taskq_t *system_taskq;
EXPORT_SYMBOL(system_taskq);
-typedef struct taskq_ent {
- spinlock_t tqent_lock;
- struct list_head tqent_list;
- taskqid_t tqent_id;
- task_func_t *tqent_func;
- void *tqent_arg;
-} taskq_ent_t;
+static int
+task_km_flags(uint_t flags)
+{
+ if (flags & TQ_NOSLEEP)
+ return KM_NOSLEEP;
+
+ if (flags & TQ_PUSHPAGE)
+ return KM_PUSHPAGE;
+
+ return KM_SLEEP;
+}
/*
* NOTE: Must be called with tq->tq_lock held, returns a list_t which
SENTRY;
ASSERT(tq);
- ASSERT(flags & (TQ_SLEEP | TQ_NOSLEEP)); /* One set */
- ASSERT(!((flags & TQ_SLEEP) && (flags & TQ_NOSLEEP))); /* Not both */
ASSERT(spin_is_locked(&tq->tq_lock));
retry:
/* Acquire taskq_ent_t's from free list if available */
if (!list_empty(&tq->tq_free_list) && !(flags & TQ_NEW)) {
t = list_entry(tq->tq_free_list.next, taskq_ent_t, tqent_list);
+
+ ASSERT(!(t->tqent_flags & TQENT_FLAG_PREALLOC));
+
list_del_init(&t->tqent_list);
SRETURN(t);
}
}
spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
- t = kmem_alloc(sizeof(taskq_ent_t), flags & (TQ_SLEEP | TQ_NOSLEEP));
+ t = kmem_alloc(sizeof(taskq_ent_t), task_km_flags(flags));
spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
if (t) {
- spin_lock_init(&t->tqent_lock);
- INIT_LIST_HEAD(&t->tqent_list);
- t->tqent_id = 0;
- t->tqent_func = NULL;
- t->tqent_arg = NULL;
+ taskq_init_ent(t);
tq->tq_nalloc++;
}
t->tqent_id = 0;
t->tqent_func = NULL;
t->tqent_arg = NULL;
+ t->tqent_flags = 0;
+
list_add_tail(&t->tqent_list, &tq->tq_free_list);
} else {
task_free(tq, t);
* monotonically increasing taskqid and added to the tail of the pending
* list. As worker threads become available the tasks are removed from
* the head of the pending or priority list, giving preference to the
- * priority list. The tasks are then added to the work list, preserving
- * the ordering by taskqid. Finally, as tasks complete they are removed
- * from the work list. This means that the pending and work lists are
- * always kept sorted by taskqid. Thus the lowest outstanding
- * incomplete taskqid can be determined simply by checking the min
- * taskqid for each head item on the pending, priority, and work list.
- * This value is stored in tq->tq_lowest_id and only updated to the new
- * lowest id when the previous lowest id completes. All taskqids lower
- * than tq->tq_lowest_id must have completed. It is also possible
- * larger taskqid's have completed because they may be processed in
- * parallel by several worker threads. However, this is not a problem
- * because the behavior of taskq_wait_id() is to block until all
- * previously submitted taskqid's have completed.
+ * priority list. The tasks are then removed from their respective
+ * list, and the taskq_thread servicing the task is added to the active
+ * list, preserving the order using the serviced task's taskqid.
+ * Finally, as tasks complete the taskq_thread servicing the task is
+ * removed from the active list. This means that the pending task and
+ * active taskq_thread lists are always kept sorted by taskqid. Thus the
+ * lowest outstanding incomplete taskqid can be determined simply by
+ * checking the min taskqid for each head item on the pending, priority,
+ * and active taskq_thread list. This value is stored in
+ * tq->tq_lowest_id and only updated to the new lowest id when the
+ * previous lowest id completes. All taskqids lower than
+ * tq->tq_lowest_id must have completed. It is also possible larger
+ * taskqid's have completed because they may be processed in parallel by
+ * several worker threads. However, this is not a problem because the
+ * behavior of taskq_wait_id() is to block until all previously
+ * submitted taskqid's have completed.
*
* XXX: Taskqid_t wrapping is not handled. However, taskqid_t's are
* 64-bit values so even if a taskq is processing 2^24 (16,777,216)
int
__taskq_member(taskq_t *tq, void *t)
{
- int i;
+ struct list_head *l;
+ taskq_thread_t *tqt;
SENTRY;
ASSERT(tq);
ASSERT(t);
- for (i = 0; i < tq->tq_nthreads; i++)
- if (tq->tq_threads[i] == (struct task_struct *)t)
- SRETURN(1);
+ list_for_each(l, &tq->tq_thread_list) {
+ tqt = list_entry(l, taskq_thread_t, tqt_thread_list);
+ if (tqt->tqt_thread == (struct task_struct *)t)
+ SRETURN(1);
+ }
SRETURN(0);
}
ASSERT(tq);
ASSERT(func);
- /* Solaris assumes TQ_SLEEP if not passed explicitly */
- if (!(flags & (TQ_SLEEP | TQ_NOSLEEP)))
- flags |= TQ_SLEEP;
-
- if (unlikely(in_atomic() && (flags & TQ_SLEEP)))
- PANIC("May schedule while atomic: %s/0x%08x/%d\n",
- current->comm, preempt_count(), current->pid);
-
spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
/* Taskq being destroyed and all tasks drained */
tq->tq_next_id++;
t->tqent_func = func;
t->tqent_arg = arg;
+
+ ASSERT(!(t->tqent_flags & TQENT_FLAG_PREALLOC));
+
spin_unlock(&t->tqent_lock);
wake_up(&tq->tq_work_waitq);
SRETURN(rc);
}
EXPORT_SYMBOL(__taskq_dispatch);
+
+void
+__taskq_dispatch_ent(taskq_t *tq, task_func_t func, void *arg, uint_t flags,
+ taskq_ent_t *t)
+{
+ SENTRY;
+
+ ASSERT(tq);
+ ASSERT(func);
+ ASSERT(!(tq->tq_flags & TASKQ_DYNAMIC));
+
+ spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
+
+ /* Taskq being destroyed and all tasks drained */
+ if (!(tq->tq_flags & TQ_ACTIVE)) {
+ t->tqent_id = 0;
+ goto out;
+ }
+
+ spin_lock(&t->tqent_lock);
+
+ /*
+ * Mark it as a prealloc'd task. This is important
+ * to ensure that we don't free it later.
+ */
+ t->tqent_flags |= TQENT_FLAG_PREALLOC;
+
+ /* Queue to the priority list instead of the pending list */
+ if (flags & TQ_FRONT)
+ list_add_tail(&t->tqent_list, &tq->tq_prio_list);
+ else
+ list_add_tail(&t->tqent_list, &tq->tq_pend_list);
+
+ t->tqent_id = tq->tq_next_id;
+ tq->tq_next_id++;
+ t->tqent_func = func;
+ t->tqent_arg = arg;
+
+ spin_unlock(&t->tqent_lock);
+
+ wake_up(&tq->tq_work_waitq);
+out:
+ spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
+ SEXIT;
+}
+EXPORT_SYMBOL(__taskq_dispatch_ent);
+
+int
+__taskq_empty_ent(taskq_ent_t *t)
+{
+ return list_empty(&t->tqent_list);
+}
+EXPORT_SYMBOL(__taskq_empty_ent);
+
+void
+__taskq_init_ent(taskq_ent_t *t)
+{
+ spin_lock_init(&t->tqent_lock);
+ INIT_LIST_HEAD(&t->tqent_list);
+ t->tqent_id = 0;
+ t->tqent_func = NULL;
+ t->tqent_arg = NULL;
+ t->tqent_flags = 0;
+}
+EXPORT_SYMBOL(__taskq_init_ent);
+
/*
* Returns the lowest incomplete taskqid_t. The taskqid_t may
* be queued on the pending list, on the priority list, or on
{
taskqid_t lowest_id = tq->tq_next_id;
taskq_ent_t *t;
+ taskq_thread_t *tqt;
SENTRY;
ASSERT(tq);
lowest_id = MIN(lowest_id, t->tqent_id);
}
- if (!list_empty(&tq->tq_work_list)) {
- t = list_entry(tq->tq_work_list.next, taskq_ent_t, tqent_list);
- lowest_id = MIN(lowest_id, t->tqent_id);
+ if (!list_empty(&tq->tq_active_list)) {
+ tqt = list_entry(tq->tq_active_list.next, taskq_thread_t,
+ tqt_active_list);
+ ASSERT(tqt->tqt_id != 0);
+ lowest_id = MIN(lowest_id, tqt->tqt_id);
}
SRETURN(lowest_id);
* taskqid.
*/
static void
-taskq_insert_in_order(taskq_t *tq, taskq_ent_t *t)
+taskq_insert_in_order(taskq_t *tq, taskq_thread_t *tqt)
{
- taskq_ent_t *w;
+ taskq_thread_t *w;
struct list_head *l;
SENTRY;
ASSERT(tq);
- ASSERT(t);
+ ASSERT(tqt);
ASSERT(spin_is_locked(&tq->tq_lock));
- list_for_each_prev(l, &tq->tq_work_list) {
- w = list_entry(l, taskq_ent_t, tqent_list);
- if (w->tqent_id < t->tqent_id) {
- list_add(&t->tqent_list, l);
+ list_for_each_prev(l, &tq->tq_active_list) {
+ w = list_entry(l, taskq_thread_t, tqt_active_list);
+ if (w->tqt_id < tqt->tqt_id) {
+ list_add(&tqt->tqt_active_list, l);
break;
}
}
- if (l == &tq->tq_work_list)
- list_add(&t->tqent_list, &tq->tq_work_list);
+ if (l == &tq->tq_active_list)
+ list_add(&tqt->tqt_active_list, &tq->tq_active_list);
SEXIT;
}
{
DECLARE_WAITQUEUE(wait, current);
sigset_t blocked;
- taskqid_t id;
- taskq_t *tq = args;
+ taskq_thread_t *tqt = args;
+ taskq_t *tq;
taskq_ent_t *t;
struct list_head *pend_list;
SENTRY;
- ASSERT(tq);
+ ASSERT(tqt);
+ tq = tqt->tqt_tq;
current->flags |= PF_NOFREEZE;
- /* Disable the direct memory reclaim path */
- if (tq->tq_flags & TASKQ_NORECLAIM)
- current->flags |= PF_MEMALLOC;
-
sigfillset(&blocked);
sigprocmask(SIG_BLOCK, &blocked, NULL);
flush_signals(current);
while (!kthread_should_stop()) {
- add_wait_queue(&tq->tq_work_waitq, &wait);
if (list_empty(&tq->tq_pend_list) &&
list_empty(&tq->tq_prio_list)) {
+ add_wait_queue_exclusive(&tq->tq_work_waitq, &wait);
spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
schedule();
spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
+ remove_wait_queue(&tq->tq_work_waitq, &wait);
} else {
__set_current_state(TASK_RUNNING);
}
- remove_wait_queue(&tq->tq_work_waitq, &wait);
if (!list_empty(&tq->tq_prio_list))
pend_list = &tq->tq_prio_list;
if (pend_list) {
t = list_entry(pend_list->next, taskq_ent_t, tqent_list);
list_del_init(&t->tqent_list);
- taskq_insert_in_order(tq, t);
+
+ /* In order to support recursively dispatching a
+ * preallocated taskq_ent_t, tqent_id must be
+ * stored prior to executing tqent_func. */
+ tqt->tqt_id = t->tqent_id;
+
+ /* We must store a copy of the flags prior to
+ * servicing the task (servicing a prealloc'd task
+ * returns the ownership of the tqent back to
+ * the caller of taskq_dispatch). Thus,
+ * tqent_flags _may_ change within the call. */
+ tqt->tqt_flags = t->tqent_flags;
+
+ taskq_insert_in_order(tq, tqt);
tq->tq_nactive++;
spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
tq->tq_nactive--;
- id = t->tqent_id;
- task_done(tq, t);
+ list_del_init(&tqt->tqt_active_list);
+
+ /* For prealloc'd tasks, we don't free anything. */
+ if ((tq->tq_flags & TASKQ_DYNAMIC) ||
+ !(tqt->tqt_flags & TQENT_FLAG_PREALLOC))
+ task_done(tq, t);
/* When the current lowest outstanding taskqid is
* done calculate the new lowest outstanding id */
- if (tq->tq_lowest_id == id) {
+ if (tq->tq_lowest_id == tqt->tqt_id) {
tq->tq_lowest_id = taskq_lowest_id(tq);
- ASSERT(tq->tq_lowest_id > id);
+ ASSERT3S(tq->tq_lowest_id, >, tqt->tqt_id);
}
+ tqt->tqt_id = 0;
+ tqt->tqt_flags = 0;
wake_up_all(&tq->tq_wait_waitq);
}
__set_current_state(TASK_RUNNING);
tq->tq_nthreads--;
+ list_del_init(&tqt->tqt_thread_list);
+ kmem_free(tqt, sizeof(taskq_thread_t));
+
spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
SRETURN(0);
int minalloc, int maxalloc, uint_t flags)
{
taskq_t *tq;
- struct task_struct *t;
+ taskq_thread_t *tqt;
int rc = 0, i, j = 0;
SENTRY;
nthreads = MAX((num_online_cpus() * nthreads) / 100, 1);
}
- tq = kmem_alloc(sizeof(*tq), KM_SLEEP);
+ tq = kmem_alloc(sizeof(*tq), KM_PUSHPAGE);
if (tq == NULL)
SRETURN(NULL);
- tq->tq_threads = kmem_alloc(nthreads * sizeof(t), KM_SLEEP);
- if (tq->tq_threads == NULL) {
- kmem_free(tq, sizeof(*tq));
- SRETURN(NULL);
- }
-
spin_lock_init(&tq->tq_lock);
spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
+ INIT_LIST_HEAD(&tq->tq_thread_list);
+ INIT_LIST_HEAD(&tq->tq_active_list);
tq->tq_name = name;
tq->tq_nactive = 0;
tq->tq_nthreads = 0;
tq->tq_next_id = 1;
tq->tq_lowest_id = 1;
INIT_LIST_HEAD(&tq->tq_free_list);
- INIT_LIST_HEAD(&tq->tq_work_list);
INIT_LIST_HEAD(&tq->tq_pend_list);
INIT_LIST_HEAD(&tq->tq_prio_list);
init_waitqueue_head(&tq->tq_work_waitq);
if (flags & TASKQ_PREPOPULATE)
for (i = 0; i < minalloc; i++)
- task_done(tq, task_alloc(tq, TQ_SLEEP | TQ_NEW));
+ task_done(tq, task_alloc(tq, TQ_PUSHPAGE | TQ_NEW));
spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
- for (i = 0; i < nthreads; i++) {
- t = kthread_create(taskq_thread, tq, "%s/%d", name, i);
- if (t) {
- tq->tq_threads[i] = t;
- kthread_bind(t, i % num_online_cpus());
- set_user_nice(t, PRIO_TO_NICE(pri));
- wake_up_process(t);
+ for (i = 0; i < nthreads; i++) {
+ tqt = kmem_alloc(sizeof(*tqt), KM_PUSHPAGE);
+ INIT_LIST_HEAD(&tqt->tqt_thread_list);
+ INIT_LIST_HEAD(&tqt->tqt_active_list);
+ tqt->tqt_tq = tq;
+ tqt->tqt_id = 0;
+
+ tqt->tqt_thread = kthread_create(taskq_thread, tqt,
+ "%s/%d", name, i);
+ if (tqt->tqt_thread) {
+ list_add(&tqt->tqt_thread_list, &tq->tq_thread_list);
+ kthread_bind(tqt->tqt_thread, i % num_online_cpus());
+ set_user_nice(tqt->tqt_thread, PRIO_TO_NICE(pri));
+ wake_up_process(tqt->tqt_thread);
j++;
- } else {
- tq->tq_threads[i] = NULL;
- rc = 1;
- }
- }
+ } else {
+ kmem_free(tqt, sizeof(taskq_thread_t));
+ rc = 1;
+ }
+ }
/* Wait for all threads to be started before potential destroy */
wait_event(tq->tq_wait_waitq, tq->tq_nthreads == j);
void
__taskq_destroy(taskq_t *tq)
{
+ struct task_struct *thread;
+ taskq_thread_t *tqt;
taskq_ent_t *t;
- int i, nthreads;
SENTRY;
ASSERT(tq);
/* TQ_ACTIVE cleared prevents new tasks being added to pending */
__taskq_wait(tq);
- nthreads = tq->tq_nthreads;
- for (i = 0; i < nthreads; i++)
- if (tq->tq_threads[i])
- kthread_stop(tq->tq_threads[i]);
-
spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
+ /*
+ * Signal each thread to exit and block until it does. Each thread
+ * is responsible for removing itself from the list and freeing its
+ * taskq_thread_t. This allows for idle threads to opt to remove
+ * themselves from the taskq. They can be recreated as needed.
+ */
+ while (!list_empty(&tq->tq_thread_list)) {
+ tqt = list_entry(tq->tq_thread_list.next,
+ taskq_thread_t, tqt_thread_list);
+ thread = tqt->tqt_thread;
+ spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
+
+ kthread_stop(thread);
+
+ spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
+ }
+
while (!list_empty(&tq->tq_free_list)) {
t = list_entry(tq->tq_free_list.next, taskq_ent_t, tqent_list);
+
+ ASSERT(!(t->tqent_flags & TQENT_FLAG_PREALLOC));
+
list_del_init(&t->tqent_list);
task_free(tq, t);
}
ASSERT(tq->tq_nthreads == 0);
ASSERT(tq->tq_nalloc == 0);
+ ASSERT(list_empty(&tq->tq_thread_list));
+ ASSERT(list_empty(&tq->tq_active_list));
ASSERT(list_empty(&tq->tq_free_list));
- ASSERT(list_empty(&tq->tq_work_list));
ASSERT(list_empty(&tq->tq_pend_list));
ASSERT(list_empty(&tq->tq_prio_list));
spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
- kmem_free(tq->tq_threads, nthreads * sizeof(taskq_ent_t *));
+
kmem_free(tq, sizeof(taskq_t));
SEXIT;
projects / mirror_spl.git / blobdiff