/*
* Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
* Use is subject to license terms.
+ *
+ * Copyright (c) 2011, 2012, Intel Corporation.
*/
/*
* This file is part of Lustre, http://www.lustre.org/
LASSERT (the_lnet.ln_init);
+ /* NB: use lnet_net_lock(0) to serialize operations on test peers */
if (threshold != 0) {
/* Adding a new entry */
LIBCFS_ALLOC(tp, sizeof(*tp));
tp->tp_nid = nid;
tp->tp_threshold = threshold;
- LNET_LOCK();
- cfs_list_add_tail (&tp->tp_list, &the_lnet.ln_test_peers);
- LNET_UNLOCK();
- return 0;
- }
+ lnet_net_lock(0);
+ cfs_list_add_tail(&tp->tp_list, &the_lnet.ln_test_peers);
+ lnet_net_unlock(0);
+ return 0;
+ }
- /* removing entries */
- CFS_INIT_LIST_HEAD (&cull);
+ /* removing entries */
+ CFS_INIT_LIST_HEAD(&cull);
- LNET_LOCK();
+ lnet_net_lock(0);
cfs_list_for_each_safe (el, next, &the_lnet.ln_test_peers) {
tp = cfs_list_entry (el, lnet_test_peer_t, tp_list);
}
}
- LNET_UNLOCK();
+ lnet_net_unlock(0);
while (!cfs_list_empty (&cull)) {
tp = cfs_list_entry (cull.next, lnet_test_peer_t, tp_list);
CFS_INIT_LIST_HEAD (&cull);
- LNET_LOCK();
+ /* NB: use lnet_net_lock(0) to serialize operations on test peers */
+ lnet_net_lock(0);
cfs_list_for_each_safe (el, next, &the_lnet.ln_test_peers) {
tp = cfs_list_entry (el, lnet_test_peer_t, tp_list);
}
}
- LNET_UNLOCK ();
+ lnet_net_unlock(0);
while (!cfs_list_empty (&cull)) {
tp = cfs_list_entry (cull.next, lnet_test_peer_t, tp_list);
return (nob);
}
+EXPORT_SYMBOL(lnet_iov_nob);
void
lnet_copy_iov2iov (unsigned int ndiov, struct iovec *diov, unsigned int doffset,
}
} while (nob > 0);
}
+EXPORT_SYMBOL(lnet_copy_iov2iov);
int
lnet_extract_iov (int dst_niov, struct iovec *dst,
offset = 0;
}
}
+EXPORT_SYMBOL(lnet_extract_iov);
#ifndef __KERNEL__
unsigned int
return (nob);
}
+EXPORT_SYMBOL(lnet_kiov_nob);
void
lnet_copy_kiov2kiov (unsigned int ndiov, lnet_kiov_t *diov, unsigned int doffset,
if (saddr != NULL)
cfs_kunmap(siov->kiov_page);
}
+EXPORT_SYMBOL(lnet_copy_kiov2kiov);
void
lnet_copy_kiov2iov (unsigned int niov, struct iovec *iov, unsigned int iovoffset,
if (addr != NULL)
cfs_kunmap(kiov->kiov_page);
}
+EXPORT_SYMBOL(lnet_copy_kiov2iov);
void
lnet_copy_iov2kiov (unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
if (addr != NULL)
cfs_kunmap(kiov->kiov_page);
}
+EXPORT_SYMBOL(lnet_copy_iov2kiov);
int
lnet_extract_kiov (int dst_niov, lnet_kiov_t *dst,
offset = 0;
}
}
+EXPORT_SYMBOL(lnet_extract_kiov);
#endif
void
LASSERT(!msg->msg_sending);
LASSERT(rlen == msg->msg_len);
LASSERT(mlen <= msg->msg_len);
+ LASSERT(msg->msg_offset == offset);
+ LASSERT(msg->msg_wanted == mlen);
- msg->msg_wanted = mlen;
- msg->msg_offset = offset;
msg->msg_receiving = 0;
if (mlen != 0) {
lnet_finalize(ni, msg, rc);
}
-int
-lnet_compare_routes(lnet_route_t *r1, lnet_route_t *r2)
-{
- lnet_peer_t *p1 = r1->lr_gateway;
- lnet_peer_t *p2 = r2->lr_gateway;
-
- if (r1->lr_hops < r2->lr_hops)
- return 1;
-
- if (r1->lr_hops > r2->lr_hops)
- return -1;
-
- if (p1->lp_txqnob < p2->lp_txqnob)
- return 1;
-
- if (p1->lp_txqnob > p2->lp_txqnob)
- return -1;
-
- if (p1->lp_txcredits > p2->lp_txcredits)
- return 1;
-
- if (p1->lp_txcredits < p2->lp_txcredits)
- return -1;
-
- return 0;
-}
-
-
void
lnet_setpayloadbuffer(lnet_msg_t *msg)
{
}
int
-lnet_eager_recv_locked(lnet_msg_t *msg)
+lnet_ni_eager_recv(lnet_ni_t *ni, lnet_msg_t *msg)
{
- lnet_peer_t *peer;
- lnet_ni_t *ni;
- int rc = 0;
-
- LASSERT (!msg->msg_delayed);
- msg->msg_delayed = 1;
-
- LASSERT (msg->msg_receiving);
- LASSERT (!msg->msg_sending);
-
- peer = msg->msg_rxpeer;
- ni = peer->lp_ni;
-
- if (ni->ni_lnd->lnd_eager_recv != NULL) {
- LNET_UNLOCK();
-
- rc = (ni->ni_lnd->lnd_eager_recv)(ni, msg->msg_private, msg,
- &msg->msg_private);
- if (rc != 0) {
- CERROR("recv from %s / send to %s aborted: "
- "eager_recv failed %d\n",
- libcfs_nid2str(peer->lp_nid),
- libcfs_id2str(msg->msg_target), rc);
- LASSERT (rc < 0); /* required by my callers */
- }
-
- LNET_LOCK();
- }
+ int rc;
+
+ LASSERT(!msg->msg_sending);
+ LASSERT(msg->msg_receiving);
+ LASSERT(!msg->msg_rx_ready_delay);
+ LASSERT(ni->ni_lnd->lnd_eager_recv != NULL);
+
+ msg->msg_rx_ready_delay = 1;
+ rc = (ni->ni_lnd->lnd_eager_recv)(ni, msg->msg_private, msg,
+ &msg->msg_private);
+ if (rc != 0) {
+ CERROR("recv from %s / send to %s aborted: "
+ "eager_recv failed %d\n",
+ libcfs_nid2str(msg->msg_rxpeer->lp_nid),
+ libcfs_id2str(msg->msg_target), rc);
+ LASSERT(rc < 0); /* required by my callers */
+ }
- return rc;
+ return rc;
}
-/* NB: caller shall hold a ref on 'lp' as I'd drop LNET_LOCK */
+/* NB: caller shall hold a ref on 'lp' as I'd drop lnet_net_lock */
void
-lnet_ni_peer_alive(lnet_peer_t *lp)
+lnet_ni_query_locked(lnet_ni_t *ni, lnet_peer_t *lp)
{
- cfs_time_t last_alive = 0;
- lnet_ni_t *ni = lp->lp_ni;
+ cfs_time_t last_alive = 0;
- LASSERT (lnet_peer_aliveness_enabled(lp));
- LASSERT (ni->ni_lnd->lnd_query != NULL);
- LASSERT (the_lnet.ln_routing == 1);
+ LASSERT(lnet_peer_aliveness_enabled(lp));
+ LASSERT(ni->ni_lnd->lnd_query != NULL);
- LNET_UNLOCK();
- (ni->ni_lnd->lnd_query)(ni, lp->lp_nid, &last_alive);
- LNET_LOCK();
+ lnet_net_unlock(lp->lp_cpt);
+ (ni->ni_lnd->lnd_query)(ni, lp->lp_nid, &last_alive);
+ lnet_net_lock(lp->lp_cpt);
- lp->lp_last_query = cfs_time_current();
+ lp->lp_last_query = cfs_time_current();
- if (last_alive != 0) /* NI has updated timestamp */
- lp->lp_last_alive = last_alive;
+ if (last_alive != 0) /* NI has updated timestamp */
+ lp->lp_last_alive = last_alive;
}
-/* NB: always called with LNET_LOCK held */
+/* NB: always called with lnet_net_lock held */
static inline int
lnet_peer_is_alive (lnet_peer_t *lp, cfs_time_t now)
{
cfs_time_t deadline;
LASSERT (lnet_peer_aliveness_enabled(lp));
- LASSERT (the_lnet.ln_routing == 1);
/* Trust lnet_notify() if it has more recent aliveness news, but
* ignore the initial assumed death (see lnet_peers_start_down()).
/* NB: returns 1 when alive, 0 when dead, negative when error;
- * may drop the LNET_LOCK */
+ * may drop the lnet_net_lock */
int
lnet_peer_alive_locked (lnet_peer_t *lp)
{
cfs_time_t now = cfs_time_current();
- /* LU-630: only router checks peer health. */
- if (the_lnet.ln_routing == 0)
- return 1;
-
if (!lnet_peer_aliveness_enabled(lp))
return -ENODEV;
}
/* query NI for latest aliveness news */
- lnet_ni_peer_alive(lp);
+ lnet_ni_query_locked(lp->lp_ni, lp);
if (lnet_peer_is_alive(lp, now))
return 1;
}
int
-lnet_post_send_locked (lnet_msg_t *msg, int do_send)
+lnet_post_send_locked(lnet_msg_t *msg, int do_send)
{
- /* lnet_send is going to LNET_UNLOCK immediately after this, so it sets
- * do_send FALSE and I don't do the unlock/send/lock bit. I return
- * EAGAIN if msg blocked, EHOSTUNREACH if msg_txpeer appears dead, and
- * 0 if sent or OK to send */
- lnet_peer_t *lp = msg->msg_txpeer;
- lnet_ni_t *ni = lp->lp_ni;
-
- /* non-lnet_send() callers have checked before */
- LASSERT (!do_send || msg->msg_delayed);
- LASSERT (!msg->msg_receiving);
-
- /* NB 'lp' is always the next hop */
- if ((msg->msg_target.pid & LNET_PID_USERFLAG) == 0 &&
- lnet_peer_alive_locked(lp) == 0) {
- the_lnet.ln_counters.drop_count++;
- the_lnet.ln_counters.drop_length += msg->msg_len;
- LNET_UNLOCK();
-
- CNETERR("Dropping message for %s: peer not alive\n",
- libcfs_id2str(msg->msg_target));
- if (do_send)
- lnet_finalize(ni, msg, -EHOSTUNREACH);
-
- LNET_LOCK();
- return EHOSTUNREACH;
- }
+ /* lnet_send is going to lnet_net_unlock immediately after this,
+ * so it sets do_send FALSE and I don't do the unlock/send/lock bit.
+ * I return EAGAIN if msg blocked, EHOSTUNREACH if msg_txpeer
+ * appears dead, and 0 if sent or OK to send */
+ struct lnet_peer *lp = msg->msg_txpeer;
+ struct lnet_ni *ni = lp->lp_ni;
+ struct lnet_tx_queue *tq;
+ int cpt;
+
+ /* non-lnet_send() callers have checked before */
+ LASSERT(!do_send || msg->msg_tx_delayed);
+ LASSERT(!msg->msg_receiving);
+ LASSERT(msg->msg_tx_committed);
+
+ cpt = msg->msg_tx_cpt;
+ tq = ni->ni_tx_queues[cpt];
+
+ /* NB 'lp' is always the next hop */
+ if ((msg->msg_target.pid & LNET_PID_USERFLAG) == 0 &&
+ lnet_peer_alive_locked(lp) == 0) {
+ the_lnet.ln_counters[cpt]->drop_count++;
+ the_lnet.ln_counters[cpt]->drop_length += msg->msg_len;
+ lnet_net_unlock(cpt);
+
+ CNETERR("Dropping message for %s: peer not alive\n",
+ libcfs_id2str(msg->msg_target));
+ if (do_send)
+ lnet_finalize(ni, msg, -EHOSTUNREACH);
+
+ lnet_net_lock(cpt);
+ return EHOSTUNREACH;
+ }
if (!msg->msg_peertxcredit) {
LASSERT ((lp->lp_txcredits < 0) ==
lp->lp_mintxcredits = lp->lp_txcredits;
if (lp->lp_txcredits < 0) {
- msg->msg_delayed = 1;
+ msg->msg_tx_delayed = 1;
cfs_list_add_tail(&msg->msg_list, &lp->lp_txq);
return EAGAIN;
}
}
if (!msg->msg_txcredit) {
- LASSERT ((ni->ni_txcredits < 0) ==
- !cfs_list_empty(&ni->ni_txq));
+ LASSERT((tq->tq_credits < 0) ==
+ !cfs_list_empty(&tq->tq_delayed));
- msg->msg_txcredit = 1;
- ni->ni_txcredits--;
+ msg->msg_txcredit = 1;
+ tq->tq_credits--;
- if (ni->ni_txcredits < ni->ni_mintxcredits)
- ni->ni_mintxcredits = ni->ni_txcredits;
+ if (tq->tq_credits < tq->tq_credits_min)
+ tq->tq_credits_min = tq->tq_credits;
- if (ni->ni_txcredits < 0) {
- msg->msg_delayed = 1;
- cfs_list_add_tail(&msg->msg_list, &ni->ni_txq);
- return EAGAIN;
- }
- }
+ if (tq->tq_credits < 0) {
+ msg->msg_tx_delayed = 1;
+ cfs_list_add_tail(&msg->msg_list, &tq->tq_delayed);
+ return EAGAIN;
+ }
+ }
- if (do_send) {
- LNET_UNLOCK();
- lnet_ni_send(ni, msg);
- LNET_LOCK();
- }
- return 0;
+ if (do_send) {
+ lnet_net_unlock(cpt);
+ lnet_ni_send(ni, msg);
+ lnet_net_lock(cpt);
+ }
+ return 0;
}
#ifdef __KERNEL__
-static void
-lnet_commit_routedmsg (lnet_msg_t *msg)
-{
- /* ALWAYS called holding the LNET_LOCK */
- LASSERT (msg->msg_routing);
-
- the_lnet.ln_counters.msgs_alloc++;
- if (the_lnet.ln_counters.msgs_alloc >
- the_lnet.ln_counters.msgs_max)
- the_lnet.ln_counters.msgs_max =
- the_lnet.ln_counters.msgs_alloc;
-
- the_lnet.ln_counters.route_count++;
- the_lnet.ln_counters.route_length += msg->msg_len;
-
- LASSERT (!msg->msg_onactivelist);
- msg->msg_onactivelist = 1;
- cfs_list_add(&msg->msg_activelist,
- &the_lnet.ln_msg_container.msc_active);
-}
lnet_rtrbufpool_t *
lnet_msg2bufpool(lnet_msg_t *msg)
{
- lnet_rtrbufpool_t *rbp = &the_lnet.ln_rtrpools[0];
+ lnet_rtrbufpool_t *rbp;
+ int cpt;
- LASSERT (msg->msg_len <= LNET_MTU);
- while (msg->msg_len > (unsigned int)rbp->rbp_npages * CFS_PAGE_SIZE) {
- rbp++;
- LASSERT (rbp < &the_lnet.ln_rtrpools[LNET_NRBPOOLS]);
- }
+ LASSERT(msg->msg_rx_committed);
+
+ cpt = msg->msg_rx_cpt;
+ rbp = &the_lnet.ln_rtrpools[cpt][0];
+
+ LASSERT(msg->msg_len <= LNET_MTU);
+ while (msg->msg_len > (unsigned int)rbp->rbp_npages * CFS_PAGE_SIZE) {
+ rbp++;
+ LASSERT(rbp < &the_lnet.ln_rtrpools[cpt][LNET_NRBPOOLS]);
+ }
- return rbp;
+ return rbp;
}
int
lnet_post_routed_recv_locked (lnet_msg_t *msg, int do_recv)
{
- /* lnet_parse is going to LNET_UNLOCK immediately after this, so it
- * sets do_recv FALSE and I don't do the unlock/send/lock bit. I
- * return EAGAIN if msg blocked and 0 if received or OK to receive */
+ /* lnet_parse is going to lnet_net_unlock immediately after this, so it
+ * sets do_recv FALSE and I don't do the unlock/send/lock bit. I
+ * return EAGAIN if msg blocked and 0 if received or OK to receive */
lnet_peer_t *lp = msg->msg_rxpeer;
lnet_rtrbufpool_t *rbp;
lnet_rtrbuf_t *rb;
LASSERT (msg->msg_receiving);
LASSERT (!msg->msg_sending);
- /* non-lnet_parse callers only send delayed messages */
- LASSERT (!do_recv || msg->msg_delayed);
+ /* non-lnet_parse callers only receive delayed messages */
+ LASSERT(!do_recv || msg->msg_rx_delayed);
if (!msg->msg_peerrtrcredit) {
LASSERT ((lp->lp_rtrcredits < 0) ==
if (lp->lp_rtrcredits < 0) {
/* must have checked eager_recv before here */
- LASSERT (msg->msg_delayed);
+ LASSERT(msg->msg_rx_ready_delay);
+ msg->msg_rx_delayed = 1;
cfs_list_add_tail(&msg->msg_list, &lp->lp_rtrq);
return EAGAIN;
}
if (rbp->rbp_credits < 0) {
/* must have checked eager_recv before here */
- LASSERT (msg->msg_delayed);
+ LASSERT(msg->msg_rx_ready_delay);
+ msg->msg_rx_delayed = 1;
cfs_list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
return EAGAIN;
}
msg->msg_kiov = &rb->rb_kiov[0];
if (do_recv) {
- LNET_UNLOCK();
- lnet_ni_recv(lp->lp_ni, msg->msg_private, msg, 1,
- 0, msg->msg_len, msg->msg_len);
- LNET_LOCK();
- }
- return 0;
+ int cpt = msg->msg_rx_cpt;
+
+ lnet_net_unlock(cpt);
+ lnet_ni_recv(lp->lp_ni, msg->msg_private, msg, 1,
+ 0, msg->msg_len, msg->msg_len);
+ lnet_net_lock(cpt);
+ }
+ return 0;
}
#endif
void
-lnet_return_credits_locked (lnet_msg_t *msg)
+lnet_return_tx_credits_locked(lnet_msg_t *msg)
{
- lnet_peer_t *txpeer = msg->msg_txpeer;
- lnet_peer_t *rxpeer = msg->msg_rxpeer;
- lnet_msg_t *msg2;
- lnet_ni_t *ni;
-
- if (msg->msg_txcredit) {
- /* give back NI txcredits */
- msg->msg_txcredit = 0;
- ni = txpeer->lp_ni;
-
- LASSERT((ni->ni_txcredits < 0) == !cfs_list_empty(&ni->ni_txq));
-
- ni->ni_txcredits++;
- if (ni->ni_txcredits <= 0) {
- msg2 = cfs_list_entry(ni->ni_txq.next, lnet_msg_t,
- msg_list);
- cfs_list_del(&msg2->msg_list);
+ lnet_peer_t *txpeer = msg->msg_txpeer;
+ lnet_msg_t *msg2;
+
+ if (msg->msg_txcredit) {
+ struct lnet_ni *ni = txpeer->lp_ni;
+ struct lnet_tx_queue *tq = ni->ni_tx_queues[msg->msg_tx_cpt];
+
+ /* give back NI txcredits */
+ msg->msg_txcredit = 0;
+
+ LASSERT((tq->tq_credits < 0) ==
+ !cfs_list_empty(&tq->tq_delayed));
- LASSERT(msg2->msg_txpeer->lp_ni == ni);
- LASSERT(msg2->msg_delayed);
+ tq->tq_credits++;
+ if (tq->tq_credits <= 0) {
+ msg2 = cfs_list_entry(tq->tq_delayed.next,
+ lnet_msg_t, msg_list);
+ cfs_list_del(&msg2->msg_list);
+
+ LASSERT(msg2->msg_txpeer->lp_ni == ni);
+ LASSERT(msg2->msg_tx_delayed);
(void) lnet_post_send_locked(msg2, 1);
}
lnet_msg_t, msg_list);
cfs_list_del(&msg2->msg_list);
- LASSERT (msg2->msg_txpeer == txpeer);
- LASSERT (msg2->msg_delayed);
+ LASSERT(msg2->msg_txpeer == txpeer);
+ LASSERT(msg2->msg_tx_delayed);
(void) lnet_post_send_locked(msg2, 1);
}
msg->msg_txpeer = NULL;
lnet_peer_decref_locked(txpeer);
}
+}
+void
+lnet_return_rx_credits_locked(lnet_msg_t *msg)
+{
+ lnet_peer_t *rxpeer = msg->msg_rxpeer;
#ifdef __KERNEL__
+ lnet_msg_t *msg2;
+
if (msg->msg_rtrcredit) {
/* give back global router credits */
lnet_rtrbuf_t *rb;
}
}
+static int
+lnet_compare_routes(lnet_route_t *r1, lnet_route_t *r2)
+{
+ lnet_peer_t *p1 = r1->lr_gateway;
+ lnet_peer_t *p2 = r2->lr_gateway;
+
+ if (r1->lr_hops < r2->lr_hops)
+ return 1;
+
+ if (r1->lr_hops > r2->lr_hops)
+ return -1;
+
+ if (p1->lp_txqnob < p2->lp_txqnob)
+ return 1;
+
+ if (p1->lp_txqnob > p2->lp_txqnob)
+ return -1;
+
+ if (p1->lp_txcredits > p2->lp_txcredits)
+ return 1;
+
+ if (p1->lp_txcredits < p2->lp_txcredits)
+ return -1;
+
+ if (r1->lr_seq - r2->lr_seq <= 0)
+ return 1;
+
+ return -1;
+}
+
static lnet_peer_t *
-lnet_find_route_locked(lnet_ni_t *ni, lnet_nid_t target)
+lnet_find_route_locked(lnet_ni_t *ni, lnet_nid_t target, lnet_nid_t rtr_nid)
{
lnet_remotenet_t *rnet;
lnet_route_t *rtr;
struct lnet_peer *lp;
int rc;
+ /* If @rtr_nid is not LNET_NID_ANY, return the gateway with
+ * rtr_nid nid, otherwise find the best gateway I can use */
+
rnet = lnet_find_net_locked(LNET_NIDNET(target));
if (rnet == NULL)
return NULL;
lp = rtr->lr_gateway;
if (!lp->lp_alive || /* gateway is down */
- (lp->lp_ping_version == LNET_PROTO_PING_VERSION &&
+ ((lp->lp_ping_feats & LNET_PING_FEAT_NI_STATUS) != 0 &&
rtr->lr_downis != 0)) /* NI to target is down */
continue;
if (ni != NULL && lp->lp_ni != ni)
continue;
+ if (lp->lp_nid == rtr_nid) /* it's pre-determined router */
+ return lp;
+
if (lp_best == NULL) {
rtr_best = rtr_last = rtr;
lp_best = lp;
continue;
}
+ /* no protection on below fields, but it's harmless */
+ if (rtr_last->lr_seq - rtr->lr_seq < 0)
+ rtr_last = rtr;
+
rc = lnet_compare_routes(rtr, rtr_best);
if (rc < 0)
continue;
lp_best = lp;
}
- if (rtr_best != NULL) {
- /* Place selected route at the end of the route list to ensure
- * fairness; everything else being equal... */
- cfs_list_del(&rtr_best->lr_list);
- cfs_list_add_tail(&rtr_best->lr_list, &rnet->lrn_routes);
- }
-
+ /* set sequence number on the best router to the latest sequence + 1
+ * so we can round-robin all routers, it's race and inaccurate but
+ * harmless and functional */
+ if (rtr_best != NULL)
+ rtr_best->lr_seq = rtr_last->lr_seq + 1;
return lp_best;
}
int
-lnet_send(lnet_nid_t src_nid, lnet_msg_t *msg)
+lnet_send(lnet_nid_t src_nid, lnet_msg_t *msg, lnet_nid_t rtr_nid)
{
- lnet_nid_t dst_nid = msg->msg_target.nid;
- lnet_ni_t *src_ni;
- lnet_ni_t *local_ni;
- lnet_peer_t *lp;
- int rc;
+ lnet_nid_t dst_nid = msg->msg_target.nid;
+ struct lnet_ni *src_ni;
+ struct lnet_ni *local_ni;
+ struct lnet_peer *lp;
+ int cpt;
+ int cpt2;
+ int rc;
+ /* NB: rtr_nid is set to LNET_NID_ANY for all current use-cases,
+ * but we might want to use pre-determined router for ACK/REPLY
+ * in the future */
+ /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
LASSERT (msg->msg_txpeer == NULL);
LASSERT (!msg->msg_sending);
LASSERT (!msg->msg_target_is_router);
msg->msg_sending = 1;
- /* NB! ni != NULL == interface pre-determined (ACK/REPLY) */
-
- LNET_LOCK();
+ LASSERT(!msg->msg_tx_committed);
+ cpt = lnet_cpt_of_nid(rtr_nid == LNET_NID_ANY ? dst_nid : rtr_nid);
+ again:
+ lnet_net_lock(cpt);
- if (the_lnet.ln_shutdown) {
- LNET_UNLOCK();
- return -ESHUTDOWN;
- }
+ if (the_lnet.ln_shutdown) {
+ lnet_net_unlock(cpt);
+ return -ESHUTDOWN;
+ }
- if (src_nid == LNET_NID_ANY) {
- src_ni = NULL;
- } else {
- src_ni = lnet_nid2ni_locked(src_nid);
- if (src_ni == NULL) {
- LNET_UNLOCK();
+ if (src_nid == LNET_NID_ANY) {
+ src_ni = NULL;
+ } else {
+ src_ni = lnet_nid2ni_locked(src_nid, cpt);
+ if (src_ni == NULL) {
+ lnet_net_unlock(cpt);
LCONSOLE_WARN("Can't send to %s: src %s is not a "
"local nid\n", libcfs_nid2str(dst_nid),
libcfs_nid2str(src_nid));
}
/* Is this for someone on a local network? */
- local_ni = lnet_net2ni_locked(LNET_NIDNET(dst_nid));
+ local_ni = lnet_net2ni_locked(LNET_NIDNET(dst_nid), cpt);
if (local_ni != NULL) {
if (src_ni == NULL) {
src_ni = local_ni;
src_nid = src_ni->ni_nid;
} else if (src_ni == local_ni) {
- lnet_ni_decref_locked(local_ni);
- } else {
- lnet_ni_decref_locked(local_ni);
- lnet_ni_decref_locked(src_ni);
- LNET_UNLOCK();
- LCONSOLE_WARN("No route to %s via from %s\n",
- libcfs_nid2str(dst_nid),
- libcfs_nid2str(src_nid));
- return -EINVAL;
- }
+ lnet_ni_decref_locked(local_ni, cpt);
+ } else {
+ lnet_ni_decref_locked(local_ni, cpt);
+ lnet_ni_decref_locked(src_ni, cpt);
+ lnet_net_unlock(cpt);
+ LCONSOLE_WARN("No route to %s via from %s\n",
+ libcfs_nid2str(dst_nid),
+ libcfs_nid2str(src_nid));
+ return -EINVAL;
+ }
- LASSERT (src_nid != LNET_NID_ANY);
+ LASSERT(src_nid != LNET_NID_ANY);
+ lnet_msg_commit(msg, cpt);
- if (!msg->msg_routing)
- msg->msg_hdr.src_nid = cpu_to_le64(src_nid);
+ if (!msg->msg_routing)
+ msg->msg_hdr.src_nid = cpu_to_le64(src_nid);
- if (src_ni == the_lnet.ln_loni) {
- /* No send credit hassles with LOLND */
- LNET_UNLOCK();
- lnet_ni_send(src_ni, msg);
- lnet_ni_decref(src_ni);
- return 0;
- }
+ if (src_ni == the_lnet.ln_loni) {
+ /* No send credit hassles with LOLND */
+ lnet_net_unlock(cpt);
+ lnet_ni_send(src_ni, msg);
- rc = lnet_nid2peer_locked(&lp, dst_nid);
- lnet_ni_decref_locked(src_ni); /* lp has ref on src_ni; lose mine */
- if (rc != 0) {
- LNET_UNLOCK();
+ lnet_net_lock(cpt);
+ lnet_ni_decref_locked(src_ni, cpt);
+ lnet_net_unlock(cpt);
+ return 0;
+ }
+
+ rc = lnet_nid2peer_locked(&lp, dst_nid, cpt);
+ /* lp has ref on src_ni; lose mine */
+ lnet_ni_decref_locked(src_ni, cpt);
+ if (rc != 0) {
+ lnet_net_unlock(cpt);
LCONSOLE_WARN("Error %d finding peer %s\n", rc,
libcfs_nid2str(dst_nid));
/* ENOMEM or shutting down */
LASSERT (lp->lp_ni == src_ni);
} else {
#ifndef __KERNEL__
- LNET_UNLOCK();
+ lnet_net_unlock(cpt);
/* NB
* - once application finishes computation, check here to update
if (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING)
lnet_router_checker();
- LNET_LOCK();
+ lnet_net_lock(cpt);
#endif
- /* sending to a remote network */
- lp = lnet_find_route_locked(src_ni, dst_nid);
- if (lp == NULL) {
- if (src_ni != NULL)
- lnet_ni_decref_locked(src_ni);
- LNET_UNLOCK();
+ /* sending to a remote network */
+ lp = lnet_find_route_locked(src_ni, dst_nid, rtr_nid);
+ if (lp == NULL) {
+ if (src_ni != NULL)
+ lnet_ni_decref_locked(src_ni, cpt);
+ lnet_net_unlock(cpt);
LCONSOLE_WARN("No route to %s via %s "
"(all routers down)\n",
return -EHOSTUNREACH;
}
+ /* rtr_nid is LNET_NID_ANY or NID of pre-determined router,
+ * it's possible that rtr_nid isn't LNET_NID_ANY and lp isn't
+ * pre-determined router, this can happen if router table
+ * was changed when we release the lock */
+ if (rtr_nid != lp->lp_nid) {
+ cpt2 = lnet_cpt_of_nid_locked(lp->lp_nid);
+ if (cpt2 != cpt) {
+ if (src_ni != NULL)
+ lnet_ni_decref_locked(src_ni, cpt);
+ lnet_net_unlock(cpt);
+
+ rtr_nid = lp->lp_nid;
+ cpt = cpt2;
+ goto again;
+ }
+ }
+
CDEBUG(D_NET, "Best route to %s via %s for %s %d\n",
libcfs_nid2str(dst_nid), libcfs_nid2str(lp->lp_nid),
lnet_msgtyp2str(msg->msg_type), msg->msg_len);
src_nid = src_ni->ni_nid;
} else {
LASSERT (src_ni == lp->lp_ni);
- lnet_ni_decref_locked(src_ni);
- }
+ lnet_ni_decref_locked(src_ni, cpt);
+ }
- lnet_peer_addref_locked(lp);
+ lnet_peer_addref_locked(lp);
- LASSERT (src_nid != LNET_NID_ANY);
+ LASSERT(src_nid != LNET_NID_ANY);
+ lnet_msg_commit(msg, cpt);
if (!msg->msg_routing) {
/* I'm the source and now I know which NI to send on */
msg->msg_txpeer = lp; /* msg takes my ref on lp */
rc = lnet_post_send_locked(msg, 0);
- LNET_UNLOCK();
+ lnet_net_unlock(cpt);
if (rc == EHOSTUNREACH)
return -EHOSTUNREACH;
return 0;
}
-void
-lnet_commit_md (lnet_libmd_t *md, lnet_msg_t *msg)
-{
- /* ALWAYS called holding the LNET_LOCK */
- /* Here, we commit the MD to a network OP by marking it busy and
- * decrementing its threshold. Come what may, the network "owns"
- * the MD until a call to lnet_finalize() signals completion. */
- LASSERT (!msg->msg_routing);
-
- msg->msg_md = md;
-
- md->md_refcount++;
- if (md->md_threshold != LNET_MD_THRESH_INF) {
- LASSERT (md->md_threshold > 0);
- md->md_threshold--;
- }
-
- the_lnet.ln_counters.msgs_alloc++;
- if (the_lnet.ln_counters.msgs_alloc >
- the_lnet.ln_counters.msgs_max)
- the_lnet.ln_counters.msgs_max =
- the_lnet.ln_counters.msgs_alloc;
-
- LASSERT (!msg->msg_onactivelist);
- msg->msg_onactivelist = 1;
- cfs_list_add(&msg->msg_activelist,
- &the_lnet.ln_msg_container.msc_active);
-}
-
static void
-lnet_drop_message (lnet_ni_t *ni, void *private, unsigned int nob)
+lnet_drop_message(lnet_ni_t *ni, int cpt, void *private, unsigned int nob)
{
- LNET_LOCK();
- the_lnet.ln_counters.drop_count++;
- the_lnet.ln_counters.drop_length += nob;
- LNET_UNLOCK();
+ lnet_net_lock(cpt);
+ the_lnet.ln_counters[cpt]->drop_count++;
+ the_lnet.ln_counters[cpt]->drop_length += nob;
+ lnet_net_unlock(cpt);
- lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
+ lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
}
static void
-lnet_recv_put(lnet_libmd_t *md, lnet_msg_t *msg, int delayed,
- unsigned int offset, unsigned int mlength)
+lnet_recv_put(lnet_ni_t *ni, lnet_msg_t *msg)
{
- lnet_hdr_t *hdr = &msg->msg_hdr;
+ lnet_hdr_t *hdr = &msg->msg_hdr;
- LNET_LOCK();
+ if (msg->msg_wanted != 0)
+ lnet_setpayloadbuffer(msg);
- the_lnet.ln_counters.recv_count++;
- the_lnet.ln_counters.recv_length += mlength;
+ lnet_build_msg_event(msg, LNET_EVENT_PUT);
- LNET_UNLOCK();
-
- if (mlength != 0)
- lnet_setpayloadbuffer(msg);
+ /* Must I ACK? If so I'll grab the ack_wmd out of the header and put
+ * it back into the ACK during lnet_finalize() */
+ msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
+ (msg->msg_md->md_options & LNET_MD_ACK_DISABLE) == 0);
- msg->msg_ev.type = LNET_EVENT_PUT;
- msg->msg_ev.target.pid = hdr->dest_pid;
- msg->msg_ev.target.nid = hdr->dest_nid;
- msg->msg_ev.hdr_data = hdr->msg.put.hdr_data;
-
- /* Must I ACK? If so I'll grab the ack_wmd out of the header and put
- * it back into the ACK during lnet_finalize() */
- msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
- (md->md_options & LNET_MD_ACK_DISABLE) == 0);
-
- lnet_ni_recv(msg->msg_rxpeer->lp_ni,
- msg->msg_private,
- msg, delayed, offset, mlength,
- hdr->payload_length);
+ lnet_ni_recv(ni, msg->msg_private, msg, msg->msg_rx_delayed,
+ msg->msg_offset, msg->msg_wanted, hdr->payload_length);
}
static int
lnet_parse_put(lnet_ni_t *ni, lnet_msg_t *msg)
{
- int rc;
- int index;
- __u64 version;
- lnet_hdr_t *hdr = &msg->msg_hdr;
- unsigned int rlength = hdr->payload_length;
- unsigned int mlength = 0;
- unsigned int offset = 0;
- lnet_process_id_t src= {0};
- lnet_libmd_t *md;
- lnet_portal_t *ptl;
-
- src.nid = hdr->src_nid;
- src.pid = hdr->src_pid;
+ lnet_hdr_t *hdr = &msg->msg_hdr;
+ struct lnet_match_info info;
+ int rc;
- /* Convert put fields to host byte order */
- hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
- hdr->msg.put.ptl_index = le32_to_cpu(hdr->msg.put.ptl_index);
- hdr->msg.put.offset = le32_to_cpu(hdr->msg.put.offset);
+ /* Convert put fields to host byte order */
+ hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
+ hdr->msg.put.ptl_index = le32_to_cpu(hdr->msg.put.ptl_index);
+ hdr->msg.put.offset = le32_to_cpu(hdr->msg.put.offset);
- index = hdr->msg.put.ptl_index;
+ info.mi_id.nid = hdr->src_nid;
+ info.mi_id.pid = hdr->src_pid;
+ info.mi_opc = LNET_MD_OP_PUT;
+ info.mi_portal = hdr->msg.put.ptl_index;
+ info.mi_rlength = hdr->payload_length;
+ info.mi_roffset = hdr->msg.put.offset;
+ info.mi_mbits = hdr->msg.put.match_bits;
- LNET_LOCK();
+ msg->msg_rx_ready_delay = ni->ni_lnd->lnd_eager_recv == NULL;
again:
- rc = lnet_match_md(index, LNET_MD_OP_PUT, src,
- rlength, hdr->msg.put.offset,
- hdr->msg.put.match_bits, msg,
- &mlength, &offset, &md);
+ rc = lnet_ptl_match_md(&info, msg);
switch (rc) {
default:
LBUG();
case LNET_MATCHMD_OK:
- LNET_UNLOCK();
- lnet_recv_put(md, msg, msg->msg_delayed, offset, mlength);
+ lnet_recv_put(ni, msg);
return 0;
case LNET_MATCHMD_NONE:
- ptl = the_lnet.ln_portals[index];
- version = ptl->ptl_ml_version;
-
- rc = 0;
- if (!msg->msg_delayed)
- rc = lnet_eager_recv_locked(msg);
-
- if (rc == 0 &&
- !the_lnet.ln_shutdown &&
- lnet_ptl_is_lazy(ptl)) {
- if (version != ptl->ptl_ml_version)
- goto again;
-
- cfs_list_add_tail(&msg->msg_list, &ptl->ptl_msgq);
- ptl->ptl_msgq_version++;
- LNET_UNLOCK();
+ if (msg->msg_rx_delayed) /* attached on delayed list */
+ return 0;
- CDEBUG(D_NET, "Delaying PUT from %s portal %d match "
- LPU64" offset %d length %d: no match \n",
- libcfs_id2str(src), index,
- hdr->msg.put.match_bits,
- hdr->msg.put.offset, rlength);
- return 0;
- }
- /* fall through */
+ rc = lnet_ni_eager_recv(ni, msg);
+ if (rc == 0)
+ goto again;
+ /* fall through */
- case LNET_MATCHMD_DROP:
- CNETERR("Dropping PUT from %s portal %d match "LPU64
- " offset %d length %d: %d\n",
- libcfs_id2str(src), index,
- hdr->msg.put.match_bits,
- hdr->msg.put.offset, rlength, rc);
- LNET_UNLOCK();
+ case LNET_MATCHMD_DROP:
+ CNETERR("Dropping PUT from %s portal %d match "LPU64
+ " offset %d length %d: %d\n",
+ libcfs_id2str(info.mi_id), info.mi_portal,
+ info.mi_mbits, info.mi_roffset, info.mi_rlength, rc);
- return ENOENT; /* +ve: OK but no match */
- }
+ return ENOENT; /* +ve: OK but no match */
+ }
}
static int
lnet_parse_get(lnet_ni_t *ni, lnet_msg_t *msg, int rdma_get)
{
- lnet_hdr_t *hdr = &msg->msg_hdr;
- unsigned int mlength = 0;
- unsigned int offset = 0;
- lnet_process_id_t src = {0};
- lnet_handle_wire_t reply_wmd;
- lnet_libmd_t *md;
- int rc;
-
- src.nid = hdr->src_nid;
- src.pid = hdr->src_pid;
-
- /* Convert get fields to host byte order */
- hdr->msg.get.match_bits = le64_to_cpu(hdr->msg.get.match_bits);
- hdr->msg.get.ptl_index = le32_to_cpu(hdr->msg.get.ptl_index);
- hdr->msg.get.sink_length = le32_to_cpu(hdr->msg.get.sink_length);
- hdr->msg.get.src_offset = le32_to_cpu(hdr->msg.get.src_offset);
-
- LNET_LOCK();
-
- rc = lnet_match_md(hdr->msg.get.ptl_index, LNET_MD_OP_GET, src,
- hdr->msg.get.sink_length, hdr->msg.get.src_offset,
- hdr->msg.get.match_bits, msg,
- &mlength, &offset, &md);
- if (rc == LNET_MATCHMD_DROP) {
- CNETERR("Dropping GET from %s portal %d match "LPU64
- " offset %d length %d\n",
- libcfs_id2str(src),
- hdr->msg.get.ptl_index,
- hdr->msg.get.match_bits,
- hdr->msg.get.src_offset,
- hdr->msg.get.sink_length);
- LNET_UNLOCK();
- return ENOENT; /* +ve: OK but no match */
- }
-
- LASSERT (rc == LNET_MATCHMD_OK);
+ struct lnet_match_info info;
+ lnet_hdr_t *hdr = &msg->msg_hdr;
+ lnet_handle_wire_t reply_wmd;
+ int rc;
- the_lnet.ln_counters.send_count++;
- the_lnet.ln_counters.send_length += mlength;
+ /* Convert get fields to host byte order */
+ hdr->msg.get.match_bits = le64_to_cpu(hdr->msg.get.match_bits);
+ hdr->msg.get.ptl_index = le32_to_cpu(hdr->msg.get.ptl_index);
+ hdr->msg.get.sink_length = le32_to_cpu(hdr->msg.get.sink_length);
+ hdr->msg.get.src_offset = le32_to_cpu(hdr->msg.get.src_offset);
+
+ info.mi_id.nid = hdr->src_nid;
+ info.mi_id.pid = hdr->src_pid;
+ info.mi_opc = LNET_MD_OP_GET;
+ info.mi_portal = hdr->msg.get.ptl_index;
+ info.mi_rlength = hdr->msg.get.sink_length;
+ info.mi_roffset = hdr->msg.get.src_offset;
+ info.mi_mbits = hdr->msg.get.match_bits;
+
+ rc = lnet_ptl_match_md(&info, msg);
+ if (rc == LNET_MATCHMD_DROP) {
+ CNETERR("Dropping GET from %s portal %d match "LPU64
+ " offset %d length %d\n",
+ libcfs_id2str(info.mi_id), info.mi_portal,
+ info.mi_mbits, info.mi_roffset, info.mi_rlength);
+ return ENOENT; /* +ve: OK but no match */
+ }
- LNET_UNLOCK();
+ LASSERT(rc == LNET_MATCHMD_OK);
- msg->msg_ev.type = LNET_EVENT_GET;
- msg->msg_ev.target.pid = hdr->dest_pid;
- msg->msg_ev.target.nid = hdr->dest_nid;
- msg->msg_ev.hdr_data = 0;
+ lnet_build_msg_event(msg, LNET_EVENT_GET);
- reply_wmd = hdr->msg.get.return_wmd;
+ reply_wmd = hdr->msg.get.return_wmd;
- lnet_prep_send(msg, LNET_MSG_REPLY, src, offset, mlength);
+ lnet_prep_send(msg, LNET_MSG_REPLY, info.mi_id,
+ msg->msg_offset, msg->msg_wanted);
msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
msg->msg_receiving = 0;
- rc = lnet_send(ni->ni_nid, msg);
- if (rc < 0) {
- /* didn't get as far as lnet_ni_send() */
- CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
- libcfs_nid2str(ni->ni_nid), libcfs_id2str(src), rc);
+ rc = lnet_send(ni->ni_nid, msg, LNET_NID_ANY);
+ if (rc < 0) {
+ /* didn't get as far as lnet_ni_send() */
+ CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
+ libcfs_nid2str(ni->ni_nid),
+ libcfs_id2str(info.mi_id), rc);
- lnet_finalize(ni, msg, rc);
- }
+ lnet_finalize(ni, msg, rc);
+ }
- return 0;
+ return 0;
}
static int
lnet_libmd_t *md;
int rlength;
int mlength;
+ int cpt;
- LNET_LOCK();
+ cpt = lnet_cpt_of_cookie(hdr->msg.reply.dst_wmd.wh_object_cookie);
+ lnet_res_lock(cpt);
src.nid = hdr->src_nid;
src.pid = hdr->src_pid;
CERROR("REPLY MD also attached to portal %d\n",
md->md_me->me_portal);
- LNET_UNLOCK();
+ lnet_res_unlock(cpt);
return ENOENT; /* +ve: OK but no match */
}
libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
mlength);
- LNET_UNLOCK();
+ lnet_res_unlock(cpt);
return ENOENT; /* +ve: OK but no match */
}
libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
- lnet_commit_md(md, msg);
+ lnet_msg_attach_md(msg, md, 0, mlength);
- if (mlength != 0)
- lnet_setpayloadbuffer(msg);
+ if (mlength != 0)
+ lnet_setpayloadbuffer(msg);
- msg->msg_ev.type = LNET_EVENT_REPLY;
- msg->msg_ev.target.pid = hdr->dest_pid;
- msg->msg_ev.target.nid = hdr->dest_nid;
- msg->msg_ev.initiator = src;
- msg->msg_ev.rlength = rlength;
- msg->msg_ev.mlength = mlength;
- msg->msg_ev.offset = 0;
+ lnet_res_unlock(cpt);
- lnet_md_deconstruct(md, &msg->msg_ev.md);
- lnet_md2handle(&msg->msg_ev.md_handle, md);
+ lnet_build_msg_event(msg, LNET_EVENT_REPLY);
- the_lnet.ln_counters.recv_count++;
- the_lnet.ln_counters.recv_length += mlength;
-
- LNET_UNLOCK();
-
- lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
- return 0;
+ lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
+ return 0;
}
static int
lnet_hdr_t *hdr = &msg->msg_hdr;
lnet_process_id_t src = {0};
lnet_libmd_t *md;
+ int cpt;
src.nid = hdr->src_nid;
src.pid = hdr->src_pid;
hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
- LNET_LOCK();
+ cpt = lnet_cpt_of_cookie(hdr->msg.ack.dst_wmd.wh_object_cookie);
+ lnet_res_lock(cpt);
/* NB handles only looked up by creator (no flips) */
md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
CERROR("Source MD also attached to portal %d\n",
md->md_me->me_portal);
- LNET_UNLOCK();
+ lnet_res_unlock(cpt);
return ENOENT; /* +ve! */
}
libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
hdr->msg.ack.dst_wmd.wh_object_cookie);
- lnet_commit_md(md, msg);
+ lnet_msg_attach_md(msg, md, 0, 0);
- msg->msg_ev.type = LNET_EVENT_ACK;
- msg->msg_ev.target.pid = hdr->dest_pid;
- msg->msg_ev.target.nid = hdr->dest_nid;
- msg->msg_ev.initiator = src;
- msg->msg_ev.mlength = hdr->msg.ack.mlength;
- msg->msg_ev.match_bits = hdr->msg.ack.match_bits;
+ lnet_res_unlock(cpt);
- lnet_md_deconstruct(md, &msg->msg_ev.md);
- lnet_md2handle(&msg->msg_ev.md_handle, md);
+ lnet_build_msg_event(msg, LNET_EVENT_ACK);
- the_lnet.ln_counters.recv_count++;
+ lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
+ return 0;
+}
- LNET_UNLOCK();
+static int
+lnet_parse_forward_locked(lnet_ni_t *ni, lnet_msg_t *msg)
+{
+ int rc = 0;
- lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
- return 0;
+#ifdef __KERNEL__
+ if (msg->msg_rxpeer->lp_rtrcredits <= 0 ||
+ lnet_msg2bufpool(msg)->rbp_credits <= 0) {
+ if (ni->ni_lnd->lnd_eager_recv == NULL) {
+ msg->msg_rx_ready_delay = 1;
+ } else {
+ lnet_net_unlock(msg->msg_rx_cpt);
+ rc = lnet_ni_eager_recv(ni, msg);
+ lnet_net_lock(msg->msg_rx_cpt);
+ }
+ }
+
+ if (rc == 0)
+ rc = lnet_post_routed_recv_locked(msg, 0);
+#else
+ LBUG();
+#endif
+ return rc;
}
char *
return ("<UNKNOWN>");
}
}
+EXPORT_SYMBOL(lnet_msgtyp2str);
void
lnet_print_hdr(lnet_hdr_t * hdr)
}
int
-lnet_parse(lnet_ni_t *ni, lnet_hdr_t *hdr, lnet_nid_t from_nid,
- void *private, int rdma_req)
+lnet_parse(lnet_ni_t *ni, lnet_hdr_t *hdr, lnet_nid_t from_nid,
+ void *private, int rdma_req)
{
- int rc = 0;
- int for_me;
- lnet_msg_t *msg;
+ int rc = 0;
+ int cpt;
+ int for_me;
+ struct lnet_msg *msg;
lnet_pid_t dest_pid;
lnet_nid_t dest_nid;
lnet_nid_t src_nid;
payload_length = le32_to_cpu(hdr->payload_length);
for_me = (ni->ni_nid == dest_nid);
+ cpt = lnet_cpt_of_nid(from_nid);
switch (type) {
case LNET_MSG_ACK:
if (the_lnet.ln_routing &&
ni->ni_last_alive != cfs_time_current_sec()) {
- LNET_LOCK();
+ lnet_ni_lock(ni);
/* NB: so far here is the only place to set NI status to "up */
ni->ni_last_alive = cfs_time_current_sec();
if (ni->ni_status != NULL &&
ni->ni_status->ns_status == LNET_NI_STATUS_DOWN)
ni->ni_status->ns_status = LNET_NI_STATUS_UP;
- LNET_UNLOCK();
+ lnet_ni_unlock(ni);
}
/* Regard a bad destination NID as a protocol error. Senders should
msg->msg_len = msg->msg_wanted = payload_length;
msg->msg_offset = 0;
msg->msg_hdr = *hdr;
+ /* for building message event */
+ msg->msg_from = from_nid;
+ if (!for_me) {
+ msg->msg_target.pid = dest_pid;
+ msg->msg_target.nid = dest_nid;
+ msg->msg_routing = 1;
+
+ } else {
+ /* convert common msg->hdr fields to host byteorder */
+ msg->msg_hdr.type = type;
+ msg->msg_hdr.src_nid = src_nid;
+ msg->msg_hdr.src_pid = le32_to_cpu(msg->msg_hdr.src_pid);
+ msg->msg_hdr.dest_nid = dest_nid;
+ msg->msg_hdr.dest_pid = dest_pid;
+ msg->msg_hdr.payload_length = payload_length;
+ }
- LNET_LOCK();
- rc = lnet_nid2peer_locked(&msg->msg_rxpeer, from_nid);
- if (rc != 0) {
- LNET_UNLOCK();
- CERROR("%s, src %s: Dropping %s "
- "(error %d looking up sender)\n",
- libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
- lnet_msgtyp2str(type), rc);
- goto free_drop;
- }
- LNET_UNLOCK();
+ lnet_net_lock(cpt);
+ rc = lnet_nid2peer_locked(&msg->msg_rxpeer, from_nid, cpt);
+ if (rc != 0) {
+ lnet_net_unlock(cpt);
+ CERROR("%s, src %s: Dropping %s "
+ "(error %d looking up sender)\n",
+ libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
+ lnet_msgtyp2str(type), rc);
+ lnet_msg_free(msg);
+ goto drop;
+ }
-#ifndef __KERNEL__
- LASSERT (for_me);
-#else
- if (!for_me) {
- msg->msg_target.pid = dest_pid;
- msg->msg_target.nid = dest_nid;
- msg->msg_routing = 1;
- msg->msg_offset = 0;
-
- LNET_LOCK();
- if (msg->msg_rxpeer->lp_rtrcredits <= 0 ||
- lnet_msg2bufpool(msg)->rbp_credits <= 0) {
- rc = lnet_eager_recv_locked(msg);
- if (rc != 0) {
- LNET_UNLOCK();
- goto free_drop;
- }
- }
- lnet_commit_routedmsg(msg);
- rc = lnet_post_routed_recv_locked(msg, 0);
- LNET_UNLOCK();
+ lnet_msg_commit(msg, cpt);
- if (rc == 0)
- lnet_ni_recv(ni, msg->msg_private, msg, 0,
- 0, payload_length, payload_length);
- return 0;
- }
-#endif
- /* convert common msg->hdr fields to host byteorder */
- msg->msg_hdr.type = type;
- msg->msg_hdr.src_nid = src_nid;
- msg->msg_hdr.src_pid = le32_to_cpu(msg->msg_hdr.src_pid);
- msg->msg_hdr.dest_nid = dest_nid;
- msg->msg_hdr.dest_pid = dest_pid;
- msg->msg_hdr.payload_length = payload_length;
+ if (!for_me) {
+ rc = lnet_parse_forward_locked(ni, msg);
+ lnet_net_unlock(cpt);
- msg->msg_ev.sender = from_nid;
+ if (rc < 0)
+ goto free_drop;
+ if (rc == 0) {
+ lnet_ni_recv(ni, msg->msg_private, msg, 0,
+ 0, payload_length, payload_length);
+ }
+ return 0;
+ }
+
+ lnet_net_unlock(cpt);
switch (type) {
case LNET_MSG_ACK:
break;
default:
LASSERT(0);
+ rc = -EPROTO;
goto free_drop; /* prevent an unused label if !kernel */
}
LASSERT (rc == ENOENT);
free_drop:
- LASSERT (msg->msg_md == NULL);
- LNET_LOCK();
- if (msg->msg_rxpeer != NULL) {
- lnet_peer_decref_locked(msg->msg_rxpeer);
- msg->msg_rxpeer = NULL;
- }
- lnet_msg_free_locked(msg); /* expects LNET_LOCK held */
- LNET_UNLOCK();
+ LASSERT(msg->msg_md == NULL);
+ lnet_finalize(ni, msg, rc);
drop:
- lnet_drop_message(ni, private, payload_length);
- return 0;
+ lnet_drop_message(ni, cpt, private, payload_length);
+ return 0;
}
+EXPORT_SYMBOL(lnet_parse);
void
lnet_drop_delayed_msg_list(cfs_list_t *head, char *reason)
id.pid = msg->msg_hdr.src_pid;
LASSERT(msg->msg_md == NULL);
- LASSERT(msg->msg_delayed);
+ LASSERT(msg->msg_rx_delayed);
LASSERT(msg->msg_rxpeer != NULL);
LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
* until that's done */
lnet_drop_message(msg->msg_rxpeer->lp_ni,
+ msg->msg_rxpeer->lp_cpt,
msg->msg_private, msg->msg_len);
-
- LNET_LOCK();
- lnet_peer_decref_locked(msg->msg_rxpeer);
- LNET_UNLOCK();
-
- lnet_msg_free(msg);
+ /*
+ * NB: message will not generate event because w/o attached MD,
+ * but we still should give error code so lnet_msg_decommit()
+ * can skip counters operations and other checks.
+ */
+ lnet_finalize(msg->msg_rxpeer->lp_ni, msg, -ENOENT);
}
}
id.nid = msg->msg_hdr.src_nid;
id.pid = msg->msg_hdr.src_pid;
- LASSERT(msg->msg_delayed);
+ LASSERT(msg->msg_rx_delayed);
LASSERT(msg->msg_md != NULL);
LASSERT(msg->msg_rxpeer != NULL);
LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
msg->msg_hdr.msg.put.offset,
msg->msg_hdr.payload_length);
- lnet_recv_put(msg->msg_md, msg, 1,
- msg->msg_ev.offset, msg->msg_ev.mlength);
+ lnet_recv_put(msg->msg_rxpeer->lp_ni, msg);
}
}
__u64 match_bits, unsigned int offset,
__u64 hdr_data)
{
- lnet_msg_t *msg;
- lnet_libmd_t *md;
- int rc;
+ struct lnet_msg *msg;
+ struct lnet_libmd *md;
+ int cpt;
+ int rc;
LASSERT (the_lnet.ln_init);
LASSERT (the_lnet.ln_refcount > 0);
}
msg->msg_vmflush = !!cfs_memory_pressure_get();
- LNET_LOCK();
+ cpt = lnet_cpt_of_cookie(mdh.cookie);
+ lnet_res_lock(cpt);
md = lnet_handle2md(&mdh);
if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
- lnet_msg_free_locked(msg);
-
CERROR("Dropping PUT ("LPU64":%d:%s): MD (%d) invalid\n",
match_bits, portal, libcfs_id2str(target),
md == NULL ? -1 : md->md_threshold);
if (md != NULL && md->md_me != NULL)
CERROR("Source MD also attached to portal %d\n",
md->md_me->me_portal);
+ lnet_res_unlock(cpt);
- LNET_UNLOCK();
- return -ENOENT;
- }
+ lnet_msg_free(msg);
+ return -ENOENT;
+ }
- CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target));
+ CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target));
- lnet_commit_md(md, msg);
+ lnet_msg_attach_md(msg, md, 0, 0);
lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
LNET_WIRE_HANDLE_COOKIE_NONE;
}
- msg->msg_ev.type = LNET_EVENT_SEND;
- msg->msg_ev.initiator.nid = LNET_NID_ANY;
- msg->msg_ev.initiator.pid = the_lnet.ln_pid;
- msg->msg_ev.target = target;
- msg->msg_ev.sender = LNET_NID_ANY;
- msg->msg_ev.pt_index = portal;
- msg->msg_ev.match_bits = match_bits;
- msg->msg_ev.rlength = md->md_length;
- msg->msg_ev.mlength = md->md_length;
- msg->msg_ev.offset = offset;
- msg->msg_ev.hdr_data = hdr_data;
-
- lnet_md_deconstruct(md, &msg->msg_ev.md);
- lnet_md2handle(&msg->msg_ev.md_handle, md);
+ lnet_res_unlock(cpt);
- the_lnet.ln_counters.send_count++;
- the_lnet.ln_counters.send_length += md->md_length;
+ lnet_build_msg_event(msg, LNET_EVENT_SEND);
- LNET_UNLOCK();
-
- rc = lnet_send(self, msg);
+ rc = lnet_send(self, msg, LNET_NID_ANY);
if (rc != 0) {
CNETERR( "Error sending PUT to %s: %d\n",
libcfs_id2str(target), rc);
/* completion will be signalled by an event */
return 0;
}
+EXPORT_SYMBOL(LNetPut);
lnet_msg_t *
lnet_create_reply_msg (lnet_ni_t *ni, lnet_msg_t *getmsg)
* CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
* lnet_finalize() is called on it, so the LND must call this first */
- lnet_msg_t *msg = lnet_msg_alloc();
- lnet_libmd_t *getmd = getmsg->msg_md;
- lnet_process_id_t peer_id = getmsg->msg_target;
+ struct lnet_msg *msg = lnet_msg_alloc();
+ struct lnet_libmd *getmd = getmsg->msg_md;
+ lnet_process_id_t peer_id = getmsg->msg_target;
+ int cpt;
- LASSERT (!getmsg->msg_target_is_router);
- LASSERT (!getmsg->msg_routing);
+ LASSERT(!getmsg->msg_target_is_router);
+ LASSERT(!getmsg->msg_routing);
- LNET_LOCK();
+ cpt = lnet_cpt_of_cookie(getmd->md_lh.lh_cookie);
+ lnet_res_lock(cpt);
LASSERT (getmd->md_refcount > 0);
CERROR ("%s: Dropping REPLY from %s for inactive MD %p\n",
libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id),
getmd);
- goto drop_msg;
- }
-
- LASSERT (getmd->md_offset == 0);
+ lnet_res_unlock(cpt);
+ goto drop;
+ }
- CDEBUG(D_NET, "%s: Reply from %s md %p\n",
- libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd);
+ LASSERT(getmd->md_offset == 0);
- lnet_commit_md (getmd, msg);
+ CDEBUG(D_NET, "%s: Reply from %s md %p\n",
+ libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd);
- msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
+ /* setup information for lnet_build_msg_event */
+ msg->msg_from = peer_id.nid;
+ msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
+ msg->msg_hdr.src_nid = peer_id.nid;
+ msg->msg_hdr.payload_length = getmd->md_length;
+ msg->msg_receiving = 1; /* required by lnet_msg_attach_md */
- msg->msg_ev.type = LNET_EVENT_REPLY;
- msg->msg_ev.initiator = peer_id;
- msg->msg_ev.sender = peer_id.nid; /* optimized GETs can't be routed */
- msg->msg_ev.rlength = msg->msg_ev.mlength = getmd->md_length;
- msg->msg_ev.offset = 0;
+ lnet_msg_attach_md(msg, getmd, getmd->md_offset, getmd->md_length);
+ lnet_res_unlock(cpt);
- lnet_md_deconstruct(getmd, &msg->msg_ev.md);
- lnet_md2handle(&msg->msg_ev.md_handle, getmd);
+ cpt = lnet_cpt_of_nid(peer_id.nid);
- the_lnet.ln_counters.recv_count++;
- the_lnet.ln_counters.recv_length += getmd->md_length;
+ lnet_net_lock(cpt);
+ lnet_msg_commit(msg, cpt);
+ lnet_net_unlock(cpt);
- LNET_UNLOCK();
+ lnet_build_msg_event(msg, LNET_EVENT_REPLY);
- return msg;
+ return msg;
- drop_msg:
- lnet_msg_free_locked(msg);
drop:
- the_lnet.ln_counters.drop_count++;
- the_lnet.ln_counters.drop_length += getmd->md_length;
+ cpt = lnet_cpt_of_nid(peer_id.nid);
+
+ lnet_net_lock(cpt);
+ the_lnet.ln_counters[cpt]->drop_count++;
+ the_lnet.ln_counters[cpt]->drop_length += getmd->md_length;
+ lnet_net_unlock(cpt);
- LNET_UNLOCK ();
+ if (msg != NULL)
+ lnet_msg_free(msg);
- return NULL;
+ return NULL;
}
+EXPORT_SYMBOL(lnet_create_reply_msg);
void
lnet_set_reply_msg_len(lnet_ni_t *ni, lnet_msg_t *reply, unsigned int len)
reply->msg_ev.mlength = len;
}
+EXPORT_SYMBOL(lnet_set_reply_msg_len);
/**
* Initiate an asynchronous GET operation.
* \retval -ENOENT Invalid MD object.
*/
int
-LNetGet(lnet_nid_t self, lnet_handle_md_t mdh,
- lnet_process_id_t target, unsigned int portal,
- __u64 match_bits, unsigned int offset)
+LNetGet(lnet_nid_t self, lnet_handle_md_t mdh,
+ lnet_process_id_t target, unsigned int portal,
+ __u64 match_bits, unsigned int offset)
{
- lnet_msg_t *msg;
- lnet_libmd_t *md;
- int rc;
+ struct lnet_msg *msg;
+ struct lnet_libmd *md;
+ int cpt;
+ int rc;
LASSERT (the_lnet.ln_init);
LASSERT (the_lnet.ln_refcount > 0);
return -ENOMEM;
}
- LNET_LOCK();
+ cpt = lnet_cpt_of_cookie(mdh.cookie);
+ lnet_res_lock(cpt);
md = lnet_handle2md(&mdh);
if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
- lnet_msg_free_locked(msg);
-
CERROR("Dropping GET ("LPU64":%d:%s): MD (%d) invalid\n",
match_bits, portal, libcfs_id2str(target),
md == NULL ? -1 : md->md_threshold);
CERROR("REPLY MD also attached to portal %d\n",
md->md_me->me_portal);
- LNET_UNLOCK();
+ lnet_res_unlock(cpt);
+
+ lnet_msg_free(msg);
+
return -ENOENT;
}
CDEBUG(D_NET, "LNetGet -> %s\n", libcfs_id2str(target));
- lnet_commit_md(md, msg);
+ lnet_msg_attach_md(msg, md, 0, 0);
lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
/* NB handles only looked up by creator (no flips) */
- msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
- the_lnet.ln_interface_cookie;
- msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
- md->md_lh.lh_cookie;
-
- msg->msg_ev.type = LNET_EVENT_SEND;
- msg->msg_ev.initiator.nid = LNET_NID_ANY;
- msg->msg_ev.initiator.pid = the_lnet.ln_pid;
- msg->msg_ev.target = target;
- msg->msg_ev.sender = LNET_NID_ANY;
- msg->msg_ev.pt_index = portal;
- msg->msg_ev.match_bits = match_bits;
- msg->msg_ev.rlength = md->md_length;
- msg->msg_ev.mlength = md->md_length;
- msg->msg_ev.offset = offset;
- msg->msg_ev.hdr_data = 0;
-
- lnet_md_deconstruct(md, &msg->msg_ev.md);
- lnet_md2handle(&msg->msg_ev.md_handle, md);
-
- the_lnet.ln_counters.send_count++;
-
- LNET_UNLOCK();
-
- rc = lnet_send(self, msg);
+ msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
+ the_lnet.ln_interface_cookie;
+ msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
+ md->md_lh.lh_cookie;
+
+ lnet_res_unlock(cpt);
+
+ lnet_build_msg_event(msg, LNET_EVENT_SEND);
+
+ rc = lnet_send(self, msg, LNET_NID_ANY);
if (rc < 0) {
CNETERR( "Error sending GET to %s: %d\n",
libcfs_id2str(target), rc);
/* completion will be signalled by an event */
return 0;
}
+EXPORT_SYMBOL(LNetGet);
/**
* Calculate distance to node at \a dstnid.
* \retval -EHOSTUNREACH If \a dstnid is not reachable.
*/
int
-LNetDist (lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
+LNetDist(lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
{
- cfs_list_t *e;
- lnet_ni_t *ni;
- lnet_remotenet_t *rnet;
- __u32 dstnet = LNET_NIDNET(dstnid);
- int hops;
- __u32 order = 2;
+ cfs_list_t *e;
+ struct lnet_ni *ni;
+ lnet_remotenet_t *rnet;
+ __u32 dstnet = LNET_NIDNET(dstnid);
+ int hops;
+ int cpt;
+ __u32 order = 2;
+ cfs_list_t *rn_list;
/* if !local_nid_dist_zero, I don't return a distance of 0 ever
* (when lustre sees a distance of 0, it substitutes 0@lo), so I
LASSERT (the_lnet.ln_init);
LASSERT (the_lnet.ln_refcount > 0);
- LNET_LOCK();
+ cpt = lnet_net_lock_current();
cfs_list_for_each (e, &the_lnet.ln_nis) {
ni = cfs_list_entry(e, lnet_ni_t, ni_list);
else
*orderp = 1;
}
- LNET_UNLOCK();
+ lnet_net_unlock(cpt);
return local_nid_dist_zero ? 0 : 1;
}
*srcnidp = ni->ni_nid;
if (orderp != NULL)
*orderp = order;
- LNET_UNLOCK();
- return 1;
- }
+ lnet_net_unlock(cpt);
+ return 1;
+ }
- order++;
- }
+ order++;
+ }
- cfs_list_for_each (e, &the_lnet.ln_remote_nets) {
+ rn_list = lnet_net2rnethash(dstnet);
+ cfs_list_for_each(e, rn_list) {
rnet = cfs_list_entry(e, lnet_remotenet_t, lrn_list);
if (rnet->lrn_net == dstnet) {
*srcnidp = shortest->lr_gateway->lp_ni->ni_nid;
if (orderp != NULL)
*orderp = order;
- LNET_UNLOCK();
- return hops + 1;
- }
- order++;
- }
+ lnet_net_unlock(cpt);
+ return hops + 1;
+ }
+ order++;
+ }
- LNET_UNLOCK();
- return -EHOSTUNREACH;
+ lnet_net_unlock(cpt);
+ return -EHOSTUNREACH;
}
+EXPORT_SYMBOL(LNetDist);
/**
* Set the number of asynchronous messages expected from a target process.
int maxnids = 256;
int rc = 0;
int rc2;
+ int cpt;
/* Target on a local network? */
ni = lnet_net2ni(LNET_NIDNET(id.nid));
nnids = 0;
/* Snapshot all the router NIDs */
- LNET_LOCK();
- rnet = lnet_find_net_locked(LNET_NIDNET(id.nid));
- if (rnet != NULL) {
- cfs_list_for_each(tmp, &rnet->lrn_routes) {
- if (nnids == maxnids) {
- LNET_UNLOCK();
+ cpt = lnet_net_lock_current();
+ rnet = lnet_find_net_locked(LNET_NIDNET(id.nid));
+ if (rnet != NULL) {
+ cfs_list_for_each(tmp, &rnet->lrn_routes) {
+ if (nnids == maxnids) {
+ lnet_net_unlock(cpt);
LIBCFS_FREE(nids, maxnids * sizeof(*nids));
maxnids *= 2;
goto again;
nids[nnids++] = route->lr_gateway->lp_nid;
}
}
- LNET_UNLOCK();
+ lnet_net_unlock(cpt);
/* set async on all the routers */
while (nnids-- > 0) {
return rc;
#endif
}
+EXPORT_SYMBOL(LNetSetAsync);