Loading...
1====================================
2Concurrency Managed Workqueue (cmwq)
3====================================
4
5:Date: September, 2010
6:Author: Tejun Heo <tj@kernel.org>
7:Author: Florian Mickler <florian@mickler.org>
8
9
10Introduction
11============
12
13There are many cases where an asynchronous process execution context
14is needed and the workqueue (wq) API is the most commonly used
15mechanism for such cases.
16
17When such an asynchronous execution context is needed, a work item
18describing which function to execute is put on a queue. An
19independent thread serves as the asynchronous execution context. The
20queue is called workqueue and the thread is called worker.
21
22While there are work items on the workqueue the worker executes the
23functions associated with the work items one after the other. When
24there is no work item left on the workqueue the worker becomes idle.
25When a new work item gets queued, the worker begins executing again.
26
27
28Why cmwq?
29=========
30
31In the original wq implementation, a multi threaded (MT) wq had one
32worker thread per CPU and a single threaded (ST) wq had one worker
33thread system-wide. A single MT wq needed to keep around the same
34number of workers as the number of CPUs. The kernel grew a lot of MT
35wq users over the years and with the number of CPU cores continuously
36rising, some systems saturated the default 32k PID space just booting
37up.
38
39Although MT wq wasted a lot of resource, the level of concurrency
40provided was unsatisfactory. The limitation was common to both ST and
41MT wq albeit less severe on MT. Each wq maintained its own separate
42worker pool. An MT wq could provide only one execution context per CPU
43while an ST wq one for the whole system. Work items had to compete for
44those very limited execution contexts leading to various problems
45including proneness to deadlocks around the single execution context.
46
47The tension between the provided level of concurrency and resource
48usage also forced its users to make unnecessary tradeoffs like libata
49choosing to use ST wq for polling PIOs and accepting an unnecessary
50limitation that no two polling PIOs can progress at the same time. As
51MT wq don't provide much better concurrency, users which require
52higher level of concurrency, like async or fscache, had to implement
53their own thread pool.
54
55Concurrency Managed Workqueue (cmwq) is a reimplementation of wq with
56focus on the following goals.
57
58* Maintain compatibility with the original workqueue API.
59
60* Use per-CPU unified worker pools shared by all wq to provide
61 flexible level of concurrency on demand without wasting a lot of
62 resource.
63
64* Automatically regulate worker pool and level of concurrency so that
65 the API users don't need to worry about such details.
66
67
68The Design
69==========
70
71In order to ease the asynchronous execution of functions a new
72abstraction, the work item, is introduced.
73
74A work item is a simple struct that holds a pointer to the function
75that is to be executed asynchronously. Whenever a driver or subsystem
76wants a function to be executed asynchronously it has to set up a work
77item pointing to that function and queue that work item on a
78workqueue.
79
80Special purpose threads, called worker threads, execute the functions
81off of the queue, one after the other. If no work is queued, the
82worker threads become idle. These worker threads are managed in so
83called worker-pools.
84
85The cmwq design differentiates between the user-facing workqueues that
86subsystems and drivers queue work items on and the backend mechanism
87which manages worker-pools and processes the queued work items.
88
89There are two worker-pools, one for normal work items and the other
90for high priority ones, for each possible CPU and some extra
91worker-pools to serve work items queued on unbound workqueues - the
92number of these backing pools is dynamic.
93
94Subsystems and drivers can create and queue work items through special
95workqueue API functions as they see fit. They can influence some
96aspects of the way the work items are executed by setting flags on the
97workqueue they are putting the work item on. These flags include
98things like CPU locality, concurrency limits, priority and more. To
99get a detailed overview refer to the API description of
100``alloc_workqueue()`` below.
101
102When a work item is queued to a workqueue, the target worker-pool is
103determined according to the queue parameters and workqueue attributes
104and appended on the shared worklist of the worker-pool. For example,
105unless specifically overridden, a work item of a bound workqueue will
106be queued on the worklist of either normal or highpri worker-pool that
107is associated to the CPU the issuer is running on.
108
109For any worker pool implementation, managing the concurrency level
110(how many execution contexts are active) is an important issue. cmwq
111tries to keep the concurrency at a minimal but sufficient level.
112Minimal to save resources and sufficient in that the system is used at
113its full capacity.
114
115Each worker-pool bound to an actual CPU implements concurrency
116management by hooking into the scheduler. The worker-pool is notified
117whenever an active worker wakes up or sleeps and keeps track of the
118number of the currently runnable workers. Generally, work items are
119not expected to hog a CPU and consume many cycles. That means
120maintaining just enough concurrency to prevent work processing from
121stalling should be optimal. As long as there are one or more runnable
122workers on the CPU, the worker-pool doesn't start execution of a new
123work, but, when the last running worker goes to sleep, it immediately
124schedules a new worker so that the CPU doesn't sit idle while there
125are pending work items. This allows using a minimal number of workers
126without losing execution bandwidth.
127
128Keeping idle workers around doesn't cost other than the memory space
129for kthreads, so cmwq holds onto idle ones for a while before killing
130them.
131
132For unbound workqueues, the number of backing pools is dynamic.
133Unbound workqueue can be assigned custom attributes using
134``apply_workqueue_attrs()`` and workqueue will automatically create
135backing worker pools matching the attributes. The responsibility of
136regulating concurrency level is on the users. There is also a flag to
137mark a bound wq to ignore the concurrency management. Please refer to
138the API section for details.
139
140Forward progress guarantee relies on that workers can be created when
141more execution contexts are necessary, which in turn is guaranteed
142through the use of rescue workers. All work items which might be used
143on code paths that handle memory reclaim are required to be queued on
144wq's that have a rescue-worker reserved for execution under memory
145pressure. Else it is possible that the worker-pool deadlocks waiting
146for execution contexts to free up.
147
148
149Application Programming Interface (API)
150=======================================
151
152``alloc_workqueue()`` allocates a wq. The original
153``create_*workqueue()`` functions are deprecated and scheduled for
154removal. ``alloc_workqueue()`` takes three arguments - ``@name``,
155``@flags`` and ``@max_active``. ``@name`` is the name of the wq and
156also used as the name of the rescuer thread if there is one.
157
158A wq no longer manages execution resources but serves as a domain for
159forward progress guarantee, flush and work item attributes. ``@flags``
160and ``@max_active`` control how work items are assigned execution
161resources, scheduled and executed.
162
163
164``flags``
165---------
166
167``WQ_UNBOUND``
168 Work items queued to an unbound wq are served by the special
169 worker-pools which host workers which are not bound to any
170 specific CPU. This makes the wq behave as a simple execution
171 context provider without concurrency management. The unbound
172 worker-pools try to start execution of work items as soon as
173 possible. Unbound wq sacrifices locality but is useful for
174 the following cases.
175
176 * Wide fluctuation in the concurrency level requirement is
177 expected and using bound wq may end up creating large number
178 of mostly unused workers across different CPUs as the issuer
179 hops through different CPUs.
180
181 * Long running CPU intensive workloads which can be better
182 managed by the system scheduler.
183
184``WQ_FREEZABLE``
185 A freezable wq participates in the freeze phase of the system
186 suspend operations. Work items on the wq are drained and no
187 new work item starts execution until thawed.
188
189``WQ_MEM_RECLAIM``
190 All wq which might be used in the memory reclaim paths **MUST**
191 have this flag set. The wq is guaranteed to have at least one
192 execution context regardless of memory pressure.
193
194``WQ_HIGHPRI``
195 Work items of a highpri wq are queued to the highpri
196 worker-pool of the target cpu. Highpri worker-pools are
197 served by worker threads with elevated nice level.
198
199 Note that normal and highpri worker-pools don't interact with
200 each other. Each maintains its separate pool of workers and
201 implements concurrency management among its workers.
202
203``WQ_CPU_INTENSIVE``
204 Work items of a CPU intensive wq do not contribute to the
205 concurrency level. In other words, runnable CPU intensive
206 work items will not prevent other work items in the same
207 worker-pool from starting execution. This is useful for bound
208 work items which are expected to hog CPU cycles so that their
209 execution is regulated by the system scheduler.
210
211 Although CPU intensive work items don't contribute to the
212 concurrency level, start of their executions is still
213 regulated by the concurrency management and runnable
214 non-CPU-intensive work items can delay execution of CPU
215 intensive work items.
216
217 This flag is meaningless for unbound wq.
218
219
220``max_active``
221--------------
222
223``@max_active`` determines the maximum number of execution contexts
224per CPU which can be assigned to the work items of a wq. For example,
225with ``@max_active`` of 16, at most 16 work items of the wq can be
226executing at the same time per CPU.
227
228Currently, for a bound wq, the maximum limit for ``@max_active`` is
229512 and the default value used when 0 is specified is 256. For an
230unbound wq, the limit is higher of 512 and 4 *
231``num_possible_cpus()``. These values are chosen sufficiently high
232such that they are not the limiting factor while providing protection
233in runaway cases.
234
235The number of active work items of a wq is usually regulated by the
236users of the wq, more specifically, by how many work items the users
237may queue at the same time. Unless there is a specific need for
238throttling the number of active work items, specifying '0' is
239recommended.
240
241Some users depend on the strict execution ordering of ST wq. The
242combination of ``@max_active`` of 1 and ``WQ_UNBOUND`` used to
243achieve this behavior. Work items on such wq were always queued to the
244unbound worker-pools and only one work item could be active at any given
245time thus achieving the same ordering property as ST wq.
246
247In the current implementation the above configuration only guarantees
248ST behavior within a given NUMA node. Instead ``alloc_ordered_queue()`` should
249be used to achieve system-wide ST behavior.
250
251
252Example Execution Scenarios
253===========================
254
255The following example execution scenarios try to illustrate how cmwq
256behave under different configurations.
257
258 Work items w0, w1, w2 are queued to a bound wq q0 on the same CPU.
259 w0 burns CPU for 5ms then sleeps for 10ms then burns CPU for 5ms
260 again before finishing. w1 and w2 burn CPU for 5ms then sleep for
261 10ms.
262
263Ignoring all other tasks, works and processing overhead, and assuming
264simple FIFO scheduling, the following is one highly simplified version
265of possible sequences of events with the original wq. ::
266
267 TIME IN MSECS EVENT
268 0 w0 starts and burns CPU
269 5 w0 sleeps
270 15 w0 wakes up and burns CPU
271 20 w0 finishes
272 20 w1 starts and burns CPU
273 25 w1 sleeps
274 35 w1 wakes up and finishes
275 35 w2 starts and burns CPU
276 40 w2 sleeps
277 50 w2 wakes up and finishes
278
279And with cmwq with ``@max_active`` >= 3, ::
280
281 TIME IN MSECS EVENT
282 0 w0 starts and burns CPU
283 5 w0 sleeps
284 5 w1 starts and burns CPU
285 10 w1 sleeps
286 10 w2 starts and burns CPU
287 15 w2 sleeps
288 15 w0 wakes up and burns CPU
289 20 w0 finishes
290 20 w1 wakes up and finishes
291 25 w2 wakes up and finishes
292
293If ``@max_active`` == 2, ::
294
295 TIME IN MSECS EVENT
296 0 w0 starts and burns CPU
297 5 w0 sleeps
298 5 w1 starts and burns CPU
299 10 w1 sleeps
300 15 w0 wakes up and burns CPU
301 20 w0 finishes
302 20 w1 wakes up and finishes
303 20 w2 starts and burns CPU
304 25 w2 sleeps
305 35 w2 wakes up and finishes
306
307Now, let's assume w1 and w2 are queued to a different wq q1 which has
308``WQ_CPU_INTENSIVE`` set, ::
309
310 TIME IN MSECS EVENT
311 0 w0 starts and burns CPU
312 5 w0 sleeps
313 5 w1 and w2 start and burn CPU
314 10 w1 sleeps
315 15 w2 sleeps
316 15 w0 wakes up and burns CPU
317 20 w0 finishes
318 20 w1 wakes up and finishes
319 25 w2 wakes up and finishes
320
321
322Guidelines
323==========
324
325* Do not forget to use ``WQ_MEM_RECLAIM`` if a wq may process work
326 items which are used during memory reclaim. Each wq with
327 ``WQ_MEM_RECLAIM`` set has an execution context reserved for it. If
328 there is dependency among multiple work items used during memory
329 reclaim, they should be queued to separate wq each with
330 ``WQ_MEM_RECLAIM``.
331
332* Unless strict ordering is required, there is no need to use ST wq.
333
334* Unless there is a specific need, using 0 for @max_active is
335 recommended. In most use cases, concurrency level usually stays
336 well under the default limit.
337
338* A wq serves as a domain for forward progress guarantee
339 (``WQ_MEM_RECLAIM``, flush and work item attributes. Work items
340 which are not involved in memory reclaim and don't need to be
341 flushed as a part of a group of work items, and don't require any
342 special attribute, can use one of the system wq. There is no
343 difference in execution characteristics between using a dedicated wq
344 and a system wq.
345
346* Unless work items are expected to consume a huge amount of CPU
347 cycles, using a bound wq is usually beneficial due to the increased
348 level of locality in wq operations and work item execution.
349
350
351Debugging
352=========
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
360 root 5671 0.0 0.0 0 0 ? S 12:07 0:00 [kworker/0:1]
361 root 5672 0.0 0.0 0 0 ? S 12:07 0:00 [kworker/1:2]
362 root 5673 0.0 0.0 0 0 ? S 12:12 0:00 [kworker/0:0]
363 root 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 being 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
390Non-reentrance Conditions
391=========================
392
393Workqueue guarantees that a work item cannot be re-entrant if the following
394conditions hold after a work item gets queued:
395
396 1. The work function hasn't been changed.
397 2. No one queues the work item to another workqueue.
398 3. The work item hasn't been reinitiated.
399
400In other words, if the above conditions hold, the work item is guaranteed to be
401executed by at most one worker system-wide at any given time.
402
403Note that requeuing the work item (to the same queue) in the self function
404doesn't break these conditions, so it's safe to do. Otherwise, caution is
405required when breaking the conditions inside a work function.
406
407
408Kernel Inline Documentations Reference
409======================================
410
411.. kernel-doc:: include/linux/workqueue.h
412
413.. kernel-doc:: kernel/workqueue.c
1=========
2Workqueue
3=========
4
5:Date: September, 2010
6:Author: Tejun Heo <tj@kernel.org>
7:Author: Florian Mickler <florian@mickler.org>
8
9
10Introduction
11============
12
13There are many cases where an asynchronous process execution context
14is needed and the workqueue (wq) API is the most commonly used
15mechanism for such cases.
16
17When such an asynchronous execution context is needed, a work item
18describing which function to execute is put on a queue. An
19independent thread serves as the asynchronous execution context. The
20queue is called workqueue and the thread is called worker.
21
22While there are work items on the workqueue the worker executes the
23functions associated with the work items one after the other. When
24there is no work item left on the workqueue the worker becomes idle.
25When a new work item gets queued, the worker begins executing again.
26
27
28Why Concurrency Managed Workqueue?
29==================================
30
31In the original wq implementation, a multi threaded (MT) wq had one
32worker thread per CPU and a single threaded (ST) wq had one worker
33thread system-wide. A single MT wq needed to keep around the same
34number of workers as the number of CPUs. The kernel grew a lot of MT
35wq users over the years and with the number of CPU cores continuously
36rising, some systems saturated the default 32k PID space just booting
37up.
38
39Although MT wq wasted a lot of resource, the level of concurrency
40provided was unsatisfactory. The limitation was common to both ST and
41MT wq albeit less severe on MT. Each wq maintained its own separate
42worker pool. An MT wq could provide only one execution context per CPU
43while an ST wq one for the whole system. Work items had to compete for
44those very limited execution contexts leading to various problems
45including proneness to deadlocks around the single execution context.
46
47The tension between the provided level of concurrency and resource
48usage also forced its users to make unnecessary tradeoffs like libata
49choosing to use ST wq for polling PIOs and accepting an unnecessary
50limitation that no two polling PIOs can progress at the same time. As
51MT wq don't provide much better concurrency, users which require
52higher level of concurrency, like async or fscache, had to implement
53their own thread pool.
54
55Concurrency Managed Workqueue (cmwq) is a reimplementation of wq with
56focus on the following goals.
57
58* Maintain compatibility with the original workqueue API.
59
60* Use per-CPU unified worker pools shared by all wq to provide
61 flexible level of concurrency on demand without wasting a lot of
62 resource.
63
64* Automatically regulate worker pool and level of concurrency so that
65 the API users don't need to worry about such details.
66
67
68The Design
69==========
70
71In order to ease the asynchronous execution of functions a new
72abstraction, the work item, is introduced.
73
74A work item is a simple struct that holds a pointer to the function
75that is to be executed asynchronously. Whenever a driver or subsystem
76wants a function to be executed asynchronously it has to set up a work
77item pointing to that function and queue that work item on a
78workqueue.
79
80A work item can be executed in either a thread or the BH (softirq) context.
81
82For threaded workqueues, special purpose threads, called [k]workers, execute
83the functions off of the queue, one after the other. If no work is queued,
84the worker threads become idle. These worker threads are managed in
85worker-pools.
86
87The cmwq design differentiates between the user-facing workqueues that
88subsystems and drivers queue work items on and the backend mechanism
89which manages worker-pools and processes the queued work items.
90
91There are two worker-pools, one for normal work items and the other
92for high priority ones, for each possible CPU and some extra
93worker-pools to serve work items queued on unbound workqueues - the
94number of these backing pools is dynamic.
95
96BH workqueues use the same framework. However, as there can only be one
97concurrent execution context, there's no need to worry about concurrency.
98Each per-CPU BH worker pool contains only one pseudo worker which represents
99the BH execution context. A BH workqueue can be considered a convenience
100interface to softirq.
101
102Subsystems and drivers can create and queue work items through special
103workqueue API functions as they see fit. They can influence some
104aspects of the way the work items are executed by setting flags on the
105workqueue they are putting the work item on. These flags include
106things like CPU locality, concurrency limits, priority and more. To
107get a detailed overview refer to the API description of
108``alloc_workqueue()`` below.
109
110When a work item is queued to a workqueue, the target worker-pool is
111determined according to the queue parameters and workqueue attributes
112and appended on the shared worklist of the worker-pool. For example,
113unless specifically overridden, a work item of a bound workqueue will
114be queued on the worklist of either normal or highpri worker-pool that
115is associated to the CPU the issuer is running on.
116
117For any thread pool implementation, managing the concurrency level
118(how many execution contexts are active) is an important issue. cmwq
119tries to keep the concurrency at a minimal but sufficient level.
120Minimal to save resources and sufficient in that the system is used at
121its full capacity.
122
123Each worker-pool bound to an actual CPU implements concurrency
124management by hooking into the scheduler. The worker-pool is notified
125whenever an active worker wakes up or sleeps and keeps track of the
126number of the currently runnable workers. Generally, work items are
127not expected to hog a CPU and consume many cycles. That means
128maintaining just enough concurrency to prevent work processing from
129stalling should be optimal. As long as there are one or more runnable
130workers on the CPU, the worker-pool doesn't start execution of a new
131work, but, when the last running worker goes to sleep, it immediately
132schedules a new worker so that the CPU doesn't sit idle while there
133are pending work items. This allows using a minimal number of workers
134without losing execution bandwidth.
135
136Keeping idle workers around doesn't cost other than the memory space
137for kthreads, so cmwq holds onto idle ones for a while before killing
138them.
139
140For unbound workqueues, the number of backing pools is dynamic.
141Unbound workqueue can be assigned custom attributes using
142``apply_workqueue_attrs()`` and workqueue will automatically create
143backing worker pools matching the attributes. The responsibility of
144regulating concurrency level is on the users. There is also a flag to
145mark a bound wq to ignore the concurrency management. Please refer to
146the API section for details.
147
148Forward progress guarantee relies on that workers can be created when
149more execution contexts are necessary, which in turn is guaranteed
150through the use of rescue workers. All work items which might be used
151on code paths that handle memory reclaim are required to be queued on
152wq's that have a rescue-worker reserved for execution under memory
153pressure. Else it is possible that the worker-pool deadlocks waiting
154for execution contexts to free up.
155
156
157Application Programming Interface (API)
158=======================================
159
160``alloc_workqueue()`` allocates a wq. The original
161``create_*workqueue()`` functions are deprecated and scheduled for
162removal. ``alloc_workqueue()`` takes three arguments - ``@name``,
163``@flags`` and ``@max_active``. ``@name`` is the name of the wq and
164also used as the name of the rescuer thread if there is one.
165
166A wq no longer manages execution resources but serves as a domain for
167forward progress guarantee, flush and work item attributes. ``@flags``
168and ``@max_active`` control how work items are assigned execution
169resources, scheduled and executed.
170
171
172``flags``
173---------
174
175``WQ_BH``
176 BH workqueues can be considered a convenience interface to softirq. BH
177 workqueues are always per-CPU and all BH work items are executed in the
178 queueing CPU's softirq context in the queueing order.
179
180 All BH workqueues must have 0 ``max_active`` and ``WQ_HIGHPRI`` is the
181 only allowed additional flag.
182
183 BH work items cannot sleep. All other features such as delayed queueing,
184 flushing and canceling are supported.
185
186``WQ_UNBOUND``
187 Work items queued to an unbound wq are served by the special
188 worker-pools which host workers which are not bound to any
189 specific CPU. This makes the wq behave as a simple execution
190 context provider without concurrency management. The unbound
191 worker-pools try to start execution of work items as soon as
192 possible. Unbound wq sacrifices locality but is useful for
193 the following cases.
194
195 * Wide fluctuation in the concurrency level requirement is
196 expected and using bound wq may end up creating large number
197 of mostly unused workers across different CPUs as the issuer
198 hops through different CPUs.
199
200 * Long running CPU intensive workloads which can be better
201 managed by the system scheduler.
202
203``WQ_FREEZABLE``
204 A freezable wq participates in the freeze phase of the system
205 suspend operations. Work items on the wq are drained and no
206 new work item starts execution until thawed.
207
208``WQ_MEM_RECLAIM``
209 All wq which might be used in the memory reclaim paths **MUST**
210 have this flag set. The wq is guaranteed to have at least one
211 execution context regardless of memory pressure.
212
213``WQ_HIGHPRI``
214 Work items of a highpri wq are queued to the highpri
215 worker-pool of the target cpu. Highpri worker-pools are
216 served by worker threads with elevated nice level.
217
218 Note that normal and highpri worker-pools don't interact with
219 each other. Each maintains its separate pool of workers and
220 implements concurrency management among its workers.
221
222``WQ_CPU_INTENSIVE``
223 Work items of a CPU intensive wq do not contribute to the
224 concurrency level. In other words, runnable CPU intensive
225 work items will not prevent other work items in the same
226 worker-pool from starting execution. This is useful for bound
227 work items which are expected to hog CPU cycles so that their
228 execution is regulated by the system scheduler.
229
230 Although CPU intensive work items don't contribute to the
231 concurrency level, start of their executions is still
232 regulated by the concurrency management and runnable
233 non-CPU-intensive work items can delay execution of CPU
234 intensive work items.
235
236 This flag is meaningless for unbound wq.
237
238
239``max_active``
240--------------
241
242``@max_active`` determines the maximum number of execution contexts per
243CPU which can be assigned to the work items of a wq. For example, with
244``@max_active`` of 16, at most 16 work items of the wq can be executing
245at the same time per CPU. This is always a per-CPU attribute, even for
246unbound workqueues.
247
248The maximum limit for ``@max_active`` is 2048 and the default value used
249when 0 is specified is 1024. These values are chosen sufficiently high
250such that they are not the limiting factor while providing protection in
251runaway cases.
252
253The number of active work items of a wq is usually regulated by the
254users of the wq, more specifically, by how many work items the users
255may queue at the same time. Unless there is a specific need for
256throttling the number of active work items, specifying '0' is
257recommended.
258
259Some users depend on strict execution ordering where only one work item
260is in flight at any given time and the work items are processed in
261queueing order. While the combination of ``@max_active`` of 1 and
262``WQ_UNBOUND`` used to achieve this behavior, this is no longer the
263case. Use alloc_ordered_workqueue() instead.
264
265
266Example Execution Scenarios
267===========================
268
269The following example execution scenarios try to illustrate how cmwq
270behave under different configurations.
271
272 Work items w0, w1, w2 are queued to a bound wq q0 on the same CPU.
273 w0 burns CPU for 5ms then sleeps for 10ms then burns CPU for 5ms
274 again before finishing. w1 and w2 burn CPU for 5ms then sleep for
275 10ms.
276
277Ignoring all other tasks, works and processing overhead, and assuming
278simple FIFO scheduling, the following is one highly simplified version
279of possible sequences of events with the original wq. ::
280
281 TIME IN MSECS EVENT
282 0 w0 starts and burns CPU
283 5 w0 sleeps
284 15 w0 wakes up and burns CPU
285 20 w0 finishes
286 20 w1 starts and burns CPU
287 25 w1 sleeps
288 35 w1 wakes up and finishes
289 35 w2 starts and burns CPU
290 40 w2 sleeps
291 50 w2 wakes up and finishes
292
293And with cmwq with ``@max_active`` >= 3, ::
294
295 TIME IN MSECS EVENT
296 0 w0 starts and burns CPU
297 5 w0 sleeps
298 5 w1 starts and burns CPU
299 10 w1 sleeps
300 10 w2 starts and burns CPU
301 15 w2 sleeps
302 15 w0 wakes up and burns CPU
303 20 w0 finishes
304 20 w1 wakes up and finishes
305 25 w2 wakes up and finishes
306
307If ``@max_active`` == 2, ::
308
309 TIME IN MSECS EVENT
310 0 w0 starts and burns CPU
311 5 w0 sleeps
312 5 w1 starts and burns CPU
313 10 w1 sleeps
314 15 w0 wakes up and burns CPU
315 20 w0 finishes
316 20 w1 wakes up and finishes
317 20 w2 starts and burns CPU
318 25 w2 sleeps
319 35 w2 wakes up and finishes
320
321Now, let's assume w1 and w2 are queued to a different wq q1 which has
322``WQ_CPU_INTENSIVE`` set, ::
323
324 TIME IN MSECS EVENT
325 0 w0 starts and burns CPU
326 5 w0 sleeps
327 5 w1 and w2 start and burn CPU
328 10 w1 sleeps
329 15 w2 sleeps
330 15 w0 wakes up and burns CPU
331 20 w0 finishes
332 20 w1 wakes up and finishes
333 25 w2 wakes up and finishes
334
335
336Guidelines
337==========
338
339* Do not forget to use ``WQ_MEM_RECLAIM`` if a wq may process work
340 items which are used during memory reclaim. Each wq with
341 ``WQ_MEM_RECLAIM`` set has an execution context reserved for it. If
342 there is dependency among multiple work items used during memory
343 reclaim, they should be queued to separate wq each with
344 ``WQ_MEM_RECLAIM``.
345
346* Unless strict ordering is required, there is no need to use ST wq.
347
348* Unless there is a specific need, using 0 for @max_active is
349 recommended. In most use cases, concurrency level usually stays
350 well under the default limit.
351
352* A wq serves as a domain for forward progress guarantee
353 (``WQ_MEM_RECLAIM``, flush and work item attributes. Work items
354 which are not involved in memory reclaim and don't need to be
355 flushed as a part of a group of work items, and don't require any
356 special attribute, can use one of the system wq. There is no
357 difference in execution characteristics between using a dedicated wq
358 and a system wq.
359
360 Note: If something may generate more than @max_active outstanding
361 work items (do stress test your producers), it may saturate a system
362 wq and potentially lead to deadlock. It should utilize its own
363 dedicated workqueue rather than the system wq.
364
365* Unless work items are expected to consume a huge amount of CPU
366 cycles, using a bound wq is usually beneficial due to the increased
367 level of locality in wq operations and work item execution.
368
369
370Affinity Scopes
371===============
372
373An unbound workqueue groups CPUs according to its affinity scope to improve
374cache locality. For example, if a workqueue is using the default affinity
375scope of "cache", it will group CPUs according to last level cache
376boundaries. A work item queued on the workqueue will be assigned to a worker
377on one of the CPUs which share the last level cache with the issuing CPU.
378Once started, the worker may or may not be allowed to move outside the scope
379depending on the ``affinity_strict`` setting of the scope.
380
381Workqueue currently supports the following affinity scopes.
382
383``default``
384 Use the scope in module parameter ``workqueue.default_affinity_scope``
385 which is always set to one of the scopes below.
386
387``cpu``
388 CPUs are not grouped. A work item issued on one CPU is processed by a
389 worker on the same CPU. This makes unbound workqueues behave as per-cpu
390 workqueues without concurrency management.
391
392``smt``
393 CPUs are grouped according to SMT boundaries. This usually means that the
394 logical threads of each physical CPU core are grouped together.
395
396``cache``
397 CPUs are grouped according to cache boundaries. Which specific cache
398 boundary is used is determined by the arch code. L3 is used in a lot of
399 cases. This is the default affinity scope.
400
401``numa``
402 CPUs are grouped according to NUMA boundaries.
403
404``system``
405 All CPUs are put in the same group. Workqueue makes no effort to process a
406 work item on a CPU close to the issuing CPU.
407
408The default affinity scope can be changed with the module parameter
409``workqueue.default_affinity_scope`` and a specific workqueue's affinity
410scope can be changed using ``apply_workqueue_attrs()``.
411
412If ``WQ_SYSFS`` is set, the workqueue will have the following affinity scope
413related interface files under its ``/sys/devices/virtual/workqueue/WQ_NAME/``
414directory.
415
416``affinity_scope``
417 Read to see the current affinity scope. Write to change.
418
419 When default is the current scope, reading this file will also show the
420 current effective scope in parentheses, for example, ``default (cache)``.
421
422``affinity_strict``
423 0 by default indicating that affinity scopes are not strict. When a work
424 item starts execution, workqueue makes a best-effort attempt to ensure
425 that the worker is inside its affinity scope, which is called
426 repatriation. Once started, the scheduler is free to move the worker
427 anywhere in the system as it sees fit. This enables benefiting from scope
428 locality while still being able to utilize other CPUs if necessary and
429 available.
430
431 If set to 1, all workers of the scope are guaranteed always to be in the
432 scope. This may be useful when crossing affinity scopes has other
433 implications, for example, in terms of power consumption or workload
434 isolation. Strict NUMA scope can also be used to match the workqueue
435 behavior of older kernels.
436
437
438Affinity Scopes and Performance
439===============================
440
441It'd be ideal if an unbound workqueue's behavior is optimal for vast
442majority of use cases without further tuning. Unfortunately, in the current
443kernel, there exists a pronounced trade-off between locality and utilization
444necessitating explicit configurations when workqueues are heavily used.
445
446Higher locality leads to higher efficiency where more work is performed for
447the same number of consumed CPU cycles. However, higher locality may also
448cause lower overall system utilization if the work items are not spread
449enough across the affinity scopes by the issuers. The following performance
450testing with dm-crypt clearly illustrates this trade-off.
451
452The tests are run on a CPU with 12-cores/24-threads split across four L3
453caches (AMD Ryzen 9 3900x). CPU clock boost is turned off for consistency.
454``/dev/dm-0`` is a dm-crypt device created on NVME SSD (Samsung 990 PRO) and
455opened with ``cryptsetup`` with default settings.
456
457
458Scenario 1: Enough issuers and work spread across the machine
459-------------------------------------------------------------
460
461The command used: ::
462
463 $ fio --filename=/dev/dm-0 --direct=1 --rw=randrw --bs=32k --ioengine=libaio \
464 --iodepth=64 --runtime=60 --numjobs=24 --time_based --group_reporting \
465 --name=iops-test-job --verify=sha512
466
467There are 24 issuers, each issuing 64 IOs concurrently. ``--verify=sha512``
468makes ``fio`` generate and read back the content each time which makes
469execution locality matter between the issuer and ``kcryptd``. The following
470are the read bandwidths and CPU utilizations depending on different affinity
471scope settings on ``kcryptd`` measured over five runs. Bandwidths are in
472MiBps, and CPU util in percents.
473
474.. list-table::
475 :widths: 16 20 20
476 :header-rows: 1
477
478 * - Affinity
479 - Bandwidth (MiBps)
480 - CPU util (%)
481
482 * - system
483 - 1159.40 ±1.34
484 - 99.31 ±0.02
485
486 * - cache
487 - 1166.40 ±0.89
488 - 99.34 ±0.01
489
490 * - cache (strict)
491 - 1166.00 ±0.71
492 - 99.35 ±0.01
493
494With enough issuers spread across the system, there is no downside to
495"cache", strict or otherwise. All three configurations saturate the whole
496machine but the cache-affine ones outperform by 0.6% thanks to improved
497locality.
498
499
500Scenario 2: Fewer issuers, enough work for saturation
501-----------------------------------------------------
502
503The command used: ::
504
505 $ fio --filename=/dev/dm-0 --direct=1 --rw=randrw --bs=32k \
506 --ioengine=libaio --iodepth=64 --runtime=60 --numjobs=8 \
507 --time_based --group_reporting --name=iops-test-job --verify=sha512
508
509The only difference from the previous scenario is ``--numjobs=8``. There are
510a third of the issuers but is still enough total work to saturate the
511system.
512
513.. list-table::
514 :widths: 16 20 20
515 :header-rows: 1
516
517 * - Affinity
518 - Bandwidth (MiBps)
519 - CPU util (%)
520
521 * - system
522 - 1155.40 ±0.89
523 - 97.41 ±0.05
524
525 * - cache
526 - 1154.40 ±1.14
527 - 96.15 ±0.09
528
529 * - cache (strict)
530 - 1112.00 ±4.64
531 - 93.26 ±0.35
532
533This is more than enough work to saturate the system. Both "system" and
534"cache" are nearly saturating the machine but not fully. "cache" is using
535less CPU but the better efficiency puts it at the same bandwidth as
536"system".
537
538Eight issuers moving around over four L3 cache scope still allow "cache
539(strict)" to mostly saturate the machine but the loss of work conservation
540is now starting to hurt with 3.7% bandwidth loss.
541
542
543Scenario 3: Even fewer issuers, not enough work to saturate
544-----------------------------------------------------------
545
546The command used: ::
547
548 $ fio --filename=/dev/dm-0 --direct=1 --rw=randrw --bs=32k \
549 --ioengine=libaio --iodepth=64 --runtime=60 --numjobs=4 \
550 --time_based --group_reporting --name=iops-test-job --verify=sha512
551
552Again, the only difference is ``--numjobs=4``. With the number of issuers
553reduced to four, there now isn't enough work to saturate the whole system
554and the bandwidth becomes dependent on completion latencies.
555
556.. list-table::
557 :widths: 16 20 20
558 :header-rows: 1
559
560 * - Affinity
561 - Bandwidth (MiBps)
562 - CPU util (%)
563
564 * - system
565 - 993.60 ±1.82
566 - 75.49 ±0.06
567
568 * - cache
569 - 973.40 ±1.52
570 - 74.90 ±0.07
571
572 * - cache (strict)
573 - 828.20 ±4.49
574 - 66.84 ±0.29
575
576Now, the tradeoff between locality and utilization is clearer. "cache" shows
5772% bandwidth loss compared to "system" and "cache (struct)" whopping 20%.
578
579
580Conclusion and Recommendations
581------------------------------
582
583In the above experiments, the efficiency advantage of the "cache" affinity
584scope over "system" is, while consistent and noticeable, small. However, the
585impact is dependent on the distances between the scopes and may be more
586pronounced in processors with more complex topologies.
587
588While the loss of work-conservation in certain scenarios hurts, it is a lot
589better than "cache (strict)" and maximizing workqueue utilization is
590unlikely to be the common case anyway. As such, "cache" is the default
591affinity scope for unbound pools.
592
593* As there is no one option which is great for most cases, workqueue usages
594 that may consume a significant amount of CPU are recommended to configure
595 the workqueues using ``apply_workqueue_attrs()`` and/or enable
596 ``WQ_SYSFS``.
597
598* An unbound workqueue with strict "cpu" affinity scope behaves the same as
599 ``WQ_CPU_INTENSIVE`` per-cpu workqueue. There is no real advanage to the
600 latter and an unbound workqueue provides a lot more flexibility.
601
602* Affinity scopes are introduced in Linux v6.5. To emulate the previous
603 behavior, use strict "numa" affinity scope.
604
605* The loss of work-conservation in non-strict affinity scopes is likely
606 originating from the scheduler. There is no theoretical reason why the
607 kernel wouldn't be able to do the right thing and maintain
608 work-conservation in most cases. As such, it is possible that future
609 scheduler improvements may make most of these tunables unnecessary.
610
611
612Examining Configuration
613=======================
614
615Use tools/workqueue/wq_dump.py to examine unbound CPU affinity
616configuration, worker pools and how workqueues map to the pools: ::
617
618 $ tools/workqueue/wq_dump.py
619 Affinity Scopes
620 ===============
621 wq_unbound_cpumask=0000000f
622
623 CPU
624 nr_pods 4
625 pod_cpus [0]=00000001 [1]=00000002 [2]=00000004 [3]=00000008
626 pod_node [0]=0 [1]=0 [2]=1 [3]=1
627 cpu_pod [0]=0 [1]=1 [2]=2 [3]=3
628
629 SMT
630 nr_pods 4
631 pod_cpus [0]=00000001 [1]=00000002 [2]=00000004 [3]=00000008
632 pod_node [0]=0 [1]=0 [2]=1 [3]=1
633 cpu_pod [0]=0 [1]=1 [2]=2 [3]=3
634
635 CACHE (default)
636 nr_pods 2
637 pod_cpus [0]=00000003 [1]=0000000c
638 pod_node [0]=0 [1]=1
639 cpu_pod [0]=0 [1]=0 [2]=1 [3]=1
640
641 NUMA
642 nr_pods 2
643 pod_cpus [0]=00000003 [1]=0000000c
644 pod_node [0]=0 [1]=1
645 cpu_pod [0]=0 [1]=0 [2]=1 [3]=1
646
647 SYSTEM
648 nr_pods 1
649 pod_cpus [0]=0000000f
650 pod_node [0]=-1
651 cpu_pod [0]=0 [1]=0 [2]=0 [3]=0
652
653 Worker Pools
654 ============
655 pool[00] ref= 1 nice= 0 idle/workers= 4/ 4 cpu= 0
656 pool[01] ref= 1 nice=-20 idle/workers= 2/ 2 cpu= 0
657 pool[02] ref= 1 nice= 0 idle/workers= 4/ 4 cpu= 1
658 pool[03] ref= 1 nice=-20 idle/workers= 2/ 2 cpu= 1
659 pool[04] ref= 1 nice= 0 idle/workers= 4/ 4 cpu= 2
660 pool[05] ref= 1 nice=-20 idle/workers= 2/ 2 cpu= 2
661 pool[06] ref= 1 nice= 0 idle/workers= 3/ 3 cpu= 3
662 pool[07] ref= 1 nice=-20 idle/workers= 2/ 2 cpu= 3
663 pool[08] ref=42 nice= 0 idle/workers= 6/ 6 cpus=0000000f
664 pool[09] ref=28 nice= 0 idle/workers= 3/ 3 cpus=00000003
665 pool[10] ref=28 nice= 0 idle/workers= 17/ 17 cpus=0000000c
666 pool[11] ref= 1 nice=-20 idle/workers= 1/ 1 cpus=0000000f
667 pool[12] ref= 2 nice=-20 idle/workers= 1/ 1 cpus=00000003
668 pool[13] ref= 2 nice=-20 idle/workers= 1/ 1 cpus=0000000c
669
670 Workqueue CPU -> pool
671 =====================
672 [ workqueue \ CPU 0 1 2 3 dfl]
673 events percpu 0 2 4 6
674 events_highpri percpu 1 3 5 7
675 events_long percpu 0 2 4 6
676 events_unbound unbound 9 9 10 10 8
677 events_freezable percpu 0 2 4 6
678 events_power_efficient percpu 0 2 4 6
679 events_freezable_pwr_ef percpu 0 2 4 6
680 rcu_gp percpu 0 2 4 6
681 rcu_par_gp percpu 0 2 4 6
682 slub_flushwq percpu 0 2 4 6
683 netns ordered 8 8 8 8 8
684 ...
685
686See the command's help message for more info.
687
688
689Monitoring
690==========
691
692Use tools/workqueue/wq_monitor.py to monitor workqueue operations: ::
693
694 $ tools/workqueue/wq_monitor.py events
695 total infl CPUtime CPUhog CMW/RPR mayday rescued
696 events 18545 0 6.1 0 5 - -
697 events_highpri 8 0 0.0 0 0 - -
698 events_long 3 0 0.0 0 0 - -
699 events_unbound 38306 0 0.1 - 7 - -
700 events_freezable 0 0 0.0 0 0 - -
701 events_power_efficient 29598 0 0.2 0 0 - -
702 events_freezable_pwr_ef 10 0 0.0 0 0 - -
703 sock_diag_events 0 0 0.0 0 0 - -
704
705 total infl CPUtime CPUhog CMW/RPR mayday rescued
706 events 18548 0 6.1 0 5 - -
707 events_highpri 8 0 0.0 0 0 - -
708 events_long 3 0 0.0 0 0 - -
709 events_unbound 38322 0 0.1 - 7 - -
710 events_freezable 0 0 0.0 0 0 - -
711 events_power_efficient 29603 0 0.2 0 0 - -
712 events_freezable_pwr_ef 10 0 0.0 0 0 - -
713 sock_diag_events 0 0 0.0 0 0 - -
714
715 ...
716
717See the command's help message for more info.
718
719
720Debugging
721=========
722
723Because the work functions are executed by generic worker threads
724there are a few tricks needed to shed some light on misbehaving
725workqueue users.
726
727Worker threads show up in the process list as: ::
728
729 root 5671 0.0 0.0 0 0 ? S 12:07 0:00 [kworker/0:1]
730 root 5672 0.0 0.0 0 0 ? S 12:07 0:00 [kworker/1:2]
731 root 5673 0.0 0.0 0 0 ? S 12:12 0:00 [kworker/0:0]
732 root 5674 0.0 0.0 0 0 ? S 12:13 0:00 [kworker/1:0]
733
734If kworkers are going crazy (using too much cpu), there are two types
735of possible problems:
736
737 1. Something being scheduled in rapid succession
738 2. A single work item that consumes lots of cpu cycles
739
740The first one can be tracked using tracing: ::
741
742 $ echo workqueue:workqueue_queue_work > /sys/kernel/tracing/set_event
743 $ cat /sys/kernel/tracing/trace_pipe > out.txt
744 (wait a few secs)
745 ^C
746
747If something is busy looping on work queueing, it would be dominating
748the output and the offender can be determined with the work item
749function.
750
751For the second type of problems it should be possible to just check
752the stack trace of the offending worker thread. ::
753
754 $ cat /proc/THE_OFFENDING_KWORKER/stack
755
756The work item's function should be trivially visible in the stack
757trace.
758
759
760Non-reentrance Conditions
761=========================
762
763Workqueue guarantees that a work item cannot be re-entrant if the following
764conditions hold after a work item gets queued:
765
766 1. The work function hasn't been changed.
767 2. No one queues the work item to another workqueue.
768 3. The work item hasn't been reinitiated.
769
770In other words, if the above conditions hold, the work item is guaranteed to be
771executed by at most one worker system-wide at any given time.
772
773Note that requeuing the work item (to the same queue) in the self function
774doesn't break these conditions, so it's safe to do. Otherwise, caution is
775required when breaking the conditions inside a work function.
776
777
778Kernel Inline Documentations Reference
779======================================
780
781.. kernel-doc:: include/linux/workqueue.h
782
783.. kernel-doc:: kernel/workqueue.c