========= Workqueue ========= :Date: September, 2010 :Author: Tejun Heo :Author: Florian Mickler 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. Why Concurrency Managed Workqueue? ================================== 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. An MT wq could provide only one execution context per CPU while an 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. 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. 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 worker-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 in the memory reclaim paths **MUST** have this flag set. The wq is guaranteed to have at least one execution context regardless of memory pressure. ``WQ_HIGHPRI`` Work items of a highpri wq are queued to the highpri worker-pool of the target cpu. Highpri worker-pools are served by worker threads with elevated nice level. Note that normal and highpri worker-pools don't interact with each other. Each maintains its separate pool of workers and implements concurrency management among its workers. ``WQ_CPU_INTENSIVE`` Work items of a CPU intensive wq do not contribute to the concurrency level. In other words, runnable CPU intensive work items will not prevent other work items in the same worker-pool from starting execution. This is useful for bound work items which are expected to hog CPU cycles so that their execution is regulated by the system scheduler. Although CPU intensive work items don't contribute to the concurrency level, start of their executions is still regulated by the concurrency management and runnable non-CPU-intensive work items can delay execution of CPU intensive work items. This flag is meaningless for unbound wq. ``max_active`` -------------- ``@max_active`` determines the maximum number of execution contexts per CPU which can be assigned to the work items of a wq. For example, with ``@max_active`` of 16, at most 16 work items of the wq can be executing at the same time per CPU. This is always a per-CPU attribute, even for unbound workqueues. The maximum limit for ``@max_active`` is 512 and the default value used when 0 is specified is 256. These values are chosen sufficiently high such that they are not the limiting factor while providing protection in runaway cases. The number of active work items of a wq is usually regulated by the users of the wq, more specifically, by how many work items the users may queue at the same time. Unless there is a specific need for throttling the number of active work items, specifying '0' is recommended. Some users depend on the strict execution ordering of ST wq. The combination of ``@max_active`` of 1 and ``WQ_UNBOUND`` used to achieve this behavior. Work items on such wq were always queued to the unbound worker-pools and only one work item could be active at any given time thus achieving the same ordering property as ST wq. In the current implementation the above configuration only guarantees ST behavior within a given NUMA node. Instead ``alloc_ordered_workqueue()`` should be used to achieve system-wide ST behavior. Example Execution Scenarios =========================== The following example execution scenarios try to illustrate how cmwq behave under different configurations. Work items w0, w1, w2 are queued to a bound wq q0 on the same CPU. w0 burns CPU for 5ms then sleeps for 10ms then burns CPU for 5ms again before finishing. w1 and w2 burn CPU for 5ms then sleep for 10ms. Ignoring all other tasks, works and processing overhead, and assuming simple FIFO scheduling, the following is one highly simplified version of possible sequences of events with the original wq. :: TIME IN MSECS EVENT 0 w0 starts and burns CPU 5 w0 sleeps 15 w0 wakes up and burns CPU 20 w0 finishes 20 w1 starts and burns CPU 25 w1 sleeps 35 w1 wakes up and finishes 35 w2 starts and burns CPU 40 w2 sleeps 50 w2 wakes up and finishes And with cmwq with ``@max_active`` >= 3, :: TIME IN MSECS EVENT 0 w0 starts and burns CPU 5 w0 sleeps 5 w1 starts and burns CPU 10 w1 sleeps 10 w2 starts and burns CPU 15 w2 sleeps 15 w0 wakes up and burns CPU 20 w0 finishes 20 w1 wakes up and finishes 25 w2 wakes up and finishes If ``@max_active`` == 2, :: TIME IN MSECS EVENT 0 w0 starts and burns CPU 5 w0 sleeps 5 w1 starts and burns CPU 10 w1 sleeps 15 w0 wakes up and burns CPU 20 w0 finishes 20 w1 wakes up and finishes 20 w2 starts and burns CPU 25 w2 sleeps 35 w2 wakes up and finishes Now, let's assume w1 and w2 are queued to a different wq q1 which has ``WQ_CPU_INTENSIVE`` set, :: TIME IN MSECS EVENT 0 w0 starts and burns CPU 5 w0 sleeps 5 w1 and w2 start and burn CPU 10 w1 sleeps 15 w2 sleeps 15 w0 wakes up and burns CPU 20 w0 finishes 20 w1 wakes up and finishes 25 w2 wakes up and finishes Guidelines ========== * Do not forget to use ``WQ_MEM_RECLAIM`` if a wq may process work items which are used during memory reclaim. Each wq with ``WQ_MEM_RECLAIM`` set has an execution context reserved for it. If there is dependency among multiple work items used during memory reclaim, they should be queued to separate wq each with ``WQ_MEM_RECLAIM``. * Unless strict ordering is required, there is no need to use ST wq. * Unless there is a specific need, using 0 for @max_active is recommended. In most use cases, concurrency level usually stays well under the default limit. * A wq serves as a domain for forward progress guarantee (``WQ_MEM_RECLAIM``, flush and work item attributes. Work items which are not involved in memory reclaim and don't need to be flushed as a part of a group of work items, and don't require any special attribute, can use one of the system wq. There is no difference in execution characteristics between using a dedicated wq and a system wq. * Unless work items are expected to consume a huge amount of CPU cycles, using a bound wq is usually beneficial due to the increased level of locality in wq operations and work item execution. Affinity Scopes =============== An unbound workqueue groups CPUs according to its affinity scope to improve cache locality. For example, if a workqueue is using the default affinity scope of "cache", it will group CPUs according to last level cache boundaries. A work item queued on the workqueue will be assigned to a worker on one of the CPUs which share the last level cache with the issuing CPU. Once started, the worker may or may not be allowed to move outside the scope depending on the ``affinity_strict`` setting of the scope. Workqueue currently supports the following affinity scopes. ``default`` Use the scope in module parameter ``workqueue.default_affinity_scope`` which is always set to one of the scopes below. ``cpu`` CPUs are not grouped. A work item issued on one CPU is processed by a worker on the same CPU. This makes unbound workqueues behave as per-cpu workqueues without concurrency management. ``smt`` CPUs are grouped according to SMT boundaries. This usually means that the logical threads of each physical CPU core are grouped together. ``cache`` CPUs are grouped according to cache boundaries. Which specific cache boundary is used is determined by the arch code. L3 is used in a lot of cases. This is the default affinity scope. ``numa`` CPUs are grouped according to NUMA boundaries. ``system`` All CPUs are put in the same group. Workqueue makes no effort to process a work item on a CPU close to the issuing CPU. The default affinity scope can be changed with the module parameter ``workqueue.default_affinity_scope`` and a specific workqueue's affinity scope can be changed using ``apply_workqueue_attrs()``. If ``WQ_SYSFS`` is set, the workqueue will have the following affinity scope related interface files under its ``/sys/devices/virtual/workqueue/WQ_NAME/`` directory. ``affinity_scope`` Read to see the current affinity scope. Write to change. When default is the current scope, reading this file will also show the current effective scope in parentheses, for example, ``default (cache)``. ``affinity_strict`` 0 by default indicating that affinity scopes are not strict. When a work item starts execution, workqueue makes a best-effort attempt to ensure that the worker is inside its affinity scope, which is called repatriation. Once started, the scheduler is free to move the worker anywhere in the system as it sees fit. This enables benefiting from scope locality while still being able to utilize other CPUs if necessary and available. If set to 1, all workers of the scope are guaranteed always to be in the scope. This may be useful when crossing affinity scopes has other implications, for example, in terms of power consumption or workload isolation. Strict NUMA scope can also be used to match the workqueue behavior of older kernels. Affinity Scopes and Performance =============================== It'd be ideal if an unbound workqueue's behavior is optimal for vast majority of use cases without further tuning. Unfortunately, in the current kernel, there exists a pronounced trade-off between locality and utilization necessitating explicit configurations when workqueues are heavily used. Higher locality leads to higher efficiency where more work is performed for the same number of consumed CPU cycles. However, higher locality may also cause lower overall system utilization if the work items are not spread enough across the affinity scopes by the issuers. The following performance testing with dm-crypt clearly illustrates this trade-off. The tests are run on a CPU with 12-cores/24-threads split across four L3 caches (AMD Ryzen 9 3900x). CPU clock boost is turned off for consistency. ``/dev/dm-0`` is a dm-crypt device created on NVME SSD (Samsung 990 PRO) and opened with ``cryptsetup`` with default settings. Scenario 1: Enough issuers and work spread across the machine ------------------------------------------------------------- The command used: :: $ fio --filename=/dev/dm-0 --direct=1 --rw=randrw --bs=32k --ioengine=libaio \ --iodepth=64 --runtime=60 --numjobs=24 --time_based --group_reporting \ --name=iops-test-job --verify=sha512 There are 24 issuers, each issuing 64 IOs concurrently. ``--verify=sha512`` makes ``fio`` generate and read back the content each time which makes execution locality matter between the issuer and ``kcryptd``. The following are the read bandwidths and CPU utilizations depending on different affinity scope settings on ``kcryptd`` measured over five runs. Bandwidths are in MiBps, and CPU util in percents. .. list-table:: :widths: 16 20 20 :header-rows: 1 * - Affinity - Bandwidth (MiBps) - CPU util (%) * - system - 1159.40 ±1.34 - 99.31 ±0.02 * - cache - 1166.40 ±0.89 - 99.34 ±0.01 * - cache (strict) - 1166.00 ±0.71 - 99.35 ±0.01 With enough issuers spread across the system, there is no downside to "cache", strict or otherwise. All three configurations saturate the whole machine but the cache-affine ones outperform by 0.6% thanks to improved locality. Scenario 2: Fewer issuers, enough work for saturation ----------------------------------------------------- The command used: :: $ fio --filename=/dev/dm-0 --direct=1 --rw=randrw --bs=32k \ --ioengine=libaio --iodepth=64 --runtime=60 --numjobs=8 \ --time_based --group_reporting --name=iops-test-job --verify=sha512 The only difference from the previous scenario is ``--numjobs=8``. There are a third of the issuers but is still enough total work to saturate the system. .. list-table:: :widths: 16 20 20 :header-rows: 1 * - Affinity - Bandwidth (MiBps) - CPU util (%) * - system - 1155.40 ±0.89 - 97.41 ±0.05 * - cache - 1154.40 ±1.14 - 96.15 ±0.09 * - cache (strict) - 1112.00 ±4.64 - 93.26 ±0.35 This is more than enough work to saturate the system. Both "system" and "cache" are nearly saturating the machine but not fully. "cache" is using less CPU but the better efficiency puts it at the same bandwidth as "system". Eight issuers moving around over four L3 cache scope still allow "cache (strict)" to mostly saturate the machine but the loss of work conservation is now starting to hurt with 3.7% bandwidth loss. Scenario 3: Even fewer issuers, not enough work to saturate ----------------------------------------------------------- The command used: :: $ fio --filename=/dev/dm-0 --direct=1 --rw=randrw --bs=32k \ --ioengine=libaio --iodepth=64 --runtime=60 --numjobs=4 \ --time_based --group_reporting --name=iops-test-job --verify=sha512 Again, the only difference is ``--numjobs=4``. With the number of issuers reduced to four, there now isn't enough work to saturate the whole system and the bandwidth becomes dependent on completion latencies. .. list-table:: :widths: 16 20 20 :header-rows: 1 * - Affinity - Bandwidth (MiBps) - CPU util (%) * - system - 993.60 ±1.82 - 75.49 ±0.06 * - cache - 973.40 ±1.52 - 74.90 ±0.07 * - cache (strict) - 828.20 ±4.49 - 66.84 ±0.29 Now, the tradeoff between locality and utilization is clearer. "cache" shows 2% bandwidth loss compared to "system" and "cache (struct)" whopping 20%. Conclusion and Recommendations ------------------------------ In the above experiments, the efficiency advantage of the "cache" affinity scope over "system" is, while consistent and noticeable, small. However, the impact is dependent on the distances between the scopes and may be more pronounced in processors with more complex topologies. While the loss of work-conservation in certain scenarios hurts, it is a lot better than "cache (strict)" and maximizing workqueue utilization is unlikely to be the common case anyway. As such, "cache" is the default affinity scope for unbound pools. * As there is no one option which is great for most cases, workqueue usages that may consume a significant amount of CPU are recommended to configure the workqueues using ``apply_workqueue_attrs()`` and/or enable ``WQ_SYSFS``. * An unbound workqueue with strict "cpu" affinity scope behaves the same as ``WQ_CPU_INTENSIVE`` per-cpu workqueue. There is no real advanage to the latter and an unbound workqueue provides a lot more flexibility. * Affinity scopes are introduced in Linux v6.5. To emulate the previous behavior, use strict "numa" affinity scope. * The loss of work-conservation in non-strict affinity scopes is likely originating from the scheduler. There is no theoretical reason why the kernel wouldn't be able to do the right thing and maintain work-conservation in most cases. As such, it is possible that future scheduler improvements may make most of these tunables unnecessary. Examining Configuration ======================= Use tools/workqueue/wq_dump.py to examine unbound CPU affinity configuration, worker pools and how workqueues map to the pools: :: $ tools/workqueue/wq_dump.py Affinity Scopes =============== wq_unbound_cpumask=0000000f CPU nr_pods 4 pod_cpus [0]=00000001 [1]=00000002 [2]=00000004 [3]=00000008 pod_node [0]=0 [1]=0 [2]=1 [3]=1 cpu_pod [0]=0 [1]=1 [2]=2 [3]=3 SMT nr_pods 4 pod_cpus [0]=00000001 [1]=00000002 [2]=00000004 [3]=00000008 pod_node [0]=0 [1]=0 [2]=1 [3]=1 cpu_pod [0]=0 [1]=1 [2]=2 [3]=3 CACHE (default) nr_pods 2 pod_cpus [0]=00000003 [1]=0000000c pod_node [0]=0 [1]=1 cpu_pod [0]=0 [1]=0 [2]=1 [3]=1 NUMA nr_pods 2 pod_cpus [0]=00000003 [1]=0000000c pod_node [0]=0 [1]=1 cpu_pod [0]=0 [1]=0 [2]=1 [3]=1 SYSTEM nr_pods 1 pod_cpus [0]=0000000f pod_node [0]=-1 cpu_pod [0]=0 [1]=0 [2]=0 [3]=0 Worker Pools ============ pool[00] ref= 1 nice= 0 idle/workers= 4/ 4 cpu= 0 pool[01] ref= 1 nice=-20 idle/workers= 2/ 2 cpu= 0 pool[02] ref= 1 nice= 0 idle/workers= 4/ 4 cpu= 1 pool[03] ref= 1 nice=-20 idle/workers= 2/ 2 cpu= 1 pool[04] ref= 1 nice= 0 idle/workers= 4/ 4 cpu= 2 pool[05] ref= 1 nice=-20 idle/workers= 2/ 2 cpu= 2 pool[06] ref= 1 nice= 0 idle/workers= 3/ 3 cpu= 3 pool[07] ref= 1 nice=-20 idle/workers= 2/ 2 cpu= 3 pool[08] ref=42 nice= 0 idle/workers= 6/ 6 cpus=0000000f pool[09] ref=28 nice= 0 idle/workers= 3/ 3 cpus=00000003 pool[10] ref=28 nice= 0 idle/workers= 17/ 17 cpus=0000000c pool[11] ref= 1 nice=-20 idle/workers= 1/ 1 cpus=0000000f pool[12] ref= 2 nice=-20 idle/workers= 1/ 1 cpus=00000003 pool[13] ref= 2 nice=-20 idle/workers= 1/ 1 cpus=0000000c Workqueue CPU -> pool ===================== [ workqueue \ CPU 0 1 2 3 dfl] events percpu 0 2 4 6 events_highpri percpu 1 3 5 7 events_long percpu 0 2 4 6 events_unbound unbound 9 9 10 10 8 events_freezable percpu 0 2 4 6 events_power_efficient percpu 0 2 4 6 events_freezable_power_ percpu 0 2 4 6 rcu_gp percpu 0 2 4 6 rcu_par_gp percpu 0 2 4 6 slub_flushwq percpu 0 2 4 6 netns ordered 8 8 8 8 8 ... See the command's help message for more info. Monitoring ========== Use tools/workqueue/wq_monitor.py to monitor workqueue operations: :: $ tools/workqueue/wq_monitor.py events total infl CPUtime CPUhog CMW/RPR mayday rescued events 18545 0 6.1 0 5 - - events_highpri 8 0 0.0 0 0 - - events_long 3 0 0.0 0 0 - - events_unbound 38306 0 0.1 - 7 - - events_freezable 0 0 0.0 0 0 - - events_power_efficient 29598 0 0.2 0 0 - - events_freezable_power_ 10 0 0.0 0 0 - - sock_diag_events 0 0 0.0 0 0 - - total infl CPUtime CPUhog CMW/RPR mayday rescued events 18548 0 6.1 0 5 - - events_highpri 8 0 0.0 0 0 - - events_long 3 0 0.0 0 0 - - events_unbound 38322 0 0.1 - 7 - - events_freezable 0 0 0.0 0 0 - - events_power_efficient 29603 0 0.2 0 0 - - events_freezable_power_ 10 0 0.0 0 0 - - sock_diag_events 0 0 0.0 0 0 - - ... See the command's help message for more info. Debugging ========= Because the work functions are executed by generic worker threads there are a few tricks needed to shed some light on misbehaving workqueue users. Worker threads show up in the process list as: :: root 5671 0.0 0.0 0 0 ? S 12:07 0:00 [kworker/0:1] root 5672 0.0 0.0 0 0 ? S 12:07 0:00 [kworker/1:2] root 5673 0.0 0.0 0 0 ? S 12:12 0:00 [kworker/0:0] root 5674 0.0 0.0 0 0 ? S 12:13 0:00 [kworker/1:0] If kworkers are going crazy (using too much cpu), there are two types of possible problems: 1. Something being scheduled in rapid succession 2. A single work item that consumes lots of cpu cycles The first one can be tracked using tracing: :: $ echo workqueue:workqueue_queue_work > /sys/kernel/tracing/set_event $ cat /sys/kernel/tracing/trace_pipe > out.txt (wait a few secs) ^C If something is busy looping on work queueing, it would be dominating the output and the offender can be determined with the work item function. For the second type of problems it should be possible to just check the stack trace of the offending worker thread. :: $ cat /proc/THE_OFFENDING_KWORKER/stack The work item's function should be trivially visible in the stack trace. Non-reentrance Conditions ========================= Workqueue guarantees that a work item cannot be re-entrant if the following conditions hold after a work item gets queued: 1. The work function hasn't been changed. 2. No one queues the work item to another workqueue. 3. The work item hasn't been reinitiated. In other words, if the above conditions hold, the work item is guaranteed to be executed by at most one worker system-wide at any given time. Note that requeuing the work item (to the same queue) in the self function doesn't break these conditions, so it's safe to do. Otherwise, caution is required when breaking the conditions inside a work function. Kernel Inline Documentations Reference ====================================== .. kernel-doc:: include/linux/workqueue.h .. kernel-doc:: kernel/workqueue.c