Change task_work_cancel() to use lockless_dereference(), this is what the code really wants but we didn't have this helper when it was written.
Also add the fast-path task->task_works == NULL check, in the likely case this task has no pending works and we can avoid spin_lock(task->pi_lock). While at it, change other users of ACCESS_ONCE() to use READ_ONCE(). Signed-off-by: Oleg Nesterov <[email protected]> --- kernel/task_work.c | 10 ++++++---- 1 file changed, 6 insertions(+), 4 deletions(-) diff --git a/kernel/task_work.c b/kernel/task_work.c index 53fa971..175f67a 100644 --- a/kernel/task_work.c +++ b/kernel/task_work.c @@ -29,7 +29,7 @@ task_work_add(struct task_struct *task, struct callback_head *work, bool notify) struct callback_head *head; do { - head = ACCESS_ONCE(task->task_works); + head = READ_ONCE(task->task_works); if (unlikely(head == &work_exited)) return -ESRCH; work->next = head; @@ -57,6 +57,9 @@ task_work_cancel(struct task_struct *task, task_work_func_t func) struct callback_head **pprev = &task->task_works; struct callback_head *work; unsigned long flags; + + if (likely(!task->task_works)) + return NULL; /* * If cmpxchg() fails we continue without updating pprev. * Either we raced with task_work_add() which added the @@ -64,8 +67,7 @@ task_work_cancel(struct task_struct *task, task_work_func_t func) * we raced with task_work_run(), *pprev == NULL/exited. */ raw_spin_lock_irqsave(&task->pi_lock, flags); - while ((work = ACCESS_ONCE(*pprev))) { - smp_read_barrier_depends(); + while ((work = lockless_dereference(*pprev))) { if (work->func != func) pprev = &work->next; else if (cmpxchg(pprev, work, work->next) == work) @@ -95,7 +97,7 @@ void task_work_run(void) * work_exited unless the list is empty. */ do { - work = ACCESS_ONCE(task->task_works); + work = READ_ONCE(task->task_works); head = !work && (task->flags & PF_EXITING) ? &work_exited : NULL; } while (cmpxchg(&task->task_works, work, head) != work); -- 2.5.0
