workqueue.rst 15 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400
  1. ====================================
  2. Concurrency Managed Workqueue (cmwq)
  3. ====================================
  4. :Date: September, 2010
  5. :Author: Tejun Heo <tj@kernel.org>
  6. :Author: Florian Mickler <florian@mickler.org>
  7. Introduction
  8. ============
  9. There are many cases where an asynchronous process execution context
  10. is needed and the workqueue (wq) API is the most commonly used
  11. mechanism for such cases.
  12. When such an asynchronous execution context is needed, a work item
  13. describing which function to execute is put on a queue. An
  14. independent thread serves as the asynchronous execution context. The
  15. queue is called workqueue and the thread is called worker.
  16. While there are work items on the workqueue the worker executes the
  17. functions associated with the work items one after the other. When
  18. there is no work item left on the workqueue the worker becomes idle.
  19. When a new work item gets queued, the worker begins executing again.
  20. Why cmwq?
  21. =========
  22. In the original wq implementation, a multi threaded (MT) wq had one
  23. worker thread per CPU and a single threaded (ST) wq had one worker
  24. thread system-wide. A single MT wq needed to keep around the same
  25. number of workers as the number of CPUs. The kernel grew a lot of MT
  26. wq users over the years and with the number of CPU cores continuously
  27. rising, some systems saturated the default 32k PID space just booting
  28. up.
  29. Although MT wq wasted a lot of resource, the level of concurrency
  30. provided was unsatisfactory. The limitation was common to both ST and
  31. MT wq albeit less severe on MT. Each wq maintained its own separate
  32. worker pool. An MT wq could provide only one execution context per CPU
  33. while an ST wq one for the whole system. Work items had to compete for
  34. those very limited execution contexts leading to various problems
  35. including proneness to deadlocks around the single execution context.
  36. The tension between the provided level of concurrency and resource
  37. usage also forced its users to make unnecessary tradeoffs like libata
  38. choosing to use ST wq for polling PIOs and accepting an unnecessary
  39. limitation that no two polling PIOs can progress at the same time. As
  40. MT wq don't provide much better concurrency, users which require
  41. higher level of concurrency, like async or fscache, had to implement
  42. their own thread pool.
  43. Concurrency Managed Workqueue (cmwq) is a reimplementation of wq with
  44. focus on the following goals.
  45. * Maintain compatibility with the original workqueue API.
  46. * Use per-CPU unified worker pools shared by all wq to provide
  47. flexible level of concurrency on demand without wasting a lot of
  48. resource.
  49. * Automatically regulate worker pool and level of concurrency so that
  50. the API users don't need to worry about such details.
  51. The Design
  52. ==========
  53. In order to ease the asynchronous execution of functions a new
  54. abstraction, the work item, is introduced.
  55. A work item is a simple struct that holds a pointer to the function
  56. that is to be executed asynchronously. Whenever a driver or subsystem
  57. wants a function to be executed asynchronously it has to set up a work
  58. item pointing to that function and queue that work item on a
  59. workqueue.
  60. Special purpose threads, called worker threads, execute the functions
  61. off of the queue, one after the other. If no work is queued, the
  62. worker threads become idle. These worker threads are managed in so
  63. called worker-pools.
  64. The cmwq design differentiates between the user-facing workqueues that
  65. subsystems and drivers queue work items on and the backend mechanism
  66. which manages worker-pools and processes the queued work items.
  67. There are two worker-pools, one for normal work items and the other
  68. for high priority ones, for each possible CPU and some extra
  69. worker-pools to serve work items queued on unbound workqueues - the
  70. number of these backing pools is dynamic.
  71. Subsystems and drivers can create and queue work items through special
  72. workqueue API functions as they see fit. They can influence some
  73. aspects of the way the work items are executed by setting flags on the
  74. workqueue they are putting the work item on. These flags include
  75. things like CPU locality, concurrency limits, priority and more. To
  76. get a detailed overview refer to the API description of
  77. ``alloc_workqueue()`` below.
  78. When a work item is queued to a workqueue, the target worker-pool is
  79. determined according to the queue parameters and workqueue attributes
  80. and appended on the shared worklist of the worker-pool. For example,
  81. unless specifically overridden, a work item of a bound workqueue will
  82. be queued on the worklist of either normal or highpri worker-pool that
  83. is associated to the CPU the issuer is running on.
  84. For any worker pool implementation, managing the concurrency level
  85. (how many execution contexts are active) is an important issue. cmwq
  86. tries to keep the concurrency at a minimal but sufficient level.
  87. Minimal to save resources and sufficient in that the system is used at
  88. its full capacity.
  89. Each worker-pool bound to an actual CPU implements concurrency
  90. management by hooking into the scheduler. The worker-pool is notified
  91. whenever an active worker wakes up or sleeps and keeps track of the
  92. number of the currently runnable workers. Generally, work items are
  93. not expected to hog a CPU and consume many cycles. That means
  94. maintaining just enough concurrency to prevent work processing from
  95. stalling should be optimal. As long as there are one or more runnable
  96. workers on the CPU, the worker-pool doesn't start execution of a new
  97. work, but, when the last running worker goes to sleep, it immediately
  98. schedules a new worker so that the CPU doesn't sit idle while there
  99. are pending work items. This allows using a minimal number of workers
  100. without losing execution bandwidth.
  101. Keeping idle workers around doesn't cost other than the memory space
  102. for kthreads, so cmwq holds onto idle ones for a while before killing
  103. them.
  104. For unbound workqueues, the number of backing pools is dynamic.
  105. Unbound workqueue can be assigned custom attributes using
  106. ``apply_workqueue_attrs()`` and workqueue will automatically create
  107. backing worker pools matching the attributes. The responsibility of
  108. regulating concurrency level is on the users. There is also a flag to
  109. mark a bound wq to ignore the concurrency management. Please refer to
  110. the API section for details.
  111. Forward progress guarantee relies on that workers can be created when
  112. more execution contexts are necessary, which in turn is guaranteed
  113. through the use of rescue workers. All work items which might be used
  114. on code paths that handle memory reclaim are required to be queued on
  115. wq's that have a rescue-worker reserved for execution under memory
  116. pressure. Else it is possible that the worker-pool deadlocks waiting
  117. for execution contexts to free up.
  118. Application Programming Interface (API)
  119. =======================================
  120. ``alloc_workqueue()`` allocates a wq. The original
  121. ``create_*workqueue()`` functions are deprecated and scheduled for
  122. removal. ``alloc_workqueue()`` takes three arguments - ``@name``,
  123. ``@flags`` and ``@max_active``. ``@name`` is the name of the wq and
  124. also used as the name of the rescuer thread if there is one.
  125. A wq no longer manages execution resources but serves as a domain for
  126. forward progress guarantee, flush and work item attributes. ``@flags``
  127. and ``@max_active`` control how work items are assigned execution
  128. resources, scheduled and executed.
  129. ``flags``
  130. ---------
  131. ``WQ_UNBOUND``
  132. Work items queued to an unbound wq are served by the special
  133. worker-pools which host workers which are not bound to any
  134. specific CPU. This makes the wq behave as a simple execution
  135. context provider without concurrency management. The unbound
  136. worker-pools try to start execution of work items as soon as
  137. possible. Unbound wq sacrifices locality but is useful for
  138. the following cases.
  139. * Wide fluctuation in the concurrency level requirement is
  140. expected and using bound wq may end up creating large number
  141. of mostly unused workers across different CPUs as the issuer
  142. hops through different CPUs.
  143. * Long running CPU intensive workloads which can be better
  144. managed by the system scheduler.
  145. ``WQ_FREEZABLE``
  146. A freezable wq participates in the freeze phase of the system
  147. suspend operations. Work items on the wq are drained and no
  148. new work item starts execution until thawed.
  149. ``WQ_MEM_RECLAIM``
  150. All wq which might be used in the memory reclaim paths **MUST**
  151. have this flag set. The wq is guaranteed to have at least one
  152. execution context regardless of memory pressure.
  153. ``WQ_HIGHPRI``
  154. Work items of a highpri wq are queued to the highpri
  155. worker-pool of the target cpu. Highpri worker-pools are
  156. served by worker threads with elevated nice level.
  157. Note that normal and highpri worker-pools don't interact with
  158. each other. Each maintains its separate pool of workers and
  159. implements concurrency management among its workers.
  160. ``WQ_CPU_INTENSIVE``
  161. Work items of a CPU intensive wq do not contribute to the
  162. concurrency level. In other words, runnable CPU intensive
  163. work items will not prevent other work items in the same
  164. worker-pool from starting execution. This is useful for bound
  165. work items which are expected to hog CPU cycles so that their
  166. execution is regulated by the system scheduler.
  167. Although CPU intensive work items don't contribute to the
  168. concurrency level, start of their executions is still
  169. regulated by the concurrency management and runnable
  170. non-CPU-intensive work items can delay execution of CPU
  171. intensive work items.
  172. This flag is meaningless for unbound wq.
  173. Note that the flag ``WQ_NON_REENTRANT`` no longer exists as all
  174. workqueues are now non-reentrant - any work item is guaranteed to be
  175. executed by at most one worker system-wide at any given time.
  176. ``max_active``
  177. --------------
  178. ``@max_active`` determines the maximum number of execution contexts
  179. per CPU which can be assigned to the work items of a wq. For example,
  180. with ``@max_active`` of 16, at most 16 work items of the wq can be
  181. executing at the same time per CPU.
  182. Currently, for a bound wq, the maximum limit for ``@max_active`` is
  183. 512 and the default value used when 0 is specified is 256. For an
  184. unbound wq, the limit is higher of 512 and 4 *
  185. ``num_possible_cpus()``. These values are chosen sufficiently high
  186. such that they are not the limiting factor while providing protection
  187. in runaway cases.
  188. The number of active work items of a wq is usually regulated by the
  189. users of the wq, more specifically, by how many work items the users
  190. may queue at the same time. Unless there is a specific need for
  191. throttling the number of active work items, specifying '0' is
  192. recommended.
  193. Some users depend on the strict execution ordering of ST wq. The
  194. combination of ``@max_active`` of 1 and ``WQ_UNBOUND`` used to
  195. achieve this behavior. Work items on such wq were always queued to the
  196. unbound worker-pools and only one work item could be active at any given
  197. time thus achieving the same ordering property as ST wq.
  198. In the current implementation the above configuration only guarantees
  199. ST behavior within a given NUMA node. Instead ``alloc_ordered_queue()`` should
  200. be used to achieve system-wide ST behavior.
  201. Example Execution Scenarios
  202. ===========================
  203. The following example execution scenarios try to illustrate how cmwq
  204. behave under different configurations.
  205. Work items w0, w1, w2 are queued to a bound wq q0 on the same CPU.
  206. w0 burns CPU for 5ms then sleeps for 10ms then burns CPU for 5ms
  207. again before finishing. w1 and w2 burn CPU for 5ms then sleep for
  208. 10ms.
  209. Ignoring all other tasks, works and processing overhead, and assuming
  210. simple FIFO scheduling, the following is one highly simplified version
  211. of possible sequences of events with the original wq. ::
  212. TIME IN MSECS EVENT
  213. 0 w0 starts and burns CPU
  214. 5 w0 sleeps
  215. 15 w0 wakes up and burns CPU
  216. 20 w0 finishes
  217. 20 w1 starts and burns CPU
  218. 25 w1 sleeps
  219. 35 w1 wakes up and finishes
  220. 35 w2 starts and burns CPU
  221. 40 w2 sleeps
  222. 50 w2 wakes up and finishes
  223. And with cmwq with ``@max_active`` >= 3, ::
  224. TIME IN MSECS EVENT
  225. 0 w0 starts and burns CPU
  226. 5 w0 sleeps
  227. 5 w1 starts and burns CPU
  228. 10 w1 sleeps
  229. 10 w2 starts and burns CPU
  230. 15 w2 sleeps
  231. 15 w0 wakes up and burns CPU
  232. 20 w0 finishes
  233. 20 w1 wakes up and finishes
  234. 25 w2 wakes up and finishes
  235. If ``@max_active`` == 2, ::
  236. TIME IN MSECS EVENT
  237. 0 w0 starts and burns CPU
  238. 5 w0 sleeps
  239. 5 w1 starts and burns CPU
  240. 10 w1 sleeps
  241. 15 w0 wakes up and burns CPU
  242. 20 w0 finishes
  243. 20 w1 wakes up and finishes
  244. 20 w2 starts and burns CPU
  245. 25 w2 sleeps
  246. 35 w2 wakes up and finishes
  247. Now, let's assume w1 and w2 are queued to a different wq q1 which has
  248. ``WQ_CPU_INTENSIVE`` set, ::
  249. TIME IN MSECS EVENT
  250. 0 w0 starts and burns CPU
  251. 5 w0 sleeps
  252. 5 w1 and w2 start and burn CPU
  253. 10 w1 sleeps
  254. 15 w2 sleeps
  255. 15 w0 wakes up and burns CPU
  256. 20 w0 finishes
  257. 20 w1 wakes up and finishes
  258. 25 w2 wakes up and finishes
  259. Guidelines
  260. ==========
  261. * Do not forget to use ``WQ_MEM_RECLAIM`` if a wq may process work
  262. items which are used during memory reclaim. Each wq with
  263. ``WQ_MEM_RECLAIM`` set has an execution context reserved for it. If
  264. there is dependency among multiple work items used during memory
  265. reclaim, they should be queued to separate wq each with
  266. ``WQ_MEM_RECLAIM``.
  267. * Unless strict ordering is required, there is no need to use ST wq.
  268. * Unless there is a specific need, using 0 for @max_active is
  269. recommended. In most use cases, concurrency level usually stays
  270. well under the default limit.
  271. * A wq serves as a domain for forward progress guarantee
  272. (``WQ_MEM_RECLAIM``, flush and work item attributes. Work items
  273. which are not involved in memory reclaim and don't need to be
  274. flushed as a part of a group of work items, and don't require any
  275. special attribute, can use one of the system wq. There is no
  276. difference in execution characteristics between using a dedicated wq
  277. and a system wq.
  278. * Unless work items are expected to consume a huge amount of CPU
  279. cycles, using a bound wq is usually beneficial due to the increased
  280. level of locality in wq operations and work item execution.
  281. Debugging
  282. =========
  283. Because the work functions are executed by generic worker threads
  284. there are a few tricks needed to shed some light on misbehaving
  285. workqueue users.
  286. Worker threads show up in the process list as: ::
  287. root 5671 0.0 0.0 0 0 ? S 12:07 0:00 [kworker/0:1]
  288. root 5672 0.0 0.0 0 0 ? S 12:07 0:00 [kworker/1:2]
  289. root 5673 0.0 0.0 0 0 ? S 12:12 0:00 [kworker/0:0]
  290. root 5674 0.0 0.0 0 0 ? S 12:13 0:00 [kworker/1:0]
  291. If kworkers are going crazy (using too much cpu), there are two types
  292. of possible problems:
  293. 1. Something being scheduled in rapid succession
  294. 2. A single work item that consumes lots of cpu cycles
  295. The first one can be tracked using tracing: ::
  296. $ echo workqueue:workqueue_queue_work > /sys/kernel/debug/tracing/set_event
  297. $ cat /sys/kernel/debug/tracing/trace_pipe > out.txt
  298. (wait a few secs)
  299. ^C
  300. If something is busy looping on work queueing, it would be dominating
  301. the output and the offender can be determined with the work item
  302. function.
  303. For the second type of problems it should be possible to just check
  304. the stack trace of the offending worker thread. ::
  305. $ cat /proc/THE_OFFENDING_KWORKER/stack
  306. The work item's function should be trivially visible in the stack
  307. trace.
  308. Kernel Inline Documentations Reference
  309. ======================================
  310. .. kernel-doc:: include/linux/workqueue.h
  311. .. kernel-doc:: kernel/workqueue.c