1 2Concurrency Managed Workqueue (cmwq) 3 4September, 2010 Tejun Heo <tj@kernel.org> 5 Florian Mickler <florian@mickler.org> 6 7CONTENTS 8 91. Introduction 102. Why cmwq? 113. The Design 124. Application Programming Interface (API) 135. Example Execution Scenarios 146. Guidelines 157. Debugging 16 17 181. Introduction 19 20There are many cases where an asynchronous process execution context 21is needed and the workqueue (wq) API is the most commonly used 22mechanism for such cases. 23 24When such an asynchronous execution context is needed, a work item 25describing which function to execute is put on a queue. An 26independent thread serves as the asynchronous execution context. The 27queue is called workqueue and the thread is called worker. 28 29While there are work items on the workqueue the worker executes the 30functions associated with the work items one after the other. When 31there is no work item left on the workqueue the worker becomes idle. 32When a new work item gets queued, the worker begins executing again. 33 34 352. Why cmwq? 36 37In the original wq implementation, a multi threaded (MT) wq had one 38worker thread per CPU and a single threaded (ST) wq had one worker 39thread system-wide. A single MT wq needed to keep around the same 40number of workers as the number of CPUs. The kernel grew a lot of MT 41wq users over the years and with the number of CPU cores continuously 42rising, some systems saturated the default 32k PID space just booting 43up. 44 45Although MT wq wasted a lot of resource, the level of concurrency 46provided was unsatisfactory. The limitation was common to both ST and 47MT wq albeit less severe on MT. Each wq maintained its own separate 48worker pool. A MT wq could provide only one execution context per CPU 49while a ST wq one for the whole system. Work items had to compete for 50those very limited execution contexts leading to various problems 51including proneness to deadlocks around the single execution context. 52 53The tension between the provided level of concurrency and resource 54usage also forced its users to make unnecessary tradeoffs like libata 55choosing to use ST wq for polling PIOs and accepting an unnecessary 56limitation that no two polling PIOs can progress at the same time. As 57MT wq don't provide much better concurrency, users which require 58higher level of concurrency, like async or fscache, had to implement 59their own thread pool. 60 61Concurrency Managed Workqueue (cmwq) is a reimplementation of wq with 62focus on the following goals. 63 64* Maintain compatibility with the original workqueue API. 65 66* Use per-CPU unified worker pools shared by all wq to provide 67 flexible level of concurrency on demand without wasting a lot of 68 resource. 69 70* Automatically regulate worker pool and level of concurrency so that 71 the API users don't need to worry about such details. 72 73 743. The Design 75 76In order to ease the asynchronous execution of functions a new 77abstraction, the work item, is introduced. 78 79A work item is a simple struct that holds a pointer to the function 80that is to be executed asynchronously. Whenever a driver or subsystem 81wants a function to be executed asynchronously it has to set up a work 82item pointing to that function and queue that work item on a 83workqueue. 84 85Special purpose threads, called worker threads, execute the functions 86off of the queue, one after the other. If no work is queued, the 87worker threads become idle. These worker threads are managed in so 88called thread-pools. 89 90The cmwq design differentiates between the user-facing workqueues that 91subsystems and drivers queue work items on and the backend mechanism 92which manages thread-pool and processes the queued work items. 93 94The backend is called gcwq. There is one gcwq for each possible CPU 95and one gcwq to serve work items queued on unbound workqueues. 96 97Subsystems and drivers can create and queue work items through special 98workqueue API functions as they see fit. They can influence some 99aspects of the way the work items are executed by setting flags on the 100workqueue they are putting the work item on. These flags include 101things like CPU locality, reentrancy, concurrency limits and more. To 102get a detailed overview refer to the API description of 103alloc_workqueue() below. 104 105When a work item is queued to a workqueue, the target gcwq is 106determined according to the queue parameters and workqueue attributes 107and appended on the shared worklist of the gcwq. For example, unless 108specifically overridden, a work item of a bound workqueue will be 109queued on the worklist of exactly that gcwq that is associated to the 110CPU the issuer is running on. 111 112For any worker pool implementation, managing the concurrency level 113(how many execution contexts are active) is an important issue. cmwq 114tries to keep the concurrency at a minimal but sufficient level. 115Minimal to save resources and sufficient in that the system is used at 116its full capacity. 117 118Each gcwq bound to an actual CPU implements concurrency management by 119hooking into the scheduler. The gcwq is notified whenever an active 120worker wakes up or sleeps and keeps track of the number of the 121currently runnable workers. Generally, work items are not expected to 122hog a CPU and consume many cycles. That means maintaining just enough 123concurrency to prevent work processing from stalling should be 124optimal. As long as there are one or more runnable workers on the 125CPU, the gcwq doesn't start execution of a new work, but, when the 126last running worker goes to sleep, it immediately schedules a new 127worker so that the CPU doesn't sit idle while there are pending work 128items. This allows using a minimal number of workers without losing 129execution bandwidth. 130 131Keeping idle workers around doesn't cost other than the memory space 132for kthreads, so cmwq holds onto idle ones for a while before killing 133them. 134 135For an unbound wq, the above concurrency management doesn't apply and 136the gcwq for the pseudo unbound CPU tries to start executing all work 137items as soon as possible. The responsibility of regulating 138concurrency level is on the users. There is also a flag to mark a 139bound wq to ignore the concurrency management. Please refer to the 140API section for details. 141 142Forward progress guarantee relies on that workers can be created when 143more execution contexts are necessary, which in turn is guaranteed 144through the use of rescue workers. All work items which might be used 145on code paths that handle memory reclaim are required to be queued on 146wq's that have a rescue-worker reserved for execution under memory 147pressure. Else it is possible that the thread-pool deadlocks waiting 148for execution contexts to free up. 149 150 1514. Application Programming Interface (API) 152 153alloc_workqueue() allocates a wq. The original create_*workqueue() 154functions are deprecated and scheduled for removal. alloc_workqueue() 155takes three arguments - @name, @flags and @max_active. @name is the 156name of the wq and also used as the name of the rescuer thread if 157there is one. 158 159A wq no longer manages execution resources but serves as a domain for 160forward progress guarantee, flush and work item attributes. @flags 161and @max_active control how work items are assigned execution 162resources, scheduled and executed. 163 164@flags: 165 166 WQ_NON_REENTRANT 167 168 By default, a wq guarantees non-reentrance only on the same 169 CPU. A work item may not be executed concurrently on the same 170 CPU by multiple workers but is allowed to be executed 171 concurrently on multiple CPUs. This flag makes sure 172 non-reentrance is enforced across all CPUs. Work items queued 173 to a non-reentrant wq are guaranteed to be executed by at most 174 one worker system-wide at any given time. 175 176 WQ_UNBOUND 177 178 Work items queued to an unbound wq are served by a special 179 gcwq which hosts workers which are not bound to any specific 180 CPU. This makes the wq behave as a simple execution context 181 provider without concurrency management. The unbound gcwq 182 tries to start execution of work items as soon as possible. 183 Unbound wq sacrifices locality but is useful for the following 184 cases. 185 186 * Wide fluctuation in the concurrency level requirement is 187 expected and using bound wq may end up creating large number 188 of mostly unused workers across different CPUs as the issuer 189 hops through different CPUs. 190 191 * Long running CPU intensive workloads which can be better 192 managed by the system scheduler. 193 194 WQ_FREEZABLE 195 196 A freezable wq participates in the freeze phase of the system 197 suspend operations. Work items on the wq are drained and no 198 new work item starts execution until thawed. 199 200 WQ_MEM_RECLAIM 201 202 All wq which might be used in the memory reclaim paths _MUST_ 203 have this flag set. The wq is guaranteed to have at least one 204 execution context regardless of memory pressure. 205 206 WQ_HIGHPRI 207 208 Work items of a highpri wq are queued at the head of the 209 worklist of the target gcwq and start execution regardless of 210 the current concurrency level. In other words, highpri work 211 items will always start execution as soon as execution 212 resource is available. 213 214 Ordering among highpri work items is preserved - a highpri 215 work item queued after another highpri work item will start 216 execution after the earlier highpri work item starts. 217 218 Although highpri work items are not held back by other 219 runnable work items, they still contribute to the concurrency 220 level. Highpri work items in runnable state will prevent 221 non-highpri work items from starting execution. 222 223 This flag is meaningless for unbound wq. 224 225 WQ_CPU_INTENSIVE 226 227 Work items of a CPU intensive wq do not contribute to the 228 concurrency level. In other words, runnable CPU intensive 229 work items will not prevent other work items from starting 230 execution. This is useful for bound work items which are 231 expected to hog CPU cycles so that their execution is 232 regulated by the system scheduler. 233 234 Although CPU intensive work items don't contribute to the 235 concurrency level, start of their executions is still 236 regulated by the concurrency management and runnable 237 non-CPU-intensive work items can delay execution of CPU 238 intensive work items. 239 240 This flag is meaningless for unbound wq. 241 242 WQ_HIGHPRI | WQ_CPU_INTENSIVE 243 244 This combination makes the wq avoid interaction with 245 concurrency management completely and behave as a simple 246 per-CPU execution context provider. Work items queued on a 247 highpri CPU-intensive wq start execution as soon as resources 248 are available and don't affect execution of other work items. 249 250@max_active: 251 252@max_active determines the maximum number of execution contexts per 253CPU which can be assigned to the work items of a wq. For example, 254with @max_active of 16, at most 16 work items of the wq can be 255executing at the same time per CPU. 256 257Currently, for a bound wq, the maximum limit for @max_active is 512 258and the default value used when 0 is specified is 256. For an unbound 259wq, the limit is higher of 512 and 4 * num_possible_cpus(). These 260values are chosen sufficiently high such that they are not the 261limiting factor while providing protection in runaway cases. 262 263The number of active work items of a wq is usually regulated by the 264users of the wq, more specifically, by how many work items the users 265may queue at the same time. Unless there is a specific need for 266throttling the number of active work items, specifying '0' is 267recommended. 268 269Some users depend on the strict execution ordering of ST wq. The 270combination of @max_active of 1 and WQ_UNBOUND is used to achieve this 271behavior. Work items on such wq are always queued to the unbound gcwq 272and only one work item can be active at any given time thus achieving 273the same ordering property as ST wq. 274 275 2765. Example Execution Scenarios 277 278The following example execution scenarios try to illustrate how cmwq 279behave under different configurations. 280 281 Work items w0, w1, w2 are queued to a bound wq q0 on the same CPU. 282 w0 burns CPU for 5ms then sleeps for 10ms then burns CPU for 5ms 283 again before finishing. w1 and w2 burn CPU for 5ms then sleep for 284 10ms. 285 286Ignoring all other tasks, works and processing overhead, and assuming 287simple FIFO scheduling, the following is one highly simplified version 288of possible sequences of events with the original wq. 289 290 TIME IN MSECS EVENT 291 0 w0 starts and burns CPU 292 5 w0 sleeps 293 15 w0 wakes up and burns CPU 294 20 w0 finishes 295 20 w1 starts and burns CPU 296 25 w1 sleeps 297 35 w1 wakes up and finishes 298 35 w2 starts and burns CPU 299 40 w2 sleeps 300 50 w2 wakes up and finishes 301 302And with cmwq with @max_active >= 3, 303 304 TIME IN MSECS EVENT 305 0 w0 starts and burns CPU 306 5 w0 sleeps 307 5 w1 starts and burns CPU 308 10 w1 sleeps 309 10 w2 starts and burns CPU 310 15 w2 sleeps 311 15 w0 wakes up and burns CPU 312 20 w0 finishes 313 20 w1 wakes up and finishes 314 25 w2 wakes up and finishes 315 316If @max_active == 2, 317 318 TIME IN MSECS EVENT 319 0 w0 starts and burns CPU 320 5 w0 sleeps 321 5 w1 starts and burns CPU 322 10 w1 sleeps 323 15 w0 wakes up and burns CPU 324 20 w0 finishes 325 20 w1 wakes up and finishes 326 20 w2 starts and burns CPU 327 25 w2 sleeps 328 35 w2 wakes up and finishes 329 330Now, let's assume w1 and w2 are queued to a different wq q1 which has 331WQ_HIGHPRI set, 332 333 TIME IN MSECS EVENT 334 0 w1 and w2 start and burn CPU 335 5 w1 sleeps 336 10 w2 sleeps 337 10 w0 starts and burns CPU 338 15 w0 sleeps 339 15 w1 wakes up and finishes 340 20 w2 wakes up and finishes 341 25 w0 wakes up and burns CPU 342 30 w0 finishes 343 344If q1 has WQ_CPU_INTENSIVE set, 345 346 TIME IN MSECS EVENT 347 0 w0 starts and burns CPU 348 5 w0 sleeps 349 5 w1 and w2 start and burn CPU 350 10 w1 sleeps 351 15 w2 sleeps 352 15 w0 wakes up and burns CPU 353 20 w0 finishes 354 20 w1 wakes up and finishes 355 25 w2 wakes up and finishes 356 357 3586. Guidelines 359 360* Do not forget to use WQ_MEM_RECLAIM if a wq may process work items 361 which are used during memory reclaim. Each wq with WQ_MEM_RECLAIM 362 set has an execution context reserved for it. If there is 363 dependency among multiple work items used during memory reclaim, 364 they should be queued to separate wq each with WQ_MEM_RECLAIM. 365 366* Unless strict ordering is required, there is no need to use ST wq. 367 368* Unless there is a specific need, using 0 for @max_active is 369 recommended. In most use cases, concurrency level usually stays 370 well under the default limit. 371 372* A wq serves as a domain for forward progress guarantee 373 (WQ_MEM_RECLAIM, flush and work item attributes. Work items which 374 are not involved in memory reclaim and don't need to be flushed as a 375 part of a group of work items, and don't require any special 376 attribute, can use one of the system wq. There is no difference in 377 execution characteristics between using a dedicated wq and a system 378 wq. 379 380* Unless work items are expected to consume a huge amount of CPU 381 cycles, using a bound wq is usually beneficial due to the increased 382 level of locality in wq operations and work item execution. 383 384 3857. Debugging 386 387Because the work functions are executed by generic worker threads 388there are a few tricks needed to shed some light on misbehaving 389workqueue users. 390 391Worker threads show up in the process list as: 392 393root 5671 0.0 0.0 0 0 ? S 12:07 0:00 [kworker/0:1] 394root 5672 0.0 0.0 0 0 ? S 12:07 0:00 [kworker/1:2] 395root 5673 0.0 0.0 0 0 ? S 12:12 0:00 [kworker/0:0] 396root 5674 0.0 0.0 0 0 ? S 12:13 0:00 [kworker/1:0] 397 398If kworkers are going crazy (using too much cpu), there are two types 399of possible problems: 400 401 1. Something beeing scheduled in rapid succession 402 2. A single work item that consumes lots of cpu cycles 403 404The first one can be tracked using tracing: 405 406 $ echo workqueue:workqueue_queue_work > /sys/kernel/debug/tracing/set_event 407 $ cat /sys/kernel/debug/tracing/trace_pipe > out.txt 408 (wait a few secs) 409 ^C 410 411If something is busy looping on work queueing, it would be dominating 412the output and the offender can be determined with the work item 413function. 414 415For the second type of problems it should be possible to just check 416the stack trace of the offending worker thread. 417 418 $ cat /proc/THE_OFFENDING_KWORKER/stack 419 420The work item's function should be trivially visible in the stack 421trace. 422