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 = kiblnd_cfg_rdma_frags(peer->ibp_ni);
341 peer->ibp_queue_depth = ni->ni_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)
680 vectors = conn->ibc_cmid->device->num_comp_vectors;
684 mask = cfs_cpt_cpumask(lnet_cpt_table(), cpt);
686 /* hash NID to CPU id in this partition... */
687 ibp_nid = conn->ibc_peer->ibp_nid;
688 off = do_div(ibp_nid, cpumask_weight(mask));
689 for_each_cpu(i, mask) {
699 kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
700 int state, int version)
703 * If the new conn is created successfully it takes over the caller's
704 * ref on 'peer'. It also "owns" 'cmid' and destroys it when it itself
705 * is destroyed. On failure, the caller's ref on 'peer' remains and
706 * she must dispose of 'cmid'. (Actually I'd block forever if I tried
707 * to destroy 'cmid' here since I'm called from the CM which still has
708 * its ref on 'cmid'). */
709 rwlock_t *glock = &kiblnd_data.kib_global_lock;
710 kib_net_t *net = peer->ibp_ni->ni_data;
712 struct ib_qp_init_attr *init_qp_attr;
713 struct kib_sched_info *sched;
714 #ifdef HAVE_IB_CQ_INIT_ATTR
715 struct ib_cq_init_attr cq_attr = {};
724 LASSERT(net != NULL);
725 LASSERT(!in_interrupt());
729 cpt = lnet_cpt_of_nid(peer->ibp_nid);
730 sched = kiblnd_data.kib_scheds[cpt];
732 LASSERT(sched->ibs_nthreads > 0);
734 LIBCFS_CPT_ALLOC(init_qp_attr, lnet_cpt_table(), cpt,
735 sizeof(*init_qp_attr));
736 if (init_qp_attr == NULL) {
737 CERROR("Can't allocate qp_attr for %s\n",
738 libcfs_nid2str(peer->ibp_nid));
742 LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
744 CERROR("Can't allocate connection for %s\n",
745 libcfs_nid2str(peer->ibp_nid));
749 conn->ibc_state = IBLND_CONN_INIT;
750 conn->ibc_version = version;
751 conn->ibc_peer = peer; /* I take the caller's ref */
752 cmid->context = conn; /* for future CM callbacks */
753 conn->ibc_cmid = cmid;
754 conn->ibc_max_frags = peer->ibp_max_frags;
755 conn->ibc_queue_depth = peer->ibp_queue_depth;
757 INIT_LIST_HEAD(&conn->ibc_early_rxs);
758 INIT_LIST_HEAD(&conn->ibc_tx_noops);
759 INIT_LIST_HEAD(&conn->ibc_tx_queue);
760 INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
761 INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
762 INIT_LIST_HEAD(&conn->ibc_active_txs);
763 spin_lock_init(&conn->ibc_lock);
765 LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
766 sizeof(*conn->ibc_connvars));
767 if (conn->ibc_connvars == NULL) {
768 CERROR("Can't allocate in-progress connection state\n");
772 write_lock_irqsave(glock, flags);
773 if (dev->ibd_failover) {
774 write_unlock_irqrestore(glock, flags);
775 CERROR("%s: failover in progress\n", dev->ibd_ifname);
779 if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
780 /* wakeup failover thread and teardown connection */
781 if (kiblnd_dev_can_failover(dev)) {
782 list_add_tail(&dev->ibd_fail_list,
783 &kiblnd_data.kib_failed_devs);
784 wake_up(&kiblnd_data.kib_failover_waitq);
787 write_unlock_irqrestore(glock, flags);
788 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
789 cmid->device->name, dev->ibd_ifname);
793 kiblnd_hdev_addref_locked(dev->ibd_hdev);
794 conn->ibc_hdev = dev->ibd_hdev;
796 kiblnd_setup_mtu_locked(cmid);
798 write_unlock_irqrestore(glock, flags);
800 LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
801 IBLND_RX_MSGS(conn) * sizeof(kib_rx_t));
802 if (conn->ibc_rxs == NULL) {
803 CERROR("Cannot allocate RX buffers\n");
807 rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
808 IBLND_RX_MSG_PAGES(conn));
812 kiblnd_map_rx_descs(conn);
814 #ifdef HAVE_IB_CQ_INIT_ATTR
815 cq_attr.cqe = IBLND_CQ_ENTRIES(conn);
816 cq_attr.comp_vector = kiblnd_get_completion_vector(conn, cpt);
817 cq = ib_create_cq(cmid->device,
818 kiblnd_cq_completion, kiblnd_cq_event, conn,
821 cq = ib_create_cq(cmid->device,
822 kiblnd_cq_completion, kiblnd_cq_event, conn,
823 IBLND_CQ_ENTRIES(conn),
824 kiblnd_get_completion_vector(conn, cpt));
827 CERROR("Failed to create CQ with %d CQEs: %ld\n",
828 IBLND_CQ_ENTRIES(conn), PTR_ERR(cq));
834 rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
836 CERROR("Can't request completion notification: %d\n", rc);
840 init_qp_attr->event_handler = kiblnd_qp_event;
841 init_qp_attr->qp_context = conn;
842 init_qp_attr->cap.max_send_wr = IBLND_SEND_WRS(conn);
843 init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(conn);
844 init_qp_attr->cap.max_send_sge = 1;
845 init_qp_attr->cap.max_recv_sge = 1;
846 init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
847 init_qp_attr->qp_type = IB_QPT_RC;
848 init_qp_attr->send_cq = cq;
849 init_qp_attr->recv_cq = cq;
851 conn->ibc_sched = sched;
854 rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr);
855 if (!rc || init_qp_attr->cap.max_send_wr < 16)
858 init_qp_attr->cap.max_send_wr -= init_qp_attr->cap.max_send_wr / 4;
862 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d\n",
863 rc, init_qp_attr->cap.max_send_wr,
864 init_qp_attr->cap.max_recv_wr);
868 if (init_qp_attr->cap.max_send_wr != IBLND_SEND_WRS(conn))
869 CDEBUG(D_NET, "original send wr %d, created with %d\n",
870 IBLND_SEND_WRS(conn), init_qp_attr->cap.max_send_wr);
872 LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
874 /* 1 ref for caller and each rxmsg */
875 atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(conn));
876 conn->ibc_nrx = IBLND_RX_MSGS(conn);
879 for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
880 rc = kiblnd_post_rx(&conn->ibc_rxs[i], IBLND_POSTRX_NO_CREDIT);
882 CERROR("Can't post rxmsg: %d\n", rc);
884 /* Make posted receives complete */
885 kiblnd_abort_receives(conn);
887 /* correct # of posted buffers
888 * NB locking needed now I'm racing with completion */
889 spin_lock_irqsave(&sched->ibs_lock, flags);
890 conn->ibc_nrx -= IBLND_RX_MSGS(conn) - i;
891 spin_unlock_irqrestore(&sched->ibs_lock, flags);
893 /* cmid will be destroyed by CM(ofed) after cm_callback
894 * returned, so we can't refer it anymore
895 * (by kiblnd_connd()->kiblnd_destroy_conn) */
896 rdma_destroy_qp(conn->ibc_cmid);
897 conn->ibc_cmid = NULL;
899 /* Drop my own and unused rxbuffer refcounts */
900 while (i++ <= IBLND_RX_MSGS(conn))
901 kiblnd_conn_decref(conn);
907 /* Init successful! */
908 LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
909 state == IBLND_CONN_PASSIVE_WAIT);
910 conn->ibc_state = state;
913 atomic_inc(&net->ibn_nconns);
917 kiblnd_destroy_conn(conn, true);
919 LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
925 kiblnd_destroy_conn(kib_conn_t *conn, bool free_conn)
927 struct rdma_cm_id *cmid = conn->ibc_cmid;
928 kib_peer_t *peer = conn->ibc_peer;
931 LASSERT (!in_interrupt());
932 LASSERT (atomic_read(&conn->ibc_refcount) == 0);
933 LASSERT(list_empty(&conn->ibc_early_rxs));
934 LASSERT(list_empty(&conn->ibc_tx_noops));
935 LASSERT(list_empty(&conn->ibc_tx_queue));
936 LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd));
937 LASSERT(list_empty(&conn->ibc_tx_queue_nocred));
938 LASSERT(list_empty(&conn->ibc_active_txs));
939 LASSERT (conn->ibc_noops_posted == 0);
940 LASSERT (conn->ibc_nsends_posted == 0);
942 switch (conn->ibc_state) {
944 /* conn must be completely disengaged from the network */
947 case IBLND_CONN_DISCONNECTED:
948 /* connvars should have been freed already */
949 LASSERT (conn->ibc_connvars == NULL);
952 case IBLND_CONN_INIT:
956 /* conn->ibc_cmid might be destroyed by CM already */
957 if (cmid != NULL && cmid->qp != NULL)
958 rdma_destroy_qp(cmid);
960 if (conn->ibc_cq != NULL) {
961 rc = ib_destroy_cq(conn->ibc_cq);
963 CWARN("Error destroying CQ: %d\n", rc);
966 if (conn->ibc_rx_pages != NULL)
967 kiblnd_unmap_rx_descs(conn);
969 if (conn->ibc_rxs != NULL) {
970 LIBCFS_FREE(conn->ibc_rxs,
971 IBLND_RX_MSGS(conn) * sizeof(kib_rx_t));
974 if (conn->ibc_connvars != NULL)
975 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
977 if (conn->ibc_hdev != NULL)
978 kiblnd_hdev_decref(conn->ibc_hdev);
980 /* See CAVEAT EMPTOR above in kiblnd_create_conn */
981 if (conn->ibc_state != IBLND_CONN_INIT) {
982 kib_net_t *net = peer->ibp_ni->ni_data;
984 kiblnd_peer_decref(peer);
985 rdma_destroy_id(cmid);
986 atomic_dec(&net->ibn_nconns);
990 LIBCFS_FREE(conn, sizeof(*conn));
994 kiblnd_close_peer_conns_locked(kib_peer_t *peer, int why)
997 struct list_head *ctmp;
998 struct list_head *cnxt;
1001 list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
1002 conn = list_entry(ctmp, kib_conn_t, ibc_list);
1004 CDEBUG(D_NET, "Closing conn -> %s, "
1005 "version: %x, reason: %d\n",
1006 libcfs_nid2str(peer->ibp_nid),
1007 conn->ibc_version, why);
1009 kiblnd_close_conn_locked(conn, why);
1017 kiblnd_close_stale_conns_locked(kib_peer_t *peer,
1018 int version, __u64 incarnation)
1021 struct list_head *ctmp;
1022 struct list_head *cnxt;
1025 list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
1026 conn = list_entry(ctmp, kib_conn_t, ibc_list);
1028 if (conn->ibc_version == version &&
1029 conn->ibc_incarnation == incarnation)
1032 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
1033 "incarnation:"LPX64"(%x, "LPX64")\n",
1034 libcfs_nid2str(peer->ibp_nid),
1035 conn->ibc_version, conn->ibc_incarnation,
1036 version, incarnation);
1038 kiblnd_close_conn_locked(conn, -ESTALE);
1046 kiblnd_close_matching_conns(lnet_ni_t *ni, lnet_nid_t nid)
1049 struct list_head *ptmp;
1050 struct list_head *pnxt;
1054 unsigned long flags;
1057 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1059 if (nid != LNET_NID_ANY)
1060 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
1063 hi = kiblnd_data.kib_peer_hash_size - 1;
1066 for (i = lo; i <= hi; i++) {
1067 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
1069 peer = list_entry(ptmp, kib_peer_t, ibp_list);
1070 LASSERT(!kiblnd_peer_idle(peer));
1072 if (peer->ibp_ni != ni)
1075 if (!(nid == LNET_NID_ANY || nid == peer->ibp_nid))
1078 count += kiblnd_close_peer_conns_locked(peer, 0);
1082 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1084 /* wildcards always succeed */
1085 if (nid == LNET_NID_ANY)
1088 return (count == 0) ? -ENOENT : 0;
1092 kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
1094 struct libcfs_ioctl_data *data = arg;
1098 case IOC_LIBCFS_GET_PEER: {
1102 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1104 data->ioc_nid = nid;
1105 data->ioc_count = count;
1109 case IOC_LIBCFS_DEL_PEER: {
1110 rc = kiblnd_del_peer(ni, data->ioc_nid);
1113 case IOC_LIBCFS_GET_CONN: {
1117 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1123 LASSERT (conn->ibc_cmid != NULL);
1124 data->ioc_nid = conn->ibc_peer->ibp_nid;
1125 if (conn->ibc_cmid->route.path_rec == NULL)
1126 data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1129 ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1130 kiblnd_conn_decref(conn);
1133 case IOC_LIBCFS_CLOSE_CONNECTION: {
1134 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1146 kiblnd_query(lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when)
1148 cfs_time_t last_alive = 0;
1149 cfs_time_t now = cfs_time_current();
1150 rwlock_t *glock = &kiblnd_data.kib_global_lock;
1152 unsigned long flags;
1154 read_lock_irqsave(glock, flags);
1156 peer = kiblnd_find_peer_locked(nid);
1158 last_alive = peer->ibp_last_alive;
1160 read_unlock_irqrestore(glock, flags);
1162 if (last_alive != 0)
1165 /* peer is not persistent in hash, trigger peer creation
1166 * and connection establishment with a NULL tx */
1168 kiblnd_launch_tx(ni, NULL, nid);
1170 CDEBUG(D_NET, "Peer %s %p, alive %ld secs ago\n",
1171 libcfs_nid2str(nid), peer,
1172 last_alive ? cfs_duration_sec(now - last_alive) : -1);
1177 kiblnd_free_pages(kib_pages_t *p)
1179 int npages = p->ibp_npages;
1182 for (i = 0; i < npages; i++) {
1183 if (p->ibp_pages[i] != NULL)
1184 __free_page(p->ibp_pages[i]);
1187 LIBCFS_FREE(p, offsetof(kib_pages_t, ibp_pages[npages]));
1191 kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages)
1196 LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1197 offsetof(kib_pages_t, ibp_pages[npages]));
1199 CERROR("Can't allocate descriptor for %d pages\n", npages);
1203 memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1204 p->ibp_npages = npages;
1206 for (i = 0; i < npages; i++) {
1207 p->ibp_pages[i] = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1209 if (p->ibp_pages[i] == NULL) {
1210 CERROR("Can't allocate page %d of %d\n", i, npages);
1211 kiblnd_free_pages(p);
1221 kiblnd_unmap_rx_descs(kib_conn_t *conn)
1226 LASSERT (conn->ibc_rxs != NULL);
1227 LASSERT (conn->ibc_hdev != NULL);
1229 for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
1230 rx = &conn->ibc_rxs[i];
1232 LASSERT(rx->rx_nob >= 0); /* not posted */
1234 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1235 KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1237 IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1240 kiblnd_free_pages(conn->ibc_rx_pages);
1242 conn->ibc_rx_pages = NULL;
1246 kiblnd_map_rx_descs(kib_conn_t *conn)
1254 for (pg_off = ipg = i = 0; i < IBLND_RX_MSGS(conn); i++) {
1255 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1256 rx = &conn->ibc_rxs[i];
1259 rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
1262 kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1263 rx->rx_msg, IBLND_MSG_SIZE,
1265 LASSERT(!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1267 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1269 CDEBUG(D_NET, "rx %d: %p "LPX64"("LPX64")\n",
1270 i, rx->rx_msg, rx->rx_msgaddr,
1271 (__u64)(page_to_phys(pg) + pg_off));
1273 pg_off += IBLND_MSG_SIZE;
1274 LASSERT(pg_off <= PAGE_SIZE);
1276 if (pg_off == PAGE_SIZE) {
1279 LASSERT(ipg <= IBLND_RX_MSG_PAGES(conn));
1285 kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo)
1287 kib_hca_dev_t *hdev = tpo->tpo_hdev;
1291 LASSERT (tpo->tpo_pool.po_allocated == 0);
1296 for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1297 tx = &tpo->tpo_tx_descs[i];
1298 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1299 KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1301 IBLND_MSG_SIZE, DMA_TO_DEVICE);
1304 kiblnd_hdev_decref(hdev);
1305 tpo->tpo_hdev = NULL;
1308 static kib_hca_dev_t *
1309 kiblnd_current_hdev(kib_dev_t *dev)
1311 kib_hca_dev_t *hdev;
1312 unsigned long flags;
1315 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1316 while (dev->ibd_failover) {
1317 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1319 CDEBUG(D_NET, "%s: Wait for failover\n",
1321 set_current_state(TASK_INTERRUPTIBLE);
1322 schedule_timeout(cfs_time_seconds(1) / 100);
1324 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1327 kiblnd_hdev_addref_locked(dev->ibd_hdev);
1328 hdev = dev->ibd_hdev;
1330 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1336 kiblnd_map_tx_pool(kib_tx_pool_t *tpo)
1338 kib_pages_t *txpgs = tpo->tpo_tx_pages;
1339 kib_pool_t *pool = &tpo->tpo_pool;
1340 kib_net_t *net = pool->po_owner->ps_net;
1348 LASSERT (net != NULL);
1352 /* pre-mapped messages are not bigger than 1 page */
1353 CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE);
1355 /* No fancy arithmetic when we do the buffer calculations */
1356 CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
1358 tpo->tpo_hdev = kiblnd_current_hdev(dev);
1360 for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1361 page = txpgs->ibp_pages[ipage];
1362 tx = &tpo->tpo_tx_descs[i];
1364 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1367 tx->tx_msgaddr = kiblnd_dma_map_single(tpo->tpo_hdev->ibh_ibdev,
1371 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1373 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1375 list_add(&tx->tx_list, &pool->po_free_list);
1377 page_offset += IBLND_MSG_SIZE;
1378 LASSERT(page_offset <= PAGE_SIZE);
1380 if (page_offset == PAGE_SIZE) {
1383 LASSERT(ipage <= txpgs->ibp_npages);
1389 kiblnd_find_rd_dma_mr(struct lnet_ni *ni, kib_rdma_desc_t *rd,
1390 int negotiated_nfrags)
1392 kib_net_t *net = ni->ni_data;
1393 kib_hca_dev_t *hdev = net->ibn_dev->ibd_hdev;
1394 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
1398 tunables = &ni->ni_lnd_tunables->lt_tun_u.lt_o2ib;
1399 mod = tunables->lnd_map_on_demand;
1400 nfrags = (negotiated_nfrags != -1) ? negotiated_nfrags : mod;
1402 LASSERT(hdev->ibh_mrs != NULL);
1404 if (mod > 0 && nfrags <= rd->rd_nfrags)
1407 return hdev->ibh_mrs;
1411 kiblnd_destroy_fmr_pool(kib_fmr_pool_t *fpo)
1413 LASSERT(fpo->fpo_map_count == 0);
1415 if (fpo->fpo_is_fmr) {
1416 if (fpo->fmr.fpo_fmr_pool)
1417 ib_destroy_fmr_pool(fpo->fmr.fpo_fmr_pool);
1419 struct kib_fast_reg_descriptor *frd, *tmp;
1422 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1424 list_del(&frd->frd_list);
1425 ib_free_fast_reg_page_list(frd->frd_frpl);
1426 ib_dereg_mr(frd->frd_mr);
1427 LIBCFS_FREE(frd, sizeof(*frd));
1430 if (i < fpo->fast_reg.fpo_pool_size)
1431 CERROR("FastReg pool still has %d regions registered\n",
1432 fpo->fast_reg.fpo_pool_size - i);
1436 kiblnd_hdev_decref(fpo->fpo_hdev);
1438 LIBCFS_FREE(fpo, sizeof(*fpo));
1442 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1444 kib_fmr_pool_t *fpo, *tmp;
1446 list_for_each_entry_safe(fpo, tmp, head, fpo_list) {
1447 list_del(&fpo->fpo_list);
1448 kiblnd_destroy_fmr_pool(fpo);
1453 kiblnd_fmr_pool_size(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1456 int size = tunables->lnd_fmr_pool_size / ncpts;
1458 return max(IBLND_FMR_POOL, size);
1462 kiblnd_fmr_flush_trigger(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1465 int size = tunables->lnd_fmr_flush_trigger / ncpts;
1467 return max(IBLND_FMR_POOL_FLUSH, size);
1470 static int kiblnd_alloc_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t *fpo)
1472 struct ib_fmr_pool_param param = {
1473 .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
1474 .page_shift = PAGE_SHIFT,
1475 .access = (IB_ACCESS_LOCAL_WRITE |
1476 IB_ACCESS_REMOTE_WRITE),
1477 .pool_size = fps->fps_pool_size,
1478 .dirty_watermark = fps->fps_flush_trigger,
1479 .flush_function = NULL,
1481 .cache = !!fps->fps_cache };
1484 fpo->fmr.fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd,
1486 if (IS_ERR(fpo->fmr.fpo_fmr_pool)) {
1487 rc = PTR_ERR(fpo->fmr.fpo_fmr_pool);
1489 CERROR("Failed to create FMR pool: %d\n", rc);
1491 CERROR("FMRs are not supported\n");
1497 static int kiblnd_alloc_freg_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t *fpo)
1499 struct kib_fast_reg_descriptor *frd, *tmp;
1502 INIT_LIST_HEAD(&fpo->fast_reg.fpo_pool_list);
1503 fpo->fast_reg.fpo_pool_size = 0;
1504 for (i = 0; i < fps->fps_pool_size; i++) {
1505 LIBCFS_CPT_ALLOC(frd, lnet_cpt_table(), fps->fps_cpt,
1508 CERROR("Failed to allocate a new fast_reg descriptor\n");
1514 frd->frd_frpl = ib_alloc_fast_reg_page_list(fpo->fpo_hdev->ibh_ibdev,
1515 LNET_MAX_PAYLOAD/PAGE_SIZE);
1516 if (IS_ERR(frd->frd_frpl)) {
1517 rc = PTR_ERR(frd->frd_frpl);
1518 CERROR("Failed to allocate ib_fast_reg_page_list: %d\n",
1520 frd->frd_frpl = NULL;
1524 frd->frd_mr = ib_alloc_fast_reg_mr(fpo->fpo_hdev->ibh_pd,
1525 LNET_MAX_PAYLOAD/PAGE_SIZE);
1526 if (IS_ERR(frd->frd_mr)) {
1527 rc = PTR_ERR(frd->frd_mr);
1528 CERROR("Failed to allocate ib_fast_reg_mr: %d\n", rc);
1533 frd->frd_valid = true;
1535 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1536 fpo->fast_reg.fpo_pool_size++;
1543 ib_dereg_mr(frd->frd_mr);
1545 ib_free_fast_reg_page_list(frd->frd_frpl);
1546 LIBCFS_FREE(frd, sizeof(*frd));
1549 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1551 list_del(&frd->frd_list);
1552 ib_free_fast_reg_page_list(frd->frd_frpl);
1553 ib_dereg_mr(frd->frd_mr);
1554 LIBCFS_FREE(frd, sizeof(*frd));
1561 kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo)
1563 struct ib_device_attr *dev_attr;
1564 kib_dev_t *dev = fps->fps_net->ibn_dev;
1565 kib_fmr_pool_t *fpo;
1568 dev_attr = kmalloc(sizeof(*dev_attr), GFP_KERNEL);
1572 LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1578 fpo->fpo_hdev = kiblnd_current_hdev(dev);
1580 rc = ib_query_device(fpo->fpo_hdev->ibh_ibdev, dev_attr);
1582 CERROR("Query device failed for %s: %d\n",
1583 fpo->fpo_hdev->ibh_ibdev->name, rc);
1587 /* Check for FMR or FastReg support */
1588 fpo->fpo_is_fmr = 0;
1589 if (fpo->fpo_hdev->ibh_ibdev->alloc_fmr &&
1590 fpo->fpo_hdev->ibh_ibdev->dealloc_fmr &&
1591 fpo->fpo_hdev->ibh_ibdev->map_phys_fmr &&
1592 fpo->fpo_hdev->ibh_ibdev->unmap_fmr) {
1593 LCONSOLE_INFO("Using FMR for registration\n");
1594 fpo->fpo_is_fmr = 1;
1595 } else if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
1596 LCONSOLE_INFO("Using FastReg for registration\n");
1599 LCONSOLE_ERROR_MSG(rc, "IB device does not support FMRs nor FastRegs, can't register memory\n");
1603 if (fpo->fpo_is_fmr)
1604 rc = kiblnd_alloc_fmr_pool(fps, fpo);
1606 rc = kiblnd_alloc_freg_pool(fps, fpo);
1611 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1612 fpo->fpo_owner = fps;
1618 kiblnd_hdev_decref(fpo->fpo_hdev);
1619 LIBCFS_FREE(fpo, sizeof(*fpo));
1628 kiblnd_fail_fmr_poolset(kib_fmr_poolset_t *fps, struct list_head *zombies)
1630 if (fps->fps_net == NULL) /* intialized? */
1633 spin_lock(&fps->fps_lock);
1635 while (!list_empty(&fps->fps_pool_list)) {
1636 kib_fmr_pool_t *fpo = list_entry(fps->fps_pool_list.next,
1637 kib_fmr_pool_t, fpo_list);
1638 fpo->fpo_failed = 1;
1639 list_del(&fpo->fpo_list);
1640 if (fpo->fpo_map_count == 0)
1641 list_add(&fpo->fpo_list, zombies);
1643 list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1646 spin_unlock(&fps->fps_lock);
1650 kiblnd_fini_fmr_poolset(kib_fmr_poolset_t *fps)
1652 if (fps->fps_net != NULL) { /* initialized? */
1653 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1654 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1659 kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt, int ncpts,
1661 struct lnet_ioctl_config_o2iblnd_tunables *tunables)
1663 kib_fmr_pool_t *fpo;
1666 memset(fps, 0, sizeof(kib_fmr_poolset_t));
1671 fps->fps_pool_size = kiblnd_fmr_pool_size(tunables, ncpts);
1672 fps->fps_flush_trigger = kiblnd_fmr_flush_trigger(tunables, ncpts);
1673 fps->fps_cache = tunables->lnd_fmr_cache;
1675 spin_lock_init(&fps->fps_lock);
1676 INIT_LIST_HEAD(&fps->fps_pool_list);
1677 INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1679 rc = kiblnd_create_fmr_pool(fps, &fpo);
1681 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1687 kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, cfs_time_t now)
1689 if (fpo->fpo_map_count != 0) /* still in use */
1691 if (fpo->fpo_failed)
1693 return cfs_time_aftereq(now, fpo->fpo_deadline);
1697 kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1699 struct list_head zombies = LIST_HEAD_INIT(zombies);
1700 kib_fmr_pool_t *fpo = fmr->fmr_pool;
1701 kib_fmr_poolset_t *fps;
1702 cfs_time_t now = cfs_time_current();
1703 kib_fmr_pool_t *tmp;
1709 fps = fpo->fpo_owner;
1710 if (fpo->fpo_is_fmr) {
1711 if (fmr->fmr_pfmr) {
1712 rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1714 fmr->fmr_pfmr = NULL;
1718 rc = ib_flush_fmr_pool(fpo->fmr.fpo_fmr_pool);
1722 struct kib_fast_reg_descriptor *frd = fmr->fmr_frd;
1725 frd->frd_valid = false;
1726 spin_lock(&fps->fps_lock);
1727 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1728 spin_unlock(&fps->fps_lock);
1729 fmr->fmr_frd = NULL;
1732 fmr->fmr_pool = NULL;
1734 spin_lock(&fps->fps_lock);
1735 fpo->fpo_map_count--; /* decref the pool */
1737 list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1738 /* the first pool is persistent */
1739 if (fps->fps_pool_list.next == &fpo->fpo_list)
1742 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1743 list_move(&fpo->fpo_list, &zombies);
1747 spin_unlock(&fps->fps_lock);
1749 if (!list_empty(&zombies))
1750 kiblnd_destroy_fmr_pool_list(&zombies);
1754 kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages,
1755 __u32 nob, __u64 iov, bool is_rx, kib_fmr_t *fmr)
1757 kib_fmr_pool_t *fpo;
1762 spin_lock(&fps->fps_lock);
1763 version = fps->fps_version;
1764 list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1765 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1766 fpo->fpo_map_count++;
1768 if (fpo->fpo_is_fmr) {
1769 struct ib_pool_fmr *pfmr;
1771 spin_unlock(&fps->fps_lock);
1772 pfmr = ib_fmr_pool_map_phys(fpo->fmr.fpo_fmr_pool,
1773 pages, npages, iov);
1774 if (likely(!IS_ERR(pfmr))) {
1775 fmr->fmr_key = is_rx ? pfmr->fmr->rkey
1777 fmr->fmr_frd = NULL;
1778 fmr->fmr_pfmr = pfmr;
1779 fmr->fmr_pool = fpo;
1784 if (!list_empty(&fpo->fast_reg.fpo_pool_list)) {
1785 struct ib_send_wr *wr;
1786 struct kib_fast_reg_descriptor *frd;
1787 struct ib_fast_reg_page_list *frpl;
1790 frd = list_first_entry(&fpo->fast_reg.fpo_pool_list,
1791 struct kib_fast_reg_descriptor,
1793 list_del(&frd->frd_list);
1794 spin_unlock(&fps->fps_lock);
1796 frpl = frd->frd_frpl;
1799 if (!frd->frd_valid) {
1800 struct ib_send_wr *inv_wr;
1801 __u32 key = is_rx ? mr->rkey : mr->lkey;
1803 inv_wr = &frd->frd_inv_wr;
1804 memset(inv_wr, 0, sizeof(*inv_wr));
1805 inv_wr->opcode = IB_WR_LOCAL_INV;
1806 inv_wr->wr_id = IBLND_WID_MR;
1807 inv_wr->ex.invalidate_rkey = key;
1810 key = ib_inc_rkey(key);
1811 ib_update_fast_reg_key(mr, key);
1814 LASSERT(npages <= frpl->max_page_list_len);
1815 memcpy(frpl->page_list, pages,
1816 sizeof(*pages) * npages);
1818 /* Prepare FastReg WR */
1819 wr = &frd->frd_fastreg_wr;
1820 memset(wr, 0, sizeof(*wr));
1821 wr->opcode = IB_WR_FAST_REG_MR;
1822 wr->wr_id = IBLND_WID_MR;
1823 wr->wr.fast_reg.iova_start = iov;
1824 wr->wr.fast_reg.page_list = frpl;
1825 wr->wr.fast_reg.page_list_len = npages;
1826 wr->wr.fast_reg.page_shift = PAGE_SHIFT;
1827 wr->wr.fast_reg.length = nob;
1828 wr->wr.fast_reg.rkey = is_rx ? mr->rkey
1830 wr->wr.fast_reg.access_flags =
1831 (IB_ACCESS_LOCAL_WRITE |
1832 IB_ACCESS_REMOTE_WRITE);
1834 fmr->fmr_key = is_rx ? mr->rkey : mr->lkey;
1836 fmr->fmr_pfmr = NULL;
1837 fmr->fmr_pool = fpo;
1840 spin_unlock(&fps->fps_lock);
1844 spin_lock(&fps->fps_lock);
1845 fpo->fpo_map_count--;
1846 if (rc != -EAGAIN) {
1847 spin_unlock(&fps->fps_lock);
1851 /* EAGAIN and ... */
1852 if (version != fps->fps_version) {
1853 spin_unlock(&fps->fps_lock);
1858 if (fps->fps_increasing) {
1859 spin_unlock(&fps->fps_lock);
1860 CDEBUG(D_NET, "Another thread is allocating new "
1861 "FMR pool, waiting for her to complete\n");
1867 if (cfs_time_before(cfs_time_current(), fps->fps_next_retry)) {
1868 /* someone failed recently */
1869 spin_unlock(&fps->fps_lock);
1873 fps->fps_increasing = 1;
1874 spin_unlock(&fps->fps_lock);
1876 CDEBUG(D_NET, "Allocate new FMR pool\n");
1877 rc = kiblnd_create_fmr_pool(fps, &fpo);
1878 spin_lock(&fps->fps_lock);
1879 fps->fps_increasing = 0;
1882 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1884 fps->fps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1886 spin_unlock(&fps->fps_lock);
1892 kiblnd_fini_pool(kib_pool_t *pool)
1894 LASSERT(list_empty(&pool->po_free_list));
1895 LASSERT(pool->po_allocated == 0);
1897 CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
1901 kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
1903 CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
1905 memset(pool, 0, sizeof(kib_pool_t));
1906 INIT_LIST_HEAD(&pool->po_free_list);
1907 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1908 pool->po_owner = ps;
1909 pool->po_size = size;
1913 kiblnd_destroy_pool_list(struct list_head *head)
1917 while (!list_empty(head)) {
1918 pool = list_entry(head->next, kib_pool_t, po_list);
1919 list_del(&pool->po_list);
1921 LASSERT(pool->po_owner != NULL);
1922 pool->po_owner->ps_pool_destroy(pool);
1927 kiblnd_fail_poolset(kib_poolset_t *ps, struct list_head *zombies)
1929 if (ps->ps_net == NULL) /* intialized? */
1932 spin_lock(&ps->ps_lock);
1933 while (!list_empty(&ps->ps_pool_list)) {
1934 kib_pool_t *po = list_entry(ps->ps_pool_list.next,
1935 kib_pool_t, po_list);
1937 list_del(&po->po_list);
1938 if (po->po_allocated == 0)
1939 list_add(&po->po_list, zombies);
1941 list_add(&po->po_list, &ps->ps_failed_pool_list);
1943 spin_unlock(&ps->ps_lock);
1947 kiblnd_fini_poolset(kib_poolset_t *ps)
1949 if (ps->ps_net != NULL) { /* initialized? */
1950 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
1951 kiblnd_destroy_pool_list(&ps->ps_pool_list);
1956 kiblnd_init_poolset(kib_poolset_t *ps, int cpt,
1957 kib_net_t *net, char *name, int size,
1958 kib_ps_pool_create_t po_create,
1959 kib_ps_pool_destroy_t po_destroy,
1960 kib_ps_node_init_t nd_init,
1961 kib_ps_node_fini_t nd_fini)
1966 memset(ps, 0, sizeof(kib_poolset_t));
1970 ps->ps_pool_create = po_create;
1971 ps->ps_pool_destroy = po_destroy;
1972 ps->ps_node_init = nd_init;
1973 ps->ps_node_fini = nd_fini;
1974 ps->ps_pool_size = size;
1975 if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
1976 >= sizeof(ps->ps_name))
1978 spin_lock_init(&ps->ps_lock);
1979 INIT_LIST_HEAD(&ps->ps_pool_list);
1980 INIT_LIST_HEAD(&ps->ps_failed_pool_list);
1982 rc = ps->ps_pool_create(ps, size, &pool);
1984 list_add(&pool->po_list, &ps->ps_pool_list);
1986 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
1992 kiblnd_pool_is_idle(kib_pool_t *pool, cfs_time_t now)
1994 if (pool->po_allocated != 0) /* still in use */
1996 if (pool->po_failed)
1998 return cfs_time_aftereq(now, pool->po_deadline);
2002 kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node)
2004 struct list_head zombies = LIST_HEAD_INIT(zombies);
2005 kib_poolset_t *ps = pool->po_owner;
2007 cfs_time_t now = cfs_time_current();
2009 spin_lock(&ps->ps_lock);
2011 if (ps->ps_node_fini != NULL)
2012 ps->ps_node_fini(pool, node);
2014 LASSERT(pool->po_allocated > 0);
2015 list_add(node, &pool->po_free_list);
2016 pool->po_allocated--;
2018 list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
2019 /* the first pool is persistent */
2020 if (ps->ps_pool_list.next == &pool->po_list)
2023 if (kiblnd_pool_is_idle(pool, now))
2024 list_move(&pool->po_list, &zombies);
2026 spin_unlock(&ps->ps_lock);
2028 if (!list_empty(&zombies))
2029 kiblnd_destroy_pool_list(&zombies);
2033 kiblnd_pool_alloc_node(kib_poolset_t *ps)
2035 struct list_head *node;
2038 unsigned int interval = 1;
2039 cfs_time_t time_before;
2040 unsigned int trips = 0;
2043 spin_lock(&ps->ps_lock);
2044 list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
2045 if (list_empty(&pool->po_free_list))
2048 pool->po_allocated++;
2049 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
2050 node = pool->po_free_list.next;
2053 if (ps->ps_node_init != NULL) {
2054 /* still hold the lock */
2055 ps->ps_node_init(pool, node);
2057 spin_unlock(&ps->ps_lock);
2061 /* no available tx pool and ... */
2062 if (ps->ps_increasing) {
2063 /* another thread is allocating a new pool */
2064 spin_unlock(&ps->ps_lock);
2066 CDEBUG(D_NET, "Another thread is allocating new "
2067 "%s pool, waiting %d HZs for her to complete."
2069 ps->ps_name, interval, trips);
2071 set_current_state(TASK_INTERRUPTIBLE);
2072 schedule_timeout(interval);
2073 if (interval < cfs_time_seconds(1))
2079 if (cfs_time_before(cfs_time_current(), ps->ps_next_retry)) {
2080 /* someone failed recently */
2081 spin_unlock(&ps->ps_lock);
2085 ps->ps_increasing = 1;
2086 spin_unlock(&ps->ps_lock);
2088 CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
2089 time_before = cfs_time_current();
2090 rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
2091 CDEBUG(D_NET, "ps_pool_create took %lu HZ to complete",
2092 cfs_time_current() - time_before);
2094 spin_lock(&ps->ps_lock);
2095 ps->ps_increasing = 0;
2097 list_add_tail(&pool->po_list, &ps->ps_pool_list);
2099 ps->ps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
2100 CERROR("Can't allocate new %s pool because out of memory\n",
2103 spin_unlock(&ps->ps_lock);
2109 kiblnd_destroy_tx_pool(kib_pool_t *pool)
2111 kib_tx_pool_t *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
2114 LASSERT (pool->po_allocated == 0);
2116 if (tpo->tpo_tx_pages != NULL) {
2117 kiblnd_unmap_tx_pool(tpo);
2118 kiblnd_free_pages(tpo->tpo_tx_pages);
2121 if (tpo->tpo_tx_descs == NULL)
2124 for (i = 0; i < pool->po_size; i++) {
2125 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2127 list_del(&tx->tx_list);
2128 if (tx->tx_pages != NULL)
2129 LIBCFS_FREE(tx->tx_pages,
2131 sizeof(*tx->tx_pages));
2132 if (tx->tx_frags != NULL)
2133 LIBCFS_FREE(tx->tx_frags,
2134 IBLND_MAX_RDMA_FRAGS *
2135 sizeof(*tx->tx_frags));
2136 if (tx->tx_wrq != NULL)
2137 LIBCFS_FREE(tx->tx_wrq,
2138 (1 + IBLND_MAX_RDMA_FRAGS) *
2139 sizeof(*tx->tx_wrq));
2140 if (tx->tx_sge != NULL)
2141 LIBCFS_FREE(tx->tx_sge,
2142 (1 + IBLND_MAX_RDMA_FRAGS) *
2143 sizeof(*tx->tx_sge));
2144 if (tx->tx_rd != NULL)
2145 LIBCFS_FREE(tx->tx_rd,
2146 offsetof(kib_rdma_desc_t,
2147 rd_frags[IBLND_MAX_RDMA_FRAGS]));
2150 LIBCFS_FREE(tpo->tpo_tx_descs,
2151 pool->po_size * sizeof(kib_tx_t));
2153 kiblnd_fini_pool(pool);
2154 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2157 static int kiblnd_tx_pool_size(int ncpts)
2159 int ntx = *kiblnd_tunables.kib_ntx / ncpts;
2161 return max(IBLND_TX_POOL, ntx);
2165 kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
2172 LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2174 CERROR("Failed to allocate TX pool\n");
2178 pool = &tpo->tpo_pool;
2179 kiblnd_init_pool(ps, pool, size);
2180 tpo->tpo_tx_descs = NULL;
2181 tpo->tpo_tx_pages = NULL;
2183 npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2184 if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
2185 CERROR("Can't allocate tx pages: %d\n", npg);
2186 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2190 LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2191 size * sizeof(kib_tx_t));
2192 if (tpo->tpo_tx_descs == NULL) {
2193 CERROR("Can't allocate %d tx descriptors\n", size);
2194 ps->ps_pool_destroy(pool);
2198 memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
2200 for (i = 0; i < size; i++) {
2201 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2204 if (ps->ps_net->ibn_fmr_ps != NULL) {
2205 LIBCFS_CPT_ALLOC(tx->tx_pages,
2206 lnet_cpt_table(), ps->ps_cpt,
2207 LNET_MAX_IOV * sizeof(*tx->tx_pages));
2208 if (tx->tx_pages == NULL)
2212 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2213 IBLND_MAX_RDMA_FRAGS * sizeof(*tx->tx_frags));
2214 if (tx->tx_frags == NULL)
2217 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS);
2219 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2220 (1 + IBLND_MAX_RDMA_FRAGS) *
2221 sizeof(*tx->tx_wrq));
2222 if (tx->tx_wrq == NULL)
2225 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2226 (1 + IBLND_MAX_RDMA_FRAGS) *
2227 sizeof(*tx->tx_sge));
2228 if (tx->tx_sge == NULL)
2231 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2232 offsetof(kib_rdma_desc_t,
2233 rd_frags[IBLND_MAX_RDMA_FRAGS]));
2234 if (tx->tx_rd == NULL)
2239 kiblnd_map_tx_pool(tpo);
2244 ps->ps_pool_destroy(pool);
2249 kiblnd_tx_init(kib_pool_t *pool, struct list_head *node)
2251 kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t,
2253 kib_tx_t *tx = list_entry(node, kib_tx_t, tx_list);
2255 tx->tx_cookie = tps->tps_next_tx_cookie++;
2259 kiblnd_net_fini_pools(kib_net_t *net)
2263 cfs_cpt_for_each(i, lnet_cpt_table()) {
2264 kib_tx_poolset_t *tps;
2265 kib_fmr_poolset_t *fps;
2267 if (net->ibn_tx_ps != NULL) {
2268 tps = net->ibn_tx_ps[i];
2269 kiblnd_fini_poolset(&tps->tps_poolset);
2272 if (net->ibn_fmr_ps != NULL) {
2273 fps = net->ibn_fmr_ps[i];
2274 kiblnd_fini_fmr_poolset(fps);
2278 if (net->ibn_tx_ps != NULL) {
2279 cfs_percpt_free(net->ibn_tx_ps);
2280 net->ibn_tx_ps = NULL;
2283 if (net->ibn_fmr_ps != NULL) {
2284 cfs_percpt_free(net->ibn_fmr_ps);
2285 net->ibn_fmr_ps = NULL;
2290 kiblnd_net_init_pools(kib_net_t *net, lnet_ni_t *ni, __u32 *cpts, int ncpts)
2292 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2293 unsigned long flags;
2298 tunables = &ni->ni_lnd_tunables->lt_tun_u.lt_o2ib;
2300 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2301 if (tunables->lnd_map_on_demand == 0) {
2302 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2304 goto create_tx_pool;
2307 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2309 if (tunables->lnd_fmr_pool_size < *kiblnd_tunables.kib_ntx / 4) {
2310 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2311 tunables->lnd_fmr_pool_size,
2312 *kiblnd_tunables.kib_ntx / 4);
2317 /* TX pool must be created later than FMR, see LU-2268
2319 LASSERT(net->ibn_tx_ps == NULL);
2321 /* premapping can fail if ibd_nmr > 1, so we always create
2322 * FMR pool and map-on-demand if premapping failed */
2324 net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2325 sizeof(kib_fmr_poolset_t));
2326 if (net->ibn_fmr_ps == NULL) {
2327 CERROR("Failed to allocate FMR pool array\n");
2332 for (i = 0; i < ncpts; i++) {
2333 cpt = (cpts == NULL) ? i : cpts[i];
2334 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, ncpts,
2337 CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2344 LASSERT(i == ncpts);
2347 net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2348 sizeof(kib_tx_poolset_t));
2349 if (net->ibn_tx_ps == NULL) {
2350 CERROR("Failed to allocate tx pool array\n");
2355 for (i = 0; i < ncpts; i++) {
2356 cpt = (cpts == NULL) ? i : cpts[i];
2357 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2359 kiblnd_tx_pool_size(ncpts),
2360 kiblnd_create_tx_pool,
2361 kiblnd_destroy_tx_pool,
2362 kiblnd_tx_init, NULL);
2364 CERROR("Can't initialize TX pool for CPT %d: %d\n",
2372 kiblnd_net_fini_pools(net);
2378 kiblnd_hdev_get_attr(kib_hca_dev_t *hdev)
2380 struct ib_device_attr *attr;
2383 /* It's safe to assume a HCA can handle a page size
2384 * matching that of the native system */
2385 hdev->ibh_page_shift = PAGE_SHIFT;
2386 hdev->ibh_page_size = 1 << PAGE_SHIFT;
2387 hdev->ibh_page_mask = ~((__u64)hdev->ibh_page_size - 1);
2389 LIBCFS_ALLOC(attr, sizeof(*attr));
2391 CERROR("Out of memory\n");
2395 rc = ib_query_device(hdev->ibh_ibdev, attr);
2397 hdev->ibh_mr_size = attr->max_mr_size;
2399 LIBCFS_FREE(attr, sizeof(*attr));
2402 CERROR("Failed to query IB device: %d\n", rc);
2406 if (hdev->ibh_mr_size == ~0ULL) {
2407 hdev->ibh_mr_shift = 64;
2411 CERROR("Invalid mr size: "LPX64"\n", hdev->ibh_mr_size);
2416 kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev)
2418 if (hdev->ibh_mrs == NULL)
2421 ib_dereg_mr(hdev->ibh_mrs);
2423 hdev->ibh_mrs = NULL;
2427 kiblnd_hdev_destroy(kib_hca_dev_t *hdev)
2429 kiblnd_hdev_cleanup_mrs(hdev);
2431 if (hdev->ibh_pd != NULL)
2432 ib_dealloc_pd(hdev->ibh_pd);
2434 if (hdev->ibh_cmid != NULL)
2435 rdma_destroy_id(hdev->ibh_cmid);
2437 LIBCFS_FREE(hdev, sizeof(*hdev));
2441 kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev)
2445 int acflags = IB_ACCESS_LOCAL_WRITE |
2446 IB_ACCESS_REMOTE_WRITE;
2448 rc = kiblnd_hdev_get_attr(hdev);
2452 mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2454 CERROR("Failed ib_get_dma_mr: %ld\n", PTR_ERR(mr));
2455 kiblnd_hdev_cleanup_mrs(hdev);
2465 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2471 kiblnd_dev_need_failover(kib_dev_t *dev)
2473 struct rdma_cm_id *cmid;
2474 struct sockaddr_in srcaddr;
2475 struct sockaddr_in dstaddr;
2478 if (dev->ibd_hdev == NULL || /* initializing */
2479 dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2480 *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2483 /* XXX: it's UGLY, but I don't have better way to find
2484 * ib-bonding HCA failover because:
2486 * a. no reliable CM event for HCA failover...
2487 * b. no OFED API to get ib_device for current net_device...
2489 * We have only two choices at this point:
2491 * a. rdma_bind_addr(), it will conflict with listener cmid
2492 * b. rdma_resolve_addr() to zero addr */
2493 cmid = kiblnd_rdma_create_id(kiblnd_dummy_callback, dev, RDMA_PS_TCP,
2497 CERROR("Failed to create cmid for failover: %d\n", rc);
2501 memset(&srcaddr, 0, sizeof(srcaddr));
2502 srcaddr.sin_family = AF_INET;
2503 srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2505 memset(&dstaddr, 0, sizeof(dstaddr));
2506 dstaddr.sin_family = AF_INET;
2507 rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2508 (struct sockaddr *)&dstaddr, 1);
2509 if (rc != 0 || cmid->device == NULL) {
2510 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2511 dev->ibd_ifname, &dev->ibd_ifip,
2513 rdma_destroy_id(cmid);
2517 rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2518 rdma_destroy_id(cmid);
2523 kiblnd_dev_failover(kib_dev_t *dev)
2525 struct list_head zombie_tpo = LIST_HEAD_INIT(zombie_tpo);
2526 struct list_head zombie_ppo = LIST_HEAD_INIT(zombie_ppo);
2527 struct list_head zombie_fpo = LIST_HEAD_INIT(zombie_fpo);
2528 struct rdma_cm_id *cmid = NULL;
2529 kib_hca_dev_t *hdev = NULL;
2533 struct sockaddr_in addr;
2534 unsigned long flags;
2538 LASSERT (*kiblnd_tunables.kib_dev_failover > 1 ||
2539 dev->ibd_can_failover ||
2540 dev->ibd_hdev == NULL);
2542 rc = kiblnd_dev_need_failover(dev);
2546 if (dev->ibd_hdev != NULL &&
2547 dev->ibd_hdev->ibh_cmid != NULL) {
2548 /* XXX it's not good to close old listener at here,
2549 * because we can fail to create new listener.
2550 * But we have to close it now, otherwise rdma_bind_addr
2551 * will return EADDRINUSE... How crap! */
2552 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2554 cmid = dev->ibd_hdev->ibh_cmid;
2555 /* make next schedule of kiblnd_dev_need_failover()
2556 * return 1 for me */
2557 dev->ibd_hdev->ibh_cmid = NULL;
2558 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2560 rdma_destroy_id(cmid);
2563 cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP,
2567 CERROR("Failed to create cmid for failover: %d\n", rc);
2571 memset(&addr, 0, sizeof(addr));
2572 addr.sin_family = AF_INET;
2573 addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2574 addr.sin_port = htons(*kiblnd_tunables.kib_service);
2576 /* Bind to failover device or port */
2577 rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2578 if (rc != 0 || cmid->device == NULL) {
2579 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2580 dev->ibd_ifname, &dev->ibd_ifip,
2582 rdma_destroy_id(cmid);
2586 LIBCFS_ALLOC(hdev, sizeof(*hdev));
2588 CERROR("Failed to allocate kib_hca_dev\n");
2589 rdma_destroy_id(cmid);
2594 atomic_set(&hdev->ibh_ref, 1);
2595 hdev->ibh_dev = dev;
2596 hdev->ibh_cmid = cmid;
2597 hdev->ibh_ibdev = cmid->device;
2599 pd = ib_alloc_pd(cmid->device);
2602 CERROR("Can't allocate PD: %d\n", rc);
2608 rc = rdma_listen(cmid, 0);
2610 CERROR("Can't start new listener: %d\n", rc);
2614 rc = kiblnd_hdev_setup_mrs(hdev);
2616 CERROR("Can't setup device: %d\n", rc);
2620 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2622 old = dev->ibd_hdev;
2623 dev->ibd_hdev = hdev; /* take over the refcount */
2626 list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2627 cfs_cpt_for_each(i, lnet_cpt_table()) {
2628 kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2631 if (net->ibn_fmr_ps != NULL)
2632 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2637 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2639 if (!list_empty(&zombie_tpo))
2640 kiblnd_destroy_pool_list(&zombie_tpo);
2641 if (!list_empty(&zombie_ppo))
2642 kiblnd_destroy_pool_list(&zombie_ppo);
2643 if (!list_empty(&zombie_fpo))
2644 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2646 kiblnd_hdev_decref(hdev);
2649 dev->ibd_failed_failover++;
2651 dev->ibd_failed_failover = 0;
2657 kiblnd_destroy_dev (kib_dev_t *dev)
2659 LASSERT (dev->ibd_nnets == 0);
2660 LASSERT(list_empty(&dev->ibd_nets));
2662 list_del(&dev->ibd_fail_list);
2663 list_del(&dev->ibd_list);
2665 if (dev->ibd_hdev != NULL)
2666 kiblnd_hdev_decref(dev->ibd_hdev);
2668 LIBCFS_FREE(dev, sizeof(*dev));
2672 kiblnd_create_dev(char *ifname)
2674 struct net_device *netdev;
2681 rc = lnet_ipif_query(ifname, &up, &ip, &netmask);
2683 CERROR("Can't query IPoIB interface %s: %d\n",
2689 CERROR("Can't query IPoIB interface %s: it's down\n", ifname);
2693 LIBCFS_ALLOC(dev, sizeof(*dev));
2697 netdev = dev_get_by_name(&init_net, ifname);
2698 if (netdev == NULL) {
2699 dev->ibd_can_failover = 0;
2701 dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER);
2705 INIT_LIST_HEAD(&dev->ibd_nets);
2706 INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */
2707 INIT_LIST_HEAD(&dev->ibd_fail_list);
2709 strcpy(&dev->ibd_ifname[0], ifname);
2711 /* initialize the device */
2712 rc = kiblnd_dev_failover(dev);
2714 CERROR("Can't initialize device: %d\n", rc);
2715 LIBCFS_FREE(dev, sizeof(*dev));
2719 list_add_tail(&dev->ibd_list,
2720 &kiblnd_data.kib_devs);
2725 kiblnd_base_shutdown(void)
2727 struct kib_sched_info *sched;
2730 LASSERT(list_empty(&kiblnd_data.kib_devs));
2732 CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2733 atomic_read(&libcfs_kmemory));
2735 switch (kiblnd_data.kib_init) {
2739 case IBLND_INIT_ALL:
2740 case IBLND_INIT_DATA:
2741 LASSERT (kiblnd_data.kib_peers != NULL);
2742 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2743 LASSERT(list_empty(&kiblnd_data.kib_peers[i]));
2745 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
2746 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
2747 LASSERT(list_empty(&kiblnd_data.kib_reconn_list));
2748 LASSERT(list_empty(&kiblnd_data.kib_reconn_wait));
2750 /* flag threads to terminate; wake and wait for them to die */
2751 kiblnd_data.kib_shutdown = 1;
2753 /* NB: we really want to stop scheduler threads net by net
2754 * instead of the whole module, this should be improved
2755 * with dynamic configuration LNet */
2756 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2757 wake_up_all(&sched->ibs_waitq);
2759 wake_up_all(&kiblnd_data.kib_connd_waitq);
2760 wake_up_all(&kiblnd_data.kib_failover_waitq);
2763 while (atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2766 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2767 "Waiting for %d threads to terminate\n",
2768 atomic_read(&kiblnd_data.kib_nthreads));
2769 set_current_state(TASK_UNINTERRUPTIBLE);
2770 schedule_timeout(cfs_time_seconds(1));
2775 case IBLND_INIT_NOTHING:
2779 if (kiblnd_data.kib_peers != NULL) {
2780 LIBCFS_FREE(kiblnd_data.kib_peers,
2781 sizeof(struct list_head) *
2782 kiblnd_data.kib_peer_hash_size);
2785 if (kiblnd_data.kib_scheds != NULL)
2786 cfs_percpt_free(kiblnd_data.kib_scheds);
2788 CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2789 atomic_read(&libcfs_kmemory));
2791 kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2792 module_put(THIS_MODULE);
2796 kiblnd_shutdown (lnet_ni_t *ni)
2798 kib_net_t *net = ni->ni_data;
2799 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2801 unsigned long flags;
2803 LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2808 CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
2809 atomic_read(&libcfs_kmemory));
2811 write_lock_irqsave(g_lock, flags);
2812 net->ibn_shutdown = 1;
2813 write_unlock_irqrestore(g_lock, flags);
2815 switch (net->ibn_init) {
2819 case IBLND_INIT_ALL:
2820 /* nuke all existing peers within this net */
2821 kiblnd_del_peer(ni, LNET_NID_ANY);
2823 /* Wait for all peer state to clean up */
2825 while (atomic_read(&net->ibn_npeers) != 0) {
2828 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2829 "%s: waiting for %d peers to disconnect\n",
2830 libcfs_nid2str(ni->ni_nid),
2831 atomic_read(&net->ibn_npeers));
2832 set_current_state(TASK_UNINTERRUPTIBLE);
2833 schedule_timeout(cfs_time_seconds(1));
2836 kiblnd_net_fini_pools(net);
2838 write_lock_irqsave(g_lock, flags);
2839 LASSERT(net->ibn_dev->ibd_nnets > 0);
2840 net->ibn_dev->ibd_nnets--;
2841 list_del(&net->ibn_list);
2842 write_unlock_irqrestore(g_lock, flags);
2846 case IBLND_INIT_NOTHING:
2847 LASSERT (atomic_read(&net->ibn_nconns) == 0);
2849 if (net->ibn_dev != NULL &&
2850 net->ibn_dev->ibd_nnets == 0)
2851 kiblnd_destroy_dev(net->ibn_dev);
2856 CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
2857 atomic_read(&libcfs_kmemory));
2859 net->ibn_init = IBLND_INIT_NOTHING;
2862 LIBCFS_FREE(net, sizeof(*net));
2865 if (list_empty(&kiblnd_data.kib_devs))
2866 kiblnd_base_shutdown();
2871 kiblnd_base_startup(void)
2873 struct kib_sched_info *sched;
2877 LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
2879 try_module_get(THIS_MODULE);
2880 memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
2882 rwlock_init(&kiblnd_data.kib_global_lock);
2884 INIT_LIST_HEAD(&kiblnd_data.kib_devs);
2885 INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
2887 kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
2888 LIBCFS_ALLOC(kiblnd_data.kib_peers,
2889 sizeof(struct list_head) *
2890 kiblnd_data.kib_peer_hash_size);
2891 if (kiblnd_data.kib_peers == NULL)
2894 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2895 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
2897 spin_lock_init(&kiblnd_data.kib_connd_lock);
2898 INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
2899 INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
2900 INIT_LIST_HEAD(&kiblnd_data.kib_reconn_list);
2901 INIT_LIST_HEAD(&kiblnd_data.kib_reconn_wait);
2903 init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
2904 init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
2906 kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
2908 if (kiblnd_data.kib_scheds == NULL)
2911 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
2914 spin_lock_init(&sched->ibs_lock);
2915 INIT_LIST_HEAD(&sched->ibs_conns);
2916 init_waitqueue_head(&sched->ibs_waitq);
2918 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2919 if (*kiblnd_tunables.kib_nscheds > 0) {
2920 nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
2922 /* max to half of CPUs, another half is reserved for
2923 * upper layer modules */
2924 nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
2927 sched->ibs_nthreads_max = nthrs;
2931 kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
2933 /* lists/ptrs/locks initialised */
2934 kiblnd_data.kib_init = IBLND_INIT_DATA;
2935 /*****************************************************/
2937 rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
2939 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
2943 if (*kiblnd_tunables.kib_dev_failover != 0)
2944 rc = kiblnd_thread_start(kiblnd_failover_thread, NULL,
2948 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
2952 /* flag everything initialised */
2953 kiblnd_data.kib_init = IBLND_INIT_ALL;
2954 /*****************************************************/
2959 kiblnd_base_shutdown();
2964 kiblnd_start_schedulers(struct kib_sched_info *sched)
2970 if (sched->ibs_nthreads == 0) {
2971 if (*kiblnd_tunables.kib_nscheds > 0) {
2972 nthrs = sched->ibs_nthreads_max;
2974 nthrs = cfs_cpt_weight(lnet_cpt_table(),
2976 nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
2977 nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
2980 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
2981 /* increase one thread if there is new interface */
2982 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
2985 for (i = 0; i < nthrs; i++) {
2988 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
2989 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
2990 KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
2991 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
2995 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
2996 sched->ibs_cpt, sched->ibs_nthreads + i, rc);
3000 sched->ibs_nthreads += i;
3005 kiblnd_dev_start_threads(kib_dev_t *dev, int newdev, __u32 *cpts, int ncpts)
3011 for (i = 0; i < ncpts; i++) {
3012 struct kib_sched_info *sched;
3014 cpt = (cpts == NULL) ? i : cpts[i];
3015 sched = kiblnd_data.kib_scheds[cpt];
3017 if (!newdev && sched->ibs_nthreads > 0)
3020 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
3022 CERROR("Failed to start scheduler threads for %s\n",
3031 kiblnd_dev_search(char *ifname)
3033 kib_dev_t *alias = NULL;
3038 colon = strchr(ifname, ':');
3039 list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3040 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3046 colon2 = strchr(dev->ibd_ifname, ':');
3052 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3064 kiblnd_startup (lnet_ni_t *ni)
3067 kib_dev_t *ibdev = NULL;
3070 unsigned long flags;
3074 LASSERT (ni->ni_lnd == &the_o2iblnd);
3076 if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
3077 rc = kiblnd_base_startup();
3082 LIBCFS_ALLOC(net, sizeof(*net));
3087 do_gettimeofday(&tv);
3088 net->ibn_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
3090 kiblnd_tunables_setup(ni);
3092 if (ni->ni_interfaces[0] != NULL) {
3093 /* Use the IPoIB interface specified in 'networks=' */
3095 CLASSERT (LNET_MAX_INTERFACES > 1);
3096 if (ni->ni_interfaces[1] != NULL) {
3097 CERROR("Multiple interfaces not supported\n");
3101 ifname = ni->ni_interfaces[0];
3103 ifname = *kiblnd_tunables.kib_default_ipif;
3106 if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3107 CERROR("IPoIB interface name too long: %s\n", ifname);
3111 ibdev = kiblnd_dev_search(ifname);
3113 newdev = ibdev == NULL;
3114 /* hmm...create kib_dev even for alias */
3115 if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0)
3116 ibdev = kiblnd_create_dev(ifname);
3121 net->ibn_dev = ibdev;
3122 ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
3124 rc = kiblnd_dev_start_threads(ibdev, newdev,
3125 ni->ni_cpts, ni->ni_ncpts);
3129 rc = kiblnd_net_init_pools(net, ni, ni->ni_cpts, ni->ni_ncpts);
3131 CERROR("Failed to initialize NI pools: %d\n", rc);
3135 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3137 list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3138 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3140 net->ibn_init = IBLND_INIT_ALL;
3145 if (net != NULL && net->ibn_dev == NULL && ibdev != NULL)
3146 kiblnd_destroy_dev(ibdev);
3148 kiblnd_shutdown(ni);
3150 CDEBUG(D_NET, "kiblnd_startup failed\n");
3154 static lnd_t the_o2iblnd = {
3155 .lnd_type = O2IBLND,
3156 .lnd_startup = kiblnd_startup,
3157 .lnd_shutdown = kiblnd_shutdown,
3158 .lnd_ctl = kiblnd_ctl,
3159 .lnd_query = kiblnd_query,
3160 .lnd_send = kiblnd_send,
3161 .lnd_recv = kiblnd_recv,
3164 static void __exit ko2iblnd_exit(void)
3166 lnet_unregister_lnd(&the_o2iblnd);
3167 kiblnd_tunables_fini();
3170 static int __init ko2iblnd_init(void)
3174 CLASSERT(sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
3175 CLASSERT(offsetof(kib_msg_t,
3176 ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) <=
3178 CLASSERT(offsetof(kib_msg_t,
3179 ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3182 rc = kiblnd_tunables_init();
3186 lnet_register_lnd(&the_o2iblnd);
3191 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3192 MODULE_DESCRIPTION("OpenIB gen2 LNet Network Driver");
3193 MODULE_VERSION("2.8.0");
3194 MODULE_LICENSE("GPL");
3196 module_init(ko2iblnd_init);
3197 module_exit(ko2iblnd_exit);