1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lnet/klnds/o2iblnd/o2iblnd.c
38 * Author: Eric Barton <eric@bartonsoftware.com>
45 .lnd_startup = kiblnd_startup,
46 .lnd_shutdown = kiblnd_shutdown,
47 .lnd_ctl = kiblnd_ctl,
48 .lnd_query = kiblnd_query,
49 .lnd_send = kiblnd_send,
50 .lnd_recv = kiblnd_recv,
53 kib_data_t kiblnd_data;
56 kiblnd_cksum (void *ptr, int nob)
62 sum = ((sum << 1) | (sum >> 31)) + *c++;
64 /* ensure I don't return 0 (== no checksum) */
65 return (sum == 0) ? 1 : sum;
69 kiblnd_msgtype2str(int type)
72 case IBLND_MSG_CONNREQ:
75 case IBLND_MSG_CONNACK:
81 case IBLND_MSG_IMMEDIATE:
84 case IBLND_MSG_PUT_REQ:
87 case IBLND_MSG_PUT_NAK:
90 case IBLND_MSG_PUT_ACK:
93 case IBLND_MSG_PUT_DONE:
96 case IBLND_MSG_GET_REQ:
99 case IBLND_MSG_GET_DONE:
108 kiblnd_msgtype2size(int type)
110 const int hdr_size = offsetof(kib_msg_t, ibm_u);
113 case IBLND_MSG_CONNREQ:
114 case IBLND_MSG_CONNACK:
115 return hdr_size + sizeof(kib_connparams_t);
120 case IBLND_MSG_IMMEDIATE:
121 return offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[0]);
123 case IBLND_MSG_PUT_REQ:
124 return hdr_size + sizeof(kib_putreq_msg_t);
126 case IBLND_MSG_PUT_ACK:
127 return hdr_size + sizeof(kib_putack_msg_t);
129 case IBLND_MSG_GET_REQ:
130 return hdr_size + sizeof(kib_get_msg_t);
132 case IBLND_MSG_PUT_NAK:
133 case IBLND_MSG_PUT_DONE:
134 case IBLND_MSG_GET_DONE:
135 return hdr_size + sizeof(kib_completion_msg_t);
142 kiblnd_unpack_rd(kib_msg_t *msg, int flip)
149 LASSERT (msg->ibm_type == IBLND_MSG_GET_REQ ||
150 msg->ibm_type == IBLND_MSG_PUT_ACK);
152 rd = msg->ibm_type == IBLND_MSG_GET_REQ ?
153 &msg->ibm_u.get.ibgm_rd :
154 &msg->ibm_u.putack.ibpam_rd;
157 __swab32s(&rd->rd_key);
158 __swab32s(&rd->rd_nfrags);
163 if (n <= 0 || n > IBLND_MAX_RDMA_FRAGS) {
164 CERROR("Bad nfrags: %d, should be 0 < n <= %d\n",
165 n, IBLND_MAX_RDMA_FRAGS);
169 nob = offsetof (kib_msg_t, ibm_u) +
170 kiblnd_rd_msg_size(rd, msg->ibm_type, n);
172 if (msg->ibm_nob < nob) {
173 CERROR("Short %s: %d(%d)\n",
174 kiblnd_msgtype2str(msg->ibm_type), msg->ibm_nob, nob);
181 for (i = 0; i < n; i++) {
182 __swab32s(&rd->rd_frags[i].rf_nob);
183 __swab64s(&rd->rd_frags[i].rf_addr);
190 kiblnd_pack_msg (lnet_ni_t *ni, kib_msg_t *msg, int version,
191 int credits, lnet_nid_t dstnid, __u64 dststamp)
193 kib_net_t *net = ni->ni_data;
195 /* CAVEAT EMPTOR! all message fields not set here should have been
196 * initialised previously. */
197 msg->ibm_magic = IBLND_MSG_MAGIC;
198 msg->ibm_version = version;
200 msg->ibm_credits = credits;
203 msg->ibm_srcnid = lnet_ptlcompat_srcnid(ni->ni_nid, dstnid);
204 msg->ibm_srcstamp = net->ibn_incarnation;
205 msg->ibm_dstnid = dstnid;
206 msg->ibm_dststamp = dststamp;
208 if (*kiblnd_tunables.kib_cksum) {
209 /* NB ibm_cksum zero while computing cksum */
210 msg->ibm_cksum = kiblnd_cksum(msg, msg->ibm_nob);
215 kiblnd_unpack_msg(kib_msg_t *msg, int nob)
217 const int hdr_size = offsetof(kib_msg_t, ibm_u);
223 /* 6 bytes are enough to have received magic + version */
225 CERROR("Short message: %d\n", nob);
229 if (msg->ibm_magic == IBLND_MSG_MAGIC) {
231 } else if (msg->ibm_magic == __swab32(IBLND_MSG_MAGIC)) {
234 CERROR("Bad magic: %08x\n", msg->ibm_magic);
238 version = flip ? __swab16(msg->ibm_version) : msg->ibm_version;
239 if (version != IBLND_MSG_VERSION &&
240 version != IBLND_MSG_VERSION_1) {
241 CERROR("Bad version: %x\n", version);
245 if (nob < hdr_size) {
246 CERROR("Short message: %d\n", nob);
250 msg_nob = flip ? __swab32(msg->ibm_nob) : msg->ibm_nob;
252 CERROR("Short message: got %d, wanted %d\n", nob, msg_nob);
256 /* checksum must be computed with ibm_cksum zero and BEFORE anything
258 msg_cksum = flip ? __swab32(msg->ibm_cksum) : msg->ibm_cksum;
260 if (msg_cksum != 0 &&
261 msg_cksum != kiblnd_cksum(msg, msg_nob)) {
262 CERROR("Bad checksum\n");
266 msg->ibm_cksum = msg_cksum;
269 /* leave magic unflipped as a clue to peer endianness */
270 msg->ibm_version = version;
271 CLASSERT (sizeof(msg->ibm_type) == 1);
272 CLASSERT (sizeof(msg->ibm_credits) == 1);
273 msg->ibm_nob = msg_nob;
274 __swab64s(&msg->ibm_srcnid);
275 __swab64s(&msg->ibm_srcstamp);
276 __swab64s(&msg->ibm_dstnid);
277 __swab64s(&msg->ibm_dststamp);
280 if (msg->ibm_srcnid == LNET_NID_ANY) {
281 CERROR("Bad src nid: %s\n", libcfs_nid2str(msg->ibm_srcnid));
285 if (msg_nob < kiblnd_msgtype2size(msg->ibm_type)) {
286 CERROR("Short %s: %d(%d)\n", kiblnd_msgtype2str(msg->ibm_type),
287 msg_nob, kiblnd_msgtype2size(msg->ibm_type));
291 switch (msg->ibm_type) {
293 CERROR("Unknown message type %x\n", msg->ibm_type);
297 case IBLND_MSG_IMMEDIATE:
298 case IBLND_MSG_PUT_REQ:
301 case IBLND_MSG_PUT_ACK:
302 case IBLND_MSG_GET_REQ:
303 if (kiblnd_unpack_rd(msg, flip))
307 case IBLND_MSG_PUT_NAK:
308 case IBLND_MSG_PUT_DONE:
309 case IBLND_MSG_GET_DONE:
311 __swab32s(&msg->ibm_u.completion.ibcm_status);
314 case IBLND_MSG_CONNREQ:
315 case IBLND_MSG_CONNACK:
317 __swab16s(&msg->ibm_u.connparams.ibcp_queue_depth);
318 __swab16s(&msg->ibm_u.connparams.ibcp_max_frags);
319 __swab32s(&msg->ibm_u.connparams.ibcp_max_msg_size);
327 kiblnd_create_peer (lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid)
330 kib_net_t *net = ni->ni_data;
333 LASSERT (net != NULL);
334 LASSERT (nid != LNET_NID_ANY);
336 LIBCFS_ALLOC(peer, sizeof(*peer));
338 CERROR("Cannot allocate peer\n");
342 memset(peer, 0, sizeof(*peer)); /* zero flags etc */
347 peer->ibp_last_alive = 0;
348 atomic_set(&peer->ibp_refcount, 1); /* 1 ref for caller */
350 INIT_LIST_HEAD(&peer->ibp_list); /* not in the peer table yet */
351 INIT_LIST_HEAD(&peer->ibp_conns);
352 INIT_LIST_HEAD(&peer->ibp_tx_queue);
354 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
356 /* always called with a ref on ni, which prevents ni being shutdown */
357 LASSERT (net->ibn_shutdown == 0);
359 /* npeers only grows with the global lock held */
360 atomic_inc(&net->ibn_npeers);
362 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
369 kiblnd_destroy_peer (kib_peer_t *peer)
371 kib_net_t *net = peer->ibp_ni->ni_data;
373 LASSERT (net != NULL);
374 LASSERT (atomic_read(&peer->ibp_refcount) == 0);
375 LASSERT (!kiblnd_peer_active(peer));
376 LASSERT (peer->ibp_connecting == 0);
377 LASSERT (peer->ibp_accepting == 0);
378 LASSERT (list_empty(&peer->ibp_conns));
379 LASSERT (list_empty(&peer->ibp_tx_queue));
381 LIBCFS_FREE(peer, sizeof(*peer));
383 /* NB a peer's connections keep a reference on their peer until
384 * they are destroyed, so we can be assured that _all_ state to do
385 * with this peer has been cleaned up when its refcount drops to
387 atomic_dec(&net->ibn_npeers);
391 kiblnd_find_peer_locked (lnet_nid_t nid)
393 /* the caller is responsible for accounting the additional reference
394 * that this creates */
395 struct list_head *peer_list = kiblnd_nid2peerlist(nid);
396 struct list_head *tmp;
399 list_for_each (tmp, peer_list) {
401 peer = list_entry(tmp, kib_peer_t, ibp_list);
403 LASSERT (peer->ibp_connecting > 0 || /* creating conns */
404 peer->ibp_accepting > 0 ||
405 !list_empty(&peer->ibp_conns)); /* active conn */
407 if (peer->ibp_nid != nid)
410 CDEBUG(D_NET, "got peer [%p] -> %s (%d) version: %x\n",
411 peer, libcfs_nid2str(nid),
412 atomic_read(&peer->ibp_refcount),
420 kiblnd_unlink_peer_locked (kib_peer_t *peer)
422 LASSERT (list_empty(&peer->ibp_conns));
424 LASSERT (kiblnd_peer_active(peer));
425 list_del_init(&peer->ibp_list);
426 /* lose peerlist's ref */
427 kiblnd_peer_decref(peer);
431 kiblnd_get_peer_info (lnet_ni_t *ni, int index,
432 lnet_nid_t *nidp, int *count)
435 struct list_head *ptmp;
439 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
441 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
443 list_for_each (ptmp, &kiblnd_data.kib_peers[i]) {
445 peer = list_entry(ptmp, kib_peer_t, ibp_list);
446 LASSERT (peer->ibp_connecting > 0 ||
447 peer->ibp_accepting > 0 ||
448 !list_empty(&peer->ibp_conns));
450 if (peer->ibp_ni != ni)
456 *nidp = peer->ibp_nid;
457 *count = atomic_read(&peer->ibp_refcount);
459 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
465 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
470 kiblnd_del_peer_locked (kib_peer_t *peer)
472 struct list_head *ctmp;
473 struct list_head *cnxt;
476 if (list_empty(&peer->ibp_conns)) {
477 kiblnd_unlink_peer_locked(peer);
479 list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
480 conn = list_entry(ctmp, kib_conn_t, ibc_list);
482 kiblnd_close_conn_locked(conn, 0);
484 /* NB closing peer's last conn unlinked it. */
486 /* NB peer now unlinked; might even be freed if the peer table had the
491 kiblnd_del_peer (lnet_ni_t *ni, lnet_nid_t nid)
493 CFS_LIST_HEAD (zombies);
494 struct list_head *ptmp;
495 struct list_head *pnxt;
503 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
505 if (nid != LNET_NID_ANY) {
506 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
509 hi = kiblnd_data.kib_peer_hash_size - 1;
512 for (i = lo; i <= hi; i++) {
513 list_for_each_safe (ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
514 peer = list_entry(ptmp, kib_peer_t, ibp_list);
515 LASSERT (peer->ibp_connecting > 0 ||
516 peer->ibp_accepting > 0 ||
517 !list_empty(&peer->ibp_conns));
519 if (peer->ibp_ni != ni)
522 if (!(nid == LNET_NID_ANY || peer->ibp_nid == nid))
525 if (!list_empty(&peer->ibp_tx_queue)) {
526 LASSERT (list_empty(&peer->ibp_conns));
528 list_splice_init(&peer->ibp_tx_queue, &zombies);
531 kiblnd_del_peer_locked(peer);
532 rc = 0; /* matched something */
536 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
538 kiblnd_txlist_done(ni, &zombies, -EIO);
544 kiblnd_get_conn_by_idx (lnet_ni_t *ni, int index)
547 struct list_head *ptmp;
549 struct list_head *ctmp;
553 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
555 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
556 list_for_each (ptmp, &kiblnd_data.kib_peers[i]) {
558 peer = list_entry(ptmp, kib_peer_t, ibp_list);
559 LASSERT (peer->ibp_connecting > 0 ||
560 peer->ibp_accepting > 0 ||
561 !list_empty(&peer->ibp_conns));
563 if (peer->ibp_ni != ni)
566 list_for_each (ctmp, &peer->ibp_conns) {
570 conn = list_entry(ctmp, kib_conn_t, ibc_list);
571 kiblnd_conn_addref(conn);
572 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
579 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
584 kiblnd_debug_rx (kib_rx_t *rx)
586 CDEBUG(D_CONSOLE, " %p status %d msg_type %x cred %d\n",
587 rx, rx->rx_status, rx->rx_msg->ibm_type,
588 rx->rx_msg->ibm_credits);
592 kiblnd_debug_tx (kib_tx_t *tx)
594 CDEBUG(D_CONSOLE, " %p snd %d q %d w %d rc %d dl %lx "
595 "cookie "LPX64" msg %s%s type %x cred %d\n",
596 tx, tx->tx_sending, tx->tx_queued, tx->tx_waiting,
597 tx->tx_status, tx->tx_deadline, tx->tx_cookie,
598 tx->tx_lntmsg[0] == NULL ? "-" : "!",
599 tx->tx_lntmsg[1] == NULL ? "-" : "!",
600 tx->tx_msg->ibm_type, tx->tx_msg->ibm_credits);
604 kiblnd_debug_conn (kib_conn_t *conn)
606 struct list_head *tmp;
609 spin_lock(&conn->ibc_lock);
611 CDEBUG(D_CONSOLE, "conn[%d] %p [version %x] -> %s: \n",
612 atomic_read(&conn->ibc_refcount), conn,
613 conn->ibc_version, libcfs_nid2str(conn->ibc_peer->ibp_nid));
614 CDEBUG(D_CONSOLE, " state %d nposted %d/%d cred %d o_cred %d r_cred %d\n",
615 conn->ibc_state, conn->ibc_noops_posted,
616 conn->ibc_nsends_posted, conn->ibc_credits,
617 conn->ibc_outstanding_credits, conn->ibc_reserved_credits);
618 CDEBUG(D_CONSOLE, " comms_err %d\n", conn->ibc_comms_error);
620 CDEBUG(D_CONSOLE, " early_rxs:\n");
621 list_for_each(tmp, &conn->ibc_early_rxs)
622 kiblnd_debug_rx(list_entry(tmp, kib_rx_t, rx_list));
624 CDEBUG(D_CONSOLE, " tx_queue_nocred:\n");
625 list_for_each(tmp, &conn->ibc_tx_queue_nocred)
626 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
628 CDEBUG(D_CONSOLE, " tx_queue_rsrvd:\n");
629 list_for_each(tmp, &conn->ibc_tx_queue_rsrvd)
630 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
632 CDEBUG(D_CONSOLE, " tx_queue:\n");
633 list_for_each(tmp, &conn->ibc_tx_queue)
634 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
636 CDEBUG(D_CONSOLE, " active_txs:\n");
637 list_for_each(tmp, &conn->ibc_active_txs)
638 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
640 CDEBUG(D_CONSOLE, " rxs:\n");
641 for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++)
642 kiblnd_debug_rx(&conn->ibc_rxs[i]);
644 spin_unlock(&conn->ibc_lock);
648 kiblnd_translate_mtu(int value)
669 kiblnd_setup_mtu(struct rdma_cm_id *cmid)
674 /* XXX There is no path record for iWARP, set by netdev->change_mtu? */
675 if (cmid->route.path_rec == NULL)
678 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
680 mtu = kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu);
683 cmid->route.path_rec->mtu = mtu;
685 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
689 kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
690 int state, int version)
693 * If the new conn is created successfully it takes over the caller's
694 * ref on 'peer'. It also "owns" 'cmid' and destroys it when it itself
695 * is destroyed. On failure, the caller's ref on 'peer' remains and
696 * she must dispose of 'cmid'. (Actually I'd block forever if I tried
697 * to destroy 'cmid' here since I'm called from the CM which still has
698 * its ref on 'cmid'). */
699 kib_net_t *net = peer->ibp_ni->ni_data;
700 struct ib_qp_init_attr *init_qp_attr;
707 LASSERT (net != NULL);
708 LASSERT (!in_interrupt());
710 LIBCFS_ALLOC(init_qp_attr, sizeof(*init_qp_attr));
711 if (init_qp_attr == NULL) {
712 CERROR("Can't allocate qp_attr for %s\n",
713 libcfs_nid2str(peer->ibp_nid));
717 LIBCFS_ALLOC(conn, sizeof(*conn));
719 CERROR("Can't allocate connection for %s\n",
720 libcfs_nid2str(peer->ibp_nid));
724 memset(conn, 0, sizeof(*conn)); /* zero flags, NULL pointers etc... */
726 conn->ibc_state = IBLND_CONN_INIT;
727 conn->ibc_version = version;
728 conn->ibc_peer = peer; /* I take the caller's ref */
729 cmid->context = conn; /* for future CM callbacks */
730 conn->ibc_cmid = cmid;
732 INIT_LIST_HEAD(&conn->ibc_early_rxs);
733 INIT_LIST_HEAD(&conn->ibc_tx_queue);
734 INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
735 INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
736 INIT_LIST_HEAD(&conn->ibc_active_txs);
737 spin_lock_init(&conn->ibc_lock);
739 LIBCFS_ALLOC(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
740 if (conn->ibc_connvars == NULL) {
741 CERROR("Can't allocate in-progress connection state\n");
744 memset(conn->ibc_connvars, 0, sizeof(*conn->ibc_connvars));
746 LIBCFS_ALLOC(conn->ibc_rxs, IBLND_RX_MSGS(version) * sizeof(kib_rx_t));
747 if (conn->ibc_rxs == NULL) {
748 CERROR("Cannot allocate RX buffers\n");
751 memset(conn->ibc_rxs, 0, IBLND_RX_MSGS(version) * sizeof(kib_rx_t));
753 rc = kiblnd_alloc_pages(&conn->ibc_rx_pages,
754 IBLND_RX_MSG_PAGES(version));
758 kiblnd_map_rx_descs(conn);
760 #ifdef HAVE_OFED_IB_COMP_VECTOR
761 cq = ib_create_cq(cmid->device,
762 kiblnd_cq_completion, kiblnd_cq_event, conn,
763 IBLND_CQ_ENTRIES(version), 0);
765 cq = ib_create_cq(cmid->device,
766 kiblnd_cq_completion, kiblnd_cq_event, conn,
767 IBLND_CQ_ENTRIES(version));
770 CERROR("Can't create CQ: %ld, cqe: %d\n",
771 PTR_ERR(cq), IBLND_CQ_ENTRIES(version));
777 rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
779 CERROR("Can't request completion notificiation: %d\n", rc);
783 kiblnd_setup_mtu(cmid);
785 memset(init_qp_attr, 0, sizeof(*init_qp_attr));
786 init_qp_attr->event_handler = kiblnd_qp_event;
787 init_qp_attr->qp_context = conn;
788 init_qp_attr->cap.max_send_wr = IBLND_SEND_WRS(version);
789 init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(version);
790 init_qp_attr->cap.max_send_sge = 1;
791 init_qp_attr->cap.max_recv_sge = 1;
792 init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
793 init_qp_attr->qp_type = IB_QPT_RC;
794 init_qp_attr->send_cq = cq;
795 init_qp_attr->recv_cq = cq;
797 rc = rdma_create_qp(cmid, net->ibn_dev->ibd_pd, init_qp_attr);
799 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d\n",
800 rc, init_qp_attr->cap.max_send_wr,
801 init_qp_attr->cap.max_recv_wr);
805 LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
807 /* 1 ref for caller and each rxmsg */
808 atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(version));
809 conn->ibc_nrx = IBLND_RX_MSGS(version);
812 for (i = 0; i < IBLND_RX_MSGS(version); i++) {
813 rc = kiblnd_post_rx(&conn->ibc_rxs[i],
814 IBLND_POSTRX_NO_CREDIT);
816 CERROR("Can't post rxmsg: %d\n", rc);
818 /* Make posted receives complete */
819 kiblnd_abort_receives(conn);
821 /* correct # of posted buffers
822 * NB locking needed now I'm racing with completion */
823 spin_lock_irqsave(&kiblnd_data.kib_sched_lock, flags);
824 conn->ibc_nrx -= IBLND_RX_MSGS(version) - i;
825 spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock,
828 /* cmid will be destroyed by CM(ofed) after cm_callback
829 * returned, so we can't refer it anymore
830 * (by kiblnd_connd()->kiblnd_destroy_conn) */
831 rdma_destroy_qp(conn->ibc_cmid);
832 conn->ibc_cmid = NULL;
834 /* Drop my own and unused rxbuffer refcounts */
835 while (i++ <= IBLND_RX_MSGS(version))
836 kiblnd_conn_decref(conn);
842 /* Init successful! */
843 LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
844 state == IBLND_CONN_PASSIVE_WAIT);
845 conn->ibc_state = state;
848 atomic_inc(&net->ibn_nconns);
852 kiblnd_destroy_conn(conn);
854 LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
860 kiblnd_destroy_conn (kib_conn_t *conn)
862 struct rdma_cm_id *cmid = conn->ibc_cmid;
863 kib_peer_t *peer = conn->ibc_peer;
866 LASSERT (!in_interrupt());
867 LASSERT (atomic_read(&conn->ibc_refcount) == 0);
868 LASSERT (list_empty(&conn->ibc_early_rxs));
869 LASSERT (list_empty(&conn->ibc_tx_queue));
870 LASSERT (list_empty(&conn->ibc_tx_queue_rsrvd));
871 LASSERT (list_empty(&conn->ibc_tx_queue_nocred));
872 LASSERT (list_empty(&conn->ibc_active_txs));
873 LASSERT (conn->ibc_noops_posted == 0);
874 LASSERT (conn->ibc_nsends_posted == 0);
876 switch (conn->ibc_state) {
878 /* conn must be completely disengaged from the network */
881 case IBLND_CONN_DISCONNECTED:
882 /* connvars should have been freed already */
883 LASSERT (conn->ibc_connvars == NULL);
886 case IBLND_CONN_INIT:
890 /* conn->ibc_cmid might be destroyed by CM already */
891 if (cmid != NULL && cmid->qp != NULL)
892 rdma_destroy_qp(cmid);
894 if (conn->ibc_cq != NULL) {
895 rc = ib_destroy_cq(conn->ibc_cq);
897 CWARN("Error destroying CQ: %d\n", rc);
900 if (conn->ibc_rx_pages != NULL)
901 kiblnd_unmap_rx_descs(conn);
903 if (conn->ibc_rxs != NULL) {
904 LIBCFS_FREE(conn->ibc_rxs,
905 IBLND_RX_MSGS(conn->ibc_version) * sizeof(kib_rx_t));
908 if (conn->ibc_connvars != NULL)
909 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
911 /* See CAVEAT EMPTOR above in kiblnd_create_conn */
912 if (conn->ibc_state != IBLND_CONN_INIT) {
913 kib_net_t *net = peer->ibp_ni->ni_data;
915 kiblnd_peer_decref(peer);
916 rdma_destroy_id(cmid);
917 atomic_dec(&net->ibn_nconns);
920 LIBCFS_FREE(conn, sizeof(*conn));
924 kiblnd_close_peer_conns_locked (kib_peer_t *peer, int why)
927 struct list_head *ctmp;
928 struct list_head *cnxt;
931 list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
932 conn = list_entry(ctmp, kib_conn_t, ibc_list);
934 CDEBUG(D_NET, "Closing conn -> %s, "
935 "version: %x, reason: %d\n",
936 libcfs_nid2str(peer->ibp_nid),
937 conn->ibc_version, why);
939 kiblnd_close_conn_locked(conn, why);
947 kiblnd_close_stale_conns_locked (kib_peer_t *peer,
948 int version, __u64 incarnation)
951 struct list_head *ctmp;
952 struct list_head *cnxt;
955 list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
956 conn = list_entry(ctmp, kib_conn_t, ibc_list);
958 if (conn->ibc_version == version &&
959 conn->ibc_incarnation == incarnation)
962 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
963 "incarnation:"LPX64"(%x, "LPX64")\n",
964 libcfs_nid2str(peer->ibp_nid),
965 conn->ibc_version, conn->ibc_incarnation,
966 version, incarnation);
968 kiblnd_close_conn_locked(conn, -ESTALE);
976 kiblnd_close_matching_conns (lnet_ni_t *ni, lnet_nid_t nid)
979 struct list_head *ptmp;
980 struct list_head *pnxt;
987 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
989 if (nid != LNET_NID_ANY)
990 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
993 hi = kiblnd_data.kib_peer_hash_size - 1;
996 for (i = lo; i <= hi; i++) {
997 list_for_each_safe (ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
999 peer = list_entry(ptmp, kib_peer_t, ibp_list);
1000 LASSERT (peer->ibp_connecting > 0 ||
1001 peer->ibp_accepting > 0 ||
1002 !list_empty(&peer->ibp_conns));
1004 if (peer->ibp_ni != ni)
1007 if (!(nid == LNET_NID_ANY || nid == peer->ibp_nid))
1010 count += kiblnd_close_peer_conns_locked(peer, 0);
1014 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1016 /* wildcards always succeed */
1017 if (nid == LNET_NID_ANY)
1020 return (count == 0) ? -ENOENT : 0;
1024 kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
1026 struct libcfs_ioctl_data *data = arg;
1030 case IOC_LIBCFS_GET_PEER: {
1034 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1036 data->ioc_nid = nid;
1037 data->ioc_count = count;
1041 case IOC_LIBCFS_DEL_PEER: {
1042 rc = kiblnd_del_peer(ni, data->ioc_nid);
1045 case IOC_LIBCFS_GET_CONN: {
1049 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1055 LASSERT (conn->ibc_cmid != NULL);
1056 data->ioc_nid = conn->ibc_peer->ibp_nid;
1057 if (conn->ibc_cmid->route.path_rec == NULL)
1058 data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1061 ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1062 kiblnd_conn_decref(conn);
1065 case IOC_LIBCFS_CLOSE_CONNECTION: {
1066 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1078 kiblnd_query (lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when)
1080 cfs_time_t last_alive = 0;
1081 cfs_time_t now = cfs_time_current();
1082 rwlock_t *glock = &kiblnd_data.kib_global_lock;
1084 unsigned long flags;
1086 read_lock_irqsave(glock, flags);
1088 peer = kiblnd_find_peer_locked(nid);
1090 LASSERT (peer->ibp_connecting > 0 || /* creating conns */
1091 peer->ibp_accepting > 0 ||
1092 !list_empty(&peer->ibp_conns)); /* active conn */
1093 last_alive = peer->ibp_last_alive;
1096 read_unlock_irqrestore(glock, flags);
1098 if (last_alive != 0)
1101 /* peer is not persistent in hash, trigger peer creation
1102 * and connection establishment with a NULL tx */
1104 kiblnd_launch_tx(ni, NULL, nid);
1106 CDEBUG(D_NET, "Peer %s %p, alive %ld secs ago\n",
1107 libcfs_nid2str(nid), peer,
1108 last_alive ? cfs_duration_sec(now - last_alive) : -1);
1113 kiblnd_free_pages (kib_pages_t *p)
1115 int npages = p->ibp_npages;
1118 LASSERT (p->ibp_device == NULL);
1120 for (i = 0; i < npages; i++)
1121 if (p->ibp_pages[i] != NULL)
1122 __free_page(p->ibp_pages[i]);
1124 LIBCFS_FREE (p, offsetof(kib_pages_t, ibp_pages[npages]));
1128 kiblnd_alloc_pages (kib_pages_t **pp, int npages)
1133 LIBCFS_ALLOC(p, offsetof(kib_pages_t, ibp_pages[npages]));
1135 CERROR("Can't allocate descriptor for %d pages\n", npages);
1139 memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1140 p->ibp_npages = npages;
1142 for (i = 0; i < npages; i++) {
1143 p->ibp_pages[i] = alloc_page(GFP_KERNEL);
1144 if (p->ibp_pages[i] == NULL) {
1145 CERROR("Can't allocate page %d of %d\n", i, npages);
1146 kiblnd_free_pages(p);
1156 kiblnd_unmap_rx_descs(kib_conn_t *conn)
1161 LASSERT (conn->ibc_rxs != NULL);
1162 LASSERT (conn->ibc_rx_pages->ibp_device != NULL);
1164 for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1165 rx = &conn->ibc_rxs[i];
1167 LASSERT (rx->rx_nob >= 0); /* not posted */
1169 kiblnd_dma_unmap_single(conn->ibc_rx_pages->ibp_device,
1170 KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1172 IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1175 conn->ibc_rx_pages->ibp_device = NULL;
1177 kiblnd_free_pages(conn->ibc_rx_pages);
1179 conn->ibc_rx_pages = NULL;
1183 kiblnd_map_rx_descs(kib_conn_t *conn)
1191 for (pg_off = ipg = i = 0;
1192 i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1193 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1194 rx = &conn->ibc_rxs[i];
1197 rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
1199 rx->rx_msgaddr = kiblnd_dma_map_single(conn->ibc_cmid->device,
1200 rx->rx_msg, IBLND_MSG_SIZE,
1202 LASSERT (!kiblnd_dma_mapping_error(conn->ibc_cmid->device,
1204 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1206 CDEBUG(D_NET,"rx %d: %p "LPX64"("LPX64")\n",
1207 i, rx->rx_msg, rx->rx_msgaddr,
1208 lnet_page2phys(pg) + pg_off);
1210 pg_off += IBLND_MSG_SIZE;
1211 LASSERT (pg_off <= PAGE_SIZE);
1213 if (pg_off == PAGE_SIZE) {
1216 LASSERT (ipg <= IBLND_RX_MSG_PAGES(conn->ibc_version));
1220 conn->ibc_rx_pages->ibp_device = conn->ibc_cmid->device;
1224 kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo)
1229 LASSERT (tpo->tpo_pool.po_allocated == 0);
1230 LASSERT (tpo->tpo_tx_pages->ibp_device != NULL);
1232 for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1233 tx = &tpo->tpo_tx_descs[i];
1234 kiblnd_dma_unmap_single(tpo->tpo_tx_pages->ibp_device,
1235 KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1237 IBLND_MSG_SIZE, DMA_TO_DEVICE);
1239 tpo->tpo_tx_pages->ibp_device = NULL;
1243 kiblnd_map_tx_pool(kib_tx_pool_t *tpo)
1245 kib_pages_t *txpgs = tpo->tpo_tx_pages;
1246 kib_pool_t *pool = &tpo->tpo_pool;
1247 kib_net_t *net = pool->po_owner->ps_net;
1254 LASSERT (net != NULL);
1255 LASSERT (net->ibn_dev->ibd_cmid != NULL &&
1256 net->ibn_dev->ibd_cmid->device != NULL);
1258 /* pre-mapped messages are not bigger than 1 page */
1259 CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE);
1261 /* No fancy arithmetic when we do the buffer calculations */
1262 CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
1264 txpgs->ibp_device = net->ibn_dev->ibd_cmid->device;
1266 for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1267 page = txpgs->ibp_pages[ipage];
1268 tx = &tpo->tpo_tx_descs[i];
1270 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1273 tx->tx_msgaddr = kiblnd_dma_map_single(
1274 txpgs->ibp_device, tx->tx_msg,
1275 IBLND_MSG_SIZE, DMA_TO_DEVICE);
1276 LASSERT (!kiblnd_dma_mapping_error(txpgs->ibp_device,
1278 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1280 list_add(&tx->tx_list, &pool->po_free_list);
1282 page_offset += IBLND_MSG_SIZE;
1283 LASSERT (page_offset <= PAGE_SIZE);
1285 if (page_offset == PAGE_SIZE) {
1288 LASSERT (ipage <= txpgs->ibp_npages);
1294 kiblnd_find_dma_mr(kib_net_t *net, __u64 addr, __u64 size)
1298 LASSERT (net->ibn_dev->ibd_mrs[0] != NULL);
1300 if (net->ibn_dev->ibd_nmrs == 1)
1301 return net->ibn_dev->ibd_mrs[0];
1303 index = addr >> net->ibn_dev->ibd_mr_shift;
1305 if (index < net->ibn_dev->ibd_nmrs &&
1306 index == ((addr + size - 1) >> net->ibn_dev->ibd_mr_shift))
1307 return net->ibn_dev->ibd_mrs[index];
1313 kiblnd_find_rd_dma_mr(kib_net_t *net, kib_rdma_desc_t *rd)
1315 struct ib_mr *prev_mr;
1319 LASSERT (net->ibn_dev->ibd_mrs[0] != NULL);
1321 if (*kiblnd_tunables.kib_map_on_demand > 0 &&
1322 *kiblnd_tunables.kib_map_on_demand <= rd->rd_nfrags)
1325 if (net->ibn_dev->ibd_nmrs == 1)
1326 return net->ibn_dev->ibd_mrs[0];
1328 for (i = 0, mr = prev_mr = NULL;
1329 i < rd->rd_nfrags; i++) {
1330 mr = kiblnd_find_dma_mr(net,
1331 rd->rd_frags[i].rf_addr,
1332 rd->rd_frags[i].rf_nob);
1333 if (prev_mr == NULL)
1336 if (mr == NULL || prev_mr != mr) {
1337 /* Can't covered by one single MR */
1347 kiblnd_destroy_fmr_pool(kib_fmr_pool_t *pool)
1349 LASSERT (pool->fpo_map_count == 0);
1351 if (pool->fpo_fmr_pool != NULL)
1352 ib_destroy_fmr_pool(pool->fpo_fmr_pool);
1354 LIBCFS_FREE(pool, sizeof(kib_fmr_pool_t));
1358 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1360 kib_fmr_pool_t *pool;
1362 while (!list_empty(head)) {
1363 pool = list_entry(head->next, kib_fmr_pool_t, fpo_list);
1364 list_del(&pool->fpo_list);
1365 kiblnd_destroy_fmr_pool(pool);
1370 kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo)
1372 /* FMR pool for RDMA */
1373 kib_fmr_pool_t *fpo;
1374 struct ib_fmr_pool_param param = {
1375 .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
1376 .page_shift = PAGE_SHIFT,
1377 .access = (IB_ACCESS_LOCAL_WRITE |
1378 IB_ACCESS_REMOTE_WRITE),
1379 .pool_size = *kiblnd_tunables.kib_fmr_pool_size,
1380 .dirty_watermark = *kiblnd_tunables.kib_fmr_flush_trigger,
1381 .flush_function = NULL,
1383 .cache = !!*kiblnd_tunables.kib_fmr_cache};
1386 LASSERT (fps->fps_net->ibn_dev != NULL &&
1387 fps->fps_net->ibn_dev->ibd_pd != NULL);
1389 LIBCFS_ALLOC(fpo, sizeof(kib_fmr_pool_t));
1393 memset(fpo, 0, sizeof(kib_fmr_pool_t));
1394 fpo->fpo_fmr_pool = ib_create_fmr_pool(fps->fps_net->ibn_dev->ibd_pd, ¶m);
1395 if (IS_ERR(fpo->fpo_fmr_pool)) {
1396 CERROR("Failed to create FMR pool: %ld\n",
1397 PTR_ERR(fpo->fpo_fmr_pool));
1398 rc = PTR_ERR(fpo->fpo_fmr_pool);
1399 LIBCFS_FREE(fpo, sizeof(kib_fmr_pool_t));
1403 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1404 fpo->fpo_owner = fps;
1411 kiblnd_fini_fmr_pool_set(kib_fmr_poolset_t *fps)
1413 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1417 kiblnd_init_fmr_pool_set(kib_fmr_poolset_t *fps, kib_net_t *net)
1419 kib_fmr_pool_t *fpo;
1422 memset(fps, 0, sizeof(kib_fmr_poolset_t));
1425 spin_lock_init(&fps->fps_lock);
1426 CFS_INIT_LIST_HEAD(&fps->fps_pool_list);
1427 rc = kiblnd_create_fmr_pool(fps, &fpo);
1429 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1435 kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1437 CFS_LIST_HEAD (zombies);
1438 kib_fmr_pool_t *fpo = fmr->fmr_pool;
1439 kib_fmr_poolset_t *fps = fpo->fpo_owner;
1440 kib_fmr_pool_t *tmp;
1443 rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1447 rc = ib_flush_fmr_pool(fpo->fpo_fmr_pool);
1451 fmr->fmr_pool = NULL;
1452 fmr->fmr_pfmr = NULL;
1454 spin_lock(&fps->fps_lock);
1455 fpo->fpo_map_count --; /* decref the pool */
1457 list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1458 /* the first pool is persistent */
1459 if (fps->fps_pool_list.next == &fpo->fpo_list)
1462 if (fpo->fpo_map_count == 0 && /* no more reference */
1463 cfs_time_aftereq(cfs_time_current(), fpo->fpo_deadline)) {
1464 list_move(&fpo->fpo_list, &zombies);
1465 fps->fps_version ++;
1468 spin_unlock(&fps->fps_lock);
1470 if (!list_empty(&zombies))
1471 kiblnd_destroy_fmr_pool_list(&zombies);
1475 kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages,
1476 __u64 iov, kib_fmr_t *fmr)
1478 struct ib_pool_fmr *pfmr;
1479 kib_fmr_pool_t *fpo;
1483 LASSERT (fps->fps_net->ibn_with_fmr);
1485 spin_lock(&fps->fps_lock);
1486 version = fps->fps_version;
1487 list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1488 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1489 fpo->fpo_map_count ++;
1490 spin_unlock(&fps->fps_lock);
1492 pfmr = ib_fmr_pool_map_phys(fpo->fpo_fmr_pool,
1493 pages, npages, iov);
1494 if (likely(!IS_ERR(pfmr))) {
1495 fmr->fmr_pool = fpo;
1496 fmr->fmr_pfmr = pfmr;
1500 spin_lock(&fps->fps_lock);
1501 fpo->fpo_map_count --;
1502 if (PTR_ERR(pfmr) != -EAGAIN) {
1503 spin_unlock(&fps->fps_lock);
1504 return PTR_ERR(pfmr);
1507 /* EAGAIN and ... */
1508 if (version != fps->fps_version) {
1509 spin_unlock(&fps->fps_lock);
1514 if (fps->fps_increasing) {
1515 spin_unlock(&fps->fps_lock);
1516 CDEBUG(D_NET, "Another thread is allocating new "
1517 "FMR pool, waiting for her to complete\n");
1523 if (cfs_time_before(cfs_time_current(), fps->fps_next_retry)) {
1524 /* someone failed recently */
1525 spin_unlock(&fps->fps_lock);
1529 fps->fps_increasing = 1;
1530 spin_unlock(&fps->fps_lock);
1532 CDEBUG(D_NET, "Allocate new FMR pool\n");
1533 rc = kiblnd_create_fmr_pool(fps, &fpo);
1534 spin_lock(&fps->fps_lock);
1535 fps->fps_increasing = 0;
1537 fps->fps_version ++;
1538 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1540 fps->fps_next_retry = cfs_time_shift(10);
1542 spin_unlock(&fps->fps_lock);
1548 kiblnd_fini_pool(kib_pool_t *pool)
1550 LASSERT (list_empty(&pool->po_free_list));
1551 LASSERT (pool->po_allocated == 0);
1553 CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
1557 kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
1559 CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
1561 memset(pool, 0, sizeof(kib_pool_t));
1562 CFS_INIT_LIST_HEAD(&pool->po_free_list);
1563 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1564 pool->po_owner = ps;
1565 pool->po_size = size;
1569 kiblnd_destroy_pool_list(kib_poolset_t *ps, struct list_head *head)
1573 while (!list_empty(head)) {
1574 pool = list_entry(head->next, kib_pool_t, po_list);
1575 list_del(&pool->po_list);
1576 ps->ps_pool_destroy(pool);
1581 kiblnd_fini_pool_set(kib_poolset_t *ps)
1583 kiblnd_destroy_pool_list(ps, &ps->ps_pool_list);
1587 kiblnd_init_pool_set(kib_poolset_t *ps, kib_net_t *net,
1588 char *name, int size,
1589 kib_ps_pool_create_t po_create,
1590 kib_ps_pool_destroy_t po_destroy,
1591 kib_ps_node_init_t nd_init,
1592 kib_ps_node_fini_t nd_fini)
1597 memset(ps, 0, sizeof(kib_poolset_t));
1600 ps->ps_pool_create = po_create;
1601 ps->ps_pool_destroy = po_destroy;
1602 ps->ps_node_init = nd_init;
1603 ps->ps_node_fini = nd_fini;
1604 ps->ps_pool_size = size;
1605 strncpy(ps->ps_name, name, IBLND_POOL_NAME_LEN);
1606 spin_lock_init(&ps->ps_lock);
1607 CFS_INIT_LIST_HEAD(&ps->ps_pool_list);
1609 rc = ps->ps_pool_create(ps, size, &pool);
1611 list_add(&pool->po_list, &ps->ps_pool_list);
1613 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
1619 kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node)
1621 CFS_LIST_HEAD (zombies);
1622 kib_poolset_t *ps = pool->po_owner;
1624 cfs_time_t now = cfs_time_current();
1626 spin_lock(&ps->ps_lock);
1628 if (ps->ps_node_fini != NULL)
1629 ps->ps_node_fini(pool, node);
1631 LASSERT (pool->po_allocated > 0);
1632 list_add(node, &pool->po_free_list);
1633 pool->po_allocated --;
1635 list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
1636 /* the first pool is persistent */
1637 if (ps->ps_pool_list.next == &pool->po_list)
1640 if (pool->po_allocated == 0 &&
1641 cfs_time_aftereq(now, pool->po_deadline))
1642 list_move(&pool->po_list, &zombies);
1644 spin_unlock(&ps->ps_lock);
1646 if (!list_empty(&zombies))
1647 kiblnd_destroy_pool_list(ps, &zombies);
1651 kiblnd_pool_alloc_node(kib_poolset_t *ps)
1653 struct list_head *node;
1658 spin_lock(&ps->ps_lock);
1659 list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
1660 if (list_empty(&pool->po_free_list))
1663 pool->po_allocated ++;
1664 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1665 node = pool->po_free_list.next;
1668 if (ps->ps_node_init != NULL) {
1669 /* still hold the lock */
1670 ps->ps_node_init(pool, node);
1672 spin_unlock(&ps->ps_lock);
1676 /* no available tx pool and ... */
1677 if (ps->ps_increasing) {
1678 /* another thread is allocating a new pool */
1679 spin_unlock(&ps->ps_lock);
1680 CDEBUG(D_NET, "Another thread is allocating new "
1681 "%s pool, waiting for her to complete\n",
1687 if (cfs_time_before(cfs_time_current(), ps->ps_next_retry)) {
1688 /* someone failed recently */
1689 spin_unlock(&ps->ps_lock);
1693 ps->ps_increasing = 1;
1694 spin_unlock(&ps->ps_lock);
1696 CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
1698 rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
1700 spin_lock(&ps->ps_lock);
1701 ps->ps_increasing = 0;
1703 list_add_tail(&pool->po_list, &ps->ps_pool_list);
1705 /* retry 10 seconds later */
1706 ps->ps_next_retry = cfs_time_shift(10);
1707 CERROR("Can't allocate new %s pool because out of memory\n",
1710 spin_unlock(&ps->ps_lock);
1716 kiblnd_pmr_pool_unmap(kib_phys_mr_t *pmr)
1718 kib_pmr_pool_t *ppo = pmr->pmr_pool;
1719 struct ib_mr *mr = pmr->pmr_mr;
1722 kiblnd_pool_free_node(&ppo->ppo_pool, &pmr->pmr_list);
1728 kiblnd_pmr_pool_map(kib_pmr_poolset_t *pps, kib_rdma_desc_t *rd,
1729 __u64 *iova, kib_phys_mr_t **pp_pmr)
1732 struct list_head *node;
1736 node = kiblnd_pool_alloc_node(&pps->pps_poolset);
1738 CERROR("Failed to allocate PMR descriptor\n");
1742 pmr = container_of(node, kib_phys_mr_t, pmr_list);
1743 for (i = 0; i < rd->rd_nfrags; i ++) {
1744 pmr->pmr_ipb[i].addr = rd->rd_frags[i].rf_addr;
1745 pmr->pmr_ipb[i].size = rd->rd_frags[i].rf_nob;
1748 pmr->pmr_mr = ib_reg_phys_mr(pps->pps_poolset.ps_net->ibn_dev->ibd_pd,
1749 pmr->pmr_ipb, rd->rd_nfrags,
1750 IB_ACCESS_LOCAL_WRITE |
1751 IB_ACCESS_REMOTE_WRITE,
1753 if (!IS_ERR(pmr->pmr_mr)) {
1754 pmr->pmr_iova = *iova;
1759 rc = PTR_ERR(pmr->pmr_mr);
1760 CERROR("Failed ib_reg_phys_mr: %d\n", rc);
1763 kiblnd_pool_free_node(&pmr->pmr_pool->ppo_pool, node);
1769 kiblnd_destroy_pmr_pool(kib_pool_t *pool)
1771 kib_pmr_pool_t *ppo = container_of(pool, kib_pmr_pool_t, ppo_pool);
1774 LASSERT (pool->po_allocated == 0);
1776 while (!list_empty(&pool->po_free_list)) {
1777 pmr = list_entry(pool->po_free_list.next,
1778 kib_phys_mr_t, pmr_list);
1780 LASSERT (pmr->pmr_mr == NULL);
1781 list_del(&pmr->pmr_list);
1783 if (pmr->pmr_ipb != NULL) {
1784 LIBCFS_FREE(pmr->pmr_ipb,
1785 IBLND_MAX_RDMA_FRAGS *
1786 sizeof(struct ib_phys_buf));
1789 LIBCFS_FREE(pmr, sizeof(kib_phys_mr_t));
1792 kiblnd_fini_pool(pool);
1793 LIBCFS_FREE(ppo, sizeof(kib_pmr_pool_t));
1797 kiblnd_create_pmr_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
1799 kib_pmr_pool_t *ppo;
1804 LIBCFS_ALLOC(ppo, sizeof(kib_pmr_pool_t));
1806 CERROR("Failed to allocate PMR pool\n");
1810 pool = &ppo->ppo_pool;
1811 kiblnd_init_pool(ps, pool, size);
1813 for (i = 0; i < size; i++) {
1814 LIBCFS_ALLOC(pmr, sizeof(kib_phys_mr_t));
1818 memset(pmr, 0, sizeof(kib_phys_mr_t));
1819 pmr->pmr_pool = ppo;
1820 LIBCFS_ALLOC(pmr->pmr_ipb,
1821 IBLND_MAX_RDMA_FRAGS *
1822 sizeof(struct ib_phys_buf));
1823 if (pmr->pmr_ipb == NULL)
1826 list_add(&pmr->pmr_list, &pool->po_free_list);
1830 ps->ps_pool_destroy(pool);
1839 kiblnd_destroy_tx_pool(kib_pool_t *pool)
1841 kib_tx_pool_t *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
1844 LASSERT (pool->po_allocated == 0);
1846 if (tpo->tpo_tx_pages != NULL) {
1847 if (tpo->tpo_tx_pages->ibp_device != NULL)
1848 kiblnd_unmap_tx_pool(tpo);
1849 kiblnd_free_pages(tpo->tpo_tx_pages);
1852 if (tpo->tpo_tx_descs == NULL)
1855 for (i = 0; i < pool->po_size; i++) {
1856 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
1858 list_del(&tx->tx_list);
1859 if (tx->tx_pages != NULL)
1860 LIBCFS_FREE(tx->tx_pages,
1862 sizeof(*tx->tx_pages));
1863 if (tx->tx_frags != NULL)
1864 LIBCFS_FREE(tx->tx_frags,
1865 IBLND_MAX_RDMA_FRAGS *
1866 sizeof(*tx->tx_frags));
1867 if (tx->tx_wrq != NULL)
1868 LIBCFS_FREE(tx->tx_wrq,
1869 (1 + IBLND_MAX_RDMA_FRAGS) *
1870 sizeof(*tx->tx_wrq));
1871 if (tx->tx_sge != NULL)
1872 LIBCFS_FREE(tx->tx_sge,
1873 (1 + IBLND_MAX_RDMA_FRAGS) *
1874 sizeof(*tx->tx_sge));
1875 if (tx->tx_rd != NULL)
1876 LIBCFS_FREE(tx->tx_rd,
1877 offsetof(kib_rdma_desc_t,
1878 rd_frags[IBLND_MAX_RDMA_FRAGS]));
1881 LIBCFS_FREE(tpo->tpo_tx_descs,
1882 pool->po_size * sizeof(kib_tx_t));
1884 kiblnd_fini_pool(pool);
1885 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
1889 kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
1896 LIBCFS_ALLOC(tpo, sizeof(kib_tx_pool_t));
1898 CERROR("Failed to allocate TX pool\n");
1902 pool = &tpo->tpo_pool;
1903 kiblnd_init_pool(ps, pool, size);
1904 tpo->tpo_tx_descs = NULL;
1905 tpo->tpo_tx_pages = NULL;
1907 npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
1908 if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, npg) != 0) {
1909 CERROR("Can't allocate tx pages: %d\n", npg);
1910 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
1914 LIBCFS_ALLOC (tpo->tpo_tx_descs, size * sizeof(kib_tx_t));
1915 if (tpo->tpo_tx_descs == NULL) {
1916 CERROR("Can't allocate %d tx descriptors\n", size);
1917 ps->ps_pool_destroy(pool);
1921 memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
1923 for (i = 0; i < size; i++) {
1924 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
1927 if (ps->ps_net->ibn_with_fmr){
1928 LIBCFS_ALLOC(tx->tx_pages, LNET_MAX_IOV *
1929 sizeof(*tx->tx_pages));
1930 if (tx->tx_pages == NULL)
1934 LIBCFS_ALLOC(tx->tx_frags,
1935 IBLND_MAX_RDMA_FRAGS *
1936 sizeof(*tx->tx_frags));
1937 if (tx->tx_frags == NULL)
1940 LIBCFS_ALLOC(tx->tx_wrq,
1941 (1 + IBLND_MAX_RDMA_FRAGS) *
1942 sizeof(*tx->tx_wrq));
1943 if (tx->tx_wrq == NULL)
1946 LIBCFS_ALLOC(tx->tx_sge,
1947 (1 + IBLND_MAX_RDMA_FRAGS) *
1948 sizeof(*tx->tx_sge));
1949 if (tx->tx_sge == NULL)
1952 LIBCFS_ALLOC(tx->tx_rd,
1953 offsetof(kib_rdma_desc_t,
1954 rd_frags[IBLND_MAX_RDMA_FRAGS]));
1955 if (tx->tx_rd == NULL)
1960 kiblnd_map_tx_pool(tpo);
1965 ps->ps_pool_destroy(pool);
1970 kiblnd_tx_init(kib_pool_t *pool, struct list_head *node)
1972 kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t, tps_poolset);
1973 kib_tx_t *tx = list_entry(node, kib_tx_t, tx_list);
1975 tx->tx_cookie = tps->tps_next_tx_cookie ++;
1979 kiblnd_ni_fini_pools(kib_net_t *net)
1981 kiblnd_fini_pool_set(&net->ibn_tx_ps.tps_poolset);
1982 if (net->ibn_with_fmr)
1983 kiblnd_fini_fmr_pool_set(&net->ibn_fmr_ps);
1984 else if (net->ibn_with_pmr)
1985 kiblnd_fini_pool_set(&net->ibn_pmr_ps.pps_poolset);
1989 kiblnd_net_init_pools(kib_net_t *net)
1991 kib_fmr_poolset_t *fps = &net->ibn_fmr_ps;
1992 kib_pmr_poolset_t *pps = &net->ibn_pmr_ps;
1993 kib_tx_poolset_t *tps = &net->ibn_tx_ps;
1996 if (*kiblnd_tunables.kib_fmr_pool_size <
1997 *kiblnd_tunables.kib_ntx / 4) {
1998 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
1999 *kiblnd_tunables.kib_fmr_pool_size,
2000 *kiblnd_tunables.kib_ntx / 4);
2004 if (*kiblnd_tunables.kib_pmr_pool_size <
2005 *kiblnd_tunables.kib_ntx / 4) {
2006 CERROR("Can't set pmr pool size (%d) < ntx / 4(%d)\n",
2007 *kiblnd_tunables.kib_pmr_pool_size,
2008 *kiblnd_tunables.kib_ntx / 4);
2012 if (*kiblnd_tunables.kib_map_on_demand > 0 ||
2013 net->ibn_dev->ibd_nmrs > 1) { /* premapping can fail if ibd_nmr > 1,
2014 * so we always create FMR/PMR pool and
2015 * map-on-demand if premapping failed */
2016 rc = kiblnd_init_fmr_pool_set(fps, net);
2018 net->ibn_with_fmr = 1;
2019 } else if (rc == -ENOSYS) {
2020 rc = kiblnd_init_pool_set(&pps->pps_poolset, net, "PMR",
2021 *kiblnd_tunables.kib_pmr_pool_size,
2022 kiblnd_create_pmr_pool,
2023 kiblnd_destroy_pmr_pool,
2026 net->ibn_with_pmr = 1;
2032 rc = kiblnd_init_pool_set(&tps->tps_poolset, net, "TX", IBLND_TX_MSGS(),
2033 kiblnd_create_tx_pool, kiblnd_destroy_tx_pool,
2034 kiblnd_tx_init, NULL);
2038 if (net->ibn_with_fmr)
2039 kiblnd_fini_fmr_pool_set(fps);
2040 else if (net->ibn_with_pmr)
2041 kiblnd_fini_pool_set(&pps->pps_poolset);
2047 kiblnd_dev_cleanup(kib_dev_t *ibdev)
2051 if (ibdev->ibd_mrs == NULL)
2054 for (i = 0; i < ibdev->ibd_nmrs; i++) {
2055 if (ibdev->ibd_mrs[i] == NULL)
2058 ib_dereg_mr(ibdev->ibd_mrs[i]);
2061 LIBCFS_FREE(ibdev->ibd_mrs, sizeof(*ibdev->ibd_mrs) * ibdev->ibd_nmrs);
2062 ibdev->ibd_mrs = NULL;
2066 kiblnd_dev_get_attr(kib_dev_t *ibdev)
2068 struct ib_device_attr *attr;
2071 /* It's safe to assume a HCA can handle a page size
2072 * matching that of the native system */
2073 ibdev->ibd_page_shift = PAGE_SHIFT;
2074 ibdev->ibd_page_size = 1 << PAGE_SHIFT;
2075 ibdev->ibd_page_mask = ~((__u64)ibdev->ibd_page_size - 1);
2077 LIBCFS_ALLOC(attr, sizeof(*attr));
2079 CERROR("Out of memory\n");
2083 rc = ib_query_device(ibdev->ibd_cmid->device, attr);
2085 ibdev->ibd_mr_size = attr->max_mr_size;
2087 LIBCFS_FREE(attr, sizeof(*attr));
2090 CERROR("Failed to query IB device: %d\n", rc);
2094 #ifdef HAVE_OFED_TRANSPORT_IWARP
2095 /* XXX We can't trust this value returned by Chelsio driver, it's wrong
2096 * and we have reported the bug, remove these in the future when Chelsio
2098 if (rdma_node_get_transport(ibdev->ibd_cmid->device->node_type) ==
2099 RDMA_TRANSPORT_IWARP)
2100 ibdev->ibd_mr_size = (1ULL << 32) - 1;
2103 if (ibdev->ibd_mr_size == ~0ULL) {
2104 ibdev->ibd_mr_shift = 64;
2108 for (ibdev->ibd_mr_shift = 0;
2109 ibdev->ibd_mr_shift < 64; ibdev->ibd_mr_shift ++) {
2110 if (ibdev->ibd_mr_size == (1ULL << ibdev->ibd_mr_shift) ||
2111 ibdev->ibd_mr_size == (1ULL << ibdev->ibd_mr_shift) - 1)
2115 CERROR("Invalid mr size: "LPX64"\n", ibdev->ibd_mr_size);
2120 kiblnd_dev_setup(kib_dev_t *ibdev)
2127 int acflags = IB_ACCESS_LOCAL_WRITE |
2128 IB_ACCESS_REMOTE_WRITE;
2130 rc = kiblnd_dev_get_attr(ibdev);
2134 if (ibdev->ibd_mr_shift == 64) {
2135 LIBCFS_ALLOC(ibdev->ibd_mrs, 1 * sizeof(*ibdev->ibd_mrs));
2136 if (ibdev->ibd_mrs == NULL) {
2137 CERROR("Failed to allocate MRs table\n");
2141 ibdev->ibd_mrs[0] = NULL;
2142 ibdev->ibd_nmrs = 1;
2144 mr = ib_get_dma_mr(ibdev->ibd_pd, acflags);
2146 CERROR("Failed ib_get_dma_mr : %ld\n", PTR_ERR(mr));
2147 kiblnd_dev_cleanup(ibdev);
2151 ibdev->ibd_mrs[0] = mr;
2156 mr_size = (1ULL << ibdev->ibd_mr_shift);
2157 mm_size = (unsigned long)high_memory - PAGE_OFFSET;
2159 ibdev->ibd_nmrs = (int)((mm_size + mr_size - 1) >> ibdev->ibd_mr_shift);
2161 if (ibdev->ibd_mr_shift < 32 || ibdev->ibd_nmrs > 1024) {
2162 /* it's 4T..., assume we will re-code at that time */
2163 CERROR("Can't support memory size: x"LPX64
2164 " with MR size: x"LPX64"\n", mm_size, mr_size);
2168 /* create an array of MRs to cover all memory */
2169 LIBCFS_ALLOC(ibdev->ibd_mrs, sizeof(*ibdev->ibd_mrs) * ibdev->ibd_nmrs);
2170 if (ibdev->ibd_mrs == NULL) {
2171 CERROR("Failed to allocate MRs' table\n");
2175 memset(ibdev->ibd_mrs, 0, sizeof(*ibdev->ibd_mrs) * ibdev->ibd_nmrs);
2177 for (i = 0; i < ibdev->ibd_nmrs; i++) {
2178 struct ib_phys_buf ipb;
2181 ipb.size = ibdev->ibd_mr_size;
2182 ipb.addr = i * mr_size;
2185 mr = ib_reg_phys_mr(ibdev->ibd_pd, &ipb, 1, acflags, &iova);
2187 CERROR("Failed ib_reg_phys_mr addr "LPX64
2188 " size "LPX64" : %ld\n",
2189 ipb.addr, ipb.size, PTR_ERR(mr));
2190 kiblnd_dev_cleanup(ibdev);
2194 LASSERT (iova == ipb.addr);
2196 ibdev->ibd_mrs[i] = mr;
2200 CDEBUG(D_CONSOLE, "Register global MR array, MR size: "
2201 LPX64", array size: %d\n",
2202 ibdev->ibd_mr_size, ibdev->ibd_nmrs);
2204 list_add_tail(&ibdev->ibd_list,
2205 &kiblnd_data.kib_devs);
2210 kiblnd_destroy_dev (kib_dev_t *dev)
2212 LASSERT (dev->ibd_nnets == 0);
2214 if (!list_empty(&dev->ibd_list)) /* on kib_devs? */
2215 list_del_init(&dev->ibd_list);
2217 kiblnd_dev_cleanup(dev);
2219 if (dev->ibd_pd != NULL)
2220 ib_dealloc_pd(dev->ibd_pd);
2222 if (dev->ibd_cmid != NULL)
2223 rdma_destroy_id(dev->ibd_cmid);
2225 LIBCFS_FREE(dev, sizeof(*dev));
2229 kiblnd_base_shutdown (void)
2233 LASSERT (list_empty(&kiblnd_data.kib_devs));
2235 CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2236 atomic_read(&libcfs_kmemory));
2238 switch (kiblnd_data.kib_init) {
2242 case IBLND_INIT_ALL:
2243 case IBLND_INIT_DATA:
2244 LASSERT (kiblnd_data.kib_peers != NULL);
2245 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2246 LASSERT (list_empty(&kiblnd_data.kib_peers[i]));
2248 LASSERT (list_empty(&kiblnd_data.kib_connd_zombies));
2249 LASSERT (list_empty(&kiblnd_data.kib_connd_conns));
2251 /* flag threads to terminate; wake and wait for them to die */
2252 kiblnd_data.kib_shutdown = 1;
2253 wake_up_all(&kiblnd_data.kib_sched_waitq);
2254 wake_up_all(&kiblnd_data.kib_connd_waitq);
2257 while (atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2259 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
2260 "Waiting for %d threads to terminate\n",
2261 atomic_read(&kiblnd_data.kib_nthreads));
2262 cfs_pause(cfs_time_seconds(1));
2267 case IBLND_INIT_NOTHING:
2271 if (kiblnd_data.kib_peers != NULL)
2272 LIBCFS_FREE(kiblnd_data.kib_peers,
2273 sizeof(struct list_head) *
2274 kiblnd_data.kib_peer_hash_size);
2276 CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2277 atomic_read(&libcfs_kmemory));
2279 kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2280 PORTAL_MODULE_UNUSE;
2284 kiblnd_shutdown (lnet_ni_t *ni)
2286 kib_net_t *net = ni->ni_data;
2287 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2289 unsigned long flags;
2291 LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2296 CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
2297 atomic_read(&libcfs_kmemory));
2299 write_lock_irqsave(g_lock, flags);
2300 net->ibn_shutdown = 1;
2301 write_unlock_irqrestore(g_lock, flags);
2303 switch (net->ibn_init) {
2307 case IBLND_INIT_ALL:
2308 /* nuke all existing peers within this net */
2309 kiblnd_del_peer(ni, LNET_NID_ANY);
2311 /* Wait for all peer state to clean up */
2313 while (atomic_read(&net->ibn_npeers) != 0) {
2315 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* 2**n? */
2316 "%s: waiting for %d peers to disconnect\n",
2317 libcfs_nid2str(ni->ni_nid),
2318 atomic_read(&net->ibn_npeers));
2319 cfs_pause(cfs_time_seconds(1));
2322 kiblnd_ni_fini_pools(net);
2324 LASSERT (net->ibn_dev->ibd_nnets > 0);
2325 net->ibn_dev->ibd_nnets--;
2329 case IBLND_INIT_NOTHING:
2330 LASSERT (atomic_read(&net->ibn_nconns) == 0);
2332 if (net->ibn_dev != NULL &&
2333 net->ibn_dev->ibd_nnets == 0)
2334 kiblnd_destroy_dev(net->ibn_dev);
2339 CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
2340 atomic_read(&libcfs_kmemory));
2342 net->ibn_init = IBLND_INIT_NOTHING;
2345 LIBCFS_FREE(net, sizeof(*net));
2348 if (list_empty(&kiblnd_data.kib_devs))
2349 kiblnd_base_shutdown();
2354 kiblnd_base_startup (void)
2359 LASSERT (kiblnd_data.kib_init == IBLND_INIT_NOTHING);
2362 memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
2364 rwlock_init(&kiblnd_data.kib_global_lock);
2366 INIT_LIST_HEAD(&kiblnd_data.kib_devs);
2368 kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
2369 LIBCFS_ALLOC(kiblnd_data.kib_peers,
2370 sizeof(struct list_head) * kiblnd_data.kib_peer_hash_size);
2371 if (kiblnd_data.kib_peers == NULL) {
2374 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2375 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
2377 spin_lock_init(&kiblnd_data.kib_connd_lock);
2378 INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
2379 INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
2380 init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
2382 spin_lock_init(&kiblnd_data.kib_sched_lock);
2383 INIT_LIST_HEAD(&kiblnd_data.kib_sched_conns);
2384 init_waitqueue_head(&kiblnd_data.kib_sched_waitq);
2386 kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
2388 /* lists/ptrs/locks initialised */
2389 kiblnd_data.kib_init = IBLND_INIT_DATA;
2390 /*****************************************************/
2392 for (i = 0; i < IBLND_N_SCHED; i++) {
2393 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)((long)i));
2395 CERROR("Can't spawn o2iblnd scheduler[%d]: %d\n",
2401 rc = kiblnd_thread_start(kiblnd_connd, NULL);
2403 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
2407 /* flag everything initialised */
2408 kiblnd_data.kib_init = IBLND_INIT_ALL;
2409 /*****************************************************/
2414 kiblnd_base_shutdown();
2419 kiblnd_startup (lnet_ni_t *ni)
2422 kib_dev_t *ibdev = NULL;
2424 struct list_head *tmp;
2428 LASSERT (ni->ni_lnd == &the_o2iblnd);
2430 if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
2431 rc = kiblnd_base_startup();
2436 LIBCFS_ALLOC(net, sizeof(*net));
2441 memset(net, 0, sizeof(*net));
2443 do_gettimeofday(&tv);
2444 net->ibn_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
2446 ni->ni_peertimeout = *kiblnd_tunables.kib_peertimeout;
2447 ni->ni_maxtxcredits = *kiblnd_tunables.kib_credits;
2448 ni->ni_peertxcredits = *kiblnd_tunables.kib_peertxcredits;
2449 ni->ni_peerrtrcredits = *kiblnd_tunables.kib_peerrtrcredits;
2451 if (ni->ni_interfaces[0] != NULL) {
2452 /* Use the IPoIB interface specified in 'networks=' */
2454 CLASSERT (LNET_MAX_INTERFACES > 1);
2455 if (ni->ni_interfaces[1] != NULL) {
2456 CERROR("Multiple interfaces not supported\n");
2460 ifname = ni->ni_interfaces[0];
2462 ifname = *kiblnd_tunables.kib_default_ipif;
2465 if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
2466 CERROR("IPoIB interface name too long: %s\n", ifname);
2470 list_for_each (tmp, &kiblnd_data.kib_devs) {
2471 ibdev = list_entry(tmp, kib_dev_t, ibd_list);
2473 if (!strcmp(&ibdev->ibd_ifname[0], ifname))
2479 if (ibdev == NULL) {
2483 struct rdma_cm_id *id;
2485 struct sockaddr_in addr;
2487 rc = libcfs_ipif_query(ifname, &up, &ip, &netmask);
2489 CERROR("Can't query IPoIB interface %s: %d\n",
2495 CERROR("Can't query IPoIB interface %s: it's down\n",
2500 LIBCFS_ALLOC(ibdev, sizeof(*ibdev));
2504 memset(ibdev, 0, sizeof(*ibdev));
2506 CFS_INIT_LIST_HEAD(&ibdev->ibd_list); /* not yet in kib_devs */
2507 ibdev->ibd_ifip = ip;
2508 strcpy(&ibdev->ibd_ifname[0], ifname);
2510 id = rdma_create_id(kiblnd_cm_callback, ibdev, RDMA_PS_TCP);
2512 CERROR("Can't create listen ID: %ld\n", PTR_ERR(id));
2516 ibdev->ibd_cmid = id;
2518 memset(&addr, 0, sizeof(addr));
2519 addr.sin_family = AF_INET;
2520 addr.sin_port = htons(*kiblnd_tunables.kib_service);
2521 addr.sin_addr.s_addr = htonl(ip);
2523 rc = rdma_bind_addr(id, (struct sockaddr *)&addr);
2525 CERROR("Can't bind to %s: %d\n", ifname, rc);
2529 /* Binding should have assigned me an IB device */
2530 LASSERT (id->device != NULL);
2531 CDEBUG(D_CONSOLE, "Listener bound to %s:%u.%u.%u.%u:%d:%s\n",
2532 ifname, HIPQUAD(ip), *kiblnd_tunables.kib_service,
2535 pd = ib_alloc_pd(id->device);
2537 CERROR("Can't allocate PD: %ld\n", PTR_ERR(pd));
2543 rc = rdma_listen(id, 256);
2545 CERROR("Can't start listener: %d\n", rc);
2549 rc = kiblnd_dev_setup(ibdev);
2551 CERROR("Can't setup device: %d\n", rc);
2556 ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
2557 net->ibn_dev = ibdev;
2559 rc = kiblnd_net_init_pools(net);
2561 CERROR("Failed to initialize NI pools: %d\n", rc);
2565 net->ibn_init = IBLND_INIT_ALL;
2570 if (net->ibn_dev == NULL && ibdev != NULL)
2571 kiblnd_destroy_dev(ibdev);
2573 kiblnd_shutdown(ni);
2575 CDEBUG(D_NET, "kiblnd_startup failed\n");
2580 kiblnd_module_fini (void)
2582 lnet_unregister_lnd(&the_o2iblnd);
2583 kiblnd_tunables_fini();
2587 kiblnd_module_init (void)
2591 CLASSERT (sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
2592 CLASSERT (offsetof(kib_msg_t, ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
2594 CLASSERT (offsetof(kib_msg_t, ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
2597 rc = kiblnd_tunables_init();
2601 lnet_register_lnd(&the_o2iblnd);
2606 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
2607 MODULE_DESCRIPTION("Kernel OpenIB gen2 LND v2.00");
2608 MODULE_LICENSE("GPL");
2610 module_init(kiblnd_module_init);
2611 module_exit(kiblnd_module_fini);