4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, Whamcloud, Inc.
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_noops:\n");
627 cfs_list_for_each(tmp, &conn->ibc_tx_noops)
628 kiblnd_debug_tx(cfs_list_entry(tmp, kib_tx_t, tx_list));
630 CDEBUG(D_CONSOLE, " tx_queue_nocred:\n");
631 cfs_list_for_each(tmp, &conn->ibc_tx_queue_nocred)
632 kiblnd_debug_tx(cfs_list_entry(tmp, kib_tx_t, tx_list));
634 CDEBUG(D_CONSOLE, " tx_queue_rsrvd:\n");
635 cfs_list_for_each(tmp, &conn->ibc_tx_queue_rsrvd)
636 kiblnd_debug_tx(cfs_list_entry(tmp, kib_tx_t, tx_list));
638 CDEBUG(D_CONSOLE, " tx_queue:\n");
639 cfs_list_for_each(tmp, &conn->ibc_tx_queue)
640 kiblnd_debug_tx(cfs_list_entry(tmp, kib_tx_t, tx_list));
642 CDEBUG(D_CONSOLE, " active_txs:\n");
643 cfs_list_for_each(tmp, &conn->ibc_active_txs)
644 kiblnd_debug_tx(cfs_list_entry(tmp, kib_tx_t, tx_list));
646 CDEBUG(D_CONSOLE, " rxs:\n");
647 for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++)
648 kiblnd_debug_rx(&conn->ibc_rxs[i]);
650 cfs_spin_unlock(&conn->ibc_lock);
654 kiblnd_translate_mtu(int value)
675 kiblnd_setup_mtu_locked(struct rdma_cm_id *cmid)
679 /* XXX There is no path record for iWARP, set by netdev->change_mtu? */
680 if (cmid->route.path_rec == NULL)
683 mtu = kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu);
686 cmid->route.path_rec->mtu = mtu;
690 kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
691 int state, int version)
694 * If the new conn is created successfully it takes over the caller's
695 * ref on 'peer'. It also "owns" 'cmid' and destroys it when it itself
696 * is destroyed. On failure, the caller's ref on 'peer' remains and
697 * she must dispose of 'cmid'. (Actually I'd block forever if I tried
698 * to destroy 'cmid' here since I'm called from the CM which still has
699 * its ref on 'cmid'). */
700 cfs_rwlock_t *glock = &kiblnd_data.kib_global_lock;
701 kib_net_t *net = peer->ibp_ni->ni_data;
702 kib_dev_t *dev = net->ibn_dev;
703 struct ib_qp_init_attr *init_qp_attr;
710 LASSERT (net != NULL);
711 LASSERT (!cfs_in_interrupt());
713 LIBCFS_ALLOC(init_qp_attr, sizeof(*init_qp_attr));
714 if (init_qp_attr == NULL) {
715 CERROR("Can't allocate qp_attr for %s\n",
716 libcfs_nid2str(peer->ibp_nid));
720 LIBCFS_ALLOC(conn, sizeof(*conn));
722 CERROR("Can't allocate connection for %s\n",
723 libcfs_nid2str(peer->ibp_nid));
727 memset(conn, 0, sizeof(*conn)); /* zero flags, NULL pointers etc... */
729 conn->ibc_state = IBLND_CONN_INIT;
730 conn->ibc_version = version;
731 conn->ibc_peer = peer; /* I take the caller's ref */
732 cmid->context = conn; /* for future CM callbacks */
733 conn->ibc_cmid = cmid;
735 CFS_INIT_LIST_HEAD(&conn->ibc_early_rxs);
736 CFS_INIT_LIST_HEAD(&conn->ibc_tx_noops);
737 CFS_INIT_LIST_HEAD(&conn->ibc_tx_queue);
738 CFS_INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
739 CFS_INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
740 CFS_INIT_LIST_HEAD(&conn->ibc_active_txs);
741 cfs_spin_lock_init(&conn->ibc_lock);
743 LIBCFS_ALLOC(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
744 if (conn->ibc_connvars == NULL) {
745 CERROR("Can't allocate in-progress connection state\n");
748 memset(conn->ibc_connvars, 0, sizeof(*conn->ibc_connvars));
750 cfs_write_lock_irqsave(glock, flags);
751 if (dev->ibd_failover) {
752 cfs_write_unlock_irqrestore(glock, flags);
753 CERROR("%s: failover in progress\n", dev->ibd_ifname);
757 if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
758 /* wakeup failover thread and teardown connection */
759 if (kiblnd_dev_can_failover(dev)) {
760 cfs_list_add_tail(&dev->ibd_fail_list,
761 &kiblnd_data.kib_failed_devs);
762 cfs_waitq_signal(&kiblnd_data.kib_failover_waitq);
765 cfs_write_unlock_irqrestore(glock, flags);
766 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
767 cmid->device->name, dev->ibd_ifname);
771 kiblnd_hdev_addref_locked(dev->ibd_hdev);
772 conn->ibc_hdev = dev->ibd_hdev;
774 kiblnd_setup_mtu_locked(cmid);
776 cfs_write_unlock_irqrestore(glock, flags);
778 LIBCFS_ALLOC(conn->ibc_rxs, IBLND_RX_MSGS(version) * sizeof(kib_rx_t));
779 if (conn->ibc_rxs == NULL) {
780 CERROR("Cannot allocate RX buffers\n");
783 memset(conn->ibc_rxs, 0, IBLND_RX_MSGS(version) * sizeof(kib_rx_t));
785 rc = kiblnd_alloc_pages(&conn->ibc_rx_pages,
786 IBLND_RX_MSG_PAGES(version));
790 kiblnd_map_rx_descs(conn);
792 #ifdef HAVE_OFED_IB_COMP_VECTOR
793 cq = ib_create_cq(cmid->device,
794 kiblnd_cq_completion, kiblnd_cq_event, conn,
795 IBLND_CQ_ENTRIES(version), 0);
797 cq = ib_create_cq(cmid->device,
798 kiblnd_cq_completion, kiblnd_cq_event, conn,
799 IBLND_CQ_ENTRIES(version));
802 CERROR("Can't create CQ: %ld, cqe: %d\n",
803 PTR_ERR(cq), IBLND_CQ_ENTRIES(version));
809 rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
811 CERROR("Can't request completion notificiation: %d\n", rc);
815 memset(init_qp_attr, 0, sizeof(*init_qp_attr));
816 init_qp_attr->event_handler = kiblnd_qp_event;
817 init_qp_attr->qp_context = conn;
818 init_qp_attr->cap.max_send_wr = IBLND_SEND_WRS(version);
819 init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(version);
820 init_qp_attr->cap.max_send_sge = 1;
821 init_qp_attr->cap.max_recv_sge = 1;
822 init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
823 init_qp_attr->qp_type = IB_QPT_RC;
824 init_qp_attr->send_cq = cq;
825 init_qp_attr->recv_cq = cq;
827 rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr);
829 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d\n",
830 rc, init_qp_attr->cap.max_send_wr,
831 init_qp_attr->cap.max_recv_wr);
835 LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
837 /* 1 ref for caller and each rxmsg */
838 cfs_atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(version));
839 conn->ibc_nrx = IBLND_RX_MSGS(version);
842 for (i = 0; i < IBLND_RX_MSGS(version); i++) {
843 rc = kiblnd_post_rx(&conn->ibc_rxs[i],
844 IBLND_POSTRX_NO_CREDIT);
846 CERROR("Can't post rxmsg: %d\n", rc);
848 /* Make posted receives complete */
849 kiblnd_abort_receives(conn);
851 /* correct # of posted buffers
852 * NB locking needed now I'm racing with completion */
853 cfs_spin_lock_irqsave(&kiblnd_data.kib_sched_lock,
855 conn->ibc_nrx -= IBLND_RX_MSGS(version) - i;
856 cfs_spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock,
859 /* cmid will be destroyed by CM(ofed) after cm_callback
860 * returned, so we can't refer it anymore
861 * (by kiblnd_connd()->kiblnd_destroy_conn) */
862 rdma_destroy_qp(conn->ibc_cmid);
863 conn->ibc_cmid = NULL;
865 /* Drop my own and unused rxbuffer refcounts */
866 while (i++ <= IBLND_RX_MSGS(version))
867 kiblnd_conn_decref(conn);
873 /* Init successful! */
874 LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
875 state == IBLND_CONN_PASSIVE_WAIT);
876 conn->ibc_state = state;
879 cfs_atomic_inc(&net->ibn_nconns);
883 kiblnd_destroy_conn(conn);
885 LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
891 kiblnd_destroy_conn (kib_conn_t *conn)
893 struct rdma_cm_id *cmid = conn->ibc_cmid;
894 kib_peer_t *peer = conn->ibc_peer;
897 LASSERT (!cfs_in_interrupt());
898 LASSERT (cfs_atomic_read(&conn->ibc_refcount) == 0);
899 LASSERT (cfs_list_empty(&conn->ibc_early_rxs));
900 LASSERT (cfs_list_empty(&conn->ibc_tx_noops));
901 LASSERT (cfs_list_empty(&conn->ibc_tx_queue));
902 LASSERT (cfs_list_empty(&conn->ibc_tx_queue_rsrvd));
903 LASSERT (cfs_list_empty(&conn->ibc_tx_queue_nocred));
904 LASSERT (cfs_list_empty(&conn->ibc_active_txs));
905 LASSERT (conn->ibc_noops_posted == 0);
906 LASSERT (conn->ibc_nsends_posted == 0);
908 switch (conn->ibc_state) {
910 /* conn must be completely disengaged from the network */
913 case IBLND_CONN_DISCONNECTED:
914 /* connvars should have been freed already */
915 LASSERT (conn->ibc_connvars == NULL);
918 case IBLND_CONN_INIT:
922 /* conn->ibc_cmid might be destroyed by CM already */
923 if (cmid != NULL && cmid->qp != NULL)
924 rdma_destroy_qp(cmid);
926 if (conn->ibc_cq != NULL) {
927 rc = ib_destroy_cq(conn->ibc_cq);
929 CWARN("Error destroying CQ: %d\n", rc);
932 if (conn->ibc_rx_pages != NULL)
933 kiblnd_unmap_rx_descs(conn);
935 if (conn->ibc_rxs != NULL) {
936 LIBCFS_FREE(conn->ibc_rxs,
937 IBLND_RX_MSGS(conn->ibc_version) * sizeof(kib_rx_t));
940 if (conn->ibc_connvars != NULL)
941 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
943 if (conn->ibc_hdev != NULL)
944 kiblnd_hdev_decref(conn->ibc_hdev);
946 /* See CAVEAT EMPTOR above in kiblnd_create_conn */
947 if (conn->ibc_state != IBLND_CONN_INIT) {
948 kib_net_t *net = peer->ibp_ni->ni_data;
950 kiblnd_peer_decref(peer);
951 rdma_destroy_id(cmid);
952 cfs_atomic_dec(&net->ibn_nconns);
955 LIBCFS_FREE(conn, sizeof(*conn));
959 kiblnd_close_peer_conns_locked (kib_peer_t *peer, int why)
966 cfs_list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
967 conn = cfs_list_entry(ctmp, kib_conn_t, ibc_list);
969 CDEBUG(D_NET, "Closing conn -> %s, "
970 "version: %x, reason: %d\n",
971 libcfs_nid2str(peer->ibp_nid),
972 conn->ibc_version, why);
974 kiblnd_close_conn_locked(conn, why);
982 kiblnd_close_stale_conns_locked (kib_peer_t *peer,
983 int version, __u64 incarnation)
990 cfs_list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
991 conn = cfs_list_entry(ctmp, kib_conn_t, ibc_list);
993 if (conn->ibc_version == version &&
994 conn->ibc_incarnation == incarnation)
997 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
998 "incarnation:"LPX64"(%x, "LPX64")\n",
999 libcfs_nid2str(peer->ibp_nid),
1000 conn->ibc_version, conn->ibc_incarnation,
1001 version, incarnation);
1003 kiblnd_close_conn_locked(conn, -ESTALE);
1011 kiblnd_close_matching_conns (lnet_ni_t *ni, lnet_nid_t nid)
1019 unsigned long flags;
1022 cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1024 if (nid != LNET_NID_ANY)
1025 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
1028 hi = kiblnd_data.kib_peer_hash_size - 1;
1031 for (i = lo; i <= hi; i++) {
1032 cfs_list_for_each_safe (ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
1034 peer = cfs_list_entry(ptmp, kib_peer_t, ibp_list);
1035 LASSERT (peer->ibp_connecting > 0 ||
1036 peer->ibp_accepting > 0 ||
1037 !cfs_list_empty(&peer->ibp_conns));
1039 if (peer->ibp_ni != ni)
1042 if (!(nid == LNET_NID_ANY || nid == peer->ibp_nid))
1045 count += kiblnd_close_peer_conns_locked(peer, 0);
1049 cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1051 /* wildcards always succeed */
1052 if (nid == LNET_NID_ANY)
1055 return (count == 0) ? -ENOENT : 0;
1059 kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
1061 struct libcfs_ioctl_data *data = arg;
1065 case IOC_LIBCFS_GET_PEER: {
1069 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1071 data->ioc_nid = nid;
1072 data->ioc_count = count;
1076 case IOC_LIBCFS_DEL_PEER: {
1077 rc = kiblnd_del_peer(ni, data->ioc_nid);
1080 case IOC_LIBCFS_GET_CONN: {
1084 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1090 LASSERT (conn->ibc_cmid != NULL);
1091 data->ioc_nid = conn->ibc_peer->ibp_nid;
1092 if (conn->ibc_cmid->route.path_rec == NULL)
1093 data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1096 ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1097 kiblnd_conn_decref(conn);
1100 case IOC_LIBCFS_CLOSE_CONNECTION: {
1101 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1113 kiblnd_query (lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when)
1115 cfs_time_t last_alive = 0;
1116 cfs_time_t now = cfs_time_current();
1117 cfs_rwlock_t *glock = &kiblnd_data.kib_global_lock;
1119 unsigned long flags;
1121 cfs_read_lock_irqsave(glock, flags);
1123 peer = kiblnd_find_peer_locked(nid);
1125 LASSERT (peer->ibp_connecting > 0 || /* creating conns */
1126 peer->ibp_accepting > 0 ||
1127 !cfs_list_empty(&peer->ibp_conns)); /* active conn */
1128 last_alive = peer->ibp_last_alive;
1131 cfs_read_unlock_irqrestore(glock, flags);
1133 if (last_alive != 0)
1136 /* peer is not persistent in hash, trigger peer creation
1137 * and connection establishment with a NULL tx */
1139 kiblnd_launch_tx(ni, NULL, nid);
1141 CDEBUG(D_NET, "Peer %s %p, alive %ld secs ago\n",
1142 libcfs_nid2str(nid), peer,
1143 last_alive ? cfs_duration_sec(now - last_alive) : -1);
1148 kiblnd_free_pages (kib_pages_t *p)
1150 int npages = p->ibp_npages;
1153 for (i = 0; i < npages; i++)
1154 if (p->ibp_pages[i] != NULL)
1155 __free_page(p->ibp_pages[i]);
1157 LIBCFS_FREE (p, offsetof(kib_pages_t, ibp_pages[npages]));
1161 kiblnd_alloc_pages (kib_pages_t **pp, int npages)
1166 LIBCFS_ALLOC(p, offsetof(kib_pages_t, ibp_pages[npages]));
1168 CERROR("Can't allocate descriptor for %d pages\n", npages);
1172 memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1173 p->ibp_npages = npages;
1175 for (i = 0; i < npages; i++) {
1176 p->ibp_pages[i] = alloc_page(GFP_KERNEL);
1177 if (p->ibp_pages[i] == NULL) {
1178 CERROR("Can't allocate page %d of %d\n", i, npages);
1179 kiblnd_free_pages(p);
1189 kiblnd_unmap_rx_descs(kib_conn_t *conn)
1194 LASSERT (conn->ibc_rxs != NULL);
1195 LASSERT (conn->ibc_hdev != NULL);
1197 for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1198 rx = &conn->ibc_rxs[i];
1200 LASSERT (rx->rx_nob >= 0); /* not posted */
1202 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1203 KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1205 IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1208 kiblnd_free_pages(conn->ibc_rx_pages);
1210 conn->ibc_rx_pages = NULL;
1214 kiblnd_map_rx_descs(kib_conn_t *conn)
1222 for (pg_off = ipg = i = 0;
1223 i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1224 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1225 rx = &conn->ibc_rxs[i];
1228 rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
1230 rx->rx_msgaddr = kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1231 rx->rx_msg, IBLND_MSG_SIZE,
1233 LASSERT (!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1235 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1237 CDEBUG(D_NET,"rx %d: %p "LPX64"("LPX64")\n",
1238 i, rx->rx_msg, rx->rx_msgaddr,
1239 lnet_page2phys(pg) + pg_off);
1241 pg_off += IBLND_MSG_SIZE;
1242 LASSERT (pg_off <= PAGE_SIZE);
1244 if (pg_off == PAGE_SIZE) {
1247 LASSERT (ipg <= IBLND_RX_MSG_PAGES(conn->ibc_version));
1253 kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo)
1255 kib_hca_dev_t *hdev = tpo->tpo_hdev;
1259 LASSERT (tpo->tpo_pool.po_allocated == 0);
1264 for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1265 tx = &tpo->tpo_tx_descs[i];
1266 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1267 KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1269 IBLND_MSG_SIZE, DMA_TO_DEVICE);
1272 kiblnd_hdev_decref(hdev);
1273 tpo->tpo_hdev = NULL;
1276 static kib_hca_dev_t *
1277 kiblnd_current_hdev(kib_dev_t *dev)
1279 kib_hca_dev_t *hdev;
1280 unsigned long flags;
1283 cfs_read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1284 while (dev->ibd_failover) {
1285 cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1287 CDEBUG(D_NET, "Wait for dev(%s) failover\n", dev->ibd_ifname);
1288 cfs_schedule_timeout(cfs_time_seconds(1) / 100);
1290 cfs_read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1293 kiblnd_hdev_addref_locked(dev->ibd_hdev);
1294 hdev = dev->ibd_hdev;
1296 cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1302 kiblnd_map_tx_pool(kib_tx_pool_t *tpo)
1304 kib_pages_t *txpgs = tpo->tpo_tx_pages;
1305 kib_pool_t *pool = &tpo->tpo_pool;
1306 kib_net_t *net = pool->po_owner->ps_net;
1307 kib_dev_t *dev = net->ibn_dev;
1314 LASSERT (net != NULL);
1316 /* pre-mapped messages are not bigger than 1 page */
1317 CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE);
1319 /* No fancy arithmetic when we do the buffer calculations */
1320 CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
1322 tpo->tpo_hdev = kiblnd_current_hdev(dev);
1324 for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1325 page = txpgs->ibp_pages[ipage];
1326 tx = &tpo->tpo_tx_descs[i];
1328 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1331 tx->tx_msgaddr = kiblnd_dma_map_single(
1332 tpo->tpo_hdev->ibh_ibdev, tx->tx_msg,
1333 IBLND_MSG_SIZE, DMA_TO_DEVICE);
1334 LASSERT (!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1336 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1338 cfs_list_add(&tx->tx_list, &pool->po_free_list);
1340 page_offset += IBLND_MSG_SIZE;
1341 LASSERT (page_offset <= PAGE_SIZE);
1343 if (page_offset == PAGE_SIZE) {
1346 LASSERT (ipage <= txpgs->ibp_npages);
1352 kiblnd_find_dma_mr(kib_hca_dev_t *hdev, __u64 addr, __u64 size)
1356 LASSERT (hdev->ibh_mrs[0] != NULL);
1358 if (hdev->ibh_nmrs == 1)
1359 return hdev->ibh_mrs[0];
1361 index = addr >> hdev->ibh_mr_shift;
1363 if (index < hdev->ibh_nmrs &&
1364 index == ((addr + size - 1) >> hdev->ibh_mr_shift))
1365 return hdev->ibh_mrs[index];
1371 kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev, kib_rdma_desc_t *rd)
1373 struct ib_mr *prev_mr;
1377 LASSERT (hdev->ibh_mrs[0] != NULL);
1379 if (*kiblnd_tunables.kib_map_on_demand > 0 &&
1380 *kiblnd_tunables.kib_map_on_demand <= rd->rd_nfrags)
1383 if (hdev->ibh_nmrs == 1)
1384 return hdev->ibh_mrs[0];
1386 for (i = 0, mr = prev_mr = NULL;
1387 i < rd->rd_nfrags; i++) {
1388 mr = kiblnd_find_dma_mr(hdev,
1389 rd->rd_frags[i].rf_addr,
1390 rd->rd_frags[i].rf_nob);
1391 if (prev_mr == NULL)
1394 if (mr == NULL || prev_mr != mr) {
1395 /* Can't covered by one single MR */
1405 kiblnd_destroy_fmr_pool(kib_fmr_pool_t *pool)
1407 LASSERT (pool->fpo_map_count == 0);
1409 if (pool->fpo_fmr_pool != NULL)
1410 ib_destroy_fmr_pool(pool->fpo_fmr_pool);
1412 if (pool->fpo_hdev != NULL)
1413 kiblnd_hdev_decref(pool->fpo_hdev);
1415 LIBCFS_FREE(pool, sizeof(kib_fmr_pool_t));
1419 kiblnd_destroy_fmr_pool_list(cfs_list_t *head)
1421 kib_fmr_pool_t *pool;
1423 while (!cfs_list_empty(head)) {
1424 pool = cfs_list_entry(head->next, kib_fmr_pool_t, fpo_list);
1425 cfs_list_del(&pool->fpo_list);
1426 kiblnd_destroy_fmr_pool(pool);
1431 kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo)
1433 /* FMR pool for RDMA */
1434 kib_dev_t *dev = fps->fps_net->ibn_dev;
1435 kib_fmr_pool_t *fpo;
1436 struct ib_fmr_pool_param param = {
1437 .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
1438 .page_shift = PAGE_SHIFT,
1439 .access = (IB_ACCESS_LOCAL_WRITE |
1440 IB_ACCESS_REMOTE_WRITE),
1441 .pool_size = *kiblnd_tunables.kib_fmr_pool_size,
1442 .dirty_watermark = *kiblnd_tunables.kib_fmr_flush_trigger,
1443 .flush_function = NULL,
1445 .cache = !!*kiblnd_tunables.kib_fmr_cache};
1448 LIBCFS_ALLOC(fpo, sizeof(kib_fmr_pool_t));
1452 memset(fpo, 0, sizeof(kib_fmr_pool_t));
1453 fpo->fpo_hdev = kiblnd_current_hdev(dev);
1454 fpo->fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd, ¶m);
1455 if (IS_ERR(fpo->fpo_fmr_pool)) {
1456 CERROR("Failed to create FMR pool: %ld\n",
1457 PTR_ERR(fpo->fpo_fmr_pool));
1458 rc = PTR_ERR(fpo->fpo_fmr_pool);
1459 kiblnd_hdev_decref(fpo->fpo_hdev);
1460 LIBCFS_FREE(fpo, sizeof(kib_fmr_pool_t));
1464 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1465 fpo->fpo_owner = fps;
1472 kiblnd_fail_fmr_pool_set(kib_fmr_poolset_t *fps, cfs_list_t *zombies)
1474 cfs_spin_lock(&fps->fps_lock);
1476 while (!cfs_list_empty(&fps->fps_pool_list)) {
1477 kib_fmr_pool_t *fpo = cfs_list_entry(fps->fps_pool_list.next,
1478 kib_fmr_pool_t, fpo_list);
1479 fpo->fpo_failed = 1;
1480 cfs_list_del(&fpo->fpo_list);
1481 if (fpo->fpo_map_count == 0)
1482 cfs_list_add(&fpo->fpo_list, zombies);
1484 cfs_list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1487 cfs_spin_unlock(&fps->fps_lock);
1491 kiblnd_fini_fmr_pool_set(kib_fmr_poolset_t *fps)
1493 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1494 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1498 kiblnd_init_fmr_pool_set(kib_fmr_poolset_t *fps, kib_net_t *net)
1500 kib_fmr_pool_t *fpo;
1503 memset(fps, 0, sizeof(kib_fmr_poolset_t));
1506 cfs_spin_lock_init(&fps->fps_lock);
1507 CFS_INIT_LIST_HEAD(&fps->fps_pool_list);
1508 CFS_INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1509 rc = kiblnd_create_fmr_pool(fps, &fpo);
1511 cfs_list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1517 kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, cfs_time_t now)
1519 if (fpo->fpo_map_count != 0) /* still in use */
1521 if (fpo->fpo_failed)
1523 return cfs_time_aftereq(now, fpo->fpo_deadline);
1527 kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1529 CFS_LIST_HEAD (zombies);
1530 kib_fmr_pool_t *fpo = fmr->fmr_pool;
1531 kib_fmr_poolset_t *fps = fpo->fpo_owner;
1532 cfs_time_t now = cfs_time_current();
1533 kib_fmr_pool_t *tmp;
1536 rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1540 rc = ib_flush_fmr_pool(fpo->fpo_fmr_pool);
1544 fmr->fmr_pool = NULL;
1545 fmr->fmr_pfmr = NULL;
1547 cfs_spin_lock(&fps->fps_lock);
1548 fpo->fpo_map_count --; /* decref the pool */
1550 cfs_list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1551 /* the first pool is persistent */
1552 if (fps->fps_pool_list.next == &fpo->fpo_list)
1555 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1556 cfs_list_move(&fpo->fpo_list, &zombies);
1557 fps->fps_version ++;
1560 cfs_spin_unlock(&fps->fps_lock);
1562 if (!cfs_list_empty(&zombies))
1563 kiblnd_destroy_fmr_pool_list(&zombies);
1567 kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages,
1568 __u64 iov, kib_fmr_t *fmr)
1570 struct ib_pool_fmr *pfmr;
1571 kib_fmr_pool_t *fpo;
1575 LASSERT (fps->fps_net->ibn_with_fmr);
1577 cfs_spin_lock(&fps->fps_lock);
1578 version = fps->fps_version;
1579 cfs_list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1580 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1581 fpo->fpo_map_count ++;
1582 cfs_spin_unlock(&fps->fps_lock);
1584 pfmr = ib_fmr_pool_map_phys(fpo->fpo_fmr_pool,
1585 pages, npages, iov);
1586 if (likely(!IS_ERR(pfmr))) {
1587 fmr->fmr_pool = fpo;
1588 fmr->fmr_pfmr = pfmr;
1592 cfs_spin_lock(&fps->fps_lock);
1593 fpo->fpo_map_count --;
1594 if (PTR_ERR(pfmr) != -EAGAIN) {
1595 cfs_spin_unlock(&fps->fps_lock);
1596 return PTR_ERR(pfmr);
1599 /* EAGAIN and ... */
1600 if (version != fps->fps_version) {
1601 cfs_spin_unlock(&fps->fps_lock);
1606 if (fps->fps_increasing) {
1607 cfs_spin_unlock(&fps->fps_lock);
1608 CDEBUG(D_NET, "Another thread is allocating new "
1609 "FMR pool, waiting for her to complete\n");
1615 if (cfs_time_before(cfs_time_current(), fps->fps_next_retry)) {
1616 /* someone failed recently */
1617 cfs_spin_unlock(&fps->fps_lock);
1621 fps->fps_increasing = 1;
1622 cfs_spin_unlock(&fps->fps_lock);
1624 CDEBUG(D_NET, "Allocate new FMR pool\n");
1625 rc = kiblnd_create_fmr_pool(fps, &fpo);
1626 cfs_spin_lock(&fps->fps_lock);
1627 fps->fps_increasing = 0;
1629 fps->fps_version ++;
1630 cfs_list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1632 fps->fps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1634 cfs_spin_unlock(&fps->fps_lock);
1640 kiblnd_fini_pool(kib_pool_t *pool)
1642 LASSERT (cfs_list_empty(&pool->po_free_list));
1643 LASSERT (pool->po_allocated == 0);
1645 CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
1649 kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
1651 CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
1653 memset(pool, 0, sizeof(kib_pool_t));
1654 CFS_INIT_LIST_HEAD(&pool->po_free_list);
1655 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1656 pool->po_owner = ps;
1657 pool->po_size = size;
1661 kiblnd_destroy_pool_list(cfs_list_t *head)
1665 while (!cfs_list_empty(head)) {
1666 pool = cfs_list_entry(head->next, kib_pool_t, po_list);
1667 cfs_list_del(&pool->po_list);
1669 LASSERT (pool->po_owner != NULL);
1670 pool->po_owner->ps_pool_destroy(pool);
1675 kiblnd_fail_pool_set(kib_poolset_t *ps, cfs_list_t *zombies)
1677 cfs_spin_lock(&ps->ps_lock);
1678 while (!cfs_list_empty(&ps->ps_pool_list)) {
1679 kib_pool_t *po = cfs_list_entry(ps->ps_pool_list.next,
1680 kib_pool_t, po_list);
1682 cfs_list_del(&po->po_list);
1683 if (po->po_allocated == 0)
1684 cfs_list_add(&po->po_list, zombies);
1686 cfs_list_add(&po->po_list, &ps->ps_failed_pool_list);
1688 cfs_spin_unlock(&ps->ps_lock);
1692 kiblnd_fini_pool_set(kib_poolset_t *ps)
1694 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
1695 kiblnd_destroy_pool_list(&ps->ps_pool_list);
1699 kiblnd_init_pool_set(kib_poolset_t *ps, kib_net_t *net,
1700 char *name, int size,
1701 kib_ps_pool_create_t po_create,
1702 kib_ps_pool_destroy_t po_destroy,
1703 kib_ps_node_init_t nd_init,
1704 kib_ps_node_fini_t nd_fini)
1709 memset(ps, 0, sizeof(kib_poolset_t));
1712 ps->ps_pool_create = po_create;
1713 ps->ps_pool_destroy = po_destroy;
1714 ps->ps_node_init = nd_init;
1715 ps->ps_node_fini = nd_fini;
1716 ps->ps_pool_size = size;
1717 strncpy(ps->ps_name, name, IBLND_POOL_NAME_LEN);
1718 cfs_spin_lock_init(&ps->ps_lock);
1719 CFS_INIT_LIST_HEAD(&ps->ps_pool_list);
1720 CFS_INIT_LIST_HEAD(&ps->ps_failed_pool_list);
1722 rc = ps->ps_pool_create(ps, size, &pool);
1724 cfs_list_add(&pool->po_list, &ps->ps_pool_list);
1726 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
1732 kiblnd_pool_is_idle(kib_pool_t *pool, cfs_time_t now)
1734 if (pool->po_allocated != 0) /* still in use */
1736 if (pool->po_failed)
1738 return cfs_time_aftereq(now, pool->po_deadline);
1742 kiblnd_pool_free_node(kib_pool_t *pool, cfs_list_t *node)
1744 CFS_LIST_HEAD (zombies);
1745 kib_poolset_t *ps = pool->po_owner;
1747 cfs_time_t now = cfs_time_current();
1749 cfs_spin_lock(&ps->ps_lock);
1751 if (ps->ps_node_fini != NULL)
1752 ps->ps_node_fini(pool, node);
1754 LASSERT (pool->po_allocated > 0);
1755 cfs_list_add(node, &pool->po_free_list);
1756 pool->po_allocated --;
1758 cfs_list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
1759 /* the first pool is persistent */
1760 if (ps->ps_pool_list.next == &pool->po_list)
1763 if (kiblnd_pool_is_idle(pool, now))
1764 cfs_list_move(&pool->po_list, &zombies);
1766 cfs_spin_unlock(&ps->ps_lock);
1768 if (!cfs_list_empty(&zombies))
1769 kiblnd_destroy_pool_list(&zombies);
1773 kiblnd_pool_alloc_node(kib_poolset_t *ps)
1780 cfs_spin_lock(&ps->ps_lock);
1781 cfs_list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
1782 if (cfs_list_empty(&pool->po_free_list))
1785 pool->po_allocated ++;
1786 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1787 node = pool->po_free_list.next;
1790 if (ps->ps_node_init != NULL) {
1791 /* still hold the lock */
1792 ps->ps_node_init(pool, node);
1794 cfs_spin_unlock(&ps->ps_lock);
1798 /* no available tx pool and ... */
1799 if (ps->ps_increasing) {
1800 /* another thread is allocating a new pool */
1801 cfs_spin_unlock(&ps->ps_lock);
1802 CDEBUG(D_NET, "Another thread is allocating new "
1803 "%s pool, waiting for her to complete\n",
1809 if (cfs_time_before(cfs_time_current(), ps->ps_next_retry)) {
1810 /* someone failed recently */
1811 cfs_spin_unlock(&ps->ps_lock);
1815 ps->ps_increasing = 1;
1816 cfs_spin_unlock(&ps->ps_lock);
1818 CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
1820 rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
1822 cfs_spin_lock(&ps->ps_lock);
1823 ps->ps_increasing = 0;
1825 cfs_list_add_tail(&pool->po_list, &ps->ps_pool_list);
1827 ps->ps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1828 CERROR("Can't allocate new %s pool because out of memory\n",
1831 cfs_spin_unlock(&ps->ps_lock);
1837 kiblnd_pmr_pool_unmap(kib_phys_mr_t *pmr)
1839 kib_pmr_pool_t *ppo = pmr->pmr_pool;
1840 struct ib_mr *mr = pmr->pmr_mr;
1843 kiblnd_pool_free_node(&ppo->ppo_pool, &pmr->pmr_list);
1849 kiblnd_pmr_pool_map(kib_pmr_poolset_t *pps, kib_hca_dev_t *hdev,
1850 kib_rdma_desc_t *rd, __u64 *iova, kib_phys_mr_t **pp_pmr)
1857 node = kiblnd_pool_alloc_node(&pps->pps_poolset);
1859 CERROR("Failed to allocate PMR descriptor\n");
1863 pmr = container_of(node, kib_phys_mr_t, pmr_list);
1864 if (pmr->pmr_pool->ppo_hdev != hdev) {
1865 kiblnd_pool_free_node(&pmr->pmr_pool->ppo_pool, node);
1869 for (i = 0; i < rd->rd_nfrags; i ++) {
1870 pmr->pmr_ipb[i].addr = rd->rd_frags[i].rf_addr;
1871 pmr->pmr_ipb[i].size = rd->rd_frags[i].rf_nob;
1874 pmr->pmr_mr = ib_reg_phys_mr(hdev->ibh_pd,
1875 pmr->pmr_ipb, rd->rd_nfrags,
1876 IB_ACCESS_LOCAL_WRITE |
1877 IB_ACCESS_REMOTE_WRITE,
1879 if (!IS_ERR(pmr->pmr_mr)) {
1880 pmr->pmr_iova = *iova;
1885 rc = PTR_ERR(pmr->pmr_mr);
1886 CERROR("Failed ib_reg_phys_mr: %d\n", rc);
1889 kiblnd_pool_free_node(&pmr->pmr_pool->ppo_pool, node);
1895 kiblnd_destroy_pmr_pool(kib_pool_t *pool)
1897 kib_pmr_pool_t *ppo = container_of(pool, kib_pmr_pool_t, ppo_pool);
1900 LASSERT (pool->po_allocated == 0);
1902 while (!cfs_list_empty(&pool->po_free_list)) {
1903 pmr = cfs_list_entry(pool->po_free_list.next,
1904 kib_phys_mr_t, pmr_list);
1906 LASSERT (pmr->pmr_mr == NULL);
1907 cfs_list_del(&pmr->pmr_list);
1909 if (pmr->pmr_ipb != NULL) {
1910 LIBCFS_FREE(pmr->pmr_ipb,
1911 IBLND_MAX_RDMA_FRAGS *
1912 sizeof(struct ib_phys_buf));
1915 LIBCFS_FREE(pmr, sizeof(kib_phys_mr_t));
1918 kiblnd_fini_pool(pool);
1919 if (ppo->ppo_hdev != NULL)
1920 kiblnd_hdev_decref(ppo->ppo_hdev);
1922 LIBCFS_FREE(ppo, sizeof(kib_pmr_pool_t));
1926 kiblnd_create_pmr_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
1928 kib_pmr_pool_t *ppo;
1933 LIBCFS_ALLOC(ppo, sizeof(kib_pmr_pool_t));
1935 CERROR("Failed to allocate PMR pool\n");
1939 pool = &ppo->ppo_pool;
1940 kiblnd_init_pool(ps, pool, size);
1942 for (i = 0; i < size; i++) {
1943 LIBCFS_ALLOC(pmr, sizeof(kib_phys_mr_t));
1947 memset(pmr, 0, sizeof(kib_phys_mr_t));
1948 pmr->pmr_pool = ppo;
1949 LIBCFS_ALLOC(pmr->pmr_ipb,
1950 IBLND_MAX_RDMA_FRAGS *
1951 sizeof(struct ib_phys_buf));
1952 if (pmr->pmr_ipb == NULL)
1955 cfs_list_add(&pmr->pmr_list, &pool->po_free_list);
1959 ps->ps_pool_destroy(pool);
1963 ppo->ppo_hdev = kiblnd_current_hdev(ps->ps_net->ibn_dev);
1969 kiblnd_destroy_tx_pool(kib_pool_t *pool)
1971 kib_tx_pool_t *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
1974 LASSERT (pool->po_allocated == 0);
1976 if (tpo->tpo_tx_pages != NULL) {
1977 kiblnd_unmap_tx_pool(tpo);
1978 kiblnd_free_pages(tpo->tpo_tx_pages);
1981 if (tpo->tpo_tx_descs == NULL)
1984 for (i = 0; i < pool->po_size; i++) {
1985 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
1987 cfs_list_del(&tx->tx_list);
1988 if (tx->tx_pages != NULL)
1989 LIBCFS_FREE(tx->tx_pages,
1991 sizeof(*tx->tx_pages));
1992 if (tx->tx_frags != NULL)
1993 LIBCFS_FREE(tx->tx_frags,
1994 IBLND_MAX_RDMA_FRAGS *
1995 sizeof(*tx->tx_frags));
1996 if (tx->tx_wrq != NULL)
1997 LIBCFS_FREE(tx->tx_wrq,
1998 (1 + IBLND_MAX_RDMA_FRAGS) *
1999 sizeof(*tx->tx_wrq));
2000 if (tx->tx_sge != NULL)
2001 LIBCFS_FREE(tx->tx_sge,
2002 (1 + IBLND_MAX_RDMA_FRAGS) *
2003 sizeof(*tx->tx_sge));
2004 if (tx->tx_rd != NULL)
2005 LIBCFS_FREE(tx->tx_rd,
2006 offsetof(kib_rdma_desc_t,
2007 rd_frags[IBLND_MAX_RDMA_FRAGS]));
2010 LIBCFS_FREE(tpo->tpo_tx_descs,
2011 pool->po_size * sizeof(kib_tx_t));
2013 kiblnd_fini_pool(pool);
2014 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2018 kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
2025 LIBCFS_ALLOC(tpo, sizeof(kib_tx_pool_t));
2027 CERROR("Failed to allocate TX pool\n");
2031 pool = &tpo->tpo_pool;
2032 kiblnd_init_pool(ps, pool, size);
2033 tpo->tpo_tx_descs = NULL;
2034 tpo->tpo_tx_pages = NULL;
2036 npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2037 if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, npg) != 0) {
2038 CERROR("Can't allocate tx pages: %d\n", npg);
2039 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2043 LIBCFS_ALLOC (tpo->tpo_tx_descs, size * sizeof(kib_tx_t));
2044 if (tpo->tpo_tx_descs == NULL) {
2045 CERROR("Can't allocate %d tx descriptors\n", size);
2046 ps->ps_pool_destroy(pool);
2050 memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
2052 for (i = 0; i < size; i++) {
2053 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2056 if (ps->ps_net->ibn_with_fmr){
2057 LIBCFS_ALLOC(tx->tx_pages, LNET_MAX_IOV *
2058 sizeof(*tx->tx_pages));
2059 if (tx->tx_pages == NULL)
2063 LIBCFS_ALLOC(tx->tx_frags,
2064 IBLND_MAX_RDMA_FRAGS *
2065 sizeof(*tx->tx_frags));
2066 if (tx->tx_frags == NULL)
2069 LIBCFS_ALLOC(tx->tx_wrq,
2070 (1 + IBLND_MAX_RDMA_FRAGS) *
2071 sizeof(*tx->tx_wrq));
2072 if (tx->tx_wrq == NULL)
2075 LIBCFS_ALLOC(tx->tx_sge,
2076 (1 + IBLND_MAX_RDMA_FRAGS) *
2077 sizeof(*tx->tx_sge));
2078 if (tx->tx_sge == NULL)
2081 LIBCFS_ALLOC(tx->tx_rd,
2082 offsetof(kib_rdma_desc_t,
2083 rd_frags[IBLND_MAX_RDMA_FRAGS]));
2084 if (tx->tx_rd == NULL)
2089 kiblnd_map_tx_pool(tpo);
2094 ps->ps_pool_destroy(pool);
2099 kiblnd_tx_init(kib_pool_t *pool, cfs_list_t *node)
2101 kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t,
2103 kib_tx_t *tx = cfs_list_entry(node, kib_tx_t, tx_list);
2105 tx->tx_cookie = tps->tps_next_tx_cookie ++;
2109 kiblnd_ni_fini_pools(kib_net_t *net)
2111 kiblnd_fini_pool_set(&net->ibn_tx_ps.tps_poolset);
2112 if (net->ibn_with_fmr)
2113 kiblnd_fini_fmr_pool_set(&net->ibn_fmr_ps);
2114 else if (net->ibn_with_pmr)
2115 kiblnd_fini_pool_set(&net->ibn_pmr_ps.pps_poolset);
2119 kiblnd_net_init_pools(kib_net_t *net)
2121 kib_fmr_poolset_t *fps = &net->ibn_fmr_ps;
2122 kib_pmr_poolset_t *pps = &net->ibn_pmr_ps;
2123 kib_tx_poolset_t *tps = &net->ibn_tx_ps;
2124 unsigned long flags;
2127 if (*kiblnd_tunables.kib_fmr_pool_size <
2128 *kiblnd_tunables.kib_ntx / 4) {
2129 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2130 *kiblnd_tunables.kib_fmr_pool_size,
2131 *kiblnd_tunables.kib_ntx / 4);
2135 if (*kiblnd_tunables.kib_pmr_pool_size <
2136 *kiblnd_tunables.kib_ntx / 4) {
2137 CERROR("Can't set pmr pool size (%d) < ntx / 4(%d)\n",
2138 *kiblnd_tunables.kib_pmr_pool_size,
2139 *kiblnd_tunables.kib_ntx / 4);
2143 cfs_read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2144 if (*kiblnd_tunables.kib_map_on_demand > 0 ||
2145 net->ibn_dev->ibd_hdev->ibh_nmrs > 1) {
2146 /* premapping can fail if ibd_nmr > 1, so we always create
2147 * FMR/PMR pool and map-on-demand if premapping failed */
2148 cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2149 rc = kiblnd_init_fmr_pool_set(fps, net);
2151 net->ibn_with_fmr = 1;
2152 } else if (rc == -ENOSYS) {
2153 rc = kiblnd_init_pool_set(&pps->pps_poolset, net, "PMR",
2154 *kiblnd_tunables.kib_pmr_pool_size,
2155 kiblnd_create_pmr_pool,
2156 kiblnd_destroy_pmr_pool,
2159 net->ibn_with_pmr = 1;
2164 cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2167 rc = kiblnd_init_pool_set(&tps->tps_poolset, net, "TX", IBLND_TX_MSGS(),
2168 kiblnd_create_tx_pool, kiblnd_destroy_tx_pool,
2169 kiblnd_tx_init, NULL);
2173 if (net->ibn_with_fmr)
2174 kiblnd_fini_fmr_pool_set(fps);
2175 else if (net->ibn_with_pmr)
2176 kiblnd_fini_pool_set(&pps->pps_poolset);
2182 kiblnd_hdev_get_attr(kib_hca_dev_t *hdev)
2184 struct ib_device_attr *attr;
2187 /* It's safe to assume a HCA can handle a page size
2188 * matching that of the native system */
2189 hdev->ibh_page_shift = PAGE_SHIFT;
2190 hdev->ibh_page_size = 1 << PAGE_SHIFT;
2191 hdev->ibh_page_mask = ~((__u64)hdev->ibh_page_size - 1);
2193 LIBCFS_ALLOC(attr, sizeof(*attr));
2195 CERROR("Out of memory\n");
2199 rc = ib_query_device(hdev->ibh_ibdev, attr);
2201 hdev->ibh_mr_size = attr->max_mr_size;
2203 LIBCFS_FREE(attr, sizeof(*attr));
2206 CERROR("Failed to query IB device: %d\n", rc);
2210 if (hdev->ibh_mr_size == ~0ULL) {
2211 hdev->ibh_mr_shift = 64;
2215 for (hdev->ibh_mr_shift = 0;
2216 hdev->ibh_mr_shift < 64; hdev->ibh_mr_shift ++) {
2217 if (hdev->ibh_mr_size == (1ULL << hdev->ibh_mr_shift) ||
2218 hdev->ibh_mr_size == (1ULL << hdev->ibh_mr_shift) - 1)
2222 CERROR("Invalid mr size: "LPX64"\n", hdev->ibh_mr_size);
2227 kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev)
2231 if (hdev->ibh_nmrs == 0 || hdev->ibh_mrs == NULL)
2234 for (i = 0; i < hdev->ibh_nmrs; i++) {
2235 if (hdev->ibh_mrs[i] == NULL)
2238 ib_dereg_mr(hdev->ibh_mrs[i]);
2241 LIBCFS_FREE(hdev->ibh_mrs, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2242 hdev->ibh_mrs = NULL;
2247 kiblnd_hdev_destroy(kib_hca_dev_t *hdev)
2249 kiblnd_hdev_cleanup_mrs(hdev);
2251 if (hdev->ibh_pd != NULL)
2252 ib_dealloc_pd(hdev->ibh_pd);
2254 if (hdev->ibh_cmid != NULL)
2255 rdma_destroy_id(hdev->ibh_cmid);
2257 LIBCFS_FREE(hdev, sizeof(*hdev));
2261 kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev)
2268 int acflags = IB_ACCESS_LOCAL_WRITE |
2269 IB_ACCESS_REMOTE_WRITE;
2271 rc = kiblnd_hdev_get_attr(hdev);
2275 if (hdev->ibh_mr_shift == 64) {
2276 LIBCFS_ALLOC(hdev->ibh_mrs, 1 * sizeof(*hdev->ibh_mrs));
2277 if (hdev->ibh_mrs == NULL) {
2278 CERROR("Failed to allocate MRs table\n");
2282 hdev->ibh_mrs[0] = NULL;
2285 mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2287 CERROR("Failed ib_get_dma_mr : %ld\n", PTR_ERR(mr));
2288 kiblnd_hdev_cleanup_mrs(hdev);
2292 hdev->ibh_mrs[0] = mr;
2297 mr_size = (1ULL << hdev->ibh_mr_shift);
2298 mm_size = (unsigned long)high_memory - PAGE_OFFSET;
2300 hdev->ibh_nmrs = (int)((mm_size + mr_size - 1) >> hdev->ibh_mr_shift);
2302 if (hdev->ibh_mr_shift < 32 || hdev->ibh_nmrs > 1024) {
2303 /* it's 4T..., assume we will re-code at that time */
2304 CERROR("Can't support memory size: x"LPX64
2305 " with MR size: x"LPX64"\n", mm_size, mr_size);
2309 /* create an array of MRs to cover all memory */
2310 LIBCFS_ALLOC(hdev->ibh_mrs, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2311 if (hdev->ibh_mrs == NULL) {
2312 CERROR("Failed to allocate MRs' table\n");
2316 memset(hdev->ibh_mrs, 0, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2318 for (i = 0; i < hdev->ibh_nmrs; i++) {
2319 struct ib_phys_buf ipb;
2322 ipb.size = hdev->ibh_mr_size;
2323 ipb.addr = i * mr_size;
2326 mr = ib_reg_phys_mr(hdev->ibh_pd, &ipb, 1, acflags, &iova);
2328 CERROR("Failed ib_reg_phys_mr addr "LPX64
2329 " size "LPX64" : %ld\n",
2330 ipb.addr, ipb.size, PTR_ERR(mr));
2331 kiblnd_hdev_cleanup_mrs(hdev);
2335 LASSERT (iova == ipb.addr);
2337 hdev->ibh_mrs[i] = mr;
2341 if (hdev->ibh_mr_size != ~0ULL || hdev->ibh_nmrs != 1)
2342 LCONSOLE_INFO("Register global MR array, MR size: "
2343 LPX64", array size: %d\n",
2344 hdev->ibh_mr_size, hdev->ibh_nmrs);
2349 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2355 kiblnd_dev_need_failover(kib_dev_t *dev)
2357 struct rdma_cm_id *cmid;
2358 struct sockaddr_in srcaddr;
2359 struct sockaddr_in dstaddr;
2362 if (dev->ibd_hdev == NULL || /* initializing */
2363 dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2364 *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2367 /* XXX: it's UGLY, but I don't have better way to find
2368 * ib-bonding HCA failover because:
2370 * a. no reliable CM event for HCA failover...
2371 * b. no OFED API to get ib_device for current net_device...
2373 * We have only two choices at this point:
2375 * a. rdma_bind_addr(), it will conflict with listener cmid
2376 * b. rdma_resolve_addr() to zero addr */
2377 cmid = kiblnd_rdma_create_id(kiblnd_dummy_callback, dev, RDMA_PS_TCP,
2381 CERROR("Failed to create cmid for failover: %d\n", rc);
2385 memset(&srcaddr, 0, sizeof(srcaddr));
2386 srcaddr.sin_family = AF_INET;
2387 srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2389 memset(&dstaddr, 0, sizeof(dstaddr));
2390 dstaddr.sin_family = AF_INET;
2391 rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2392 (struct sockaddr *)&dstaddr, 1);
2394 CERROR("Failed to bind %s to device: %d\n",
2395 dev->ibd_ifname, rc);
2396 rdma_destroy_id(cmid);
2400 LASSERT (cmid->device != NULL);
2401 if (dev->ibd_hdev->ibh_ibdev == cmid->device) {
2402 /* don't need device failover */
2403 rdma_destroy_id(cmid);
2411 kiblnd_dev_failover(kib_dev_t *dev)
2413 CFS_LIST_HEAD (zombie_tpo);
2414 CFS_LIST_HEAD (zombie_ppo);
2415 CFS_LIST_HEAD (zombie_fpo);
2416 struct rdma_cm_id *cmid = NULL;
2417 kib_hca_dev_t *hdev = NULL;
2421 struct sockaddr_in addr;
2422 unsigned long flags;
2425 LASSERT (*kiblnd_tunables.kib_dev_failover > 1 ||
2426 dev->ibd_can_failover ||
2427 dev->ibd_hdev == NULL);
2429 rc = kiblnd_dev_need_failover(dev);
2433 if (dev->ibd_hdev != NULL &&
2434 dev->ibd_hdev->ibh_cmid != NULL) {
2435 /* XXX it's not good to close old listener at here,
2436 * because we can fail to create new listener.
2437 * But we have to close it now, otherwise rdma_bind_addr
2438 * will return EADDRINUSE... How crap! */
2439 cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2441 cmid = dev->ibd_hdev->ibh_cmid;
2442 /* make next schedule of kiblnd_dev_need_failover
2443 * will return 1 for me */
2444 dev->ibd_hdev->ibh_cmid = NULL;
2445 cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2447 rdma_destroy_id(cmid);
2450 cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP,
2454 CERROR("Failed to create cmid for failover: %d\n", rc);
2458 memset(&addr, 0, sizeof(addr));
2459 addr.sin_family = AF_INET;
2460 addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2461 addr.sin_port = htons(*kiblnd_tunables.kib_service);
2463 /* Bind to failover device or port */
2464 rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2466 CERROR("Failed to bind %s to device: %d\n",
2467 dev->ibd_ifname, rc);
2468 rdma_destroy_id(cmid);
2472 LIBCFS_ALLOC(hdev, sizeof(*hdev));
2474 CERROR("Failed to allocate kib_hca_dev\n");
2475 rdma_destroy_id(cmid);
2479 memset(hdev, 0, sizeof(*hdev));
2480 atomic_set(&hdev->ibh_ref, 1);
2481 hdev->ibh_dev = dev;
2482 hdev->ibh_cmid = cmid;
2483 hdev->ibh_ibdev = cmid->device;
2485 pd = ib_alloc_pd(cmid->device);
2488 CERROR("Can't allocate PD: %d\n", rc);
2494 rc = rdma_listen(cmid, 0);
2496 CERROR("Can't start new listener: %d\n", rc);
2500 rc = kiblnd_hdev_setup_mrs(hdev);
2502 CERROR("Can't setup device: %d\n", rc);
2506 cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2508 old = dev->ibd_hdev;
2509 dev->ibd_hdev = hdev; /* take over the refcount */
2512 cfs_list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2513 kiblnd_fail_pool_set(&net->ibn_tx_ps.tps_poolset, &zombie_tpo);
2514 if (net->ibn_with_pmr)
2515 kiblnd_fail_pool_set(&net->ibn_pmr_ps.pps_poolset, &zombie_ppo);
2516 if (net->ibn_with_fmr)
2517 kiblnd_fail_fmr_pool_set(&net->ibn_fmr_ps, &zombie_fpo);
2520 cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2522 if (!cfs_list_empty(&zombie_tpo))
2523 kiblnd_destroy_pool_list(&zombie_tpo);
2524 if (!cfs_list_empty(&zombie_ppo))
2525 kiblnd_destroy_pool_list(&zombie_ppo);
2526 if (!cfs_list_empty(&zombie_fpo))
2527 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2529 kiblnd_hdev_decref(hdev);
2532 dev->ibd_failed_failover++;
2534 dev->ibd_failed_failover = 0;
2540 kiblnd_destroy_dev (kib_dev_t *dev)
2542 LASSERT (dev->ibd_nnets == 0);
2543 LASSERT (cfs_list_empty(&dev->ibd_nets));
2545 cfs_list_del(&dev->ibd_fail_list);
2546 cfs_list_del(&dev->ibd_list);
2548 if (dev->ibd_hdev != NULL)
2549 kiblnd_hdev_decref(dev->ibd_hdev);
2551 LIBCFS_FREE(dev, sizeof(*dev));
2555 kiblnd_create_dev(char *ifname)
2557 struct net_device *netdev;
2564 rc = libcfs_ipif_query(ifname, &up, &ip, &netmask);
2566 CERROR("Can't query IPoIB interface %s: %d\n",
2572 CERROR("Can't query IPoIB interface %s: it's down\n", ifname);
2576 LIBCFS_ALLOC(dev, sizeof(*dev));
2580 memset(dev, 0, sizeof(*dev));
2581 #ifdef HAVE_DEV_GET_BY_NAME_2ARG
2582 netdev = dev_get_by_name(&init_net, ifname);
2584 netdev = dev_get_by_name(ifname);
2586 if (netdev == NULL) {
2587 dev->ibd_can_failover = 0;
2589 dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER);
2593 CFS_INIT_LIST_HEAD(&dev->ibd_nets);
2594 CFS_INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */
2595 CFS_INIT_LIST_HEAD(&dev->ibd_fail_list);
2597 strcpy(&dev->ibd_ifname[0], ifname);
2599 /* initialize the device */
2600 rc = kiblnd_dev_failover(dev);
2602 CERROR("Can't initialize device: %d\n", rc);
2603 LIBCFS_FREE(dev, sizeof(*dev));
2607 cfs_list_add_tail(&dev->ibd_list,
2608 &kiblnd_data.kib_devs);
2613 kiblnd_base_shutdown (void)
2617 LASSERT (cfs_list_empty(&kiblnd_data.kib_devs));
2619 CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2620 cfs_atomic_read(&libcfs_kmemory));
2622 switch (kiblnd_data.kib_init) {
2626 case IBLND_INIT_ALL:
2627 case IBLND_INIT_DATA:
2628 LASSERT (kiblnd_data.kib_peers != NULL);
2629 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2630 LASSERT (cfs_list_empty(&kiblnd_data.kib_peers[i]));
2632 LASSERT (cfs_list_empty(&kiblnd_data.kib_connd_zombies));
2633 LASSERT (cfs_list_empty(&kiblnd_data.kib_connd_conns));
2635 /* flag threads to terminate; wake and wait for them to die */
2636 kiblnd_data.kib_shutdown = 1;
2637 cfs_waitq_broadcast(&kiblnd_data.kib_sched_waitq);
2638 cfs_waitq_broadcast(&kiblnd_data.kib_connd_waitq);
2639 cfs_waitq_broadcast(&kiblnd_data.kib_failover_waitq);
2642 while (cfs_atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2644 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
2645 "Waiting for %d threads to terminate\n",
2646 cfs_atomic_read(&kiblnd_data.kib_nthreads));
2647 cfs_pause(cfs_time_seconds(1));
2652 case IBLND_INIT_NOTHING:
2656 if (kiblnd_data.kib_peers != NULL)
2657 LIBCFS_FREE(kiblnd_data.kib_peers,
2658 sizeof(cfs_list_t) *
2659 kiblnd_data.kib_peer_hash_size);
2661 CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2662 cfs_atomic_read(&libcfs_kmemory));
2664 kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2665 PORTAL_MODULE_UNUSE;
2669 kiblnd_shutdown (lnet_ni_t *ni)
2671 kib_net_t *net = ni->ni_data;
2672 cfs_rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2674 unsigned long flags;
2676 LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2681 CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
2682 cfs_atomic_read(&libcfs_kmemory));
2684 cfs_write_lock_irqsave(g_lock, flags);
2685 net->ibn_shutdown = 1;
2686 cfs_write_unlock_irqrestore(g_lock, flags);
2688 switch (net->ibn_init) {
2692 case IBLND_INIT_ALL:
2693 /* nuke all existing peers within this net */
2694 kiblnd_del_peer(ni, LNET_NID_ANY);
2696 /* Wait for all peer state to clean up */
2698 while (cfs_atomic_read(&net->ibn_npeers) != 0) {
2700 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* 2**n? */
2701 "%s: waiting for %d peers to disconnect\n",
2702 libcfs_nid2str(ni->ni_nid),
2703 cfs_atomic_read(&net->ibn_npeers));
2704 cfs_pause(cfs_time_seconds(1));
2707 kiblnd_ni_fini_pools(net);
2709 cfs_write_lock_irqsave(g_lock, flags);
2710 LASSERT (net->ibn_dev->ibd_nnets > 0);
2711 net->ibn_dev->ibd_nnets--;
2712 cfs_list_del(&net->ibn_list);
2713 cfs_write_unlock_irqrestore(g_lock, flags);
2717 case IBLND_INIT_NOTHING:
2718 LASSERT (cfs_atomic_read(&net->ibn_nconns) == 0);
2720 if (net->ibn_dev != NULL &&
2721 net->ibn_dev->ibd_nnets == 0)
2722 kiblnd_destroy_dev(net->ibn_dev);
2727 CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
2728 cfs_atomic_read(&libcfs_kmemory));
2730 net->ibn_init = IBLND_INIT_NOTHING;
2733 LIBCFS_FREE(net, sizeof(*net));
2736 if (cfs_list_empty(&kiblnd_data.kib_devs))
2737 kiblnd_base_shutdown();
2742 kiblnd_base_startup (void)
2747 LASSERT (kiblnd_data.kib_init == IBLND_INIT_NOTHING);
2750 memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
2752 cfs_rwlock_init(&kiblnd_data.kib_global_lock);
2754 CFS_INIT_LIST_HEAD(&kiblnd_data.kib_devs);
2755 CFS_INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
2757 kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
2758 LIBCFS_ALLOC(kiblnd_data.kib_peers,
2759 sizeof(cfs_list_t) *
2760 kiblnd_data.kib_peer_hash_size);
2761 if (kiblnd_data.kib_peers == NULL) {
2764 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2765 CFS_INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
2767 cfs_spin_lock_init(&kiblnd_data.kib_connd_lock);
2768 CFS_INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
2769 CFS_INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
2770 cfs_waitq_init(&kiblnd_data.kib_connd_waitq);
2772 cfs_spin_lock_init(&kiblnd_data.kib_sched_lock);
2773 CFS_INIT_LIST_HEAD(&kiblnd_data.kib_sched_conns);
2774 cfs_waitq_init(&kiblnd_data.kib_sched_waitq);
2775 cfs_waitq_init(&kiblnd_data.kib_failover_waitq);
2777 kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
2779 /* lists/ptrs/locks initialised */
2780 kiblnd_data.kib_init = IBLND_INIT_DATA;
2781 /*****************************************************/
2783 for (i = 0; i < IBLND_N_SCHED; i++) {
2784 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)((long)i));
2786 CERROR("Can't spawn o2iblnd scheduler[%d]: %d\n",
2792 rc = kiblnd_thread_start(kiblnd_connd, NULL);
2794 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
2798 if (*kiblnd_tunables.kib_dev_failover != 0)
2799 rc = kiblnd_thread_start(kiblnd_failover_thread, NULL);
2802 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
2806 /* flag everything initialised */
2807 kiblnd_data.kib_init = IBLND_INIT_ALL;
2808 /*****************************************************/
2813 kiblnd_base_shutdown();
2818 kiblnd_startup (lnet_ni_t *ni)
2821 kib_dev_t *ibdev = NULL;
2825 unsigned long flags;
2828 LASSERT (ni->ni_lnd == &the_o2iblnd);
2830 if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
2831 rc = kiblnd_base_startup();
2836 LIBCFS_ALLOC(net, sizeof(*net));
2841 memset(net, 0, sizeof(*net));
2843 cfs_gettimeofday(&tv);
2844 net->ibn_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
2846 ni->ni_peertimeout = *kiblnd_tunables.kib_peertimeout;
2847 ni->ni_maxtxcredits = *kiblnd_tunables.kib_credits;
2848 ni->ni_peertxcredits = *kiblnd_tunables.kib_peertxcredits;
2849 ni->ni_peerrtrcredits = *kiblnd_tunables.kib_peerrtrcredits;
2851 if (ni->ni_interfaces[0] != NULL) {
2852 /* Use the IPoIB interface specified in 'networks=' */
2854 CLASSERT (LNET_MAX_INTERFACES > 1);
2855 if (ni->ni_interfaces[1] != NULL) {
2856 CERROR("Multiple interfaces not supported\n");
2860 ifname = ni->ni_interfaces[0];
2862 ifname = *kiblnd_tunables.kib_default_ipif;
2865 if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
2866 CERROR("IPoIB interface name too long: %s\n", ifname);
2870 cfs_list_for_each (tmp, &kiblnd_data.kib_devs) {
2871 ibdev = cfs_list_entry(tmp, kib_dev_t, ibd_list);
2873 if (!strcmp(&ibdev->ibd_ifname[0], ifname))
2880 ibdev = kiblnd_create_dev(ifname);
2885 net->ibn_dev = ibdev;
2886 ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
2888 rc = kiblnd_net_init_pools(net);
2890 CERROR("Failed to initialize NI pools: %d\n", rc);
2894 cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2896 cfs_list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
2897 cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2899 net->ibn_init = IBLND_INIT_ALL;
2904 if (net->ibn_dev == NULL && ibdev != NULL)
2905 kiblnd_destroy_dev(ibdev);
2907 kiblnd_shutdown(ni);
2909 CDEBUG(D_NET, "kiblnd_startup failed\n");
2914 kiblnd_module_fini (void)
2916 lnet_unregister_lnd(&the_o2iblnd);
2917 kiblnd_tunables_fini();
2921 kiblnd_module_init (void)
2925 CLASSERT (sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
2926 CLASSERT (offsetof(kib_msg_t, ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
2928 CLASSERT (offsetof(kib_msg_t, ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
2931 rc = kiblnd_tunables_init();
2935 lnet_register_lnd(&the_o2iblnd);
2940 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
2941 MODULE_DESCRIPTION("Kernel OpenIB gen2 LND v2.00");
2942 MODULE_LICENSE("GPL");
2944 module_init(kiblnd_module_init);
2945 module_exit(kiblnd_module_fini);