Document User Managed Concurrency Groups syscalls, data structures, state transitions, etc. This is a text version of umcg.rst. Signed-off-by: Peter Oskolkov <posk@xxxxxxxxxx> --- Documentation/userspace-api/umcg.txt | 537 +++++++++++++++++++++++++++ 1 file changed, 537 insertions(+) create mode 100644 Documentation/userspace-api/umcg.txt diff --git a/Documentation/userspace-api/umcg.txt b/Documentation/userspace-api/umcg.txt new file mode 100644 index 000000000000..ef055bc40fc6 --- /dev/null +++ b/Documentation/userspace-api/umcg.txt @@ -0,0 +1,537 @@ +UMCG USERSPACE API + +User Managed Concurrency Groups (UMCG) is an M:N threading +subsystem/toolkit that lets user space application developers implement +in-process user space schedulers. + + +CONTENTS + + WHY? HETEROGENEOUS IN-PROCESS WORKLOADS + REQUIREMENTS + UMCG KERNEL API + SERVERS + WORKERS + UMCG TASK STATES + STRUCT UMCG_TASK + SYS_UMCG_CTL() + SYS_UMCG_WAIT() + STATE TRANSITIONS + SERVER-ONLY USE CASES + + +WHY? HETEROGENEOUS IN-PROCESS WORKLOADS + +Linux kernel's CFS scheduler is designed for the "common" use case, with +efficiency/throughput in mind. Work isolation and workloads of different +"urgency" are addressed by tools such as cgroups, CPU affinity, priorities, +etc., which are difficult or impossible to efficiently use in-process. + +For example, a single DBMS process may receive tens of thousands requests +per second; some of these requests may have strong response latency +requirements as they serve live user requests (e.g. login authentication); +some of these requests may not care much about latency but must be served +within a certain time period (e.g. an hourly aggregate usage report); some +of these requests are to be served only on a best-effort basis and can be +NACKed under high load (e.g. an exploratory research/hypothesis testing +workload). + +Beyond different work item latency/throughput requirements as outlined +above, the DBMS may need to provide certain guarantees to different users; +for example, user A may "reserve" 1 CPU for their high-priority/low latency +requests, 2 CPUs for mid-level throughput workloads, and be allowed to send +as many best-effort requests as possible, which may or may not be served, +depending on the DBMS load. Besides, the best-effort work, started when the +load was low, may need to be delayed if suddenly a large amount of +higher-priority work arrives. With hundreds or thousands of users like +this, it is very difficult to guarantee the application's responsiveness +using standard Linux tools while maintaining high CPU utilization. + +Gaming is another use case: some in-process work must be completed before a +certain deadline dictated by frame rendering schedule, while other work +items can be delayed; some work may need to be cancelled/discarded because +the deadline has passed; etc. + +User Managed Concurrency Groups is an M:N threading toolkit that allows +constructing user space schedulers designed to efficiently manage +heterogeneous in-process workloads described above while maintaining high +CPU utilization (95%+). + + +REQUIREMENTS + +One relatively established way to design high-efficiency, low-latency +systems is to split all work into small on-cpu work items, with +asynchronous I/O and continuations, all executed on a thread pool with the +number of threads not exceeding the number of available CPUs. Although this +approach works, it is quite difficult to develop and maintain such a +system, as, for example, small continuations are difficult to piece +together when debugging. Besides, such asynchronous callback-based systems +tend to be somewhat cache-inefficient, as continuations can get scheduled +on any CPU regardless of cache locality. + +M:N threading and cooperative user space scheduling enables controlled CPU +usage (minimal OS preemption), synchronous coding style, and better cache +locality. + +Specifically: + +* a variable/fluctuating number M of "application" threads should be + "scheduled over" a relatively fixed number N of "kernel" threads, where + N is less than or equal to the number of CPUs available; +* only those application threads that are attached to kernel threads are + scheduled "on CPU"; +* application threads should be able to cooperatively + yield to each other; +* when an application thread blocks in kernel (e.g. in I/O), this becomes + a scheduling event ("block") that the userspace scheduler should be able + to efficiently detect, and reassign a waiting application thread to the + freeded "kernel" thread; +* when a blocked application thread wakes (e.g. its I/O operation + completes), this even ("wake") should also be detectable by the + userspace scheduler, which should be able to either quickly dispatch the + newly woken thread to an idle "kernel" thread or, if all "kernel" + threads are busy, put it in the waiting queue; +* in addition to the above, it would be extremely useful for a separate + in-process "watchdog" facility to be able to monitor the state of each + of the M+N threads, and to intervene in case of runaway workloads + (interrupt/preempt). + + +UMCG KERNEL API + +Based on the requrements above, UMCG kernel API is build around the +following ideas: + +* UMCG server: a task/thread representing "kernel threads", or CPUs from + the requirements above; +* UMCG worker: a task/thread representing "application threads", to be + scheduled over servers; +* UMCG task state: (NONE), RUNNING, BLOCKED, IDLE: states a UMCG task (a + server or a worker) can be in; +* UMCG task state flag: LOCKED, PREEMPTED: additional state flags that + can be ORed with the task state to communicate additional information to + the kernel; +* struct umcg_task: a per-task userspace set of data fields, usually + residing in the TLS, that fully reflects the current task's UMCG state + and controls the way the kernel manages the task; +* sys_umcg_ctl(): a syscall used to register the current task/thread as a + server or a worker, or to unregister a UMCG task; +* sys_umcg_wait(): a syscall used to put the current task to sleep and/or + wake another task, pontentially context-switching between the two tasks + on-CPU synchronously. + + +SERVERS + +When a task/thread is registered as a server, it is in RUNNING state and +behaves like any other normal task/thread. In addition, servers can +interact with other UMCG tasks via sys_umcg_wait(): + +* servers can voluntarily suspend their execution (wait), becoming IDLE; +* servers can wake other IDLE servers; +* servers can context-switch between each other. + +Note that if a server blocks in the kernel not via sys_umcg_wait(), it +still retains its RUNNING state. + +Also note that servers can be used for fast on-CPU context switching across +process boundaries; server-worker interactions assume they belong to the +same mm. + +See the next section on how servers interact with workers. + + +WORKERS + +A worker cannot be RUNNING without having a server associated with it, so +when a task is first registered as a worker, it enters the IDLE state. + +* a worker becomes RUNNING when a server calls sys_umcg_wait to + context-switch into it; the server goes IDLE, and the worker becomes + RUNNING in its place; +* when a running worker blocks in the kernel, it becomes BLOCKED, its + associated server becomes RUNNING and the server's sys_umcg_wait() call + from the bullet above returns; this transition is sometimes called + "block detection"; +* when the syscall on which a BLOCKED worker completes, the worker + becomes IDLE and is added to the list of idle workers; if there is an + idle server waiting, the kernel wakes it; this transition is sometimes + called "wake detection"; +* running workers can voluntarily suspend their execution (wait), + becoming IDLE; their associated servers are woken; +* a RUNNING worker can context-switch with an IDLE worker; the server of + the switched-out worker is transferred to the switched-in worker; +* any UMCG task can "wake" an IDLE worker via sys_umcg_wait(); unless + this is a server running the worker as described in the first bullet in + this list, the worker remain IDLE but is added to the idle workers list; + this "wake" operation exists for completeness, to make sure + wait/wake/context-switch operations are available for all UMCG tasks; +* the userspace can preempt a RUNNING worker by marking it + RUNNING|PREEMPTED and sending a signal to it; the userspace should have + installed a NOP signal handler for the signal; the kernel will then + transition the worker into IDLE|PREEMPTED state and wake its associated + server. + + +UMCG TASK STATES + +Important: all state transitions described below involve at least two +steps: the change of the state field in struct umcg_task, for example +RUNNING to IDLE, and the corresponding change in struct task_struct state, +for example a transition between the task running on CPU and being +descheduled and removed from the kernel runqueue. The key principle of UMCG +API design is that the party initiating the state transition modifies the +state variable. + +For example, a task going IDLE first changes its state from RUNNING to IDLE +in the userpace and then calls sys_umcg_wait(), which completes the +transition. + +Note on documentation: in include/uapi/linux/umcg.h, task states have the +form UMCG_TASK_RUNNING, UMCG_TASK_BLOCKED, etc. In this document these are +usually referred to simply RUNNING and BLOCKED, unless it creates +ambiguity. Task state flags, e.g. UMCG_TF_PREEMPTED, are treated similarly. + +UMCG task states reflect the view from the userspace, rather than from the +kernel. There are three fundamental task states: + +* RUNNING: indicates that the task is schedulable by the kernel; applies + to both servers and workers; +* IDLE: indicates that the task is not schedulable by the kernel (see + umcg_idle_loop() in kernel/sched/umcg.c); applies to both servers and + workers; +* BLOCKED: indicates that the worker is blocked in the kernel; does not + apply to servers. + +In addition to the three states above, two state flags help with state +transitions: + +* LOCKED: the userspace is preparing the worker for a state transition + and "locks" the worker until the worker is ready for the kernel to act + on the state transition; used similarly to preempt_disable or + irq_disable in the kernel; applies only to workers in RUNNING or IDLE + state; RUNNING|LOCKED means "this worker is about to become RUNNING, + while IDLE|LOCKED means "this worker is about to become IDLE or + unregister; +* PREEMPTED: the userspace indicates it wants the worker to be preempted; + there are no situations when both LOCKED and PREEMPTED flags are set at + the same time. + + +STRUCT UMCG_TASK + +From include/uapi/linux/umcg.h: + +struct umcg_task { + uint32_t state; /* r/w */ + uint32_t next_tid; /* r */ + uint64_t idle_workers_ptr; /* r/w */ + uint64_t idle_server_tid_ptr; /* r* */ +}; + +Each UMCG task is identified by struct umcg_task, which is provided to the +kernel when the task is registered via sys_umcg_ctl(). + +* uint32_t state: the current state of the task this struct identifies, + as described in the previous section. Readable/writable by both the + kernel and the userspace. + + bits 0 - 7: task state (RUNNING, IDLE, BLOCKED); + bits 8 - 15: state flags (LOCKED, PREEMPTED); + bits 16 - 23: reserved; must be zeroes; + bits 24 - 31: for userspace use. + +* uint32_t next_tid: contains the TID of the task to context-switch-into + in sys_umcg_wait(); can be zero; writable by the userspace, readable by + the kernel; if this is a RUNNING worker, this field contains the TID of + the server that should be woken when this worker blocks; see + sys_umcg_wait() for more details; + +* uint64_t idle_workers_ptr: this field forms a single-linked list of + idle workers: all RUNNING workers have this field set to point to the + head of the list (a pointer variable in the userspace). + + When a worker's blocking operation in the kernel completes, the kernel + changes the worker's state from BLOCKED to IDLE and adds the worker to + the top of the list of idle workers using this logic: + + /* kernel side */ + /** + * enqueue_idle_worker - push an idle worker onto idle_workers_ptr + * list/stack. + * + * Returns true on success, false on a fatal failure. + */ + static bool enqueue_idle_worker(struct umcg_task __user *ut_worker) + { + u64 __user *node = &ut_worker->idle_workers_ptr; + u64 __user *head_ptr; + u64 first = (u64)node; + u64 head; + + if (get_user_nosleep(head, node) || !head) + return false; + + head_ptr = (u64 __user *)head; + + if (put_user_nosleep(UMCG_IDLE_NODE_PENDING, node)) + return false; + + if (xchg_user_64(head_ptr, &first)) + return false; + + if (put_user_nosleep(first, node)) + return false; + + return true; + } + + In the userspace the list is cleared atomically using this logic: + + /* userspace side */ + uint64_t *idle_workers = (uint64_t *)*head; + + atomic_exchange(&idle_workers, NULL); + + The userspace re-points workers' idle_workers_ptr to the list head + variable before the worker is allowed to become RUNNING again. + + When processing the idle workers list, the userspace should wait for + workers marked as UMCG_IDLE_NODE_PENDING to have the flag cleared (see + enqueue_idle_worker() above). + +* uint64_t idle_server_tid_ptr: points to a variable in the userspace + that points to an idle server, i.e. a server in IDLE state waiting in + sys_umcg_wait(); read-only; workers must have this field set; not used + in servers. + + When a worker's blocking operation in the kernel completes, the kernel + changes the worker's state from BLOCKED to IDLE, adds the worker to the + list of idle workers, and wakes the idle server if present; the kernel + atomically exchanges (*idle_server_tid_ptr) with 0, thus waking the idle + server, if present, only once. See State transitions below for more + details. + + +SYS_UMCG_CTL() + +int sys_umcg_ctl(uint32_t flags, struct umcg_task *self) is used to +register or unregister the current task as a worker or server. Flags can be +one of the following: + + UMCG_CTL_REGISTER: register a server; + UMCG_CTL_REGISTER | UMCG_CTL_WORKER: register a worker; + UMCG_CTL_UNREGISTER: unregister the current server or worker. + +When registering a task, self must point to struct umcg_task describing +this server or worker; the pointer must remain valid until the task is +unregistered. + +When registering a server, self->state must be RUNNING; all other fields in +self must be zeroes. + +When registering a worker, self->state must be BLOCKED; +self->idle_server_tid_ptr and self->idle_workers_ptr must be valid pointers +as described in struct umcg_task; self->next_tid must be zero. + +When unregistering a task, self must be NULL. + + +SYS_UMCG_WAIT() + +int sys_umcg_wait(uint32_t flags, uint64_t abs_timeout) operates on +registered UMCG servers and workers: struct umcg_task *self provided to +sys_umcg_ctl() when registering the current task is consulted in addition +to flags and abs_timeout parameters. + +The function can be used to perform one of the three operations: + +* wait: if self->next_tid is zero, sys_umcg_wait() puts the current + server or worker to sleep; +* wake: if self->next_tid is not zero, and flags & UMCG_WAIT_WAKE_ONLY, + the task identified by next_tid is woken; +* context switch: if self->next_tid is not zero, and !(flags & + UMCG_WAIT_WAKE_ONLY), the current task is put to sleep and the next task + is woken, synchronously switching between the tasks on the current CPU + on the fast path. + +Flags can be zero or a combination of the following values: + +* UMCG_WAIT_WAKE_ONLY: wake the next task, don't put the current task to + sleep; +* UMCG_WAIT_WF_CURRENT_CPU: wake the next task on the curent CPU; this + flag has an effect only if UMCG_WAIT_WAKE_ONLY is set: context switching + is always attempted to happen on the curent CPU. + +The section below provides more details on how servers and workers interact +via sys_umcg_wait(), during worker block/wake events, and during worker +preemption. + + +STATE TRANSITIONS + +As mentioned above, the key principle of UMCG state transitions is that the +party initiating the state transition modifies the state of affected tasks. + +Below, "TASK:STATE" indicates a task T, where T can be either W for worker +or S for server, in state S, where S can be one of the three states, +potentially ORed with a state flag. Each individual state transition is an +atomic operation (cmpxchg) unless indicated otherwise. Also note that the +order of state transitions is important and is part of the contract between +the userspace and the kernel. The kernel is free to kill the task (SIGKILL) +if the contract is broken. + +Some worker state transitions below include adding LOCKED flag to worker +state. This is done to indicate to the kernel that the worker is +transitioning state and should not participate in the block/wake detection +routines, which can happen due to interrupts/pagefaults/signals. + +IDLE|LOCKED means that a running worker is preparing to sleep, so +interrupts should not lead to server wakeup; RUNNING|LOCKED means that an +idle worker is going to be "scheduled to run", but may not yet have its +server set up properly. + +Key state transitions: + +* server to worker context switch ("schedule a worker to run"): + S:RUNNING+W:IDLE => S:IDLE+W:RUNNING: + in the userspace, in the context of the server S running: + S:RUNNING => S:IDLE (mark self as idle) + W:IDLE => W:RUNNING|LOCKED (mark the worker as running) + W.next_tid := S.tid; S.next_tid := W.tid (link the server with + the worker) + W:RUNNING|LOCKED => W:RUNNING (unlock the worker) + S: sys_umcg_wait() (make the syscall) + the kernel context switches from the server to the worker; the + server sleeps until it becomes RUNNING during one of the + transitions below; + +* worker to server context switch (worker "yields"): S:IDLE+W:RUNNING => +S:RUNNING+W:IDLE: + in the userspace, in the context of the worker W running (note that + a running worker has its next_tid set to point to its server): + W:RUNNING => W:IDLE|LOCKED (mark self as idle) + S:IDLE => S:RUNNING (mark the server as running) + W: sys_umcg_wait() (make the syscall) + the kernel removes the LOCKED flag from the worker's state and + context switches from the worker to the server; the worker sleeps + until it becomes RUNNING; + +* worker to worker context switch: W1:RUNNING+W2:IDLE => + W1:IDLE+W2:RUNNING: + in the userspace, in the context of W1 running: + W2:IDLE => W2:RUNNING|LOCKED (mark W2 as running) + W1:RUNNING => W1:IDLE|LOCKED (mark self as idle) + W2.next_tid := W1.next_tid; S.next_tid := W2.tid (transfer the + server W1 => W2) + W1:next_tid := W2.tid (indicate that W1 should context-switch + into W2) + W2:RUNNING|LOCKED => W2:RUNNING (unlock W2) + W1: sys_umcg_wait() (make the syscall) + same as above, the kernel removes the LOCKED flag from the W1's + state and context switches to next_tid; + +* worker wakeup: W:IDLE => W:RUNNING: + in the userspace, a server S can wake a worker W without "running" + it: + S:next_tid :=W.tid + W:next_tid := 0 + W:IDLE => W:RUNNING + sys_umcg_wait(UMCG_WAIT_WAKE_ONLY) (make the syscall) + the kernel will wake the worker W; as the worker does not have a + server assigned, "wake detection" will happen, the worker will be + immediately marked as IDLE and added to idle workers list; an idle + server, if any, will be woken (see 'wake detection' below); + + Note: if needed, it is possible for a worker to wake another + worker: the waker marks itself "IDLE|LOCKED", points its next_tid + to the wakee, makes the syscall, restores its server in next_tid, + marks itself as RUNNING. + +* block detection: worker blocks in the kernel: S:IDLE+W:RUNNING => + S:RUNNING+W:BLOCKED: + when a worker blocks in the kernel in RUNNING state (not LOCKED), + before descheduling the task from the CPU the kernel performs + these operations: + W:RUNNING => W:BLOCKED + S := W.next_tid + S:IDLE => S:RUNNING + try_to_wake_up(S) + if any of the first three operations above fail, the worker is + killed via SIGKILL. Note that ttwu(S) is not required to succeed, + as the server may still be transitioning to sleep in + sys_umcg_wait(); before actually putting the server to sleep its + UMCG state is checked and, if it is RUNNING, sys_umcg_wait() + returns to the userspace; + if the worker has its LOCKED flag set, block detection does not + trigger, as the worker is assumed to be in the userspace + scheduling code. + +* wake detection: worker wakes in the kernel: W:BLOCKED => W:IDLE: + all workers' returns to the userspace are intercepted: + start: (a label) + if W:RUNNING & W.next_tid != 0: let the worker exit to the + userspace, as this is a RUNNING worker with a server; + W:* => W:IDLE (previously blocked or woken without servers + workers are not allowed to return to the userspace); + the worker is appended to W.idle_workers_ptr idle workers list; + S := *W.idle_server_tid_ptr; if (S != 0) S:IDLE => S.RUNNING; + ttwu(S) + idle_loop(W): this is the same idle loop that sys_umcg_wait() + uses: it breaks only when the worker becomes RUNNING; when + the idle loop exits, it is assumed that the userspace has + properly removed the worker from the idle workers list + before marking it RUNNING; + goto start; (repeat from the beginning). + + the logic above is a bit more complicated in the presence of + LOCKED or PREEMPTED flags, but the main invariants + stay the same: + only RUNNING workers with servers assigned are allowed to run + in the userspace (unless LOCKED); + newly IDLE workers are added to the idle workers list; any + user-initiated state change assumes the userspace + properly removed the worker from the list; + as with wake detection, any "breach of contract" by the + userspace will result in the task termination via SIGKILL. + +* worker preemption: S:IDLE+W:RUNNING => S:RUNNING+W:IDLE|PREEMPTED: + when the userspace wants to preempt a RUNNING worker, it changes it + state, atomically, RUNNING => RUNNING|PREEMPTED and sends a + signal to the worker via tgkill(); the signal handler, previously + set up by the userspace, can be a NOP (note that only RUNNING + workers can be preempted); + + if the worker, at the moment the signal arrived, continued to be + running on-CPU in the userspace, the "wake detection" code will be + triggered that, in addition to what was described above, will + check if the worker is in RUNNING|PREEMPTED state: + W:RUNNING|PREEMPTED => W:IDLE|PREEMPTED + S := W.next_tid + S:IDLE => S:RUNNING + try_to_wakeup(S) + + if the signal arrives after the worker blocks in the kernel, + the "block detection" happened as described above, with the + following change: + W:RUNNING|PREEMPTED => W:BLOCKED|PREEMPTED + S := W.next_tid + S:IDLE => S:RUNNING + try_to_wake_up(S) + + in any case, the worker's server is woken, with its attached + worker (S.next_tid) either in BLOCKED|PREEMPTED or IDLE|PREEMPTED + state. + + +SERVER-ONLY USE CASES + +Some workloads/applications may benefit from fast and synchronous on-CPU +user-initiated context switches without the need for full userspace +scheduling (block/wake detection). These applications can use "standalone" +UMCG servers to wait/wake/context-switch, including across process +boundaries. + +These "worker-less" operations involve trivial RUNNING <==> IDLE state +changes, not discussed here for brevity. -- 2.25.1