2 Concurrency Managed Workqueue (cmwq)
4 September, 2010 Tejun Heo <tj@kernel.org>
5 Florian Mickler <florian@mickler.org>
12 4. Application Programming Interface (API)
13 5. Example Execution Scenarios
20 There are many cases where an asynchronous process execution context
21 is needed and the workqueue (wq) API is the most commonly used
22 mechanism for such cases.
24 When such an asynchronous execution context is needed, a work item
25 describing which function to execute is put on a queue. An
26 independent thread serves as the asynchronous execution context. The
27 queue is called workqueue and the thread is called worker.
29 While there are work items on the workqueue the worker executes the
30 functions associated with the work items one after the other. When
31 there is no work item left on the workqueue the worker becomes idle.
32 When a new work item gets queued, the worker begins executing again.
37 In the original wq implementation, a multi threaded (MT) wq had one
38 worker thread per CPU and a single threaded (ST) wq had one worker
39 thread system-wide. A single MT wq needed to keep around the same
40 number of workers as the number of CPUs. The kernel grew a lot of MT
41 wq users over the years and with the number of CPU cores continuously
42 rising, some systems saturated the default 32k PID space just booting
45 Although MT wq wasted a lot of resource, the level of concurrency
46 provided was unsatisfactory. The limitation was common to both ST and
47 MT wq albeit less severe on MT. Each wq maintained its own separate
48 worker pool. A MT wq could provide only one execution context per CPU
49 while a ST wq one for the whole system. Work items had to compete for
50 those very limited execution contexts leading to various problems
51 including proneness to deadlocks around the single execution context.
53 The tension between the provided level of concurrency and resource
54 usage also forced its users to make unnecessary tradeoffs like libata
55 choosing to use ST wq for polling PIOs and accepting an unnecessary
56 limitation that no two polling PIOs can progress at the same time. As
57 MT wq don't provide much better concurrency, users which require
58 higher level of concurrency, like async or fscache, had to implement
59 their own thread pool.
61 Concurrency Managed Workqueue (cmwq) is a reimplementation of wq with
62 focus on the following goals.
64 * Maintain compatibility with the original workqueue API.
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
70 * Automatically regulate worker pool and level of concurrency so that
71 the API users don't need to worry about such details.
76 In order to ease the asynchronous execution of functions a new
77 abstraction, the work item, is introduced.
79 A work item is a simple struct that holds a pointer to the function
80 that is to be executed asynchronously. Whenever a driver or subsystem
81 wants a function to be executed asynchronously it has to set up a work
82 item pointing to that function and queue that work item on a
85 Special purpose threads, called worker threads, execute the functions
86 off of the queue, one after the other. If no work is queued, the
87 worker threads become idle. These worker threads are managed in so
90 The cmwq design differentiates between the user-facing workqueues that
91 subsystems and drivers queue work items on and the backend mechanism
92 which manages thread-pools and processes the queued work items.
94 The backend is called gcwq. There is one gcwq for each possible CPU
95 and one gcwq to serve work items queued on unbound workqueues. Each
96 gcwq has two thread-pools - one for normal work items and the other
97 for high priority ones.
99 Subsystems and drivers can create and queue work items through special
100 workqueue API functions as they see fit. They can influence some
101 aspects of the way the work items are executed by setting flags on the
102 workqueue they are putting the work item on. These flags include
103 things like CPU locality, reentrancy, concurrency limits, priority and
104 more. To get a detailed overview refer to the API description of
105 alloc_workqueue() below.
107 When a work item is queued to a workqueue, the target gcwq and
108 thread-pool is determined according to the queue parameters and
109 workqueue attributes and appended on the shared worklist of the
110 thread-pool. For example, unless specifically overridden, a work item
111 of a bound workqueue will be queued on the worklist of either normal
112 or highpri thread-pool of the gcwq that is associated to the CPU the
113 issuer is running on.
115 For any worker pool implementation, managing the concurrency level
116 (how many execution contexts are active) is an important issue. cmwq
117 tries to keep the concurrency at a minimal but sufficient level.
118 Minimal to save resources and sufficient in that the system is used at
121 Each thread-pool bound to an actual CPU implements concurrency
122 management by hooking into the scheduler. The thread-pool is notified
123 whenever an active worker wakes up or sleeps and keeps track of the
124 number of the currently runnable workers. Generally, work items are
125 not expected to hog a CPU and consume many cycles. That means
126 maintaining just enough concurrency to prevent work processing from
127 stalling should be optimal. As long as there are one or more runnable
128 workers on the CPU, the thread-pool doesn't start execution of a new
129 work, but, when the last running worker goes to sleep, it immediately
130 schedules a new worker so that the CPU doesn't sit idle while there
131 are pending work items. This allows using a minimal number of workers
132 without losing execution bandwidth.
134 Keeping idle workers around doesn't cost other than the memory space
135 for kthreads, so cmwq holds onto idle ones for a while before killing
138 For an unbound wq, the above concurrency management doesn't apply and
139 the thread-pools for the pseudo unbound CPU try to start executing all
140 work items as soon as possible. The responsibility of regulating
141 concurrency level is on the users. There is also a flag to mark a
142 bound wq to ignore the concurrency management. Please refer to the
143 API section for details.
145 Forward progress guarantee relies on that workers can be created when
146 more execution contexts are necessary, which in turn is guaranteed
147 through the use of rescue workers. All work items which might be used
148 on code paths that handle memory reclaim are required to be queued on
149 wq's that have a rescue-worker reserved for execution under memory
150 pressure. Else it is possible that the thread-pool deadlocks waiting
151 for execution contexts to free up.
154 4. Application Programming Interface (API)
156 alloc_workqueue() allocates a wq. The original create_*workqueue()
157 functions are deprecated and scheduled for removal. alloc_workqueue()
158 takes three arguments - @name, @flags and @max_active. @name is the
159 name of the wq and also used as the name of the rescuer thread if
162 A wq no longer manages execution resources but serves as a domain for
163 forward progress guarantee, flush and work item attributes. @flags
164 and @max_active control how work items are assigned execution
165 resources, scheduled and executed.
171 By default, a wq guarantees non-reentrance only on the same
172 CPU. A work item may not be executed concurrently on the same
173 CPU by multiple workers but is allowed to be executed
174 concurrently on multiple CPUs. This flag makes sure
175 non-reentrance is enforced across all CPUs. Work items queued
176 to a non-reentrant wq are guaranteed to be executed by at most
177 one worker system-wide at any given time.
181 Work items queued to an unbound wq are served by a special
182 gcwq which hosts workers which are not bound to any specific
183 CPU. This makes the wq behave as a simple execution context
184 provider without concurrency management. The unbound gcwq
185 tries to start execution of work items as soon as possible.
186 Unbound wq sacrifices locality but is useful for the following
189 * Wide fluctuation in the concurrency level requirement is
190 expected and using bound wq may end up creating large number
191 of mostly unused workers across different CPUs as the issuer
192 hops through different CPUs.
194 * Long running CPU intensive workloads which can be better
195 managed by the system scheduler.
199 A freezable wq participates in the freeze phase of the system
200 suspend operations. Work items on the wq are drained and no
201 new work item starts execution until thawed.
205 All wq which might be used in the memory reclaim paths _MUST_
206 have this flag set. The wq is guaranteed to have at least one
207 execution context regardless of memory pressure.
211 Work items of a highpri wq are queued to the highpri
212 thread-pool of the target gcwq. Highpri thread-pools are
213 served by worker threads with elevated nice level.
215 Note that normal and highpri thread-pools don't interact with
216 each other. Each maintain its separate pool of workers and
217 implements concurrency management among its workers.
221 Work items of a CPU intensive wq do not contribute to the
222 concurrency level. In other words, runnable CPU intensive
223 work items will not prevent other work items in the same
224 thread-pool from starting execution. This is useful for bound
225 work items which are expected to hog CPU cycles so that their
226 execution is regulated by the system scheduler.
228 Although CPU intensive work items don't contribute to the
229 concurrency level, start of their executions is still
230 regulated by the concurrency management and runnable
231 non-CPU-intensive work items can delay execution of CPU
232 intensive work items.
234 This flag is meaningless for unbound wq.
238 @max_active determines the maximum number of execution contexts per
239 CPU which can be assigned to the work items of a wq. For example,
240 with @max_active of 16, at most 16 work items of the wq can be
241 executing at the same time per CPU.
243 Currently, for a bound wq, the maximum limit for @max_active is 512
244 and the default value used when 0 is specified is 256. For an unbound
245 wq, the limit is higher of 512 and 4 * num_possible_cpus(). These
246 values are chosen sufficiently high such that they are not the
247 limiting factor while providing protection in runaway cases.
249 The number of active work items of a wq is usually regulated by the
250 users of the wq, more specifically, by how many work items the users
251 may queue at the same time. Unless there is a specific need for
252 throttling the number of active work items, specifying '0' is
255 Some users depend on the strict execution ordering of ST wq. The
256 combination of @max_active of 1 and WQ_UNBOUND is used to achieve this
257 behavior. Work items on such wq are always queued to the unbound gcwq
258 and only one work item can be active at any given time thus achieving
259 the same ordering property as ST wq.
262 5. Example Execution Scenarios
264 The following example execution scenarios try to illustrate how cmwq
265 behave under different configurations.
267 Work items w0, w1, w2 are queued to a bound wq q0 on the same CPU.
268 w0 burns CPU for 5ms then sleeps for 10ms then burns CPU for 5ms
269 again before finishing. w1 and w2 burn CPU for 5ms then sleep for
272 Ignoring all other tasks, works and processing overhead, and assuming
273 simple FIFO scheduling, the following is one highly simplified version
274 of possible sequences of events with the original wq.
277 0 w0 starts and burns CPU
279 15 w0 wakes up and burns CPU
281 20 w1 starts and burns CPU
283 35 w1 wakes up and finishes
284 35 w2 starts and burns CPU
286 50 w2 wakes up and finishes
288 And with cmwq with @max_active >= 3,
291 0 w0 starts and burns CPU
293 5 w1 starts and burns CPU
295 10 w2 starts and burns CPU
297 15 w0 wakes up and burns CPU
299 20 w1 wakes up and finishes
300 25 w2 wakes up and finishes
305 0 w0 starts and burns CPU
307 5 w1 starts and burns CPU
309 15 w0 wakes up and burns CPU
311 20 w1 wakes up and finishes
312 20 w2 starts and burns CPU
314 35 w2 wakes up and finishes
316 Now, let's assume w1 and w2 are queued to a different wq q1 which has
317 WQ_CPU_INTENSIVE set,
320 0 w0 starts and burns CPU
322 5 w1 and w2 start and burn CPU
325 15 w0 wakes up and burns CPU
327 20 w1 wakes up and finishes
328 25 w2 wakes up and finishes
333 * Do not forget to use WQ_MEM_RECLAIM if a wq may process work items
334 which are used during memory reclaim. Each wq with WQ_MEM_RECLAIM
335 set has an execution context reserved for it. If there is
336 dependency among multiple work items used during memory reclaim,
337 they should be queued to separate wq each with WQ_MEM_RECLAIM.
339 * Unless strict ordering is required, there is no need to use ST wq.
341 * Unless there is a specific need, using 0 for @max_active is
342 recommended. In most use cases, concurrency level usually stays
343 well under the default limit.
345 * A wq serves as a domain for forward progress guarantee
346 (WQ_MEM_RECLAIM, flush and work item attributes. Work items which
347 are not involved in memory reclaim and don't need to be flushed as a
348 part of a group of work items, and don't require any special
349 attribute, can use one of the system wq. There is no difference in
350 execution characteristics between using a dedicated wq and a system
353 * Unless work items are expected to consume a huge amount of CPU
354 cycles, using a bound wq is usually beneficial due to the increased
355 level of locality in wq operations and work item execution.
360 Because the work functions are executed by generic worker threads
361 there are a few tricks needed to shed some light on misbehaving
364 Worker threads show up in the process list as:
366 root 5671 0.0 0.0 0 0 ? S 12:07 0:00 [kworker/0:1]
367 root 5672 0.0 0.0 0 0 ? S 12:07 0:00 [kworker/1:2]
368 root 5673 0.0 0.0 0 0 ? S 12:12 0:00 [kworker/0:0]
369 root 5674 0.0 0.0 0 0 ? S 12:13 0:00 [kworker/1:0]
371 If kworkers are going crazy (using too much cpu), there are two types
372 of possible problems:
374 1. Something beeing scheduled in rapid succession
375 2. A single work item that consumes lots of cpu cycles
377 The first one can be tracked using tracing:
379 $ echo workqueue:workqueue_queue_work > /sys/kernel/debug/tracing/set_event
380 $ cat /sys/kernel/debug/tracing/trace_pipe > out.txt
384 If something is busy looping on work queueing, it would be dominating
385 the output and the offender can be determined with the work item
388 For the second type of problems it should be possible to just check
389 the stack trace of the offending worker thread.
391 $ cat /proc/THE_OFFENDING_KWORKER/stack
393 The work item's function should be trivially visible in the stack