*
* You should have received a copy of the GNU General Public License
* version 2 along with this program; If not, see
- * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
- *
- * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
- * CA 95054 USA or visit www.sun.com if you need additional information or
- * have any questions.
+ * http://www.gnu.org/licenses/gpl-2.0.html
*
* GPL HEADER END
*/
/*
* Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
* Use is subject to license terms.
+ *
+ * Copyright (c) 2012, 2016, Intel Corporation.
*/
/*
* This file is part of Lustre, http://www.lustre.org/
* \param handle On successful return, this location will hold a handle for
* the newly created EQ.
*
- * \retval 0 On success.
+ * \retval 0 On success.
* \retval -EINVAL If an parameter is not valid.
* \retval -ENOMEM If memory for the EQ can't be allocated.
*
*/
int
LNetEQAlloc(unsigned int count, lnet_eq_handler_t callback,
- lnet_handle_eq_t *handle)
+ lnet_handle_eq_t *handle)
{
- lnet_eq_t *eq;
+ lnet_eq_t *eq;
- LASSERT (the_lnet.ln_init);
- LASSERT (the_lnet.ln_refcount > 0);
+ LASSERT(the_lnet.ln_refcount > 0);
- /* We need count to be a power of 2 so that when eq_{enq,deq}_seq
- * overflow, they don't skip entries, so the queue has the same
- * apparent capacity at all times */
+ /* We need count to be a power of 2 so that when eq_{enq,deq}_seq
+ * overflow, they don't skip entries, so the queue has the same
+ * apparent capacity at all times */
- count = cfs_power2_roundup(count);
+ if (count)
+ count = roundup_pow_of_two(count);
if (callback != LNET_EQ_HANDLER_NONE && count != 0) {
CWARN("EQ callback is guaranteed to get every event, "
if (count != 0) {
LIBCFS_ALLOC(eq->eq_events, count * sizeof(lnet_event_t));
- if (eq->eq_events == NULL) {
- lnet_eq_free(eq);
- return -ENOMEM;
- }
+ if (eq->eq_events == NULL)
+ goto failed;
/* NB allocator has set all event sequence numbers to 0,
* so all them should be earlier than eq_deq_seq */
}
- eq->eq_deq_seq = 1;
- eq->eq_enq_seq = 1;
- eq->eq_size = count;
- eq->eq_refcount = 0;
- eq->eq_callback = callback;
+ eq->eq_deq_seq = 1;
+ eq->eq_enq_seq = 1;
+ eq->eq_size = count;
+ eq->eq_callback = callback;
+
+ eq->eq_refs = cfs_percpt_alloc(lnet_cpt_table(),
+ sizeof(*eq->eq_refs[0]));
+ if (eq->eq_refs == NULL)
+ goto failed;
- lnet_res_lock();
+ /* MUST hold both exclusive lnet_res_lock */
+ lnet_res_lock(LNET_LOCK_EX);
+ /* NB: hold lnet_eq_wait_lock for EQ link/unlink, so we can do
+ * both EQ lookup and poll event with only lnet_eq_wait_lock */
+ lnet_eq_wait_lock();
lnet_res_lh_initialize(&the_lnet.ln_eq_container, &eq->eq_lh);
- cfs_list_add(&eq->eq_list, &the_lnet.ln_eq_container.rec_active);
+ list_add(&eq->eq_list, &the_lnet.ln_eq_container.rec_active);
- lnet_res_unlock();
+ lnet_eq_wait_unlock();
+ lnet_res_unlock(LNET_LOCK_EX);
lnet_eq2handle(handle, eq);
return 0;
+
+failed:
+ if (eq->eq_events != NULL)
+ LIBCFS_FREE(eq->eq_events, count * sizeof(lnet_event_t));
+
+ if (eq->eq_refs != NULL)
+ cfs_percpt_free(eq->eq_refs);
+
+ lnet_eq_free(eq);
+ return -ENOMEM;
}
+EXPORT_SYMBOL(LNetEQAlloc);
/**
* Release the resources associated with an event queue if it's idle;
int
LNetEQFree(lnet_handle_eq_t eqh)
{
- lnet_eq_t *eq;
- int size;
- lnet_event_t *events;
+ struct lnet_eq *eq;
+ lnet_event_t *events = NULL;
+ int **refs = NULL;
+ int *ref;
+ int rc = 0;
+ int size = 0;
+ int i;
- LASSERT (the_lnet.ln_init);
- LASSERT (the_lnet.ln_refcount > 0);
+ LASSERT(the_lnet.ln_refcount > 0);
- lnet_res_lock();
+ lnet_res_lock(LNET_LOCK_EX);
+ /* NB: hold lnet_eq_wait_lock for EQ link/unlink, so we can do
+ * both EQ lookup and poll event with only lnet_eq_wait_lock */
+ lnet_eq_wait_lock();
eq = lnet_handle2eq(&eqh);
if (eq == NULL) {
- lnet_res_unlock();
- return -ENOENT;
+ rc = -ENOENT;
+ goto out;
}
- if (eq->eq_refcount != 0) {
- CDEBUG(D_NET, "Event queue (%d) busy on destroy.\n",
- eq->eq_refcount);
- lnet_res_unlock();
- return -EBUSY;
+ cfs_percpt_for_each(ref, i, eq->eq_refs) {
+ LASSERT(*ref >= 0);
+ if (*ref == 0)
+ continue;
+
+ CDEBUG(D_NET, "Event equeue (%d: %d) busy on destroy.\n",
+ i, *ref);
+ rc = -EBUSY;
+ goto out;
}
/* stash for free after lock dropped */
events = eq->eq_events;
size = eq->eq_size;
+ refs = eq->eq_refs;
lnet_res_lh_invalidate(&eq->eq_lh);
- cfs_list_del(&eq->eq_list);
- lnet_eq_free_locked(eq);
-
- lnet_res_unlock();
+ list_del(&eq->eq_list);
+ lnet_eq_free(eq);
+ out:
+ lnet_eq_wait_unlock();
+ lnet_res_unlock(LNET_LOCK_EX);
if (events != NULL)
LIBCFS_FREE(events, size * sizeof(lnet_event_t));
+ if (refs != NULL)
+ cfs_percpt_free(refs);
- return 0;
+ return rc;
}
+EXPORT_SYMBOL(LNetEQFree);
void
lnet_eq_enqueue_event(lnet_eq_t *eq, lnet_event_t *ev)
{
- /* MUST called with resource lock hold */
+ /* MUST called with resource lock hold but w/o lnet_eq_wait_lock */
int index;
if (eq->eq_size == 0) {
return;
}
+ lnet_eq_wait_lock();
ev->sequence = eq->eq_enq_seq++;
LASSERT(eq->eq_size == LOWEST_BIT_SET(eq->eq_size));
if (eq->eq_callback != LNET_EQ_HANDLER_NONE)
eq->eq_callback(ev);
-#ifdef __KERNEL__
- /* Wake anyone waiting in LNetEQPoll() */
- if (cfs_waitq_active(&the_lnet.ln_eq_waitq))
- cfs_waitq_broadcast(&the_lnet.ln_eq_waitq);
-#else
-# ifndef HAVE_LIBPTHREAD
- /* LNetEQPoll() calls into _the_ LND to wait for action */
-# else
/* Wake anyone waiting in LNetEQPoll() */
- pthread_cond_broadcast(&the_lnet.ln_eq_cond);
-# endif
-#endif
+ if (waitqueue_active(&the_lnet.ln_eq_waitq))
+ wake_up_all(&the_lnet.ln_eq_waitq);
+ lnet_eq_wait_unlock();
}
-int
+static int
lnet_eq_dequeue_event(lnet_eq_t *eq, lnet_event_t *ev)
{
- int new_index = eq->eq_deq_seq & (eq->eq_size - 1);
- lnet_event_t *new_event = &eq->eq_events[new_index];
- int rc;
- ENTRY;
+ int new_index = eq->eq_deq_seq & (eq->eq_size - 1);
+ lnet_event_t *new_event = &eq->eq_events[new_index];
+ int rc;
+ ENTRY;
- if (LNET_SEQ_GT (eq->eq_deq_seq, new_event->sequence)) {
- RETURN(0);
- }
+ /* must called with lnet_eq_wait_lock hold */
+ if (LNET_SEQ_GT(eq->eq_deq_seq, new_event->sequence))
+ RETURN(0);
- /* We've got a new event... */
- *ev = *new_event;
+ /* We've got a new event... */
+ *ev = *new_event;
CDEBUG(D_INFO, "event: %p, sequence: %lu, eq->size: %u\n",
new_event, eq->eq_deq_seq, eq->eq_size);
- /* ...but did it overwrite an event we've not seen yet? */
- if (eq->eq_deq_seq == new_event->sequence) {
- rc = 1;
- } else {
- /* don't complain with CERROR: some EQs are sized small
- * anyway; if it's important, the caller should complain */
- CDEBUG(D_NET, "Event Queue Overflow: eq seq %lu ev seq %lu\n",
- eq->eq_deq_seq, new_event->sequence);
- rc = -EOVERFLOW;
- }
-
- eq->eq_deq_seq = new_event->sequence + 1;
- RETURN(rc);
+ /* ...but did it overwrite an event we've not seen yet? */
+ if (eq->eq_deq_seq == new_event->sequence) {
+ rc = 1;
+ } else {
+ /* don't complain with CERROR: some EQs are sized small
+ * anyway; if it's important, the caller should complain */
+ CDEBUG(D_NET, "Event Queue Overflow: eq seq %lu ev seq %lu\n",
+ eq->eq_deq_seq, new_event->sequence);
+ rc = -EOVERFLOW;
+ }
+
+ eq->eq_deq_seq = new_event->sequence + 1;
+ RETURN(rc);
}
/**
* \param event On successful return (1 or -EOVERFLOW), this location will
* hold the next event in the EQ.
*
- * \retval 0 No pending event in the EQ.
- * \retval 1 Indicates success.
+ * \retval 0 No pending event in the EQ.
+ * \retval 1 Indicates success.
* \retval -ENOENT If \a eventq does not point to a valid EQ.
* \retval -EOVERFLOW Indicates success (i.e., an event is returned) and that
* at least one event between this event and the last event obtained from the
int
LNetEQGet (lnet_handle_eq_t eventq, lnet_event_t *event)
{
- int which;
+ int which;
- return LNetEQPoll(&eventq, 1, 0,
- event, &which);
+ return LNetEQPoll(&eventq, 1, 0,
+ event, &which);
}
+EXPORT_SYMBOL(LNetEQGet);
/**
* Block the calling process until there is an event in the EQ.
* \param event On successful return (1 or -EOVERFLOW), this location will
* hold the next event in the EQ.
*
- * \retval 1 Indicates success.
+ * \retval 1 Indicates success.
* \retval -ENOENT If \a eventq does not point to a valid EQ.
* \retval -EOVERFLOW Indicates success (i.e., an event is returned) and that
* at least one event between this event and the last event obtained from the
int
LNetEQWait (lnet_handle_eq_t eventq, lnet_event_t *event)
{
- int which;
+ int which;
- return LNetEQPoll(&eventq, 1, LNET_TIME_FOREVER,
- event, &which);
+ return LNetEQPoll(&eventq, 1, MAX_SCHEDULE_TIMEOUT,
+ event, &which);
}
-
-#ifdef __KERNEL__
+EXPORT_SYMBOL(LNetEQWait);
static int
-lnet_eq_wait_locked(int *timeout_ms)
+lnet_eq_wait_locked(signed long *timeout)
+__must_hold(&the_lnet.ln_eq_wait_lock)
{
- int tms = *timeout_ms;
- int wait;
- cfs_waitlink_t wl;
- cfs_time_t now;
+ signed long tms = *timeout;
+ wait_queue_t wl;
+ int wait;
if (tms == 0)
- return -1; /* don't want to wait and no new event */
-
- cfs_waitlink_init(&wl);
- cfs_set_current_state(CFS_TASK_INTERRUPTIBLE);
- cfs_waitq_add(&the_lnet.ln_eq_waitq, &wl);
+ return -ENXIO; /* don't want to wait and no new event */
- lnet_res_unlock();
+ init_waitqueue_entry(&wl, current);
+ add_wait_queue(&the_lnet.ln_eq_waitq, &wl);
- if (tms < 0) {
- cfs_waitq_wait(&wl, CFS_TASK_INTERRUPTIBLE);
-
- } else {
- struct timeval tv;
-
- now = cfs_time_current();
- cfs_waitq_timedwait(&wl, CFS_TASK_INTERRUPTIBLE,
- cfs_time_seconds(tms) / 1000);
- cfs_duration_usec(cfs_time_sub(cfs_time_current(), now), &tv);
- tms -= (int)(tv.tv_sec * 1000 + tv.tv_usec / 1000);
- if (tms < 0) /* no more wait but may have new event */
- tms = 0;
- }
+ lnet_eq_wait_unlock();
+ tms = schedule_timeout_interruptible(tms);
wait = tms != 0; /* might need to call here again */
- *timeout_ms = tms;
-
- lnet_res_lock();
- cfs_waitq_del(&the_lnet.ln_eq_waitq, &wl);
-
- return wait;
-}
-
-#else /* !__KERNEL__ */
-
-# ifdef HAVE_LIBPTHREAD
-static void
-lnet_eq_cond_wait(struct timespec *ts)
-{
- if (ts == NULL) {
- pthread_cond_wait(&the_lnet.ln_eq_cond, &the_lnet.ln_res_lock);
- } else {
- pthread_cond_timedwait(&the_lnet.ln_eq_cond,
- &the_lnet.ln_res_lock, ts);
- }
-}
-# endif
-
-static int
-lnet_eq_wait_locked(int *timeout_ms)
-{
- lnet_ni_t *eq_waitni = NULL;
- int tms = *timeout_ms;
- int wait;
- struct timeval then;
- struct timeval now;
-
- if (the_lnet.ln_eq_waitni != NULL) {
- /* I have a single NI that I have to call into, to get
- * events queued, or to block. */
- lnet_res_unlock();
-
- LNET_LOCK();
- eq_waitni = the_lnet.ln_eq_waitni;
- if (unlikely(eq_waitni == NULL)) {
- LNET_UNLOCK();
-
- lnet_res_lock();
- return -1;
- }
-
- lnet_ni_addref_locked(eq_waitni);
- LNET_UNLOCK();
-
- if (tms <= 0) { /* even for tms == 0 */
- (eq_waitni->ni_lnd->lnd_wait)(eq_waitni, tms);
-
- } else {
- gettimeofday(&then, NULL);
-
- (eq_waitni->ni_lnd->lnd_wait)(eq_waitni, tms);
-
- gettimeofday(&now, NULL);
- tms -= (now.tv_sec - then.tv_sec) * 1000 +
- (now.tv_usec - then.tv_usec) / 1000;
- if (tms < 0)
- tms = 0;
- }
-
- lnet_ni_decref(eq_waitni);
- lnet_res_lock();
- } else { /* w/o eq_waitni */
-# ifndef HAVE_LIBPTHREAD
- /* If I'm single-threaded, LNET fails at startup if it can't
- * set the_lnet.ln_eqwaitni correctly. */
- LBUG();
-# else /* HAVE_LIBPTHREAD */
- struct timespec ts;
+ *timeout = tms;
- if (tms == 0) /* don't want to wait and new event */
- return -1;
-
- if (tms < 0) {
- lnet_eq_cond_wait(NULL);
-
- } else {
-
- gettimeofday(&then, NULL);
-
- ts.tv_sec = then.tv_sec + tms / 1000;
- ts.tv_nsec = then.tv_usec * 1000 +
- (tms % 1000) * 1000000;
- if (ts.tv_nsec >= 1000000000) {
- ts.tv_sec++;
- ts.tv_nsec -= 1000000000;
- }
-
- lnet_eq_cond_wait(&ts);
-
- gettimeofday(&now, NULL);
- tms -= (now.tv_sec - then.tv_sec) * 1000 +
- (now.tv_usec - then.tv_usec) / 1000;
- if (tms < 0)
- tms = 0;
- }
-# endif /* HAVE_LIBPTHREAD */
- }
-
- wait = tms != 0;
- *timeout_ms = tms;
+ lnet_eq_wait_lock();
+ remove_wait_queue(&the_lnet.ln_eq_waitq, &wl);
return wait;
}
-#endif /* __KERNEL__ */
-
-
/**
* Block the calling process until there's an event from a set of EQs or
* timeout happens.
* fixed period, or block indefinitely.
*
* \param eventqs,neq An array of EQ handles, and size of the array.
- * \param timeout_ms Time in milliseconds to wait for an event to occur on
- * one of the EQs. The constant LNET_TIME_FOREVER can be used to indicate an
+ * \param timeout Time in jiffies to wait for an event to occur on
+ * one of the EQs. The constant MAX_SCHEDULE_TIMEOUT can be used to indicate an
* infinite timeout.
* \param event,which On successful return (1 or -EOVERFLOW), \a event will
* hold the next event in the EQs, and \a which will contain the index of the
* EQ from which the event was taken.
*
- * \retval 0 No pending event in the EQs after timeout.
- * \retval 1 Indicates success.
+ * \retval 0 No pending event in the EQs after timeout.
+ * \retval 1 Indicates success.
* \retval -EOVERFLOW Indicates success (i.e., an event is returned) and that
* at least one event between this event and the last event obtained from the
* EQ indicated by \a which has been dropped due to limited space in the EQ.
* \retval -ENOENT If there's an invalid handle in \a eventqs.
*/
int
-LNetEQPoll (lnet_handle_eq_t *eventqs, int neq, int timeout_ms,
- lnet_event_t *event, int *which)
+LNetEQPoll(lnet_handle_eq_t *eventqs, int neq, signed long timeout,
+ lnet_event_t *event, int *which)
{
int wait = 1;
int rc;
int i;
- ENTRY;
-
- LASSERT (the_lnet.ln_init);
- LASSERT (the_lnet.ln_refcount > 0);
-
- if (neq < 1)
- RETURN(-ENOENT);
+ ENTRY;
- lnet_res_lock();
+ LASSERT(the_lnet.ln_refcount > 0);
- for (;;) {
-#ifndef __KERNEL__
- lnet_res_unlock();
+ if (neq < 1)
+ RETURN(-ENOENT);
- /* Recursion breaker */
- if (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING &&
- !LNetHandleIsEqual(eventqs[0], the_lnet.ln_rc_eqh))
- lnet_router_checker();
+ lnet_eq_wait_lock();
- lnet_res_lock();
-#endif
+ for (;;) {
for (i = 0; i < neq; i++) {
lnet_eq_t *eq = lnet_handle2eq(&eventqs[i]);
if (eq == NULL) {
- lnet_res_unlock();
+ lnet_eq_wait_unlock();
RETURN(-ENOENT);
}
rc = lnet_eq_dequeue_event(eq, event);
if (rc != 0) {
- lnet_res_unlock();
+ lnet_eq_wait_unlock();
*which = i;
RETURN(rc);
}
* -1 : did nothing and it's sure no new event
* 1 : sleep inside and wait until new event
* 0 : don't want to wait anymore, but might have new event
- * so need to call dequeue again
+ * so need to call dequeue again
*/
- wait = lnet_eq_wait_locked(&timeout_ms);
+ wait = lnet_eq_wait_locked(&timeout);
if (wait < 0) /* no new event */
break;
}
- lnet_res_unlock();
+ lnet_eq_wait_unlock();
RETURN(0);
}
git://git.whamcloud.com / fs / lustre-release.git / blobdiff