workqueue.txt 15 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388
  1. Concurrency Managed Workqueue (cmwq)
  2. September, 2010 Tejun Heo <tj@kernel.org>
  3. Florian Mickler <florian@mickler.org>
  4. CONTENTS
  5. 1. Introduction
  6. 2. Why cmwq?
  7. 3. The Design
  8. 4. Application Programming Interface (API)
  9. 5. Example Execution Scenarios
  10. 6. Guidelines
  11. 7. Debugging
  12. 1. Introduction
  13. There are many cases where an asynchronous process execution context
  14. is needed and the workqueue (wq) API is the most commonly used
  15. mechanism for such cases.
  16. When such an asynchronous execution context is needed, a work item
  17. describing which function to execute is put on a queue. An
  18. independent thread serves as the asynchronous execution context. The
  19. queue is called workqueue and the thread is called worker.
  20. While there are work items on the workqueue the worker executes the
  21. functions associated with the work items one after the other. When
  22. there is no work item left on the workqueue the worker becomes idle.
  23. When a new work item gets queued, the worker begins executing again.
  24. 2. Why cmwq?
  25. In the original wq implementation, a multi threaded (MT) wq had one
  26. worker thread per CPU and a single threaded (ST) wq had one worker
  27. thread system-wide. A single MT wq needed to keep around the same
  28. number of workers as the number of CPUs. The kernel grew a lot of MT
  29. wq users over the years and with the number of CPU cores continuously
  30. rising, some systems saturated the default 32k PID space just booting
  31. up.
  32. Although MT wq wasted a lot of resource, the level of concurrency
  33. provided was unsatisfactory. The limitation was common to both ST and
  34. MT wq albeit less severe on MT. Each wq maintained its own separate
  35. worker pool. A MT wq could provide only one execution context per CPU
  36. while a ST wq one for the whole system. Work items had to compete for
  37. those very limited execution contexts leading to various problems
  38. including proneness to deadlocks around the single execution context.
  39. The tension between the provided level of concurrency and resource
  40. usage also forced its users to make unnecessary tradeoffs like libata
  41. choosing to use ST wq for polling PIOs and accepting an unnecessary
  42. limitation that no two polling PIOs can progress at the same time. As
  43. MT wq don't provide much better concurrency, users which require
  44. higher level of concurrency, like async or fscache, had to implement
  45. their own thread pool.
  46. Concurrency Managed Workqueue (cmwq) is a reimplementation of wq with
  47. focus on the following goals.
  48. * Maintain compatibility with the original workqueue API.
  49. * Use per-CPU unified worker pools shared by all wq to provide
  50. flexible level of concurrency on demand without wasting a lot of
  51. resource.
  52. * Automatically regulate worker pool and level of concurrency so that
  53. the API users don't need to worry about such details.
  54. 3. The Design
  55. In order to ease the asynchronous execution of functions a new
  56. abstraction, the work item, is introduced.
  57. A work item is a simple struct that holds a pointer to the function
  58. that is to be executed asynchronously. Whenever a driver or subsystem
  59. wants a function to be executed asynchronously it has to set up a work
  60. item pointing to that function and queue that work item on a
  61. workqueue.
  62. Special purpose threads, called worker threads, execute the functions
  63. off of the queue, one after the other. If no work is queued, the
  64. worker threads become idle. These worker threads are managed in so
  65. called worker-pools.
  66. The cmwq design differentiates between the user-facing workqueues that
  67. subsystems and drivers queue work items on and the backend mechanism
  68. which manages worker-pools and processes the queued work items.
  69. There are two worker-pools, one for normal work items and the other
  70. for high priority ones, for each possible CPU and some extra
  71. worker-pools to serve work items queued on unbound workqueues - the
  72. number of these backing pools is dynamic.
  73. Subsystems and drivers can create and queue work items through special
  74. workqueue API functions as they see fit. They can influence some
  75. aspects of the way the work items are executed by setting flags on the
  76. workqueue they are putting the work item on. These flags include
  77. things like CPU locality, concurrency limits, priority and more. To
  78. get a detailed overview refer to the API description of
  79. alloc_workqueue() below.
  80. When a work item is queued to a workqueue, the target worker-pool is
  81. determined according to the queue parameters and workqueue attributes
  82. and appended on the shared worklist of the worker-pool. For example,
  83. unless specifically overridden, a work item of a bound workqueue will
  84. be queued on the worklist of either normal or highpri worker-pool that
  85. is associated to the CPU the issuer is running on.
  86. For any worker pool implementation, managing the concurrency level
  87. (how many execution contexts are active) is an important issue. cmwq
  88. tries to keep the concurrency at a minimal but sufficient level.
  89. Minimal to save resources and sufficient in that the system is used at
  90. its full capacity.
  91. Each worker-pool bound to an actual CPU implements concurrency
  92. management by hooking into the scheduler. The worker-pool is notified
  93. whenever an active worker wakes up or sleeps and keeps track of the
  94. number of the currently runnable workers. Generally, work items are
  95. not expected to hog a CPU and consume many cycles. That means
  96. maintaining just enough concurrency to prevent work processing from
  97. stalling should be optimal. As long as there are one or more runnable
  98. workers on the CPU, the worker-pool doesn't start execution of a new
  99. work, but, when the last running worker goes to sleep, it immediately
  100. schedules a new worker so that the CPU doesn't sit idle while there
  101. are pending work items. This allows using a minimal number of workers
  102. without losing execution bandwidth.
  103. Keeping idle workers around doesn't cost other than the memory space
  104. for kthreads, so cmwq holds onto idle ones for a while before killing
  105. them.
  106. For unbound workqueues, the number of backing pools is dynamic.
  107. Unbound workqueue can be assigned custom attributes using
  108. apply_workqueue_attrs() and workqueue will automatically create
  109. backing worker pools matching the attributes. The responsibility of
  110. regulating concurrency level is on the users. There is also a flag to
  111. mark a bound wq to ignore the concurrency management. Please refer to
  112. the API section for details.
  113. Forward progress guarantee relies on that workers can be created when
  114. more execution contexts are necessary, which in turn is guaranteed
  115. through the use of rescue workers. All work items which might be used
  116. on code paths that handle memory reclaim are required to be queued on
  117. wq's that have a rescue-worker reserved for execution under memory
  118. pressure. Else it is possible that the worker-pool deadlocks waiting
  119. for execution contexts to free up.
  120. 4. Application Programming Interface (API)
  121. alloc_workqueue() allocates a wq. The original create_*workqueue()
  122. functions are deprecated and scheduled for removal. alloc_workqueue()
  123. takes three arguments - @name, @flags and @max_active. @name is the
  124. name of the wq and also used as the name of the rescuer thread if
  125. there is one.
  126. A wq no longer manages execution resources but serves as a domain for
  127. forward progress guarantee, flush and work item attributes. @flags
  128. and @max_active control how work items are assigned execution
  129. resources, scheduled and executed.
  130. @flags:
  131. WQ_UNBOUND
  132. Work items queued to an unbound wq are served by the special
  133. woker-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 maintain 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 workqueues
  174. are now non-reentrant - any work item is guaranteed to be executed by
  175. at most one worker system-wide at any given time.
  176. @max_active:
  177. @max_active determines the maximum number of execution contexts per
  178. CPU which can be assigned to the work items of a wq. For example,
  179. with @max_active of 16, at most 16 work items of the wq can be
  180. executing at the same time per CPU.
  181. Currently, for a bound wq, the maximum limit for @max_active is 512
  182. and the default value used when 0 is specified is 256. For an unbound
  183. wq, the limit is higher of 512 and 4 * num_possible_cpus(). These
  184. values are chosen sufficiently high such that they are not the
  185. limiting factor while providing protection in runaway cases.
  186. The number of active work items of a wq is usually regulated by the
  187. users of the wq, more specifically, by how many work items the users
  188. may queue at the same time. Unless there is a specific need for
  189. throttling the number of active work items, specifying '0' is
  190. recommended.
  191. Some users depend on the strict execution ordering of ST wq. The
  192. combination of @max_active of 1 and WQ_UNBOUND is used to achieve this
  193. behavior. Work items on such wq are always queued to the unbound
  194. worker-pools and only one work item can be active at any given time thus
  195. achieving the same ordering property as ST wq.
  196. 5. Example Execution Scenarios
  197. The following example execution scenarios try to illustrate how cmwq
  198. behave under different configurations.
  199. Work items w0, w1, w2 are queued to a bound wq q0 on the same CPU.
  200. w0 burns CPU for 5ms then sleeps for 10ms then burns CPU for 5ms
  201. again before finishing. w1 and w2 burn CPU for 5ms then sleep for
  202. 10ms.
  203. Ignoring all other tasks, works and processing overhead, and assuming
  204. simple FIFO scheduling, the following is one highly simplified version
  205. of possible sequences of events with the original wq.
  206. TIME IN MSECS EVENT
  207. 0 w0 starts and burns CPU
  208. 5 w0 sleeps
  209. 15 w0 wakes up and burns CPU
  210. 20 w0 finishes
  211. 20 w1 starts and burns CPU
  212. 25 w1 sleeps
  213. 35 w1 wakes up and finishes
  214. 35 w2 starts and burns CPU
  215. 40 w2 sleeps
  216. 50 w2 wakes up and finishes
  217. And with cmwq with @max_active >= 3,
  218. TIME IN MSECS EVENT
  219. 0 w0 starts and burns CPU
  220. 5 w0 sleeps
  221. 5 w1 starts and burns CPU
  222. 10 w1 sleeps
  223. 10 w2 starts and burns CPU
  224. 15 w2 sleeps
  225. 15 w0 wakes up and burns CPU
  226. 20 w0 finishes
  227. 20 w1 wakes up and finishes
  228. 25 w2 wakes up and finishes
  229. If @max_active == 2,
  230. TIME IN MSECS EVENT
  231. 0 w0 starts and burns CPU
  232. 5 w0 sleeps
  233. 5 w1 starts and burns CPU
  234. 10 w1 sleeps
  235. 15 w0 wakes up and burns CPU
  236. 20 w0 finishes
  237. 20 w1 wakes up and finishes
  238. 20 w2 starts and burns CPU
  239. 25 w2 sleeps
  240. 35 w2 wakes up and finishes
  241. Now, let's assume w1 and w2 are queued to a different wq q1 which has
  242. WQ_CPU_INTENSIVE set,
  243. TIME IN MSECS EVENT
  244. 0 w0 starts and burns CPU
  245. 5 w0 sleeps
  246. 5 w1 and w2 start and burn CPU
  247. 10 w1 sleeps
  248. 15 w2 sleeps
  249. 15 w0 wakes up and burns CPU
  250. 20 w0 finishes
  251. 20 w1 wakes up and finishes
  252. 25 w2 wakes up and finishes
  253. 6. Guidelines
  254. * Do not forget to use WQ_MEM_RECLAIM if a wq may process work items
  255. which are used during memory reclaim. Each wq with WQ_MEM_RECLAIM
  256. set has an execution context reserved for it. If there is
  257. dependency among multiple work items used during memory reclaim,
  258. they should be queued to separate wq each with WQ_MEM_RECLAIM.
  259. * Unless strict ordering is required, there is no need to use ST wq.
  260. * Unless there is a specific need, using 0 for @max_active is
  261. recommended. In most use cases, concurrency level usually stays
  262. well under the default limit.
  263. * A wq serves as a domain for forward progress guarantee
  264. (WQ_MEM_RECLAIM, flush and work item attributes. Work items which
  265. are not involved in memory reclaim and don't need to be flushed as a
  266. part of a group of work items, and don't require any special
  267. attribute, can use one of the system wq. There is no difference in
  268. execution characteristics between using a dedicated wq and a system
  269. wq.
  270. * Unless work items are expected to consume a huge amount of CPU
  271. cycles, using a bound wq is usually beneficial due to the increased
  272. level of locality in wq operations and work item execution.
  273. 7. Debugging
  274. Because the work functions are executed by generic worker threads
  275. there are a few tricks needed to shed some light on misbehaving
  276. workqueue users.
  277. Worker threads show up in the process list as:
  278. root 5671 0.0 0.0 0 0 ? S 12:07 0:00 [kworker/0:1]
  279. root 5672 0.0 0.0 0 0 ? S 12:07 0:00 [kworker/1:2]
  280. root 5673 0.0 0.0 0 0 ? S 12:12 0:00 [kworker/0:0]
  281. root 5674 0.0 0.0 0 0 ? S 12:13 0:00 [kworker/1:0]
  282. If kworkers are going crazy (using too much cpu), there are two types
  283. of possible problems:
  284. 1. Something being scheduled in rapid succession
  285. 2. A single work item that consumes lots of cpu cycles
  286. The first one can be tracked using tracing:
  287. $ echo workqueue:workqueue_queue_work > /sys/kernel/debug/tracing/set_event
  288. $ cat /sys/kernel/debug/tracing/trace_pipe > out.txt
  289. (wait a few secs)
  290. ^C
  291. If something is busy looping on work queueing, it would be dominating
  292. the output and the offender can be determined with the work item
  293. function.
  294. For the second type of problems it should be possible to just check
  295. the stack trace of the offending worker thread.
  296. $ cat /proc/THE_OFFENDING_KWORKER/stack
  297. The work item's function should be trivially visible in the stack
  298. trace.