1=====================================================
2High resolution timers and dynamic ticks design notes
3=====================================================
4
5Further information can be found in the paper of the OLS 2006 talk "hrtimers
6and beyond". The paper is part of the OLS 2006 Proceedings Volume 1, which can
7be found on the OLS website:
8https://www.kernel.org/doc/ols/2006/ols2006v1-pages-333-346.pdf
9
10The slides to this talk are available from:
11http://www.cs.columbia.edu/~nahum/w6998/papers/ols2006-hrtimers-slides.pdf
12
13The slides contain five figures (pages 2, 15, 18, 20, 22), which illustrate the
14changes in the time(r) related Linux subsystems. Figure #1 (p. 2) shows the
15design of the Linux time(r) system before hrtimers and other building blocks
16got merged into mainline.
17
18Note: the paper and the slides are talking about "clock event source", while we
19switched to the name "clock event devices" in meantime.
20
21The design contains the following basic building blocks:
22
23- hrtimer base infrastructure
24- timeofday and clock source management
25- clock event management
26- high resolution timer functionality
27- dynamic ticks
28
29
30hrtimer base infrastructure
31---------------------------
32
33The hrtimer base infrastructure was merged into the 2.6.16 kernel. Details of
34the base implementation are covered in Documentation/timers/hrtimers.rst. See
35also figure #2 (OLS slides p. 15)
36
37The main differences to the timer wheel, which holds the armed timer_list type
38timers are:
39
40       - time ordered enqueueing into a rb-tree
41       - independent of ticks (the processing is based on nanoseconds)
42
43
44timeofday and clock source management
45-------------------------------------
46
47John Stultz's Generic Time Of Day (GTOD) framework moves a large portion of
48code out of the architecture-specific areas into a generic management
49framework, as illustrated in figure #3 (OLS slides p. 18). The architecture
50specific portion is reduced to the low level hardware details of the clock
51sources, which are registered in the framework and selected on a quality based
52decision. The low level code provides hardware setup and readout routines and
53initializes data structures, which are used by the generic time keeping code to
54convert the clock ticks to nanosecond based time values. All other time keeping
55related functionality is moved into the generic code. The GTOD base patch got
56merged into the 2.6.18 kernel.
57
58Further information about the Generic Time Of Day framework is available in the
59OLS 2005 Proceedings Volume 1:
60
61	http://www.linuxsymposium.org/2005/linuxsymposium_procv1.pdf
62
63The paper "We Are Not Getting Any Younger: A New Approach to Time and
64Timers" was written by J. Stultz, D.V. Hart, & N. Aravamudan.
65
66Figure #3 (OLS slides p.18) illustrates the transformation.
67
68
69clock event management
70----------------------
71
72While clock sources provide read access to the monotonically increasing time
73value, clock event devices are used to schedule the next event
74interrupt(s). The next event is currently defined to be periodic, with its
75period defined at compile time. The setup and selection of the event device
76for various event driven functionalities is hardwired into the architecture
77dependent code. This results in duplicated code across all architectures and
78makes it extremely difficult to change the configuration of the system to use
79event interrupt devices other than those already built into the
80architecture. Another implication of the current design is that it is necessary
81to touch all the architecture-specific implementations in order to provide new
82functionality like high resolution timers or dynamic ticks.
83
84The clock events subsystem tries to address this problem by providing a generic
85solution to manage clock event devices and their usage for the various clock
86event driven kernel functionalities. The goal of the clock event subsystem is
87to minimize the clock event related architecture dependent code to the pure
88hardware related handling and to allow easy addition and utilization of new
89clock event devices. It also minimizes the duplicated code across the
90architectures as it provides generic functionality down to the interrupt
91service handler, which is almost inherently hardware dependent.
92
93Clock event devices are registered either by the architecture dependent boot
94code or at module insertion time. Each clock event device fills a data
95structure with clock-specific property parameters and callback functions. The
96clock event management decides, by using the specified property parameters, the
97set of system functions a clock event device will be used to support. This
98includes the distinction of per-CPU and per-system global event devices.
99
100System-level global event devices are used for the Linux periodic tick. Per-CPU
101event devices are used to provide local CPU functionality such as process
102accounting, profiling, and high resolution timers.
103
104The management layer assigns one or more of the following functions to a clock
105event device:
106
107      - system global periodic tick (jiffies update)
108      - cpu local update_process_times
109      - cpu local profiling
110      - cpu local next event interrupt (non periodic mode)
111
112The clock event device delegates the selection of those timer interrupt related
113functions completely to the management layer. The clock management layer stores
114a function pointer in the device description structure, which has to be called
115from the hardware level handler. This removes a lot of duplicated code from the
116architecture specific timer interrupt handlers and hands the control over the
117clock event devices and the assignment of timer interrupt related functionality
118to the core code.
119
120The clock event layer API is rather small. Aside from the clock event device
121registration interface it provides functions to schedule the next event
122interrupt, clock event device notification service and support for suspend and
123resume.
124
125The framework adds about 700 lines of code which results in a 2KB increase of
126the kernel binary size. The conversion of i386 removes about 100 lines of
127code. The binary size decrease is in the range of 400 byte. We believe that the
128increase of flexibility and the avoidance of duplicated code across
129architectures justifies the slight increase of the binary size.
130
131The conversion of an architecture has no functional impact, but allows to
132utilize the high resolution and dynamic tick functionalities without any change
133to the clock event device and timer interrupt code. After the conversion the
134enabling of high resolution timers and dynamic ticks is simply provided by
135adding the kernel/time/Kconfig file to the architecture specific Kconfig and
136adding the dynamic tick specific calls to the idle routine (a total of 3 lines
137added to the idle function and the Kconfig file)
138
139Figure #4 (OLS slides p.20) illustrates the transformation.
140
141
142high resolution timer functionality
143-----------------------------------
144
145During system boot it is not possible to use the high resolution timer
146functionality, while making it possible would be difficult and would serve no
147useful function. The initialization of the clock event device framework, the
148clock source framework (GTOD) and hrtimers itself has to be done and
149appropriate clock sources and clock event devices have to be registered before
150the high resolution functionality can work. Up to the point where hrtimers are
151initialized, the system works in the usual low resolution periodic mode. The
152clock source and the clock event device layers provide notification functions
153which inform hrtimers about availability of new hardware. hrtimers validates
154the usability of the registered clock sources and clock event devices before
155switching to high resolution mode. This ensures also that a kernel which is
156configured for high resolution timers can run on a system which lacks the
157necessary hardware support.
158
159The high resolution timer code does not support SMP machines which have only
160global clock event devices. The support of such hardware would involve IPI
161calls when an interrupt happens. The overhead would be much larger than the
162benefit. This is the reason why we currently disable high resolution and
163dynamic ticks on i386 SMP systems which stop the local APIC in C3 power
164state. A workaround is available as an idea, but the problem has not been
165tackled yet.
166
167The time ordered insertion of timers provides all the infrastructure to decide
168whether the event device has to be reprogrammed when a timer is added. The
169decision is made per timer base and synchronized across per-cpu timer bases in
170a support function. The design allows the system to utilize separate per-CPU
171clock event devices for the per-CPU timer bases, but currently only one
172reprogrammable clock event device per-CPU is utilized.
173
174When the timer interrupt happens, the next event interrupt handler is called
175from the clock event distribution code and moves expired timers from the
176red-black tree to a separate double linked list and invokes the softirq
177handler. An additional mode field in the hrtimer structure allows the system to
178execute callback functions directly from the next event interrupt handler. This
179is restricted to code which can safely be executed in the hard interrupt
180context. This applies, for example, to the common case of a wakeup function as
181used by nanosleep. The advantage of executing the handler in the interrupt
182context is the avoidance of up to two context switches - from the interrupted
183context to the softirq and to the task which is woken up by the expired
184timer.
185
186Once a system has switched to high resolution mode, the periodic tick is
187switched off. This disables the per system global periodic clock event device -
188e.g. the PIT on i386 SMP systems.
189
190The periodic tick functionality is provided by an per-cpu hrtimer. The callback
191function is executed in the next event interrupt context and updates jiffies
192and calls update_process_times and profiling. The implementation of the hrtimer
193based periodic tick is designed to be extended with dynamic tick functionality.
194This allows to use a single clock event device to schedule high resolution
195timer and periodic events (jiffies tick, profiling, process accounting) on UP
196systems. This has been proved to work with the PIT on i386 and the Incrementer
197on PPC.
198
199The softirq for running the hrtimer queues and executing the callbacks has been
200separated from the tick bound timer softirq to allow accurate delivery of high
201resolution timer signals which are used by itimer and POSIX interval
202timers. The execution of this softirq can still be delayed by other softirqs,
203but the overall latencies have been significantly improved by this separation.
204
205Figure #5 (OLS slides p.22) illustrates the transformation.
206
207
208dynamic ticks
209-------------
210
211Dynamic ticks are the logical consequence of the hrtimer based periodic tick
212replacement (sched_tick). The functionality of the sched_tick hrtimer is
213extended by three functions:
214
215- hrtimer_stop_sched_tick
216- hrtimer_restart_sched_tick
217- hrtimer_update_jiffies
218
219hrtimer_stop_sched_tick() is called when a CPU goes into idle state. The code
220evaluates the next scheduled timer event (from both hrtimers and the timer
221wheel) and in case that the next event is further away than the next tick it
222reprograms the sched_tick to this future event, to allow longer idle sleeps
223without worthless interruption by the periodic tick. The function is also
224called when an interrupt happens during the idle period, which does not cause a
225reschedule. The call is necessary as the interrupt handler might have armed a
226new timer whose expiry time is before the time which was identified as the
227nearest event in the previous call to hrtimer_stop_sched_tick.
228
229hrtimer_restart_sched_tick() is called when the CPU leaves the idle state before
230it calls schedule(). hrtimer_restart_sched_tick() resumes the periodic tick,
231which is kept active until the next call to hrtimer_stop_sched_tick().
232
233hrtimer_update_jiffies() is called from irq_enter() when an interrupt happens
234in the idle period to make sure that jiffies are up to date and the interrupt
235handler has not to deal with an eventually stale jiffy value.
236
237The dynamic tick feature provides statistical values which are exported to
238userspace via /proc/stat and can be made available for enhanced power
239management control.
240
241The implementation leaves room for further development like full tickless
242systems, where the time slice is controlled by the scheduler, variable
243frequency profiling, and a complete removal of jiffies in the future.
244
245
246Aside the current initial submission of i386 support, the patchset has been
247extended to x86_64 and ARM already. Initial (work in progress) support is also
248available for MIPS and PowerPC.
249
250	  Thomas, Ingo
251