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, 2015, Intel Corporation.
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>
44 static lnd_t the_o2iblnd;
46 kib_data_t kiblnd_data;
49 kiblnd_cksum (void *ptr, int nob)
55 sum = ((sum << 1) | (sum >> 31)) + *c++;
57 /* ensure I don't return 0 (== no checksum) */
58 return (sum == 0) ? 1 : sum;
62 kiblnd_msgtype2str(int type)
65 case IBLND_MSG_CONNREQ:
68 case IBLND_MSG_CONNACK:
74 case IBLND_MSG_IMMEDIATE:
77 case IBLND_MSG_PUT_REQ:
80 case IBLND_MSG_PUT_NAK:
83 case IBLND_MSG_PUT_ACK:
86 case IBLND_MSG_PUT_DONE:
89 case IBLND_MSG_GET_REQ:
92 case IBLND_MSG_GET_DONE:
101 kiblnd_msgtype2size(int type)
103 const int hdr_size = offsetof(kib_msg_t, ibm_u);
106 case IBLND_MSG_CONNREQ:
107 case IBLND_MSG_CONNACK:
108 return hdr_size + sizeof(kib_connparams_t);
113 case IBLND_MSG_IMMEDIATE:
114 return offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[0]);
116 case IBLND_MSG_PUT_REQ:
117 return hdr_size + sizeof(kib_putreq_msg_t);
119 case IBLND_MSG_PUT_ACK:
120 return hdr_size + sizeof(kib_putack_msg_t);
122 case IBLND_MSG_GET_REQ:
123 return hdr_size + sizeof(kib_get_msg_t);
125 case IBLND_MSG_PUT_NAK:
126 case IBLND_MSG_PUT_DONE:
127 case IBLND_MSG_GET_DONE:
128 return hdr_size + sizeof(kib_completion_msg_t);
135 kiblnd_unpack_rd(kib_msg_t *msg, int flip)
142 LASSERT (msg->ibm_type == IBLND_MSG_GET_REQ ||
143 msg->ibm_type == IBLND_MSG_PUT_ACK);
145 rd = msg->ibm_type == IBLND_MSG_GET_REQ ?
146 &msg->ibm_u.get.ibgm_rd :
147 &msg->ibm_u.putack.ibpam_rd;
150 __swab32s(&rd->rd_key);
151 __swab32s(&rd->rd_nfrags);
156 if (n <= 0 || n > IBLND_MAX_RDMA_FRAGS) {
157 CERROR("Bad nfrags: %d, should be 0 < n <= %d\n",
158 n, IBLND_MAX_RDMA_FRAGS);
162 nob = offsetof (kib_msg_t, ibm_u) +
163 kiblnd_rd_msg_size(rd, msg->ibm_type, n);
165 if (msg->ibm_nob < nob) {
166 CERROR("Short %s: %d(%d)\n",
167 kiblnd_msgtype2str(msg->ibm_type), msg->ibm_nob, nob);
174 for (i = 0; i < n; i++) {
175 __swab32s(&rd->rd_frags[i].rf_nob);
176 __swab64s(&rd->rd_frags[i].rf_addr);
183 kiblnd_pack_msg (lnet_ni_t *ni, kib_msg_t *msg, int version,
184 int credits, lnet_nid_t dstnid, __u64 dststamp)
186 kib_net_t *net = ni->ni_data;
188 /* CAVEAT EMPTOR! all message fields not set here should have been
189 * initialised previously. */
190 msg->ibm_magic = IBLND_MSG_MAGIC;
191 msg->ibm_version = version;
193 msg->ibm_credits = credits;
196 msg->ibm_srcnid = ni->ni_nid;
197 msg->ibm_srcstamp = net->ibn_incarnation;
198 msg->ibm_dstnid = dstnid;
199 msg->ibm_dststamp = dststamp;
201 if (*kiblnd_tunables.kib_cksum) {
202 /* NB ibm_cksum zero while computing cksum */
203 msg->ibm_cksum = kiblnd_cksum(msg, msg->ibm_nob);
208 kiblnd_unpack_msg(kib_msg_t *msg, int nob)
210 const int hdr_size = offsetof(kib_msg_t, ibm_u);
216 /* 6 bytes are enough to have received magic + version */
218 CERROR("Short message: %d\n", nob);
222 if (msg->ibm_magic == IBLND_MSG_MAGIC) {
224 } else if (msg->ibm_magic == __swab32(IBLND_MSG_MAGIC)) {
227 CERROR("Bad magic: %08x\n", msg->ibm_magic);
231 version = flip ? __swab16(msg->ibm_version) : msg->ibm_version;
232 if (version != IBLND_MSG_VERSION &&
233 version != IBLND_MSG_VERSION_1) {
234 CERROR("Bad version: %x\n", version);
238 if (nob < hdr_size) {
239 CERROR("Short message: %d\n", nob);
243 msg_nob = flip ? __swab32(msg->ibm_nob) : msg->ibm_nob;
245 CERROR("Short message: got %d, wanted %d\n", nob, msg_nob);
249 /* checksum must be computed with ibm_cksum zero and BEFORE anything
251 msg_cksum = flip ? __swab32(msg->ibm_cksum) : msg->ibm_cksum;
253 if (msg_cksum != 0 &&
254 msg_cksum != kiblnd_cksum(msg, msg_nob)) {
255 CERROR("Bad checksum\n");
259 msg->ibm_cksum = msg_cksum;
262 /* leave magic unflipped as a clue to peer endianness */
263 msg->ibm_version = version;
264 CLASSERT (sizeof(msg->ibm_type) == 1);
265 CLASSERT (sizeof(msg->ibm_credits) == 1);
266 msg->ibm_nob = msg_nob;
267 __swab64s(&msg->ibm_srcnid);
268 __swab64s(&msg->ibm_srcstamp);
269 __swab64s(&msg->ibm_dstnid);
270 __swab64s(&msg->ibm_dststamp);
273 if (msg->ibm_srcnid == LNET_NID_ANY) {
274 CERROR("Bad src nid: %s\n", libcfs_nid2str(msg->ibm_srcnid));
278 if (msg_nob < kiblnd_msgtype2size(msg->ibm_type)) {
279 CERROR("Short %s: %d(%d)\n", kiblnd_msgtype2str(msg->ibm_type),
280 msg_nob, kiblnd_msgtype2size(msg->ibm_type));
284 switch (msg->ibm_type) {
286 CERROR("Unknown message type %x\n", msg->ibm_type);
290 case IBLND_MSG_IMMEDIATE:
291 case IBLND_MSG_PUT_REQ:
294 case IBLND_MSG_PUT_ACK:
295 case IBLND_MSG_GET_REQ:
296 if (kiblnd_unpack_rd(msg, flip))
300 case IBLND_MSG_PUT_NAK:
301 case IBLND_MSG_PUT_DONE:
302 case IBLND_MSG_GET_DONE:
304 __swab32s(&msg->ibm_u.completion.ibcm_status);
307 case IBLND_MSG_CONNREQ:
308 case IBLND_MSG_CONNACK:
310 __swab16s(&msg->ibm_u.connparams.ibcp_queue_depth);
311 __swab16s(&msg->ibm_u.connparams.ibcp_max_frags);
312 __swab32s(&msg->ibm_u.connparams.ibcp_max_msg_size);
320 kiblnd_create_peer(lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid)
323 kib_net_t *net = ni->ni_data;
324 int cpt = lnet_cpt_of_nid(nid);
327 LASSERT(net != NULL);
328 LASSERT(nid != LNET_NID_ANY);
330 LIBCFS_CPT_ALLOC(peer, lnet_cpt_table(), cpt, sizeof(*peer));
332 CERROR("Cannot allocate peer\n");
339 peer->ibp_last_alive = 0;
340 peer->ibp_max_frags = IBLND_CFG_RDMA_FRAGS;
341 peer->ibp_queue_depth = *kiblnd_tunables.kib_peertxcredits;
342 atomic_set(&peer->ibp_refcount, 1); /* 1 ref for caller */
344 INIT_LIST_HEAD(&peer->ibp_list); /* not in the peer table yet */
345 INIT_LIST_HEAD(&peer->ibp_conns);
346 INIT_LIST_HEAD(&peer->ibp_tx_queue);
348 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
350 /* always called with a ref on ni, which prevents ni being shutdown */
351 LASSERT(net->ibn_shutdown == 0);
353 /* npeers only grows with the global lock held */
354 atomic_inc(&net->ibn_npeers);
356 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
363 kiblnd_destroy_peer (kib_peer_t *peer)
365 kib_net_t *net = peer->ibp_ni->ni_data;
367 LASSERT(net != NULL);
368 LASSERT (atomic_read(&peer->ibp_refcount) == 0);
369 LASSERT(!kiblnd_peer_active(peer));
370 LASSERT(kiblnd_peer_idle(peer));
371 LASSERT(list_empty(&peer->ibp_tx_queue));
373 LIBCFS_FREE(peer, sizeof(*peer));
375 /* NB a peer's connections keep a reference on their peer until
376 * they are destroyed, so we can be assured that _all_ state to do
377 * with this peer has been cleaned up when its refcount drops to
379 atomic_dec(&net->ibn_npeers);
383 kiblnd_find_peer_locked (lnet_nid_t nid)
385 /* the caller is responsible for accounting the additional reference
386 * that this creates */
387 struct list_head *peer_list = kiblnd_nid2peerlist(nid);
388 struct list_head *tmp;
391 list_for_each(tmp, peer_list) {
393 peer = list_entry(tmp, kib_peer_t, ibp_list);
394 LASSERT(!kiblnd_peer_idle(peer));
396 if (peer->ibp_nid != nid)
399 CDEBUG(D_NET, "got peer [%p] -> %s (%d) version: %x\n",
400 peer, libcfs_nid2str(nid),
401 atomic_read(&peer->ibp_refcount),
409 kiblnd_unlink_peer_locked (kib_peer_t *peer)
411 LASSERT(list_empty(&peer->ibp_conns));
413 LASSERT (kiblnd_peer_active(peer));
414 list_del_init(&peer->ibp_list);
415 /* lose peerlist's ref */
416 kiblnd_peer_decref(peer);
420 kiblnd_get_peer_info(lnet_ni_t *ni, int index,
421 lnet_nid_t *nidp, int *count)
424 struct list_head *ptmp;
428 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
430 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
432 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
434 peer = list_entry(ptmp, kib_peer_t, ibp_list);
435 LASSERT(!kiblnd_peer_idle(peer));
437 if (peer->ibp_ni != ni)
443 *nidp = peer->ibp_nid;
444 *count = atomic_read(&peer->ibp_refcount);
446 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
452 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
457 kiblnd_del_peer_locked (kib_peer_t *peer)
459 struct list_head *ctmp;
460 struct list_head *cnxt;
463 if (list_empty(&peer->ibp_conns)) {
464 kiblnd_unlink_peer_locked(peer);
466 list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
467 conn = list_entry(ctmp, kib_conn_t, ibc_list);
469 kiblnd_close_conn_locked(conn, 0);
471 /* NB closing peer's last conn unlinked it. */
473 /* NB peer now unlinked; might even be freed if the peer table had the
478 kiblnd_del_peer (lnet_ni_t *ni, lnet_nid_t nid)
480 struct list_head zombies = LIST_HEAD_INIT(zombies);
481 struct list_head *ptmp;
482 struct list_head *pnxt;
490 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
492 if (nid != LNET_NID_ANY) {
493 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
496 hi = kiblnd_data.kib_peer_hash_size - 1;
499 for (i = lo; i <= hi; i++) {
500 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
501 peer = list_entry(ptmp, kib_peer_t, ibp_list);
502 LASSERT(!kiblnd_peer_idle(peer));
504 if (peer->ibp_ni != ni)
507 if (!(nid == LNET_NID_ANY || peer->ibp_nid == nid))
510 if (!list_empty(&peer->ibp_tx_queue)) {
511 LASSERT(list_empty(&peer->ibp_conns));
513 list_splice_init(&peer->ibp_tx_queue,
517 kiblnd_del_peer_locked(peer);
518 rc = 0; /* matched something */
522 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
524 kiblnd_txlist_done(ni, &zombies, -EIO);
530 kiblnd_get_conn_by_idx(lnet_ni_t *ni, int index)
533 struct list_head *ptmp;
535 struct list_head *ctmp;
539 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
541 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
542 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
544 peer = list_entry(ptmp, kib_peer_t, ibp_list);
545 LASSERT(!kiblnd_peer_idle(peer));
547 if (peer->ibp_ni != ni)
550 list_for_each(ctmp, &peer->ibp_conns) {
554 conn = list_entry(ctmp, kib_conn_t, ibc_list);
555 kiblnd_conn_addref(conn);
556 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
563 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
568 kiblnd_debug_rx (kib_rx_t *rx)
570 CDEBUG(D_CONSOLE, " %p status %d msg_type %x cred %d\n",
571 rx, rx->rx_status, rx->rx_msg->ibm_type,
572 rx->rx_msg->ibm_credits);
576 kiblnd_debug_tx (kib_tx_t *tx)
578 CDEBUG(D_CONSOLE, " %p snd %d q %d w %d rc %d dl %lx "
579 "cookie "LPX64" msg %s%s type %x cred %d\n",
580 tx, tx->tx_sending, tx->tx_queued, tx->tx_waiting,
581 tx->tx_status, tx->tx_deadline, tx->tx_cookie,
582 tx->tx_lntmsg[0] == NULL ? "-" : "!",
583 tx->tx_lntmsg[1] == NULL ? "-" : "!",
584 tx->tx_msg->ibm_type, tx->tx_msg->ibm_credits);
588 kiblnd_debug_conn (kib_conn_t *conn)
590 struct list_head *tmp;
593 spin_lock(&conn->ibc_lock);
595 CDEBUG(D_CONSOLE, "conn[%d] %p [version %x] -> %s:\n",
596 atomic_read(&conn->ibc_refcount), conn,
597 conn->ibc_version, libcfs_nid2str(conn->ibc_peer->ibp_nid));
598 CDEBUG(D_CONSOLE, " state %d nposted %d/%d cred %d o_cred %d "
599 " r_cred %d\n", conn->ibc_state, conn->ibc_noops_posted,
600 conn->ibc_nsends_posted, conn->ibc_credits,
601 conn->ibc_outstanding_credits, conn->ibc_reserved_credits);
602 CDEBUG(D_CONSOLE, " comms_err %d\n", conn->ibc_comms_error);
604 CDEBUG(D_CONSOLE, " early_rxs:\n");
605 list_for_each(tmp, &conn->ibc_early_rxs)
606 kiblnd_debug_rx(list_entry(tmp, kib_rx_t, rx_list));
608 CDEBUG(D_CONSOLE, " tx_noops:\n");
609 list_for_each(tmp, &conn->ibc_tx_noops)
610 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
612 CDEBUG(D_CONSOLE, " tx_queue_nocred:\n");
613 list_for_each(tmp, &conn->ibc_tx_queue_nocred)
614 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
616 CDEBUG(D_CONSOLE, " tx_queue_rsrvd:\n");
617 list_for_each(tmp, &conn->ibc_tx_queue_rsrvd)
618 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
620 CDEBUG(D_CONSOLE, " tx_queue:\n");
621 list_for_each(tmp, &conn->ibc_tx_queue)
622 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
624 CDEBUG(D_CONSOLE, " active_txs:\n");
625 list_for_each(tmp, &conn->ibc_active_txs)
626 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
628 CDEBUG(D_CONSOLE, " rxs:\n");
629 for (i = 0; i < IBLND_RX_MSGS(conn); i++)
630 kiblnd_debug_rx(&conn->ibc_rxs[i]);
632 spin_unlock(&conn->ibc_lock);
636 kiblnd_translate_mtu(int value)
657 kiblnd_setup_mtu_locked(struct rdma_cm_id *cmid)
661 /* XXX There is no path record for iWARP, set by netdev->change_mtu? */
662 if (cmid->route.path_rec == NULL)
665 mtu = kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu);
668 cmid->route.path_rec->mtu = mtu;
672 kiblnd_get_completion_vector(kib_conn_t *conn, int cpt)
679 vectors = conn->ibc_cmid->device->num_comp_vectors;
683 mask = cfs_cpt_cpumask(lnet_cpt_table(), cpt);
685 /* hash NID to CPU id in this partition... */
686 off = conn->ibc_peer->ibp_nid % cpumask_weight(mask);
687 for_each_cpu(i, mask) {
697 kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
698 int state, int version)
701 * If the new conn is created successfully it takes over the caller's
702 * ref on 'peer'. It also "owns" 'cmid' and destroys it when it itself
703 * is destroyed. On failure, the caller's ref on 'peer' remains and
704 * she must dispose of 'cmid'. (Actually I'd block forever if I tried
705 * to destroy 'cmid' here since I'm called from the CM which still has
706 * its ref on 'cmid'). */
707 rwlock_t *glock = &kiblnd_data.kib_global_lock;
708 kib_net_t *net = peer->ibp_ni->ni_data;
710 struct ib_qp_init_attr *init_qp_attr;
711 struct kib_sched_info *sched;
712 #ifdef HAVE_IB_CQ_INIT_ATTR
713 struct ib_cq_init_attr cq_attr = {};
722 LASSERT(net != NULL);
723 LASSERT(!in_interrupt());
727 cpt = lnet_cpt_of_nid(peer->ibp_nid);
728 sched = kiblnd_data.kib_scheds[cpt];
730 LASSERT(sched->ibs_nthreads > 0);
732 LIBCFS_CPT_ALLOC(init_qp_attr, lnet_cpt_table(), cpt,
733 sizeof(*init_qp_attr));
734 if (init_qp_attr == NULL) {
735 CERROR("Can't allocate qp_attr for %s\n",
736 libcfs_nid2str(peer->ibp_nid));
740 LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
742 CERROR("Can't allocate connection for %s\n",
743 libcfs_nid2str(peer->ibp_nid));
747 conn->ibc_state = IBLND_CONN_INIT;
748 conn->ibc_version = version;
749 conn->ibc_peer = peer; /* I take the caller's ref */
750 cmid->context = conn; /* for future CM callbacks */
751 conn->ibc_cmid = cmid;
752 conn->ibc_max_frags = peer->ibp_max_frags;
753 conn->ibc_queue_depth = peer->ibp_queue_depth;
755 INIT_LIST_HEAD(&conn->ibc_early_rxs);
756 INIT_LIST_HEAD(&conn->ibc_tx_noops);
757 INIT_LIST_HEAD(&conn->ibc_tx_queue);
758 INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
759 INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
760 INIT_LIST_HEAD(&conn->ibc_active_txs);
761 spin_lock_init(&conn->ibc_lock);
763 LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
764 sizeof(*conn->ibc_connvars));
765 if (conn->ibc_connvars == NULL) {
766 CERROR("Can't allocate in-progress connection state\n");
770 write_lock_irqsave(glock, flags);
771 if (dev->ibd_failover) {
772 write_unlock_irqrestore(glock, flags);
773 CERROR("%s: failover in progress\n", dev->ibd_ifname);
777 if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
778 /* wakeup failover thread and teardown connection */
779 if (kiblnd_dev_can_failover(dev)) {
780 list_add_tail(&dev->ibd_fail_list,
781 &kiblnd_data.kib_failed_devs);
782 wake_up(&kiblnd_data.kib_failover_waitq);
785 write_unlock_irqrestore(glock, flags);
786 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
787 cmid->device->name, dev->ibd_ifname);
791 kiblnd_hdev_addref_locked(dev->ibd_hdev);
792 conn->ibc_hdev = dev->ibd_hdev;
794 kiblnd_setup_mtu_locked(cmid);
796 write_unlock_irqrestore(glock, flags);
798 LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
799 IBLND_RX_MSGS(conn) * sizeof(kib_rx_t));
800 if (conn->ibc_rxs == NULL) {
801 CERROR("Cannot allocate RX buffers\n");
805 rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
806 IBLND_RX_MSG_PAGES(conn));
810 kiblnd_map_rx_descs(conn);
812 #ifdef HAVE_IB_CQ_INIT_ATTR
813 cq_attr.cqe = IBLND_CQ_ENTRIES(conn);
814 cq_attr.comp_vector = kiblnd_get_completion_vector(conn, cpt);
815 cq = ib_create_cq(cmid->device,
816 kiblnd_cq_completion, kiblnd_cq_event, conn,
819 cq = ib_create_cq(cmid->device,
820 kiblnd_cq_completion, kiblnd_cq_event, conn,
821 IBLND_CQ_ENTRIES(conn),
822 kiblnd_get_completion_vector(conn, cpt));
825 CERROR("Failed to create CQ with %d CQEs: %ld\n",
826 IBLND_CQ_ENTRIES(conn), PTR_ERR(cq));
832 rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
834 CERROR("Can't request completion notification: %d\n", rc);
838 init_qp_attr->event_handler = kiblnd_qp_event;
839 init_qp_attr->qp_context = conn;
840 init_qp_attr->cap.max_send_wr = IBLND_SEND_WRS(conn);
841 init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(conn);
842 init_qp_attr->cap.max_send_sge = 1;
843 init_qp_attr->cap.max_recv_sge = 1;
844 init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
845 init_qp_attr->qp_type = IB_QPT_RC;
846 init_qp_attr->send_cq = cq;
847 init_qp_attr->recv_cq = cq;
849 conn->ibc_sched = sched;
851 rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr);
853 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d\n",
854 rc, init_qp_attr->cap.max_send_wr,
855 init_qp_attr->cap.max_recv_wr);
859 LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
861 /* 1 ref for caller and each rxmsg */
862 atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(conn));
863 conn->ibc_nrx = IBLND_RX_MSGS(conn);
866 for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
867 rc = kiblnd_post_rx(&conn->ibc_rxs[i],
868 IBLND_POSTRX_NO_CREDIT);
870 CERROR("Can't post rxmsg: %d\n", rc);
872 /* Make posted receives complete */
873 kiblnd_abort_receives(conn);
875 /* correct # of posted buffers
876 * NB locking needed now I'm racing with completion */
877 spin_lock_irqsave(&sched->ibs_lock, flags);
878 conn->ibc_nrx -= IBLND_RX_MSGS(conn) - i;
879 spin_unlock_irqrestore(&sched->ibs_lock, flags);
881 /* cmid will be destroyed by CM(ofed) after cm_callback
882 * returned, so we can't refer it anymore
883 * (by kiblnd_connd()->kiblnd_destroy_conn) */
884 rdma_destroy_qp(conn->ibc_cmid);
885 conn->ibc_cmid = NULL;
887 /* Drop my own and unused rxbuffer refcounts */
888 while (i++ <= IBLND_RX_MSGS(conn))
889 kiblnd_conn_decref(conn);
895 /* Init successful! */
896 LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
897 state == IBLND_CONN_PASSIVE_WAIT);
898 conn->ibc_state = state;
901 atomic_inc(&net->ibn_nconns);
905 kiblnd_destroy_conn(conn, true);
907 LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
913 kiblnd_destroy_conn(kib_conn_t *conn, bool free_conn)
915 struct rdma_cm_id *cmid = conn->ibc_cmid;
916 kib_peer_t *peer = conn->ibc_peer;
919 LASSERT (!in_interrupt());
920 LASSERT (atomic_read(&conn->ibc_refcount) == 0);
921 LASSERT(list_empty(&conn->ibc_early_rxs));
922 LASSERT(list_empty(&conn->ibc_tx_noops));
923 LASSERT(list_empty(&conn->ibc_tx_queue));
924 LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd));
925 LASSERT(list_empty(&conn->ibc_tx_queue_nocred));
926 LASSERT(list_empty(&conn->ibc_active_txs));
927 LASSERT (conn->ibc_noops_posted == 0);
928 LASSERT (conn->ibc_nsends_posted == 0);
930 switch (conn->ibc_state) {
932 /* conn must be completely disengaged from the network */
935 case IBLND_CONN_DISCONNECTED:
936 /* connvars should have been freed already */
937 LASSERT (conn->ibc_connvars == NULL);
940 case IBLND_CONN_INIT:
944 /* conn->ibc_cmid might be destroyed by CM already */
945 if (cmid != NULL && cmid->qp != NULL)
946 rdma_destroy_qp(cmid);
948 if (conn->ibc_cq != NULL) {
949 rc = ib_destroy_cq(conn->ibc_cq);
951 CWARN("Error destroying CQ: %d\n", rc);
954 if (conn->ibc_rx_pages != NULL)
955 kiblnd_unmap_rx_descs(conn);
957 if (conn->ibc_rxs != NULL) {
958 LIBCFS_FREE(conn->ibc_rxs,
959 IBLND_RX_MSGS(conn) * sizeof(kib_rx_t));
962 if (conn->ibc_connvars != NULL)
963 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
965 if (conn->ibc_hdev != NULL)
966 kiblnd_hdev_decref(conn->ibc_hdev);
968 /* See CAVEAT EMPTOR above in kiblnd_create_conn */
969 if (conn->ibc_state != IBLND_CONN_INIT) {
970 kib_net_t *net = peer->ibp_ni->ni_data;
972 kiblnd_peer_decref(peer);
973 rdma_destroy_id(cmid);
974 atomic_dec(&net->ibn_nconns);
978 LIBCFS_FREE(conn, sizeof(*conn));
982 kiblnd_close_peer_conns_locked(kib_peer_t *peer, int why)
985 struct list_head *ctmp;
986 struct list_head *cnxt;
989 list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
990 conn = list_entry(ctmp, kib_conn_t, ibc_list);
992 CDEBUG(D_NET, "Closing conn -> %s, "
993 "version: %x, reason: %d\n",
994 libcfs_nid2str(peer->ibp_nid),
995 conn->ibc_version, why);
997 kiblnd_close_conn_locked(conn, why);
1005 kiblnd_close_stale_conns_locked(kib_peer_t *peer,
1006 int version, __u64 incarnation)
1009 struct list_head *ctmp;
1010 struct list_head *cnxt;
1013 list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
1014 conn = list_entry(ctmp, kib_conn_t, ibc_list);
1016 if (conn->ibc_version == version &&
1017 conn->ibc_incarnation == incarnation)
1020 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
1021 "incarnation:"LPX64"(%x, "LPX64")\n",
1022 libcfs_nid2str(peer->ibp_nid),
1023 conn->ibc_version, conn->ibc_incarnation,
1024 version, incarnation);
1026 kiblnd_close_conn_locked(conn, -ESTALE);
1034 kiblnd_close_matching_conns(lnet_ni_t *ni, lnet_nid_t nid)
1037 struct list_head *ptmp;
1038 struct list_head *pnxt;
1042 unsigned long flags;
1045 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1047 if (nid != LNET_NID_ANY)
1048 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
1051 hi = kiblnd_data.kib_peer_hash_size - 1;
1054 for (i = lo; i <= hi; i++) {
1055 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
1057 peer = list_entry(ptmp, kib_peer_t, ibp_list);
1058 LASSERT(!kiblnd_peer_idle(peer));
1060 if (peer->ibp_ni != ni)
1063 if (!(nid == LNET_NID_ANY || nid == peer->ibp_nid))
1066 count += kiblnd_close_peer_conns_locked(peer, 0);
1070 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1072 /* wildcards always succeed */
1073 if (nid == LNET_NID_ANY)
1076 return (count == 0) ? -ENOENT : 0;
1080 kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
1082 struct libcfs_ioctl_data *data = arg;
1086 case IOC_LIBCFS_GET_PEER: {
1090 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1092 data->ioc_nid = nid;
1093 data->ioc_count = count;
1097 case IOC_LIBCFS_DEL_PEER: {
1098 rc = kiblnd_del_peer(ni, data->ioc_nid);
1101 case IOC_LIBCFS_GET_CONN: {
1105 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1111 LASSERT (conn->ibc_cmid != NULL);
1112 data->ioc_nid = conn->ibc_peer->ibp_nid;
1113 if (conn->ibc_cmid->route.path_rec == NULL)
1114 data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1117 ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1118 kiblnd_conn_decref(conn);
1121 case IOC_LIBCFS_CLOSE_CONNECTION: {
1122 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1134 kiblnd_query(lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when)
1136 cfs_time_t last_alive = 0;
1137 cfs_time_t now = cfs_time_current();
1138 rwlock_t *glock = &kiblnd_data.kib_global_lock;
1140 unsigned long flags;
1142 read_lock_irqsave(glock, flags);
1144 peer = kiblnd_find_peer_locked(nid);
1146 last_alive = peer->ibp_last_alive;
1148 read_unlock_irqrestore(glock, flags);
1150 if (last_alive != 0)
1153 /* peer is not persistent in hash, trigger peer creation
1154 * and connection establishment with a NULL tx */
1156 kiblnd_launch_tx(ni, NULL, nid);
1158 CDEBUG(D_NET, "Peer %s %p, alive %ld secs ago\n",
1159 libcfs_nid2str(nid), peer,
1160 last_alive ? cfs_duration_sec(now - last_alive) : -1);
1165 kiblnd_free_pages(kib_pages_t *p)
1167 int npages = p->ibp_npages;
1170 for (i = 0; i < npages; i++) {
1171 if (p->ibp_pages[i] != NULL)
1172 __free_page(p->ibp_pages[i]);
1175 LIBCFS_FREE(p, offsetof(kib_pages_t, ibp_pages[npages]));
1179 kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages)
1184 LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1185 offsetof(kib_pages_t, ibp_pages[npages]));
1187 CERROR("Can't allocate descriptor for %d pages\n", npages);
1191 memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1192 p->ibp_npages = npages;
1194 for (i = 0; i < npages; i++) {
1195 p->ibp_pages[i] = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1197 if (p->ibp_pages[i] == NULL) {
1198 CERROR("Can't allocate page %d of %d\n", i, npages);
1199 kiblnd_free_pages(p);
1209 kiblnd_unmap_rx_descs(kib_conn_t *conn)
1214 LASSERT (conn->ibc_rxs != NULL);
1215 LASSERT (conn->ibc_hdev != NULL);
1217 for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
1218 rx = &conn->ibc_rxs[i];
1220 LASSERT(rx->rx_nob >= 0); /* not posted */
1222 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1223 KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1225 IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1228 kiblnd_free_pages(conn->ibc_rx_pages);
1230 conn->ibc_rx_pages = NULL;
1234 kiblnd_map_rx_descs(kib_conn_t *conn)
1242 for (pg_off = ipg = i = 0; i < IBLND_RX_MSGS(conn); i++) {
1243 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1244 rx = &conn->ibc_rxs[i];
1247 rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
1250 kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1251 rx->rx_msg, IBLND_MSG_SIZE,
1253 LASSERT(!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1255 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1257 CDEBUG(D_NET, "rx %d: %p "LPX64"("LPX64")\n",
1258 i, rx->rx_msg, rx->rx_msgaddr,
1259 (__u64)(page_to_phys(pg) + pg_off));
1261 pg_off += IBLND_MSG_SIZE;
1262 LASSERT(pg_off <= PAGE_SIZE);
1264 if (pg_off == PAGE_SIZE) {
1267 LASSERT(ipg <= IBLND_RX_MSG_PAGES(conn));
1273 kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo)
1275 kib_hca_dev_t *hdev = tpo->tpo_hdev;
1279 LASSERT (tpo->tpo_pool.po_allocated == 0);
1284 for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1285 tx = &tpo->tpo_tx_descs[i];
1286 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1287 KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1289 IBLND_MSG_SIZE, DMA_TO_DEVICE);
1292 kiblnd_hdev_decref(hdev);
1293 tpo->tpo_hdev = NULL;
1296 static kib_hca_dev_t *
1297 kiblnd_current_hdev(kib_dev_t *dev)
1299 kib_hca_dev_t *hdev;
1300 unsigned long flags;
1303 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1304 while (dev->ibd_failover) {
1305 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1307 CDEBUG(D_NET, "%s: Wait for failover\n",
1309 set_current_state(TASK_INTERRUPTIBLE);
1310 schedule_timeout(cfs_time_seconds(1) / 100);
1312 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1315 kiblnd_hdev_addref_locked(dev->ibd_hdev);
1316 hdev = dev->ibd_hdev;
1318 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1324 kiblnd_map_tx_pool(kib_tx_pool_t *tpo)
1326 kib_pages_t *txpgs = tpo->tpo_tx_pages;
1327 kib_pool_t *pool = &tpo->tpo_pool;
1328 kib_net_t *net = pool->po_owner->ps_net;
1336 LASSERT (net != NULL);
1340 /* pre-mapped messages are not bigger than 1 page */
1341 CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE);
1343 /* No fancy arithmetic when we do the buffer calculations */
1344 CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
1346 tpo->tpo_hdev = kiblnd_current_hdev(dev);
1348 for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1349 page = txpgs->ibp_pages[ipage];
1350 tx = &tpo->tpo_tx_descs[i];
1352 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1355 tx->tx_msgaddr = kiblnd_dma_map_single(tpo->tpo_hdev->ibh_ibdev,
1359 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1361 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1363 list_add(&tx->tx_list, &pool->po_free_list);
1365 page_offset += IBLND_MSG_SIZE;
1366 LASSERT(page_offset <= PAGE_SIZE);
1368 if (page_offset == PAGE_SIZE) {
1371 LASSERT(ipage <= txpgs->ibp_npages);
1377 kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev, kib_rdma_desc_t *rd,
1378 int negotiated_nfrags)
1380 __u16 nfrags = (negotiated_nfrags != -1) ?
1381 negotiated_nfrags : *kiblnd_tunables.kib_map_on_demand;
1383 LASSERT(hdev->ibh_mrs != NULL);
1385 if (*kiblnd_tunables.kib_map_on_demand > 0 &&
1386 nfrags <= rd->rd_nfrags)
1389 return hdev->ibh_mrs;
1393 kiblnd_destroy_fmr_pool(kib_fmr_pool_t *fpo)
1395 LASSERT(fpo->fpo_map_count == 0);
1397 if (fpo->fpo_is_fmr) {
1398 if (fpo->fmr.fpo_fmr_pool)
1399 ib_destroy_fmr_pool(fpo->fmr.fpo_fmr_pool);
1401 struct kib_fast_reg_descriptor *frd, *tmp;
1404 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1406 list_del(&frd->frd_list);
1407 ib_free_fast_reg_page_list(frd->frd_frpl);
1408 ib_dereg_mr(frd->frd_mr);
1409 LIBCFS_FREE(frd, sizeof(*frd));
1412 if (i < fpo->fast_reg.fpo_pool_size)
1413 CERROR("FastReg pool still has %d regions registered\n",
1414 fpo->fast_reg.fpo_pool_size - i);
1418 kiblnd_hdev_decref(fpo->fpo_hdev);
1420 LIBCFS_FREE(fpo, sizeof(*fpo));
1424 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1426 kib_fmr_pool_t *fpo, *tmp;
1428 list_for_each_entry_safe(fpo, tmp, head, fpo_list) {
1429 list_del(&fpo->fpo_list);
1430 kiblnd_destroy_fmr_pool(fpo);
1434 static int kiblnd_fmr_pool_size(int ncpts)
1436 int size = *kiblnd_tunables.kib_fmr_pool_size / ncpts;
1438 return max(IBLND_FMR_POOL, size);
1441 static int kiblnd_fmr_flush_trigger(int ncpts)
1443 int size = *kiblnd_tunables.kib_fmr_flush_trigger / ncpts;
1445 return max(IBLND_FMR_POOL_FLUSH, size);
1448 static int kiblnd_alloc_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t *fpo)
1450 struct ib_fmr_pool_param param = {
1451 .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
1452 .page_shift = PAGE_SHIFT,
1453 .access = (IB_ACCESS_LOCAL_WRITE |
1454 IB_ACCESS_REMOTE_WRITE),
1455 .pool_size = fps->fps_pool_size,
1456 .dirty_watermark = fps->fps_flush_trigger,
1457 .flush_function = NULL,
1459 .cache = !!*kiblnd_tunables.kib_fmr_cache};
1462 fpo->fmr.fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd,
1464 if (IS_ERR(fpo->fmr.fpo_fmr_pool)) {
1465 rc = PTR_ERR(fpo->fmr.fpo_fmr_pool);
1467 CERROR("Failed to create FMR pool: %d\n", rc);
1469 CERROR("FMRs are not supported\n");
1475 static int kiblnd_alloc_freg_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t *fpo)
1477 struct kib_fast_reg_descriptor *frd, *tmp;
1480 INIT_LIST_HEAD(&fpo->fast_reg.fpo_pool_list);
1481 fpo->fast_reg.fpo_pool_size = 0;
1482 for (i = 0; i < fps->fps_pool_size; i++) {
1483 LIBCFS_CPT_ALLOC(frd, lnet_cpt_table(), fps->fps_cpt,
1486 CERROR("Failed to allocate a new fast_reg descriptor\n");
1492 frd->frd_frpl = ib_alloc_fast_reg_page_list(fpo->fpo_hdev->ibh_ibdev,
1493 LNET_MAX_PAYLOAD/PAGE_SIZE);
1494 if (IS_ERR(frd->frd_frpl)) {
1495 rc = PTR_ERR(frd->frd_frpl);
1496 CERROR("Failed to allocate ib_fast_reg_page_list: %d\n",
1501 frd->frd_mr = ib_alloc_fast_reg_mr(fpo->fpo_hdev->ibh_pd,
1502 LNET_MAX_PAYLOAD/PAGE_SIZE);
1503 if (IS_ERR(frd->frd_mr)) {
1504 rc = PTR_ERR(frd->frd_mr);
1505 CERROR("Failed to allocate ib_fast_reg_mr: %d\n", rc);
1509 frd->frd_valid = true;
1511 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1512 fpo->fast_reg.fpo_pool_size++;
1519 ib_dereg_mr(frd->frd_mr);
1521 ib_free_fast_reg_page_list(frd->frd_frpl);
1522 LIBCFS_FREE(frd, sizeof(*frd));
1525 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1527 list_del(&frd->frd_list);
1528 ib_free_fast_reg_page_list(frd->frd_frpl);
1529 ib_dereg_mr(frd->frd_mr);
1530 LIBCFS_FREE(frd, sizeof(*frd));
1537 kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo)
1539 struct ib_device_attr *dev_attr;
1540 kib_dev_t *dev = fps->fps_net->ibn_dev;
1541 kib_fmr_pool_t *fpo;
1544 dev_attr = kmalloc(sizeof(*dev_attr), GFP_KERNEL);
1548 LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1554 fpo->fpo_hdev = kiblnd_current_hdev(dev);
1556 rc = ib_query_device(fpo->fpo_hdev->ibh_ibdev, dev_attr);
1558 CERROR("Query device failed for %s: %d\n",
1559 fpo->fpo_hdev->ibh_ibdev->name, rc);
1563 /* Check for FMR or FastReg support */
1564 fpo->fpo_is_fmr = 0;
1565 if (fpo->fpo_hdev->ibh_ibdev->alloc_fmr &&
1566 fpo->fpo_hdev->ibh_ibdev->dealloc_fmr &&
1567 fpo->fpo_hdev->ibh_ibdev->map_phys_fmr &&
1568 fpo->fpo_hdev->ibh_ibdev->unmap_fmr) {
1569 LCONSOLE_INFO("Using FMR for registration\n");
1570 fpo->fpo_is_fmr = 1;
1571 } else if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
1572 LCONSOLE_INFO("Using FastReg for registration\n");
1575 LCONSOLE_ERROR_MSG(rc, "IB device does not support FMRs nor FastRegs, can't register memory\n");
1579 if (fpo->fpo_is_fmr)
1580 rc = kiblnd_alloc_fmr_pool(fps, fpo);
1582 rc = kiblnd_alloc_freg_pool(fps, fpo);
1587 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1588 fpo->fpo_owner = fps;
1594 kiblnd_hdev_decref(fpo->fpo_hdev);
1595 LIBCFS_FREE(fpo, sizeof(*fpo));
1604 kiblnd_fail_fmr_poolset(kib_fmr_poolset_t *fps, struct list_head *zombies)
1606 if (fps->fps_net == NULL) /* intialized? */
1609 spin_lock(&fps->fps_lock);
1611 while (!list_empty(&fps->fps_pool_list)) {
1612 kib_fmr_pool_t *fpo = list_entry(fps->fps_pool_list.next,
1613 kib_fmr_pool_t, fpo_list);
1614 fpo->fpo_failed = 1;
1615 list_del(&fpo->fpo_list);
1616 if (fpo->fpo_map_count == 0)
1617 list_add(&fpo->fpo_list, zombies);
1619 list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1622 spin_unlock(&fps->fps_lock);
1626 kiblnd_fini_fmr_poolset(kib_fmr_poolset_t *fps)
1628 if (fps->fps_net != NULL) { /* initialized? */
1629 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1630 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1635 kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt, kib_net_t *net,
1636 int pool_size, int flush_trigger)
1638 kib_fmr_pool_t *fpo;
1641 memset(fps, 0, sizeof(kib_fmr_poolset_t));
1645 fps->fps_pool_size = pool_size;
1646 fps->fps_flush_trigger = flush_trigger;
1647 spin_lock_init(&fps->fps_lock);
1648 INIT_LIST_HEAD(&fps->fps_pool_list);
1649 INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1651 rc = kiblnd_create_fmr_pool(fps, &fpo);
1653 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1659 kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, cfs_time_t now)
1661 if (fpo->fpo_map_count != 0) /* still in use */
1663 if (fpo->fpo_failed)
1665 return cfs_time_aftereq(now, fpo->fpo_deadline);
1669 kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1671 struct list_head zombies = LIST_HEAD_INIT(zombies);
1672 kib_fmr_pool_t *fpo = fmr->fmr_pool;
1673 kib_fmr_poolset_t *fps;
1674 cfs_time_t now = cfs_time_current();
1675 kib_fmr_pool_t *tmp;
1681 fps = fpo->fpo_owner;
1682 if (fpo->fpo_is_fmr) {
1683 if (fmr->fmr_pfmr) {
1684 rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1686 fmr->fmr_pfmr = NULL;
1690 rc = ib_flush_fmr_pool(fpo->fmr.fpo_fmr_pool);
1694 struct kib_fast_reg_descriptor *frd = fmr->fmr_frd;
1697 frd->frd_valid = false;
1698 spin_lock(&fps->fps_lock);
1699 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1700 spin_unlock(&fps->fps_lock);
1701 fmr->fmr_frd = NULL;
1704 fmr->fmr_pool = NULL;
1706 spin_lock(&fps->fps_lock);
1707 fpo->fpo_map_count--; /* decref the pool */
1709 list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1710 /* the first pool is persistent */
1711 if (fps->fps_pool_list.next == &fpo->fpo_list)
1714 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1715 list_move(&fpo->fpo_list, &zombies);
1719 spin_unlock(&fps->fps_lock);
1721 if (!list_empty(&zombies))
1722 kiblnd_destroy_fmr_pool_list(&zombies);
1726 kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages,
1727 __u32 nob, __u64 iov, bool is_rx, kib_fmr_t *fmr)
1729 kib_fmr_pool_t *fpo;
1734 spin_lock(&fps->fps_lock);
1735 version = fps->fps_version;
1736 list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1737 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1738 fpo->fpo_map_count++;
1740 if (fpo->fpo_is_fmr) {
1741 struct ib_pool_fmr *pfmr;
1743 spin_unlock(&fps->fps_lock);
1744 pfmr = ib_fmr_pool_map_phys(fpo->fmr.fpo_fmr_pool,
1745 pages, npages, iov);
1746 if (likely(!IS_ERR(pfmr))) {
1747 fmr->fmr_key = is_rx ? pfmr->fmr->rkey
1749 fmr->fmr_frd = NULL;
1750 fmr->fmr_pfmr = pfmr;
1751 fmr->fmr_pool = fpo;
1756 if (!list_empty(&fpo->fast_reg.fpo_pool_list)) {
1757 struct ib_send_wr *wr;
1758 struct kib_fast_reg_descriptor *frd;
1759 struct ib_fast_reg_page_list *frpl;
1762 frd = list_first_entry(&fpo->fast_reg.fpo_pool_list,
1763 struct kib_fast_reg_descriptor,
1765 list_del(&frd->frd_list);
1766 spin_unlock(&fps->fps_lock);
1768 frpl = frd->frd_frpl;
1771 if (!frd->frd_valid) {
1772 struct ib_send_wr *inv_wr;
1773 __u32 key = is_rx ? mr->rkey : mr->lkey;
1775 inv_wr = &frd->frd_inv_wr;
1776 memset(inv_wr, 0, sizeof(*inv_wr));
1777 inv_wr->opcode = IB_WR_LOCAL_INV;
1778 inv_wr->wr_id = IBLND_WID_MR;
1779 inv_wr->ex.invalidate_rkey = key;
1782 key = ib_inc_rkey(key);
1783 ib_update_fast_reg_key(mr, key);
1786 LASSERT(npages <= frpl->max_page_list_len);
1787 memcpy(frpl->page_list, pages,
1788 sizeof(*pages) * npages);
1790 /* Prepare FastReg WR */
1791 wr = &frd->frd_fastreg_wr;
1792 memset(wr, 0, sizeof(*wr));
1793 wr->opcode = IB_WR_FAST_REG_MR;
1794 wr->wr_id = IBLND_WID_MR;
1795 wr->wr.fast_reg.iova_start = iov;
1796 wr->wr.fast_reg.page_list = frpl;
1797 wr->wr.fast_reg.page_list_len = npages;
1798 wr->wr.fast_reg.page_shift = PAGE_SHIFT;
1799 wr->wr.fast_reg.length = nob;
1800 wr->wr.fast_reg.rkey = is_rx ? mr->rkey
1802 wr->wr.fast_reg.access_flags =
1803 (IB_ACCESS_LOCAL_WRITE |
1804 IB_ACCESS_REMOTE_WRITE);
1806 fmr->fmr_key = is_rx ? mr->rkey : mr->lkey;
1808 fmr->fmr_pfmr = NULL;
1809 fmr->fmr_pool = fpo;
1812 spin_unlock(&fps->fps_lock);
1816 spin_lock(&fps->fps_lock);
1817 fpo->fpo_map_count--;
1818 if (rc != -EAGAIN) {
1819 spin_unlock(&fps->fps_lock);
1823 /* EAGAIN and ... */
1824 if (version != fps->fps_version) {
1825 spin_unlock(&fps->fps_lock);
1830 if (fps->fps_increasing) {
1831 spin_unlock(&fps->fps_lock);
1832 CDEBUG(D_NET, "Another thread is allocating new "
1833 "FMR pool, waiting for her to complete\n");
1839 if (cfs_time_before(cfs_time_current(), fps->fps_next_retry)) {
1840 /* someone failed recently */
1841 spin_unlock(&fps->fps_lock);
1845 fps->fps_increasing = 1;
1846 spin_unlock(&fps->fps_lock);
1848 CDEBUG(D_NET, "Allocate new FMR pool\n");
1849 rc = kiblnd_create_fmr_pool(fps, &fpo);
1850 spin_lock(&fps->fps_lock);
1851 fps->fps_increasing = 0;
1854 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1856 fps->fps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1858 spin_unlock(&fps->fps_lock);
1864 kiblnd_fini_pool(kib_pool_t *pool)
1866 LASSERT(list_empty(&pool->po_free_list));
1867 LASSERT(pool->po_allocated == 0);
1869 CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
1873 kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
1875 CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
1877 memset(pool, 0, sizeof(kib_pool_t));
1878 INIT_LIST_HEAD(&pool->po_free_list);
1879 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1880 pool->po_owner = ps;
1881 pool->po_size = size;
1885 kiblnd_destroy_pool_list(struct list_head *head)
1889 while (!list_empty(head)) {
1890 pool = list_entry(head->next, kib_pool_t, po_list);
1891 list_del(&pool->po_list);
1893 LASSERT(pool->po_owner != NULL);
1894 pool->po_owner->ps_pool_destroy(pool);
1899 kiblnd_fail_poolset(kib_poolset_t *ps, struct list_head *zombies)
1901 if (ps->ps_net == NULL) /* intialized? */
1904 spin_lock(&ps->ps_lock);
1905 while (!list_empty(&ps->ps_pool_list)) {
1906 kib_pool_t *po = list_entry(ps->ps_pool_list.next,
1907 kib_pool_t, po_list);
1909 list_del(&po->po_list);
1910 if (po->po_allocated == 0)
1911 list_add(&po->po_list, zombies);
1913 list_add(&po->po_list, &ps->ps_failed_pool_list);
1915 spin_unlock(&ps->ps_lock);
1919 kiblnd_fini_poolset(kib_poolset_t *ps)
1921 if (ps->ps_net != NULL) { /* initialized? */
1922 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
1923 kiblnd_destroy_pool_list(&ps->ps_pool_list);
1928 kiblnd_init_poolset(kib_poolset_t *ps, int cpt,
1929 kib_net_t *net, char *name, int size,
1930 kib_ps_pool_create_t po_create,
1931 kib_ps_pool_destroy_t po_destroy,
1932 kib_ps_node_init_t nd_init,
1933 kib_ps_node_fini_t nd_fini)
1938 memset(ps, 0, sizeof(kib_poolset_t));
1942 ps->ps_pool_create = po_create;
1943 ps->ps_pool_destroy = po_destroy;
1944 ps->ps_node_init = nd_init;
1945 ps->ps_node_fini = nd_fini;
1946 ps->ps_pool_size = size;
1947 if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
1948 >= sizeof(ps->ps_name))
1950 spin_lock_init(&ps->ps_lock);
1951 INIT_LIST_HEAD(&ps->ps_pool_list);
1952 INIT_LIST_HEAD(&ps->ps_failed_pool_list);
1954 rc = ps->ps_pool_create(ps, size, &pool);
1956 list_add(&pool->po_list, &ps->ps_pool_list);
1958 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
1964 kiblnd_pool_is_idle(kib_pool_t *pool, cfs_time_t now)
1966 if (pool->po_allocated != 0) /* still in use */
1968 if (pool->po_failed)
1970 return cfs_time_aftereq(now, pool->po_deadline);
1974 kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node)
1976 struct list_head zombies = LIST_HEAD_INIT(zombies);
1977 kib_poolset_t *ps = pool->po_owner;
1979 cfs_time_t now = cfs_time_current();
1981 spin_lock(&ps->ps_lock);
1983 if (ps->ps_node_fini != NULL)
1984 ps->ps_node_fini(pool, node);
1986 LASSERT(pool->po_allocated > 0);
1987 list_add(node, &pool->po_free_list);
1988 pool->po_allocated--;
1990 list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
1991 /* the first pool is persistent */
1992 if (ps->ps_pool_list.next == &pool->po_list)
1995 if (kiblnd_pool_is_idle(pool, now))
1996 list_move(&pool->po_list, &zombies);
1998 spin_unlock(&ps->ps_lock);
2000 if (!list_empty(&zombies))
2001 kiblnd_destroy_pool_list(&zombies);
2005 kiblnd_pool_alloc_node(kib_poolset_t *ps)
2007 struct list_head *node;
2010 unsigned int interval = 1;
2011 cfs_time_t time_before;
2012 unsigned int trips = 0;
2015 spin_lock(&ps->ps_lock);
2016 list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
2017 if (list_empty(&pool->po_free_list))
2020 pool->po_allocated++;
2021 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
2022 node = pool->po_free_list.next;
2025 if (ps->ps_node_init != NULL) {
2026 /* still hold the lock */
2027 ps->ps_node_init(pool, node);
2029 spin_unlock(&ps->ps_lock);
2033 /* no available tx pool and ... */
2034 if (ps->ps_increasing) {
2035 /* another thread is allocating a new pool */
2036 spin_unlock(&ps->ps_lock);
2038 CDEBUG(D_NET, "Another thread is allocating new "
2039 "%s pool, waiting %d HZs for her to complete."
2041 ps->ps_name, interval, trips);
2043 set_current_state(TASK_INTERRUPTIBLE);
2044 schedule_timeout(interval);
2045 if (interval < cfs_time_seconds(1))
2051 if (cfs_time_before(cfs_time_current(), ps->ps_next_retry)) {
2052 /* someone failed recently */
2053 spin_unlock(&ps->ps_lock);
2057 ps->ps_increasing = 1;
2058 spin_unlock(&ps->ps_lock);
2060 CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
2061 time_before = cfs_time_current();
2062 rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
2063 CDEBUG(D_NET, "ps_pool_create took %lu HZ to complete",
2064 cfs_time_current() - time_before);
2066 spin_lock(&ps->ps_lock);
2067 ps->ps_increasing = 0;
2069 list_add_tail(&pool->po_list, &ps->ps_pool_list);
2071 ps->ps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
2072 CERROR("Can't allocate new %s pool because out of memory\n",
2075 spin_unlock(&ps->ps_lock);
2081 kiblnd_destroy_tx_pool(kib_pool_t *pool)
2083 kib_tx_pool_t *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
2086 LASSERT (pool->po_allocated == 0);
2088 if (tpo->tpo_tx_pages != NULL) {
2089 kiblnd_unmap_tx_pool(tpo);
2090 kiblnd_free_pages(tpo->tpo_tx_pages);
2093 if (tpo->tpo_tx_descs == NULL)
2096 for (i = 0; i < pool->po_size; i++) {
2097 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2099 list_del(&tx->tx_list);
2100 if (tx->tx_pages != NULL)
2101 LIBCFS_FREE(tx->tx_pages,
2103 sizeof(*tx->tx_pages));
2104 if (tx->tx_frags != NULL)
2105 LIBCFS_FREE(tx->tx_frags,
2106 IBLND_MAX_RDMA_FRAGS *
2107 sizeof(*tx->tx_frags));
2108 if (tx->tx_wrq != NULL)
2109 LIBCFS_FREE(tx->tx_wrq,
2110 (1 + IBLND_MAX_RDMA_FRAGS) *
2111 sizeof(*tx->tx_wrq));
2112 if (tx->tx_sge != NULL)
2113 LIBCFS_FREE(tx->tx_sge,
2114 (1 + IBLND_MAX_RDMA_FRAGS) *
2115 sizeof(*tx->tx_sge));
2116 if (tx->tx_rd != NULL)
2117 LIBCFS_FREE(tx->tx_rd,
2118 offsetof(kib_rdma_desc_t,
2119 rd_frags[IBLND_MAX_RDMA_FRAGS]));
2122 LIBCFS_FREE(tpo->tpo_tx_descs,
2123 pool->po_size * sizeof(kib_tx_t));
2125 kiblnd_fini_pool(pool);
2126 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2129 static int kiblnd_tx_pool_size(int ncpts)
2131 int ntx = *kiblnd_tunables.kib_ntx / ncpts;
2133 return max(IBLND_TX_POOL, ntx);
2137 kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
2144 LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2146 CERROR("Failed to allocate TX pool\n");
2150 pool = &tpo->tpo_pool;
2151 kiblnd_init_pool(ps, pool, size);
2152 tpo->tpo_tx_descs = NULL;
2153 tpo->tpo_tx_pages = NULL;
2155 npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2156 if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
2157 CERROR("Can't allocate tx pages: %d\n", npg);
2158 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2162 LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2163 size * sizeof(kib_tx_t));
2164 if (tpo->tpo_tx_descs == NULL) {
2165 CERROR("Can't allocate %d tx descriptors\n", size);
2166 ps->ps_pool_destroy(pool);
2170 memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
2172 for (i = 0; i < size; i++) {
2173 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2176 if (ps->ps_net->ibn_fmr_ps != NULL) {
2177 LIBCFS_CPT_ALLOC(tx->tx_pages,
2178 lnet_cpt_table(), ps->ps_cpt,
2179 LNET_MAX_IOV * sizeof(*tx->tx_pages));
2180 if (tx->tx_pages == NULL)
2184 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2185 IBLND_MAX_RDMA_FRAGS * sizeof(*tx->tx_frags));
2186 if (tx->tx_frags == NULL)
2189 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS);
2191 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2192 (1 + IBLND_MAX_RDMA_FRAGS) *
2193 sizeof(*tx->tx_wrq));
2194 if (tx->tx_wrq == NULL)
2197 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2198 (1 + IBLND_MAX_RDMA_FRAGS) *
2199 sizeof(*tx->tx_sge));
2200 if (tx->tx_sge == NULL)
2203 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2204 offsetof(kib_rdma_desc_t,
2205 rd_frags[IBLND_MAX_RDMA_FRAGS]));
2206 if (tx->tx_rd == NULL)
2211 kiblnd_map_tx_pool(tpo);
2216 ps->ps_pool_destroy(pool);
2221 kiblnd_tx_init(kib_pool_t *pool, struct list_head *node)
2223 kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t,
2225 kib_tx_t *tx = list_entry(node, kib_tx_t, tx_list);
2227 tx->tx_cookie = tps->tps_next_tx_cookie++;
2231 kiblnd_net_fini_pools(kib_net_t *net)
2235 cfs_cpt_for_each(i, lnet_cpt_table()) {
2236 kib_tx_poolset_t *tps;
2237 kib_fmr_poolset_t *fps;
2239 if (net->ibn_tx_ps != NULL) {
2240 tps = net->ibn_tx_ps[i];
2241 kiblnd_fini_poolset(&tps->tps_poolset);
2244 if (net->ibn_fmr_ps != NULL) {
2245 fps = net->ibn_fmr_ps[i];
2246 kiblnd_fini_fmr_poolset(fps);
2250 if (net->ibn_tx_ps != NULL) {
2251 cfs_percpt_free(net->ibn_tx_ps);
2252 net->ibn_tx_ps = NULL;
2255 if (net->ibn_fmr_ps != NULL) {
2256 cfs_percpt_free(net->ibn_fmr_ps);
2257 net->ibn_fmr_ps = NULL;
2262 kiblnd_net_init_pools(kib_net_t *net, __u32 *cpts, int ncpts)
2264 unsigned long flags;
2269 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2270 if (*kiblnd_tunables.kib_map_on_demand == 0) {
2271 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2273 goto create_tx_pool;
2276 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2278 if (*kiblnd_tunables.kib_fmr_pool_size <
2279 *kiblnd_tunables.kib_ntx / 4) {
2280 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2281 *kiblnd_tunables.kib_fmr_pool_size,
2282 *kiblnd_tunables.kib_ntx / 4);
2287 /* TX pool must be created later than FMR, see LU-2268
2289 LASSERT(net->ibn_tx_ps == NULL);
2291 /* premapping can fail if ibd_nmr > 1, so we always create
2292 * FMR pool and map-on-demand if premapping failed */
2294 net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2295 sizeof(kib_fmr_poolset_t));
2296 if (net->ibn_fmr_ps == NULL) {
2297 CERROR("Failed to allocate FMR pool array\n");
2302 for (i = 0; i < ncpts; i++) {
2303 cpt = (cpts == NULL) ? i : cpts[i];
2304 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, net,
2305 kiblnd_fmr_pool_size(ncpts),
2306 kiblnd_fmr_flush_trigger(ncpts));
2308 CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2315 LASSERT(i == ncpts);
2318 net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2319 sizeof(kib_tx_poolset_t));
2320 if (net->ibn_tx_ps == NULL) {
2321 CERROR("Failed to allocate tx pool array\n");
2326 for (i = 0; i < ncpts; i++) {
2327 cpt = (cpts == NULL) ? i : cpts[i];
2328 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2330 kiblnd_tx_pool_size(ncpts),
2331 kiblnd_create_tx_pool,
2332 kiblnd_destroy_tx_pool,
2333 kiblnd_tx_init, NULL);
2335 CERROR("Can't initialize TX pool for CPT %d: %d\n",
2343 kiblnd_net_fini_pools(net);
2349 kiblnd_hdev_get_attr(kib_hca_dev_t *hdev)
2351 struct ib_device_attr *attr;
2354 /* It's safe to assume a HCA can handle a page size
2355 * matching that of the native system */
2356 hdev->ibh_page_shift = PAGE_SHIFT;
2357 hdev->ibh_page_size = 1 << PAGE_SHIFT;
2358 hdev->ibh_page_mask = ~((__u64)hdev->ibh_page_size - 1);
2360 LIBCFS_ALLOC(attr, sizeof(*attr));
2362 CERROR("Out of memory\n");
2366 rc = ib_query_device(hdev->ibh_ibdev, attr);
2368 hdev->ibh_mr_size = attr->max_mr_size;
2370 LIBCFS_FREE(attr, sizeof(*attr));
2373 CERROR("Failed to query IB device: %d\n", rc);
2377 if (hdev->ibh_mr_size == ~0ULL) {
2378 hdev->ibh_mr_shift = 64;
2382 CERROR("Invalid mr size: "LPX64"\n", hdev->ibh_mr_size);
2387 kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev)
2389 if (hdev->ibh_mrs == NULL)
2392 ib_dereg_mr(hdev->ibh_mrs);
2394 hdev->ibh_mrs = NULL;
2398 kiblnd_hdev_destroy(kib_hca_dev_t *hdev)
2400 kiblnd_hdev_cleanup_mrs(hdev);
2402 if (hdev->ibh_pd != NULL)
2403 ib_dealloc_pd(hdev->ibh_pd);
2405 if (hdev->ibh_cmid != NULL)
2406 rdma_destroy_id(hdev->ibh_cmid);
2408 LIBCFS_FREE(hdev, sizeof(*hdev));
2412 kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev)
2416 int acflags = IB_ACCESS_LOCAL_WRITE |
2417 IB_ACCESS_REMOTE_WRITE;
2419 rc = kiblnd_hdev_get_attr(hdev);
2423 mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2425 CERROR("Failed ib_get_dma_mr: %ld\n", PTR_ERR(mr));
2426 kiblnd_hdev_cleanup_mrs(hdev);
2436 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2442 kiblnd_dev_need_failover(kib_dev_t *dev)
2444 struct rdma_cm_id *cmid;
2445 struct sockaddr_in srcaddr;
2446 struct sockaddr_in dstaddr;
2449 if (dev->ibd_hdev == NULL || /* initializing */
2450 dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2451 *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2454 /* XXX: it's UGLY, but I don't have better way to find
2455 * ib-bonding HCA failover because:
2457 * a. no reliable CM event for HCA failover...
2458 * b. no OFED API to get ib_device for current net_device...
2460 * We have only two choices at this point:
2462 * a. rdma_bind_addr(), it will conflict with listener cmid
2463 * b. rdma_resolve_addr() to zero addr */
2464 cmid = kiblnd_rdma_create_id(kiblnd_dummy_callback, dev, RDMA_PS_TCP,
2468 CERROR("Failed to create cmid for failover: %d\n", rc);
2472 memset(&srcaddr, 0, sizeof(srcaddr));
2473 srcaddr.sin_family = AF_INET;
2474 srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2476 memset(&dstaddr, 0, sizeof(dstaddr));
2477 dstaddr.sin_family = AF_INET;
2478 rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2479 (struct sockaddr *)&dstaddr, 1);
2480 if (rc != 0 || cmid->device == NULL) {
2481 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2482 dev->ibd_ifname, &dev->ibd_ifip,
2484 rdma_destroy_id(cmid);
2488 rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2489 rdma_destroy_id(cmid);
2494 kiblnd_dev_failover(kib_dev_t *dev)
2496 struct list_head zombie_tpo = LIST_HEAD_INIT(zombie_tpo);
2497 struct list_head zombie_ppo = LIST_HEAD_INIT(zombie_ppo);
2498 struct list_head zombie_fpo = LIST_HEAD_INIT(zombie_fpo);
2499 struct rdma_cm_id *cmid = NULL;
2500 kib_hca_dev_t *hdev = NULL;
2504 struct sockaddr_in addr;
2505 unsigned long flags;
2509 LASSERT (*kiblnd_tunables.kib_dev_failover > 1 ||
2510 dev->ibd_can_failover ||
2511 dev->ibd_hdev == NULL);
2513 rc = kiblnd_dev_need_failover(dev);
2517 if (dev->ibd_hdev != NULL &&
2518 dev->ibd_hdev->ibh_cmid != NULL) {
2519 /* XXX it's not good to close old listener at here,
2520 * because we can fail to create new listener.
2521 * But we have to close it now, otherwise rdma_bind_addr
2522 * will return EADDRINUSE... How crap! */
2523 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2525 cmid = dev->ibd_hdev->ibh_cmid;
2526 /* make next schedule of kiblnd_dev_need_failover()
2527 * return 1 for me */
2528 dev->ibd_hdev->ibh_cmid = NULL;
2529 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2531 rdma_destroy_id(cmid);
2534 cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP,
2538 CERROR("Failed to create cmid for failover: %d\n", rc);
2542 memset(&addr, 0, sizeof(addr));
2543 addr.sin_family = AF_INET;
2544 addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2545 addr.sin_port = htons(*kiblnd_tunables.kib_service);
2547 /* Bind to failover device or port */
2548 rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2549 if (rc != 0 || cmid->device == NULL) {
2550 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2551 dev->ibd_ifname, &dev->ibd_ifip,
2553 rdma_destroy_id(cmid);
2557 LIBCFS_ALLOC(hdev, sizeof(*hdev));
2559 CERROR("Failed to allocate kib_hca_dev\n");
2560 rdma_destroy_id(cmid);
2565 atomic_set(&hdev->ibh_ref, 1);
2566 hdev->ibh_dev = dev;
2567 hdev->ibh_cmid = cmid;
2568 hdev->ibh_ibdev = cmid->device;
2570 pd = ib_alloc_pd(cmid->device);
2573 CERROR("Can't allocate PD: %d\n", rc);
2579 rc = rdma_listen(cmid, 0);
2581 CERROR("Can't start new listener: %d\n", rc);
2585 rc = kiblnd_hdev_setup_mrs(hdev);
2587 CERROR("Can't setup device: %d\n", rc);
2591 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2593 old = dev->ibd_hdev;
2594 dev->ibd_hdev = hdev; /* take over the refcount */
2597 list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2598 cfs_cpt_for_each(i, lnet_cpt_table()) {
2599 kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2602 if (net->ibn_fmr_ps != NULL)
2603 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2608 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2610 if (!list_empty(&zombie_tpo))
2611 kiblnd_destroy_pool_list(&zombie_tpo);
2612 if (!list_empty(&zombie_ppo))
2613 kiblnd_destroy_pool_list(&zombie_ppo);
2614 if (!list_empty(&zombie_fpo))
2615 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2617 kiblnd_hdev_decref(hdev);
2620 dev->ibd_failed_failover++;
2622 dev->ibd_failed_failover = 0;
2628 kiblnd_destroy_dev (kib_dev_t *dev)
2630 LASSERT (dev->ibd_nnets == 0);
2631 LASSERT(list_empty(&dev->ibd_nets));
2633 list_del(&dev->ibd_fail_list);
2634 list_del(&dev->ibd_list);
2636 if (dev->ibd_hdev != NULL)
2637 kiblnd_hdev_decref(dev->ibd_hdev);
2639 LIBCFS_FREE(dev, sizeof(*dev));
2643 kiblnd_create_dev(char *ifname)
2645 struct net_device *netdev;
2652 rc = lnet_ipif_query(ifname, &up, &ip, &netmask);
2654 CERROR("Can't query IPoIB interface %s: %d\n",
2660 CERROR("Can't query IPoIB interface %s: it's down\n", ifname);
2664 LIBCFS_ALLOC(dev, sizeof(*dev));
2668 memset(dev, 0, sizeof(*dev));
2669 netdev = dev_get_by_name(&init_net, ifname);
2670 if (netdev == NULL) {
2671 dev->ibd_can_failover = 0;
2673 dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER);
2677 INIT_LIST_HEAD(&dev->ibd_nets);
2678 INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */
2679 INIT_LIST_HEAD(&dev->ibd_fail_list);
2681 strcpy(&dev->ibd_ifname[0], ifname);
2683 /* initialize the device */
2684 rc = kiblnd_dev_failover(dev);
2686 CERROR("Can't initialize device: %d\n", rc);
2687 LIBCFS_FREE(dev, sizeof(*dev));
2691 list_add_tail(&dev->ibd_list,
2692 &kiblnd_data.kib_devs);
2697 kiblnd_base_shutdown(void)
2699 struct kib_sched_info *sched;
2702 LASSERT(list_empty(&kiblnd_data.kib_devs));
2704 CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2705 atomic_read(&libcfs_kmemory));
2707 switch (kiblnd_data.kib_init) {
2711 case IBLND_INIT_ALL:
2712 case IBLND_INIT_DATA:
2713 LASSERT (kiblnd_data.kib_peers != NULL);
2714 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2715 LASSERT(list_empty(&kiblnd_data.kib_peers[i]));
2717 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
2718 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
2719 LASSERT(list_empty(&kiblnd_data.kib_reconn_list));
2720 LASSERT(list_empty(&kiblnd_data.kib_reconn_wait));
2722 /* flag threads to terminate; wake and wait for them to die */
2723 kiblnd_data.kib_shutdown = 1;
2725 /* NB: we really want to stop scheduler threads net by net
2726 * instead of the whole module, this should be improved
2727 * with dynamic configuration LNet */
2728 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2729 wake_up_all(&sched->ibs_waitq);
2731 wake_up_all(&kiblnd_data.kib_connd_waitq);
2732 wake_up_all(&kiblnd_data.kib_failover_waitq);
2735 while (atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2738 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2739 "Waiting for %d threads to terminate\n",
2740 atomic_read(&kiblnd_data.kib_nthreads));
2741 set_current_state(TASK_UNINTERRUPTIBLE);
2742 schedule_timeout(cfs_time_seconds(1));
2747 case IBLND_INIT_NOTHING:
2751 if (kiblnd_data.kib_peers != NULL) {
2752 LIBCFS_FREE(kiblnd_data.kib_peers,
2753 sizeof(struct list_head) *
2754 kiblnd_data.kib_peer_hash_size);
2757 if (kiblnd_data.kib_scheds != NULL)
2758 cfs_percpt_free(kiblnd_data.kib_scheds);
2760 CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2761 atomic_read(&libcfs_kmemory));
2763 kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2764 module_put(THIS_MODULE);
2768 kiblnd_shutdown (lnet_ni_t *ni)
2770 kib_net_t *net = ni->ni_data;
2771 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2773 unsigned long flags;
2775 LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2780 CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
2781 atomic_read(&libcfs_kmemory));
2783 write_lock_irqsave(g_lock, flags);
2784 net->ibn_shutdown = 1;
2785 write_unlock_irqrestore(g_lock, flags);
2787 switch (net->ibn_init) {
2791 case IBLND_INIT_ALL:
2792 /* nuke all existing peers within this net */
2793 kiblnd_del_peer(ni, LNET_NID_ANY);
2795 /* Wait for all peer state to clean up */
2797 while (atomic_read(&net->ibn_npeers) != 0) {
2800 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2801 "%s: waiting for %d peers to disconnect\n",
2802 libcfs_nid2str(ni->ni_nid),
2803 atomic_read(&net->ibn_npeers));
2804 set_current_state(TASK_UNINTERRUPTIBLE);
2805 schedule_timeout(cfs_time_seconds(1));
2808 kiblnd_net_fini_pools(net);
2810 write_lock_irqsave(g_lock, flags);
2811 LASSERT(net->ibn_dev->ibd_nnets > 0);
2812 net->ibn_dev->ibd_nnets--;
2813 list_del(&net->ibn_list);
2814 write_unlock_irqrestore(g_lock, flags);
2818 case IBLND_INIT_NOTHING:
2819 LASSERT (atomic_read(&net->ibn_nconns) == 0);
2821 if (net->ibn_dev != NULL &&
2822 net->ibn_dev->ibd_nnets == 0)
2823 kiblnd_destroy_dev(net->ibn_dev);
2828 CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
2829 atomic_read(&libcfs_kmemory));
2831 net->ibn_init = IBLND_INIT_NOTHING;
2834 LIBCFS_FREE(net, sizeof(*net));
2837 if (list_empty(&kiblnd_data.kib_devs))
2838 kiblnd_base_shutdown();
2843 kiblnd_base_startup(void)
2845 struct kib_sched_info *sched;
2849 LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
2851 try_module_get(THIS_MODULE);
2852 memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
2854 rwlock_init(&kiblnd_data.kib_global_lock);
2856 INIT_LIST_HEAD(&kiblnd_data.kib_devs);
2857 INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
2859 kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
2860 LIBCFS_ALLOC(kiblnd_data.kib_peers,
2861 sizeof(struct list_head) *
2862 kiblnd_data.kib_peer_hash_size);
2863 if (kiblnd_data.kib_peers == NULL)
2866 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2867 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
2869 spin_lock_init(&kiblnd_data.kib_connd_lock);
2870 INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
2871 INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
2872 INIT_LIST_HEAD(&kiblnd_data.kib_reconn_list);
2873 INIT_LIST_HEAD(&kiblnd_data.kib_reconn_wait);
2875 init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
2876 init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
2878 kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
2880 if (kiblnd_data.kib_scheds == NULL)
2883 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
2886 spin_lock_init(&sched->ibs_lock);
2887 INIT_LIST_HEAD(&sched->ibs_conns);
2888 init_waitqueue_head(&sched->ibs_waitq);
2890 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2891 if (*kiblnd_tunables.kib_nscheds > 0) {
2892 nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
2894 /* max to half of CPUs, another half is reserved for
2895 * upper layer modules */
2896 nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
2899 sched->ibs_nthreads_max = nthrs;
2903 kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
2905 /* lists/ptrs/locks initialised */
2906 kiblnd_data.kib_init = IBLND_INIT_DATA;
2907 /*****************************************************/
2909 rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
2911 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
2915 if (*kiblnd_tunables.kib_dev_failover != 0)
2916 rc = kiblnd_thread_start(kiblnd_failover_thread, NULL,
2920 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
2924 /* flag everything initialised */
2925 kiblnd_data.kib_init = IBLND_INIT_ALL;
2926 /*****************************************************/
2931 kiblnd_base_shutdown();
2936 kiblnd_start_schedulers(struct kib_sched_info *sched)
2942 if (sched->ibs_nthreads == 0) {
2943 if (*kiblnd_tunables.kib_nscheds > 0) {
2944 nthrs = sched->ibs_nthreads_max;
2946 nthrs = cfs_cpt_weight(lnet_cpt_table(),
2948 nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
2949 nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
2952 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
2953 /* increase one thread if there is new interface */
2954 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
2957 for (i = 0; i < nthrs; i++) {
2960 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
2961 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
2962 KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
2963 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
2967 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
2968 sched->ibs_cpt, sched->ibs_nthreads + i, rc);
2972 sched->ibs_nthreads += i;
2977 kiblnd_dev_start_threads(kib_dev_t *dev, int newdev, __u32 *cpts, int ncpts)
2983 for (i = 0; i < ncpts; i++) {
2984 struct kib_sched_info *sched;
2986 cpt = (cpts == NULL) ? i : cpts[i];
2987 sched = kiblnd_data.kib_scheds[cpt];
2989 if (!newdev && sched->ibs_nthreads > 0)
2992 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
2994 CERROR("Failed to start scheduler threads for %s\n",
3003 kiblnd_dev_search(char *ifname)
3005 kib_dev_t *alias = NULL;
3010 colon = strchr(ifname, ':');
3011 list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3012 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3018 colon2 = strchr(dev->ibd_ifname, ':');
3024 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3036 kiblnd_startup (lnet_ni_t *ni)
3039 kib_dev_t *ibdev = NULL;
3042 unsigned long flags;
3046 LASSERT (ni->ni_lnd == &the_o2iblnd);
3048 if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
3049 rc = kiblnd_base_startup();
3054 LIBCFS_ALLOC(net, sizeof(*net));
3059 memset(net, 0, sizeof(*net));
3061 do_gettimeofday(&tv);
3062 net->ibn_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
3064 ni->ni_peertimeout = *kiblnd_tunables.kib_peertimeout;
3065 ni->ni_maxtxcredits = *kiblnd_tunables.kib_credits;
3066 ni->ni_peertxcredits = *kiblnd_tunables.kib_peertxcredits;
3067 ni->ni_peerrtrcredits = *kiblnd_tunables.kib_peerrtrcredits;
3069 if (ni->ni_interfaces[0] != NULL) {
3070 /* Use the IPoIB interface specified in 'networks=' */
3072 CLASSERT (LNET_MAX_INTERFACES > 1);
3073 if (ni->ni_interfaces[1] != NULL) {
3074 CERROR("Multiple interfaces not supported\n");
3078 ifname = ni->ni_interfaces[0];
3080 ifname = *kiblnd_tunables.kib_default_ipif;
3083 if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3084 CERROR("IPoIB interface name too long: %s\n", ifname);
3088 ibdev = kiblnd_dev_search(ifname);
3090 newdev = ibdev == NULL;
3091 /* hmm...create kib_dev even for alias */
3092 if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0)
3093 ibdev = kiblnd_create_dev(ifname);
3098 net->ibn_dev = ibdev;
3099 ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
3101 rc = kiblnd_dev_start_threads(ibdev, newdev,
3102 ni->ni_cpts, ni->ni_ncpts);
3106 rc = kiblnd_net_init_pools(net, ni->ni_cpts, ni->ni_ncpts);
3108 CERROR("Failed to initialize NI pools: %d\n", rc);
3112 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3114 list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3115 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3117 net->ibn_init = IBLND_INIT_ALL;
3122 if (net != NULL && net->ibn_dev == NULL && ibdev != NULL)
3123 kiblnd_destroy_dev(ibdev);
3125 kiblnd_shutdown(ni);
3127 CDEBUG(D_NET, "kiblnd_startup failed\n");
3131 static lnd_t the_o2iblnd = {
3132 .lnd_type = O2IBLND,
3133 .lnd_startup = kiblnd_startup,
3134 .lnd_shutdown = kiblnd_shutdown,
3135 .lnd_ctl = kiblnd_ctl,
3136 .lnd_query = kiblnd_query,
3137 .lnd_send = kiblnd_send,
3138 .lnd_recv = kiblnd_recv,
3141 static void __exit ko2iblnd_exit(void)
3143 lnet_unregister_lnd(&the_o2iblnd);
3144 kiblnd_tunables_fini();
3147 static int __init ko2iblnd_init(void)
3151 CLASSERT(sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
3152 CLASSERT(offsetof(kib_msg_t,
3153 ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) <=
3155 CLASSERT(offsetof(kib_msg_t,
3156 ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3159 rc = kiblnd_tunables_init();
3163 lnet_register_lnd(&the_o2iblnd);
3168 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3169 MODULE_DESCRIPTION("OpenIB gen2 LNet Network Driver");
3170 MODULE_VERSION("2.8.0");
3171 MODULE_LICENSE("GPL");
3173 module_init(ko2iblnd_init);
3174 module_exit(ko2iblnd_exit);