-/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
- * vim:expandtab:shiftwidth=8:tabstop=8:
+/*
+ * GPL HEADER START
+ *
+ * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License version 2 only,
+ * as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * General Public License version 2 for more details (a copy is included
+ * in the LICENSE file that accompanied this code).
+ *
+ * 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.
+ *
+ * GPL HEADER END
+ */
+/*
+ * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
+ * Use is subject to license terms.
+ *
+ * Copyright (c) 2012, Intel Corporation.
+ */
+/*
+ * This file is part of Lustre, http://www.lustre.org/
+ * Lustre is a trademark of Sun Microsystems, Inc.
+ *
+ * lnet/lnet/lib-eq.c
*
- * lib/lib-eq.c
* Library level Event queue management routines
+ */
+
+#define DEBUG_SUBSYSTEM S_LNET
+#include <lnet/lib-lnet.h>
+
+/**
+ * Create an event queue that has room for \a count number of events.
+ *
+ * The event queue is circular and older events will be overwritten by new
+ * ones if they are not removed in time by the user using the functions
+ * LNetEQGet(), LNetEQWait(), or LNetEQPoll(). It is up to the user to
+ * determine the appropriate size of the event queue to prevent this loss
+ * of events. Note that when EQ handler is specified in \a callback, no
+ * event loss can happen, since the handler is run for each event deposited
+ * into the EQ.
*
- * Copyright (c) 2001-2003 Cluster File Systems, Inc.
- * Copyright (c) 2001-2002 Sandia National Laboratories
+ * \param count The number of events to be stored in the event queue. It
+ * will be rounded up to the next power of two.
+ * \param callback A handler function that runs when an event is deposited
+ * into the EQ. The constant value LNET_EQ_HANDLER_NONE can be used to
+ * indicate that no event handler is desired.
+ * \param handle On successful return, this location will hold a handle for
+ * the newly created EQ.
*
- * This file is part of Lustre, http://www.sf.net/projects/lustre/
+ * \retval 0 On success.
+ * \retval -EINVAL If an parameter is not valid.
+ * \retval -ENOMEM If memory for the EQ can't be allocated.
*
- * Lustre is free software; you can redistribute it and/or
- * modify it under the terms of version 2 of the GNU General Public
- * License as published by the Free Software Foundation.
+ * \see lnet_eq_handler_t for the discussion on EQ handler semantics.
+ */
+int
+LNetEQAlloc(unsigned int count, lnet_eq_handler_t callback,
+ lnet_handle_eq_t *handle)
+{
+ lnet_eq_t *eq;
+
+ LASSERT (the_lnet.ln_init);
+ 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 */
+
+ count = cfs_power2_roundup(count);
+
+ if (callback != LNET_EQ_HANDLER_NONE && count != 0) {
+ CWARN("EQ callback is guaranteed to get every event, "
+ "do you still want to set eqcount %d for polling "
+ "event which will have locking overhead? "
+ "Please contact with developer to confirm\n", count);
+ }
+
+ /* count can be 0 if only need callback, we can eliminate
+ * overhead of enqueue event */
+ if (count == 0 && callback == LNET_EQ_HANDLER_NONE)
+ return -EINVAL;
+
+ eq = lnet_eq_alloc();
+ if (eq == NULL)
+ return -ENOMEM;
+
+ if (count != 0) {
+ LIBCFS_ALLOC(eq->eq_events, count * sizeof(lnet_event_t));
+ 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_callback = callback;
+
+ eq->eq_refs = cfs_percpt_alloc(lnet_cpt_table(),
+ sizeof(*eq->eq_refs[0]));
+ if (eq->eq_refs == NULL)
+ goto failed;
+
+ /* 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);
+ list_add(&eq->eq_list, &the_lnet.ln_eq_container.rec_active);
+
+ 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;
+ * otherwise do nothing and it's up to the user to try again.
*
- * Lustre is distributed in the hope that it will be useful,
- * but WITHOUT ANY WARRANTY; without even the implied warranty of
- * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
- * GNU General Public License for more details.
+ * \param eqh A handle for the event queue to be released.
*
- * You should have received a copy of the GNU General Public License
- * along with Lustre; if not, write to the Free Software
- * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+ * \retval 0 If the EQ is not in use and freed.
+ * \retval -ENOENT If \a eqh does not point to a valid EQ.
+ * \retval -EBUSY If the EQ is still in use by some MDs.
*/
+int
+LNetEQFree(lnet_handle_eq_t eqh)
+{
+ 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);
+
+ 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) {
+ rc = -ENOENT;
+ goto out;
+ }
+
+ cfs_percpt_for_each(ref, i, eq->eq_refs) {
+ LASSERT(*ref >= 0);
+ if (*ref == 0)
+ continue;
-#define DEBUG_SUBSYSTEM S_PORTALS
-#include <portals/lib-p30.h>
-#include <portals/arg-blocks.h>
+ CDEBUG(D_NET, "Event equeue (%d: %d) busy on destroy.\n",
+ i, *ref);
+ rc = -EBUSY;
+ goto out;
+ }
-int do_PtlEQAlloc_internal(nal_cb_t * nal, void *private, void *v_args,
- void *v_ret)
+ /* 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);
+ list_del(&eq->eq_list);
+ lnet_eq_free_locked(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 rc;
+}
+EXPORT_SYMBOL(LNetEQFree);
+
+void
+lnet_eq_enqueue_event(lnet_eq_t *eq, lnet_event_t *ev)
{
- /*
- * Incoming:
- * ptl_handle_ni_t ni_in
- * ptl_size_t count_in
- * void * base_in
- *
- * Outgoing:
- * ptl_handle_eq_t * handle_out
- */
-
- PtlEQAlloc_in *args = v_args;
- PtlEQAlloc_out *ret = v_ret;
-
- lib_eq_t *eq;
- unsigned long flags;
-
- /* api should have rounded up */
- if (args->count_in != LOWEST_BIT_SET (args->count_in))
- return ret->rc = PTL_VAL_FAILED;
-
- eq = lib_eq_alloc (nal);
- if (eq == NULL)
- return (ret->rc = PTL_NO_SPACE);
-
- state_lock(nal, &flags);
-
- if (nal->cb_map != NULL) {
- struct iovec iov = {
- .iov_base = args->base_in,
- .iov_len = args->count_in * sizeof (ptl_event_t) };
-
- ret->rc = nal->cb_map (nal, 1, &iov, &eq->eq_addrkey);
- if (ret->rc != PTL_OK) {
- lib_eq_free (nal, eq);
-
- state_unlock (nal, &flags);
- return (ret->rc);
- }
- }
+ /* MUST called with resource lock hold but w/o lnet_eq_wait_lock */
+ int index;
+
+ if (eq->eq_size == 0) {
+ LASSERT(eq->eq_callback != LNET_EQ_HANDLER_NONE);
+ eq->eq_callback(ev);
+ return;
+ }
+
+ lnet_eq_wait_lock();
+ ev->sequence = eq->eq_enq_seq++;
- eq->sequence = 1;
- eq->base = args->base_in;
- eq->size = args->count_in;
- eq->eq_refcount = 0;
- eq->event_callback = args->callback_in;
+ LASSERT(eq->eq_size == LOWEST_BIT_SET(eq->eq_size));
+ index = ev->sequence & (eq->eq_size - 1);
- lib_initialise_handle (nal, &eq->eq_lh, PTL_COOKIE_TYPE_EQ);
- list_add (&eq->eq_list, &nal->ni.ni_active_eqs);
+ eq->eq_events[index] = *ev;
- state_unlock(nal, &flags);
+ if (eq->eq_callback != LNET_EQ_HANDLER_NONE)
+ eq->eq_callback(ev);
- ptl_eq2handle(&ret->handle_out, eq);
- return (ret->rc = PTL_OK);
+#ifdef __KERNEL__
+ /* Wake anyone waiting in LNetEQPoll() */
+ if (waitqueue_active(&the_lnet.ln_eq_waitq))
+ wake_up_all(&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
+ lnet_eq_wait_unlock();
}
-int do_PtlEQFree_internal(nal_cb_t * nal, void *private, void *v_args,
- void *v_ret)
+int
+lnet_eq_dequeue_event(lnet_eq_t *eq, lnet_event_t *ev)
{
- /*
- * Incoming:
- * ptl_handle_eq_t eventq_in
- *
- * Outgoing:
- */
-
- PtlEQFree_in *args = v_args;
- PtlEQFree_out *ret = v_ret;
- lib_eq_t *eq;
- long flags;
-
- state_lock (nal, &flags);
-
- eq = ptl_handle2eq(&args->eventq_in, nal);
- if (eq == NULL) {
- ret->rc = PTL_EQ_INVALID;
- } else if (eq->eq_refcount != 0) {
- ret->rc = PTL_EQ_IN_USE;
+ int new_index = eq->eq_deq_seq & (eq->eq_size - 1);
+ lnet_event_t *new_event = &eq->eq_events[new_index];
+ int rc;
+ ENTRY;
+
+ /* 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;
+
+ 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 {
- if (nal->cb_unmap != NULL) {
- struct iovec iov = {
- .iov_base = eq->base,
- .iov_len = eq->size * sizeof (ptl_event_t) };
-
- nal->cb_unmap(nal, 1, &iov, &eq->eq_addrkey);
- }
-
- lib_invalidate_handle (nal, &eq->eq_lh);
- list_del (&eq->eq_list);
- lib_eq_free (nal, eq);
- ret->rc = PTL_OK;
+ /* 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;
}
- state_unlock (nal, &flags);
+ eq->eq_deq_seq = new_event->sequence + 1;
+ RETURN(rc);
+}
+
+/**
+ * A nonblocking function that can be used to get the next event in an EQ.
+ * If an event handler is associated with the EQ, the handler will run before
+ * this function returns successfully. The event is removed from the queue.
+ *
+ * \param eventq A handle for the event queue.
+ * \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 -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
+ * EQ has been dropped due to limited space in the EQ.
+ */
+int
+LNetEQGet (lnet_handle_eq_t eventq, lnet_event_t *event)
+{
+ int which;
+
+ return LNetEQPoll(&eventq, 1, 0,
+ event, &which);
+}
+EXPORT_SYMBOL(LNetEQGet);
+
+/**
+ * Block the calling process until there is an event in the EQ.
+ * If an event handler is associated with the EQ, the handler will run before
+ * this function returns successfully. This function returns the next event
+ * in the EQ and removes it from the EQ.
+ *
+ * \param eventq A handle for the event queue.
+ * \param event On successful return (1 or -EOVERFLOW), this location will
+ * hold the next 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
+ * EQ has been dropped due to limited space in the EQ.
+ */
+int
+LNetEQWait (lnet_handle_eq_t eventq, lnet_event_t *event)
+{
+ int which;
+
+ return LNetEQPoll(&eventq, 1, LNET_TIME_FOREVER,
+ event, &which);
+}
+EXPORT_SYMBOL(LNetEQWait);
+
+#ifdef __KERNEL__
+
+static int
+lnet_eq_wait_locked(int *timeout_ms)
+__must_hold(&the_lnet.ln_eq_wait_lock)
+{
+ int tms = *timeout_ms;
+ int wait;
+ wait_queue_t wl;
+ cfs_time_t now;
+
+ if (tms == 0)
+ return -1; /* don't want to wait and no new event */
+
+ init_waitqueue_entry_current(&wl);
+ set_current_state(TASK_INTERRUPTIBLE);
+ add_wait_queue(&the_lnet.ln_eq_waitq, &wl);
+
+ lnet_eq_wait_unlock();
+
+ if (tms < 0) {
+ waitq_wait(&wl, TASK_INTERRUPTIBLE);
+
+ } else {
+ struct timeval tv;
+
+ now = cfs_time_current();
+ waitq_timedwait(&wl, 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;
+ }
+
+ wait = tms != 0; /* might need to call here again */
+ *timeout_ms = tms;
+
+ lnet_eq_wait_lock();
+ remove_wait_queue(&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_eq_wait_lock);
+ } else {
+ pthread_cond_timedwait(&the_lnet.ln_eq_cond,
+ &the_lnet.ln_eq_wait_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_eq_wait_unlock();
+
+ lnet_net_lock(0);
+ eq_waitni = the_lnet.ln_eq_waitni;
+ if (unlikely(eq_waitni == NULL)) {
+ lnet_net_unlock(0);
+
+ lnet_eq_wait_lock();
+ return -1;
+ }
+
+ lnet_ni_addref_locked(eq_waitni, 0);
+ lnet_net_unlock(0);
+
+ 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_eq_wait_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;
+
+ 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;
+
+ return wait;
+}
+
+#endif /* __KERNEL__ */
+
+
+/**
+ * Block the calling process until there's an event from a set of EQs or
+ * timeout happens.
+ *
+ * If an event handler is associated with the EQ, the handler will run before
+ * this function returns successfully, in which case the corresponding event
+ * is consumed.
+ *
+ * LNetEQPoll() provides a timeout to allow applications to poll, block for a
+ * 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
+ * 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 -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)
+{
+ int wait = 1;
+ int rc;
+ int i;
+ ENTRY;
+
+ LASSERT (the_lnet.ln_init);
+ LASSERT (the_lnet.ln_refcount > 0);
+
+ if (neq < 1)
+ RETURN(-ENOENT);
+
+ lnet_eq_wait_lock();
+
+ for (;;) {
+#ifndef __KERNEL__
+ lnet_eq_wait_unlock();
+
+ /* 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();
+#endif
+ for (i = 0; i < neq; i++) {
+ lnet_eq_t *eq = lnet_handle2eq(&eventqs[i]);
+
+ if (eq == NULL) {
+ lnet_eq_wait_unlock();
+ RETURN(-ENOENT);
+ }
+
+ rc = lnet_eq_dequeue_event(eq, event);
+ if (rc != 0) {
+ lnet_eq_wait_unlock();
+ *which = i;
+ RETURN(rc);
+ }
+ }
+
+ if (wait == 0)
+ break;
+
+ /*
+ * return value of lnet_eq_wait_locked:
+ * -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
+ */
+ wait = lnet_eq_wait_locked(&timeout_ms);
+ if (wait < 0) /* no new event */
+ break;
+ }
- return (ret->rc);
+ lnet_eq_wait_unlock();
+ RETURN(0);
}