4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
31 * This file is part of Lustre, http://www.lustre.org/
32 * Lustre is a trademark of Sun Microsystems, Inc.
36 * Library level Event queue management routines
39 #define DEBUG_SUBSYSTEM S_LNET
40 #include <lnet/lib-lnet.h>
43 * Create an event queue that has room for \a count number of events.
45 * The event queue is circular and older events will be overwritten by new
46 * ones if they are not removed in time by the user using the functions
47 * LNetEQGet(), LNetEQWait(), or LNetEQPoll(). It is up to the user to
48 * determine the appropriate size of the event queue to prevent this loss
49 * of events. Note that when EQ handler is specified in \a callback, no
50 * event loss can happen, since the handler is run for each event deposited
53 * \param count The number of events to be stored in the event queue. It
54 * will be rounded up to the next power of two.
55 * \param callback A handler function that runs when an event is deposited
56 * into the EQ. The constant value LNET_EQ_HANDLER_NONE can be used to
57 * indicate that no event handler is desired.
58 * \param handle On successful return, this location will hold a handle for
59 * the newly created EQ.
61 * \retval 0 On success.
62 * \retval -EINVAL If an parameter is not valid.
63 * \retval -ENOMEM If memory for the EQ can't be allocated.
65 * \see lnet_eq_handler_t for the discussion on EQ handler semantics.
68 LNetEQAlloc(unsigned int count, lnet_eq_handler_t callback,
69 lnet_handle_eq_t *handle)
73 LASSERT (the_lnet.ln_init);
74 LASSERT (the_lnet.ln_refcount > 0);
76 /* We need count to be a power of 2 so that when eq_{enq,deq}_seq
77 * overflow, they don't skip entries, so the queue has the same
78 * apparent capacity at all times */
80 count = cfs_power2_roundup(count);
82 if (callback != LNET_EQ_HANDLER_NONE && count != 0) {
83 CWARN("EQ callback is guaranteed to get every event, "
84 "do you still want to set eqcount %d for polling "
85 "event which will have locking overhead? "
86 "Please contact with developer to confirm\n", count);
89 /* count can be 0 if only need callback, we can eliminate
90 * overhead of enqueue event */
91 if (count == 0 && callback == LNET_EQ_HANDLER_NONE)
99 LIBCFS_ALLOC(eq->eq_events, count * sizeof(lnet_event_t));
100 if (eq->eq_events == NULL)
102 /* NB allocator has set all event sequence numbers to 0,
103 * so all them should be earlier than eq_deq_seq */
109 eq->eq_callback = callback;
111 eq->eq_refs = cfs_percpt_alloc(lnet_cpt_table(),
112 sizeof(*eq->eq_refs[0]));
113 if (eq->eq_refs == NULL)
116 /* MUST hold both exclusive lnet_res_lock */
117 lnet_res_lock(LNET_LOCK_EX);
118 /* NB: hold lnet_eq_wait_lock for EQ link/unlink, so we can do
119 * both EQ lookup and poll event with only lnet_eq_wait_lock */
122 lnet_res_lh_initialize(&the_lnet.ln_eq_container, &eq->eq_lh);
123 cfs_list_add(&eq->eq_list, &the_lnet.ln_eq_container.rec_active);
125 lnet_eq_wait_unlock();
126 lnet_res_unlock(LNET_LOCK_EX);
128 lnet_eq2handle(handle, eq);
132 if (eq->eq_events != NULL)
133 LIBCFS_FREE(eq->eq_events, count * sizeof(lnet_event_t));
135 if (eq->eq_refs != NULL)
136 cfs_percpt_free(eq->eq_refs);
143 * Release the resources associated with an event queue if it's idle;
144 * otherwise do nothing and it's up to the user to try again.
146 * \param eqh A handle for the event queue to be released.
148 * \retval 0 If the EQ is not in use and freed.
149 * \retval -ENOENT If \a eqh does not point to a valid EQ.
150 * \retval -EBUSY If the EQ is still in use by some MDs.
153 LNetEQFree(lnet_handle_eq_t eqh)
156 lnet_event_t *events = NULL;
163 LASSERT(the_lnet.ln_init);
164 LASSERT(the_lnet.ln_refcount > 0);
166 lnet_res_lock(LNET_LOCK_EX);
167 /* NB: hold lnet_eq_wait_lock for EQ link/unlink, so we can do
168 * both EQ lookup and poll event with only lnet_eq_wait_lock */
171 eq = lnet_handle2eq(&eqh);
177 cfs_percpt_for_each(ref, i, eq->eq_refs) {
182 CDEBUG(D_NET, "Event equeue (%d: %d) busy on destroy.\n",
188 /* stash for free after lock dropped */
189 events = eq->eq_events;
193 lnet_res_lh_invalidate(&eq->eq_lh);
194 cfs_list_del(&eq->eq_list);
195 lnet_eq_free_locked(eq);
197 lnet_eq_wait_unlock();
198 lnet_res_unlock(LNET_LOCK_EX);
201 LIBCFS_FREE(events, size * sizeof(lnet_event_t));
203 cfs_percpt_free(refs);
209 lnet_eq_enqueue_event(lnet_eq_t *eq, lnet_event_t *ev)
211 /* MUST called with resource lock hold but w/o lnet_eq_wait_lock */
214 if (eq->eq_size == 0) {
215 LASSERT(eq->eq_callback != LNET_EQ_HANDLER_NONE);
221 ev->sequence = eq->eq_enq_seq++;
223 LASSERT(eq->eq_size == LOWEST_BIT_SET(eq->eq_size));
224 index = ev->sequence & (eq->eq_size - 1);
226 eq->eq_events[index] = *ev;
228 if (eq->eq_callback != LNET_EQ_HANDLER_NONE)
232 /* Wake anyone waiting in LNetEQPoll() */
233 if (cfs_waitq_active(&the_lnet.ln_eq_waitq))
234 cfs_waitq_broadcast(&the_lnet.ln_eq_waitq);
236 # ifndef HAVE_LIBPTHREAD
237 /* LNetEQPoll() calls into _the_ LND to wait for action */
239 /* Wake anyone waiting in LNetEQPoll() */
240 pthread_cond_broadcast(&the_lnet.ln_eq_cond);
243 lnet_eq_wait_unlock();
247 lnet_eq_dequeue_event(lnet_eq_t *eq, lnet_event_t *ev)
249 int new_index = eq->eq_deq_seq & (eq->eq_size - 1);
250 lnet_event_t *new_event = &eq->eq_events[new_index];
254 /* must called with lnet_eq_wait_lock hold */
255 if (LNET_SEQ_GT(eq->eq_deq_seq, new_event->sequence))
258 /* We've got a new event... */
261 CDEBUG(D_INFO, "event: %p, sequence: %lu, eq->size: %u\n",
262 new_event, eq->eq_deq_seq, eq->eq_size);
264 /* ...but did it overwrite an event we've not seen yet? */
265 if (eq->eq_deq_seq == new_event->sequence) {
268 /* don't complain with CERROR: some EQs are sized small
269 * anyway; if it's important, the caller should complain */
270 CDEBUG(D_NET, "Event Queue Overflow: eq seq %lu ev seq %lu\n",
271 eq->eq_deq_seq, new_event->sequence);
275 eq->eq_deq_seq = new_event->sequence + 1;
280 * A nonblocking function that can be used to get the next event in an EQ.
281 * If an event handler is associated with the EQ, the handler will run before
282 * this function returns successfully. The event is removed from the queue.
284 * \param eventq A handle for the event queue.
285 * \param event On successful return (1 or -EOVERFLOW), this location will
286 * hold the next event in the EQ.
288 * \retval 0 No pending event in the EQ.
289 * \retval 1 Indicates success.
290 * \retval -ENOENT If \a eventq does not point to a valid EQ.
291 * \retval -EOVERFLOW Indicates success (i.e., an event is returned) and that
292 * at least one event between this event and the last event obtained from the
293 * EQ has been dropped due to limited space in the EQ.
296 LNetEQGet (lnet_handle_eq_t eventq, lnet_event_t *event)
300 return LNetEQPoll(&eventq, 1, 0,
305 * Block the calling process until there is an event in the EQ.
306 * If an event handler is associated with the EQ, the handler will run before
307 * this function returns successfully. This function returns the next event
308 * in the EQ and removes it from the EQ.
310 * \param eventq A handle for the event queue.
311 * \param event On successful return (1 or -EOVERFLOW), this location will
312 * hold the next event in the EQ.
314 * \retval 1 Indicates success.
315 * \retval -ENOENT If \a eventq does not point to a valid EQ.
316 * \retval -EOVERFLOW Indicates success (i.e., an event is returned) and that
317 * at least one event between this event and the last event obtained from the
318 * EQ has been dropped due to limited space in the EQ.
321 LNetEQWait (lnet_handle_eq_t eventq, lnet_event_t *event)
325 return LNetEQPoll(&eventq, 1, LNET_TIME_FOREVER,
332 lnet_eq_wait_locked(int *timeout_ms)
334 int tms = *timeout_ms;
340 return -1; /* don't want to wait and no new event */
342 cfs_waitlink_init(&wl);
343 cfs_set_current_state(CFS_TASK_INTERRUPTIBLE);
344 cfs_waitq_add(&the_lnet.ln_eq_waitq, &wl);
346 lnet_eq_wait_unlock();
349 cfs_waitq_wait(&wl, CFS_TASK_INTERRUPTIBLE);
354 now = cfs_time_current();
355 cfs_waitq_timedwait(&wl, CFS_TASK_INTERRUPTIBLE,
356 cfs_time_seconds(tms) / 1000);
357 cfs_duration_usec(cfs_time_sub(cfs_time_current(), now), &tv);
358 tms -= (int)(tv.tv_sec * 1000 + tv.tv_usec / 1000);
359 if (tms < 0) /* no more wait but may have new event */
363 wait = tms != 0; /* might need to call here again */
367 cfs_waitq_del(&the_lnet.ln_eq_waitq, &wl);
372 #else /* !__KERNEL__ */
374 # ifdef HAVE_LIBPTHREAD
376 lnet_eq_cond_wait(struct timespec *ts)
379 pthread_cond_wait(&the_lnet.ln_eq_cond,
380 &the_lnet.ln_eq_wait_lock);
382 pthread_cond_timedwait(&the_lnet.ln_eq_cond,
383 &the_lnet.ln_eq_wait_lock, ts);
389 lnet_eq_wait_locked(int *timeout_ms)
391 lnet_ni_t *eq_waitni = NULL;
392 int tms = *timeout_ms;
397 if (the_lnet.ln_eq_waitni != NULL) {
398 /* I have a single NI that I have to call into, to get
399 * events queued, or to block. */
400 lnet_eq_wait_unlock();
403 eq_waitni = the_lnet.ln_eq_waitni;
404 if (unlikely(eq_waitni == NULL)) {
411 lnet_ni_addref_locked(eq_waitni, 0);
414 if (tms <= 0) { /* even for tms == 0 */
415 (eq_waitni->ni_lnd->lnd_wait)(eq_waitni, tms);
418 gettimeofday(&then, NULL);
420 (eq_waitni->ni_lnd->lnd_wait)(eq_waitni, tms);
422 gettimeofday(&now, NULL);
423 tms -= (now.tv_sec - then.tv_sec) * 1000 +
424 (now.tv_usec - then.tv_usec) / 1000;
429 lnet_ni_decref(eq_waitni);
431 } else { /* w/o eq_waitni */
432 # ifndef HAVE_LIBPTHREAD
433 /* If I'm single-threaded, LNET fails at startup if it can't
434 * set the_lnet.ln_eqwaitni correctly. */
436 # else /* HAVE_LIBPTHREAD */
439 if (tms == 0) /* don't want to wait and new event */
443 lnet_eq_cond_wait(NULL);
447 gettimeofday(&then, NULL);
449 ts.tv_sec = then.tv_sec + tms / 1000;
450 ts.tv_nsec = then.tv_usec * 1000 +
451 (tms % 1000) * 1000000;
452 if (ts.tv_nsec >= 1000000000) {
454 ts.tv_nsec -= 1000000000;
457 lnet_eq_cond_wait(&ts);
459 gettimeofday(&now, NULL);
460 tms -= (now.tv_sec - then.tv_sec) * 1000 +
461 (now.tv_usec - then.tv_usec) / 1000;
465 # endif /* HAVE_LIBPTHREAD */
474 #endif /* __KERNEL__ */
478 * Block the calling process until there's an event from a set of EQs or
481 * If an event handler is associated with the EQ, the handler will run before
482 * this function returns successfully, in which case the corresponding event
485 * LNetEQPoll() provides a timeout to allow applications to poll, block for a
486 * fixed period, or block indefinitely.
488 * \param eventqs,neq An array of EQ handles, and size of the array.
489 * \param timeout_ms Time in milliseconds to wait for an event to occur on
490 * one of the EQs. The constant LNET_TIME_FOREVER can be used to indicate an
492 * \param event,which On successful return (1 or -EOVERFLOW), \a event will
493 * hold the next event in the EQs, and \a which will contain the index of the
494 * EQ from which the event was taken.
496 * \retval 0 No pending event in the EQs after timeout.
497 * \retval 1 Indicates success.
498 * \retval -EOVERFLOW Indicates success (i.e., an event is returned) and that
499 * at least one event between this event and the last event obtained from the
500 * EQ indicated by \a which has been dropped due to limited space in the EQ.
501 * \retval -ENOENT If there's an invalid handle in \a eventqs.
504 LNetEQPoll(lnet_handle_eq_t *eventqs, int neq, int timeout_ms,
505 lnet_event_t *event, int *which)
512 LASSERT (the_lnet.ln_init);
513 LASSERT (the_lnet.ln_refcount > 0);
522 lnet_eq_wait_unlock();
524 /* Recursion breaker */
525 if (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING &&
526 !LNetHandleIsEqual(eventqs[0], the_lnet.ln_rc_eqh))
527 lnet_router_checker();
531 for (i = 0; i < neq; i++) {
532 lnet_eq_t *eq = lnet_handle2eq(&eventqs[i]);
535 lnet_eq_wait_unlock();
539 rc = lnet_eq_dequeue_event(eq, event);
541 lnet_eq_wait_unlock();
551 * return value of lnet_eq_wait_locked:
552 * -1 : did nothing and it's sure no new event
553 * 1 : sleep inside and wait until new event
554 * 0 : don't want to wait anymore, but might have new event
555 * so need to call dequeue again
557 wait = lnet_eq_wait_locked(&timeout_ms);
558 if (wait < 0) /* no new event */
562 lnet_eq_wait_unlock();