On Thu, Oct 21, 2010 at 1:10 PM, Arun R Bharadwaj
<a...@linux.vnet.ibm.com> wrote:
> static ssize_t qemu_paio_return(struct qemu_paiocb *aiocb)
> {
> ssize_t ret;
>
> - mutex_lock(&lock);
> ret = aiocb->ret;
> - mutex_unlock(&lock);
> -
> return ret;
> }
>
> @@ -536,14 +442,15 @@ static void paio_cancel(BlockDriverAIOCB *blockacb)
> struct qemu_paiocb *acb = (struct qemu_paiocb *)blockacb;
> int active = 0;
>
> - mutex_lock(&lock);
> if (!acb->active) {
> - QTAILQ_REMOVE(&request_list, acb, node);
> - acb->ret = -ECANCELED;
> + if (!deque_threadletwork(&acb->work)) {
> + acb->ret = -ECANCELED;
> + } else {
> + active = 1;
> + }
> } else if (acb->ret == -EINPROGRESS) {
> active = 1;
> }
> - mutex_unlock(&lock);
>
> if (active) {
> /* fail safe: if the aio could not be canceled, we wait for
Here is the assembly listing of what happens next:
454:posix-aio-compat.c **** while (qemu_paio_error(acb) == EINPROGRESS)
539 0347 48F7D8 negq %rax
540 034a 83F873 cmpl $115, %eax
541 034d 7581 jne .L46
543 .L58:
545 0350 EBFD jmp .L58
This while loop is an infinite loop. The compiler doesn't need to
load acb->ret from memory.
The reason this loop worked before threadlets was because
qemu_paio_return() used to acquire a lock to access acb->ret, forcing
the return value to be loaded each iteration:
static ssize_t qemu_paio_return(struct qemu_paiocb *aiocb)
{
ssize_t ret;
mutex_lock(&lock);
ret = aiocb->ret;
mutex_unlock(&lock);
return ret;
}
Please use synchronization to wait on the request like Anthony suggested.
Stefan
