Concurrency Managed Workqueue (cmwq) September, 2010 Tejun Heo Florian Mickler CONTENTS 1. Introduction 2. Why cmwq? 3. The Design 4. Application Programming Interface (API) 5. Example Execution Scenarios 6. Guidelines 7. Debugging 1. Introduction There are many cases where an asynchronous process execution context is needed and the workqueue (wq) API is the most commonly used mechanism for such cases. When such an asynchronous execution context is needed, a work item describing which function to execute is put on a queue. An independent thread serves as the asynchronous execution context. The queue is called workqueue and the thread is called worker. While there are work items on the workqueue the worker executes the functions associated with the work items one after the other. When there is no work item left on the workqueue the worker becomes idle. When a new work item gets queued, the worker begins executing again. 2. Why cmwq? In the original wq implementation, a multi threaded (MT) wq had one worker thread per CPU and a single threaded (ST) wq had one worker thread system-wide. A single MT wq needed to keep around the same number of workers as the number of CPUs. The kernel grew a lot of MT wq users over the years and with the number of CPU cores continuously rising, some systems saturated the default 32k PID space just booting up. Although MT wq wasted a lot of resource, the level of concurrency provided was unsatisfactory. The limitation was common to both ST and MT wq albeit less severe on MT. Each wq maintained its own separate worker pool. A MT wq could provide only one execution context per CPU while a ST wq one for the whole system. Work items had to compete for those very limited execution contexts leading to various problems including proneness to deadlocks around the single execution context. The tension between the provided level of concurrency and resource usage also forced its users to make unnecessary tradeoffs like libata choosing to use ST wq for polling PIOs and accepting an unnecessary limitation that no two polling PIOs can progress at the same time. As MT wq don't provide much better concurrency, users which require higher level of concurrency, like async or fscache, had to implement their own thread pool. Concurrency Managed Workqueue (cmwq) is a reimplementation of wq with focus on the following goals. * Maintain compatibility with the original workqueue API. * Use per-CPU unified worker pools shared by all wq to provide flexible level of concurrency on demand without wasting a lot of resource. * Automatically regulate worker pool and level of concurrency so that the API users don't need to worry about such details. 3. The Design In order to ease the asynchronous execution of functions a new abstraction, the work item, is introduced. A work item is a simple struct that holds a pointer to the function that is to be executed asynchronously. Whenever a driver or subsystem wants a function to be executed asynchronously it has to set up a work item pointing to that function and queue that work item on a workqueue. Special purpose threads, called worker threads, execute the functions off of the queue, one after the other. If no work is queued, the worker threads become idle. These worker threads are managed in so called worker-pools. The cmwq design differentiates between the user-facing workqueues that subsystems and drivers queue work items on and the backend mechanism which manages worker-pools and processes the queued work items. There are two worker-pools, one for normal work items and the other for high priority ones, for each possible CPU and some extra worker-pools to serve work items queued on unbound workqueues - the number of these backing pools is dynamic. Subsystems and drivers can create and queue work items through special workqueue API functions as they see fit. They can influence some aspects of the way the work items are executed by setting flags on the workqueue they are putting the work item on. These flags include things like CPU locality, concurrency limits, priority and more. To get a detailed overview refer to the API description of alloc_workqueue() below. When a work item is queued to a workqueue, the target worker-pool is determined according to the queue parameters and workqueue attributes and appended on the shared worklist of the worker-pool. For example, unless specifically overridden, a work item of a bound workqueue will be queued on the worklist of either normal or highpri worker-pool that is associated to the CPU the issuer is running on. For any worker pool implementation, managing the concurrency level (how many execution contexts are active) is an important issue. cmwq tries to keep the concurrency at a minimal but sufficient level. Minimal to save resources and sufficient in that the system is used at its full capacity. Each worker-pool bound to an actual CPU implements concurrency management by hooking into the scheduler. The worker-pool is notified whenever an active worker wakes up or sleeps and keeps track of the number of the currently runnable workers. Generally, work items are not expected to hog a CPU and consume many cycles. That means maintaining just enough concurrency to prevent work processing from stalling should be optimal. As long as there are one or more runnable workers on the CPU, the worker-pool doesn't start execution of a new work, but, when the last running worker goes to sleep, it immediately schedules a new worker so that the CPU doesn't sit idle while there are pending work items. This allows using a minimal number of workers without losing execution bandwidth. Keeping idle workers around doesn't cost other than the memory space for kthreads, so cmwq holds onto idle ones for a while before killing them. For unbound workqueues, the number of backing pools is dynamic. Unbound workqueue can be assigned custom attributes using apply_workqueue_attrs() and workqueue will automatically create backing worker pools matching the attributes. The responsibility of regulating concurrency level is on the users. There is also a flag to mark a bound wq to ignore the concurrency management. Please refer to the API section for details. Forward progress guarantee relies on that workers can be created when more execution contexts are necessary, which in turn is guaranteed through the use of rescue workers. All work items which might be used on code paths that handle memory reclaim are required to be queued on wq's that have a rescue-worker reserved for execution under memory pressure. Else it is possible that the worker-pool deadlocks waiting for execution contexts to free up. 4. Application Programming Interface (API) alloc_workqueue() allocates a wq. The original create_*workqueue() functions are deprecated and scheduled for removal. alloc_workqueue() takes three arguments - @name, @flags and @max_active. @name is the name of the wq and also used as the name of the rescuer thread if there is one. A wq no longer manages execution resources but serves as a domain for forward progress guarantee, flush and work item attributes. @flags and @max_active control how work items are assigned execution resources, scheduled and executed. @flags: WQ_UNBOUND Work items queued to an unbound wq are served by the special woker-pools which host workers which are not bound to any specific CPU. This makes the wq behave as a simple execution context provider without concurrency management. The unbound worker-pools try to start execution of work items as soon as possible. Unbound wq sacrifices locality but is useful for the following cases. * Wide fluctuation in the concurrency level requirement is expected and using bound wq may end up creating large number of mostly unused workers across different CPUs as the issuer hops through different CPUs. * Long running CPU intensive workloads which can be better managed by the system scheduler. WQ_FREEZABLE A freezable wq participates in the freeze phase of the system suspend operations. Work items on the wq are drained and no new work item starts execution until thawed. WQ_MEM_RECLAIM All wq which might be used i