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 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, time_t *when)
1080 cfs_time_t last_alive = 0;
1081 rwlock_t *glock = &kiblnd_data.kib_global_lock;
1083 unsigned long flags;
1085 read_lock_irqsave(glock, flags);
1087 peer = kiblnd_find_peer_locked(nid);
1089 LASSERT (peer->ibp_connecting > 0 || /* creating conns */
1090 peer->ibp_accepting > 0 ||
1091 !list_empty(&peer->ibp_conns)); /* active conn */
1092 last_alive = peer->ibp_last_alive;
1095 read_unlock_irqrestore(glock, flags);
1097 if (last_alive != 0)
1098 *when = cfs_time_current_sec() -
1099 cfs_duration_sec(cfs_time_current() - last_alive);
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);
1109 kiblnd_free_pages (kib_pages_t *p)
1111 int npages = p->ibp_npages;
1114 LASSERT (p->ibp_device == NULL);
1116 for (i = 0; i < npages; i++)
1117 if (p->ibp_pages[i] != NULL)
1118 __free_page(p->ibp_pages[i]);
1120 LIBCFS_FREE (p, offsetof(kib_pages_t, ibp_pages[npages]));
1124 kiblnd_alloc_pages (kib_pages_t **pp, int npages)
1129 LIBCFS_ALLOC(p, offsetof(kib_pages_t, ibp_pages[npages]));
1131 CERROR("Can't allocate descriptor for %d pages\n", npages);
1135 memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1136 p->ibp_npages = npages;
1138 for (i = 0; i < npages; i++) {
1139 p->ibp_pages[i] = alloc_page(GFP_KERNEL);
1140 if (p->ibp_pages[i] == NULL) {
1141 CERROR("Can't allocate page %d of %d\n", i, npages);
1142 kiblnd_free_pages(p);
1152 kiblnd_unmap_rx_descs(kib_conn_t *conn)
1157 LASSERT (conn->ibc_rxs != NULL);
1158 LASSERT (conn->ibc_rx_pages->ibp_device != NULL);
1160 for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1161 rx = &conn->ibc_rxs[i];
1163 LASSERT (rx->rx_nob >= 0); /* not posted */
1165 kiblnd_dma_unmap_single(conn->ibc_rx_pages->ibp_device,
1166 KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1168 IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1171 conn->ibc_rx_pages->ibp_device = NULL;
1173 kiblnd_free_pages(conn->ibc_rx_pages);
1175 conn->ibc_rx_pages = NULL;
1179 kiblnd_map_rx_descs(kib_conn_t *conn)
1187 for (pg_off = ipg = i = 0;
1188 i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1189 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1190 rx = &conn->ibc_rxs[i];
1193 rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
1195 rx->rx_msgaddr = kiblnd_dma_map_single(conn->ibc_cmid->device,
1196 rx->rx_msg, IBLND_MSG_SIZE,
1198 LASSERT (!kiblnd_dma_mapping_error(conn->ibc_cmid->device,
1200 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1202 CDEBUG(D_NET,"rx %d: %p "LPX64"("LPX64")\n",
1203 i, rx->rx_msg, rx->rx_msgaddr,
1204 lnet_page2phys(pg) + pg_off);
1206 pg_off += IBLND_MSG_SIZE;
1207 LASSERT (pg_off <= PAGE_SIZE);
1209 if (pg_off == PAGE_SIZE) {
1212 LASSERT (ipg <= IBLND_RX_MSG_PAGES(conn->ibc_version));
1216 conn->ibc_rx_pages->ibp_device = conn->ibc_cmid->device;
1220 kiblnd_unmap_tx_descs(lnet_ni_t *ni)
1224 kib_net_t *net = ni->ni_data;
1226 LASSERT (net->ibn_tx_pages != NULL);
1227 LASSERT (net->ibn_tx_pages->ibp_device != NULL);
1229 for (i = 0; i < IBLND_TX_MSGS(); i++) {
1230 tx = &net->ibn_tx_descs[i];
1232 kiblnd_dma_unmap_single(net->ibn_tx_pages->ibp_device,
1233 KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1235 IBLND_MSG_SIZE, DMA_TO_DEVICE);
1238 net->ibn_tx_pages->ibp_device = NULL;
1242 kiblnd_map_tx_descs (lnet_ni_t *ni)
1244 kib_net_t *net = ni->ni_data;
1251 LASSERT (net != NULL);
1253 /* pre-mapped messages are not bigger than 1 page */
1254 CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE);
1256 /* No fancy arithmetic when we do the buffer calculations */
1257 CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
1260 for (ipage = page_offset = i = 0; i < IBLND_TX_MSGS(); i++) {
1261 page = net->ibn_tx_pages->ibp_pages[ipage];
1262 tx = &net->ibn_tx_descs[i];
1264 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1267 tx->tx_msgaddr = kiblnd_dma_map_single(
1268 net->ibn_dev->ibd_cmid->device,
1269 tx->tx_msg, IBLND_MSG_SIZE, DMA_TO_DEVICE);
1270 LASSERT (!kiblnd_dma_mapping_error(net->ibn_dev->ibd_cmid->device,
1272 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1274 list_add(&tx->tx_list, &net->ibn_idle_txs);
1276 page_offset += IBLND_MSG_SIZE;
1277 LASSERT (page_offset <= PAGE_SIZE);
1279 if (page_offset == PAGE_SIZE) {
1282 LASSERT (ipage <= IBLND_TX_MSG_PAGES());
1286 net->ibn_tx_pages->ibp_device = net->ibn_dev->ibd_cmid->device;
1290 kiblnd_free_tx_descs (lnet_ni_t *ni)
1293 kib_net_t *net = ni->ni_data;
1295 LASSERT (net != NULL);
1297 if (net->ibn_tx_pages != NULL)
1298 kiblnd_free_pages(net->ibn_tx_pages);
1300 if (net->ibn_tx_descs == NULL)
1303 for (i = 0; i < IBLND_TX_MSGS(); i++) {
1304 kib_tx_t *tx = &net->ibn_tx_descs[i];
1306 if (tx->tx_pages != NULL)
1307 LIBCFS_FREE(tx->tx_pages,
1309 sizeof(*tx->tx_pages));
1311 if (tx->tx_ipb != NULL)
1312 LIBCFS_FREE(tx->tx_ipb,
1313 IBLND_MAX_RDMA_FRAGS *
1314 sizeof(*tx->tx_ipb));
1316 if (tx->tx_frags != NULL)
1317 LIBCFS_FREE(tx->tx_frags,
1318 IBLND_MAX_RDMA_FRAGS *
1319 sizeof(*tx->tx_frags));
1321 if (tx->tx_wrq != NULL)
1322 LIBCFS_FREE(tx->tx_wrq,
1323 (1 + IBLND_MAX_RDMA_FRAGS) *
1324 sizeof(*tx->tx_wrq));
1326 if (tx->tx_sge != NULL)
1327 LIBCFS_FREE(tx->tx_sge,
1328 (1 + IBLND_MAX_RDMA_FRAGS) *
1329 sizeof(*tx->tx_sge));
1331 if (tx->tx_rd != NULL)
1332 LIBCFS_FREE(tx->tx_rd,
1333 offsetof(kib_rdma_desc_t,
1334 rd_frags[IBLND_MAX_RDMA_FRAGS]));
1337 LIBCFS_FREE(net->ibn_tx_descs,
1338 IBLND_TX_MSGS() * sizeof(kib_tx_t));
1342 kiblnd_alloc_tx_descs (lnet_ni_t *ni)
1346 kib_net_t *net = ni->ni_data;
1348 LASSERT (net != NULL);
1350 rc = kiblnd_alloc_pages(&net->ibn_tx_pages, IBLND_TX_MSG_PAGES());
1353 CERROR("Can't allocate tx pages\n");
1357 LIBCFS_ALLOC (net->ibn_tx_descs,
1358 IBLND_TX_MSGS() * sizeof(kib_tx_t));
1359 if (net->ibn_tx_descs == NULL) {
1360 CERROR("Can't allocate %d tx descriptors\n", IBLND_TX_MSGS());
1364 memset(net->ibn_tx_descs, 0,
1365 IBLND_TX_MSGS() * sizeof(kib_tx_t));
1367 for (i = 0; i < IBLND_TX_MSGS(); i++) {
1368 kib_tx_t *tx = &net->ibn_tx_descs[i];
1370 if (net->ibn_fmrpool != NULL) {
1371 LIBCFS_ALLOC(tx->tx_pages, LNET_MAX_IOV *
1372 sizeof(*tx->tx_pages));
1373 if (tx->tx_pages == NULL)
1377 if (net->ibn_pmrpool != NULL) {
1378 LIBCFS_ALLOC(tx->tx_ipb,
1379 IBLND_MAX_RDMA_FRAGS *
1380 sizeof(*tx->tx_ipb));
1381 if (tx->tx_ipb == NULL)
1385 LIBCFS_ALLOC(tx->tx_frags,
1386 IBLND_MAX_RDMA_FRAGS *
1387 sizeof(*tx->tx_frags));
1388 if (tx->tx_frags == NULL)
1391 LIBCFS_ALLOC(tx->tx_wrq,
1392 (1 + IBLND_MAX_RDMA_FRAGS) *
1393 sizeof(*tx->tx_wrq));
1394 if (tx->tx_wrq == NULL)
1397 LIBCFS_ALLOC(tx->tx_sge,
1398 (1 + IBLND_MAX_RDMA_FRAGS) *
1399 sizeof(*tx->tx_sge));
1400 if (tx->tx_sge == NULL)
1403 LIBCFS_ALLOC(tx->tx_rd,
1404 offsetof(kib_rdma_desc_t,
1405 rd_frags[IBLND_MAX_RDMA_FRAGS]));
1406 if (tx->tx_rd == NULL)
1414 kiblnd_find_dma_mr(kib_net_t *net, __u64 addr, __u64 size)
1418 LASSERT (net->ibn_dev->ibd_mrs[0] != NULL);
1420 if (net->ibn_dev->ibd_nmrs == 1)
1421 return net->ibn_dev->ibd_mrs[0];
1423 index = addr >> net->ibn_dev->ibd_mr_shift;
1425 if (index < net->ibn_dev->ibd_nmrs &&
1426 index == ((addr + size - 1) >> net->ibn_dev->ibd_mr_shift))
1427 return net->ibn_dev->ibd_mrs[index];
1433 kiblnd_find_rd_dma_mr(kib_net_t *net, kib_rdma_desc_t *rd)
1435 struct ib_mr *prev_mr;
1439 LASSERT (net->ibn_dev->ibd_mrs[0] != NULL);
1441 if (*kiblnd_tunables.kib_map_on_demand > 0 &&
1442 *kiblnd_tunables.kib_map_on_demand <= rd->rd_nfrags)
1445 if (net->ibn_dev->ibd_nmrs == 1)
1446 return net->ibn_dev->ibd_mrs[0];
1448 for (i = 0, mr = prev_mr = NULL;
1449 i < rd->rd_nfrags; i++) {
1450 mr = kiblnd_find_dma_mr(net,
1451 rd->rd_frags[i].rf_addr,
1452 rd->rd_frags[i].rf_nob);
1453 if (prev_mr == NULL)
1456 if (mr == NULL || prev_mr != mr) {
1457 /* Can't covered by one single MR */
1467 kiblnd_dev_cleanup(kib_dev_t *ibdev)
1471 if (ibdev->ibd_mrs == NULL)
1474 for (i = 0; i < ibdev->ibd_nmrs; i++) {
1475 if (ibdev->ibd_mrs[i] == NULL)
1478 ib_dereg_mr(ibdev->ibd_mrs[i]);
1481 LIBCFS_FREE(ibdev->ibd_mrs, sizeof(*ibdev->ibd_mrs) * ibdev->ibd_nmrs);
1482 ibdev->ibd_mrs = NULL;
1486 kiblnd_ib_create_fmr_pool(kib_dev_t *ibdev, struct ib_fmr_pool **fmrpp)
1488 /* FMR pool for RDMA */
1489 struct ib_fmr_pool *fmrpool;
1490 struct ib_fmr_pool_param param = {
1491 .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
1492 .page_shift = PAGE_SHIFT,
1493 .access = (IB_ACCESS_LOCAL_WRITE |
1494 IB_ACCESS_REMOTE_WRITE),
1495 .pool_size = *kiblnd_tunables.kib_fmr_pool_size,
1496 .dirty_watermark = *kiblnd_tunables.kib_fmr_flush_trigger,
1497 .flush_function = NULL,
1499 .cache = *kiblnd_tunables.kib_fmr_cache};
1501 if (*kiblnd_tunables.kib_fmr_pool_size <
1502 *kiblnd_tunables.kib_ntx) {
1503 CERROR("Can't set fmr pool size (%d) < ntx(%d)\n",
1504 *kiblnd_tunables.kib_fmr_pool_size,
1505 *kiblnd_tunables.kib_ntx);
1509 fmrpool = ib_create_fmr_pool(ibdev->ibd_pd, ¶m);
1510 if (IS_ERR(fmrpool))
1511 return PTR_ERR(fmrpool);
1519 kiblnd_phys_mr_unmap(kib_net_t *net, kib_phys_mr_t *pmr)
1521 kib_phys_mr_pool_t *pool = net->ibn_pmrpool;
1524 spin_lock(&pool->ibmp_lock);
1528 list_add(&pmr->ibpm_link, &pool->ibmp_free_list);
1529 pool->ibmp_allocated --;
1531 spin_unlock(&pool->ibmp_lock);
1538 kiblnd_phys_mr_map(kib_net_t *net, kib_rdma_desc_t *rd,
1539 struct ib_phys_buf *ipb, __u64 *iova)
1541 kib_phys_mr_pool_t *pool = net->ibn_pmrpool;
1545 LASSERT (ipb != NULL);
1547 spin_lock(&pool->ibmp_lock);
1548 if (list_empty(&pool->ibmp_free_list)) {
1549 spin_unlock(&pool->ibmp_lock);
1550 CERROR("pre-allocated MRs is not enough\n");
1555 pmr = list_entry(pool->ibmp_free_list.next,
1556 kib_phys_mr_t, ibpm_link);
1557 list_del_init(&pmr->ibpm_link);
1558 pool->ibmp_allocated ++;
1560 spin_unlock(&pool->ibmp_lock);
1562 for (i = 0; i < rd->rd_nfrags; i ++) {
1563 ipb[i].addr = rd->rd_frags[i].rf_addr;
1564 ipb[i].size = rd->rd_frags[i].rf_nob;
1567 pmr->ibpm_mr = ib_reg_phys_mr(net->ibn_dev->ibd_pd, ipb,
1569 IB_ACCESS_LOCAL_WRITE |
1570 IB_ACCESS_REMOTE_WRITE,
1572 if (!IS_ERR(pmr->ibpm_mr)) {
1573 pmr->ibpm_iova = *iova;
1577 CERROR("Failed ib_reg_phys_mr: %ld\n", PTR_ERR(pmr->ibpm_mr));
1578 pmr->ibpm_mr = NULL;
1580 spin_lock(&pool->ibmp_lock);
1582 list_add(&pmr->ibpm_link, &pool->ibmp_free_list);
1583 pool->ibmp_allocated --;
1585 spin_unlock(&pool->ibmp_lock);
1591 kiblnd_destroy_pmr_pool(kib_phys_mr_pool_t *pool)
1595 LASSERT (pool->ibmp_allocated == 0);
1597 while (!list_empty(&pool->ibmp_free_list)) {
1598 pmr = list_entry(pool->ibmp_free_list.next,
1599 kib_phys_mr_t, ibpm_link);
1601 LASSERT (pmr->ibpm_mr == NULL);
1603 list_del(&pmr->ibpm_link);
1605 LIBCFS_FREE(pmr, sizeof(kib_phys_mr_t));
1608 LIBCFS_FREE(pool, sizeof(kib_phys_mr_pool_t));
1612 kiblnd_create_pmr_pool(kib_dev_t *ibdev, kib_phys_mr_pool_t **poolpp)
1614 kib_phys_mr_pool_t *pool;
1618 if (*kiblnd_tunables.kib_pmr_pool_size <
1619 *kiblnd_tunables.kib_ntx) {
1620 CERROR("Can't set pmr pool size (%d) < ntx(%d)\n",
1621 *kiblnd_tunables.kib_pmr_pool_size,
1622 *kiblnd_tunables.kib_ntx);
1626 LIBCFS_ALLOC(pool, sizeof(kib_phys_mr_pool_t));
1630 spin_lock_init(&pool->ibmp_lock);
1632 pool->ibmp_allocated = 0;
1633 CFS_INIT_LIST_HEAD(&pool->ibmp_free_list);
1635 for (i = 0; i < *kiblnd_tunables.kib_pmr_pool_size; i++) {
1636 LIBCFS_ALLOC(pmr, sizeof(kib_phys_mr_t));
1639 kiblnd_destroy_pmr_pool(pool);
1643 memset(pmr, 0, sizeof(kib_phys_mr_t));
1645 list_add(&pmr->ibpm_link, &pool->ibmp_free_list);
1654 kiblnd_dev_get_attr(kib_dev_t *ibdev)
1656 struct ib_device_attr *attr;
1659 /* XXX here should be HCA's page shift/size/mask in the future? */
1660 ibdev->ibd_page_shift = PAGE_SHIFT;
1661 ibdev->ibd_page_size = 1 << PAGE_SHIFT;
1662 ibdev->ibd_page_mask = ~((__u64)ibdev->ibd_page_size - 1);
1664 LIBCFS_ALLOC(attr, sizeof(*attr));
1666 CERROR("Out of memory\n");
1670 rc = ib_query_device(ibdev->ibd_cmid->device, attr);
1672 ibdev->ibd_mr_size = attr->max_mr_size;
1674 LIBCFS_FREE(attr, sizeof(*attr));
1677 CERROR("Failed to query IB device: %d\n", rc);
1681 #ifdef HAVE_OFED_TRANSPORT_IWARP
1682 /* XXX We can't trust this value returned by Chelsio driver, it's wrong
1683 * and we have reported the bug, remove these in the future when Chelsio
1685 if (rdma_node_get_transport(ibdev->ibd_cmid->device->node_type) ==
1686 RDMA_TRANSPORT_IWARP)
1687 ibdev->ibd_mr_size = (1ULL << 32) - 1;
1690 if (ibdev->ibd_mr_size == ~0ULL) {
1691 ibdev->ibd_mr_shift = 64;
1695 for (ibdev->ibd_mr_shift = 0;
1696 ibdev->ibd_mr_shift < 64; ibdev->ibd_mr_shift ++) {
1697 if (ibdev->ibd_mr_size == (1ULL << ibdev->ibd_mr_shift) ||
1698 ibdev->ibd_mr_size == (1ULL << ibdev->ibd_mr_shift) - 1)
1702 CERROR("Invalid mr size: "LPX64"\n", ibdev->ibd_mr_size);
1707 kiblnd_dev_setup(kib_dev_t *ibdev)
1714 int acflags = IB_ACCESS_LOCAL_WRITE |
1715 IB_ACCESS_REMOTE_WRITE;
1717 rc = kiblnd_dev_get_attr(ibdev);
1721 if (ibdev->ibd_mr_shift == 64) {
1722 LIBCFS_ALLOC(ibdev->ibd_mrs, 1 * sizeof(*ibdev->ibd_mrs));
1723 if (ibdev->ibd_mrs == NULL) {
1724 CERROR("Failed to allocate MRs table\n");
1728 ibdev->ibd_mrs[0] = NULL;
1729 ibdev->ibd_nmrs = 1;
1731 mr = ib_get_dma_mr(ibdev->ibd_pd, acflags);
1733 CERROR("Failed ib_get_dma_mr : %ld\n", PTR_ERR(mr));
1734 kiblnd_dev_cleanup(ibdev);
1738 ibdev->ibd_mrs[0] = mr;
1743 mr_size = (1ULL << ibdev->ibd_mr_shift);
1744 mm_size = (unsigned long)high_memory - PAGE_OFFSET;
1746 ibdev->ibd_nmrs = (int)((mm_size + mr_size - 1) >> ibdev->ibd_mr_shift);
1748 if (ibdev->ibd_mr_shift < 32 || ibdev->ibd_nmrs > 1024) {
1749 /* it's 4T..., assume we will re-code at that time */
1750 CERROR("Can't support memory size: x"LPX64
1751 " with MR size: x"LPX64"\n", mm_size, mr_size);
1755 /* create an array of MRs to cover all memory */
1756 LIBCFS_ALLOC(ibdev->ibd_mrs, sizeof(*ibdev->ibd_mrs) * ibdev->ibd_nmrs);
1757 if (ibdev->ibd_mrs == NULL) {
1758 CERROR("Failed to allocate MRs' table\n");
1762 memset(ibdev->ibd_mrs, 0, sizeof(*ibdev->ibd_mrs) * ibdev->ibd_nmrs);
1764 for (i = 0; i < ibdev->ibd_nmrs; i++) {
1765 struct ib_phys_buf ipb;
1768 ipb.size = ibdev->ibd_mr_size;
1769 ipb.addr = i * mr_size;
1772 mr = ib_reg_phys_mr(ibdev->ibd_pd, &ipb, 1, acflags, &iova);
1774 CERROR("Failed ib_reg_phys_mr addr "LPX64
1775 " size "LPX64" : %ld\n",
1776 ipb.addr, ipb.size, PTR_ERR(mr));
1777 kiblnd_dev_cleanup(ibdev);
1781 LASSERT (iova == ipb.addr);
1783 ibdev->ibd_mrs[i] = mr;
1787 CDEBUG(D_CONSOLE, "Register global MR array, MR size: "
1788 LPX64", array size: %d\n",
1789 ibdev->ibd_mr_size, ibdev->ibd_nmrs);
1791 list_add_tail(&ibdev->ibd_list,
1792 &kiblnd_data.kib_devs);
1797 kiblnd_destroy_dev (kib_dev_t *dev)
1799 LASSERT (dev->ibd_nnets == 0);
1801 if (!list_empty(&dev->ibd_list)) /* on kib_devs? */
1802 list_del_init(&dev->ibd_list);
1804 kiblnd_dev_cleanup(dev);
1806 if (dev->ibd_pd != NULL)
1807 ib_dealloc_pd(dev->ibd_pd);
1809 if (dev->ibd_cmid != NULL)
1810 rdma_destroy_id(dev->ibd_cmid);
1812 LIBCFS_FREE(dev, sizeof(*dev));
1816 kiblnd_base_shutdown (void)
1820 LASSERT (list_empty(&kiblnd_data.kib_devs));
1822 CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
1823 atomic_read(&libcfs_kmemory));
1825 switch (kiblnd_data.kib_init) {
1829 case IBLND_INIT_ALL:
1830 case IBLND_INIT_DATA:
1831 LASSERT (kiblnd_data.kib_peers != NULL);
1832 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
1833 LASSERT (list_empty(&kiblnd_data.kib_peers[i]));
1835 LASSERT (list_empty(&kiblnd_data.kib_connd_zombies));
1836 LASSERT (list_empty(&kiblnd_data.kib_connd_conns));
1838 /* flag threads to terminate; wake and wait for them to die */
1839 kiblnd_data.kib_shutdown = 1;
1840 wake_up_all(&kiblnd_data.kib_sched_waitq);
1841 wake_up_all(&kiblnd_data.kib_connd_waitq);
1844 while (atomic_read(&kiblnd_data.kib_nthreads) != 0) {
1846 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
1847 "Waiting for %d threads to terminate\n",
1848 atomic_read(&kiblnd_data.kib_nthreads));
1849 cfs_pause(cfs_time_seconds(1));
1854 case IBLND_INIT_NOTHING:
1858 if (kiblnd_data.kib_peers != NULL)
1859 LIBCFS_FREE(kiblnd_data.kib_peers,
1860 sizeof(struct list_head) *
1861 kiblnd_data.kib_peer_hash_size);
1863 CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
1864 atomic_read(&libcfs_kmemory));
1866 kiblnd_data.kib_init = IBLND_INIT_NOTHING;
1867 PORTAL_MODULE_UNUSE;
1871 kiblnd_shutdown (lnet_ni_t *ni)
1873 kib_net_t *net = ni->ni_data;
1874 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
1876 unsigned long flags;
1878 LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
1883 CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
1884 atomic_read(&libcfs_kmemory));
1886 write_lock_irqsave(g_lock, flags);
1887 net->ibn_shutdown = 1;
1888 write_unlock_irqrestore(g_lock, flags);
1890 switch (net->ibn_init) {
1894 case IBLND_INIT_ALL:
1895 /* nuke all existing peers within this net */
1896 kiblnd_del_peer(ni, LNET_NID_ANY);
1898 /* Wait for all peer state to clean up */
1900 while (atomic_read(&net->ibn_npeers) != 0) {
1902 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* 2**n? */
1903 "%s: waiting for %d peers to disconnect\n",
1904 libcfs_nid2str(ni->ni_nid),
1905 atomic_read(&net->ibn_npeers));
1906 cfs_pause(cfs_time_seconds(1));
1909 kiblnd_unmap_tx_descs(ni);
1911 LASSERT (net->ibn_dev->ibd_nnets > 0);
1912 net->ibn_dev->ibd_nnets--;
1916 case IBLND_INIT_NOTHING:
1917 LASSERT (atomic_read(&net->ibn_nconns) == 0);
1919 if (net->ibn_fmrpool != NULL)
1920 ib_destroy_fmr_pool(net->ibn_fmrpool);
1921 if (net->ibn_pmrpool != NULL)
1922 kiblnd_destroy_pmr_pool(net->ibn_pmrpool);
1924 if (net->ibn_dev != NULL &&
1925 net->ibn_dev->ibd_nnets == 0)
1926 kiblnd_destroy_dev(net->ibn_dev);
1931 kiblnd_free_tx_descs(ni);
1933 CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
1934 atomic_read(&libcfs_kmemory));
1936 net->ibn_init = IBLND_INIT_NOTHING;
1939 LIBCFS_FREE(net, sizeof(*net));
1942 if (list_empty(&kiblnd_data.kib_devs))
1943 kiblnd_base_shutdown();
1948 kiblnd_base_startup (void)
1953 LASSERT (kiblnd_data.kib_init == IBLND_INIT_NOTHING);
1956 memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
1958 rwlock_init(&kiblnd_data.kib_global_lock);
1960 INIT_LIST_HEAD(&kiblnd_data.kib_devs);
1962 kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
1963 LIBCFS_ALLOC(kiblnd_data.kib_peers,
1964 sizeof(struct list_head) * kiblnd_data.kib_peer_hash_size);
1965 if (kiblnd_data.kib_peers == NULL) {
1968 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
1969 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
1971 spin_lock_init(&kiblnd_data.kib_connd_lock);
1972 INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
1973 INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
1974 init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
1976 spin_lock_init(&kiblnd_data.kib_sched_lock);
1977 INIT_LIST_HEAD(&kiblnd_data.kib_sched_conns);
1978 init_waitqueue_head(&kiblnd_data.kib_sched_waitq);
1980 kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
1982 /* lists/ptrs/locks initialised */
1983 kiblnd_data.kib_init = IBLND_INIT_DATA;
1984 /*****************************************************/
1986 for (i = 0; i < IBLND_N_SCHED; i++) {
1987 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)((long)i));
1989 CERROR("Can't spawn o2iblnd scheduler[%d]: %d\n",
1995 rc = kiblnd_thread_start(kiblnd_connd, NULL);
1997 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
2001 /* flag everything initialised */
2002 kiblnd_data.kib_init = IBLND_INIT_ALL;
2003 /*****************************************************/
2008 kiblnd_base_shutdown();
2013 kiblnd_startup (lnet_ni_t *ni)
2016 kib_dev_t *ibdev = NULL;
2018 struct list_head *tmp;
2022 LASSERT (ni->ni_lnd == &the_o2iblnd);
2024 if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
2025 rc = kiblnd_base_startup();
2030 LIBCFS_ALLOC(net, sizeof(*net));
2035 memset(net, 0, sizeof(*net));
2037 do_gettimeofday(&tv);
2038 net->ibn_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
2040 ni->ni_peertimeout = *kiblnd_tunables.kib_peertimeout;
2041 ni->ni_maxtxcredits = *kiblnd_tunables.kib_credits;
2042 ni->ni_peertxcredits = *kiblnd_tunables.kib_peertxcredits;
2043 ni->ni_peerrtrcredits = *kiblnd_tunables.kib_peerrtrcredits;
2045 spin_lock_init(&net->ibn_tx_lock);
2046 INIT_LIST_HEAD(&net->ibn_idle_txs);
2048 if (ni->ni_interfaces[0] != NULL) {
2049 /* Use the IPoIB interface specified in 'networks=' */
2051 CLASSERT (LNET_MAX_INTERFACES > 1);
2052 if (ni->ni_interfaces[1] != NULL) {
2053 CERROR("Multiple interfaces not supported\n");
2057 ifname = ni->ni_interfaces[0];
2059 ifname = *kiblnd_tunables.kib_default_ipif;
2062 if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
2063 CERROR("IPoIB interface name too long: %s\n", ifname);
2067 list_for_each (tmp, &kiblnd_data.kib_devs) {
2068 ibdev = list_entry(tmp, kib_dev_t, ibd_list);
2070 if (!strcmp(&ibdev->ibd_ifname[0], ifname))
2076 if (ibdev == NULL) {
2080 struct rdma_cm_id *id;
2082 struct sockaddr_in addr;
2084 rc = libcfs_ipif_query(ifname, &up, &ip, &netmask);
2086 CERROR("Can't query IPoIB interface %s: %d\n",
2092 CERROR("Can't query IPoIB interface %s: it's down\n",
2097 LIBCFS_ALLOC(ibdev, sizeof(*ibdev));
2101 memset(ibdev, 0, sizeof(*ibdev));
2103 CFS_INIT_LIST_HEAD(&ibdev->ibd_list); /* not yet in kib_devs */
2104 ibdev->ibd_ifip = ip;
2105 strcpy(&ibdev->ibd_ifname[0], ifname);
2107 id = rdma_create_id(kiblnd_cm_callback, ibdev, RDMA_PS_TCP);
2109 CERROR("Can't create listen ID: %ld\n", PTR_ERR(id));
2113 ibdev->ibd_cmid = id;
2115 memset(&addr, 0, sizeof(addr));
2116 addr.sin_family = AF_INET;
2117 addr.sin_port = htons(*kiblnd_tunables.kib_service);
2118 addr.sin_addr.s_addr = htonl(ip);
2120 rc = rdma_bind_addr(id, (struct sockaddr *)&addr);
2122 CERROR("Can't bind to %s: %d\n", ifname, rc);
2126 /* Binding should have assigned me an IB device */
2127 LASSERT (id->device != NULL);
2129 pd = ib_alloc_pd(id->device);
2131 CERROR("Can't allocate PD: %ld\n", PTR_ERR(pd));
2137 rc = rdma_listen(id, 256);
2139 CERROR("Can't start listener: %d\n", rc);
2143 rc = kiblnd_dev_setup(ibdev);
2145 CERROR("Can't setup device: %d\n", rc);
2150 ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
2151 net->ibn_dev = ibdev;
2153 if (*kiblnd_tunables.kib_map_on_demand > 0 ||
2154 ibdev->ibd_nmrs > 1) { /* premapping can fail if ibd_nmr > 1,
2155 * so we always create FMR/PMR pool and
2156 * map-on-demand if premapping failed */
2158 rc = kiblnd_ib_create_fmr_pool(ibdev, &net->ibn_fmrpool);
2159 if (rc == -ENOSYS) {
2160 CDEBUG(D_CONSOLE, "No FMR, creating physical mapping\n");
2162 rc = kiblnd_create_pmr_pool(ibdev, &net->ibn_pmrpool);
2166 CERROR("Can't create FMR or physical mapping pool: %d, "
2167 "please disable map_on_demand and retry\n", rc);
2173 rc = kiblnd_alloc_tx_descs(ni);
2175 CERROR("Can't allocate tx descs\n");
2179 kiblnd_map_tx_descs(ni);
2182 net->ibn_init = IBLND_INIT_ALL;
2187 if (net->ibn_dev == NULL && ibdev != NULL)
2188 kiblnd_destroy_dev(ibdev);
2190 kiblnd_shutdown(ni);
2192 CDEBUG(D_NET, "kiblnd_startup failed\n");
2197 kiblnd_module_fini (void)
2199 lnet_unregister_lnd(&the_o2iblnd);
2200 kiblnd_tunables_fini();
2204 kiblnd_module_init (void)
2208 CLASSERT (sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
2209 CLASSERT (offsetof(kib_msg_t, ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
2211 CLASSERT (offsetof(kib_msg_t, ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
2214 rc = kiblnd_tunables_init();
2218 lnet_register_lnd(&the_o2iblnd);
2223 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
2224 MODULE_DESCRIPTION("Kernel OpenIB gen2 LND v2.00");
2225 MODULE_LICENSE("GPL");
2227 module_init(kiblnd_module_init);
2228 module_exit(kiblnd_module_fini);