization of the existing timer wheel for the low
resolution and low precision use cases - once the precision-sensitive
APIs are separated from the timer wheel and are migrated over to
hrtimers. E.g. we could decrease the frequency of the timeout subsystem
from 250 Hz to 100 HZ (or even smaller).
hrtimer subsystem implementation details
----------------------------------------
the basic design considerations were:
- simplicity
- data structure not bound to jiffies or any other granularity. All the
kernel logic works at 64-bit nanoseconds resolution - no compromises.
- simplification of existing, timing related kernel code
another basic requirement was the immediate enqueueing and ordering of
timers at activation time. After looking at several possible solutions
such as radix trees and hashes, we chose the red black tree as the basic
data structure. Rbtrees are available as a library in the kernel and are
used in various performance-critical areas of e.g. memory management and
file systems. The rbtree is solely used for time sorted ordering, while
a separate list is used to give the expiry code fast access to the
queued timers, without having to walk the rbtree.
(This separate list is also useful for later when we'll introduce
high-resolution clocks, where we need separate pending and expired
queues while keeping the time-order intact.)
Time-ordered enqueueing is not purely for the purposes of
high-resolution clocks though, it also simplifies the handling of
absolute timers based on a low-resolution CLOCK_REALTIME. The existing
implementation needed to keep an extra list of all armed absolute
CLOCK_REALTIME timers along with complex locking. In case of
settimeofday and NTP, all the timers (!) had to be dequeued, the
time-changing code had to fix them up one by one, and all of them had to
be enqueued again. The time-ordered enqueueing and the storage of the
expiry time in absolute time units removes all this complex and poorly
scaling code from the posix-timer implementation - the clock can simply
be set without having to touch the rbtree. This also makes the handling
of posix-timers simpler in general.
The locking and per-CPU behavior of hrtimers was mostly taken from the
existing timer wheel code, as it is mature and well suited. Sharing code
was not really a win, due to the different data structures. Also, the
hrtimer functions now have clearer behavior and clearer names - such as
hrtimer_try_to_cancel() and hrtimer_cancel() [which are roughly
equivalent to del_timer() and del_timer_sync()] - so there's no direct
1:1 mapping between them on the algorithmic level, and thus no real
potential for code sharing either.
Basic data types: every time value, absolute or relative, is in a
special nanosecond-resolution type: ktime_t. The kernel-internal
representation of ktime_t values and operations is implemented via
macros and inline functions, and can be switched between a "hybrid
union" type and a plain "scalar" 64bit nanoseconds representation (at
compile time). The hybrid union type optimizes time conversions on 32bit
CPUs. This build-time-selectable ktime_t storage format was implemented
to avoid the performance impact of 64-bit multiplications and divisions
on 32bit CPUs. Such operations are frequently necessary to convert
between the storage formats provided by kernel and userspace interfaces
and the internal time format. (See include/linux/ktime.h for further
details.)
hrtimers - rounding of timer values
-----------------------------------
the hrtimer code will round timer events to lower-resolution clocks
because it has to. Otherwise it will do no artificial rounding at all.
one question is, what resolution value should be returned to the user by
the clock_getres() interface. This will return whatever real resolution
a given clock has - be it low-res, high-res, or artificially-low-res.
hrtimers - testing and verification
----------------------------------
We used the high-resolution clock subsystem ontop of hrtimers to verify
the hrtimer implementation details in praxis, and we also ran the posix
timer tests in order to ensure specification compliance. We also ran
tests on low-resolution clocks.
The hrtimer patch converts the following kernel functionality to use
hrtimers:
- nanosleep
- itimers
- posix-timers
The conversion of nanosleep and posix-timers enabled the unification of
nanosleep and clock_nanosleep.
The code was successfully compiled for the following platforms:
i386, x86_64, ARM, PPC, PPC64, IA64
The code was run-tested on the following platforms:
i386(UP/SMP), x86_64(UP/SMP), ARM, PPC
hrtimers were also integrated into the -rt tree, along with a
hrtimers-based high-resolution clock implementation, so the hrtimers
code got a healthy amount of testing and use in practice.
Thomas Gleixner, Ingo Molnar
|
