X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lnet%2Flnet%2Flib-eq.c;h=8810409ad53a4394249b3dc81c60811abb14d4ff;hp=f24758dcbbb258e9081ab183834bf243549175b6;hb=92c51841c50cc4061c20b277d3f7c4468f2a80cc;hpb=e5c35288861cf8ca4aecc30747b61069f9d80bb5 diff --git a/lnet/lnet/lib-eq.c b/lnet/lnet/lib-eq.c index f24758d..8810409 100644 --- a/lnet/lnet/lib-eq.c +++ b/lnet/lnet/lib-eq.c @@ -1,6 +1,4 @@ -/* -*- 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. @@ -26,8 +24,10 @@ * GPL HEADER END */ /* - * Copyright 2008 Sun Microsystems, Inc. All rights reserved + * 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/ @@ -41,6 +41,31 @@ #define DEBUG_SUBSYSTEM S_LNET #include +/** + * 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. + * + * \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. + * + * \retval 0 On success. + * \retval -EINVAL If an parameter is not valid. + * \retval -ENOMEM If memory for the EQ can't be allocated. + * + * \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) @@ -52,110 +77,193 @@ LNetEQAlloc(unsigned int count, lnet_eq_handler_t callback, /* 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 - * apparant capacity at all times */ - - if (count != LOWEST_BIT_SET(count)) { /* not a power of 2 already */ - do { /* knock off all but the top bit... */ - count &= ~LOWEST_BIT_SET (count); - } while (count != LOWEST_BIT_SET(count)); - - count <<= 1; /* ...and round up */ - } - - if (count == 0) /* catch bad parameter / overflow on roundup */ - return (-EINVAL); - - eq = lnet_eq_alloc(); - if (eq == NULL) - return (-ENOMEM); - - LIBCFS_ALLOC(eq->eq_events, count * sizeof(lnet_event_t)); - if (eq->eq_events == NULL) { - LNET_LOCK(); - lnet_eq_free (eq); - LNET_UNLOCK(); - - return -ENOMEM; - } - - /* NB this resets all event sequence numbers to 0, to be earlier - * than eq_deq_seq */ - memset(eq->eq_events, 0, count * sizeof(lnet_event_t)); + * 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_refcount = 0; eq->eq_callback = callback; - LNET_LOCK(); + 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_initialise_handle (&eq->eq_lh, LNET_COOKIE_TYPE_EQ); - list_add (&eq->eq_list, &the_lnet.ln_active_eqs); + lnet_eq_wait_unlock(); + lnet_res_unlock(LNET_LOCK_EX); - LNET_UNLOCK(); + lnet_eq2handle(handle, eq); + return 0; - 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. + * + * \param eqh A handle for the event queue to be released. + * + * \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) { - lnet_eq_t *eq; - int size; - lnet_event_t *events; - - LASSERT (the_lnet.ln_init); - LASSERT (the_lnet.ln_refcount > 0); + 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; + + 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); + 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); - LNET_LOCK(); +void +lnet_eq_enqueue_event(lnet_eq_t *eq, lnet_event_t *ev) +{ + /* MUST called with resource lock hold but w/o lnet_eq_wait_lock */ + int index; - eq = lnet_handle2eq(&eqh); - if (eq == NULL) { - LNET_UNLOCK(); - return (-ENOENT); - } + if (eq->eq_size == 0) { + LASSERT(eq->eq_callback != LNET_EQ_HANDLER_NONE); + eq->eq_callback(ev); + return; + } - if (eq->eq_refcount != 0) { - CDEBUG(D_NET, "Event queue (%d) busy on destroy.\n", - eq->eq_refcount); - LNET_UNLOCK(); - return (-EBUSY); - } + lnet_eq_wait_lock(); + ev->sequence = eq->eq_enq_seq++; - /* stash for free after lock dropped */ - events = eq->eq_events; - size = eq->eq_size; + LASSERT(eq->eq_size == LOWEST_BIT_SET(eq->eq_size)); + index = ev->sequence & (eq->eq_size - 1); - lnet_invalidate_handle (&eq->eq_lh); - list_del (&eq->eq_list); - lnet_eq_free (eq); + eq->eq_events[index] = *ev; - LNET_UNLOCK(); + if (eq->eq_callback != LNET_EQ_HANDLER_NONE) + eq->eq_callback(ev); - LIBCFS_FREE(events, size * sizeof (lnet_event_t)); - - return 0; +#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 -lib_get_event (lnet_eq_t *eq, lnet_event_t *ev) +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; - CDEBUG(D_INFO, "event: %p, sequence: %lu, eq->size: %u\n", - new_event, eq->eq_deq_seq, eq->eq_size); + /* must called with lnet_eq_wait_lock hold */ + if (LNET_SEQ_GT(eq->eq_deq_seq, new_event->sequence)) + RETURN(0); - 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) { @@ -172,7 +280,22 @@ lib_get_event (lnet_eq_t *eq, lnet_event_t *ev) 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) { @@ -181,7 +304,24 @@ LNetEQGet (lnet_handle_eq_t eventq, lnet_event_t *event) 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) { @@ -190,24 +330,189 @@ LNetEQWait (lnet_handle_eq_t eventq, lnet_event_t *event) return LNetEQPoll(&eventq, 1, LNET_TIME_FOREVER, event, &which); } +EXPORT_SYMBOL(LNetEQWait); -int -LNetEQPoll (lnet_handle_eq_t *eventqs, int neq, int timeout_ms, - lnet_event_t *event, int *which) -{ - int i; - int rc; #ifdef __KERNEL__ - cfs_waitlink_t wl; - cfs_time_t now; -#else - struct timeval then; - struct timeval now; + +static int +lnet_eq_wait_locked(int *timeout_ms) +{ + 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 - struct timespec ts; +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 - lnet_ni_t *eqwaitni = the_lnet.ln_eqwaitni; -#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); @@ -216,121 +521,50 @@ LNetEQPoll (lnet_handle_eq_t *eventqs, int neq, int timeout_ms, if (neq < 1) RETURN(-ENOENT); - LNET_LOCK(); - - for (;;) { - for (i = 0; i < neq; i++) { - lnet_eq_t *eq = lnet_handle2eq(&eventqs[i]); - - if (eq == NULL) { - LNET_UNLOCK(); - RETURN(-ENOENT); - } + lnet_eq_wait_lock(); - rc = lib_get_event (eq, event); - if (rc != 0) { - LNET_UNLOCK(); - *which = i; - RETURN(rc); - } - } + for (;;) { +#ifndef __KERNEL__ + lnet_eq_wait_unlock(); -#ifdef __KERNEL__ - if (timeout_ms == 0) { - LNET_UNLOCK(); - RETURN (0); - } - - cfs_waitlink_init(&wl); - set_current_state(TASK_INTERRUPTIBLE); - cfs_waitq_add(&the_lnet.ln_waitq, &wl); - - LNET_UNLOCK(); - - if (timeout_ms < 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(timeout_ms)/1000); - cfs_duration_usec(cfs_time_sub(cfs_time_current(), now), - &tv); - timeout_ms -= (int)(tv.tv_sec * 1000 + tv.tv_usec / 1000); - if (timeout_ms < 0) - timeout_ms = 0; - } - - LNET_LOCK(); - cfs_waitq_del(&the_lnet.ln_waitq, &wl); -#else - if (eqwaitni != NULL) { - /* I have a single NI that I have to call into, to get - * events queued, or to block. */ - lnet_ni_addref_locked(eqwaitni); - LNET_UNLOCK(); - - if (timeout_ms <= 0) { - (eqwaitni->ni_lnd->lnd_wait)(eqwaitni, timeout_ms); - } else { - gettimeofday(&then, NULL); - - (eqwaitni->ni_lnd->lnd_wait)(eqwaitni, timeout_ms); - - gettimeofday(&now, NULL); - timeout_ms -= (now.tv_sec - then.tv_sec) * 1000 + - (now.tv_usec - then.tv_usec) / 1000; - if (timeout_ms < 0) - timeout_ms = 0; - } - - LNET_LOCK(); - lnet_ni_decref_locked(eqwaitni); - - /* don't call into eqwaitni again if timeout has - * expired */ - if (timeout_ms == 0) - eqwaitni = NULL; - - continue; /* go back and check for events */ - } - - if (timeout_ms == 0) { - LNET_UNLOCK(); - RETURN (0); - } + /* Recursion breaker */ + if (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING && + !LNetHandleIsEqual(eventqs[0], the_lnet.ln_rc_eqh)) + lnet_router_checker(); -# ifndef HAVE_LIBPTHREAD - /* If I'm single-threaded, LNET fails at startup if it can't - * set the_lnet.ln_eqwaitni correctly. */ - LBUG(); -# else - if (timeout_ms < 0) { - pthread_cond_wait(&the_lnet.ln_cond, - &the_lnet.ln_lock); - } else { - gettimeofday(&then, NULL); - - ts.tv_sec = then.tv_sec + timeout_ms/1000; - ts.tv_nsec = then.tv_usec * 1000 + - (timeout_ms%1000) * 1000000; - if (ts.tv_nsec >= 1000000000) { - ts.tv_sec++; - ts.tv_nsec -= 1000000000; - } - - pthread_cond_timedwait(&the_lnet.ln_cond, - &the_lnet.ln_lock, &ts); - - gettimeofday(&now, NULL); - timeout_ms -= (now.tv_sec - then.tv_sec) * 1000 + - (now.tv_usec - then.tv_usec) / 1000; - - if (timeout_ms < 0) - timeout_ms = 0; - } -# endif + 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; + } + + lnet_eq_wait_unlock(); + RETURN(0); }