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 2008 Sun Microsystems, Inc. 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 = ni->ni_nid;
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 cfs_atomic_set(&peer->ibp_refcount, 1); /* 1 ref for caller */
350 CFS_INIT_LIST_HEAD(&peer->ibp_list); /* not in the peer table yet */
351 CFS_INIT_LIST_HEAD(&peer->ibp_conns);
352 CFS_INIT_LIST_HEAD(&peer->ibp_tx_queue);
354 cfs_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 cfs_atomic_inc(&net->ibn_npeers);
362 cfs_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 (cfs_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 (cfs_list_empty(&peer->ibp_conns));
379 LASSERT (cfs_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 cfs_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 cfs_list_t *peer_list = kiblnd_nid2peerlist(nid);
399 cfs_list_for_each (tmp, peer_list) {
401 peer = cfs_list_entry(tmp, kib_peer_t, ibp_list);
403 LASSERT (peer->ibp_connecting > 0 || /* creating conns */
404 peer->ibp_accepting > 0 ||
405 !cfs_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 cfs_atomic_read(&peer->ibp_refcount),
420 kiblnd_unlink_peer_locked (kib_peer_t *peer)
422 LASSERT (cfs_list_empty(&peer->ibp_conns));
424 LASSERT (kiblnd_peer_active(peer));
425 cfs_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)
439 cfs_read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
441 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
443 cfs_list_for_each (ptmp, &kiblnd_data.kib_peers[i]) {
445 peer = cfs_list_entry(ptmp, kib_peer_t, ibp_list);
446 LASSERT (peer->ibp_connecting > 0 ||
447 peer->ibp_accepting > 0 ||
448 !cfs_list_empty(&peer->ibp_conns));
450 if (peer->ibp_ni != ni)
456 *nidp = peer->ibp_nid;
457 *count = cfs_atomic_read(&peer->ibp_refcount);
459 cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
465 cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
470 kiblnd_del_peer_locked (kib_peer_t *peer)
476 if (cfs_list_empty(&peer->ibp_conns)) {
477 kiblnd_unlink_peer_locked(peer);
479 cfs_list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
480 conn = cfs_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);
503 cfs_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 cfs_list_for_each_safe (ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
514 peer = cfs_list_entry(ptmp, kib_peer_t, ibp_list);
515 LASSERT (peer->ibp_connecting > 0 ||
516 peer->ibp_accepting > 0 ||
517 !cfs_list_empty(&peer->ibp_conns));
519 if (peer->ibp_ni != ni)
522 if (!(nid == LNET_NID_ANY || peer->ibp_nid == nid))
525 if (!cfs_list_empty(&peer->ibp_tx_queue)) {
526 LASSERT (cfs_list_empty(&peer->ibp_conns));
528 cfs_list_splice_init(&peer->ibp_tx_queue,
532 kiblnd_del_peer_locked(peer);
533 rc = 0; /* matched something */
537 cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
539 kiblnd_txlist_done(ni, &zombies, -EIO);
545 kiblnd_get_conn_by_idx (lnet_ni_t *ni, int index)
554 cfs_read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
556 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
557 cfs_list_for_each (ptmp, &kiblnd_data.kib_peers[i]) {
559 peer = cfs_list_entry(ptmp, kib_peer_t, ibp_list);
560 LASSERT (peer->ibp_connecting > 0 ||
561 peer->ibp_accepting > 0 ||
562 !cfs_list_empty(&peer->ibp_conns));
564 if (peer->ibp_ni != ni)
567 cfs_list_for_each (ctmp, &peer->ibp_conns) {
571 conn = cfs_list_entry(ctmp, kib_conn_t,
573 kiblnd_conn_addref(conn);
574 cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
581 cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
586 kiblnd_debug_rx (kib_rx_t *rx)
588 CDEBUG(D_CONSOLE, " %p status %d msg_type %x cred %d\n",
589 rx, rx->rx_status, rx->rx_msg->ibm_type,
590 rx->rx_msg->ibm_credits);
594 kiblnd_debug_tx (kib_tx_t *tx)
596 CDEBUG(D_CONSOLE, " %p snd %d q %d w %d rc %d dl %lx "
597 "cookie "LPX64" msg %s%s type %x cred %d\n",
598 tx, tx->tx_sending, tx->tx_queued, tx->tx_waiting,
599 tx->tx_status, tx->tx_deadline, tx->tx_cookie,
600 tx->tx_lntmsg[0] == NULL ? "-" : "!",
601 tx->tx_lntmsg[1] == NULL ? "-" : "!",
602 tx->tx_msg->ibm_type, tx->tx_msg->ibm_credits);
606 kiblnd_debug_conn (kib_conn_t *conn)
611 cfs_spin_lock(&conn->ibc_lock);
613 CDEBUG(D_CONSOLE, "conn[%d] %p [version %x] -> %s: \n",
614 cfs_atomic_read(&conn->ibc_refcount), conn,
615 conn->ibc_version, libcfs_nid2str(conn->ibc_peer->ibp_nid));
616 CDEBUG(D_CONSOLE, " state %d nposted %d/%d cred %d o_cred %d r_cred %d\n",
617 conn->ibc_state, conn->ibc_noops_posted,
618 conn->ibc_nsends_posted, conn->ibc_credits,
619 conn->ibc_outstanding_credits, conn->ibc_reserved_credits);
620 CDEBUG(D_CONSOLE, " comms_err %d\n", conn->ibc_comms_error);
622 CDEBUG(D_CONSOLE, " early_rxs:\n");
623 cfs_list_for_each(tmp, &conn->ibc_early_rxs)
624 kiblnd_debug_rx(cfs_list_entry(tmp, kib_rx_t, rx_list));
626 CDEBUG(D_CONSOLE, " tx_queue_nocred:\n");
627 cfs_list_for_each(tmp, &conn->ibc_tx_queue_nocred)
628 kiblnd_debug_tx(cfs_list_entry(tmp, kib_tx_t, tx_list));
630 CDEBUG(D_CONSOLE, " tx_queue_rsrvd:\n");
631 cfs_list_for_each(tmp, &conn->ibc_tx_queue_rsrvd)
632 kiblnd_debug_tx(cfs_list_entry(tmp, kib_tx_t, tx_list));
634 CDEBUG(D_CONSOLE, " tx_queue:\n");
635 cfs_list_for_each(tmp, &conn->ibc_tx_queue)
636 kiblnd_debug_tx(cfs_list_entry(tmp, kib_tx_t, tx_list));
638 CDEBUG(D_CONSOLE, " active_txs:\n");
639 cfs_list_for_each(tmp, &conn->ibc_active_txs)
640 kiblnd_debug_tx(cfs_list_entry(tmp, kib_tx_t, tx_list));
642 CDEBUG(D_CONSOLE, " rxs:\n");
643 for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++)
644 kiblnd_debug_rx(&conn->ibc_rxs[i]);
646 cfs_spin_unlock(&conn->ibc_lock);
650 kiblnd_translate_mtu(int value)
671 kiblnd_setup_mtu(struct rdma_cm_id *cmid)
676 /* XXX There is no path record for iWARP, set by netdev->change_mtu? */
677 if (cmid->route.path_rec == NULL)
680 cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
682 mtu = kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu);
685 cmid->route.path_rec->mtu = mtu;
687 cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
691 kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
692 int state, int version)
695 * If the new conn is created successfully it takes over the caller's
696 * ref on 'peer'. It also "owns" 'cmid' and destroys it when it itself
697 * is destroyed. On failure, the caller's ref on 'peer' remains and
698 * she must dispose of 'cmid'. (Actually I'd block forever if I tried
699 * to destroy 'cmid' here since I'm called from the CM which still has
700 * its ref on 'cmid'). */
701 kib_net_t *net = peer->ibp_ni->ni_data;
702 struct ib_qp_init_attr *init_qp_attr;
709 LASSERT (net != NULL);
710 LASSERT (!cfs_in_interrupt());
712 LIBCFS_ALLOC(init_qp_attr, sizeof(*init_qp_attr));
713 if (init_qp_attr == NULL) {
714 CERROR("Can't allocate qp_attr for %s\n",
715 libcfs_nid2str(peer->ibp_nid));
719 LIBCFS_ALLOC(conn, sizeof(*conn));
721 CERROR("Can't allocate connection for %s\n",
722 libcfs_nid2str(peer->ibp_nid));
726 memset(conn, 0, sizeof(*conn)); /* zero flags, NULL pointers etc... */
728 conn->ibc_state = IBLND_CONN_INIT;
729 conn->ibc_version = version;
730 conn->ibc_peer = peer; /* I take the caller's ref */
731 cmid->context = conn; /* for future CM callbacks */
732 conn->ibc_cmid = cmid;
734 CFS_INIT_LIST_HEAD(&conn->ibc_early_rxs);
735 CFS_INIT_LIST_HEAD(&conn->ibc_tx_queue);
736 CFS_INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
737 CFS_INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
738 CFS_INIT_LIST_HEAD(&conn->ibc_active_txs);
739 cfs_spin_lock_init(&conn->ibc_lock);
741 LIBCFS_ALLOC(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
742 if (conn->ibc_connvars == NULL) {
743 CERROR("Can't allocate in-progress connection state\n");
746 memset(conn->ibc_connvars, 0, sizeof(*conn->ibc_connvars));
748 LIBCFS_ALLOC(conn->ibc_rxs, IBLND_RX_MSGS(version) * sizeof(kib_rx_t));
749 if (conn->ibc_rxs == NULL) {
750 CERROR("Cannot allocate RX buffers\n");
753 memset(conn->ibc_rxs, 0, IBLND_RX_MSGS(version) * sizeof(kib_rx_t));
755 rc = kiblnd_alloc_pages(&conn->ibc_rx_pages,
756 IBLND_RX_MSG_PAGES(version));
760 kiblnd_map_rx_descs(conn);
762 #ifdef HAVE_OFED_IB_COMP_VECTOR
763 cq = ib_create_cq(cmid->device,
764 kiblnd_cq_completion, kiblnd_cq_event, conn,
765 IBLND_CQ_ENTRIES(version), 0);
767 cq = ib_create_cq(cmid->device,
768 kiblnd_cq_completion, kiblnd_cq_event, conn,
769 IBLND_CQ_ENTRIES(version));
772 CERROR("Can't create CQ: %ld, cqe: %d\n",
773 PTR_ERR(cq), IBLND_CQ_ENTRIES(version));
779 rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
781 CERROR("Can't request completion notificiation: %d\n", rc);
785 kiblnd_setup_mtu(cmid);
787 memset(init_qp_attr, 0, sizeof(*init_qp_attr));
788 init_qp_attr->event_handler = kiblnd_qp_event;
789 init_qp_attr->qp_context = conn;
790 init_qp_attr->cap.max_send_wr = IBLND_SEND_WRS(version);
791 init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(version);
792 init_qp_attr->cap.max_send_sge = 1;
793 init_qp_attr->cap.max_recv_sge = 1;
794 init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
795 init_qp_attr->qp_type = IB_QPT_RC;
796 init_qp_attr->send_cq = cq;
797 init_qp_attr->recv_cq = cq;
799 rc = rdma_create_qp(cmid, net->ibn_dev->ibd_pd, init_qp_attr);
801 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d\n",
802 rc, init_qp_attr->cap.max_send_wr,
803 init_qp_attr->cap.max_recv_wr);
807 LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
809 /* 1 ref for caller and each rxmsg */
810 cfs_atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(version));
811 conn->ibc_nrx = IBLND_RX_MSGS(version);
814 for (i = 0; i < IBLND_RX_MSGS(version); i++) {
815 rc = kiblnd_post_rx(&conn->ibc_rxs[i],
816 IBLND_POSTRX_NO_CREDIT);
818 CERROR("Can't post rxmsg: %d\n", rc);
820 /* Make posted receives complete */
821 kiblnd_abort_receives(conn);
823 /* correct # of posted buffers
824 * NB locking needed now I'm racing with completion */
825 cfs_spin_lock_irqsave(&kiblnd_data.kib_sched_lock,
827 conn->ibc_nrx -= IBLND_RX_MSGS(version) - i;
828 cfs_spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock,
831 /* cmid will be destroyed by CM(ofed) after cm_callback
832 * returned, so we can't refer it anymore
833 * (by kiblnd_connd()->kiblnd_destroy_conn) */
834 rdma_destroy_qp(conn->ibc_cmid);
835 conn->ibc_cmid = NULL;
837 /* Drop my own and unused rxbuffer refcounts */
838 while (i++ <= IBLND_RX_MSGS(version))
839 kiblnd_conn_decref(conn);
845 /* Init successful! */
846 LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
847 state == IBLND_CONN_PASSIVE_WAIT);
848 conn->ibc_state = state;
851 cfs_atomic_inc(&net->ibn_nconns);
855 kiblnd_destroy_conn(conn);
857 LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
863 kiblnd_destroy_conn (kib_conn_t *conn)
865 struct rdma_cm_id *cmid = conn->ibc_cmid;
866 kib_peer_t *peer = conn->ibc_peer;
869 LASSERT (!cfs_in_interrupt());
870 LASSERT (cfs_atomic_read(&conn->ibc_refcount) == 0);
871 LASSERT (cfs_list_empty(&conn->ibc_early_rxs));
872 LASSERT (cfs_list_empty(&conn->ibc_tx_queue));
873 LASSERT (cfs_list_empty(&conn->ibc_tx_queue_rsrvd));
874 LASSERT (cfs_list_empty(&conn->ibc_tx_queue_nocred));
875 LASSERT (cfs_list_empty(&conn->ibc_active_txs));
876 LASSERT (conn->ibc_noops_posted == 0);
877 LASSERT (conn->ibc_nsends_posted == 0);
879 switch (conn->ibc_state) {
881 /* conn must be completely disengaged from the network */
884 case IBLND_CONN_DISCONNECTED:
885 /* connvars should have been freed already */
886 LASSERT (conn->ibc_connvars == NULL);
889 case IBLND_CONN_INIT:
893 /* conn->ibc_cmid might be destroyed by CM already */
894 if (cmid != NULL && cmid->qp != NULL)
895 rdma_destroy_qp(cmid);
897 if (conn->ibc_cq != NULL) {
898 rc = ib_destroy_cq(conn->ibc_cq);
900 CWARN("Error destroying CQ: %d\n", rc);
903 if (conn->ibc_rx_pages != NULL)
904 kiblnd_unmap_rx_descs(conn);
906 if (conn->ibc_rxs != NULL) {
907 LIBCFS_FREE(conn->ibc_rxs,
908 IBLND_RX_MSGS(conn->ibc_version) * sizeof(kib_rx_t));
911 if (conn->ibc_connvars != NULL)
912 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
914 /* See CAVEAT EMPTOR above in kiblnd_create_conn */
915 if (conn->ibc_state != IBLND_CONN_INIT) {
916 kib_net_t *net = peer->ibp_ni->ni_data;
918 kiblnd_peer_decref(peer);
919 rdma_destroy_id(cmid);
920 cfs_atomic_dec(&net->ibn_nconns);
923 LIBCFS_FREE(conn, sizeof(*conn));
927 kiblnd_close_peer_conns_locked (kib_peer_t *peer, int why)
934 cfs_list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
935 conn = cfs_list_entry(ctmp, kib_conn_t, ibc_list);
937 CDEBUG(D_NET, "Closing conn -> %s, "
938 "version: %x, reason: %d\n",
939 libcfs_nid2str(peer->ibp_nid),
940 conn->ibc_version, why);
942 kiblnd_close_conn_locked(conn, why);
950 kiblnd_close_stale_conns_locked (kib_peer_t *peer,
951 int version, __u64 incarnation)
958 cfs_list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
959 conn = cfs_list_entry(ctmp, kib_conn_t, ibc_list);
961 if (conn->ibc_version == version &&
962 conn->ibc_incarnation == incarnation)
965 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
966 "incarnation:"LPX64"(%x, "LPX64")\n",
967 libcfs_nid2str(peer->ibp_nid),
968 conn->ibc_version, conn->ibc_incarnation,
969 version, incarnation);
971 kiblnd_close_conn_locked(conn, -ESTALE);
979 kiblnd_close_matching_conns (lnet_ni_t *ni, lnet_nid_t nid)
990 cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
992 if (nid != LNET_NID_ANY)
993 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
996 hi = kiblnd_data.kib_peer_hash_size - 1;
999 for (i = lo; i <= hi; i++) {
1000 cfs_list_for_each_safe (ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
1002 peer = cfs_list_entry(ptmp, kib_peer_t, ibp_list);
1003 LASSERT (peer->ibp_connecting > 0 ||
1004 peer->ibp_accepting > 0 ||
1005 !cfs_list_empty(&peer->ibp_conns));
1007 if (peer->ibp_ni != ni)
1010 if (!(nid == LNET_NID_ANY || nid == peer->ibp_nid))
1013 count += kiblnd_close_peer_conns_locked(peer, 0);
1017 cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1019 /* wildcards always succeed */
1020 if (nid == LNET_NID_ANY)
1023 return (count == 0) ? -ENOENT : 0;
1027 kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
1029 struct libcfs_ioctl_data *data = arg;
1033 case IOC_LIBCFS_GET_PEER: {
1037 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1039 data->ioc_nid = nid;
1040 data->ioc_count = count;
1044 case IOC_LIBCFS_DEL_PEER: {
1045 rc = kiblnd_del_peer(ni, data->ioc_nid);
1048 case IOC_LIBCFS_GET_CONN: {
1052 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1058 LASSERT (conn->ibc_cmid != NULL);
1059 data->ioc_nid = conn->ibc_peer->ibp_nid;
1060 if (conn->ibc_cmid->route.path_rec == NULL)
1061 data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1064 ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1065 kiblnd_conn_decref(conn);
1068 case IOC_LIBCFS_CLOSE_CONNECTION: {
1069 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1081 kiblnd_query (lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when)
1083 cfs_time_t last_alive = 0;
1084 cfs_rwlock_t *glock = &kiblnd_data.kib_global_lock;
1086 unsigned long flags;
1088 cfs_read_lock_irqsave(glock, flags);
1090 peer = kiblnd_find_peer_locked(nid);
1092 LASSERT (peer->ibp_connecting > 0 || /* creating conns */
1093 peer->ibp_accepting > 0 ||
1094 !cfs_list_empty(&peer->ibp_conns)); /* active conn */
1095 last_alive = peer->ibp_last_alive;
1098 cfs_read_unlock_irqrestore(glock, flags);
1100 if (last_alive != 0)
1103 /* peer is not persistent in hash, trigger peer creation
1104 * and connection establishment with a NULL tx */
1106 kiblnd_launch_tx(ni, NULL, nid);
1111 kiblnd_free_pages (kib_pages_t *p)
1113 int npages = p->ibp_npages;
1116 LASSERT (p->ibp_device == NULL);
1118 for (i = 0; i < npages; i++)
1119 if (p->ibp_pages[i] != NULL)
1120 __free_page(p->ibp_pages[i]);
1122 LIBCFS_FREE (p, offsetof(kib_pages_t, ibp_pages[npages]));
1126 kiblnd_alloc_pages (kib_pages_t **pp, int npages)
1131 LIBCFS_ALLOC(p, offsetof(kib_pages_t, ibp_pages[npages]));
1133 CERROR("Can't allocate descriptor for %d pages\n", npages);
1137 memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1138 p->ibp_npages = npages;
1140 for (i = 0; i < npages; i++) {
1141 p->ibp_pages[i] = alloc_page(GFP_KERNEL);
1142 if (p->ibp_pages[i] == NULL) {
1143 CERROR("Can't allocate page %d of %d\n", i, npages);
1144 kiblnd_free_pages(p);
1154 kiblnd_unmap_rx_descs(kib_conn_t *conn)
1159 LASSERT (conn->ibc_rxs != NULL);
1160 LASSERT (conn->ibc_rx_pages->ibp_device != NULL);
1162 for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1163 rx = &conn->ibc_rxs[i];
1165 LASSERT (rx->rx_nob >= 0); /* not posted */
1167 kiblnd_dma_unmap_single(conn->ibc_rx_pages->ibp_device,
1168 KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1170 IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1173 conn->ibc_rx_pages->ibp_device = NULL;
1175 kiblnd_free_pages(conn->ibc_rx_pages);
1177 conn->ibc_rx_pages = NULL;
1181 kiblnd_map_rx_descs(kib_conn_t *conn)
1189 for (pg_off = ipg = i = 0;
1190 i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1191 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1192 rx = &conn->ibc_rxs[i];
1195 rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
1197 rx->rx_msgaddr = kiblnd_dma_map_single(conn->ibc_cmid->device,
1198 rx->rx_msg, IBLND_MSG_SIZE,
1200 LASSERT (!kiblnd_dma_mapping_error(conn->ibc_cmid->device,
1202 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1204 CDEBUG(D_NET,"rx %d: %p "LPX64"("LPX64")\n",
1205 i, rx->rx_msg, rx->rx_msgaddr,
1206 lnet_page2phys(pg) + pg_off);
1208 pg_off += IBLND_MSG_SIZE;
1209 LASSERT (pg_off <= PAGE_SIZE);
1211 if (pg_off == PAGE_SIZE) {
1214 LASSERT (ipg <= IBLND_RX_MSG_PAGES(conn->ibc_version));
1218 conn->ibc_rx_pages->ibp_device = conn->ibc_cmid->device;
1222 kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo)
1227 LASSERT (tpo->tpo_pool.po_allocated == 0);
1228 LASSERT (tpo->tpo_tx_pages->ibp_device != NULL);
1230 for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1231 tx = &tpo->tpo_tx_descs[i];
1232 kiblnd_dma_unmap_single(tpo->tpo_tx_pages->ibp_device,
1233 KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1235 IBLND_MSG_SIZE, DMA_TO_DEVICE);
1237 tpo->tpo_tx_pages->ibp_device = NULL;
1241 kiblnd_map_tx_pool(kib_tx_pool_t *tpo)
1243 kib_pages_t *txpgs = tpo->tpo_tx_pages;
1244 kib_pool_t *pool = &tpo->tpo_pool;
1245 kib_net_t *net = pool->po_owner->ps_net;
1252 LASSERT (net != NULL);
1253 LASSERT (net->ibn_dev->ibd_cmid != NULL &&
1254 net->ibn_dev->ibd_cmid->device != NULL);
1256 /* pre-mapped messages are not bigger than 1 page */
1257 CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE);
1259 /* No fancy arithmetic when we do the buffer calculations */
1260 CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
1262 txpgs->ibp_device = net->ibn_dev->ibd_cmid->device;
1264 for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1265 page = txpgs->ibp_pages[ipage];
1266 tx = &tpo->tpo_tx_descs[i];
1268 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1271 tx->tx_msgaddr = kiblnd_dma_map_single(
1272 txpgs->ibp_device, tx->tx_msg,
1273 IBLND_MSG_SIZE, DMA_TO_DEVICE);
1274 LASSERT (!kiblnd_dma_mapping_error(txpgs->ibp_device,
1276 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1278 cfs_list_add(&tx->tx_list, &pool->po_free_list);
1280 page_offset += IBLND_MSG_SIZE;
1281 LASSERT (page_offset <= PAGE_SIZE);
1283 if (page_offset == PAGE_SIZE) {
1286 LASSERT (ipage <= txpgs->ibp_npages);
1292 kiblnd_find_dma_mr(kib_net_t *net, __u64 addr, __u64 size)
1296 LASSERT (net->ibn_dev->ibd_mrs[0] != NULL);
1298 if (net->ibn_dev->ibd_nmrs == 1)
1299 return net->ibn_dev->ibd_mrs[0];
1301 index = addr >> net->ibn_dev->ibd_mr_shift;
1303 if (index < net->ibn_dev->ibd_nmrs &&
1304 index == ((addr + size - 1) >> net->ibn_dev->ibd_mr_shift))
1305 return net->ibn_dev->ibd_mrs[index];
1311 kiblnd_find_rd_dma_mr(kib_net_t *net, kib_rdma_desc_t *rd)
1313 struct ib_mr *prev_mr;
1317 LASSERT (net->ibn_dev->ibd_mrs[0] != NULL);
1319 if (*kiblnd_tunables.kib_map_on_demand > 0 &&
1320 *kiblnd_tunables.kib_map_on_demand <= rd->rd_nfrags)
1323 if (net->ibn_dev->ibd_nmrs == 1)
1324 return net->ibn_dev->ibd_mrs[0];
1326 for (i = 0, mr = prev_mr = NULL;
1327 i < rd->rd_nfrags; i++) {
1328 mr = kiblnd_find_dma_mr(net,
1329 rd->rd_frags[i].rf_addr,
1330 rd->rd_frags[i].rf_nob);
1331 if (prev_mr == NULL)
1334 if (mr == NULL || prev_mr != mr) {
1335 /* Can't covered by one single MR */
1345 kiblnd_destroy_fmr_pool(kib_fmr_pool_t *pool)
1347 LASSERT (pool->fpo_map_count == 0);
1349 if (pool->fpo_fmr_pool != NULL)
1350 ib_destroy_fmr_pool(pool->fpo_fmr_pool);
1352 LIBCFS_FREE(pool, sizeof(kib_fmr_pool_t));
1356 kiblnd_destroy_fmr_pool_list(cfs_list_t *head)
1358 kib_fmr_pool_t *pool;
1360 while (!cfs_list_empty(head)) {
1361 pool = cfs_list_entry(head->next, kib_fmr_pool_t, fpo_list);
1362 cfs_list_del(&pool->fpo_list);
1363 kiblnd_destroy_fmr_pool(pool);
1368 kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo)
1370 /* FMR pool for RDMA */
1371 kib_fmr_pool_t *fpo;
1372 struct ib_fmr_pool_param param = {
1373 .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
1374 .page_shift = PAGE_SHIFT,
1375 .access = (IB_ACCESS_LOCAL_WRITE |
1376 IB_ACCESS_REMOTE_WRITE),
1377 .pool_size = *kiblnd_tunables.kib_fmr_pool_size,
1378 .dirty_watermark = *kiblnd_tunables.kib_fmr_flush_trigger,
1379 .flush_function = NULL,
1381 .cache = !!*kiblnd_tunables.kib_fmr_cache};
1384 LASSERT (fps->fps_net->ibn_dev != NULL &&
1385 fps->fps_net->ibn_dev->ibd_pd != NULL);
1387 LIBCFS_ALLOC(fpo, sizeof(kib_fmr_pool_t));
1391 memset(fpo, 0, sizeof(kib_fmr_pool_t));
1392 fpo->fpo_fmr_pool = ib_create_fmr_pool(fps->fps_net->ibn_dev->ibd_pd, ¶m);
1393 if (IS_ERR(fpo->fpo_fmr_pool)) {
1394 CERROR("Failed to create FMR pool: %ld\n",
1395 PTR_ERR(fpo->fpo_fmr_pool));
1396 rc = PTR_ERR(fpo->fpo_fmr_pool);
1397 LIBCFS_FREE(fpo, sizeof(kib_fmr_pool_t));
1401 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1402 fpo->fpo_owner = fps;
1409 kiblnd_fini_fmr_pool_set(kib_fmr_poolset_t *fps)
1411 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1415 kiblnd_init_fmr_pool_set(kib_fmr_poolset_t *fps, kib_net_t *net)
1417 kib_fmr_pool_t *fpo;
1420 memset(fps, 0, sizeof(kib_fmr_poolset_t));
1423 cfs_spin_lock_init(&fps->fps_lock);
1424 CFS_INIT_LIST_HEAD(&fps->fps_pool_list);
1425 rc = kiblnd_create_fmr_pool(fps, &fpo);
1427 cfs_list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1433 kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1435 CFS_LIST_HEAD (zombies);
1436 kib_fmr_pool_t *fpo = fmr->fmr_pool;
1437 kib_fmr_poolset_t *fps = fpo->fpo_owner;
1438 kib_fmr_pool_t *tmp;
1441 rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1445 rc = ib_flush_fmr_pool(fpo->fpo_fmr_pool);
1449 fmr->fmr_pool = NULL;
1450 fmr->fmr_pfmr = NULL;
1452 cfs_spin_lock(&fps->fps_lock);
1453 fpo->fpo_map_count --; /* decref the pool */
1455 cfs_list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1456 /* the first pool is persistent */
1457 if (fps->fps_pool_list.next == &fpo->fpo_list)
1460 if (fpo->fpo_map_count == 0 && /* no more reference */
1461 cfs_time_aftereq(cfs_time_current(), fpo->fpo_deadline)) {
1462 cfs_list_move(&fpo->fpo_list, &zombies);
1463 fps->fps_version ++;
1466 cfs_spin_unlock(&fps->fps_lock);
1468 if (!cfs_list_empty(&zombies))
1469 kiblnd_destroy_fmr_pool_list(&zombies);
1473 kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages,
1474 __u64 iov, kib_fmr_t *fmr)
1476 struct ib_pool_fmr *pfmr;
1477 kib_fmr_pool_t *fpo;
1481 LASSERT (fps->fps_net->ibn_with_fmr);
1483 cfs_spin_lock(&fps->fps_lock);
1484 version = fps->fps_version;
1485 cfs_list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1486 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1487 fpo->fpo_map_count ++;
1488 cfs_spin_unlock(&fps->fps_lock);
1490 pfmr = ib_fmr_pool_map_phys(fpo->fpo_fmr_pool,
1491 pages, npages, iov);
1492 if (likely(!IS_ERR(pfmr))) {
1493 fmr->fmr_pool = fpo;
1494 fmr->fmr_pfmr = pfmr;
1498 cfs_spin_lock(&fps->fps_lock);
1499 fpo->fpo_map_count --;
1500 if (PTR_ERR(pfmr) != -EAGAIN) {
1501 cfs_spin_unlock(&fps->fps_lock);
1502 return PTR_ERR(pfmr);
1505 /* EAGAIN and ... */
1506 if (version != fps->fps_version) {
1507 cfs_spin_unlock(&fps->fps_lock);
1512 if (fps->fps_increasing) {
1513 cfs_spin_unlock(&fps->fps_lock);
1514 CDEBUG(D_NET, "Another thread is allocating new "
1515 "FMR pool, waiting for her to complete\n");
1521 if (cfs_time_before(cfs_time_current(), fps->fps_next_retry)) {
1522 /* someone failed recently */
1523 cfs_spin_unlock(&fps->fps_lock);
1527 fps->fps_increasing = 1;
1528 cfs_spin_unlock(&fps->fps_lock);
1530 CDEBUG(D_NET, "Allocate new FMR pool\n");
1531 rc = kiblnd_create_fmr_pool(fps, &fpo);
1532 cfs_spin_lock(&fps->fps_lock);
1533 fps->fps_increasing = 0;
1535 fps->fps_version ++;
1536 cfs_list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1538 fps->fps_next_retry = cfs_time_shift(10);
1540 cfs_spin_unlock(&fps->fps_lock);
1546 kiblnd_fini_pool(kib_pool_t *pool)
1548 LASSERT (cfs_list_empty(&pool->po_free_list));
1549 LASSERT (pool->po_allocated == 0);
1551 CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
1555 kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
1557 CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
1559 memset(pool, 0, sizeof(kib_pool_t));
1560 CFS_INIT_LIST_HEAD(&pool->po_free_list);
1561 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1562 pool->po_owner = ps;
1563 pool->po_size = size;
1567 kiblnd_destroy_pool_list(kib_poolset_t *ps, cfs_list_t *head)
1571 while (!cfs_list_empty(head)) {
1572 pool = cfs_list_entry(head->next, kib_pool_t, po_list);
1573 cfs_list_del(&pool->po_list);
1574 ps->ps_pool_destroy(pool);
1579 kiblnd_fini_pool_set(kib_poolset_t *ps)
1581 kiblnd_destroy_pool_list(ps, &ps->ps_pool_list);
1585 kiblnd_init_pool_set(kib_poolset_t *ps, kib_net_t *net,
1586 char *name, int size,
1587 kib_ps_pool_create_t po_create,
1588 kib_ps_pool_destroy_t po_destroy,
1589 kib_ps_node_init_t nd_init,
1590 kib_ps_node_fini_t nd_fini)
1595 memset(ps, 0, sizeof(kib_poolset_t));
1598 ps->ps_pool_create = po_create;
1599 ps->ps_pool_destroy = po_destroy;
1600 ps->ps_node_init = nd_init;
1601 ps->ps_node_fini = nd_fini;
1602 ps->ps_pool_size = size;
1603 strncpy(ps->ps_name, name, IBLND_POOL_NAME_LEN);
1604 cfs_spin_lock_init(&ps->ps_lock);
1605 CFS_INIT_LIST_HEAD(&ps->ps_pool_list);
1607 rc = ps->ps_pool_create(ps, size, &pool);
1609 cfs_list_add(&pool->po_list, &ps->ps_pool_list);
1611 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
1617 kiblnd_pool_free_node(kib_pool_t *pool, cfs_list_t *node)
1619 CFS_LIST_HEAD (zombies);
1620 kib_poolset_t *ps = pool->po_owner;
1622 cfs_time_t now = cfs_time_current();
1624 cfs_spin_lock(&ps->ps_lock);
1626 if (ps->ps_node_fini != NULL)
1627 ps->ps_node_fini(pool, node);
1629 LASSERT (pool->po_allocated > 0);
1630 cfs_list_add(node, &pool->po_free_list);
1631 pool->po_allocated --;
1633 cfs_list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
1634 /* the first pool is persistent */
1635 if (ps->ps_pool_list.next == &pool->po_list)
1638 if (pool->po_allocated == 0 &&
1639 cfs_time_aftereq(now, pool->po_deadline))
1640 cfs_list_move(&pool->po_list, &zombies);
1642 cfs_spin_unlock(&ps->ps_lock);
1644 if (!cfs_list_empty(&zombies))
1645 kiblnd_destroy_pool_list(ps, &zombies);
1649 kiblnd_pool_alloc_node(kib_poolset_t *ps)
1656 cfs_spin_lock(&ps->ps_lock);
1657 cfs_list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
1658 if (cfs_list_empty(&pool->po_free_list))
1661 pool->po_allocated ++;
1662 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1663 node = pool->po_free_list.next;
1666 if (ps->ps_node_init != NULL) {
1667 /* still hold the lock */
1668 ps->ps_node_init(pool, node);
1670 cfs_spin_unlock(&ps->ps_lock);
1674 /* no available tx pool and ... */
1675 if (ps->ps_increasing) {
1676 /* another thread is allocating a new pool */
1677 cfs_spin_unlock(&ps->ps_lock);
1678 CDEBUG(D_NET, "Another thread is allocating new "
1679 "%s pool, waiting for her to complete\n",
1685 if (cfs_time_before(cfs_time_current(), ps->ps_next_retry)) {
1686 /* someone failed recently */
1687 cfs_spin_unlock(&ps->ps_lock);
1691 ps->ps_increasing = 1;
1692 cfs_spin_unlock(&ps->ps_lock);
1694 CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
1696 rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
1698 cfs_spin_lock(&ps->ps_lock);
1699 ps->ps_increasing = 0;
1701 cfs_list_add_tail(&pool->po_list, &ps->ps_pool_list);
1703 /* retry 10 seconds later */
1704 ps->ps_next_retry = cfs_time_shift(10);
1705 CERROR("Can't allocate new %s pool because out of memory\n",
1708 cfs_spin_unlock(&ps->ps_lock);
1714 kiblnd_pmr_pool_unmap(kib_phys_mr_t *pmr)
1716 kib_pmr_pool_t *ppo = pmr->pmr_pool;
1717 struct ib_mr *mr = pmr->pmr_mr;
1720 kiblnd_pool_free_node(&ppo->ppo_pool, &pmr->pmr_list);
1726 kiblnd_pmr_pool_map(kib_pmr_poolset_t *pps, kib_rdma_desc_t *rd,
1727 __u64 *iova, kib_phys_mr_t **pp_pmr)
1734 node = kiblnd_pool_alloc_node(&pps->pps_poolset);
1736 CERROR("Failed to allocate PMR descriptor\n");
1740 pmr = container_of(node, kib_phys_mr_t, pmr_list);
1741 for (i = 0; i < rd->rd_nfrags; i ++) {
1742 pmr->pmr_ipb[i].addr = rd->rd_frags[i].rf_addr;
1743 pmr->pmr_ipb[i].size = rd->rd_frags[i].rf_nob;
1746 pmr->pmr_mr = ib_reg_phys_mr(pps->pps_poolset.ps_net->ibn_dev->ibd_pd,
1747 pmr->pmr_ipb, rd->rd_nfrags,
1748 IB_ACCESS_LOCAL_WRITE |
1749 IB_ACCESS_REMOTE_WRITE,
1751 if (!IS_ERR(pmr->pmr_mr)) {
1752 pmr->pmr_iova = *iova;
1757 rc = PTR_ERR(pmr->pmr_mr);
1758 CERROR("Failed ib_reg_phys_mr: %d\n", rc);
1761 kiblnd_pool_free_node(&pmr->pmr_pool->ppo_pool, node);
1767 kiblnd_destroy_pmr_pool(kib_pool_t *pool)
1769 kib_pmr_pool_t *ppo = container_of(pool, kib_pmr_pool_t, ppo_pool);
1772 LASSERT (pool->po_allocated == 0);
1774 while (!cfs_list_empty(&pool->po_free_list)) {
1775 pmr = cfs_list_entry(pool->po_free_list.next,
1776 kib_phys_mr_t, pmr_list);
1778 LASSERT (pmr->pmr_mr == NULL);
1779 cfs_list_del(&pmr->pmr_list);
1781 if (pmr->pmr_ipb != NULL) {
1782 LIBCFS_FREE(pmr->pmr_ipb,
1783 IBLND_MAX_RDMA_FRAGS *
1784 sizeof(struct ib_phys_buf));
1787 LIBCFS_FREE(pmr, sizeof(kib_phys_mr_t));
1790 kiblnd_fini_pool(pool);
1791 LIBCFS_FREE(ppo, sizeof(kib_pmr_pool_t));
1795 kiblnd_create_pmr_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
1797 kib_pmr_pool_t *ppo;
1802 LIBCFS_ALLOC(ppo, sizeof(kib_pmr_pool_t));
1804 CERROR("Failed to allocate PMR pool\n");
1808 pool = &ppo->ppo_pool;
1809 kiblnd_init_pool(ps, pool, size);
1811 for (i = 0; i < size; i++) {
1812 LIBCFS_ALLOC(pmr, sizeof(kib_phys_mr_t));
1816 memset(pmr, 0, sizeof(kib_phys_mr_t));
1817 pmr->pmr_pool = ppo;
1818 LIBCFS_ALLOC(pmr->pmr_ipb,
1819 IBLND_MAX_RDMA_FRAGS *
1820 sizeof(struct ib_phys_buf));
1821 if (pmr->pmr_ipb == NULL)
1824 cfs_list_add(&pmr->pmr_list, &pool->po_free_list);
1828 ps->ps_pool_destroy(pool);
1837 kiblnd_destroy_tx_pool(kib_pool_t *pool)
1839 kib_tx_pool_t *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
1842 LASSERT (pool->po_allocated == 0);
1844 if (tpo->tpo_tx_pages != NULL) {
1845 if (tpo->tpo_tx_pages->ibp_device != NULL)
1846 kiblnd_unmap_tx_pool(tpo);
1847 kiblnd_free_pages(tpo->tpo_tx_pages);
1850 if (tpo->tpo_tx_descs == NULL)
1853 for (i = 0; i < pool->po_size; i++) {
1854 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
1856 cfs_list_del(&tx->tx_list);
1857 if (tx->tx_pages != NULL)
1858 LIBCFS_FREE(tx->tx_pages,
1860 sizeof(*tx->tx_pages));
1861 if (tx->tx_frags != NULL)
1862 LIBCFS_FREE(tx->tx_frags,
1863 IBLND_MAX_RDMA_FRAGS *
1864 sizeof(*tx->tx_frags));
1865 if (tx->tx_wrq != NULL)
1866 LIBCFS_FREE(tx->tx_wrq,
1867 (1 + IBLND_MAX_RDMA_FRAGS) *
1868 sizeof(*tx->tx_wrq));
1869 if (tx->tx_sge != NULL)
1870 LIBCFS_FREE(tx->tx_sge,
1871 (1 + IBLND_MAX_RDMA_FRAGS) *
1872 sizeof(*tx->tx_sge));
1873 if (tx->tx_rd != NULL)
1874 LIBCFS_FREE(tx->tx_rd,
1875 offsetof(kib_rdma_desc_t,
1876 rd_frags[IBLND_MAX_RDMA_FRAGS]));
1879 LIBCFS_FREE(tpo->tpo_tx_descs,
1880 pool->po_size * sizeof(kib_tx_t));
1882 kiblnd_fini_pool(pool);
1883 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
1887 kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
1894 LIBCFS_ALLOC(tpo, sizeof(kib_tx_pool_t));
1896 CERROR("Failed to allocate TX pool\n");
1900 pool = &tpo->tpo_pool;
1901 kiblnd_init_pool(ps, pool, size);
1902 tpo->tpo_tx_descs = NULL;
1903 tpo->tpo_tx_pages = NULL;
1905 npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
1906 if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, npg) != 0) {
1907 CERROR("Can't allocate tx pages: %d\n", npg);
1908 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
1912 LIBCFS_ALLOC (tpo->tpo_tx_descs, size * sizeof(kib_tx_t));
1913 if (tpo->tpo_tx_descs == NULL) {
1914 CERROR("Can't allocate %d tx descriptors\n", size);
1915 ps->ps_pool_destroy(pool);
1919 memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
1921 for (i = 0; i < size; i++) {
1922 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
1925 if (ps->ps_net->ibn_with_fmr){
1926 LIBCFS_ALLOC(tx->tx_pages, LNET_MAX_IOV *
1927 sizeof(*tx->tx_pages));
1928 if (tx->tx_pages == NULL)
1932 LIBCFS_ALLOC(tx->tx_frags,
1933 IBLND_MAX_RDMA_FRAGS *
1934 sizeof(*tx->tx_frags));
1935 if (tx->tx_frags == NULL)
1938 LIBCFS_ALLOC(tx->tx_wrq,
1939 (1 + IBLND_MAX_RDMA_FRAGS) *
1940 sizeof(*tx->tx_wrq));
1941 if (tx->tx_wrq == NULL)
1944 LIBCFS_ALLOC(tx->tx_sge,
1945 (1 + IBLND_MAX_RDMA_FRAGS) *
1946 sizeof(*tx->tx_sge));
1947 if (tx->tx_sge == NULL)
1950 LIBCFS_ALLOC(tx->tx_rd,
1951 offsetof(kib_rdma_desc_t,
1952 rd_frags[IBLND_MAX_RDMA_FRAGS]));
1953 if (tx->tx_rd == NULL)
1958 kiblnd_map_tx_pool(tpo);
1963 ps->ps_pool_destroy(pool);
1968 kiblnd_tx_init(kib_pool_t *pool, cfs_list_t *node)
1970 kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t,
1972 kib_tx_t *tx = cfs_list_entry(node, kib_tx_t, tx_list);
1974 tx->tx_cookie = tps->tps_next_tx_cookie ++;
1978 kiblnd_ni_fini_pools(kib_net_t *net)
1980 kiblnd_fini_pool_set(&net->ibn_tx_ps.tps_poolset);
1981 if (net->ibn_with_fmr)
1982 kiblnd_fini_fmr_pool_set(&net->ibn_fmr_ps);
1983 else if (net->ibn_with_pmr)
1984 kiblnd_fini_pool_set(&net->ibn_pmr_ps.pps_poolset);
1988 kiblnd_net_init_pools(kib_net_t *net)
1990 kib_fmr_poolset_t *fps = &net->ibn_fmr_ps;
1991 kib_pmr_poolset_t *pps = &net->ibn_pmr_ps;
1992 kib_tx_poolset_t *tps = &net->ibn_tx_ps;
1995 if (*kiblnd_tunables.kib_fmr_pool_size <
1996 *kiblnd_tunables.kib_ntx / 4) {
1997 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
1998 *kiblnd_tunables.kib_fmr_pool_size,
1999 *kiblnd_tunables.kib_ntx / 4);
2003 if (*kiblnd_tunables.kib_pmr_pool_size <
2004 *kiblnd_tunables.kib_ntx / 4) {
2005 CERROR("Can't set pmr pool size (%d) < ntx / 4(%d)\n",
2006 *kiblnd_tunables.kib_pmr_pool_size,
2007 *kiblnd_tunables.kib_ntx / 4);
2011 if (*kiblnd_tunables.kib_map_on_demand > 0 ||
2012 net->ibn_dev->ibd_nmrs > 1) { /* premapping can fail if ibd_nmr > 1,
2013 * so we always create FMR/PMR pool and
2014 * map-on-demand if premapping failed */
2015 rc = kiblnd_init_fmr_pool_set(fps, net);
2017 net->ibn_with_fmr = 1;
2018 } else if (rc == -ENOSYS) {
2019 rc = kiblnd_init_pool_set(&pps->pps_poolset, net, "PMR",
2020 *kiblnd_tunables.kib_pmr_pool_size,
2021 kiblnd_create_pmr_pool,
2022 kiblnd_destroy_pmr_pool,
2025 net->ibn_with_pmr = 1;
2031 rc = kiblnd_init_pool_set(&tps->tps_poolset, net, "TX", IBLND_TX_MSGS(),
2032 kiblnd_create_tx_pool, kiblnd_destroy_tx_pool,
2033 kiblnd_tx_init, NULL);
2037 if (net->ibn_with_fmr)
2038 kiblnd_fini_fmr_pool_set(fps);
2039 else if (net->ibn_with_pmr)
2040 kiblnd_fini_pool_set(&pps->pps_poolset);
2046 kiblnd_dev_cleanup(kib_dev_t *ibdev)
2050 if (ibdev->ibd_mrs == NULL)
2053 for (i = 0; i < ibdev->ibd_nmrs; i++) {
2054 if (ibdev->ibd_mrs[i] == NULL)
2057 ib_dereg_mr(ibdev->ibd_mrs[i]);
2060 LIBCFS_FREE(ibdev->ibd_mrs, sizeof(*ibdev->ibd_mrs) * ibdev->ibd_nmrs);
2061 ibdev->ibd_mrs = NULL;
2065 kiblnd_dev_get_attr(kib_dev_t *ibdev)
2067 struct ib_device_attr *attr;
2070 /* It's safe to assume a HCA can handle a page size
2071 * matching that of the native system */
2072 ibdev->ibd_page_shift = PAGE_SHIFT;
2073 ibdev->ibd_page_size = 1 << PAGE_SHIFT;
2074 ibdev->ibd_page_mask = ~((__u64)ibdev->ibd_page_size - 1);
2076 LIBCFS_ALLOC(attr, sizeof(*attr));
2078 CERROR("Out of memory\n");
2082 rc = ib_query_device(ibdev->ibd_cmid->device, attr);
2084 ibdev->ibd_mr_size = attr->max_mr_size;
2086 LIBCFS_FREE(attr, sizeof(*attr));
2089 CERROR("Failed to query IB device: %d\n", rc);
2093 #ifdef HAVE_OFED_TRANSPORT_IWARP
2094 /* XXX We can't trust this value returned by Chelsio driver, it's wrong
2095 * and we have reported the bug, remove these in the future when Chelsio
2097 if (rdma_node_get_transport(ibdev->ibd_cmid->device->node_type) ==
2098 RDMA_TRANSPORT_IWARP)
2099 ibdev->ibd_mr_size = (1ULL << 32) - 1;
2102 if (ibdev->ibd_mr_size == ~0ULL) {
2103 ibdev->ibd_mr_shift = 64;
2107 for (ibdev->ibd_mr_shift = 0;
2108 ibdev->ibd_mr_shift < 64; ibdev->ibd_mr_shift ++) {
2109 if (ibdev->ibd_mr_size == (1ULL << ibdev->ibd_mr_shift) ||
2110 ibdev->ibd_mr_size == (1ULL << ibdev->ibd_mr_shift) - 1)
2114 CERROR("Invalid mr size: "LPX64"\n", ibdev->ibd_mr_size);
2119 kiblnd_dev_setup(kib_dev_t *ibdev)
2126 int acflags = IB_ACCESS_LOCAL_WRITE |
2127 IB_ACCESS_REMOTE_WRITE;
2129 rc = kiblnd_dev_get_attr(ibdev);
2133 if (ibdev->ibd_mr_shift == 64) {
2134 LIBCFS_ALLOC(ibdev->ibd_mrs, 1 * sizeof(*ibdev->ibd_mrs));
2135 if (ibdev->ibd_mrs == NULL) {
2136 CERROR("Failed to allocate MRs table\n");
2140 ibdev->ibd_mrs[0] = NULL;
2141 ibdev->ibd_nmrs = 1;
2143 mr = ib_get_dma_mr(ibdev->ibd_pd, acflags);
2145 CERROR("Failed ib_get_dma_mr : %ld\n", PTR_ERR(mr));
2146 kiblnd_dev_cleanup(ibdev);
2150 ibdev->ibd_mrs[0] = mr;
2155 mr_size = (1ULL << ibdev->ibd_mr_shift);
2156 mm_size = (unsigned long)high_memory - PAGE_OFFSET;
2158 ibdev->ibd_nmrs = (int)((mm_size + mr_size - 1) >> ibdev->ibd_mr_shift);
2160 if (ibdev->ibd_mr_shift < 32 || ibdev->ibd_nmrs > 1024) {
2161 /* it's 4T..., assume we will re-code at that time */
2162 CERROR("Can't support memory size: x"LPX64
2163 " with MR size: x"LPX64"\n", mm_size, mr_size);
2167 /* create an array of MRs to cover all memory */
2168 LIBCFS_ALLOC(ibdev->ibd_mrs, sizeof(*ibdev->ibd_mrs) * ibdev->ibd_nmrs);
2169 if (ibdev->ibd_mrs == NULL) {
2170 CERROR("Failed to allocate MRs' table\n");
2174 memset(ibdev->ibd_mrs, 0, sizeof(*ibdev->ibd_mrs) * ibdev->ibd_nmrs);
2176 for (i = 0; i < ibdev->ibd_nmrs; i++) {
2177 struct ib_phys_buf ipb;
2180 ipb.size = ibdev->ibd_mr_size;
2181 ipb.addr = i * mr_size;
2184 mr = ib_reg_phys_mr(ibdev->ibd_pd, &ipb, 1, acflags, &iova);
2186 CERROR("Failed ib_reg_phys_mr addr "LPX64
2187 " size "LPX64" : %ld\n",
2188 ipb.addr, ipb.size, PTR_ERR(mr));
2189 kiblnd_dev_cleanup(ibdev);
2193 LASSERT (iova == ipb.addr);
2195 ibdev->ibd_mrs[i] = mr;
2199 CDEBUG(D_CONSOLE, "Register global MR array, MR size: "
2200 LPX64", array size: %d\n",
2201 ibdev->ibd_mr_size, ibdev->ibd_nmrs);
2203 cfs_list_add_tail(&ibdev->ibd_list,
2204 &kiblnd_data.kib_devs);
2209 kiblnd_destroy_dev (kib_dev_t *dev)
2211 LASSERT (dev->ibd_nnets == 0);
2213 if (!cfs_list_empty(&dev->ibd_list)) /* on kib_devs? */
2214 cfs_list_del_init(&dev->ibd_list);
2216 kiblnd_dev_cleanup(dev);
2218 if (dev->ibd_pd != NULL)
2219 ib_dealloc_pd(dev->ibd_pd);
2221 if (dev->ibd_cmid != NULL)
2222 rdma_destroy_id(dev->ibd_cmid);
2224 LIBCFS_FREE(dev, sizeof(*dev));
2228 kiblnd_base_shutdown (void)
2232 LASSERT (cfs_list_empty(&kiblnd_data.kib_devs));
2234 CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2235 cfs_atomic_read(&libcfs_kmemory));
2237 switch (kiblnd_data.kib_init) {
2241 case IBLND_INIT_ALL:
2242 case IBLND_INIT_DATA:
2243 LASSERT (kiblnd_data.kib_peers != NULL);
2244 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2245 LASSERT (cfs_list_empty(&kiblnd_data.kib_peers[i]));
2247 LASSERT (cfs_list_empty(&kiblnd_data.kib_connd_zombies));
2248 LASSERT (cfs_list_empty(&kiblnd_data.kib_connd_conns));
2250 /* flag threads to terminate; wake and wait for them to die */
2251 kiblnd_data.kib_shutdown = 1;
2252 cfs_waitq_broadcast(&kiblnd_data.kib_sched_waitq);
2253 cfs_waitq_broadcast(&kiblnd_data.kib_connd_waitq);
2256 while (cfs_atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2258 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
2259 "Waiting for %d threads to terminate\n",
2260 cfs_atomic_read(&kiblnd_data.kib_nthreads));
2261 cfs_pause(cfs_time_seconds(1));
2266 case IBLND_INIT_NOTHING:
2270 if (kiblnd_data.kib_peers != NULL)
2271 LIBCFS_FREE(kiblnd_data.kib_peers,
2272 sizeof(cfs_list_t) *
2273 kiblnd_data.kib_peer_hash_size);
2275 CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2276 cfs_atomic_read(&libcfs_kmemory));
2278 kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2279 PORTAL_MODULE_UNUSE;
2283 kiblnd_shutdown (lnet_ni_t *ni)
2285 kib_net_t *net = ni->ni_data;
2286 cfs_rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2288 unsigned long flags;
2290 LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2295 CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
2296 cfs_atomic_read(&libcfs_kmemory));
2298 cfs_write_lock_irqsave(g_lock, flags);
2299 net->ibn_shutdown = 1;
2300 cfs_write_unlock_irqrestore(g_lock, flags);
2302 switch (net->ibn_init) {
2306 case IBLND_INIT_ALL:
2307 /* nuke all existing peers within this net */
2308 kiblnd_del_peer(ni, LNET_NID_ANY);
2310 /* Wait for all peer state to clean up */
2312 while (cfs_atomic_read(&net->ibn_npeers) != 0) {
2314 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* 2**n? */
2315 "%s: waiting for %d peers to disconnect\n",
2316 libcfs_nid2str(ni->ni_nid),
2317 cfs_atomic_read(&net->ibn_npeers));
2318 cfs_pause(cfs_time_seconds(1));
2321 kiblnd_ni_fini_pools(net);
2323 LASSERT (net->ibn_dev->ibd_nnets > 0);
2324 net->ibn_dev->ibd_nnets--;
2328 case IBLND_INIT_NOTHING:
2329 LASSERT (cfs_atomic_read(&net->ibn_nconns) == 0);
2331 if (net->ibn_dev != NULL &&
2332 net->ibn_dev->ibd_nnets == 0)
2333 kiblnd_destroy_dev(net->ibn_dev);
2338 CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
2339 cfs_atomic_read(&libcfs_kmemory));
2341 net->ibn_init = IBLND_INIT_NOTHING;
2344 LIBCFS_FREE(net, sizeof(*net));
2347 if (cfs_list_empty(&kiblnd_data.kib_devs))
2348 kiblnd_base_shutdown();
2353 kiblnd_base_startup (void)
2358 LASSERT (kiblnd_data.kib_init == IBLND_INIT_NOTHING);
2361 memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
2363 cfs_rwlock_init(&kiblnd_data.kib_global_lock);
2365 CFS_INIT_LIST_HEAD(&kiblnd_data.kib_devs);
2367 kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
2368 LIBCFS_ALLOC(kiblnd_data.kib_peers,
2369 sizeof(cfs_list_t) *
2370 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 CFS_INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
2377 cfs_spin_lock_init(&kiblnd_data.kib_connd_lock);
2378 CFS_INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
2379 CFS_INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
2380 cfs_waitq_init(&kiblnd_data.kib_connd_waitq);
2382 cfs_spin_lock_init(&kiblnd_data.kib_sched_lock);
2383 CFS_INIT_LIST_HEAD(&kiblnd_data.kib_sched_conns);
2384 cfs_waitq_init(&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;
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 cfs_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 cfs_list_for_each (tmp, &kiblnd_data.kib_devs) {
2471 ibdev = cfs_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);