internal_function
__aio_enqueue_request (aiocb_union *aiocbp, int operation)
{
int result = 0;
int policy, prio;
struct sched_param param;
struct requestlist *last, *runp, *newp;
int running = no;
if (operation == LIO_SYNC || operation == LIO_DSYNC)
aiocbp->aiocb.aio_reqprio = 0;
else if (aiocbp->aiocb.aio_reqprio < 0
|| aiocbp->aiocb.aio_reqprio > AIO_PRIO_DELTA_MAX)
{
/* Invalid priority value. */
__set_errno (EINVAL);
aiocbp->aiocb.__error_code = EINVAL;
aiocbp->aiocb.__return_value = -1;
return NULL;
}
/* Compute priority for this request. */
pthread_getschedparam (pthread_self (), &policy, ¶m);
prio = param.sched_priority - aiocbp->aiocb.aio_reqprio;
/* Get the mutex. */
pthread_mutex_lock (&__aio_requests_mutex);
last = NULL;
runp = requests;
/* First look whether the current file descriptor is currently
worked with. */
while (runp != NULL
&& runp->aiocbp->aiocb.aio_fildes < aiocbp->aiocb.aio_fildes)
{
last = runp;
runp = runp->next_fd;
}
/* Get a new element for the waiting list. */
newp = get_elem ();
if (newp == NULL)
{
pthread_mutex_unlock (&__aio_requests_mutex);
__set_errno (EAGAIN);
return NULL;
}
newp->aiocbp = aiocbp;
#ifdef BROKEN_THREAD_SIGNALS
newp->caller_pid = (aiocbp->aiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL
? getpid () : 0);
#endif
newp->waiting = NULL;
aiocbp->aiocb.__abs_prio = prio;
aiocbp->aiocb.__policy = policy;
aiocbp->aiocb.aio_lio_opcode = operation;
aiocbp->aiocb.__error_code = EINPROGRESS;
aiocbp->aiocb.__return_value = 0;
if (runp != NULL
&& runp->aiocbp->aiocb.aio_fildes == aiocbp->aiocb.aio_fildes)
{
/* The current file descriptor is worked on. It makes no sense
to start another thread since this new thread would fight
with the running thread for the resources. But we also cannot
say that the thread processing this desriptor shall immediately
after finishing the current job process this request if there
are other threads in the running queue which have a higher
priority. */
/* Simply enqueue it after the running one according to the
priority. */
last = NULL;
while (runp->next_prio != NULL
&& runp->next_prio->aiocbp->aiocb.__abs_prio >= prio)
{
last = runp;
runp = runp->next_prio;
}
newp->next_prio = runp->next_prio;
runp->next_prio = newp;
running = queued;
}
else
{
running = yes;
/* Enqueue this request for a new descriptor. */
if (last == NULL)
{
newp->last_fd = NULL;
newp->next_fd = requests;
if (requests != NULL)
requests->last_fd = newp;
requests = newp;
}
else
{
newp->next_fd = last->next_fd;
newp->last_fd = last;
last->next_fd = newp;
if (newp->next_fd != NULL)
newp->next_fd->last_fd = newp;
}
newp->next_prio = NULL;
last = NULL;
}
if (running == yes)
{
/* We try to create a new thread for this file descriptor. The
function which gets called will handle all available requests
for this descriptor and when all are processed it will
terminate.
If no new thread can be created or if the specified limit of
threads for AIO is reached we queue the request. */
/* See if we need to and are able to create a thread. */
if (nthreads < optim.aio_threads && idle_thread_count == 0)
{
pthread_t thid;
running = newp->running = allocated;
/* Now try to start a thread. */
result = aio_create_helper_thread (&thid, handle_fildes_io, newp);
if (result == 0)
/* We managed to enqueue the request. All errors which can
happen now can be recognized by calls to `aio_return' and
`aio_error'. */
++nthreads;
else
{
/* Reset the running flag. The new request is not running. */
running = newp->running = yes;
if (nthreads == 0)
{
/* We cannot create a thread in the moment and there is
also no thread running. This is a problem. `errno' is
set to EAGAIN if this is only a temporary problem. */
__aio_remove_request (last, newp, 0);
}
else
result = 0;
}
}
}
/* Enqueue the request in the run queue if it is not yet running. */
if (running == yes && result == 0)
{
add_request_to_runlist (newp);
/* If there is a thread waiting for work, then let it know that we
have just given it something to do. */
if (idle_thread_count > 0)
pthread_cond_signal (&__aio_new_request_notification);
}
if (result == 0)
newp->running = running;
else
{
/* Something went wrong. */
__aio_free_request (newp);
aiocbp->aiocb.__error_code = result;
__set_errno (result);
newp = NULL;
}
/* Release the mutex. */
pthread_mutex_unlock (&__aio_requests_mutex);
return newp;
}
static void *
handle_fildes_io (void *arg)
{
pthread_t self = pthread_self ();
struct sched_param param;
struct requestlist *runp = (struct requestlist *) arg;
aiocb_union *aiocbp;
int policy;
int fildes;
pthread_getschedparam (self, &policy, ¶m);
do
{
/* If runp is NULL, then we were created to service the work queue
in general, not to handle any particular request. In that case we
skip the "do work" stuff on the first pass, and go directly to the
"get work off the work queue" part of this loop, which is near the
end. */
if (runp == NULL)
pthread_mutex_lock (&__aio_requests_mutex);
else
{
/* Hopefully this request is marked as running. */
assert (runp->running == allocated);
/* Update our variables. */
aiocbp = runp->aiocbp;
fildes = aiocbp->aiocb.aio_fildes;
/* Change the priority to the requested value (if necessary). */
if (aiocbp->aiocb.__abs_prio != param.sched_priority
|| aiocbp->aiocb.__policy != policy)
{
param.sched_priority = aiocbp->aiocb.__abs_prio;
policy = aiocbp->aiocb.__policy;
pthread_setschedparam (self, policy, ¶m);
}
/* Process request pointed to by RUNP. We must not be disturbed
by signals. */
if ((aiocbp->aiocb.aio_lio_opcode & 127) == LIO_READ)
{
if (sizeof (off_t) != sizeof (off64_t)
&& aiocbp->aiocb.aio_lio_opcode & 128)
aiocbp->aiocb.__return_value =
TEMP_FAILURE_RETRY (__pread64 (fildes, (void *)
aiocbp->aiocb64.aio_buf,
aiocbp->aiocb64.aio_nbytes,
aiocbp->aiocb64.aio_offset));
else
aiocbp->aiocb.__return_value =
TEMP_FAILURE_RETRY (pread (fildes,
(void *) aiocbp->aiocb.aio_buf,
aiocbp->aiocb.aio_nbytes,
aiocbp->aiocb.aio_offset));
if (aiocbp->aiocb.__return_value == -1 && errno == ESPIPE)
/* The Linux kernel is different from others. It returns
ESPIPE if using pread on a socket. Other platforms
simply ignore the offset parameter and behave like
read. */
aiocbp->aiocb.__return_value =
TEMP_FAILURE_RETRY (read (fildes,
(void *) aiocbp->aiocb64.aio_buf,
aiocbp->aiocb64.aio_nbytes));
}
else if ((aiocbp->aiocb.aio_lio_opcode & 127) == LIO_WRITE)
{
if (sizeof (off_t) != sizeof (off64_t)
&& aiocbp->aiocb.aio_lio_opcode & 128)
aiocbp->aiocb.__return_value =
TEMP_FAILURE_RETRY (__pwrite64 (fildes, (const void *)
aiocbp->aiocb64.aio_buf,
aiocbp->aiocb64.aio_nbytes,
aiocbp->aiocb64.aio_offset));
else
aiocbp->aiocb.__return_value =
TEMP_FAILURE_RETRY (__libc_pwrite (fildes, (const void *)
aiocbp->aiocb.aio_buf,
aiocbp->aiocb.aio_nbytes,
aiocbp->aiocb.aio_offset));
if (aiocbp->aiocb.__return_value == -1 && errno == ESPIPE)
/* The Linux kernel is different from others. It returns
ESPIPE if using pwrite on a socket. Other platforms
simply ignore the offset parameter and behave like
write. */
aiocbp->aiocb.__return_value =
TEMP_FAILURE_RETRY (write (fildes,
(void *) aiocbp->aiocb64.aio_buf,
aiocbp->aiocb64.aio_nbytes));
}
else if (aiocbp->aiocb.aio_lio_opcode == LIO_DSYNC)
aiocbp->aiocb.__return_value =
TEMP_FAILURE_RETRY (fdatasync (fildes));
else if (aiocbp->aiocb.aio_lio_opcode == LIO_SYNC)
aiocbp->aiocb.__return_value =
TEMP_FAILURE_RETRY (fsync (fildes));
else
{
/* This is an invalid opcode. */
aiocbp->aiocb.__return_value = -1;
__set_errno (EINVAL);
}
/* Get the mutex. */
pthread_mutex_lock (&__aio_requests_mutex);
/* In theory we would need here a write memory barrier since the
callers test using aio_error() whether the request finished
and once this value != EINPROGRESS the field __return_value
must be committed to memory.
But since the pthread_mutex_lock call involves write memory
barriers as well it is not necessary. */
if (aiocbp->aiocb.__return_value == -1)
aiocbp->aiocb.__error_code = errno;
else
aiocbp->aiocb.__error_code = 0;
/* Send the signal to notify about finished processing of the
request. */
__aio_notify (runp);
/* For debugging purposes we reset the running flag of the
finished request. */
assert (runp->running == allocated);
runp->running = done;
/* Now dequeue the current request. */
__aio_remove_request (NULL, runp, 0);
if (runp->next_prio != NULL)
add_request_to_runlist (runp->next_prio);
/* Free the old element. */
__aio_free_request (runp);
}
runp = runlist;
/* If the runlist is empty, then we sleep for a while, waiting for
something to arrive in it. */
if (runp == NULL && optim.aio_idle_time >= 0)
{
struct timeval now;
struct timespec wakeup_time;
++idle_thread_count;
gettimeofday (&now, NULL);
wakeup_time.tv_sec = now.tv_sec + optim.aio_idle_time;
wakeup_time.tv_nsec = now.tv_usec * 1000;
if (wakeup_time.tv_nsec >= 1000000000)
{
wakeup_time.tv_nsec -= 1000000000;
++wakeup_time.tv_sec;
}
pthread_cond_timedwait (&__aio_new_request_notification,
&__aio_requests_mutex,
&wakeup_time);
--idle_thread_count;
runp = runlist;
}
if (runp == NULL)
--nthreads;
else
{
assert (runp->running == yes);
runp->running = allocated;
runlist = runp->next_run;
/* If we have a request to process, and there's still another in
the run list, then we need to either wake up or create a new
thread to service the request that is still in the run list. */
if (runlist != NULL)
{
/* There are at least two items in the work queue to work on.
If there are other idle threads, then we should wake them
up for these other work elements; otherwise, we should try
to create a new thread. */
if (idle_thread_count > 0)
pthread_cond_signal (&__aio_new_request_notification);
else if (nthreads < optim.aio_threads)
{
pthread_t thid;
pthread_attr_t attr;
/* Make sure the thread is created detached. */
pthread_attr_init (&attr);
pthread_attr_setdetachstate (&attr, PTHREAD_CREATE_DETACHED);
/* Now try to start a thread. If we fail, no big deal,
because we know that there is at least one thread (us)
that is working on AIO operations. */
if (pthread_create (&thid, &attr, handle_fildes_io, NULL)
== 0)
++nthreads;
}
}
}
/* Release the mutex. */
pthread_mutex_unlock (&__aio_requests_mutex);
}
while (runp != NULL);
return NULL;
}
int
aio_cancel (int fildes, struct aiocb *aiocbp)
{
struct requestlist *req = NULL;
int result = AIO_ALLDONE;
/* If fildes is invalid, error. */
if (fcntl (fildes, F_GETFL) < 0)
{
__set_errno (EBADF);
return -1;
}
/* Request the mutex. */
pthread_mutex_lock (&__aio_requests_mutex);
/* We are asked to cancel a specific AIO request. */
if (aiocbp != NULL)
{
/* If the AIO request is not for this descriptor it has no value
to look for the request block. */
if (aiocbp->aio_fildes != fildes)
{
pthread_mutex_unlock (&__aio_requests_mutex);
__set_errno (EINVAL);
return -1;
}
else if (aiocbp->__error_code == EINPROGRESS)
{
struct requestlist *last = NULL;
req = __aio_find_req_fd (fildes);
if (req == NULL)
{
not_found:
pthread_mutex_unlock (&__aio_requests_mutex);
__set_errno (EINVAL);
return -1;
}
while (req->aiocbp != (aiocb_union *) aiocbp)
{
last = req;
req = req->next_prio;
if (req == NULL)
goto not_found;
}
/* Don't remove the entry if a thread is already working on it. */
if (req->running == allocated)
{
result = AIO_NOTCANCELED;
req = NULL;
}
else
{
/* We can remove the entry. */
__aio_remove_request (last, req, 0);
result = AIO_CANCELED;
req->next_prio = NULL;
}
}
}
else
{
/* Find the beginning of the list of all requests for this
desriptor. */
req = __aio_find_req_fd (fildes);
/* If any request is worked on by a thread it must be the first.
So either we can delete all requests or all but the first. */
if (req != NULL)
{
if (req->running == allocated)
{
struct requestlist *old = req;
req = req->next_prio;
old->next_prio = NULL;
result = AIO_NOTCANCELED;
if (req != NULL)
__aio_remove_request (old, req, 1);
}
else
{
result = AIO_CANCELED;
/* We can remove the entry. */
__aio_remove_request (NULL, req, 1);
}
}
}
/* Mark requests as canceled and send signal. */
while (req != NULL)
{
struct requestlist *old = req;
assert (req->running == yes || req->running == queued);
req->aiocbp->aiocb.__error_code = ECANCELED;
req->aiocbp->aiocb.__return_value = -1;
__aio_notify (req);
req = req->next_prio;
__aio_free_request (old);
}
/* Release the mutex. */
pthread_mutex_unlock (&__aio_requests_mutex);
return result;
}
