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, 2012, 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>
42 #include <asm/div64.h>
44 static lnd_t the_o2iblnd = {
46 .lnd_startup = kiblnd_startup,
47 .lnd_shutdown = kiblnd_shutdown,
48 .lnd_ctl = kiblnd_ctl,
49 .lnd_query = kiblnd_query,
50 .lnd_send = kiblnd_send,
51 .lnd_recv = kiblnd_recv,
54 kib_data_t kiblnd_data;
56 static __u32 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;
68 static char *kiblnd_msgtype2str(int type)
71 case IBLND_MSG_CONNREQ:
74 case IBLND_MSG_CONNACK:
80 case IBLND_MSG_IMMEDIATE:
83 case IBLND_MSG_PUT_REQ:
86 case IBLND_MSG_PUT_NAK:
89 case IBLND_MSG_PUT_ACK:
92 case IBLND_MSG_PUT_DONE:
95 case IBLND_MSG_GET_REQ:
98 case IBLND_MSG_GET_DONE:
106 static int kiblnd_msgtype2size(int type)
108 const int hdr_size = offsetof(kib_msg_t, ibm_u);
111 case IBLND_MSG_CONNREQ:
112 case IBLND_MSG_CONNACK:
113 return hdr_size + sizeof(kib_connparams_t);
118 case IBLND_MSG_IMMEDIATE:
119 return offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[0]);
121 case IBLND_MSG_PUT_REQ:
122 return hdr_size + sizeof(kib_putreq_msg_t);
124 case IBLND_MSG_PUT_ACK:
125 return hdr_size + sizeof(kib_putack_msg_t);
127 case IBLND_MSG_GET_REQ:
128 return hdr_size + sizeof(kib_get_msg_t);
130 case IBLND_MSG_PUT_NAK:
131 case IBLND_MSG_PUT_DONE:
132 case IBLND_MSG_GET_DONE:
133 return hdr_size + sizeof(kib_completion_msg_t);
139 static int kiblnd_unpack_rd(kib_msg_t *msg, int flip)
146 LASSERT(msg->ibm_type == IBLND_MSG_GET_REQ ||
147 msg->ibm_type == IBLND_MSG_PUT_ACK);
149 rd = msg->ibm_type == IBLND_MSG_GET_REQ ?
150 &msg->ibm_u.get.ibgm_rd :
151 &msg->ibm_u.putack.ibpam_rd;
154 __swab32s(&rd->rd_key);
155 __swab32s(&rd->rd_nfrags);
160 if (n <= 0 || n > IBLND_MAX_RDMA_FRAGS) {
161 CERROR("Bad nfrags: %d, should be 0 < n <= %d\n",
162 n, IBLND_MAX_RDMA_FRAGS);
166 nob = offsetof(kib_msg_t, ibm_u) +
167 kiblnd_rd_msg_size(rd, msg->ibm_type, n);
169 if (msg->ibm_nob < nob) {
170 CERROR("Short %s: %d(%d)\n",
171 kiblnd_msgtype2str(msg->ibm_type), msg->ibm_nob, nob);
178 for (i = 0; i < n; i++) {
179 __swab32s(&rd->rd_frags[i].rf_nob);
180 __swab64s(&rd->rd_frags[i].rf_addr);
186 void kiblnd_pack_msg(lnet_ni_t *ni, kib_msg_t *msg, int version,
187 int credits, lnet_nid_t dstnid, __u64 dststamp)
189 kib_net_t *net = ni->ni_data;
191 /* CAVEAT EMPTOR! all message fields not set here should have been
192 * initialised previously. */
193 msg->ibm_magic = IBLND_MSG_MAGIC;
194 msg->ibm_version = version;
196 msg->ibm_credits = credits;
199 msg->ibm_srcnid = ni->ni_nid;
200 msg->ibm_srcstamp = net->ibn_incarnation;
201 msg->ibm_dstnid = dstnid;
202 msg->ibm_dststamp = dststamp;
204 if (*kiblnd_tunables.kib_cksum) {
205 /* NB ibm_cksum zero while computing cksum */
206 msg->ibm_cksum = kiblnd_cksum(msg, msg->ibm_nob);
210 int kiblnd_unpack_msg(kib_msg_t *msg, int nob)
212 const int hdr_size = offsetof(kib_msg_t, ibm_u);
218 /* 6 bytes are enough to have received magic + version */
220 CERROR("Short message: %d\n", nob);
224 if (msg->ibm_magic == IBLND_MSG_MAGIC) {
226 } else if (msg->ibm_magic == __swab32(IBLND_MSG_MAGIC)) {
229 CERROR("Bad magic: %08x\n", msg->ibm_magic);
233 version = flip ? __swab16(msg->ibm_version) : msg->ibm_version;
234 if (version != IBLND_MSG_VERSION &&
235 version != IBLND_MSG_VERSION_1) {
236 CERROR("Bad version: %x\n", version);
240 if (nob < hdr_size) {
241 CERROR("Short message: %d\n", nob);
245 msg_nob = flip ? __swab32(msg->ibm_nob) : msg->ibm_nob;
247 CERROR("Short message: got %d, wanted %d\n", nob, msg_nob);
251 /* checksum must be computed with ibm_cksum zero and BEFORE anything
253 msg_cksum = flip ? __swab32(msg->ibm_cksum) : msg->ibm_cksum;
255 if (msg_cksum != 0 &&
256 msg_cksum != kiblnd_cksum(msg, msg_nob)) {
257 CERROR("Bad checksum\n");
261 msg->ibm_cksum = msg_cksum;
264 /* leave magic unflipped as a clue to peer endianness */
265 msg->ibm_version = version;
266 CLASSERT(sizeof(msg->ibm_type) == 1);
267 CLASSERT(sizeof(msg->ibm_credits) == 1);
268 msg->ibm_nob = msg_nob;
269 __swab64s(&msg->ibm_srcnid);
270 __swab64s(&msg->ibm_srcstamp);
271 __swab64s(&msg->ibm_dstnid);
272 __swab64s(&msg->ibm_dststamp);
275 if (msg->ibm_srcnid == LNET_NID_ANY) {
276 CERROR("Bad src nid: %s\n", libcfs_nid2str(msg->ibm_srcnid));
280 if (msg_nob < kiblnd_msgtype2size(msg->ibm_type)) {
281 CERROR("Short %s: %d(%d)\n", kiblnd_msgtype2str(msg->ibm_type),
282 msg_nob, kiblnd_msgtype2size(msg->ibm_type));
286 switch (msg->ibm_type) {
288 CERROR("Unknown message type %x\n", msg->ibm_type);
292 case IBLND_MSG_IMMEDIATE:
293 case IBLND_MSG_PUT_REQ:
296 case IBLND_MSG_PUT_ACK:
297 case IBLND_MSG_GET_REQ:
298 if (kiblnd_unpack_rd(msg, flip))
302 case IBLND_MSG_PUT_NAK:
303 case IBLND_MSG_PUT_DONE:
304 case IBLND_MSG_GET_DONE:
306 __swab32s(&msg->ibm_u.completion.ibcm_status);
309 case IBLND_MSG_CONNREQ:
310 case IBLND_MSG_CONNACK:
312 __swab16s(&msg->ibm_u.connparams.ibcp_queue_depth);
313 __swab16s(&msg->ibm_u.connparams.ibcp_max_frags);
314 __swab32s(&msg->ibm_u.connparams.ibcp_max_msg_size);
321 int kiblnd_create_peer(lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid)
324 kib_net_t *net = ni->ni_data;
325 int cpt = lnet_cpt_of_nid(nid);
328 LASSERT(net != NULL);
329 LASSERT(nid != LNET_NID_ANY);
331 LIBCFS_CPT_ALLOC(peer, lnet_cpt_table(), cpt, sizeof(*peer));
333 CERROR("Cannot allocate peer\n");
337 memset(peer, 0, sizeof(*peer)); /* zero flags etc */
342 peer->ibp_last_alive = 0;
343 atomic_set(&peer->ibp_refcount, 1); /* 1 ref for caller */
345 INIT_LIST_HEAD(&peer->ibp_list); /* not in the peer table yet */
346 INIT_LIST_HEAD(&peer->ibp_conns);
347 INIT_LIST_HEAD(&peer->ibp_tx_queue);
349 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
351 /* always called with a ref on ni, which prevents ni being shutdown */
352 LASSERT(net->ibn_shutdown == 0);
354 /* npeers only grows with the global lock held */
355 atomic_inc(&net->ibn_npeers);
357 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
363 void 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(peer->ibp_connecting == 0);
371 LASSERT(peer->ibp_accepting == 0);
372 LASSERT(list_empty(&peer->ibp_conns));
373 LASSERT(list_empty(&peer->ibp_tx_queue));
375 LIBCFS_FREE(peer, sizeof(*peer));
377 /* NB a peer's connections keep a reference on their peer until
378 * they are destroyed, so we can be assured that _all_ state to do
379 * with this peer has been cleaned up when its refcount drops to
381 atomic_dec(&net->ibn_npeers);
384 kib_peer_t *kiblnd_find_peer_locked(lnet_nid_t nid)
386 /* the caller is responsible for accounting the additional reference
387 * that this creates */
388 struct list_head *peer_list = kiblnd_nid2peerlist(nid);
389 struct list_head *tmp;
392 list_for_each(tmp, peer_list) {
394 peer = list_entry(tmp, kib_peer_t, ibp_list);
396 LASSERT(peer->ibp_connecting > 0 || /* creating conns */
397 peer->ibp_accepting > 0 ||
398 !list_empty(&peer->ibp_conns)); /* active conn */
400 if (peer->ibp_nid != nid)
403 CDEBUG(D_NET, "got peer [%p] -> %s (%d) version: %x\n",
404 peer, libcfs_nid2str(nid),
405 atomic_read(&peer->ibp_refcount),
412 void kiblnd_unlink_peer_locked(kib_peer_t *peer)
414 LASSERT(list_empty(&peer->ibp_conns));
416 LASSERT(kiblnd_peer_active(peer));
417 list_del_init(&peer->ibp_list);
418 /* lose peerlist's ref */
419 kiblnd_peer_decref(peer);
422 static int kiblnd_get_peer_info(lnet_ni_t *ni, int index,
423 lnet_nid_t *nidp, int *count)
426 struct list_head *ptmp;
430 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
432 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
434 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
436 peer = list_entry(ptmp, kib_peer_t, ibp_list);
437 LASSERT(peer->ibp_connecting > 0 ||
438 peer->ibp_accepting > 0 ||
439 !list_empty(&peer->ibp_conns));
441 if (peer->ibp_ni != ni)
447 *nidp = peer->ibp_nid;
448 *count = atomic_read(&peer->ibp_refcount);
450 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
456 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
460 static void kiblnd_del_peer_locked(kib_peer_t *peer)
462 struct list_head *ctmp;
463 struct list_head *cnxt;
466 if (list_empty(&peer->ibp_conns)) {
467 kiblnd_unlink_peer_locked(peer);
469 list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
470 conn = list_entry(ctmp, kib_conn_t, ibc_list);
472 kiblnd_close_conn_locked(conn, 0);
474 /* NB closing peer's last conn unlinked it. */
476 /* NB peer now unlinked; might even be freed if the peer table had the
480 static int kiblnd_del_peer(lnet_ni_t *ni, lnet_nid_t nid)
483 struct list_head *ptmp;
484 struct list_head *pnxt;
492 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
494 if (nid != LNET_NID_ANY) {
495 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
498 hi = kiblnd_data.kib_peer_hash_size - 1;
501 for (i = lo; i <= hi; i++) {
502 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
503 peer = list_entry(ptmp, kib_peer_t, ibp_list);
504 LASSERT(peer->ibp_connecting > 0 ||
505 peer->ibp_accepting > 0 ||
506 !list_empty(&peer->ibp_conns));
508 if (peer->ibp_ni != ni)
511 if (!(nid == LNET_NID_ANY || peer->ibp_nid == nid))
514 if (!list_empty(&peer->ibp_tx_queue)) {
515 LASSERT(list_empty(&peer->ibp_conns));
517 list_splice_init(&peer->ibp_tx_queue,
521 kiblnd_del_peer_locked(peer);
522 rc = 0; /* matched something */
526 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
528 kiblnd_txlist_done(ni, &zombies, -EIO);
533 static kib_conn_t *kiblnd_get_conn_by_idx(lnet_ni_t *ni, int index)
536 struct list_head *ptmp;
538 struct list_head *ctmp;
542 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
544 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
545 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
547 peer = list_entry(ptmp, kib_peer_t, ibp_list);
548 LASSERT(peer->ibp_connecting > 0 ||
549 peer->ibp_accepting > 0 ||
550 !list_empty(&peer->ibp_conns));
552 if (peer->ibp_ni != ni)
555 list_for_each(ctmp, &peer->ibp_conns) {
559 conn = list_entry(ctmp, kib_conn_t,
561 kiblnd_conn_addref(conn);
562 read_unlock_irqrestore(
563 &kiblnd_data.kib_global_lock,
570 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
574 int kiblnd_translate_mtu(int value)
594 static void kiblnd_setup_mtu_locked(struct rdma_cm_id *cmid)
598 /* XXX There is no path record for iWARP, set by netdev->change_mtu? */
599 if (cmid->route.path_rec == NULL)
602 mtu = kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu);
605 cmid->route.path_rec->mtu = mtu;
608 static int kiblnd_get_completion_vector(kib_conn_t *conn, int cpt)
614 lnet_nid_t nid = conn->ibc_peer->ibp_nid;
616 vectors = conn->ibc_cmid->device->num_comp_vectors;
620 mask = cfs_cpt_cpumask(lnet_cpt_table(), cpt);
624 /* hash NID to CPU id in this partition... */
625 off = do_div(nid, cpumask_weight(mask));
626 for_each_cpu(i, mask) {
635 kib_conn_t *kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
636 int state, int version)
639 * If the new conn is created successfully it takes over the caller's
640 * ref on 'peer'. It also "owns" 'cmid' and destroys it when it itself
641 * is destroyed. On failure, the caller's ref on 'peer' remains and
642 * she must dispose of 'cmid'. (Actually I'd block forever if I tried
643 * to destroy 'cmid' here since I'm called from the CM which still has
644 * its ref on 'cmid'). */
645 rwlock_t *glock = &kiblnd_data.kib_global_lock;
646 kib_net_t *net = peer->ibp_ni->ni_data;
648 struct ib_qp_init_attr *init_qp_attr;
649 struct kib_sched_info *sched;
657 LASSERT(net != NULL);
658 LASSERT(!in_interrupt());
662 cpt = lnet_cpt_of_nid(peer->ibp_nid);
663 sched = kiblnd_data.kib_scheds[cpt];
665 LASSERT(sched->ibs_nthreads > 0);
667 LIBCFS_CPT_ALLOC(init_qp_attr, lnet_cpt_table(), cpt,
668 sizeof(*init_qp_attr));
669 if (init_qp_attr == NULL) {
670 CERROR("Can't allocate qp_attr for %s\n",
671 libcfs_nid2str(peer->ibp_nid));
675 LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
677 CERROR("Can't allocate connection for %s\n",
678 libcfs_nid2str(peer->ibp_nid));
682 conn->ibc_state = IBLND_CONN_INIT;
683 conn->ibc_version = version;
684 conn->ibc_peer = peer; /* I take the caller's ref */
685 cmid->context = conn; /* for future CM callbacks */
686 conn->ibc_cmid = cmid;
688 INIT_LIST_HEAD(&conn->ibc_early_rxs);
689 INIT_LIST_HEAD(&conn->ibc_tx_noops);
690 INIT_LIST_HEAD(&conn->ibc_tx_queue);
691 INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
692 INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
693 INIT_LIST_HEAD(&conn->ibc_active_txs);
694 spin_lock_init(&conn->ibc_lock);
696 LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
697 sizeof(*conn->ibc_connvars));
698 if (conn->ibc_connvars == NULL) {
699 CERROR("Can't allocate in-progress connection state\n");
703 write_lock_irqsave(glock, flags);
704 if (dev->ibd_failover) {
705 write_unlock_irqrestore(glock, flags);
706 CERROR("%s: failover in progress\n", dev->ibd_ifname);
710 if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
711 /* wakeup failover thread and teardown connection */
712 if (kiblnd_dev_can_failover(dev)) {
713 list_add_tail(&dev->ibd_fail_list,
714 &kiblnd_data.kib_failed_devs);
715 wake_up(&kiblnd_data.kib_failover_waitq);
718 write_unlock_irqrestore(glock, flags);
719 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
720 cmid->device->name, dev->ibd_ifname);
724 kiblnd_hdev_addref_locked(dev->ibd_hdev);
725 conn->ibc_hdev = dev->ibd_hdev;
727 kiblnd_setup_mtu_locked(cmid);
729 write_unlock_irqrestore(glock, flags);
731 LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
732 IBLND_RX_MSGS(version) * sizeof(kib_rx_t));
733 if (conn->ibc_rxs == NULL) {
734 CERROR("Cannot allocate RX buffers\n");
738 rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
739 IBLND_RX_MSG_PAGES(version));
743 kiblnd_map_rx_descs(conn);
745 cq = ib_create_cq(cmid->device,
746 kiblnd_cq_completion, kiblnd_cq_event, conn,
747 IBLND_CQ_ENTRIES(version),
748 kiblnd_get_completion_vector(conn, cpt));
750 CERROR("Can't create CQ: %ld, cqe: %d\n",
751 PTR_ERR(cq), IBLND_CQ_ENTRIES(version));
757 rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
759 CERROR("Can't request completion notificiation: %d\n", rc);
763 init_qp_attr->event_handler = kiblnd_qp_event;
764 init_qp_attr->qp_context = conn;
765 init_qp_attr->cap.max_send_wr = IBLND_SEND_WRS(version);
766 init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(version);
767 init_qp_attr->cap.max_send_sge = 1;
768 init_qp_attr->cap.max_recv_sge = 1;
769 init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
770 init_qp_attr->qp_type = IB_QPT_RC;
771 init_qp_attr->send_cq = cq;
772 init_qp_attr->recv_cq = cq;
774 conn->ibc_sched = sched;
776 rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr);
778 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d\n",
779 rc, init_qp_attr->cap.max_send_wr,
780 init_qp_attr->cap.max_recv_wr);
784 LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
786 /* 1 ref for caller and each rxmsg */
787 atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(version));
788 conn->ibc_nrx = IBLND_RX_MSGS(version);
791 for (i = 0; i < IBLND_RX_MSGS(version); i++) {
792 rc = kiblnd_post_rx(&conn->ibc_rxs[i],
793 IBLND_POSTRX_NO_CREDIT);
795 CERROR("Can't post rxmsg: %d\n", rc);
797 /* Make posted receives complete */
798 kiblnd_abort_receives(conn);
800 /* correct # of posted buffers
801 * NB locking needed now I'm racing with completion */
802 spin_lock_irqsave(&sched->ibs_lock, flags);
803 conn->ibc_nrx -= IBLND_RX_MSGS(version) - i;
804 spin_unlock_irqrestore(&sched->ibs_lock, flags);
806 /* cmid will be destroyed by CM(ofed) after cm_callback
807 * returned, so we can't refer it anymore
808 * (by kiblnd_connd()->kiblnd_destroy_conn) */
809 rdma_destroy_qp(conn->ibc_cmid);
810 conn->ibc_cmid = NULL;
812 /* Drop my own and unused rxbuffer refcounts */
813 while (i++ <= IBLND_RX_MSGS(version))
814 kiblnd_conn_decref(conn);
820 /* Init successful! */
821 LASSERT(state == IBLND_CONN_ACTIVE_CONNECT ||
822 state == IBLND_CONN_PASSIVE_WAIT);
823 conn->ibc_state = state;
826 atomic_inc(&net->ibn_nconns);
830 kiblnd_destroy_conn(conn);
832 LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
837 void kiblnd_destroy_conn(kib_conn_t *conn)
839 struct rdma_cm_id *cmid = conn->ibc_cmid;
840 kib_peer_t *peer = conn->ibc_peer;
843 LASSERT(!in_interrupt());
844 LASSERT(atomic_read(&conn->ibc_refcount) == 0);
845 LASSERT(list_empty(&conn->ibc_early_rxs));
846 LASSERT(list_empty(&conn->ibc_tx_noops));
847 LASSERT(list_empty(&conn->ibc_tx_queue));
848 LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd));
849 LASSERT(list_empty(&conn->ibc_tx_queue_nocred));
850 LASSERT(list_empty(&conn->ibc_active_txs));
851 LASSERT(conn->ibc_noops_posted == 0);
852 LASSERT(conn->ibc_nsends_posted == 0);
854 switch (conn->ibc_state) {
856 /* conn must be completely disengaged from the network */
859 case IBLND_CONN_DISCONNECTED:
860 /* connvars should have been freed already */
861 LASSERT(conn->ibc_connvars == NULL);
864 case IBLND_CONN_INIT:
868 /* conn->ibc_cmid might be destroyed by CM already */
869 if (cmid != NULL && cmid->qp != NULL)
870 rdma_destroy_qp(cmid);
872 if (conn->ibc_cq != NULL) {
873 rc = ib_destroy_cq(conn->ibc_cq);
875 CWARN("Error destroying CQ: %d\n", rc);
878 if (conn->ibc_rx_pages != NULL)
879 kiblnd_unmap_rx_descs(conn);
881 if (conn->ibc_rxs != NULL) {
882 LIBCFS_FREE(conn->ibc_rxs,
883 IBLND_RX_MSGS(conn->ibc_version)
887 if (conn->ibc_connvars != NULL)
888 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
890 if (conn->ibc_hdev != NULL)
891 kiblnd_hdev_decref(conn->ibc_hdev);
893 /* See CAVEAT EMPTOR above in kiblnd_create_conn */
894 if (conn->ibc_state != IBLND_CONN_INIT) {
895 kib_net_t *net = peer->ibp_ni->ni_data;
897 kiblnd_peer_decref(peer);
898 rdma_destroy_id(cmid);
899 atomic_dec(&net->ibn_nconns);
902 LIBCFS_FREE(conn, sizeof(*conn));
905 int kiblnd_close_peer_conns_locked(kib_peer_t *peer, int why)
908 struct list_head *ctmp;
909 struct list_head *cnxt;
912 list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
913 conn = list_entry(ctmp, kib_conn_t, ibc_list);
915 CDEBUG(D_NET, "Closing conn -> %s, version: %x, reason: %d\n",
916 libcfs_nid2str(peer->ibp_nid),
917 conn->ibc_version, why);
919 kiblnd_close_conn_locked(conn, why);
926 int kiblnd_close_stale_conns_locked(kib_peer_t *peer,
927 int version, __u64 incarnation)
930 struct list_head *ctmp;
931 struct list_head *cnxt;
934 list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
935 conn = list_entry(ctmp, kib_conn_t, ibc_list);
937 if (conn->ibc_version == version &&
938 conn->ibc_incarnation == incarnation)
942 "Closing stale conn -> %s version: %x, incarnation:%#llx(%x, %#llx)\n",
943 libcfs_nid2str(peer->ibp_nid),
944 conn->ibc_version, conn->ibc_incarnation,
945 version, incarnation);
947 kiblnd_close_conn_locked(conn, -ESTALE);
954 static int kiblnd_close_matching_conns(lnet_ni_t *ni, lnet_nid_t nid)
957 struct list_head *ptmp;
958 struct list_head *pnxt;
965 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
967 if (nid != LNET_NID_ANY)
968 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
971 hi = kiblnd_data.kib_peer_hash_size - 1;
974 for (i = lo; i <= hi; i++) {
975 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
977 peer = list_entry(ptmp, kib_peer_t, ibp_list);
978 LASSERT(peer->ibp_connecting > 0 ||
979 peer->ibp_accepting > 0 ||
980 !list_empty(&peer->ibp_conns));
982 if (peer->ibp_ni != ni)
985 if (!(nid == LNET_NID_ANY || nid == peer->ibp_nid))
988 count += kiblnd_close_peer_conns_locked(peer, 0);
992 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
994 /* wildcards always succeed */
995 if (nid == LNET_NID_ANY)
998 return (count == 0) ? -ENOENT : 0;
1001 int kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
1003 struct libcfs_ioctl_data *data = arg;
1007 case IOC_LIBCFS_GET_PEER: {
1011 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1013 data->ioc_nid = nid;
1014 data->ioc_count = count;
1018 case IOC_LIBCFS_DEL_PEER: {
1019 rc = kiblnd_del_peer(ni, data->ioc_nid);
1022 case IOC_LIBCFS_GET_CONN: {
1026 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1032 LASSERT(conn->ibc_cmid != NULL);
1033 data->ioc_nid = conn->ibc_peer->ibp_nid;
1034 if (conn->ibc_cmid->route.path_rec == NULL)
1035 data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1038 ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1039 kiblnd_conn_decref(conn);
1042 case IOC_LIBCFS_CLOSE_CONNECTION: {
1043 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1054 void kiblnd_query(lnet_ni_t *ni, lnet_nid_t nid, unsigned long *when)
1056 unsigned long last_alive = 0;
1057 unsigned long now = cfs_time_current();
1058 rwlock_t *glock = &kiblnd_data.kib_global_lock;
1060 unsigned long flags;
1062 read_lock_irqsave(glock, flags);
1064 peer = kiblnd_find_peer_locked(nid);
1066 LASSERT(peer->ibp_connecting > 0 || /* creating conns */
1067 peer->ibp_accepting > 0 ||
1068 !list_empty(&peer->ibp_conns)); /* active conn */
1069 last_alive = peer->ibp_last_alive;
1072 read_unlock_irqrestore(glock, flags);
1074 if (last_alive != 0)
1077 /* peer is not persistent in hash, trigger peer creation
1078 * and connection establishment with a NULL tx */
1080 kiblnd_launch_tx(ni, NULL, nid);
1082 CDEBUG(D_NET, "Peer %s %p, alive %ld secs ago\n",
1083 libcfs_nid2str(nid), peer,
1084 last_alive ? cfs_duration_sec(now - last_alive) : -1);
1087 void kiblnd_free_pages(kib_pages_t *p)
1089 int npages = p->ibp_npages;
1092 for (i = 0; i < npages; i++) {
1093 if (p->ibp_pages[i] != NULL)
1094 __free_page(p->ibp_pages[i]);
1097 LIBCFS_FREE(p, offsetof(kib_pages_t, ibp_pages[npages]));
1100 int kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages)
1105 LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1106 offsetof(kib_pages_t, ibp_pages[npages]));
1108 CERROR("Can't allocate descriptor for %d pages\n", npages);
1112 memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1113 p->ibp_npages = npages;
1115 for (i = 0; i < npages; i++) {
1116 p->ibp_pages[i] = alloc_pages_node(
1117 cfs_cpt_spread_node(lnet_cpt_table(), cpt),
1119 if (p->ibp_pages[i] == NULL) {
1120 CERROR("Can't allocate page %d of %d\n", i, npages);
1121 kiblnd_free_pages(p);
1130 void kiblnd_unmap_rx_descs(kib_conn_t *conn)
1135 LASSERT(conn->ibc_rxs != NULL);
1136 LASSERT(conn->ibc_hdev != NULL);
1138 for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1139 rx = &conn->ibc_rxs[i];
1141 LASSERT(rx->rx_nob >= 0); /* not posted */
1143 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1144 KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1146 IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1149 kiblnd_free_pages(conn->ibc_rx_pages);
1151 conn->ibc_rx_pages = NULL;
1154 void kiblnd_map_rx_descs(kib_conn_t *conn)
1162 for (pg_off = ipg = i = 0;
1163 i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1164 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1165 rx = &conn->ibc_rxs[i];
1168 rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
1170 rx->rx_msgaddr = kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1174 LASSERT(!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1176 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1178 CDEBUG(D_NET, "rx %d: %p %#llx(%#llx)\n",
1179 i, rx->rx_msg, rx->rx_msgaddr,
1180 lnet_page2phys(pg) + pg_off);
1182 pg_off += IBLND_MSG_SIZE;
1183 LASSERT(pg_off <= PAGE_SIZE);
1185 if (pg_off == PAGE_SIZE) {
1188 LASSERT(ipg <= IBLND_RX_MSG_PAGES(conn->ibc_version));
1193 static void kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo)
1195 kib_hca_dev_t *hdev = tpo->tpo_hdev;
1199 LASSERT(tpo->tpo_pool.po_allocated == 0);
1204 for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1205 tx = &tpo->tpo_tx_descs[i];
1206 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1207 KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1209 IBLND_MSG_SIZE, DMA_TO_DEVICE);
1212 kiblnd_hdev_decref(hdev);
1213 tpo->tpo_hdev = NULL;
1216 static kib_hca_dev_t *kiblnd_current_hdev(kib_dev_t *dev)
1218 kib_hca_dev_t *hdev;
1219 unsigned long flags;
1222 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1223 while (dev->ibd_failover) {
1224 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1226 CDEBUG(D_NET, "%s: Wait for failover\n",
1228 schedule_timeout(cfs_time_seconds(1) / 100);
1230 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1233 kiblnd_hdev_addref_locked(dev->ibd_hdev);
1234 hdev = dev->ibd_hdev;
1236 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1241 static void kiblnd_map_tx_pool(kib_tx_pool_t *tpo)
1243 kib_pages_t *txpgs = tpo->tpo_tx_pages;
1244 kib_pool_t *pool = &tpo->tpo_pool;
1245 kib_net_t *net = pool->po_owner->ps_net;
1253 LASSERT(net != NULL);
1257 /* pre-mapped messages are not bigger than 1 page */
1258 CLASSERT(IBLND_MSG_SIZE <= PAGE_SIZE);
1260 /* No fancy arithmetic when we do the buffer calculations */
1261 CLASSERT(PAGE_SIZE % IBLND_MSG_SIZE == 0);
1263 tpo->tpo_hdev = kiblnd_current_hdev(dev);
1265 for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1266 page = txpgs->ibp_pages[ipage];
1267 tx = &tpo->tpo_tx_descs[i];
1269 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1272 tx->tx_msgaddr = kiblnd_dma_map_single(
1273 tpo->tpo_hdev->ibh_ibdev, tx->tx_msg,
1274 IBLND_MSG_SIZE, DMA_TO_DEVICE);
1275 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1277 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1279 list_add(&tx->tx_list, &pool->po_free_list);
1281 page_offset += IBLND_MSG_SIZE;
1282 LASSERT(page_offset <= PAGE_SIZE);
1284 if (page_offset == PAGE_SIZE) {
1287 LASSERT(ipage <= txpgs->ibp_npages);
1292 struct ib_mr *kiblnd_find_dma_mr(kib_hca_dev_t *hdev, __u64 addr, __u64 size)
1296 LASSERT(hdev->ibh_mrs[0] != NULL);
1298 if (hdev->ibh_nmrs == 1)
1299 return hdev->ibh_mrs[0];
1301 index = addr >> hdev->ibh_mr_shift;
1303 if (index < hdev->ibh_nmrs &&
1304 index == ((addr + size - 1) >> hdev->ibh_mr_shift))
1305 return hdev->ibh_mrs[index];
1310 struct ib_mr *kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev, kib_rdma_desc_t *rd)
1312 struct ib_mr *prev_mr;
1316 LASSERT(hdev->ibh_mrs[0] != NULL);
1318 if (*kiblnd_tunables.kib_map_on_demand > 0 &&
1319 *kiblnd_tunables.kib_map_on_demand <= rd->rd_nfrags)
1322 if (hdev->ibh_nmrs == 1)
1323 return hdev->ibh_mrs[0];
1325 for (i = 0, mr = prev_mr = NULL;
1326 i < rd->rd_nfrags; i++) {
1327 mr = kiblnd_find_dma_mr(hdev,
1328 rd->rd_frags[i].rf_addr,
1329 rd->rd_frags[i].rf_nob);
1330 if (prev_mr == NULL)
1333 if (mr == NULL || prev_mr != mr) {
1334 /* Can't covered by one single MR */
1343 static void kiblnd_destroy_fmr_pool(kib_fmr_pool_t *pool)
1345 LASSERT(pool->fpo_map_count == 0);
1347 if (pool->fpo_fmr_pool != NULL)
1348 ib_destroy_fmr_pool(pool->fpo_fmr_pool);
1350 if (pool->fpo_hdev != NULL)
1351 kiblnd_hdev_decref(pool->fpo_hdev);
1353 LIBCFS_FREE(pool, sizeof(kib_fmr_pool_t));
1356 static void kiblnd_destroy_fmr_pool_list(struct list_head *head)
1358 kib_fmr_pool_t *pool;
1360 while (!list_empty(head)) {
1361 pool = list_entry(head->next, kib_fmr_pool_t, fpo_list);
1362 list_del(&pool->fpo_list);
1363 kiblnd_destroy_fmr_pool(pool);
1367 static int kiblnd_fmr_pool_size(int ncpts)
1369 int size = *kiblnd_tunables.kib_fmr_pool_size / ncpts;
1371 return max(IBLND_FMR_POOL, size);
1374 static int kiblnd_fmr_flush_trigger(int ncpts)
1376 int size = *kiblnd_tunables.kib_fmr_flush_trigger / ncpts;
1378 return max(IBLND_FMR_POOL_FLUSH, size);
1381 static int kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps,
1382 kib_fmr_pool_t **pp_fpo)
1384 /* FMR pool for RDMA */
1385 kib_dev_t *dev = fps->fps_net->ibn_dev;
1386 kib_fmr_pool_t *fpo;
1387 struct ib_fmr_pool_param param = {
1388 .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
1389 .page_shift = PAGE_SHIFT,
1390 .access = (IB_ACCESS_LOCAL_WRITE |
1391 IB_ACCESS_REMOTE_WRITE),
1392 .pool_size = fps->fps_pool_size,
1393 .dirty_watermark = fps->fps_flush_trigger,
1394 .flush_function = NULL,
1396 .cache = !!*kiblnd_tunables.kib_fmr_cache};
1399 LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1403 fpo->fpo_hdev = kiblnd_current_hdev(dev);
1405 fpo->fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd, ¶m);
1406 if (IS_ERR(fpo->fpo_fmr_pool)) {
1407 rc = PTR_ERR(fpo->fpo_fmr_pool);
1408 CERROR("Failed to create FMR pool: %d\n", rc);
1410 kiblnd_hdev_decref(fpo->fpo_hdev);
1411 LIBCFS_FREE(fpo, sizeof(kib_fmr_pool_t));
1415 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1416 fpo->fpo_owner = fps;
1422 static void kiblnd_fail_fmr_poolset(kib_fmr_poolset_t *fps,
1423 struct list_head *zombies)
1425 if (fps->fps_net == NULL) /* intialized? */
1428 spin_lock(&fps->fps_lock);
1430 while (!list_empty(&fps->fps_pool_list)) {
1431 kib_fmr_pool_t *fpo = list_entry(fps->fps_pool_list.next,
1432 kib_fmr_pool_t, fpo_list);
1433 fpo->fpo_failed = 1;
1434 list_del(&fpo->fpo_list);
1435 if (fpo->fpo_map_count == 0)
1436 list_add(&fpo->fpo_list, zombies);
1438 list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1441 spin_unlock(&fps->fps_lock);
1444 static void kiblnd_fini_fmr_poolset(kib_fmr_poolset_t *fps)
1446 if (fps->fps_net != NULL) { /* initialized? */
1447 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1448 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1452 static int kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt,
1453 kib_net_t *net, int pool_size,
1456 kib_fmr_pool_t *fpo;
1459 memset(fps, 0, sizeof(kib_fmr_poolset_t));
1463 fps->fps_pool_size = pool_size;
1464 fps->fps_flush_trigger = flush_trigger;
1465 spin_lock_init(&fps->fps_lock);
1466 INIT_LIST_HEAD(&fps->fps_pool_list);
1467 INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1469 rc = kiblnd_create_fmr_pool(fps, &fpo);
1471 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1476 static int kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, unsigned long now)
1478 if (fpo->fpo_map_count != 0) /* still in use */
1480 if (fpo->fpo_failed)
1482 return cfs_time_aftereq(now, fpo->fpo_deadline);
1485 void kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1488 kib_fmr_pool_t *fpo = fmr->fmr_pool;
1489 kib_fmr_poolset_t *fps = fpo->fpo_owner;
1490 unsigned long now = cfs_time_current();
1491 kib_fmr_pool_t *tmp;
1494 rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1498 rc = ib_flush_fmr_pool(fpo->fpo_fmr_pool);
1502 fmr->fmr_pool = NULL;
1503 fmr->fmr_pfmr = NULL;
1505 spin_lock(&fps->fps_lock);
1506 fpo->fpo_map_count--; /* decref the pool */
1508 list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1509 /* the first pool is persistent */
1510 if (fps->fps_pool_list.next == &fpo->fpo_list)
1513 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1514 list_move(&fpo->fpo_list, &zombies);
1518 spin_unlock(&fps->fps_lock);
1520 if (!list_empty(&zombies))
1521 kiblnd_destroy_fmr_pool_list(&zombies);
1524 int kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages,
1525 __u64 iov, kib_fmr_t *fmr)
1527 struct ib_pool_fmr *pfmr;
1528 kib_fmr_pool_t *fpo;
1533 spin_lock(&fps->fps_lock);
1534 version = fps->fps_version;
1535 list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1536 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1537 fpo->fpo_map_count++;
1538 spin_unlock(&fps->fps_lock);
1540 pfmr = ib_fmr_pool_map_phys(fpo->fpo_fmr_pool,
1541 pages, npages, iov);
1542 if (likely(!IS_ERR(pfmr))) {
1543 fmr->fmr_pool = fpo;
1544 fmr->fmr_pfmr = pfmr;
1548 spin_lock(&fps->fps_lock);
1549 fpo->fpo_map_count--;
1550 if (PTR_ERR(pfmr) != -EAGAIN) {
1551 spin_unlock(&fps->fps_lock);
1552 return PTR_ERR(pfmr);
1555 /* EAGAIN and ... */
1556 if (version != fps->fps_version) {
1557 spin_unlock(&fps->fps_lock);
1562 if (fps->fps_increasing) {
1563 spin_unlock(&fps->fps_lock);
1565 "Another thread is allocating new FMR pool, waiting for her to complete\n");
1571 if (time_before(cfs_time_current(), fps->fps_next_retry)) {
1572 /* someone failed recently */
1573 spin_unlock(&fps->fps_lock);
1577 fps->fps_increasing = 1;
1578 spin_unlock(&fps->fps_lock);
1580 CDEBUG(D_NET, "Allocate new FMR pool\n");
1581 rc = kiblnd_create_fmr_pool(fps, &fpo);
1582 spin_lock(&fps->fps_lock);
1583 fps->fps_increasing = 0;
1586 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1588 fps->fps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1590 spin_unlock(&fps->fps_lock);
1595 static void kiblnd_fini_pool(kib_pool_t *pool)
1597 LASSERT(list_empty(&pool->po_free_list));
1598 LASSERT(pool->po_allocated == 0);
1600 CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
1603 static void kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
1605 CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
1607 memset(pool, 0, sizeof(kib_pool_t));
1608 INIT_LIST_HEAD(&pool->po_free_list);
1609 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1610 pool->po_owner = ps;
1611 pool->po_size = size;
1614 static void kiblnd_destroy_pool_list(struct list_head *head)
1618 while (!list_empty(head)) {
1619 pool = list_entry(head->next, kib_pool_t, po_list);
1620 list_del(&pool->po_list);
1622 LASSERT(pool->po_owner != NULL);
1623 pool->po_owner->ps_pool_destroy(pool);
1627 static void kiblnd_fail_poolset(kib_poolset_t *ps, struct list_head *zombies)
1629 if (ps->ps_net == NULL) /* intialized? */
1632 spin_lock(&ps->ps_lock);
1633 while (!list_empty(&ps->ps_pool_list)) {
1634 kib_pool_t *po = list_entry(ps->ps_pool_list.next,
1635 kib_pool_t, po_list);
1637 list_del(&po->po_list);
1638 if (po->po_allocated == 0)
1639 list_add(&po->po_list, zombies);
1641 list_add(&po->po_list, &ps->ps_failed_pool_list);
1643 spin_unlock(&ps->ps_lock);
1646 static void kiblnd_fini_poolset(kib_poolset_t *ps)
1648 if (ps->ps_net != NULL) { /* initialized? */
1649 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
1650 kiblnd_destroy_pool_list(&ps->ps_pool_list);
1654 static int kiblnd_init_poolset(kib_poolset_t *ps, int cpt,
1655 kib_net_t *net, char *name, int size,
1656 kib_ps_pool_create_t po_create,
1657 kib_ps_pool_destroy_t po_destroy,
1658 kib_ps_node_init_t nd_init,
1659 kib_ps_node_fini_t nd_fini)
1664 memset(ps, 0, sizeof(kib_poolset_t));
1668 ps->ps_pool_create = po_create;
1669 ps->ps_pool_destroy = po_destroy;
1670 ps->ps_node_init = nd_init;
1671 ps->ps_node_fini = nd_fini;
1672 ps->ps_pool_size = size;
1673 if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
1674 >= sizeof(ps->ps_name))
1676 spin_lock_init(&ps->ps_lock);
1677 INIT_LIST_HEAD(&ps->ps_pool_list);
1678 INIT_LIST_HEAD(&ps->ps_failed_pool_list);
1680 rc = ps->ps_pool_create(ps, size, &pool);
1682 list_add(&pool->po_list, &ps->ps_pool_list);
1684 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
1689 static int kiblnd_pool_is_idle(kib_pool_t *pool, unsigned long now)
1691 if (pool->po_allocated != 0) /* still in use */
1693 if (pool->po_failed)
1695 return cfs_time_aftereq(now, pool->po_deadline);
1698 void kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node)
1701 kib_poolset_t *ps = pool->po_owner;
1703 unsigned long now = cfs_time_current();
1705 spin_lock(&ps->ps_lock);
1707 if (ps->ps_node_fini != NULL)
1708 ps->ps_node_fini(pool, node);
1710 LASSERT(pool->po_allocated > 0);
1711 list_add(node, &pool->po_free_list);
1712 pool->po_allocated--;
1714 list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
1715 /* the first pool is persistent */
1716 if (ps->ps_pool_list.next == &pool->po_list)
1719 if (kiblnd_pool_is_idle(pool, now))
1720 list_move(&pool->po_list, &zombies);
1722 spin_unlock(&ps->ps_lock);
1724 if (!list_empty(&zombies))
1725 kiblnd_destroy_pool_list(&zombies);
1728 struct list_head *kiblnd_pool_alloc_node(kib_poolset_t *ps)
1730 struct list_head *node;
1735 spin_lock(&ps->ps_lock);
1736 list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
1737 if (list_empty(&pool->po_free_list))
1740 pool->po_allocated++;
1741 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1742 node = pool->po_free_list.next;
1745 if (ps->ps_node_init != NULL) {
1746 /* still hold the lock */
1747 ps->ps_node_init(pool, node);
1749 spin_unlock(&ps->ps_lock);
1753 /* no available tx pool and ... */
1754 if (ps->ps_increasing) {
1755 /* another thread is allocating a new pool */
1756 spin_unlock(&ps->ps_lock);
1757 CDEBUG(D_NET, "Another thread is allocating new %s pool, waiting for her to complete\n",
1763 if (time_before(cfs_time_current(), ps->ps_next_retry)) {
1764 /* someone failed recently */
1765 spin_unlock(&ps->ps_lock);
1769 ps->ps_increasing = 1;
1770 spin_unlock(&ps->ps_lock);
1772 CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
1774 rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
1776 spin_lock(&ps->ps_lock);
1777 ps->ps_increasing = 0;
1779 list_add_tail(&pool->po_list, &ps->ps_pool_list);
1781 ps->ps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1782 CERROR("Can't allocate new %s pool because out of memory\n",
1785 spin_unlock(&ps->ps_lock);
1790 void kiblnd_pmr_pool_unmap(kib_phys_mr_t *pmr)
1792 kib_pmr_pool_t *ppo = pmr->pmr_pool;
1793 struct ib_mr *mr = pmr->pmr_mr;
1796 kiblnd_pool_free_node(&ppo->ppo_pool, &pmr->pmr_list);
1801 int kiblnd_pmr_pool_map(kib_pmr_poolset_t *pps, kib_hca_dev_t *hdev,
1802 kib_rdma_desc_t *rd, __u64 *iova, kib_phys_mr_t **pp_pmr)
1805 struct list_head *node;
1809 node = kiblnd_pool_alloc_node(&pps->pps_poolset);
1811 CERROR("Failed to allocate PMR descriptor\n");
1815 pmr = container_of(node, kib_phys_mr_t, pmr_list);
1816 if (pmr->pmr_pool->ppo_hdev != hdev) {
1817 kiblnd_pool_free_node(&pmr->pmr_pool->ppo_pool, node);
1821 for (i = 0; i < rd->rd_nfrags; i++) {
1822 pmr->pmr_ipb[i].addr = rd->rd_frags[i].rf_addr;
1823 pmr->pmr_ipb[i].size = rd->rd_frags[i].rf_nob;
1826 pmr->pmr_mr = ib_reg_phys_mr(hdev->ibh_pd,
1827 pmr->pmr_ipb, rd->rd_nfrags,
1828 IB_ACCESS_LOCAL_WRITE |
1829 IB_ACCESS_REMOTE_WRITE,
1831 if (!IS_ERR(pmr->pmr_mr)) {
1832 pmr->pmr_iova = *iova;
1837 rc = PTR_ERR(pmr->pmr_mr);
1838 CERROR("Failed ib_reg_phys_mr: %d\n", rc);
1841 kiblnd_pool_free_node(&pmr->pmr_pool->ppo_pool, node);
1846 static void kiblnd_destroy_pmr_pool(kib_pool_t *pool)
1848 kib_pmr_pool_t *ppo = container_of(pool, kib_pmr_pool_t, ppo_pool);
1852 LASSERT(pool->po_allocated == 0);
1854 list_for_each_entry_safe(pmr, tmp, &pool->po_free_list, pmr_list) {
1855 LASSERT(pmr->pmr_mr == NULL);
1856 list_del(&pmr->pmr_list);
1858 if (pmr->pmr_ipb != NULL) {
1859 LIBCFS_FREE(pmr->pmr_ipb,
1860 IBLND_MAX_RDMA_FRAGS *
1861 sizeof(struct ib_phys_buf));
1864 LIBCFS_FREE(pmr, sizeof(kib_phys_mr_t));
1867 kiblnd_fini_pool(pool);
1868 if (ppo->ppo_hdev != NULL)
1869 kiblnd_hdev_decref(ppo->ppo_hdev);
1871 LIBCFS_FREE(ppo, sizeof(kib_pmr_pool_t));
1874 static inline int kiblnd_pmr_pool_size(int ncpts)
1876 int size = *kiblnd_tunables.kib_pmr_pool_size / ncpts;
1878 return max(IBLND_PMR_POOL, size);
1881 static int kiblnd_create_pmr_pool(kib_poolset_t *ps, int size,
1884 struct kib_pmr_pool *ppo;
1885 struct kib_pool *pool;
1889 LIBCFS_CPT_ALLOC(ppo, lnet_cpt_table(),
1890 ps->ps_cpt, sizeof(kib_pmr_pool_t));
1892 CERROR("Failed to allocate PMR pool\n");
1896 pool = &ppo->ppo_pool;
1897 kiblnd_init_pool(ps, pool, size);
1899 for (i = 0; i < size; i++) {
1900 LIBCFS_CPT_ALLOC(pmr, lnet_cpt_table(),
1901 ps->ps_cpt, sizeof(kib_phys_mr_t));
1905 pmr->pmr_pool = ppo;
1906 LIBCFS_CPT_ALLOC(pmr->pmr_ipb, lnet_cpt_table(), ps->ps_cpt,
1907 IBLND_MAX_RDMA_FRAGS * sizeof(*pmr->pmr_ipb));
1908 if (pmr->pmr_ipb == NULL)
1911 list_add(&pmr->pmr_list, &pool->po_free_list);
1915 ps->ps_pool_destroy(pool);
1919 ppo->ppo_hdev = kiblnd_current_hdev(ps->ps_net->ibn_dev);
1924 static void kiblnd_destroy_tx_pool(kib_pool_t *pool)
1926 kib_tx_pool_t *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
1929 LASSERT(pool->po_allocated == 0);
1931 if (tpo->tpo_tx_pages != NULL) {
1932 kiblnd_unmap_tx_pool(tpo);
1933 kiblnd_free_pages(tpo->tpo_tx_pages);
1936 if (tpo->tpo_tx_descs == NULL)
1939 for (i = 0; i < pool->po_size; i++) {
1940 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
1942 list_del(&tx->tx_list);
1943 if (tx->tx_pages != NULL)
1944 LIBCFS_FREE(tx->tx_pages,
1946 sizeof(*tx->tx_pages));
1947 if (tx->tx_frags != NULL)
1948 LIBCFS_FREE(tx->tx_frags,
1949 IBLND_MAX_RDMA_FRAGS *
1950 sizeof(*tx->tx_frags));
1951 if (tx->tx_wrq != NULL)
1952 LIBCFS_FREE(tx->tx_wrq,
1953 (1 + IBLND_MAX_RDMA_FRAGS) *
1954 sizeof(*tx->tx_wrq));
1955 if (tx->tx_sge != NULL)
1956 LIBCFS_FREE(tx->tx_sge,
1957 (1 + IBLND_MAX_RDMA_FRAGS) *
1958 sizeof(*tx->tx_sge));
1959 if (tx->tx_rd != NULL)
1960 LIBCFS_FREE(tx->tx_rd,
1961 offsetof(kib_rdma_desc_t,
1962 rd_frags[IBLND_MAX_RDMA_FRAGS]));
1965 LIBCFS_FREE(tpo->tpo_tx_descs,
1966 pool->po_size * sizeof(kib_tx_t));
1968 kiblnd_fini_pool(pool);
1969 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
1972 static int kiblnd_tx_pool_size(int ncpts)
1974 int ntx = *kiblnd_tunables.kib_ntx / ncpts;
1976 return max(IBLND_TX_POOL, ntx);
1979 static int kiblnd_create_tx_pool(kib_poolset_t *ps, int size,
1987 LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
1989 CERROR("Failed to allocate TX pool\n");
1993 pool = &tpo->tpo_pool;
1994 kiblnd_init_pool(ps, pool, size);
1995 tpo->tpo_tx_descs = NULL;
1996 tpo->tpo_tx_pages = NULL;
1998 npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
1999 if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
2000 CERROR("Can't allocate tx pages: %d\n", npg);
2001 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2005 LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2006 size * sizeof(kib_tx_t));
2007 if (tpo->tpo_tx_descs == NULL) {
2008 CERROR("Can't allocate %d tx descriptors\n", size);
2009 ps->ps_pool_destroy(pool);
2013 memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
2015 for (i = 0; i < size; i++) {
2016 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2019 if (ps->ps_net->ibn_fmr_ps != NULL) {
2020 LIBCFS_CPT_ALLOC(tx->tx_pages,
2021 lnet_cpt_table(), ps->ps_cpt,
2022 LNET_MAX_IOV * sizeof(*tx->tx_pages));
2023 if (tx->tx_pages == NULL)
2027 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2028 IBLND_MAX_RDMA_FRAGS * sizeof(*tx->tx_frags));
2029 if (tx->tx_frags == NULL)
2032 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS);
2034 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2035 (1 + IBLND_MAX_RDMA_FRAGS) *
2036 sizeof(*tx->tx_wrq));
2037 if (tx->tx_wrq == NULL)
2040 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2041 (1 + IBLND_MAX_RDMA_FRAGS) *
2042 sizeof(*tx->tx_sge));
2043 if (tx->tx_sge == NULL)
2046 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2047 offsetof(kib_rdma_desc_t,
2048 rd_frags[IBLND_MAX_RDMA_FRAGS]));
2049 if (tx->tx_rd == NULL)
2054 kiblnd_map_tx_pool(tpo);
2059 ps->ps_pool_destroy(pool);
2063 static void kiblnd_tx_init(kib_pool_t *pool, struct list_head *node)
2065 kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t,
2067 kib_tx_t *tx = list_entry(node, kib_tx_t, tx_list);
2069 tx->tx_cookie = tps->tps_next_tx_cookie++;
2072 static void kiblnd_net_fini_pools(kib_net_t *net)
2076 cfs_cpt_for_each(i, lnet_cpt_table()) {
2077 kib_tx_poolset_t *tps;
2078 kib_fmr_poolset_t *fps;
2079 kib_pmr_poolset_t *pps;
2081 if (net->ibn_tx_ps != NULL) {
2082 tps = net->ibn_tx_ps[i];
2083 kiblnd_fini_poolset(&tps->tps_poolset);
2086 if (net->ibn_fmr_ps != NULL) {
2087 fps = net->ibn_fmr_ps[i];
2088 kiblnd_fini_fmr_poolset(fps);
2091 if (net->ibn_pmr_ps != NULL) {
2092 pps = net->ibn_pmr_ps[i];
2093 kiblnd_fini_poolset(&pps->pps_poolset);
2097 if (net->ibn_tx_ps != NULL) {
2098 cfs_percpt_free(net->ibn_tx_ps);
2099 net->ibn_tx_ps = NULL;
2102 if (net->ibn_fmr_ps != NULL) {
2103 cfs_percpt_free(net->ibn_fmr_ps);
2104 net->ibn_fmr_ps = NULL;
2107 if (net->ibn_pmr_ps != NULL) {
2108 cfs_percpt_free(net->ibn_pmr_ps);
2109 net->ibn_pmr_ps = NULL;
2113 static int kiblnd_net_init_pools(kib_net_t *net, __u32 *cpts, int ncpts)
2115 unsigned long flags;
2120 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2121 if (*kiblnd_tunables.kib_map_on_demand == 0 &&
2122 net->ibn_dev->ibd_hdev->ibh_nmrs == 1) {
2123 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2125 goto create_tx_pool;
2128 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2130 if (*kiblnd_tunables.kib_fmr_pool_size <
2131 *kiblnd_tunables.kib_ntx / 4) {
2132 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2133 *kiblnd_tunables.kib_fmr_pool_size,
2134 *kiblnd_tunables.kib_ntx / 4);
2139 /* TX pool must be created later than FMR/PMR, see LU-2268
2141 LASSERT(net->ibn_tx_ps == NULL);
2143 /* premapping can fail if ibd_nmr > 1, so we always create
2144 * FMR/PMR pool and map-on-demand if premapping failed */
2146 net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2147 sizeof(kib_fmr_poolset_t));
2148 if (net->ibn_fmr_ps == NULL) {
2149 CERROR("Failed to allocate FMR pool array\n");
2154 for (i = 0; i < ncpts; i++) {
2155 cpt = (cpts == NULL) ? i : cpts[i];
2156 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, net,
2157 kiblnd_fmr_pool_size(ncpts),
2158 kiblnd_fmr_flush_trigger(ncpts));
2159 if (rc == -ENOSYS && i == 0) /* no FMR */
2160 break; /* create PMR pool */
2162 if (rc != 0) { /* a real error */
2163 CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2170 LASSERT(i == ncpts);
2171 goto create_tx_pool;
2174 cfs_percpt_free(net->ibn_fmr_ps);
2175 net->ibn_fmr_ps = NULL;
2177 CWARN("Device does not support FMR, failing back to PMR\n");
2179 if (*kiblnd_tunables.kib_pmr_pool_size <
2180 *kiblnd_tunables.kib_ntx / 4) {
2181 CERROR("Can't set pmr pool size (%d) < ntx / 4(%d)\n",
2182 *kiblnd_tunables.kib_pmr_pool_size,
2183 *kiblnd_tunables.kib_ntx / 4);
2188 net->ibn_pmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2189 sizeof(kib_pmr_poolset_t));
2190 if (net->ibn_pmr_ps == NULL) {
2191 CERROR("Failed to allocate PMR pool array\n");
2196 for (i = 0; i < ncpts; i++) {
2197 cpt = (cpts == NULL) ? i : cpts[i];
2198 rc = kiblnd_init_poolset(&net->ibn_pmr_ps[cpt]->pps_poolset,
2200 kiblnd_pmr_pool_size(ncpts),
2201 kiblnd_create_pmr_pool,
2202 kiblnd_destroy_pmr_pool, NULL, NULL);
2204 CERROR("Can't initialize PMR pool for CPT %d: %d\n",
2211 net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2212 sizeof(kib_tx_poolset_t));
2213 if (net->ibn_tx_ps == NULL) {
2214 CERROR("Failed to allocate tx pool array\n");
2219 for (i = 0; i < ncpts; i++) {
2220 cpt = (cpts == NULL) ? i : cpts[i];
2221 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2223 kiblnd_tx_pool_size(ncpts),
2224 kiblnd_create_tx_pool,
2225 kiblnd_destroy_tx_pool,
2226 kiblnd_tx_init, NULL);
2228 CERROR("Can't initialize TX pool for CPT %d: %d\n",
2236 kiblnd_net_fini_pools(net);
2241 static int kiblnd_hdev_get_attr(kib_hca_dev_t *hdev)
2243 struct ib_device_attr *attr;
2246 /* It's safe to assume a HCA can handle a page size
2247 * matching that of the native system */
2248 hdev->ibh_page_shift = PAGE_SHIFT;
2249 hdev->ibh_page_size = 1 << PAGE_SHIFT;
2250 hdev->ibh_page_mask = ~((__u64)hdev->ibh_page_size - 1);
2252 LIBCFS_ALLOC(attr, sizeof(*attr));
2254 CERROR("Out of memory\n");
2258 rc = ib_query_device(hdev->ibh_ibdev, attr);
2260 hdev->ibh_mr_size = attr->max_mr_size;
2262 LIBCFS_FREE(attr, sizeof(*attr));
2265 CERROR("Failed to query IB device: %d\n", rc);
2269 if (hdev->ibh_mr_size == ~0ULL) {
2270 hdev->ibh_mr_shift = 64;
2274 for (hdev->ibh_mr_shift = 0;
2275 hdev->ibh_mr_shift < 64; hdev->ibh_mr_shift++) {
2276 if (hdev->ibh_mr_size == (1ULL << hdev->ibh_mr_shift) ||
2277 hdev->ibh_mr_size == (1ULL << hdev->ibh_mr_shift) - 1)
2281 CERROR("Invalid mr size: %#llx\n", hdev->ibh_mr_size);
2285 static void kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev)
2289 if (hdev->ibh_nmrs == 0 || hdev->ibh_mrs == NULL)
2292 for (i = 0; i < hdev->ibh_nmrs; i++) {
2293 if (hdev->ibh_mrs[i] == NULL)
2296 ib_dereg_mr(hdev->ibh_mrs[i]);
2299 LIBCFS_FREE(hdev->ibh_mrs, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2300 hdev->ibh_mrs = NULL;
2304 void kiblnd_hdev_destroy(kib_hca_dev_t *hdev)
2306 kiblnd_hdev_cleanup_mrs(hdev);
2308 if (hdev->ibh_pd != NULL)
2309 ib_dealloc_pd(hdev->ibh_pd);
2311 if (hdev->ibh_cmid != NULL)
2312 rdma_destroy_id(hdev->ibh_cmid);
2314 LIBCFS_FREE(hdev, sizeof(*hdev));
2317 static int kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev)
2324 int acflags = IB_ACCESS_LOCAL_WRITE |
2325 IB_ACCESS_REMOTE_WRITE;
2327 rc = kiblnd_hdev_get_attr(hdev);
2331 if (hdev->ibh_mr_shift == 64) {
2332 LIBCFS_ALLOC(hdev->ibh_mrs, 1 * sizeof(*hdev->ibh_mrs));
2333 if (hdev->ibh_mrs == NULL) {
2334 CERROR("Failed to allocate MRs table\n");
2338 hdev->ibh_mrs[0] = NULL;
2341 mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2343 CERROR("Failed ib_get_dma_mr : %ld\n", PTR_ERR(mr));
2344 kiblnd_hdev_cleanup_mrs(hdev);
2348 hdev->ibh_mrs[0] = mr;
2353 mr_size = 1ULL << hdev->ibh_mr_shift;
2354 mm_size = (unsigned long)high_memory - PAGE_OFFSET;
2356 hdev->ibh_nmrs = (int)((mm_size + mr_size - 1) >> hdev->ibh_mr_shift);
2358 if (hdev->ibh_mr_shift < 32 || hdev->ibh_nmrs > 1024) {
2359 /* it's 4T..., assume we will re-code at that time */
2360 CERROR("Can't support memory size: x%#llx with MR size: x%#llx\n",
2365 /* create an array of MRs to cover all memory */
2366 LIBCFS_ALLOC(hdev->ibh_mrs, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2367 if (hdev->ibh_mrs == NULL) {
2368 CERROR("Failed to allocate MRs' table\n");
2372 for (i = 0; i < hdev->ibh_nmrs; i++) {
2373 struct ib_phys_buf ipb;
2376 ipb.size = hdev->ibh_mr_size;
2377 ipb.addr = i * mr_size;
2380 mr = ib_reg_phys_mr(hdev->ibh_pd, &ipb, 1, acflags, &iova);
2382 CERROR("Failed ib_reg_phys_mr addr %#llx size %#llx : %ld\n",
2383 ipb.addr, ipb.size, PTR_ERR(mr));
2384 kiblnd_hdev_cleanup_mrs(hdev);
2388 LASSERT(iova == ipb.addr);
2390 hdev->ibh_mrs[i] = mr;
2394 if (hdev->ibh_mr_size != ~0ULL || hdev->ibh_nmrs != 1)
2395 LCONSOLE_INFO("Register global MR array, MR size: %#llx, array size: %d\n",
2396 hdev->ibh_mr_size, hdev->ibh_nmrs);
2401 static int kiblnd_dummy_callback(struct rdma_cm_id *cmid,
2402 struct rdma_cm_event *event)
2407 static int kiblnd_dev_need_failover(kib_dev_t *dev)
2409 struct rdma_cm_id *cmid;
2410 struct sockaddr_in srcaddr;
2411 struct sockaddr_in dstaddr;
2414 if (dev->ibd_hdev == NULL || /* initializing */
2415 dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2416 *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2419 /* XXX: it's UGLY, but I don't have better way to find
2420 * ib-bonding HCA failover because:
2422 * a. no reliable CM event for HCA failover...
2423 * b. no OFED API to get ib_device for current net_device...
2425 * We have only two choices at this point:
2427 * a. rdma_bind_addr(), it will conflict with listener cmid
2428 * b. rdma_resolve_addr() to zero addr */
2429 cmid = kiblnd_rdma_create_id(kiblnd_dummy_callback, dev, RDMA_PS_TCP,
2433 CERROR("Failed to create cmid for failover: %d\n", rc);
2437 memset(&srcaddr, 0, sizeof(srcaddr));
2438 srcaddr.sin_family = AF_INET;
2439 srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2441 memset(&dstaddr, 0, sizeof(dstaddr));
2442 dstaddr.sin_family = AF_INET;
2443 rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2444 (struct sockaddr *)&dstaddr, 1);
2445 if (rc != 0 || cmid->device == NULL) {
2446 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2447 dev->ibd_ifname, &dev->ibd_ifip,
2449 rdma_destroy_id(cmid);
2453 if (dev->ibd_hdev->ibh_ibdev == cmid->device) {
2454 /* don't need device failover */
2455 rdma_destroy_id(cmid);
2462 int kiblnd_dev_failover(kib_dev_t *dev)
2464 LIST_HEAD(zombie_tpo);
2465 LIST_HEAD(zombie_ppo);
2466 LIST_HEAD(zombie_fpo);
2467 struct rdma_cm_id *cmid = NULL;
2468 kib_hca_dev_t *hdev = NULL;
2472 struct sockaddr_in addr;
2473 unsigned long flags;
2477 LASSERT(*kiblnd_tunables.kib_dev_failover > 1 ||
2478 dev->ibd_can_failover ||
2479 dev->ibd_hdev == NULL);
2481 rc = kiblnd_dev_need_failover(dev);
2485 if (dev->ibd_hdev != NULL &&
2486 dev->ibd_hdev->ibh_cmid != NULL) {
2487 /* XXX it's not good to close old listener at here,
2488 * because we can fail to create new listener.
2489 * But we have to close it now, otherwise rdma_bind_addr
2490 * will return EADDRINUSE... How crap! */
2491 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2493 cmid = dev->ibd_hdev->ibh_cmid;
2494 /* make next schedule of kiblnd_dev_need_failover()
2495 * return 1 for me */
2496 dev->ibd_hdev->ibh_cmid = NULL;
2497 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2499 rdma_destroy_id(cmid);
2502 cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP,
2506 CERROR("Failed to create cmid for failover: %d\n", rc);
2510 memset(&addr, 0, sizeof(addr));
2511 addr.sin_family = AF_INET;
2512 addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2513 addr.sin_port = htons(*kiblnd_tunables.kib_service);
2515 /* Bind to failover device or port */
2516 rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2517 if (rc != 0 || cmid->device == NULL) {
2518 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2519 dev->ibd_ifname, &dev->ibd_ifip,
2521 rdma_destroy_id(cmid);
2525 LIBCFS_ALLOC(hdev, sizeof(*hdev));
2527 CERROR("Failed to allocate kib_hca_dev\n");
2528 rdma_destroy_id(cmid);
2533 atomic_set(&hdev->ibh_ref, 1);
2534 hdev->ibh_dev = dev;
2535 hdev->ibh_cmid = cmid;
2536 hdev->ibh_ibdev = cmid->device;
2538 pd = ib_alloc_pd(cmid->device);
2541 CERROR("Can't allocate PD: %d\n", rc);
2547 rc = rdma_listen(cmid, 0);
2549 CERROR("Can't start new listener: %d\n", rc);
2553 rc = kiblnd_hdev_setup_mrs(hdev);
2555 CERROR("Can't setup device: %d\n", rc);
2559 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2561 old = dev->ibd_hdev;
2562 dev->ibd_hdev = hdev; /* take over the refcount */
2565 list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2566 cfs_cpt_for_each(i, lnet_cpt_table()) {
2567 kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2570 if (net->ibn_fmr_ps != NULL) {
2571 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2574 } else if (net->ibn_pmr_ps != NULL) {
2575 kiblnd_fail_poolset(&net->ibn_pmr_ps[i]->
2576 pps_poolset, &zombie_ppo);
2581 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2583 if (!list_empty(&zombie_tpo))
2584 kiblnd_destroy_pool_list(&zombie_tpo);
2585 if (!list_empty(&zombie_ppo))
2586 kiblnd_destroy_pool_list(&zombie_ppo);
2587 if (!list_empty(&zombie_fpo))
2588 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2590 kiblnd_hdev_decref(hdev);
2593 dev->ibd_failed_failover++;
2595 dev->ibd_failed_failover = 0;
2600 void kiblnd_destroy_dev(kib_dev_t *dev)
2602 LASSERT(dev->ibd_nnets == 0);
2603 LASSERT(list_empty(&dev->ibd_nets));
2605 list_del(&dev->ibd_fail_list);
2606 list_del(&dev->ibd_list);
2608 if (dev->ibd_hdev != NULL)
2609 kiblnd_hdev_decref(dev->ibd_hdev);
2611 LIBCFS_FREE(dev, sizeof(*dev));
2614 static kib_dev_t *kiblnd_create_dev(char *ifname)
2616 struct net_device *netdev;
2623 rc = libcfs_ipif_query(ifname, &up, &ip, &netmask);
2625 CERROR("Can't query IPoIB interface %s: %d\n",
2631 CERROR("Can't query IPoIB interface %s: it's down\n", ifname);
2635 LIBCFS_ALLOC(dev, sizeof(*dev));
2639 netdev = dev_get_by_name(&init_net, ifname);
2640 if (netdev == NULL) {
2641 dev->ibd_can_failover = 0;
2643 dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER);
2647 INIT_LIST_HEAD(&dev->ibd_nets);
2648 INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */
2649 INIT_LIST_HEAD(&dev->ibd_fail_list);
2651 strcpy(&dev->ibd_ifname[0], ifname);
2653 /* initialize the device */
2654 rc = kiblnd_dev_failover(dev);
2656 CERROR("Can't initialize device: %d\n", rc);
2657 LIBCFS_FREE(dev, sizeof(*dev));
2661 list_add_tail(&dev->ibd_list,
2662 &kiblnd_data.kib_devs);
2666 static void kiblnd_base_shutdown(void)
2668 struct kib_sched_info *sched;
2671 LASSERT(list_empty(&kiblnd_data.kib_devs));
2673 CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2674 atomic_read(&libcfs_kmemory));
2676 switch (kiblnd_data.kib_init) {
2680 case IBLND_INIT_ALL:
2681 case IBLND_INIT_DATA:
2682 LASSERT(kiblnd_data.kib_peers != NULL);
2683 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2684 LASSERT(list_empty(&kiblnd_data.kib_peers[i]));
2685 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
2686 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
2688 /* flag threads to terminate; wake and wait for them to die */
2689 kiblnd_data.kib_shutdown = 1;
2691 /* NB: we really want to stop scheduler threads net by net
2692 * instead of the whole module, this should be improved
2693 * with dynamic configuration LNet */
2694 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2695 wake_up_all(&sched->ibs_waitq);
2697 wake_up_all(&kiblnd_data.kib_connd_waitq);
2698 wake_up_all(&kiblnd_data.kib_failover_waitq);
2701 while (atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2704 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2705 "Waiting for %d threads to terminate\n",
2706 atomic_read(&kiblnd_data.kib_nthreads));
2707 set_current_state(TASK_UNINTERRUPTIBLE);
2708 schedule_timeout(cfs_time_seconds(1));
2713 case IBLND_INIT_NOTHING:
2717 if (kiblnd_data.kib_peers != NULL) {
2718 LIBCFS_FREE(kiblnd_data.kib_peers,
2719 sizeof(struct list_head) *
2720 kiblnd_data.kib_peer_hash_size);
2723 if (kiblnd_data.kib_scheds != NULL)
2724 cfs_percpt_free(kiblnd_data.kib_scheds);
2726 CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2727 atomic_read(&libcfs_kmemory));
2729 kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2730 module_put(THIS_MODULE);
2733 void kiblnd_shutdown(lnet_ni_t *ni)
2735 kib_net_t *net = ni->ni_data;
2736 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2738 unsigned long flags;
2740 LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2745 CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
2746 atomic_read(&libcfs_kmemory));
2748 write_lock_irqsave(g_lock, flags);
2749 net->ibn_shutdown = 1;
2750 write_unlock_irqrestore(g_lock, flags);
2752 switch (net->ibn_init) {
2756 case IBLND_INIT_ALL:
2757 /* nuke all existing peers within this net */
2758 kiblnd_del_peer(ni, LNET_NID_ANY);
2760 /* Wait for all peer state to clean up */
2762 while (atomic_read(&net->ibn_npeers) != 0) {
2764 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* 2**n? */
2765 "%s: waiting for %d peers to disconnect\n",
2766 libcfs_nid2str(ni->ni_nid),
2767 atomic_read(&net->ibn_npeers));
2768 set_current_state(TASK_UNINTERRUPTIBLE);
2769 schedule_timeout(cfs_time_seconds(1));
2772 kiblnd_net_fini_pools(net);
2774 write_lock_irqsave(g_lock, flags);
2775 LASSERT(net->ibn_dev->ibd_nnets > 0);
2776 net->ibn_dev->ibd_nnets--;
2777 list_del(&net->ibn_list);
2778 write_unlock_irqrestore(g_lock, flags);
2782 case IBLND_INIT_NOTHING:
2783 LASSERT(atomic_read(&net->ibn_nconns) == 0);
2785 if (net->ibn_dev != NULL &&
2786 net->ibn_dev->ibd_nnets == 0)
2787 kiblnd_destroy_dev(net->ibn_dev);
2792 CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
2793 atomic_read(&libcfs_kmemory));
2795 net->ibn_init = IBLND_INIT_NOTHING;
2798 LIBCFS_FREE(net, sizeof(*net));
2801 if (list_empty(&kiblnd_data.kib_devs))
2802 kiblnd_base_shutdown();
2805 static int kiblnd_base_startup(void)
2807 struct kib_sched_info *sched;
2811 LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
2813 try_module_get(THIS_MODULE);
2814 /* zero pointers, flags etc */
2815 memset(&kiblnd_data, 0, sizeof(kiblnd_data));
2817 rwlock_init(&kiblnd_data.kib_global_lock);
2819 INIT_LIST_HEAD(&kiblnd_data.kib_devs);
2820 INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
2822 kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
2823 LIBCFS_ALLOC(kiblnd_data.kib_peers,
2824 sizeof(struct list_head) *
2825 kiblnd_data.kib_peer_hash_size);
2826 if (kiblnd_data.kib_peers == NULL)
2828 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2829 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
2831 spin_lock_init(&kiblnd_data.kib_connd_lock);
2832 INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
2833 INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
2834 init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
2835 init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
2837 kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
2839 if (kiblnd_data.kib_scheds == NULL)
2842 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
2845 spin_lock_init(&sched->ibs_lock);
2846 INIT_LIST_HEAD(&sched->ibs_conns);
2847 init_waitqueue_head(&sched->ibs_waitq);
2849 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2850 if (*kiblnd_tunables.kib_nscheds > 0) {
2851 nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
2853 /* max to half of CPUs, another half is reserved for
2854 * upper layer modules */
2855 nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
2858 sched->ibs_nthreads_max = nthrs;
2862 kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
2864 /* lists/ptrs/locks initialised */
2865 kiblnd_data.kib_init = IBLND_INIT_DATA;
2866 /*****************************************************/
2868 rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
2870 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
2874 if (*kiblnd_tunables.kib_dev_failover != 0)
2875 rc = kiblnd_thread_start(kiblnd_failover_thread, NULL,
2879 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
2883 /* flag everything initialised */
2884 kiblnd_data.kib_init = IBLND_INIT_ALL;
2885 /*****************************************************/
2890 kiblnd_base_shutdown();
2894 static int kiblnd_start_schedulers(struct kib_sched_info *sched)
2900 if (sched->ibs_nthreads == 0) {
2901 if (*kiblnd_tunables.kib_nscheds > 0) {
2902 nthrs = sched->ibs_nthreads_max;
2904 nthrs = cfs_cpt_weight(lnet_cpt_table(),
2906 nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
2907 nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
2910 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
2911 /* increase one thread if there is new interface */
2912 nthrs = sched->ibs_nthreads < sched->ibs_nthreads_max;
2915 for (i = 0; i < nthrs; i++) {
2919 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
2920 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
2921 KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
2922 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
2926 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
2927 sched->ibs_cpt, sched->ibs_nthreads + i, rc);
2931 sched->ibs_nthreads += i;
2935 static int kiblnd_dev_start_threads(kib_dev_t *dev, int newdev, __u32 *cpts,
2942 for (i = 0; i < ncpts; i++) {
2943 struct kib_sched_info *sched;
2945 cpt = (cpts == NULL) ? i : cpts[i];
2946 sched = kiblnd_data.kib_scheds[cpt];
2948 if (!newdev && sched->ibs_nthreads > 0)
2951 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
2953 CERROR("Failed to start scheduler threads for %s\n",
2961 static kib_dev_t *kiblnd_dev_search(char *ifname)
2963 kib_dev_t *alias = NULL;
2968 colon = strchr(ifname, ':');
2969 list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
2970 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
2976 colon2 = strchr(dev->ibd_ifname, ':');
2982 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
2993 int kiblnd_startup(lnet_ni_t *ni)
2996 kib_dev_t *ibdev = NULL;
2999 unsigned long flags;
3003 LASSERT(ni->ni_lnd == &the_o2iblnd);
3005 if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
3006 rc = kiblnd_base_startup();
3011 LIBCFS_ALLOC(net, sizeof(*net));
3016 do_gettimeofday(&tv);
3017 net->ibn_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
3019 ni->ni_peertimeout = *kiblnd_tunables.kib_peertimeout;
3020 ni->ni_maxtxcredits = *kiblnd_tunables.kib_credits;
3021 ni->ni_peertxcredits = *kiblnd_tunables.kib_peertxcredits;
3022 ni->ni_peerrtrcredits = *kiblnd_tunables.kib_peerrtrcredits;
3024 if (ni->ni_interfaces[0] != NULL) {
3025 /* Use the IPoIB interface specified in 'networks=' */
3027 CLASSERT(LNET_MAX_INTERFACES > 1);
3028 if (ni->ni_interfaces[1] != NULL) {
3029 CERROR("Multiple interfaces not supported\n");
3033 ifname = ni->ni_interfaces[0];
3035 ifname = *kiblnd_tunables.kib_default_ipif;
3038 if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3039 CERROR("IPoIB interface name too long: %s\n", ifname);
3043 ibdev = kiblnd_dev_search(ifname);
3045 newdev = ibdev == NULL;
3046 /* hmm...create kib_dev even for alias */
3047 if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0)
3048 ibdev = kiblnd_create_dev(ifname);
3053 net->ibn_dev = ibdev;
3054 ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
3056 rc = kiblnd_dev_start_threads(ibdev, newdev,
3057 ni->ni_cpts, ni->ni_ncpts);
3061 rc = kiblnd_net_init_pools(net, ni->ni_cpts, ni->ni_ncpts);
3063 CERROR("Failed to initialize NI pools: %d\n", rc);
3067 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3069 list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3070 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3072 net->ibn_init = IBLND_INIT_ALL;
3077 if (net->ibn_dev == NULL && ibdev != NULL)
3078 kiblnd_destroy_dev(ibdev);
3081 kiblnd_shutdown(ni);
3083 CDEBUG(D_NET, "kiblnd_startup failed\n");
3087 static void __exit kiblnd_module_fini(void)
3089 lnet_unregister_lnd(&the_o2iblnd);
3092 static int __init kiblnd_module_init(void)
3096 CLASSERT(sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
3097 CLASSERT(offsetof(kib_msg_t,
3098 ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3100 CLASSERT(offsetof(kib_msg_t,
3101 ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3104 rc = kiblnd_tunables_init();
3108 lnet_register_lnd(&the_o2iblnd);
3113 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
3114 MODULE_DESCRIPTION("Kernel OpenIB gen2 LND v2.00");
3115 MODULE_LICENSE("GPL");
3117 module_init(kiblnd_module_init);
3118 module_exit(kiblnd_module_fini);