taskq.c revision 5aeb94743e3be0c51e86f73096334611ae3a058e
/*
* CDDL HEADER START
*
* The contents of this file are subject to the terms of the
* Common Development and Distribution License (the "License").
* You may not use this file except in compliance with the License.
*
* You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
* See the License for the specific language governing permissions
* and limitations under the License.
*
* When distributing Covered Code, include this CDDL HEADER in each
* file and include the License file at usr/src/OPENSOLARIS.LICENSE.
* If applicable, add the following below this CDDL HEADER, with the
* fields enclosed by brackets "[]" replaced with your own identifying
* information: Portions Copyright [yyyy] [name of copyright owner]
*
* CDDL HEADER END
*/
/*
* Copyright 2010 Sun Microsystems, Inc. All rights reserved.
* Use is subject to license terms.
*/
/*
* Copyright 2011 Nexenta Systems, Inc. All rights reserved.
*/
/*
* Kernel task queues: general-purpose asynchronous task scheduling.
*
* A common problem in kernel programming is the need to schedule tasks
* to be performed later, by another thread. There are several reasons
* you may want or need to do this:
*
* (1) The task isn't time-critical, but your current code path is.
*
* (2) The task may require grabbing locks that you already hold.
*
* (3) The task may need to block (e.g. to wait for memory), but you
* cannot block in your current context.
*
* (4) Your code path can't complete because of some condition, but you can't
* sleep or fail, so you queue the task for later execution when condition
* disappears.
*
* (5) You just want a simple way to launch multiple tasks in parallel.
*
* Task queues provide such a facility. In its simplest form (used when
* performance is not a critical consideration) a task queue consists of a
* single list of tasks, together with one or more threads to service the
* list. There are some cases when this simple queue is not sufficient:
*
* (1) The task queues are very hot and there is a need to avoid data and lock
* contention over global resources.
*
* (2) Some tasks may depend on other tasks to complete, so they can't be put in
* the same list managed by the same thread.
*
* (3) Some tasks may block for a long time, and this should not block other
* tasks in the queue.
*
* To provide useful service in such cases we define a "dynamic task queue"
* which has an individual thread for each of the tasks. These threads are
* dynamically created as they are needed and destroyed when they are not in
* use. The API for managing task pools is the same as for managing task queues
* with the exception of a taskq creation flag TASKQ_DYNAMIC which tells that
* dynamic task pool behavior is desired.
*
* Dynamic task queues may also place tasks in the normal queue (called "backing
* queue") when task pool runs out of resources. Users of task queues may
* disallow such queued scheduling by specifying TQ_NOQUEUE in the dispatch
* flags.
*
* The backing task queue is also used for scheduling internal tasks needed for
* dynamic task queue maintenance.
*
* INTERFACES ==================================================================
*
* taskq_t *taskq_create(name, nthreads, pri, minalloc, maxall, flags);
*
* Create a taskq with specified properties.
* Possible 'flags':
*
* TASKQ_DYNAMIC: Create task pool for task management. If this flag is
* specified, 'nthreads' specifies the maximum number of threads in
* the task queue. Task execution order for dynamic task queues is
* not predictable.
*
* If this flag is not specified (default case) a
* single-list task queue is created with 'nthreads' threads
* servicing it. Entries in this queue are managed by
* taskq_ent_alloc() and taskq_ent_free() which try to keep the
* task population between 'minalloc' and 'maxalloc', but the
* latter limit is only advisory for TQ_SLEEP dispatches and the
* former limit is only advisory for TQ_NOALLOC dispatches. If
* TASKQ_PREPOPULATE is set in 'flags', the taskq will be
* prepopulated with 'minalloc' task structures.
*
* Since non-DYNAMIC taskqs are queues, tasks are guaranteed to be
* executed in the order they are scheduled if nthreads == 1.
* If nthreads > 1, task execution order is not predictable.
*
* TASKQ_PREPOPULATE: Prepopulate task queue with threads.
* Also prepopulate the task queue with 'minalloc' task structures.
*
* TASKQ_THREADS_CPU_PCT: This flag specifies that 'nthreads' should be
* interpreted as a percentage of the # of online CPUs on the
* system. The taskq subsystem will automatically adjust the
* number of threads in the taskq in response to CPU online
* and offline events, to keep the ratio. nthreads must be in
* the range [0,100].
*
* The calculation used is:
*
* MAX((ncpus_online * percentage)/100, 1)
*
* This flag is not supported for DYNAMIC task queues.
* This flag is not compatible with TASKQ_CPR_SAFE.
*
* TASKQ_CPR_SAFE: This flag specifies that users of the task queue will
* use their own protocol for handling CPR issues. This flag is not
* supported for DYNAMIC task queues. This flag is not compatible
* with TASKQ_THREADS_CPU_PCT.
*
* The 'pri' field specifies the default priority for the threads that
* service all scheduled tasks.
*
* taskq_t *taskq_create_instance(name, instance, nthreads, pri, minalloc,
* maxall, flags);
*
* Like taskq_create(), but takes an instance number (or -1 to indicate
* no instance).
*
* taskq_t *taskq_create_proc(name, nthreads, pri, minalloc, maxall, proc,
* flags);
*
* Like taskq_create(), but creates the taskq threads in the specified
* system process. If proc != &p0, this must be called from a thread
* in that process.
*
* taskq_t *taskq_create_sysdc(name, nthreads, minalloc, maxall, proc,
* dc, flags);
*
* Like taskq_create_proc(), but the taskq threads will use the
* System Duty Cycle (SDC) scheduling class with a duty cycle of dc.
*
* void taskq_destroy(tap):
*
* Waits for any scheduled tasks to complete, then destroys the taskq.
* Caller should guarantee that no new tasks are scheduled in the closing
* taskq.
*
* taskqid_t taskq_dispatch(tq, func, arg, flags):
*
* Dispatches the task "func(arg)" to taskq. The 'flags' indicates whether
* the caller is willing to block for memory. The function returns an
* opaque value which is zero iff dispatch fails. If flags is TQ_NOSLEEP
* or TQ_NOALLOC and the task can't be dispatched, taskq_dispatch() fails
* and returns (taskqid_t)0.
*
* ASSUMES: func != NULL.
*
* Possible flags:
* TQ_NOSLEEP: Do not wait for resources; may fail.
*
* TQ_NOALLOC: Do not allocate memory; may fail. May only be used with
* non-dynamic task queues.
*
* TQ_NOQUEUE: Do not enqueue a task if it can't dispatch it due to
* lack of available resources and fail. If this flag is not
* set, and the task pool is exhausted, the task may be scheduled
* in the backing queue. This flag may ONLY be used with dynamic
* task queues.
*
* NOTE: This flag should always be used when a task queue is used
* for tasks that may depend on each other for completion.
* Enqueueing dependent tasks may create deadlocks.
*
* TQ_SLEEP: May block waiting for resources. May still fail for
* dynamic task queues if TQ_NOQUEUE is also specified, otherwise
* always succeed.
*
* TQ_FRONT: Puts the new task at the front of the queue. Be careful.
*
* NOTE: Dynamic task queues are much more likely to fail in
* taskq_dispatch() (especially if TQ_NOQUEUE was specified), so it
* is important to have backup strategies handling such failures.
*
* void taskq_dispatch_ent(tq, func, arg, flags, tqent)
*
* This is a light-weight form of taskq_dispatch(), that uses a
* preallocated taskq_ent_t structure for scheduling. As a
* result, it does not perform allocations and cannot ever fail.
* Note especially that it cannot be used with TASKQ_DYNAMIC
* taskqs. The memory for the tqent must not be modified or used
* until the function (func) is called. (However, func itself
* may safely modify or free this memory, once it is called.)
* Note that the taskq framework will NOT free this memory.
*
* void taskq_wait(tq):
*
* Waits for all previously scheduled tasks to complete.
*
* NOTE: It does not stop any new task dispatches.
* Do NOT call taskq_wait() from a task: it will cause deadlock.
*
* void taskq_suspend(tq)
*
* Suspend all task execution. Tasks already scheduled for a dynamic task
* queue will still be executed, but all new scheduled tasks will be
* suspended until taskq_resume() is called.
*
* int taskq_suspended(tq)
*
* Returns 1 if taskq is suspended and 0 otherwise. It is intended to
* ASSERT that the task queue is suspended.
*
* void taskq_resume(tq)
*
* Resume task queue execution.
*
* int taskq_member(tq, thread)
*
* Returns 1 if 'thread' belongs to taskq 'tq' and 0 otherwise. The
* intended use is to ASSERT that a given function is called in taskq
* context only.
*
* system_taskq
*
* Global system-wide dynamic task queue for common uses. It may be used by
* any subsystem that needs to schedule tasks and does not need to manage
* its own task queues. It is initialized quite early during system boot.
*
* IMPLEMENTATION ==============================================================
*
* This is schematic representation of the task queue structures.
*
* taskq:
* +-------------+
* | tq_lock | +---< taskq_ent_free()
* +-------------+ |
* |... | | tqent: tqent:
* +-------------+ | +------------+ +------------+
* | tq_freelist |-->| tqent_next |--> ... ->| tqent_next |
* +-------------+ +------------+ +------------+
* |... | | ... | | ... |
* +-------------+ +------------+ +------------+
* | tq_task | |
* | | +-------------->taskq_ent_alloc()
* +--------------------------------------------------------------------------+
* | | | tqent tqent |
* | +---------------------+ +--> +------------+ +--> +------------+ |
* | | ... | | | func, arg | | | func, arg | |
* +>+---------------------+ <---|-+ +------------+ <---|-+ +------------+ |
* | tq_taskq.tqent_next | ----+ | | tqent_next | --->+ | | tqent_next |--+
* +---------------------+ | +------------+ ^ | +------------+
* +-| tq_task.tqent_prev | +--| tqent_prev | | +--| tqent_prev | ^
* | +---------------------+ +------------+ | +------------+ |
* | |... | | ... | | | ... | |
* | +---------------------+ +------------+ | +------------+ |
* | ^ | |
* | | | |
* +--------------------------------------+--------------+ TQ_APPEND() -+
* | | |
* |... | taskq_thread()-----+
* +-------------+
* | tq_buckets |--+-------> [ NULL ] (for regular task queues)
* +-------------+ |
* | DYNAMIC TASK QUEUES:
* |
* +-> taskq_bucket[nCPU] taskq_bucket_dispatch()
* +-------------------+ ^
* +--->| tqbucket_lock | |
* | +-------------------+ +--------+ +--------+
* | | tqbucket_freelist |-->| tqent |-->...| tqent | ^
* | +-------------------+<--+--------+<--...+--------+ |
* | | ... | | thread | | thread | |
* | +-------------------+ +--------+ +--------+ |
* | +-------------------+ |
* taskq_dispatch()--+--->| tqbucket_lock | TQ_APPEND()------+
* TQ_HASH() | +-------------------+ +--------+ +--------+
* | | tqbucket_freelist |-->| tqent |-->...| tqent |
* | +-------------------+<--+--------+<--...+--------+
* | | ... | | thread | | thread |
* | +-------------------+ +--------+ +--------+
* +---> ...
*
*
* Task queues use tq_task field to link new entry in the queue. The queue is a
* circular doubly-linked list. Entries are put in the end of the list with
* TQ_APPEND() and processed from the front of the list by taskq_thread() in
* FIFO order. Task queue entries are cached in the free list managed by
* taskq_ent_alloc() and taskq_ent_free() functions.
*
* All threads used by task queues mark t_taskq field of the thread to
* point to the task queue.
*
* Taskq Thread Management -----------------------------------------------------
*
* Taskq's non-dynamic threads are managed with several variables and flags:
*
* * tq_nthreads - The number of threads in taskq_thread() for the
* taskq.
*
* * tq_active - The number of threads not waiting on a CV in
* taskq_thread(); includes newly created threads
* not yet counted in tq_nthreads.
*
* * tq_nthreads_target
* - The number of threads desired for the taskq.
*
* * tq_flags & TASKQ_CHANGING
* - Indicates that tq_nthreads != tq_nthreads_target.
*
* * tq_flags & TASKQ_THREAD_CREATED
* - Indicates that a thread is being created in the taskq.
*
* During creation, tq_nthreads and tq_active are set to 0, and
* tq_nthreads_target is set to the number of threads desired. The
* TASKQ_CHANGING flag is set, and taskq_thread_create() is called to
* create the first thread. taskq_thread_create() increments tq_active,
* sets TASKQ_THREAD_CREATED, and creates the new thread.
*
* Each thread starts in taskq_thread(), clears the TASKQ_THREAD_CREATED
* flag, and increments tq_nthreads. It stores the new value of
* tq_nthreads as its "thread_id", and stores its thread pointer in the
* tq_threadlist at the (thread_id - 1). We keep the thread_id space
* densely packed by requiring that only the largest thread_id can exit during
* normal adjustment. The exception is during the destruction of the
* taskq; once tq_nthreads_target is set to zero, no new threads will be created
* for the taskq queue, so every thread can exit without any ordering being
* necessary.
*
* Threads will only process work if their thread id is <= tq_nthreads_target.
*
* When TASKQ_CHANGING is set, threads will check the current thread target
* whenever they wake up, and do whatever they can to apply its effects.
*
* TASKQ_THREAD_CPU_PCT --------------------------------------------------------
*
* When a taskq is created with TASKQ_THREAD_CPU_PCT, we store their requested
* percentage in tq_threads_ncpus_pct, start them off with the correct thread
* target, and add them to the taskq_cpupct_list for later adjustment.
*
* We register taskq_cpu_setup() to be called whenever a CPU changes state. It
* walks the list of TASKQ_THREAD_CPU_PCT taskqs, adjusts their nthread_target
* if need be, and wakes up all of the threads to process the change.
*
* Dynamic Task Queues Implementation ------------------------------------------
*
* For a dynamic task queues there is a 1-to-1 mapping between a thread and
* taskq_ent_structure. Each entry is serviced by its own thread and each thread
* is controlled by a single entry.
*
* Entries are distributed over a set of buckets. To avoid using modulo
* arithmetics the number of buckets is 2^n and is determined as the nearest
* power of two roundown of the number of CPUs in the system. Tunable
* variable 'taskq_maxbuckets' limits the maximum number of buckets. Each entry
* is attached to a bucket for its lifetime and can't migrate to other buckets.
*
* Entries that have scheduled tasks are not placed in any list. The dispatch
* function sets their "func" and "arg" fields and signals the corresponding
* thread to execute the task. Once the thread executes the task it clears the
* "func" field and places an entry on the bucket cache of free entries pointed
* by "tqbucket_freelist" field. ALL entries on the free list should have "func"
* field equal to NULL. The free list is a circular doubly-linked list identical
* in structure to the tq_task list above, but entries are taken from it in LIFO
* order - the last freed entry is the first to be allocated. The
* taskq_bucket_dispatch() function gets the most recently used entry from the
* free list, sets its "func" and "arg" fields and signals a worker thread.
*
* After executing each task a per-entry thread taskq_d_thread() places its
* entry on the bucket free list and goes to a timed sleep. If it wakes up
* without getting new task it removes the entry from the free list and destroys
* itself. The thread sleep time is controlled by a tunable variable
* `taskq_thread_timeout'.
*
* There are various statistics kept in the bucket which allows for later
* analysis of taskq usage patterns. Also, a global copy of taskq creation and
* death statistics is kept in the global taskq data structure. Since thread
* creation and death happen rarely, updating such global data does not present
* a performance problem.
*
* NOTE: Threads are not bound to any CPU and there is absolutely no association
* between the bucket and actual thread CPU, so buckets are used only to
* split resources and reduce resource contention. Having threads attached
* to the CPU denoted by a bucket may reduce number of times the job
* switches between CPUs.
*
* Current algorithm creates a thread whenever a bucket has no free
* entries. It would be nice to know how many threads are in the running
* state and don't create threads if all CPUs are busy with existing
* tasks, but it is unclear how such strategy can be implemented.
*
* Currently buckets are created statically as an array attached to task
* queue. On some system with nCPUs < max_ncpus it may waste system
* memory. One solution may be allocation of buckets when they are first
* touched, but it is not clear how useful it is.
*
*
* Before executing a task taskq_thread() (executing non-dynamic task
* queues) obtains taskq's thread lock as a reader. The taskq_suspend()
* function gets the same lock as a writer blocking all non-dynamic task
* execution. The taskq_resume() function releases the lock allowing
* taskq_thread to continue execution.
*
* For dynamic task queues, each bucket is marked as TQBUCKET_SUSPEND by
* taskq_suspend() function. After that taskq_bucket_dispatch() always
* fails, so that taskq_dispatch() will either enqueue tasks for a
* suspended backing queue or fail if TQ_NOQUEUE is specified in dispatch
* flags.
*
* NOTE: taskq_suspend() does not immediately block any tasks already
* scheduled for dynamic task queues. It only suspends new tasks
* scheduled after taskq_suspend() was called.
*
* taskq_member() function works by comparing a thread t_taskq pointer with
* the passed thread pointer.
*
* LOCKS and LOCK Hierarchy ----------------------------------------------------
*
* There are three locks used in task queues:
*
* 1) The taskq_t's tq_lock, protecting global task queue state.
*
* 2) Each per-CPU bucket has a lock for bucket management.
*
* 3) The global taskq_cpupct_lock, which protects the list of
* TASKQ_THREADS_CPU_PCT taskqs.
*
* If both (1) and (2) are needed, tq_lock should be taken *after* the bucket
* lock.
*
* If both (1) and (3) are needed, tq_lock should be taken *after*
* taskq_cpupct_lock.
*
* DEBUG FACILITIES ------------------------------------------------------------
*
* For DEBUG kernels it is possible to induce random failures to
* taskq_dispatch() function when it is given TQ_NOSLEEP argument. The value of
* taskq_dmtbf and taskq_smtbf tunables control the mean time between induced
* failures for dynamic and static task queues respectively.
*
* Setting TASKQ_STATISTIC to 0 will disable per-bucket statistics.
*
* TUNABLES --------------------------------------------------------------------
*
* system_taskq_size - Size of the global system_taskq.
* This value is multiplied by nCPUs to determine
* actual size.
* Default value: 64
*
* taskq_minimum_nthreads_max
* - Minimum size of the thread list for a taskq.
* Useful for testing different thread pool
* sizes by overwriting tq_nthreads_target.
*
* taskq_thread_timeout - Maximum idle time for taskq_d_thread()
* Default value: 5 minutes
*
* taskq_maxbuckets - Maximum number of buckets in any task queue
* Default value: 128
*
* taskq_search_depth - Maximum # of buckets searched for a free entry
* Default value: 4
*
* taskq_dmtbf - Mean time between induced dispatch failures
* for dynamic task queues.
* Default value: UINT_MAX (no induced failures)
*
* taskq_smtbf - Mean time between induced dispatch failures
* for static task queues.
* Default value: UINT_MAX (no induced failures)
*
* CONDITIONAL compilation -----------------------------------------------------
*
* TASKQ_STATISTIC - If set will enable bucket statistic (default).
*
*/
#include <sys/taskq_impl.h>
#include <sys/sysmacros.h>
/*
* Pseudo instance numbers for taskqs without explicitly provided instance.
*/
static vmem_t *taskq_id_arena;
/* Global system task queue for common use */
/*
* Maximum number of entries in global system taskq is
* system_taskq_size * max_ncpus
*/
#define SYSTEM_TASKQ_SIZE 64
/*
* Minimum size for tq_nthreads_max; useful for those who want to play around
* with increasing a taskq's tq_nthreads_target.
*/
int taskq_minimum_nthreads_max = 1;
/*
* We want to ensure that when taskq_create() returns, there is at least
* one thread ready to handle requests. To guarantee this, we have to wait
* for the second thread, since the first one cannot process requests until
* the second thread has been created.
*/
#define TASKQ_CREATE_ACTIVE_THREADS 2
/* Maximum percentage allowed for TASKQ_THREADS_CPU_PCT */
#define TASKQ_CPUPCT_MAX_PERCENT 1000
/*
* Dynamic task queue threads that don't get any work within
* taskq_thread_timeout destroy themselves
*/
#define TASKQ_MAXBUCKETS 128
int taskq_maxbuckets = TASKQ_MAXBUCKETS;
/*
* When a bucket has no available entries another buckets are tried.
* taskq_search_depth parameter limits the amount of buckets that we search
* before failing. This is mostly useful in systems with many CPUs where we may
* spend too much time scanning busy buckets.
*/
#define TASKQ_SEARCH_DEPTH 4
/*
* Hashing function: mix various bits of x. May be pretty much anything.
*/
/*
* We do not create any new threads when the system is low on memory and start
* throttling memory allocations. The following macro tries to estimate such
* condition.
*/
/*
* Static functions.
*/
static void taskq_thread(void *);
static void taskq_d_thread(taskq_ent_t *);
static void taskq_bucket_extend(void *);
static int taskq_constructor(void *, void *, int);
static void taskq_destructor(void *, void *);
static int taskq_ent_constructor(void *, void *, int);
static void taskq_ent_destructor(void *, void *);
void *);
/*
* Task queues kstats.
*/
struct taskq_kstat {
} taskq_kstat = {
{ "pid", KSTAT_DATA_UINT64 },
{ "tasks", KSTAT_DATA_UINT64 },
{ "executed", KSTAT_DATA_UINT64 },
{ "maxtasks", KSTAT_DATA_UINT64 },
{ "totaltime", KSTAT_DATA_UINT64 },
{ "nactive", KSTAT_DATA_UINT64 },
{ "nalloc", KSTAT_DATA_UINT64 },
{ "priority", KSTAT_DATA_UINT64 },
{ "threads", KSTAT_DATA_UINT64 },
};
struct taskq_d_kstat {
} taskq_d_kstat = {
{ "priority", KSTAT_DATA_UINT64 },
{ "btasks", KSTAT_DATA_UINT64 },
{ "bexecuted", KSTAT_DATA_UINT64 },
{ "bmaxtasks", KSTAT_DATA_UINT64 },
{ "bnalloc", KSTAT_DATA_UINT64 },
{ "bnactive", KSTAT_DATA_UINT64 },
{ "btotaltime", KSTAT_DATA_UINT64 },
{ "hits", KSTAT_DATA_UINT64 },
{ "misses", KSTAT_DATA_UINT64 },
{ "overflows", KSTAT_DATA_UINT64 },
{ "tcreates", KSTAT_DATA_UINT64 },
{ "tdeaths", KSTAT_DATA_UINT64 },
{ "maxthreads", KSTAT_DATA_UINT64 },
{ "nomem", KSTAT_DATA_UINT64 },
{ "disptcreates", KSTAT_DATA_UINT64 },
{ "totaltime", KSTAT_DATA_UINT64 },
{ "nalloc", KSTAT_DATA_UINT64 },
{ "nfree", KSTAT_DATA_UINT64 },
};
static kmutex_t taskq_kstat_lock;
static kmutex_t taskq_d_kstat_lock;
static int taskq_kstat_update(kstat_t *, int);
static int taskq_d_kstat_update(kstat_t *, int);
/*
* List of all TASKQ_THREADS_CPU_PCT taskqs.
*/
/*
* Collect per-bucket statistic when TASKQ_STATISTIC is defined.
*/
#define TASKQ_STATISTIC 1
#if TASKQ_STATISTIC
#define TQ_STAT(b, x) b->tqbucket_stat.x++
#else
#define TQ_STAT(b, x)
#endif
/*
* Random fault injection.
*/
/*
* TQ_NOSLEEP dispatches on dynamic task queues are always allowed to fail.
*
* TQ_NOSLEEP dispatches on static task queues can't arbitrarily fail because
* they could prepopulate the cache and make sure that they do not use more
* then minalloc entries. So, fault injection in this case insures that
* either TASKQ_PREPOPULATE is not set or there are more entries allocated
* than is specified by minalloc. TQ_NOALLOC dispatches are always allowed
* to fail, but for simplicity we treat them identically to TQ_NOSLEEP
* dispatches.
*/
#ifdef DEBUG
if ((flag & TQ_NOSLEEP) && \
return (NULL); \
}
return (NULL); \
}
#else
#endif
((l).tqent_prev == &(l)))
/*
* Append `tqe' in the end of the doubly-linked list denoted by l.
*/
tqe->tqent_next = &l; \
}
/*
* Prepend 'tqe' to the beginning of l
*/
#define TQ_PREPEND(l, tqe) { \
tqe->tqent_prev = &l; \
}
/*
* Schedule a task specified by func and arg into the task queue entry tqe.
*/
if (front) { \
} else { \
} \
}
/*
* Do-nothing task which may be used to prepopulate thread caches.
*/
/*ARGSUSED*/
void
{
}
/*ARGSUSED*/
static int
{
return (0);
}
/*ARGSUSED*/
static void
{
}
/*ARGSUSED*/
static int
{
return (0);
}
/*ARGSUSED*/
static void
{
}
void
taskq_init(void)
{
sizeof (taskq_ent_t), 0, taskq_ent_constructor,
}
static void
{
/* We must be going from non-zero to non-zero; no exiting. */
}
}
/* called during task queue creation */
static void
{
}
static void
{
}
/*ARGSUSED*/
static int
{
switch (what) {
case CPU_OFF:
case CPU_CPUPART_OUT:
/* offlines are called *before* the cpu is offlined. */
if (ncpus > 1)
ncpus--;
break;
case CPU_ON:
case CPU_CPUPART_IN:
break;
default:
return (0); /* doesn't affect cpu count */
}
/*
* If the taskq is part of the cpuset which is changing,
* update its nthreads_target.
*/
}
}
return (0);
}
void
taskq_mp_init(void)
{
/*
* Make sure we're up to date. At this point in boot, there is only
* one processor set, so we only have to update the current CPU.
*/
}
/*
* Create global system dynamic task queue.
*/
void
system_taskq_init(void)
{
}
/*
* taskq_ent_alloc()
*
* Allocates a new taskq_ent_t structure either from the free list or from the
* cache. Returns NULL if it can't be allocated.
*
* Assumes: tq->tq_lock is held.
*/
static taskq_ent_t *
{
/*
* TQ_NOALLOC allocations are allowed to use the freelist, even if
* we are below tq_minalloc.
*/
} else {
if (flags & TQ_NOALLOC)
return (NULL);
if (kmflags & KM_NOSLEEP)
return (NULL);
/*
* We don't want to exceed tq_maxalloc, but we can't
* wait for other tasks to complete (and thus free up
* task structures) without risking deadlock with
* the caller. So, we just delay for one second
* to throttle the allocation rate. If we have tasks
* complete before one second timeout expires then
* taskq_ent_free will signal us and we will
* immediately retry the allocation (reap free).
*/
tq->tq_maxalloc_wait++;
tq->tq_maxalloc_wait--;
if (wait_rv == -1)
break;
}
if (tq->tq_freelist)
goto again; /* reap freelist */
}
}
return (tqe);
}
/*
* taskq_ent_free()
*
* Free taskq_ent_t structure by either putting it on the free list or freeing
* it to the cache.
*
* Assumes: tq->tq_lock is held.
*/
static void
{
} else {
}
if (tq->tq_maxalloc_wait)
}
/*
* taskq_ent_exists()
*
* Return 1 if taskq already has entry for calling 'func(arg)'.
*
* Assumes: tq->tq_lock is held.
*/
static int
{
return (1);
return (0);
}
/*
* Dispatch a task "func(arg)" to a free entry of bucket b.
*
* Assumes: no bucket locks is held.
*
* Returns: a pointer to an entry if dispatch was successful.
* NULL if there are no free entries or if the bucket is suspended.
*/
static taskq_ent_t *
{
mutex_enter(&b->tqbucket_lock);
/*
* Get en entry from the freelist if there is one.
* Schedule task into the entry.
*/
if ((b->tqbucket_nfree != 0) &&
!(b->tqbucket_flags & TQBUCKET_SUSPEND)) {
b->tqbucket_nalloc++;
b->tqbucket_nfree--;
taskq_ent_t *, tqe);
} else {
TQ_STAT(b, tqs_misses);
}
mutex_exit(&b->tqbucket_lock);
return (tqe);
}
/*
* Dispatch a task.
*
* Assumes: func != NULL
*
* Returns: NULL if dispatch failed.
* non-NULL if task dispatched successfully.
* Actual return value is the pointer to taskq entry that was used to
* dispatch a task. This is useful for debugging.
*/
{
/*
* TQ_NOQUEUE flag can't be used with non-dynamic task queues.
*/
/*
* Enqueue the task to the underlying queue.
*/
return (NULL);
}
/* Make sure we start without any flags */
} else {
}
}
/*
* Dynamic taskq dispatching.
*/
if (bsize == 1) {
/*
* In a single-CPU case there is only one bucket, so get
* entry directly from there.
*/
!= NULL)
} else {
int loopcount;
taskq_bucket_t *b;
h = TQ_HASH(h);
/*
* The 'bucket' points to the original bucket that we hit. If we
* can't allocate from it, we search other buckets, but only
* extend this one.
*/
/*
* Do a quick check before grabbing the lock. If the bucket does
* not have free entries now, chances are very small that it
* will after we take the lock, so we just skip it.
*/
if (b->tqbucket_nfree != 0) {
} else {
TQ_STAT(b, tqs_misses);
}
bucket = b;
/*
* If bucket dispatch failed, search loopcount number of buckets
* before we give up and fail.
*/
do {
loopcount--;
if (b->tqbucket_nfree != 0) {
} else {
TQ_STAT(b, tqs_misses);
}
}
/*
* At this point we either scheduled a task and (tqe != NULL) or failed
* (tqe == NULL). Try to recover from fails.
*/
/*
* For KM_SLEEP dispatches, try to extend the bucket and retry dispatch.
*/
/*
* taskq_bucket_extend() may fail to do anything, but this is
* fine - we deal with it later. If the bucket was successfully
* extended, there is a good chance that taskq_bucket_dispatch()
* will get this new entry, unless someone is racing with us and
* stealing the new entry from under our nose.
* taskq_bucket_extend() may sleep.
*/
}
/*
* Since there are not enough free entries in the bucket, add a
* taskq entry to extend it in the background using backing queue
* (unless we already have a taskq entry to perform that extension).
*/
} else {
}
}
/*
* Dispatch failed and we can't find an entry to schedule a task.
* Revert to the backing queue unless TQ_NOQUEUE was asked.
*/
} else {
}
}
}
void
{
/*
* Mark it as a prealloc'd task. This is important
* to ensure that we don't free it later.
*/
/*
* Enqueue the task to the underlying queue.
*/
} else {
}
}
/*
* Wait for all pending tasks to complete.
* Calling taskq_wait from a task will cause deadlock.
*/
void
{
int bid = 0;
mutex_enter(&b->tqbucket_lock);
while (b->tqbucket_nalloc > 0)
mutex_exit(&b->tqbucket_lock);
}
}
}
/*
* Suspend execution of tasks.
*
* Tasks in the queue part will be suspended immediately upon return from this
* function. Pending tasks in the dynamic part will continue to execute, but all
* new tasks will be suspended.
*/
void
{
int bid = 0;
mutex_enter(&b->tqbucket_lock);
b->tqbucket_flags |= TQBUCKET_SUSPEND;
mutex_exit(&b->tqbucket_lock);
}
}
/*
* Mark task queue as being suspended. Needed for taskq_suspended().
*/
}
/*
* returns: 1 if tq is suspended, 0 otherwise.
*/
int
{
}
/*
* Resume taskq execution.
*/
void
{
int bid = 0;
mutex_enter(&b->tqbucket_lock);
b->tqbucket_flags &= ~TQBUCKET_SUSPEND;
mutex_exit(&b->tqbucket_lock);
}
}
}
int
{
}
/*
* Creates a thread in the taskq. We only allow one outstanding create at
* a time. We drop and reacquire the tq_lock in order to avoid blocking other
* taskq activity while thread_create() or lwp_kernel_create() run.
*
* The first time we're called, we do some additional setup, and do not
* return until there are enough threads to start servicing requests.
*/
static void
{
kthread_t *t;
} else {
}
if (!first) {
return;
}
/*
* We know the thread cannot go away, since tq cannot be
* destroyed until creation has completed. We can therefore
* safely dereference t.
*/
}
/* Wait until we can service requests. */
}
}
/*
* Common "sleep taskq thread" function, which handles CPR stuff, as well
* as giving a nice common point for debuggers to find inactive threads.
*/
static clock_t
{
}
if (timeout < 0)
else
}
return (ret);
}
/*
* Worker thread for processing task queue.
*/
static void
taskq_thread(void *arg)
{
int thread_id;
}
} else {
}
else
/* Allow taskq_create_common()'s taskq_thread_create() to return. */
for (;;) {
/* See if we're no longer needed */
/*
* To preserve the one-to-one mapping between
* thread_id and thread, we must exit from
* highest thread ID to least.
*
* However, if everyone is exiting, the order
* doesn't matter, so just exit immediately.
* (this is safe, since you must wait for
* nthreads to reach 0 after setting
* tq_nthreads_target to 0)
*/
tq->tq_nthreads_target == 0)
break;
/* Wait for higher thread_ids to exit */
continue;
}
/*
* If no thread is starting taskq_thread(), we can
* do some bookkeeping.
*/
/* Check if we've reached our target */
}
/* Check if we need to create a thread */
continue; /* tq_lock was dropped */
}
}
}
continue;
}
/*
* For prealloc'd tasks, we don't free anything. We
* have to check this now, because once we call the
* function for a prealloc'd taskq, we can't touch the
* tqent any longer (calling the function returns the
* ownershp of the tqent back to caller of
* taskq_dispatch.)
*/
/* clear pointers to assist assertion checks */
} else {
}
taskq_ent_t *, tqe);
taskq_ent_t *, tqe);
tq->tq_executed++;
if (freeit)
}
else
/* We're exiting, and therefore no longer active */
tq->tq_nthreads--;
/* Wake up anyone waiting for us to exit */
}
lwp_exit();
} else {
thread_exit();
}
}
/*
* Worker per-entry thread for dynamic dispatches.
*/
static void
{
clock_t w;
for (;;) {
/*
* If a task is scheduled (func != NULL), execute it, otherwise
* sleep, waiting for a job.
*/
/*
* It is possible to free the entry right away before
* actually executing the task so that subsequent
* dispatches may immediately reuse it. But this,
* effectively, creates a two-length queue in the entry
* and may lead to a deadlock if the execution of the
* current task depends on the execution of the next
* scheduled task. So, we keep the entry busy until the
* task is processed.
*/
/*
* Return the entry to the bucket free list.
*/
bucket->tqbucket_nfree++;
/*
* taskq_wait() waits for nalloc to drop to zero on
* tqbucket_cv.
*/
}
/*
* At this point the entry must be in the bucket free list -
* either because it was there initially or because it just
* finished executing a task and put itself on the free list.
*/
/*
* Go to sleep unless we are closing.
* If a thread is sleeping too long, it dies.
*/
}
/*
* At this point we may be in two different states:
*
* (1) tqent_func is set which means that a new task is
* dispatched and we need to execute it.
*
* (2) Thread is sleeping for too long or we are closing. In
* both cases destroy the thread and the entry.
*/
/* If func is NULL we should be on the freelist. */
(bucket->tqbucket_nfree > 0));
/* If func is non-NULL we should be allocated */
(bucket->tqbucket_nalloc > 0));
/* Check freelist consistency */
/*
* This thread is sleeping for too long or we are
* closing - time to die.
* Thread creation/destruction happens rarely,
* so grabbing the lock is not a big performance issue.
* The bucket lock is dropped by CALLB_CPR_EXIT().
*/
/* Remove the entry from the free list. */
bucket->tqbucket_nfree--;
tq->tq_tdeaths++;
thread_exit();
}
}
}
/*
* Taskq creation. May sleep for memory.
* Always use automatically generated instances to avoid kstat name space
* collisions.
*/
taskq_t *
{
}
/*
* Create an instance of task queue. It is legal to create task queues with the
* same name and different instances.
*
* taskq_create_instance is used by ddi_taskq_create() where it gets the
* instance from ddi_get_instance(). In some cases the instance is not
* initialized and is set to -1. This case is handled as if no instance was
* passed at all.
*/
taskq_t *
{
if (instance < 0) {
}
}
taskq_t *
{
}
taskq_t *
{
}
static taskq_t *
{
int max_nthreads;
/*
* TASKQ_DYNAMIC, TASKQ_CPR_SAFE and TASKQ_THREADS_CPU_PCT are all
* mutually incompatible.
*/
/* Cannot have DUTY_CYCLE without a non-p0 kernel process */
/* Cannot have DC_BATCH without DUTY_CYCLE */
if (flags & TASKQ_DYNAMIC) {
/* For dynamic task queues use just one backup thread */
} else if (flags & TASKQ_THREADS_CPU_PCT) {
if (pct > taskq_cpupct_max_percent)
/*
* If you're using THREADS_CPU_PCT, the process for the
* taskq threads must be curproc. This allows any pset
* binding to be inherited correctly. If proc is &p0,
* we won't be creating LWPs, so new threads will be assigned
* to the default processor set.
*/
} else {
}
/*
* Make sure the name is 0-terminated, and conforms to the rules for
* C indentifiers
*/
if (max_nthreads > 1)
if (flags & TASKQ_PREPOPULATE) {
while (minalloc-- > 0)
}
/*
* Create the first thread, which will create any other threads
* necessary. taskq_thread_create will not return until we have
* enough threads to be able to process requests.
*/
if (flags & TASKQ_DYNAMIC) {
int b_id;
/* Initialize each bucket */
NULL);
if (flags & TASKQ_PREPOPULATE)
}
}
/*
* Install kstats.
* We have two cases:
* 1) Instance is provided to taskq_create_instance(). In this case it
* should be >= 0 and we use it.
*
* 2) Instance is not provided and is automatically generated
*/
if (flags & TASKQ_NOINSTANCE) {
}
if (flags & TASKQ_DYNAMIC) {
sizeof (taskq_d_kstat) / sizeof (kstat_named_t),
KSTAT_FLAG_VIRTUAL)) != NULL) {
}
} else {
"taskq", KSTAT_TYPE_NAMED,
sizeof (taskq_kstat) / sizeof (kstat_named_t),
KSTAT_FLAG_VIRTUAL)) != NULL) {
}
}
return (tq);
}
/*
* taskq_destroy().
*
* Assumes: by the time taskq_destroy is called no one will use this task queue
* in any way and no one will try to dispatch entries in it.
*/
void
{
int bid = 0;
/*
* Destroy kstats.
*/
}
/*
* Destroy instance if needed.
*/
1);
tq->tq_instance = 0;
}
/*
* Unregister from the cpupct list.
*/
}
/*
* Wait for any pending entries to complete.
*/
taskq_wait(tq);
/* notify all the threads that they need to exit */
tq->tq_nthreads_target = 0;
while (tq->tq_nthreads != 0)
tq->tq_minalloc = 0;
/*
* Mark each bucket as closing and wakeup all sleeping threads.
*/
mutex_enter(&b->tqbucket_lock);
b->tqbucket_flags |= TQBUCKET_CLOSE;
/* Wakeup all sleeping threads */
ASSERT(b->tqbucket_nalloc == 0);
/*
* At this point we waited for all pending jobs to complete (in
* both the task queue and the bucket and no new jobs should
* arrive. Wait for all threads to die.
*/
while (b->tqbucket_nfree > 0)
mutex_exit(&b->tqbucket_lock);
mutex_destroy(&b->tqbucket_lock);
cv_destroy(&b->tqbucket_cv);
}
/* Cleanup fields before returning tq to the cache */
tq->tq_tcreates = 0;
tq->tq_tdeaths = 0;
} else {
}
tq->tq_threads_ncpus_pct = 0;
tq->tq_totaltime = 0;
tq->tq_maxtasks = 0;
tq->tq_executed = 0;
}
/*
* Extend a bucket with a new entry on the free list and attach a worker thread
* to it.
*
* Argument: pointer to the bucket.
*
* This function may quietly fail. It is only used by taskq_dispatch() which
* handles such failures properly.
*/
static void
taskq_bucket_extend(void *arg)
{
int nthreads;
if (! ENOUGH_MEMORY()) {
return;
}
/*
* Observe global taskq limits on the number of threads.
*/
tq->tq_tcreates--;
return;
}
tq->tq_tcreates--;
return;
}
/*
* Create a thread in a TS_STOPPED state first. If it is successfully
* created, place the entry on the free list and start the thread.
*/
/*
* Once the entry is ready, link it to the the bucket free list.
*/
mutex_enter(&b->tqbucket_lock);
b->tqbucket_nfree++;
TQ_STAT(b, tqs_tcreates);
#if TASKQ_STATISTIC
#endif
mutex_exit(&b->tqbucket_lock);
/*
* Start the stopped thread.
*/
}
static int
{
if (rw == KSTAT_WRITE)
return (EACCES);
return (0);
}
static int
{
int bid = 0;
if (rw == KSTAT_WRITE)
return (EACCES);
}
return (0);
}