1
2Concurrency Managed Workqueue (cmwq)
3
4September, 2010		Tejun Heo <tj@kernel.org>
5			Florian Mickler <florian@mickler.org>
6
7CONTENTS
8
91. Introduction
102. Why cmwq?
113. The Design
124. Application Programming Interface (API)
135. Example Execution Scenarios
146. Guidelines
157. Debugging
16
17
181. Introduction
19
20There are many cases where an asynchronous process execution context
21is needed and the workqueue (wq) API is the most commonly used
22mechanism for such cases.
23
24When such an asynchronous execution context is needed, a work item
25describing which function to execute is put on a queue.  An
26independent thread serves as the asynchronous execution context.  The
27queue is called workqueue and the thread is called worker.
28
29While there are work items on the workqueue the worker executes the
30functions associated with the work items one after the other.  When
31there is no work item left on the workqueue the worker becomes idle.
32When a new work item gets queued, the worker begins executing again.
33
34
352. Why cmwq?
36
37In the original wq implementation, a multi threaded (MT) wq had one
38worker thread per CPU and a single threaded (ST) wq had one worker
39thread system-wide.  A single MT wq needed to keep around the same
40number of workers as the number of CPUs.  The kernel grew a lot of MT
41wq users over the years and with the number of CPU cores continuously
42rising, some systems saturated the default 32k PID space just booting
43up.
44
45Although MT wq wasted a lot of resource, the level of concurrency
46provided was unsatisfactory.  The limitation was common to both ST and
47MT wq albeit less severe on MT.  Each wq maintained its own separate
48worker pool.  A MT wq could provide only one execution context per CPU
49while a ST wq one for the whole system.  Work items had to compete for
50those very limited execution contexts leading to various problems
51including proneness to deadlocks around the single execution context.
52
53The tension between the provided level of concurrency and resource
54usage also forced its users to make unnecessary tradeoffs like libata
55choosing to use ST wq for polling PIOs and accepting an unnecessary
56limitation that no two polling PIOs can progress at the same time.  As
57MT wq don't provide much better concurrency, users which require
58higher level of concurrency, like async or fscache, had to implement
59their own thread pool.
60
61Concurrency Managed Workqueue (cmwq) is a reimplementation of wq with
62focus on the following goals.
63
64* Maintain compatibility with the original workqueue API.
65
66* Use per-CPU unified worker pools shared by all wq to provide
67  flexible level of concurrency on demand without wasting a lot of
68  resource.
69
70* Automatically regulate worker pool and level of concurrency so that
71  the API users don't need to worry about such details.
72
73
743. The Design
75
76In order to ease the asynchronous execution of functions a new
77abstraction, the work item, is introduced.
78
79A work item is a simple struct that holds a pointer to the function
80that is to be executed asynchronously.  Whenever a driver or subsystem
81wants a function to be executed asynchronously it has to set up a work
82item pointing to that function and queue that work item on a
83workqueue.
84
85Special purpose threads, called worker threads, execute the functions
86off of the queue, one after the other.  If no work is queued, the
87worker threads become idle.  These worker threads are managed in so
88called worker-pools.
89
90The cmwq design differentiates between the user-facing workqueues that
91subsystems and drivers queue work items on and the backend mechanism
92which manages worker-pools and processes the queued work items.
93
94There are two worker-pools, one for normal work items and the other
95for high priority ones, for each possible CPU and some extra
96worker-pools to serve work items queued on unbound workqueues - the
97number of these backing pools is dynamic.
98
99Subsystems and drivers can create and queue work items through special
100workqueue API functions as they see fit. They can influence some
101aspects of the way the work items are executed by setting flags on the
102workqueue they are putting the work item on. These flags include
103things like CPU locality, concurrency limits, priority and more.  To
104get a detailed overview refer to the API description of
105alloc_workqueue() below.
106
107When a work item is queued to a workqueue, the target worker-pool is
108determined according to the queue parameters and workqueue attributes
109and appended on the shared worklist of the worker-pool.  For example,
110unless specifically overridden, a work item of a bound workqueue will
111be queued on the worklist of either normal or highpri worker-pool that
112is associated to the CPU the issuer is running on.
113
114For any worker pool implementation, managing the concurrency level
115(how many execution contexts are active) is an important issue.  cmwq
116tries to keep the concurrency at a minimal but sufficient level.
117Minimal to save resources and sufficient in that the system is used at
118its full capacity.
119
120Each worker-pool bound to an actual CPU implements concurrency
121management by hooking into the scheduler.  The worker-pool is notified
122whenever an active worker wakes up or sleeps and keeps track of the
123number of the currently runnable workers.  Generally, work items are
124not expected to hog a CPU and consume many cycles.  That means
125maintaining just enough concurrency to prevent work processing from
126stalling should be optimal.  As long as there are one or more runnable
127workers on the CPU, the worker-pool doesn't start execution of a new
128work, but, when the last running worker goes to sleep, it immediately
129schedules a new worker so that the CPU doesn't sit idle while there
130are pending work items.  This allows using a minimal number of workers
131without losing execution bandwidth.
132
133Keeping idle workers around doesn't cost other than the memory space
134for kthreads, so cmwq holds onto idle ones for a while before killing
135them.
136
137For unbound workqueues, the number of backing pools is dynamic.
138Unbound workqueue can be assigned custom attributes using
139apply_workqueue_attrs() and workqueue will automatically create
140backing worker pools matching the attributes.  The responsibility of
141regulating concurrency level is on the users.  There is also a flag to
142mark a bound wq to ignore the concurrency management.  Please refer to
143the API section for details.
144
145Forward progress guarantee relies on that workers can be created when
146more execution contexts are necessary, which in turn is guaranteed
147through the use of rescue workers.  All work items which might be used
148on code paths that handle memory reclaim are required to be queued on
149wq's that have a rescue-worker reserved for execution under memory
150pressure.  Else it is possible that the worker-pool deadlocks waiting
151for execution contexts to free up.
152
153
1544. Application Programming Interface (API)
155
156alloc_workqueue() allocates a wq.  The original create_*workqueue()
157functions are deprecated and scheduled for removal.  alloc_workqueue()
158takes three arguments - @name, @flags and @max_active.  @name is the
159name of the wq and also used as the name of the rescuer thread if
160there is one.
161
162A wq no longer manages execution resources but serves as a domain for
163forward progress guarantee, flush and work item attributes.  @flags
164and @max_active control how work items are assigned execution
165resources, scheduled and executed.
166
167@flags:
168
169  WQ_UNBOUND
170
171	Work items queued to an unbound wq are served by the special
172	woker-pools which host workers which are not bound to any
173	specific CPU.  This makes the wq behave as a simple execution
174	context provider without concurrency management.  The unbound
175	worker-pools try to start execution of work items as soon as
176	possible.  Unbound wq sacrifices locality but is useful for
177	the following cases.
178
179	* Wide fluctuation in the concurrency level requirement is
180	  expected and using bound wq may end up creating large number
181	  of mostly unused workers across different CPUs as the issuer
182	  hops through different CPUs.
183
184	* Long running CPU intensive workloads which can be better
185	  managed by the system scheduler.
186
187  WQ_FREEZABLE
188
189	A freezable wq participates in the freeze phase of the system
190	suspend operations.  Work items on the wq are drained and no
191	new work item starts execution until thawed.
192
193  WQ_MEM_RECLAIM
194
195	All wq which might be used in the memory reclaim paths _MUST_
196	have this flag set.  The wq is guaranteed to have at least one
197	execution context regardless of memory pressure.
198
199  WQ_HIGHPRI
200
201	Work items of a highpri wq are queued to the highpri
202	worker-pool of the target cpu.  Highpri worker-pools are
203	served by worker threads with elevated nice level.
204
205	Note that normal and highpri worker-pools don't interact with
206	each other.  Each maintain its separate pool of workers and
207	implements concurrency management among its workers.
208
209  WQ_CPU_INTENSIVE
210
211	Work items of a CPU intensive wq do not contribute to the
212	concurrency level.  In other words, runnable CPU intensive
213	work items will not prevent other work items in the same
214	worker-pool from starting execution.  This is useful for bound
215	work items which are expected to hog CPU cycles so that their
216	execution is regulated by the system scheduler.
217
218	Although CPU intensive work items don't contribute to the
219	concurrency level, start of their executions is still
220	regulated by the concurrency management and runnable
221	non-CPU-intensive work items can delay execution of CPU
222	intensive work items.
223
224	This flag is meaningless for unbound wq.
225
226Note that the flag WQ_NON_REENTRANT no longer exists as all workqueues
227are now non-reentrant - any work item is guaranteed to be executed by
228at most one worker system-wide at any given time.
229
230@max_active:
231
232@max_active determines the maximum number of execution contexts per
233CPU which can be assigned to the work items of a wq.  For example,
234with @max_active of 16, at most 16 work items of the wq can be
235executing at the same time per CPU.
236
237Currently, for a bound wq, the maximum limit for @max_active is 512
238and the default value used when 0 is specified is 256.  For an unbound
239wq, the limit is higher of 512 and 4 * num_possible_cpus().  These
240values are chosen sufficiently high such that they are not the
241limiting factor while providing protection in runaway cases.
242
243The number of active work items of a wq is usually regulated by the
244users of the wq, more specifically, by how many work items the users
245may queue at the same time.  Unless there is a specific need for
246throttling the number of active work items, specifying '0' is
247recommended.
248
249Some users depend on the strict execution ordering of ST wq.  The
250combination of @max_active of 1 and WQ_UNBOUND is used to achieve this
251behavior.  Work items on such wq are always queued to the unbound
252worker-pools and only one work item can be active at any given time thus
253achieving the same ordering property as ST wq.
254
255
2565. Example Execution Scenarios
257
258The following example execution scenarios try to illustrate how cmwq
259behave under different configurations.
260
261 Work items w0, w1, w2 are queued to a bound wq q0 on the same CPU.
262 w0 burns CPU for 5ms then sleeps for 10ms then burns CPU for 5ms
263 again before finishing.  w1 and w2 burn CPU for 5ms then sleep for
264 10ms.
265
266Ignoring all other tasks, works and processing overhead, and assuming
267simple FIFO scheduling, the following is one highly simplified version
268of possible sequences of events with the original wq.
269
270 TIME IN MSECS	EVENT
271 0		w0 starts and burns CPU
272 5		w0 sleeps
273 15		w0 wakes up and burns CPU
274 20		w0 finishes
275 20		w1 starts and burns CPU
276 25		w1 sleeps
277 35		w1 wakes up and finishes
278 35		w2 starts and burns CPU
279 40		w2 sleeps
280 50		w2 wakes up and finishes
281
282And with cmwq with @max_active >= 3,
283
284 TIME IN MSECS	EVENT
285 0		w0 starts and burns CPU
286 5		w0 sleeps
287 5		w1 starts and burns CPU
288 10		w1 sleeps
289 10		w2 starts and burns CPU
290 15		w2 sleeps
291 15		w0 wakes up and burns CPU
292 20		w0 finishes
293 20		w1 wakes up and finishes
294 25		w2 wakes up and finishes
295
296If @max_active == 2,
297
298 TIME IN MSECS	EVENT
299 0		w0 starts and burns CPU
300 5		w0 sleeps
301 5		w1 starts and burns CPU
302 10		w1 sleeps
303 15		w0 wakes up and burns CPU
304 20		w0 finishes
305 20		w1 wakes up and finishes
306 20		w2 starts and burns CPU
307 25		w2 sleeps
308 35		w2 wakes up and finishes
309
310Now, let's assume w1 and w2 are queued to a different wq q1 which has
311WQ_CPU_INTENSIVE set,
312
313 TIME IN MSECS	EVENT
314 0		w0 starts and burns CPU
315 5		w0 sleeps
316 5		w1 and w2 start and burn CPU
317 10		w1 sleeps
318 15		w2 sleeps
319 15		w0 wakes up and burns CPU
320 20		w0 finishes
321 20		w1 wakes up and finishes
322 25		w2 wakes up and finishes
323
324
3256. Guidelines
326
327* Do not forget to use WQ_MEM_RECLAIM if a wq may process work items
328  which are used during memory reclaim.  Each wq with WQ_MEM_RECLAIM
329  set has an execution context reserved for it.  If there is
330  dependency among multiple work items used during memory reclaim,
331  they should be queued to separate wq each with WQ_MEM_RECLAIM.
332
333* Unless strict ordering is required, there is no need to use ST wq.
334
335* Unless there is a specific need, using 0 for @max_active is
336  recommended.  In most use cases, concurrency level usually stays
337  well under the default limit.
338
339* A wq serves as a domain for forward progress guarantee
340  (WQ_MEM_RECLAIM, flush and work item attributes.  Work items which
341  are not involved in memory reclaim and don't need to be flushed as a
342  part of a group of work items, and don't require any special
343  attribute, can use one of the system wq.  There is no difference in
344  execution characteristics between using a dedicated wq and a system
345  wq.
346
347* Unless work items are expected to consume a huge amount of CPU
348  cycles, using a bound wq is usually beneficial due to the increased
349  level of locality in wq operations and work item execution.
350
351
3527. Debugging
353
354Because the work functions are executed by generic worker threads
355there are a few tricks needed to shed some light on misbehaving
356workqueue users.
357
358Worker threads show up in the process list as:
359
360root      5671  0.0  0.0      0     0 ?        S    12:07   0:00 [kworker/0:1]
361root      5672  0.0  0.0      0     0 ?        S    12:07   0:00 [kworker/1:2]
362root      5673  0.0  0.0      0     0 ?        S    12:12   0:00 [kworker/0:0]
363root      5674  0.0  0.0      0     0 ?        S    12:13   0:00 [kworker/1:0]
364
365If kworkers are going crazy (using too much cpu), there are two types
366of possible problems:
367
368	1. Something beeing scheduled in rapid succession
369	2. A single work item that consumes lots of cpu cycles
370
371The first one can be tracked using tracing:
372
373	$ echo workqueue:workqueue_queue_work > /sys/kernel/debug/tracing/set_event
374	$ cat /sys/kernel/debug/tracing/trace_pipe > out.txt
375	(wait a few secs)
376	^C
377
378If something is busy looping on work queueing, it would be dominating
379the output and the offender can be determined with the work item
380function.
381
382For the second type of problems it should be possible to just check
383the stack trace of the offending worker thread.
384
385	$ cat /proc/THE_OFFENDING_KWORKER/stack
386
387The work item's function should be trivially visible in the stack
388trace.
389