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.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lnet/klnds/o2iblnd/o2iblnd_cb.c
34 * Author: Eric Barton <eric@bartonsoftware.com>
39 #define MAX_CONN_RACES_BEFORE_ABORT 20
41 static void kiblnd_peer_alive(struct kib_peer_ni *peer_ni);
42 static void kiblnd_peer_connect_failed(struct kib_peer_ni *peer_ni, int active,
44 static void kiblnd_init_tx_msg(struct lnet_ni *ni, struct kib_tx *tx,
45 int type, int body_nob);
46 static int kiblnd_init_rdma(struct kib_conn *conn, struct kib_tx *tx, int type,
47 int resid, struct kib_rdma_desc *dstrd, u64 dstcookie);
48 static void kiblnd_queue_tx_locked(struct kib_tx *tx, struct kib_conn *conn);
49 static void kiblnd_queue_tx(struct kib_tx *tx, struct kib_conn *conn);
51 static void kiblnd_unmap_tx(struct kib_tx *tx);
52 static void kiblnd_check_sends_locked(struct kib_conn *conn);
55 kiblnd_tx_done(struct kib_tx *tx)
57 struct lnet_msg *lntmsg[2];
61 LASSERT (!in_interrupt());
62 LASSERT (!tx->tx_queued); /* mustn't be queued for sending */
63 LASSERT (tx->tx_sending == 0); /* mustn't be awaiting sent callback */
64 LASSERT (!tx->tx_waiting); /* mustn't be awaiting peer_ni response */
65 LASSERT (tx->tx_pool != NULL);
69 /* tx may have up to 2 lnet msgs to finalise */
70 lntmsg[0] = tx->tx_lntmsg[0]; tx->tx_lntmsg[0] = NULL;
71 lntmsg[1] = tx->tx_lntmsg[1]; tx->tx_lntmsg[1] = NULL;
74 if (tx->tx_conn != NULL) {
75 kiblnd_conn_decref(tx->tx_conn);
79 tx->tx_nwrq = tx->tx_nsge = 0;
82 kiblnd_pool_free_node(&tx->tx_pool->tpo_pool, &tx->tx_list);
84 /* delay finalize until my descs have been freed */
85 for (i = 0; i < 2; i++) {
86 if (lntmsg[i] == NULL)
89 /* propagate health status to LNet for requests */
90 if (i == 0 && lntmsg[i])
91 lntmsg[i]->msg_health_status = tx->tx_hstatus;
93 lnet_finalize(lntmsg[i], rc);
98 kiblnd_txlist_done(struct list_head *txlist, int status,
99 enum lnet_msg_hstatus hstatus)
103 while (!list_empty(txlist)) {
104 tx = list_entry(txlist->next, struct kib_tx, tx_list);
106 list_del(&tx->tx_list);
109 tx->tx_status = status;
110 if (hstatus != LNET_MSG_STATUS_OK)
111 tx->tx_hstatus = hstatus;
116 static struct kib_tx *
117 kiblnd_get_idle_tx(struct lnet_ni *ni, lnet_nid_t target)
119 struct kib_net *net = ni->ni_data;
120 struct list_head *node;
122 struct kib_tx_poolset *tps;
124 tps = net->ibn_tx_ps[lnet_cpt_of_nid(target, ni)];
125 node = kiblnd_pool_alloc_node(&tps->tps_poolset);
128 tx = container_of(node, struct kib_tx, tx_list);
130 LASSERT (tx->tx_nwrq == 0);
131 LASSERT (!tx->tx_queued);
132 LASSERT (tx->tx_sending == 0);
133 LASSERT (!tx->tx_waiting);
134 LASSERT (tx->tx_status == 0);
135 LASSERT (tx->tx_conn == NULL);
136 LASSERT (tx->tx_lntmsg[0] == NULL);
137 LASSERT (tx->tx_lntmsg[1] == NULL);
138 LASSERT (tx->tx_nfrags == 0);
141 tx->tx_hstatus = LNET_MSG_STATUS_OK;
147 kiblnd_drop_rx(struct kib_rx *rx)
149 struct kib_conn *conn = rx->rx_conn;
150 struct kib_sched_info *sched = conn->ibc_sched;
153 spin_lock_irqsave(&sched->ibs_lock, flags);
154 LASSERT(conn->ibc_nrx > 0);
156 spin_unlock_irqrestore(&sched->ibs_lock, flags);
158 kiblnd_conn_decref(conn);
162 kiblnd_post_rx(struct kib_rx *rx, int credit)
164 struct kib_conn *conn = rx->rx_conn;
165 struct kib_net *net = conn->ibc_peer->ibp_ni->ni_data;
166 struct ib_recv_wr *bad_wrq = NULL;
167 #ifdef HAVE_IB_GET_DMA_MR
168 struct ib_mr *mr = conn->ibc_hdev->ibh_mrs;
172 LASSERT (net != NULL);
173 LASSERT (!in_interrupt());
174 LASSERT (credit == IBLND_POSTRX_NO_CREDIT ||
175 credit == IBLND_POSTRX_PEER_CREDIT ||
176 credit == IBLND_POSTRX_RSRVD_CREDIT);
177 #ifdef HAVE_IB_GET_DMA_MR
180 rx->rx_sge.lkey = mr->lkey;
182 rx->rx_sge.lkey = conn->ibc_hdev->ibh_pd->local_dma_lkey;
184 rx->rx_sge.addr = rx->rx_msgaddr;
185 rx->rx_sge.length = IBLND_MSG_SIZE;
187 rx->rx_wrq.next = NULL;
188 rx->rx_wrq.sg_list = &rx->rx_sge;
189 rx->rx_wrq.num_sge = 1;
190 rx->rx_wrq.wr_id = kiblnd_ptr2wreqid(rx, IBLND_WID_RX);
192 LASSERT (conn->ibc_state >= IBLND_CONN_INIT);
193 LASSERT (rx->rx_nob >= 0); /* not posted */
195 if (conn->ibc_state > IBLND_CONN_ESTABLISHED) {
196 kiblnd_drop_rx(rx); /* No more posts for this rx */
200 rx->rx_nob = -1; /* flag posted */
202 /* NB: need an extra reference after ib_post_recv because we don't
203 * own this rx (and rx::rx_conn) anymore, LU-5678.
205 kiblnd_conn_addref(conn);
206 #ifdef HAVE_IB_POST_SEND_RECV_CONST
207 rc = ib_post_recv(conn->ibc_cmid->qp, &rx->rx_wrq,
208 (const struct ib_recv_wr **)&bad_wrq);
210 rc = ib_post_recv(conn->ibc_cmid->qp, &rx->rx_wrq, &bad_wrq);
212 if (unlikely(rc != 0)) {
213 CERROR("Can't post rx for %s: %d, bad_wrq: %p\n",
214 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc, bad_wrq);
218 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) /* Initial post */
221 if (unlikely(rc != 0)) {
222 kiblnd_close_conn(conn, rc);
223 kiblnd_drop_rx(rx); /* No more posts for this rx */
227 if (credit == IBLND_POSTRX_NO_CREDIT)
230 spin_lock(&conn->ibc_lock);
231 if (credit == IBLND_POSTRX_PEER_CREDIT)
232 conn->ibc_outstanding_credits++;
234 conn->ibc_reserved_credits++;
235 kiblnd_check_sends_locked(conn);
236 spin_unlock(&conn->ibc_lock);
239 kiblnd_conn_decref(conn);
243 static struct kib_tx *
244 kiblnd_find_waiting_tx_locked(struct kib_conn *conn, int txtype, u64 cookie)
246 struct list_head *tmp;
248 list_for_each(tmp, &conn->ibc_active_txs) {
249 struct kib_tx *tx = list_entry(tmp, struct kib_tx, tx_list);
251 LASSERT(!tx->tx_queued);
252 LASSERT(tx->tx_sending != 0 || tx->tx_waiting);
254 if (tx->tx_cookie != cookie)
257 if (tx->tx_waiting &&
258 tx->tx_msg->ibm_type == txtype)
261 CWARN("Bad completion: %swaiting, type %x (wanted %x)\n",
262 tx->tx_waiting ? "" : "NOT ",
263 tx->tx_msg->ibm_type, txtype);
269 kiblnd_handle_completion(struct kib_conn *conn, int txtype, int status, u64 cookie)
272 struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
275 spin_lock(&conn->ibc_lock);
277 tx = kiblnd_find_waiting_tx_locked(conn, txtype, cookie);
279 spin_unlock(&conn->ibc_lock);
281 CWARN("Unmatched completion type %x cookie %#llx from %s\n",
282 txtype, cookie, libcfs_nid2str(conn->ibc_peer->ibp_nid));
283 kiblnd_close_conn(conn, -EPROTO);
287 if (tx->tx_status == 0) { /* success so far */
288 if (status < 0) { /* failed? */
289 tx->tx_status = status;
290 tx->tx_hstatus = LNET_MSG_STATUS_REMOTE_ERROR;
291 } else if (txtype == IBLND_MSG_GET_REQ) {
292 lnet_set_reply_msg_len(ni, tx->tx_lntmsg[1], status);
298 idle = !tx->tx_queued && (tx->tx_sending == 0);
300 list_del(&tx->tx_list);
302 spin_unlock(&conn->ibc_lock);
309 kiblnd_send_completion(struct kib_conn *conn, int type, int status, u64 cookie)
311 struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
312 struct kib_tx *tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
315 CERROR("Can't get tx for completion %x for %s\n",
316 type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
320 tx->tx_msg->ibm_u.completion.ibcm_status = status;
321 tx->tx_msg->ibm_u.completion.ibcm_cookie = cookie;
322 kiblnd_init_tx_msg(ni, tx, type, sizeof(struct kib_completion_msg));
324 kiblnd_queue_tx(tx, conn);
328 kiblnd_handle_rx(struct kib_rx *rx)
330 struct kib_msg *msg = rx->rx_msg;
331 struct kib_conn *conn = rx->rx_conn;
332 struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
333 int credits = msg->ibm_credits;
339 LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
341 CDEBUG (D_NET, "Received %x[%d] from %s\n",
342 msg->ibm_type, credits,
343 libcfs_nid2str(conn->ibc_peer->ibp_nid));
346 /* Have I received credits that will let me send? */
347 spin_lock(&conn->ibc_lock);
349 if (conn->ibc_credits + credits >
350 conn->ibc_queue_depth) {
351 rc2 = conn->ibc_credits;
352 spin_unlock(&conn->ibc_lock);
354 CERROR("Bad credits from %s: %d + %d > %d\n",
355 libcfs_nid2str(conn->ibc_peer->ibp_nid),
357 conn->ibc_queue_depth);
359 kiblnd_close_conn(conn, -EPROTO);
360 kiblnd_post_rx(rx, IBLND_POSTRX_NO_CREDIT);
364 conn->ibc_credits += credits;
366 /* This ensures the credit taken by NOOP can be returned */
367 if (msg->ibm_type == IBLND_MSG_NOOP &&
368 !IBLND_OOB_CAPABLE(conn->ibc_version)) /* v1 only */
369 conn->ibc_outstanding_credits++;
371 kiblnd_check_sends_locked(conn);
372 spin_unlock(&conn->ibc_lock);
375 switch (msg->ibm_type) {
377 CERROR("Bad IBLND message type %x from %s\n",
378 msg->ibm_type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
379 post_credit = IBLND_POSTRX_NO_CREDIT;
384 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
385 post_credit = IBLND_POSTRX_NO_CREDIT;
389 if (credits != 0) /* credit already posted */
390 post_credit = IBLND_POSTRX_NO_CREDIT;
391 else /* a keepalive NOOP */
392 post_credit = IBLND_POSTRX_PEER_CREDIT;
395 case IBLND_MSG_IMMEDIATE:
396 post_credit = IBLND_POSTRX_DONT_POST;
397 rc = lnet_parse(ni, &msg->ibm_u.immediate.ibim_hdr,
398 msg->ibm_srcnid, rx, 0);
399 if (rc < 0) /* repost on error */
400 post_credit = IBLND_POSTRX_PEER_CREDIT;
403 case IBLND_MSG_PUT_REQ:
404 post_credit = IBLND_POSTRX_DONT_POST;
405 rc = lnet_parse(ni, &msg->ibm_u.putreq.ibprm_hdr,
406 msg->ibm_srcnid, rx, 1);
407 if (rc < 0) /* repost on error */
408 post_credit = IBLND_POSTRX_PEER_CREDIT;
411 case IBLND_MSG_PUT_NAK:
412 CWARN ("PUT_NACK from %s\n",
413 libcfs_nid2str(conn->ibc_peer->ibp_nid));
414 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
415 kiblnd_handle_completion(conn, IBLND_MSG_PUT_REQ,
416 msg->ibm_u.completion.ibcm_status,
417 msg->ibm_u.completion.ibcm_cookie);
420 case IBLND_MSG_PUT_ACK:
421 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
423 spin_lock(&conn->ibc_lock);
424 tx = kiblnd_find_waiting_tx_locked(conn, IBLND_MSG_PUT_REQ,
425 msg->ibm_u.putack.ibpam_src_cookie);
427 list_del(&tx->tx_list);
428 spin_unlock(&conn->ibc_lock);
431 CERROR("Unmatched PUT_ACK from %s\n",
432 libcfs_nid2str(conn->ibc_peer->ibp_nid));
437 LASSERT (tx->tx_waiting);
438 /* CAVEAT EMPTOR: I could be racing with tx_complete, but...
439 * (a) I can overwrite tx_msg since my peer_ni has received it!
440 * (b) tx_waiting set tells tx_complete() it's not done. */
442 tx->tx_nwrq = tx->tx_nsge = 0; /* overwrite PUT_REQ */
444 rc2 = kiblnd_init_rdma(conn, tx, IBLND_MSG_PUT_DONE,
445 kiblnd_rd_size(&msg->ibm_u.putack.ibpam_rd),
446 &msg->ibm_u.putack.ibpam_rd,
447 msg->ibm_u.putack.ibpam_dst_cookie);
449 CERROR("Can't setup rdma for PUT to %s: %d\n",
450 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc2);
452 spin_lock(&conn->ibc_lock);
453 tx->tx_waiting = 0; /* clear waiting and queue atomically */
454 kiblnd_queue_tx_locked(tx, conn);
455 spin_unlock(&conn->ibc_lock);
458 case IBLND_MSG_PUT_DONE:
459 post_credit = IBLND_POSTRX_PEER_CREDIT;
460 kiblnd_handle_completion(conn, IBLND_MSG_PUT_ACK,
461 msg->ibm_u.completion.ibcm_status,
462 msg->ibm_u.completion.ibcm_cookie);
465 case IBLND_MSG_GET_REQ:
466 post_credit = IBLND_POSTRX_DONT_POST;
467 rc = lnet_parse(ni, &msg->ibm_u.get.ibgm_hdr,
468 msg->ibm_srcnid, rx, 1);
469 if (rc < 0) /* repost on error */
470 post_credit = IBLND_POSTRX_PEER_CREDIT;
473 case IBLND_MSG_GET_DONE:
474 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
475 kiblnd_handle_completion(conn, IBLND_MSG_GET_REQ,
476 msg->ibm_u.completion.ibcm_status,
477 msg->ibm_u.completion.ibcm_cookie);
481 if (rc < 0) /* protocol error */
482 kiblnd_close_conn(conn, rc);
484 if (post_credit != IBLND_POSTRX_DONT_POST)
485 kiblnd_post_rx(rx, post_credit);
489 kiblnd_rx_complete(struct kib_rx *rx, int status, int nob)
491 struct kib_msg *msg = rx->rx_msg;
492 struct kib_conn *conn = rx->rx_conn;
493 struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
494 struct kib_net *net = ni->ni_data;
498 LASSERT (net != NULL);
499 LASSERT (rx->rx_nob < 0); /* was posted */
500 rx->rx_nob = 0; /* isn't now */
502 if (conn->ibc_state > IBLND_CONN_ESTABLISHED)
505 if (status != IB_WC_SUCCESS) {
506 CNETERR("Rx from %s failed: %d\n",
507 libcfs_nid2str(conn->ibc_peer->ibp_nid), status);
514 rc = kiblnd_unpack_msg(msg, rx->rx_nob);
516 CERROR ("Error %d unpacking rx from %s\n",
517 rc, libcfs_nid2str(conn->ibc_peer->ibp_nid));
521 if (msg->ibm_srcnid != conn->ibc_peer->ibp_nid ||
522 msg->ibm_dstnid != ni->ni_nid ||
523 msg->ibm_srcstamp != conn->ibc_incarnation ||
524 msg->ibm_dststamp != net->ibn_incarnation) {
525 CERROR ("Stale rx from %s\n",
526 libcfs_nid2str(conn->ibc_peer->ibp_nid));
531 /* set time last known alive */
532 kiblnd_peer_alive(conn->ibc_peer);
534 /* racing with connection establishment/teardown! */
536 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
537 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
540 write_lock_irqsave(g_lock, flags);
541 /* must check holding global lock to eliminate race */
542 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
543 list_add_tail(&rx->rx_list, &conn->ibc_early_rxs);
544 write_unlock_irqrestore(g_lock, flags);
547 write_unlock_irqrestore(g_lock, flags);
549 kiblnd_handle_rx(rx);
553 CDEBUG(D_NET, "rx %p conn %p\n", rx, conn);
554 kiblnd_close_conn(conn, err);
556 kiblnd_drop_rx(rx); /* Don't re-post rx. */
560 kiblnd_fmr_map_tx(struct kib_net *net, struct kib_tx *tx,
561 struct kib_rdma_desc *rd, u32 nob)
563 struct kib_hca_dev *hdev;
565 struct kib_fmr_poolset *fps;
570 LASSERT(tx->tx_pool != NULL);
571 LASSERT(tx->tx_pool->tpo_pool.po_owner != NULL);
574 hdev = tx->tx_pool->tpo_hdev;
575 cpt = tx->tx_pool->tpo_pool.po_owner->ps_cpt;
578 * If we're dealing with FastReg, but the device doesn't
579 * support GAPS and the tx has GAPS, then there is no real point
580 * in trying to map the memory, because it'll just fail. So
581 * preemptively fail with an appropriate message
583 if ((dev->ibd_dev_caps & IBLND_DEV_CAPS_FASTREG_ENABLED) &&
584 !(dev->ibd_dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT) &&
586 CERROR("Using FastReg with no GAPS support, but tx has gaps. "
587 "Try setting use_fastreg_gaps to 1\n");
588 return -EPROTONOSUPPORT;
592 * FMR does not support gaps but the tx has gaps then
593 * we should make sure that the number of fragments we'll be sending
594 * over fits within the number of fragments negotiated on the
595 * connection, otherwise, we won't be able to RDMA the data.
596 * We need to maintain the number of fragments negotiation on the
597 * connection for backwards compatibility.
599 if (tx->tx_gaps && (dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED)) {
601 tx->tx_conn->ibc_max_frags <= rd->rd_nfrags) {
602 CERROR("TX number of frags (%d) is <= than connection"
603 " number of frags (%d). Consider setting peer's"
604 " map_on_demand to 256\n", tx->tx_nfrags,
605 tx->tx_conn->ibc_max_frags);
610 fps = net->ibn_fmr_ps[cpt];
611 rc = kiblnd_fmr_pool_map(fps, tx, rd, nob, 0, &tx->tx_fmr);
613 CERROR("Can't map %u bytes (%u/%u)s: %d\n", nob,
614 tx->tx_nfrags, rd->rd_nfrags, rc);
619 * If rd is not tx_rd, it's going to get sent to a peer_ni, who will
622 rd->rd_key = tx->tx_fmr.fmr_key;
624 * for FastReg or FMR with no gaps we can accumulate all
625 * the fragments in one FastReg or FMR fragment.
627 if (((dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED) && !tx->tx_gaps) ||
628 (dev->ibd_dev_caps & IBLND_DEV_CAPS_FASTREG_ENABLED)) {
629 /* FMR requires zero based address */
630 if (dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED)
631 rd->rd_frags[0].rf_addr &= ~hdev->ibh_page_mask;
632 rd->rd_frags[0].rf_nob = nob;
636 * We're transmitting with gaps using FMR.
637 * We'll need to use multiple fragments and identify the
638 * zero based address of each fragment.
640 for (i = 0; i < rd->rd_nfrags; i++) {
641 rd->rd_frags[i].rf_addr &= ~hdev->ibh_page_mask;
642 rd->rd_frags[i].rf_addr += i << hdev->ibh_page_shift;
650 kiblnd_unmap_tx(struct kib_tx *tx)
652 if (tx->tx_fmr.fmr_pfmr || tx->tx_fmr.fmr_frd)
653 kiblnd_fmr_pool_unmap(&tx->tx_fmr, tx->tx_status);
655 if (tx->tx_nfrags != 0) {
656 kiblnd_dma_unmap_sg(tx->tx_pool->tpo_hdev->ibh_ibdev,
657 tx->tx_frags, tx->tx_nfrags, tx->tx_dmadir);
662 #ifdef HAVE_IB_GET_DMA_MR
663 static struct ib_mr *
664 kiblnd_find_rd_dma_mr(struct lnet_ni *ni, struct kib_rdma_desc *rd)
666 struct kib_net *net = ni->ni_data;
667 struct kib_hca_dev *hdev = net->ibn_dev->ibd_hdev;
668 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
670 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
673 * if map-on-demand is turned on and the device supports
674 * either FMR or FastReg then use that. Otherwise use global
675 * memory regions. If that's not available either, then you're
676 * dead in the water and fail the operation.
678 if (tunables->lnd_map_on_demand &&
679 (net->ibn_dev->ibd_dev_caps & IBLND_DEV_CAPS_FASTREG_ENABLED ||
680 net->ibn_dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED))
684 * hdev->ibh_mrs can be NULL. This case is dealt with gracefully
685 * in the call chain. The mapping will fail with appropriate error
688 return hdev->ibh_mrs;
692 static int kiblnd_map_tx(struct lnet_ni *ni, struct kib_tx *tx,
693 struct kib_rdma_desc *rd, int nfrags)
695 struct kib_net *net = ni->ni_data;
696 struct kib_hca_dev *hdev = net->ibn_dev->ibd_hdev;
697 #ifdef HAVE_IB_GET_DMA_MR
698 struct ib_mr *mr = NULL;
703 /* If rd is not tx_rd, it's going to get sent to a peer_ni and I'm the
705 tx->tx_dmadir = (rd != tx->tx_rd) ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
706 tx->tx_nfrags = nfrags;
708 rd->rd_nfrags = kiblnd_dma_map_sg(hdev->ibh_ibdev, tx->tx_frags,
709 tx->tx_nfrags, tx->tx_dmadir);
711 for (i = 0, nob = 0; i < rd->rd_nfrags; i++) {
712 rd->rd_frags[i].rf_nob = kiblnd_sg_dma_len(
713 hdev->ibh_ibdev, &tx->tx_frags[i]);
714 rd->rd_frags[i].rf_addr = kiblnd_sg_dma_address(
715 hdev->ibh_ibdev, &tx->tx_frags[i]);
716 nob += rd->rd_frags[i].rf_nob;
719 #ifdef HAVE_IB_GET_DMA_MR
720 mr = kiblnd_find_rd_dma_mr(ni, rd);
722 /* found pre-mapping MR */
723 rd->rd_key = (rd != tx->tx_rd) ? mr->rkey : mr->lkey;
728 if (net->ibn_fmr_ps != NULL)
729 return kiblnd_fmr_map_tx(net, tx, rd, nob);
734 static int kiblnd_setup_rd_kiov(struct lnet_ni *ni, struct kib_tx *tx,
735 struct kib_rdma_desc *rd, int nkiov,
736 struct bio_vec *kiov, int offset, int nob)
738 struct kib_net *net = ni->ni_data;
739 struct scatterlist *sg;
743 CDEBUG(D_NET, "niov %d offset %d nob %d\n", nkiov, offset, nob);
747 LASSERT(net != NULL);
749 while (offset >= kiov->bv_len) {
750 offset -= kiov->bv_len;
762 fragnob = min((int)(kiov->bv_len - offset), nob);
765 * We're allowed to start at a non-aligned page offset in
766 * the first fragment and end at a non-aligned page offset
767 * in the last fragment.
769 if ((fragnob < (int)(kiov->bv_len - offset)) &&
770 nkiov < max_nkiov && nob > fragnob) {
771 CDEBUG(D_NET, "fragnob %d < available page %d: with"
772 " remaining %d kiovs with %d nob left\n",
773 fragnob, (int)(kiov->bv_len - offset),
778 sg_set_page(sg, kiov->bv_page, fragnob,
779 kiov->bv_offset + offset);
782 CERROR("lacking enough sg entries to map tx\n");
792 return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags);
796 kiblnd_post_tx_locked(struct kib_conn *conn, struct kib_tx *tx, int credit)
797 __must_hold(&conn->ibc_lock)
799 struct kib_msg *msg = tx->tx_msg;
800 struct kib_peer_ni *peer_ni = conn->ibc_peer;
801 struct lnet_ni *ni = peer_ni->ibp_ni;
802 int ver = conn->ibc_version;
806 LASSERT(tx->tx_queued);
807 /* We rely on this for QP sizing */
808 LASSERT(tx->tx_nwrq > 0 && tx->tx_nsge >= 0);
809 LASSERT(tx->tx_nwrq <= 1 + conn->ibc_max_frags);
811 LASSERT(credit == 0 || credit == 1);
812 LASSERT(conn->ibc_outstanding_credits >= 0);
813 LASSERT(conn->ibc_outstanding_credits <= conn->ibc_queue_depth);
814 LASSERT(conn->ibc_credits >= 0);
815 LASSERT(conn->ibc_credits <= conn->ibc_queue_depth);
817 if (conn->ibc_nsends_posted ==
818 kiblnd_concurrent_sends(ver, ni)) {
819 /* tx completions outstanding... */
820 CDEBUG(D_NET, "%s: posted enough\n",
821 libcfs_nid2str(peer_ni->ibp_nid));
825 if (credit != 0 && conn->ibc_credits == 0) { /* no credits */
826 CDEBUG(D_NET, "%s: no credits\n",
827 libcfs_nid2str(peer_ni->ibp_nid));
831 if (credit != 0 && !IBLND_OOB_CAPABLE(ver) &&
832 conn->ibc_credits == 1 && /* last credit reserved */
833 msg->ibm_type != IBLND_MSG_NOOP) { /* for NOOP */
834 CDEBUG(D_NET, "%s: not using last credit\n",
835 libcfs_nid2str(peer_ni->ibp_nid));
839 /* NB don't drop ibc_lock before bumping tx_sending */
840 list_del(&tx->tx_list);
843 if (msg->ibm_type == IBLND_MSG_NOOP &&
844 (!kiblnd_need_noop(conn) || /* redundant NOOP */
845 (IBLND_OOB_CAPABLE(ver) && /* posted enough NOOP */
846 conn->ibc_noops_posted == IBLND_OOB_MSGS(ver)))) {
847 /* OK to drop when posted enough NOOPs, since
848 * kiblnd_check_sends_locked will queue NOOP again when
849 * posted NOOPs complete */
850 spin_unlock(&conn->ibc_lock);
851 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
853 spin_lock(&conn->ibc_lock);
854 CDEBUG(D_NET, "%s(%d): redundant or enough NOOP\n",
855 libcfs_nid2str(peer_ni->ibp_nid),
856 conn->ibc_noops_posted);
860 kiblnd_pack_msg(peer_ni->ibp_ni, msg, ver, conn->ibc_outstanding_credits,
861 peer_ni->ibp_nid, conn->ibc_incarnation);
863 conn->ibc_credits -= credit;
864 conn->ibc_outstanding_credits = 0;
865 conn->ibc_nsends_posted++;
866 if (msg->ibm_type == IBLND_MSG_NOOP)
867 conn->ibc_noops_posted++;
869 /* CAVEAT EMPTOR! This tx could be the PUT_DONE of an RDMA
870 * PUT. If so, it was first queued here as a PUT_REQ, sent and
871 * stashed on ibc_active_txs, matched by an incoming PUT_ACK,
872 * and then re-queued here. It's (just) possible that
873 * tx_sending is non-zero if we've not done the tx_complete()
874 * from the first send; hence the ++ rather than = below. */
876 list_add(&tx->tx_list, &conn->ibc_active_txs);
878 /* I'm still holding ibc_lock! */
879 if (conn->ibc_state != IBLND_CONN_ESTABLISHED) {
881 } else if (tx->tx_pool->tpo_pool.po_failed ||
882 conn->ibc_hdev != tx->tx_pool->tpo_hdev) {
883 /* close_conn will launch failover */
886 struct kib_fast_reg_descriptor *frd = tx->tx_fmr.fmr_frd;
887 struct ib_send_wr *bad = &tx->tx_wrq[tx->tx_nwrq - 1].wr;
888 struct ib_send_wr *wr = &tx->tx_wrq[0].wr;
891 if (!frd->frd_valid) {
892 wr = &frd->frd_inv_wr.wr;
893 wr->next = &frd->frd_fastreg_wr.wr;
895 wr = &frd->frd_fastreg_wr.wr;
897 frd->frd_fastreg_wr.wr.next = &tx->tx_wrq[0].wr;
900 LASSERTF(bad->wr_id == kiblnd_ptr2wreqid(tx, IBLND_WID_TX),
901 "bad wr_id %#llx, opc %d, flags %d, peer_ni: %s\n",
902 bad->wr_id, bad->opcode, bad->send_flags,
903 libcfs_nid2str(conn->ibc_peer->ibp_nid));
906 if (lnet_send_error_simulation(tx->tx_lntmsg[0], &tx->tx_hstatus))
909 #ifdef HAVE_IB_POST_SEND_RECV_CONST
910 rc = ib_post_send(conn->ibc_cmid->qp, wr,
911 (const struct ib_send_wr **)&bad);
913 rc = ib_post_send(conn->ibc_cmid->qp, wr, &bad);
917 conn->ibc_last_send = ktime_get();
922 /* NB credits are transferred in the actual
923 * message, which can only be the last work item */
924 conn->ibc_credits += credit;
925 conn->ibc_outstanding_credits += msg->ibm_credits;
926 conn->ibc_nsends_posted--;
927 if (msg->ibm_type == IBLND_MSG_NOOP)
928 conn->ibc_noops_posted--;
934 done = (tx->tx_sending == 0);
936 list_del(&tx->tx_list);
938 spin_unlock(&conn->ibc_lock);
940 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
941 CERROR("Error %d posting transmit to %s\n",
942 rc, libcfs_nid2str(peer_ni->ibp_nid));
944 CDEBUG(D_NET, "Error %d posting transmit to %s\n",
945 rc, libcfs_nid2str(peer_ni->ibp_nid));
947 kiblnd_close_conn(conn, rc);
952 spin_lock(&conn->ibc_lock);
958 kiblnd_check_sends_locked(struct kib_conn *conn)
960 int ver = conn->ibc_version;
961 struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
964 /* Don't send anything until after the connection is established */
965 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
966 CDEBUG(D_NET, "%s too soon\n",
967 libcfs_nid2str(conn->ibc_peer->ibp_nid));
971 LASSERT(conn->ibc_nsends_posted <=
972 kiblnd_concurrent_sends(ver, ni));
973 LASSERT (!IBLND_OOB_CAPABLE(ver) ||
974 conn->ibc_noops_posted <= IBLND_OOB_MSGS(ver));
975 LASSERT (conn->ibc_reserved_credits >= 0);
977 while (conn->ibc_reserved_credits > 0 &&
978 !list_empty(&conn->ibc_tx_queue_rsrvd)) {
979 tx = list_entry(conn->ibc_tx_queue_rsrvd.next,
980 struct kib_tx, tx_list);
981 list_move_tail(&tx->tx_list, &conn->ibc_tx_queue);
982 conn->ibc_reserved_credits--;
985 if (kiblnd_need_noop(conn)) {
986 spin_unlock(&conn->ibc_lock);
988 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
990 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_NOOP, 0);
992 spin_lock(&conn->ibc_lock);
994 kiblnd_queue_tx_locked(tx, conn);
1000 if (!list_empty(&conn->ibc_tx_queue_nocred)) {
1002 tx = list_entry(conn->ibc_tx_queue_nocred.next,
1003 struct kib_tx, tx_list);
1004 } else if (!list_empty(&conn->ibc_tx_noops)) {
1005 LASSERT (!IBLND_OOB_CAPABLE(ver));
1007 tx = list_entry(conn->ibc_tx_noops.next,
1008 struct kib_tx, tx_list);
1009 } else if (!list_empty(&conn->ibc_tx_queue)) {
1011 tx = list_entry(conn->ibc_tx_queue.next,
1012 struct kib_tx, tx_list);
1016 if (kiblnd_post_tx_locked(conn, tx, credit) != 0)
1022 kiblnd_tx_complete(struct kib_tx *tx, int status)
1024 int failed = (status != IB_WC_SUCCESS);
1025 struct kib_conn *conn = tx->tx_conn;
1028 if (tx->tx_sending <= 0) {
1029 CERROR("Received an event on a freed tx: %p status %d\n",
1035 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
1036 CNETERR("Tx -> %s cookie %#llx"
1037 " sending %d waiting %d: failed %d\n",
1038 libcfs_nid2str(conn->ibc_peer->ibp_nid),
1039 tx->tx_cookie, tx->tx_sending, tx->tx_waiting,
1042 kiblnd_close_conn(conn, -EIO);
1044 kiblnd_peer_alive(conn->ibc_peer);
1047 spin_lock(&conn->ibc_lock);
1049 /* I could be racing with rdma completion. Whoever makes 'tx' idle
1050 * gets to free it, which also drops its ref on 'conn'. */
1053 conn->ibc_nsends_posted--;
1054 if (tx->tx_msg->ibm_type == IBLND_MSG_NOOP)
1055 conn->ibc_noops_posted--;
1058 tx->tx_hstatus = LNET_MSG_STATUS_REMOTE_DROPPED;
1059 tx->tx_waiting = 0; /* don't wait for peer_ni */
1060 tx->tx_status = -EIO;
1063 idle = (tx->tx_sending == 0) && /* This is the final callback */
1064 !tx->tx_waiting && /* Not waiting for peer_ni */
1065 !tx->tx_queued; /* Not re-queued (PUT_DONE) */
1067 list_del(&tx->tx_list);
1069 kiblnd_check_sends_locked(conn);
1070 spin_unlock(&conn->ibc_lock);
1077 kiblnd_init_tx_msg(struct lnet_ni *ni, struct kib_tx *tx, int type,
1080 struct kib_hca_dev *hdev = tx->tx_pool->tpo_hdev;
1081 struct ib_sge *sge = &tx->tx_msgsge;
1082 struct ib_rdma_wr *wrq;
1083 int nob = offsetof(struct kib_msg, ibm_u) + body_nob;
1084 #ifdef HAVE_IB_GET_DMA_MR
1085 struct ib_mr *mr = hdev->ibh_mrs;
1088 LASSERT(tx->tx_nwrq >= 0);
1089 LASSERT(tx->tx_nwrq < IBLND_MAX_RDMA_FRAGS + 1);
1090 LASSERT(nob <= IBLND_MSG_SIZE);
1091 #ifdef HAVE_IB_GET_DMA_MR
1092 LASSERT(mr != NULL);
1095 kiblnd_init_msg(tx->tx_msg, type, body_nob);
1097 #ifdef HAVE_IB_GET_DMA_MR
1098 sge->lkey = mr->lkey;
1100 sge->lkey = hdev->ibh_pd->local_dma_lkey;
1102 sge->addr = tx->tx_msgaddr;
1105 wrq = &tx->tx_wrq[tx->tx_nwrq];
1106 memset(wrq, 0, sizeof(*wrq));
1108 wrq->wr.next = NULL;
1109 wrq->wr.wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_TX);
1110 wrq->wr.sg_list = sge;
1111 wrq->wr.num_sge = 1;
1112 wrq->wr.opcode = IB_WR_SEND;
1113 wrq->wr.send_flags = IB_SEND_SIGNALED;
1119 kiblnd_init_rdma(struct kib_conn *conn, struct kib_tx *tx, int type,
1120 int resid, struct kib_rdma_desc *dstrd, u64 dstcookie)
1122 struct kib_msg *ibmsg = tx->tx_msg;
1123 struct kib_rdma_desc *srcrd = tx->tx_rd;
1124 struct ib_rdma_wr *wrq = NULL;
1132 LASSERT(!in_interrupt());
1133 LASSERT(tx->tx_nwrq == 0 && tx->tx_nsge == 0);
1134 LASSERT(type == IBLND_MSG_GET_DONE || type == IBLND_MSG_PUT_DONE);
1136 for (srcidx = dstidx = wrq_sge = sge_nob = 0;
1137 resid > 0; resid -= sge_nob) {
1140 if (srcidx >= srcrd->rd_nfrags) {
1141 CERROR("Src buffer exhausted: %d frags\n", srcidx);
1146 if (dstidx >= dstrd->rd_nfrags) {
1147 CERROR("Dst buffer exhausted: %d frags\n", dstidx);
1152 if (tx->tx_nwrq >= conn->ibc_max_frags) {
1153 CERROR("RDMA has too many fragments for peer_ni %s (%d), "
1154 "src idx/frags: %d/%d dst idx/frags: %d/%d\n",
1155 libcfs_nid2str(conn->ibc_peer->ibp_nid),
1156 conn->ibc_max_frags,
1157 srcidx, srcrd->rd_nfrags,
1158 dstidx, dstrd->rd_nfrags);
1163 sge_nob = min3(kiblnd_rd_frag_size(srcrd, srcidx),
1164 kiblnd_rd_frag_size(dstrd, dstidx),
1167 sge = &tx->tx_sge[tx->tx_nsge];
1168 sge->addr = kiblnd_rd_frag_addr(srcrd, srcidx);
1169 sge->lkey = kiblnd_rd_frag_key(srcrd, srcidx);
1170 sge->length = sge_nob;
1173 wrq = &tx->tx_wrq[tx->tx_nwrq];
1175 wrq->wr.next = &(wrq + 1)->wr;
1176 wrq->wr.wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_RDMA);
1177 wrq->wr.sg_list = sge;
1178 wrq->wr.opcode = IB_WR_RDMA_WRITE;
1179 wrq->wr.send_flags = 0;
1181 #ifdef HAVE_IB_RDMA_WR
1182 wrq->remote_addr = kiblnd_rd_frag_addr(dstrd,
1184 wrq->rkey = kiblnd_rd_frag_key(dstrd,
1187 wrq->wr.wr.rdma.remote_addr = kiblnd_rd_frag_addr(dstrd,
1189 wrq->wr.wr.rdma.rkey = kiblnd_rd_frag_key(dstrd,
1194 srcidx = kiblnd_rd_consume_frag(srcrd, srcidx, sge_nob);
1195 dstidx = kiblnd_rd_consume_frag(dstrd, dstidx, sge_nob);
1198 if (wrq_sge == *kiblnd_tunables.kib_wrq_sge || dstidx != prev) {
1200 wrq->wr.num_sge = wrq_sge;
1206 if (rc < 0) /* no RDMA if completing with failure */
1207 tx->tx_nwrq = tx->tx_nsge = 0;
1209 ibmsg->ibm_u.completion.ibcm_status = rc;
1210 ibmsg->ibm_u.completion.ibcm_cookie = dstcookie;
1211 kiblnd_init_tx_msg(conn->ibc_peer->ibp_ni, tx,
1212 type, sizeof(struct kib_completion_msg));
1218 kiblnd_queue_tx_locked(struct kib_tx *tx, struct kib_conn *conn)
1220 struct list_head *q;
1223 LASSERT(tx->tx_nwrq > 0); /* work items set up */
1224 LASSERT(!tx->tx_queued); /* not queued for sending already */
1225 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1227 if (conn->ibc_state >= IBLND_CONN_DISCONNECTED) {
1228 tx->tx_status = -ECONNABORTED;
1230 if (tx->tx_conn != NULL) {
1231 /* PUT_DONE first attached to conn as a PUT_REQ */
1232 LASSERT(tx->tx_conn == conn);
1233 LASSERT(tx->tx_msg->ibm_type == IBLND_MSG_PUT_DONE);
1235 kiblnd_conn_decref(conn);
1237 list_add(&tx->tx_list, &conn->ibc_zombie_txs);
1242 timeout_ns = kiblnd_timeout() * NSEC_PER_SEC;
1244 tx->tx_deadline = ktime_add_ns(ktime_get(), timeout_ns);
1246 if (tx->tx_conn == NULL) {
1247 kiblnd_conn_addref(conn);
1249 LASSERT (tx->tx_msg->ibm_type != IBLND_MSG_PUT_DONE);
1251 /* PUT_DONE first attached to conn as a PUT_REQ */
1252 LASSERT (tx->tx_conn == conn);
1253 LASSERT (tx->tx_msg->ibm_type == IBLND_MSG_PUT_DONE);
1256 switch (tx->tx_msg->ibm_type) {
1260 case IBLND_MSG_PUT_REQ:
1261 case IBLND_MSG_GET_REQ:
1262 q = &conn->ibc_tx_queue_rsrvd;
1265 case IBLND_MSG_PUT_NAK:
1266 case IBLND_MSG_PUT_ACK:
1267 case IBLND_MSG_PUT_DONE:
1268 case IBLND_MSG_GET_DONE:
1269 q = &conn->ibc_tx_queue_nocred;
1272 case IBLND_MSG_NOOP:
1273 if (IBLND_OOB_CAPABLE(conn->ibc_version))
1274 q = &conn->ibc_tx_queue_nocred;
1276 q = &conn->ibc_tx_noops;
1279 case IBLND_MSG_IMMEDIATE:
1280 q = &conn->ibc_tx_queue;
1284 list_add_tail(&tx->tx_list, q);
1288 kiblnd_queue_tx(struct kib_tx *tx, struct kib_conn *conn)
1290 spin_lock(&conn->ibc_lock);
1291 kiblnd_queue_tx_locked(tx, conn);
1292 kiblnd_check_sends_locked(conn);
1293 spin_unlock(&conn->ibc_lock);
1297 kiblnd_resolve_addr_cap(struct rdma_cm_id *cmid,
1298 struct sockaddr_in *srcaddr,
1299 struct sockaddr_in *dstaddr,
1302 unsigned short port;
1305 LASSERT(capable(CAP_NET_BIND_SERVICE));
1307 /* allow the port to be reused */
1308 rc = rdma_set_reuseaddr(cmid, 1);
1310 CERROR("Unable to set reuse on cmid: %d\n", rc);
1314 /* look for a free privileged port */
1315 for (port = PROT_SOCK-1; port > 0; port--) {
1316 srcaddr->sin_port = htons(port);
1317 rc = rdma_resolve_addr(cmid,
1318 (struct sockaddr *)srcaddr,
1319 (struct sockaddr *)dstaddr,
1322 CDEBUG(D_NET, "bound to port %hu\n", port);
1324 } else if (rc == -EADDRINUSE || rc == -EADDRNOTAVAIL) {
1325 CDEBUG(D_NET, "bind to port %hu failed: %d\n",
1332 CERROR("Failed to bind to a free privileged port\n");
1337 kiblnd_resolve_addr(struct rdma_cm_id *cmid,
1338 struct sockaddr_in *srcaddr,
1339 struct sockaddr_in *dstaddr,
1342 const struct cred *old_creds = NULL;
1343 struct cred *new_creds;
1346 if (!capable(CAP_NET_BIND_SERVICE)) {
1347 new_creds = prepare_creds();
1351 cap_raise(new_creds->cap_effective, CAP_NET_BIND_SERVICE);
1352 old_creds = override_creds(new_creds);
1355 rc = kiblnd_resolve_addr_cap(cmid, srcaddr, dstaddr, timeout_ms);
1358 revert_creds(old_creds);
1364 kiblnd_connect_peer(struct kib_peer_ni *peer_ni)
1366 struct rdma_cm_id *cmid;
1367 struct kib_dev *dev;
1368 struct kib_net *net = peer_ni->ibp_ni->ni_data;
1369 struct sockaddr_in srcaddr;
1370 struct sockaddr_in dstaddr;
1373 LASSERT (net != NULL);
1374 LASSERT (peer_ni->ibp_connecting > 0);
1376 cmid = kiblnd_rdma_create_id(peer_ni->ibp_ni->ni_net_ns,
1377 kiblnd_cm_callback, peer_ni,
1378 RDMA_PS_TCP, IB_QPT_RC);
1381 CERROR("Can't create CMID for %s: %ld\n",
1382 libcfs_nid2str(peer_ni->ibp_nid), PTR_ERR(cmid));
1388 memset(&srcaddr, 0, sizeof(srcaddr));
1389 srcaddr.sin_family = AF_INET;
1390 srcaddr.sin_addr.s_addr = htonl(dev->ibd_ifip);
1392 memset(&dstaddr, 0, sizeof(dstaddr));
1393 dstaddr.sin_family = AF_INET;
1394 dstaddr.sin_port = htons(*kiblnd_tunables.kib_service);
1395 dstaddr.sin_addr.s_addr = htonl(LNET_NIDADDR(peer_ni->ibp_nid));
1397 kiblnd_peer_addref(peer_ni); /* cmid's ref */
1399 if (*kiblnd_tunables.kib_use_priv_port) {
1400 rc = kiblnd_resolve_addr(cmid, &srcaddr, &dstaddr,
1401 kiblnd_timeout() * 1000);
1403 rc = rdma_resolve_addr(cmid,
1404 (struct sockaddr *)&srcaddr,
1405 (struct sockaddr *)&dstaddr,
1406 kiblnd_timeout() * 1000);
1409 /* Can't initiate address resolution: */
1410 CERROR("Can't resolve addr for %s: %d\n",
1411 libcfs_nid2str(peer_ni->ibp_nid), rc);
1418 kiblnd_peer_connect_failed(peer_ni, 1, rc);
1419 kiblnd_peer_decref(peer_ni); /* cmid's ref */
1420 rdma_destroy_id(cmid);
1423 kiblnd_peer_connect_failed(peer_ni, 1, rc);
1427 kiblnd_reconnect_peer(struct kib_peer_ni *peer_ni)
1429 rwlock_t *glock = &kiblnd_data.kib_global_lock;
1430 char *reason = NULL;
1432 unsigned long flags;
1434 write_lock_irqsave(glock, flags);
1435 if (peer_ni->ibp_reconnecting == 0) {
1436 if (peer_ni->ibp_accepting)
1437 reason = "accepting";
1438 else if (peer_ni->ibp_connecting)
1439 reason = "connecting";
1440 else if (!list_empty(&peer_ni->ibp_conns))
1441 reason = "connected";
1442 else /* connected then closed */
1448 if (peer_ni->ibp_accepting)
1449 CNETERR("Detecting race between accepting and reconnecting\n");
1450 peer_ni->ibp_reconnecting--;
1452 if (!kiblnd_peer_active(peer_ni)) {
1453 list_splice_init(&peer_ni->ibp_tx_queue, &txs);
1454 reason = "unlinked";
1458 peer_ni->ibp_connecting++;
1459 peer_ni->ibp_reconnected++;
1461 write_unlock_irqrestore(glock, flags);
1463 kiblnd_connect_peer(peer_ni);
1467 write_unlock_irqrestore(glock, flags);
1469 CWARN("Abort reconnection of %s: %s\n",
1470 libcfs_nid2str(peer_ni->ibp_nid), reason);
1471 kiblnd_txlist_done(&txs, -ECONNABORTED,
1472 LNET_MSG_STATUS_LOCAL_ABORTED);
1477 kiblnd_launch_tx(struct lnet_ni *ni, struct kib_tx *tx, lnet_nid_t nid)
1479 struct kib_peer_ni *peer_ni;
1480 struct kib_peer_ni *peer2;
1481 struct kib_conn *conn;
1482 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
1483 unsigned long flags;
1486 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
1488 /* If I get here, I've committed to send, so I complete the tx with
1489 * failure on any problems */
1491 LASSERT (tx == NULL || tx->tx_conn == NULL); /* only set when assigned a conn */
1492 LASSERT (tx == NULL || tx->tx_nwrq > 0); /* work items have been set up */
1494 /* First time, just use a read lock since I expect to find my peer_ni
1496 read_lock_irqsave(g_lock, flags);
1498 peer_ni = kiblnd_find_peer_locked(ni, nid);
1499 if (peer_ni != NULL && !list_empty(&peer_ni->ibp_conns)) {
1500 /* Found a peer_ni with an established connection */
1501 conn = kiblnd_get_conn_locked(peer_ni);
1502 kiblnd_conn_addref(conn); /* 1 ref for me... */
1504 read_unlock_irqrestore(g_lock, flags);
1507 kiblnd_queue_tx(tx, conn);
1508 kiblnd_conn_decref(conn); /* ...to here */
1512 read_unlock(g_lock);
1513 /* Re-try with a write lock */
1516 peer_ni = kiblnd_find_peer_locked(ni, nid);
1517 if (peer_ni != NULL) {
1518 if (list_empty(&peer_ni->ibp_conns)) {
1519 /* found a peer_ni, but it's still connecting... */
1520 LASSERT(kiblnd_peer_connecting(peer_ni));
1522 list_add_tail(&tx->tx_list,
1523 &peer_ni->ibp_tx_queue);
1524 write_unlock_irqrestore(g_lock, flags);
1526 conn = kiblnd_get_conn_locked(peer_ni);
1527 kiblnd_conn_addref(conn); /* 1 ref for me... */
1529 write_unlock_irqrestore(g_lock, flags);
1532 kiblnd_queue_tx(tx, conn);
1533 kiblnd_conn_decref(conn); /* ...to here */
1538 write_unlock_irqrestore(g_lock, flags);
1540 /* Allocate a peer_ni ready to add to the peer_ni table and retry */
1541 rc = kiblnd_create_peer(ni, &peer_ni, nid);
1543 CERROR("Can't create peer_ni %s\n", libcfs_nid2str(nid));
1545 tx->tx_status = -EHOSTUNREACH;
1547 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
1553 write_lock_irqsave(g_lock, flags);
1555 peer2 = kiblnd_find_peer_locked(ni, nid);
1556 if (peer2 != NULL) {
1557 if (list_empty(&peer2->ibp_conns)) {
1558 /* found a peer_ni, but it's still connecting... */
1559 LASSERT(kiblnd_peer_connecting(peer2));
1561 list_add_tail(&tx->tx_list,
1562 &peer2->ibp_tx_queue);
1563 write_unlock_irqrestore(g_lock, flags);
1565 conn = kiblnd_get_conn_locked(peer2);
1566 kiblnd_conn_addref(conn); /* 1 ref for me... */
1568 write_unlock_irqrestore(g_lock, flags);
1571 kiblnd_queue_tx(tx, conn);
1572 kiblnd_conn_decref(conn); /* ...to here */
1575 kiblnd_peer_decref(peer_ni);
1579 /* Brand new peer_ni */
1580 LASSERT(peer_ni->ibp_connecting == 0);
1581 tunables = &peer_ni->ibp_ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
1582 peer_ni->ibp_connecting = tunables->lnd_conns_per_peer;
1584 /* always called with a ref on ni, which prevents ni being shutdown */
1585 LASSERT(((struct kib_net *)ni->ni_data)->ibn_shutdown == 0);
1588 list_add_tail(&tx->tx_list, &peer_ni->ibp_tx_queue);
1590 kiblnd_peer_addref(peer_ni);
1591 list_add_tail(&peer_ni->ibp_list, kiblnd_nid2peerlist(nid));
1593 write_unlock_irqrestore(g_lock, flags);
1595 for (i = 0; i < tunables->lnd_conns_per_peer; i++)
1596 kiblnd_connect_peer(peer_ni);
1597 kiblnd_peer_decref(peer_ni);
1601 kiblnd_send(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg)
1603 struct lnet_hdr *hdr = &lntmsg->msg_hdr;
1604 int type = lntmsg->msg_type;
1605 struct lnet_process_id target = lntmsg->msg_target;
1606 int target_is_router = lntmsg->msg_target_is_router;
1607 int routing = lntmsg->msg_routing;
1608 unsigned int payload_niov = lntmsg->msg_niov;
1609 struct bio_vec *payload_kiov = lntmsg->msg_kiov;
1610 unsigned int payload_offset = lntmsg->msg_offset;
1611 unsigned int payload_nob = lntmsg->msg_len;
1612 struct kib_msg *ibmsg;
1613 struct kib_rdma_desc *rd;
1618 /* NB 'private' is different depending on what we're sending.... */
1620 CDEBUG(D_NET, "sending %d bytes in %d frags to %s\n",
1621 payload_nob, payload_niov, libcfs_id2str(target));
1623 LASSERT (payload_nob == 0 || payload_niov > 0);
1624 LASSERT (payload_niov <= LNET_MAX_IOV);
1626 /* Thread context */
1627 LASSERT (!in_interrupt());
1635 LASSERT (payload_nob == 0);
1639 if (routing || target_is_router)
1640 break; /* send IMMEDIATE */
1642 /* is the REPLY message too small for RDMA? */
1643 nob = offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[lntmsg->msg_md->md_length]);
1644 if (nob <= IBLND_MSG_SIZE)
1645 break; /* send IMMEDIATE */
1647 tx = kiblnd_get_idle_tx(ni, target.nid);
1649 CERROR("Can't allocate txd for GET to %s\n",
1650 libcfs_nid2str(target.nid));
1655 rd = &ibmsg->ibm_u.get.ibgm_rd;
1656 rc = kiblnd_setup_rd_kiov(ni, tx, rd,
1657 lntmsg->msg_md->md_niov,
1658 lntmsg->msg_md->md_kiov,
1659 0, lntmsg->msg_md->md_length);
1661 CERROR("Can't setup GET sink for %s: %d\n",
1662 libcfs_nid2str(target.nid), rc);
1663 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
1668 nob = offsetof(struct kib_get_msg, ibgm_rd.rd_frags[rd->rd_nfrags]);
1669 ibmsg->ibm_u.get.ibgm_cookie = tx->tx_cookie;
1670 ibmsg->ibm_u.get.ibgm_hdr = *hdr;
1672 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_GET_REQ, nob);
1674 tx->tx_lntmsg[1] = lnet_create_reply_msg(ni, lntmsg);
1675 if (tx->tx_lntmsg[1] == NULL) {
1676 CERROR("Can't create reply for GET -> %s\n",
1677 libcfs_nid2str(target.nid));
1682 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg[0,1] on completion */
1683 tx->tx_waiting = 1; /* waiting for GET_DONE */
1684 kiblnd_launch_tx(ni, tx, target.nid);
1687 case LNET_MSG_REPLY:
1689 /* Is the payload small enough not to need RDMA? */
1690 nob = offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[payload_nob]);
1691 if (nob <= IBLND_MSG_SIZE)
1692 break; /* send IMMEDIATE */
1694 tx = kiblnd_get_idle_tx(ni, target.nid);
1696 CERROR("Can't allocate %s txd for %s\n",
1697 type == LNET_MSG_PUT ? "PUT" : "REPLY",
1698 libcfs_nid2str(target.nid));
1702 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1703 payload_niov, payload_kiov,
1704 payload_offset, payload_nob);
1706 CERROR("Can't setup PUT src for %s: %d\n",
1707 libcfs_nid2str(target.nid), rc);
1713 ibmsg->ibm_u.putreq.ibprm_hdr = *hdr;
1714 ibmsg->ibm_u.putreq.ibprm_cookie = tx->tx_cookie;
1715 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_REQ,
1716 sizeof(struct kib_putreq_msg));
1718 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1719 tx->tx_waiting = 1; /* waiting for PUT_{ACK,NAK} */
1720 kiblnd_launch_tx(ni, tx, target.nid);
1724 /* send IMMEDIATE */
1725 LASSERT(offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[payload_nob])
1728 tx = kiblnd_get_idle_tx(ni, target.nid);
1730 CERROR ("Can't send %d to %s: tx descs exhausted\n",
1731 type, libcfs_nid2str(target.nid));
1736 ibmsg->ibm_u.immediate.ibim_hdr = *hdr;
1738 lnet_copy_kiov2flat(IBLND_MSG_SIZE, ibmsg,
1739 offsetof(struct kib_msg,
1740 ibm_u.immediate.ibim_payload),
1741 payload_niov, payload_kiov,
1742 payload_offset, payload_nob);
1744 nob = offsetof(struct kib_immediate_msg, ibim_payload[payload_nob]);
1745 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_IMMEDIATE, nob);
1747 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1748 kiblnd_launch_tx(ni, tx, target.nid);
1753 kiblnd_reply(struct lnet_ni *ni, struct kib_rx *rx, struct lnet_msg *lntmsg)
1755 struct lnet_process_id target = lntmsg->msg_target;
1756 unsigned int niov = lntmsg->msg_niov;
1757 struct bio_vec *kiov = lntmsg->msg_kiov;
1758 unsigned int offset = lntmsg->msg_offset;
1759 unsigned int nob = lntmsg->msg_len;
1763 tx = kiblnd_get_idle_tx(ni, rx->rx_conn->ibc_peer->ibp_nid);
1765 CERROR("Can't get tx for REPLY to %s\n",
1766 libcfs_nid2str(target.nid));
1773 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1774 niov, kiov, offset, nob);
1777 CERROR("Can't setup GET src for %s: %d\n",
1778 libcfs_nid2str(target.nid), rc);
1782 rc = kiblnd_init_rdma(rx->rx_conn, tx,
1783 IBLND_MSG_GET_DONE, nob,
1784 &rx->rx_msg->ibm_u.get.ibgm_rd,
1785 rx->rx_msg->ibm_u.get.ibgm_cookie);
1787 CERROR("Can't setup rdma for GET from %s: %d\n",
1788 libcfs_nid2str(target.nid), rc);
1793 /* No RDMA: local completion may happen now! */
1794 lnet_finalize(lntmsg, 0);
1796 /* RDMA: lnet_finalize(lntmsg) when it
1798 tx->tx_lntmsg[0] = lntmsg;
1801 kiblnd_queue_tx(tx, rx->rx_conn);
1806 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
1809 lnet_finalize(lntmsg, -EIO);
1813 kiblnd_recv(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg,
1814 int delayed, unsigned int niov, struct bio_vec *kiov,
1815 unsigned int offset, unsigned int mlen, unsigned int rlen)
1817 struct kib_rx *rx = private;
1818 struct kib_msg *rxmsg = rx->rx_msg;
1819 struct kib_conn *conn = rx->rx_conn;
1823 int post_credit = IBLND_POSTRX_PEER_CREDIT;
1826 LASSERT (mlen <= rlen);
1827 LASSERT (!in_interrupt());
1829 switch (rxmsg->ibm_type) {
1833 case IBLND_MSG_IMMEDIATE:
1834 nob = offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[rlen]);
1835 if (nob > rx->rx_nob) {
1836 CERROR ("Immediate message from %s too big: %d(%d)\n",
1837 libcfs_nid2str(rxmsg->ibm_u.immediate.ibim_hdr.src_nid),
1843 lnet_copy_flat2kiov(niov, kiov, offset,
1844 IBLND_MSG_SIZE, rxmsg,
1845 offsetof(struct kib_msg,
1846 ibm_u.immediate.ibim_payload),
1848 lnet_finalize(lntmsg, 0);
1851 case IBLND_MSG_PUT_REQ: {
1852 struct kib_msg *txmsg;
1853 struct kib_rdma_desc *rd;
1854 ibprm_cookie = rxmsg->ibm_u.putreq.ibprm_cookie;
1857 lnet_finalize(lntmsg, 0);
1858 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK,
1863 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
1865 CERROR("Can't allocate tx for %s\n",
1866 libcfs_nid2str(conn->ibc_peer->ibp_nid));
1867 /* Not replying will break the connection */
1873 rd = &txmsg->ibm_u.putack.ibpam_rd;
1874 rc = kiblnd_setup_rd_kiov(ni, tx, rd,
1875 niov, kiov, offset, mlen);
1877 CERROR("Can't setup PUT sink for %s: %d\n",
1878 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
1879 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
1881 /* tell peer_ni it's over */
1882 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK,
1887 nob = offsetof(struct kib_putack_msg, ibpam_rd.rd_frags[rd->rd_nfrags]);
1888 txmsg->ibm_u.putack.ibpam_src_cookie = ibprm_cookie;
1889 txmsg->ibm_u.putack.ibpam_dst_cookie = tx->tx_cookie;
1891 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_ACK, nob);
1893 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1894 tx->tx_waiting = 1; /* waiting for PUT_DONE */
1895 kiblnd_queue_tx(tx, conn);
1897 /* reposted buffer reserved for PUT_DONE */
1898 post_credit = IBLND_POSTRX_NO_CREDIT;
1902 case IBLND_MSG_GET_REQ:
1903 if (lntmsg != NULL) {
1904 /* Optimized GET; RDMA lntmsg's payload */
1905 kiblnd_reply(ni, rx, lntmsg);
1907 /* GET didn't match anything */
1908 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_GET_DONE,
1910 rxmsg->ibm_u.get.ibgm_cookie);
1915 kiblnd_post_rx(rx, post_credit);
1920 kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name)
1922 struct task_struct *task = kthread_run(fn, arg, name);
1925 return PTR_ERR(task);
1927 atomic_inc(&kiblnd_data.kib_nthreads);
1932 kiblnd_thread_fini (void)
1934 atomic_dec (&kiblnd_data.kib_nthreads);
1938 kiblnd_peer_alive(struct kib_peer_ni *peer_ni)
1940 /* This is racy, but everyone's only writing ktime_get_seconds() */
1941 peer_ni->ibp_last_alive = ktime_get_seconds();
1946 kiblnd_peer_notify(struct kib_peer_ni *peer_ni)
1949 time64_t last_alive = 0;
1950 unsigned long flags;
1952 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1954 if (kiblnd_peer_idle(peer_ni) && peer_ni->ibp_error != 0) {
1955 error = peer_ni->ibp_error;
1956 peer_ni->ibp_error = 0;
1958 last_alive = peer_ni->ibp_last_alive;
1961 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1964 lnet_notify(peer_ni->ibp_ni,
1965 peer_ni->ibp_nid, false, false, last_alive);
1969 kiblnd_close_conn_locked(struct kib_conn *conn, int error)
1971 /* This just does the immediate housekeeping. 'error' is zero for a
1972 * normal shutdown which can happen only after the connection has been
1973 * established. If the connection is established, schedule the
1974 * connection to be finished off by the connd. Otherwise the connd is
1975 * already dealing with it (either to set it up or tear it down).
1976 * Caller holds kib_global_lock exclusively in irq context */
1977 struct kib_peer_ni *peer_ni = conn->ibc_peer;
1978 struct kib_dev *dev;
1979 unsigned long flags;
1981 LASSERT (error != 0 || conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1983 if (error != 0 && conn->ibc_comms_error == 0)
1984 conn->ibc_comms_error = error;
1986 if (conn->ibc_state != IBLND_CONN_ESTABLISHED)
1987 return; /* already being handled */
1990 list_empty(&conn->ibc_tx_noops) &&
1991 list_empty(&conn->ibc_tx_queue) &&
1992 list_empty(&conn->ibc_tx_queue_rsrvd) &&
1993 list_empty(&conn->ibc_tx_queue_nocred) &&
1994 list_empty(&conn->ibc_active_txs)) {
1995 CDEBUG(D_NET, "closing conn to %s\n",
1996 libcfs_nid2str(peer_ni->ibp_nid));
1998 CNETERR("Closing conn to %s: error %d%s%s%s%s%s\n",
1999 libcfs_nid2str(peer_ni->ibp_nid), error,
2000 list_empty(&conn->ibc_tx_queue) ? "" : "(sending)",
2001 list_empty(&conn->ibc_tx_noops) ? "" : "(sending_noops)",
2002 list_empty(&conn->ibc_tx_queue_rsrvd) ?
2003 "" : "(sending_rsrvd)",
2004 list_empty(&conn->ibc_tx_queue_nocred) ?
2005 "" : "(sending_nocred)",
2006 list_empty(&conn->ibc_active_txs) ? "" : "(waiting)");
2009 dev = ((struct kib_net *)peer_ni->ibp_ni->ni_data)->ibn_dev;
2010 if (peer_ni->ibp_next_conn == conn)
2011 /* clear next_conn so it won't be used */
2012 peer_ni->ibp_next_conn = NULL;
2013 list_del(&conn->ibc_list);
2014 /* connd (see below) takes over ibc_list's ref */
2016 if (list_empty(&peer_ni->ibp_conns) && /* no more conns */
2017 kiblnd_peer_active(peer_ni)) { /* still in peer_ni table */
2018 kiblnd_unlink_peer_locked(peer_ni);
2020 /* set/clear error on last conn */
2021 peer_ni->ibp_error = conn->ibc_comms_error;
2024 kiblnd_set_conn_state(conn, IBLND_CONN_CLOSING);
2027 kiblnd_dev_can_failover(dev)) {
2028 list_add_tail(&dev->ibd_fail_list,
2029 &kiblnd_data.kib_failed_devs);
2030 wake_up(&kiblnd_data.kib_failover_waitq);
2033 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
2035 list_add_tail(&conn->ibc_list, &kiblnd_data.kib_connd_conns);
2036 wake_up(&kiblnd_data.kib_connd_waitq);
2038 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
2042 kiblnd_close_conn(struct kib_conn *conn, int error)
2044 unsigned long flags;
2046 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2048 kiblnd_close_conn_locked(conn, error);
2050 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2054 kiblnd_handle_early_rxs(struct kib_conn *conn)
2056 unsigned long flags;
2059 LASSERT(!in_interrupt());
2060 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
2062 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2063 while (!list_empty(&conn->ibc_early_rxs)) {
2064 rx = list_entry(conn->ibc_early_rxs.next,
2065 struct kib_rx, rx_list);
2066 list_del(&rx->rx_list);
2067 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2069 kiblnd_handle_rx(rx);
2071 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2073 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2077 kiblnd_abort_txs(struct kib_conn *conn, struct list_head *txs)
2083 spin_lock(&conn->ibc_lock);
2085 list_for_each_entry_safe(tx, nxt, txs, tx_list) {
2086 if (txs == &conn->ibc_active_txs) {
2087 LASSERT(!tx->tx_queued);
2088 LASSERT(tx->tx_waiting ||
2089 tx->tx_sending != 0);
2090 if (conn->ibc_comms_error == -ETIMEDOUT) {
2091 if (tx->tx_waiting && !tx->tx_sending)
2093 LNET_MSG_STATUS_REMOTE_TIMEOUT;
2094 else if (tx->tx_sending)
2096 LNET_MSG_STATUS_NETWORK_TIMEOUT;
2099 LASSERT(tx->tx_queued);
2100 if (conn->ibc_comms_error == -ETIMEDOUT)
2101 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_TIMEOUT;
2103 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
2106 tx->tx_status = -ECONNABORTED;
2110 * TODO: This makes an assumption that
2111 * kiblnd_tx_complete() will be called for each tx. If
2112 * that event is dropped we could end up with stale
2113 * connections floating around. We'd like to deal with
2114 * that in a better way.
2116 * Also that means we can exceed the timeout by many
2119 if (tx->tx_sending == 0) {
2121 list_move(&tx->tx_list, &zombies);
2125 spin_unlock(&conn->ibc_lock);
2128 * aborting transmits occurs when finalizing the connection.
2129 * The connection is finalized on error.
2130 * Passing LNET_MSG_STATUS_OK to txlist_done() will not
2131 * override the value already set in tx->tx_hstatus above.
2133 kiblnd_txlist_done(&zombies, -ECONNABORTED, LNET_MSG_STATUS_OK);
2137 kiblnd_finalise_conn(struct kib_conn *conn)
2139 LASSERT (!in_interrupt());
2140 LASSERT (conn->ibc_state > IBLND_CONN_INIT);
2142 /* abort_receives moves QP state to IB_QPS_ERR. This is only required
2143 * for connections that didn't get as far as being connected, because
2144 * rdma_disconnect() does this for free. */
2145 kiblnd_abort_receives(conn);
2147 kiblnd_set_conn_state(conn, IBLND_CONN_DISCONNECTED);
2149 /* Complete all tx descs not waiting for sends to complete.
2150 * NB we should be safe from RDMA now that the QP has changed state */
2152 kiblnd_abort_txs(conn, &conn->ibc_tx_noops);
2153 kiblnd_abort_txs(conn, &conn->ibc_tx_queue);
2154 kiblnd_abort_txs(conn, &conn->ibc_tx_queue_rsrvd);
2155 kiblnd_abort_txs(conn, &conn->ibc_tx_queue_nocred);
2156 kiblnd_abort_txs(conn, &conn->ibc_active_txs);
2158 kiblnd_handle_early_rxs(conn);
2162 kiblnd_peer_connect_failed(struct kib_peer_ni *peer_ni, int active,
2166 unsigned long flags;
2168 LASSERT (error != 0);
2169 LASSERT (!in_interrupt());
2171 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2174 LASSERT(peer_ni->ibp_connecting > 0);
2175 peer_ni->ibp_connecting--;
2177 LASSERT (peer_ni->ibp_accepting > 0);
2178 peer_ni->ibp_accepting--;
2181 if (kiblnd_peer_connecting(peer_ni)) {
2182 /* another connection attempt under way... */
2183 write_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2188 peer_ni->ibp_reconnected = 0;
2189 if (list_empty(&peer_ni->ibp_conns)) {
2190 /* Take peer_ni's blocked transmits to complete with error */
2191 list_splice_init(&peer_ni->ibp_tx_queue, &zombies);
2193 if (kiblnd_peer_active(peer_ni))
2194 kiblnd_unlink_peer_locked(peer_ni);
2196 peer_ni->ibp_error = error;
2198 /* Can't have blocked transmits if there are connections */
2199 LASSERT(list_empty(&peer_ni->ibp_tx_queue));
2202 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2204 kiblnd_peer_notify(peer_ni);
2206 if (list_empty(&zombies))
2209 CNETERR("Deleting messages for %s: connection failed\n",
2210 libcfs_nid2str(peer_ni->ibp_nid));
2212 kiblnd_txlist_done(&zombies, error,
2213 LNET_MSG_STATUS_LOCAL_DROPPED);
2217 kiblnd_connreq_done(struct kib_conn *conn, int status)
2219 struct kib_peer_ni *peer_ni = conn->ibc_peer;
2222 unsigned long flags;
2225 active = (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2227 CDEBUG(D_NET,"%s: active(%d), version(%x), status(%d)\n",
2228 libcfs_nid2str(peer_ni->ibp_nid), active,
2229 conn->ibc_version, status);
2231 LASSERT (!in_interrupt());
2232 LASSERT ((conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT &&
2233 peer_ni->ibp_connecting > 0) ||
2234 (conn->ibc_state == IBLND_CONN_PASSIVE_WAIT &&
2235 peer_ni->ibp_accepting > 0));
2237 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
2238 conn->ibc_connvars = NULL;
2241 /* failed to establish connection */
2242 kiblnd_peer_connect_failed(peer_ni, active, status);
2243 kiblnd_finalise_conn(conn);
2247 /* connection established */
2248 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2250 /* reset retry count */
2251 peer_ni->ibp_retries = 0;
2253 conn->ibc_last_send = ktime_get();
2254 kiblnd_set_conn_state(conn, IBLND_CONN_ESTABLISHED);
2255 kiblnd_peer_alive(peer_ni);
2257 /* Add conn to peer_ni's list and nuke any dangling conns from a different
2258 * peer_ni instance... */
2259 kiblnd_conn_addref(conn); /* +1 ref for ibc_list */
2260 list_add(&conn->ibc_list, &peer_ni->ibp_conns);
2261 peer_ni->ibp_reconnected = 0;
2263 peer_ni->ibp_connecting--;
2265 peer_ni->ibp_accepting--;
2267 if (peer_ni->ibp_version == 0) {
2268 peer_ni->ibp_version = conn->ibc_version;
2269 peer_ni->ibp_incarnation = conn->ibc_incarnation;
2272 if (peer_ni->ibp_version != conn->ibc_version ||
2273 peer_ni->ibp_incarnation != conn->ibc_incarnation) {
2274 kiblnd_close_stale_conns_locked(peer_ni, conn->ibc_version,
2275 conn->ibc_incarnation);
2276 peer_ni->ibp_version = conn->ibc_version;
2277 peer_ni->ibp_incarnation = conn->ibc_incarnation;
2280 /* grab pending txs while I have the lock */
2281 list_splice_init(&peer_ni->ibp_tx_queue, &txs);
2283 if (!kiblnd_peer_active(peer_ni) || /* peer_ni has been deleted */
2284 conn->ibc_comms_error != 0) { /* error has happened already */
2286 /* start to shut down connection */
2287 kiblnd_close_conn_locked(conn, -ECONNABORTED);
2288 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2290 kiblnd_txlist_done(&txs, -ECONNABORTED,
2291 LNET_MSG_STATUS_LOCAL_ERROR);
2296 /* +1 ref for myself, this connection is visible to other threads
2297 * now, refcount of peer:ibp_conns can be released by connection
2298 * close from either a different thread, or the calling of
2299 * kiblnd_check_sends_locked() below. See bz21911 for details.
2301 kiblnd_conn_addref(conn);
2302 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2304 /* Schedule blocked txs
2305 * Note: if we are running with conns_per_peer > 1, these blocked
2306 * txs will all get scheduled to the first connection which gets
2307 * scheduled. We won't be using round robin on this first batch.
2309 spin_lock(&conn->ibc_lock);
2310 while (!list_empty(&txs)) {
2311 tx = list_entry(txs.next, struct kib_tx, tx_list);
2312 list_del(&tx->tx_list);
2314 kiblnd_queue_tx_locked(tx, conn);
2316 kiblnd_check_sends_locked(conn);
2317 spin_unlock(&conn->ibc_lock);
2319 /* schedule blocked rxs */
2320 kiblnd_handle_early_rxs(conn);
2321 kiblnd_conn_decref(conn);
2325 kiblnd_reject(struct rdma_cm_id *cmid, struct kib_rej *rej)
2329 #ifdef HAVE_RDMA_REJECT_4ARGS
2330 rc = rdma_reject(cmid, rej, sizeof(*rej), IB_CM_REJ_CONSUMER_DEFINED);
2332 rc = rdma_reject(cmid, rej, sizeof(*rej));
2336 CWARN("Error %d sending reject\n", rc);
2340 kiblnd_passive_connect(struct rdma_cm_id *cmid, void *priv, int priv_nob)
2342 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2343 struct kib_msg *reqmsg = priv;
2344 struct kib_msg *ackmsg;
2345 struct kib_dev *ibdev;
2346 struct kib_peer_ni *peer_ni;
2347 struct kib_peer_ni *peer2;
2348 struct kib_conn *conn;
2349 struct lnet_ni *ni = NULL;
2350 struct kib_net *net = NULL;
2352 struct rdma_conn_param cp;
2354 int version = IBLND_MSG_VERSION;
2355 unsigned long flags;
2357 struct sockaddr_in *peer_addr;
2358 LASSERT (!in_interrupt());
2360 /* cmid inherits 'context' from the corresponding listener id */
2361 ibdev = cmid->context;
2364 memset(&rej, 0, sizeof(rej));
2365 rej.ibr_magic = IBLND_MSG_MAGIC;
2366 rej.ibr_why = IBLND_REJECT_FATAL;
2367 rej.ibr_cp.ibcp_max_msg_size = IBLND_MSG_SIZE;
2369 peer_addr = (struct sockaddr_in *)&(cmid->route.addr.dst_addr);
2370 if (*kiblnd_tunables.kib_require_priv_port &&
2371 ntohs(peer_addr->sin_port) >= PROT_SOCK) {
2372 __u32 ip = ntohl(peer_addr->sin_addr.s_addr);
2373 CERROR("peer_ni's port (%pI4h:%hu) is not privileged\n",
2374 &ip, ntohs(peer_addr->sin_port));
2378 if (priv_nob < offsetof(struct kib_msg, ibm_type)) {
2379 CERROR("Short connection request\n");
2383 /* Future protocol version compatibility support! If the
2384 * o2iblnd-specific protocol changes, or when LNET unifies
2385 * protocols over all LNDs, the initial connection will
2386 * negotiate a protocol version. I trap this here to avoid
2387 * console errors; the reject tells the peer_ni which protocol I
2389 if (reqmsg->ibm_magic == LNET_PROTO_MAGIC ||
2390 reqmsg->ibm_magic == __swab32(LNET_PROTO_MAGIC))
2392 if (reqmsg->ibm_magic == IBLND_MSG_MAGIC &&
2393 reqmsg->ibm_version != IBLND_MSG_VERSION &&
2394 reqmsg->ibm_version != IBLND_MSG_VERSION_1)
2396 if (reqmsg->ibm_magic == __swab32(IBLND_MSG_MAGIC) &&
2397 reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION) &&
2398 reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION_1))
2401 rc = kiblnd_unpack_msg(reqmsg, priv_nob);
2403 CERROR("Can't parse connection request: %d\n", rc);
2407 nid = reqmsg->ibm_srcnid;
2408 ni = lnet_nid2ni_addref(reqmsg->ibm_dstnid);
2411 net = (struct kib_net *)ni->ni_data;
2412 rej.ibr_incarnation = net->ibn_incarnation;
2415 if (ni == NULL || /* no matching net */
2416 ni->ni_nid != reqmsg->ibm_dstnid || /* right NET, wrong NID! */
2417 net->ibn_dev != ibdev) { /* wrong device */
2418 CERROR("Can't accept conn from %s on %s (%s:%d:%pI4h): "
2419 "bad dst nid %s\n", libcfs_nid2str(nid),
2420 ni == NULL ? "NA" : libcfs_nid2str(ni->ni_nid),
2421 ibdev->ibd_ifname, ibdev->ibd_nnets,
2423 libcfs_nid2str(reqmsg->ibm_dstnid));
2428 /* check time stamp as soon as possible */
2429 if (reqmsg->ibm_dststamp != 0 &&
2430 reqmsg->ibm_dststamp != net->ibn_incarnation) {
2431 CWARN("Stale connection request\n");
2432 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2436 /* I can accept peer_ni's version */
2437 version = reqmsg->ibm_version;
2439 if (reqmsg->ibm_type != IBLND_MSG_CONNREQ) {
2440 CERROR("Unexpected connreq msg type: %x from %s\n",
2441 reqmsg->ibm_type, libcfs_nid2str(nid));
2445 if (reqmsg->ibm_u.connparams.ibcp_queue_depth >
2446 kiblnd_msg_queue_size(version, ni)) {
2447 CERROR("Can't accept conn from %s, queue depth too large: "
2448 " %d (<=%d wanted)\n",
2449 libcfs_nid2str(nid),
2450 reqmsg->ibm_u.connparams.ibcp_queue_depth,
2451 kiblnd_msg_queue_size(version, ni));
2453 if (version == IBLND_MSG_VERSION)
2454 rej.ibr_why = IBLND_REJECT_MSG_QUEUE_SIZE;
2459 if (reqmsg->ibm_u.connparams.ibcp_max_frags >
2460 IBLND_MAX_RDMA_FRAGS) {
2461 CWARN("Can't accept conn from %s (version %x): "
2462 "max_frags %d too large (%d wanted)\n",
2463 libcfs_nid2str(nid), version,
2464 reqmsg->ibm_u.connparams.ibcp_max_frags,
2465 IBLND_MAX_RDMA_FRAGS);
2467 if (version >= IBLND_MSG_VERSION)
2468 rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2471 } else if (reqmsg->ibm_u.connparams.ibcp_max_frags <
2472 IBLND_MAX_RDMA_FRAGS &&
2473 net->ibn_fmr_ps == NULL) {
2474 CWARN("Can't accept conn from %s (version %x): "
2475 "max_frags %d incompatible without FMR pool "
2477 libcfs_nid2str(nid), version,
2478 reqmsg->ibm_u.connparams.ibcp_max_frags,
2479 IBLND_MAX_RDMA_FRAGS);
2481 if (version == IBLND_MSG_VERSION)
2482 rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2487 if (reqmsg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2488 CERROR("Can't accept %s: message size %d too big (%d max)\n",
2489 libcfs_nid2str(nid),
2490 reqmsg->ibm_u.connparams.ibcp_max_msg_size,
2495 /* assume 'nid' is a new peer_ni; create */
2496 rc = kiblnd_create_peer(ni, &peer_ni, nid);
2498 CERROR("Can't create peer_ni for %s\n", libcfs_nid2str(nid));
2499 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2503 /* We have validated the peer's parameters so use those */
2504 peer_ni->ibp_max_frags = reqmsg->ibm_u.connparams.ibcp_max_frags;
2505 peer_ni->ibp_queue_depth = reqmsg->ibm_u.connparams.ibcp_queue_depth;
2507 write_lock_irqsave(g_lock, flags);
2509 peer2 = kiblnd_find_peer_locked(ni, nid);
2510 if (peer2 != NULL) {
2511 if (peer2->ibp_version == 0) {
2512 peer2->ibp_version = version;
2513 peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2516 /* not the guy I've talked with */
2517 if (peer2->ibp_incarnation != reqmsg->ibm_srcstamp ||
2518 peer2->ibp_version != version) {
2519 kiblnd_close_peer_conns_locked(peer2, -ESTALE);
2521 if (kiblnd_peer_active(peer2)) {
2522 peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2523 peer2->ibp_version = version;
2525 write_unlock_irqrestore(g_lock, flags);
2527 CWARN("Conn stale %s version %x/%x incarnation %llu/%llu\n",
2528 libcfs_nid2str(nid), peer2->ibp_version, version,
2529 peer2->ibp_incarnation, reqmsg->ibm_srcstamp);
2531 kiblnd_peer_decref(peer_ni);
2532 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2536 /* Tie-break connection race in favour of the higher NID.
2537 * If we keep running into a race condition multiple times,
2538 * we have to assume that the connection attempt with the
2539 * higher NID is stuck in a connecting state and will never
2540 * recover. As such, we pass through this if-block and let
2541 * the lower NID connection win so we can move forward.
2543 if (peer2->ibp_connecting != 0 &&
2544 nid < ni->ni_nid && peer2->ibp_races <
2545 MAX_CONN_RACES_BEFORE_ABORT) {
2547 write_unlock_irqrestore(g_lock, flags);
2549 CDEBUG(D_NET, "Conn race %s\n",
2550 libcfs_nid2str(peer2->ibp_nid));
2552 kiblnd_peer_decref(peer_ni);
2553 rej.ibr_why = IBLND_REJECT_CONN_RACE;
2556 if (peer2->ibp_races >= MAX_CONN_RACES_BEFORE_ABORT)
2557 CNETERR("Conn race %s: unresolved after %d attempts, letting lower NID win\n",
2558 libcfs_nid2str(peer2->ibp_nid),
2559 MAX_CONN_RACES_BEFORE_ABORT);
2561 * passive connection is allowed even this peer_ni is waiting for
2564 peer2->ibp_reconnecting = 0;
2565 peer2->ibp_races = 0;
2566 peer2->ibp_accepting++;
2567 kiblnd_peer_addref(peer2);
2569 /* Race with kiblnd_launch_tx (active connect) to create peer_ni
2570 * so copy validated parameters since we now know what the
2571 * peer_ni's limits are */
2572 peer2->ibp_max_frags = peer_ni->ibp_max_frags;
2573 peer2->ibp_queue_depth = peer_ni->ibp_queue_depth;
2575 write_unlock_irqrestore(g_lock, flags);
2576 kiblnd_peer_decref(peer_ni);
2579 /* Brand new peer_ni */
2580 LASSERT (peer_ni->ibp_accepting == 0);
2581 LASSERT (peer_ni->ibp_version == 0 &&
2582 peer_ni->ibp_incarnation == 0);
2584 peer_ni->ibp_accepting = 1;
2585 peer_ni->ibp_version = version;
2586 peer_ni->ibp_incarnation = reqmsg->ibm_srcstamp;
2588 /* I have a ref on ni that prevents it being shutdown */
2589 LASSERT (net->ibn_shutdown == 0);
2591 kiblnd_peer_addref(peer_ni);
2592 list_add_tail(&peer_ni->ibp_list, kiblnd_nid2peerlist(nid));
2594 write_unlock_irqrestore(g_lock, flags);
2597 conn = kiblnd_create_conn(peer_ni, cmid, IBLND_CONN_PASSIVE_WAIT, version);
2599 kiblnd_peer_connect_failed(peer_ni, 0, -ENOMEM);
2600 kiblnd_peer_decref(peer_ni);
2601 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2605 /* conn now "owns" cmid, so I return success from here on to ensure the
2606 * CM callback doesn't destroy cmid. */
2607 conn->ibc_incarnation = reqmsg->ibm_srcstamp;
2608 conn->ibc_credits = conn->ibc_queue_depth;
2609 conn->ibc_reserved_credits = conn->ibc_queue_depth;
2610 LASSERT(conn->ibc_credits + conn->ibc_reserved_credits +
2611 IBLND_OOB_MSGS(version) <= IBLND_RX_MSGS(conn));
2613 ackmsg = &conn->ibc_connvars->cv_msg;
2614 memset(ackmsg, 0, sizeof(*ackmsg));
2616 kiblnd_init_msg(ackmsg, IBLND_MSG_CONNACK,
2617 sizeof(ackmsg->ibm_u.connparams));
2618 ackmsg->ibm_u.connparams.ibcp_queue_depth = conn->ibc_queue_depth;
2619 ackmsg->ibm_u.connparams.ibcp_max_frags = conn->ibc_max_frags;
2620 ackmsg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2622 kiblnd_pack_msg(ni, ackmsg, version, 0, nid, reqmsg->ibm_srcstamp);
2624 memset(&cp, 0, sizeof(cp));
2625 cp.private_data = ackmsg;
2626 cp.private_data_len = ackmsg->ibm_nob;
2627 cp.responder_resources = 0; /* No atomic ops or RDMA reads */
2628 cp.initiator_depth = 0;
2629 cp.flow_control = 1;
2630 cp.retry_count = *kiblnd_tunables.kib_retry_count;
2631 cp.rnr_retry_count = *kiblnd_tunables.kib_rnr_retry_count;
2633 CDEBUG(D_NET, "Accept %s\n", libcfs_nid2str(nid));
2635 rc = rdma_accept(cmid, &cp);
2637 CERROR("Can't accept %s: %d\n", libcfs_nid2str(nid), rc);
2638 rej.ibr_version = version;
2639 rej.ibr_why = IBLND_REJECT_FATAL;
2641 kiblnd_reject(cmid, &rej);
2642 kiblnd_connreq_done(conn, rc);
2643 kiblnd_conn_decref(conn);
2651 rej.ibr_cp.ibcp_queue_depth =
2652 kiblnd_msg_queue_size(version, ni);
2653 rej.ibr_cp.ibcp_max_frags = IBLND_MAX_RDMA_FRAGS;
2657 rej.ibr_version = version;
2658 kiblnd_reject(cmid, &rej);
2660 return -ECONNREFUSED;
2664 kiblnd_check_reconnect(struct kib_conn *conn, int version,
2665 u64 incarnation, int why, struct kib_connparams *cp)
2667 rwlock_t *glock = &kiblnd_data.kib_global_lock;
2668 struct kib_peer_ni *peer_ni = conn->ibc_peer;
2670 int msg_size = IBLND_MSG_SIZE;
2674 unsigned long flags;
2676 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2677 LASSERT(peer_ni->ibp_connecting > 0); /* 'conn' at least */
2680 msg_size = cp->ibcp_max_msg_size;
2681 frag_num = cp->ibcp_max_frags;
2682 queue_dep = cp->ibcp_queue_depth;
2685 write_lock_irqsave(glock, flags);
2686 /* retry connection if it's still needed and no other connection
2687 * attempts (active or passive) are in progress
2688 * NB: reconnect is still needed even when ibp_tx_queue is
2689 * empty if ibp_version != version because reconnect may be
2692 reconnect = (!list_empty(&peer_ni->ibp_tx_queue) ||
2693 peer_ni->ibp_version != version) &&
2694 peer_ni->ibp_connecting &&
2695 peer_ni->ibp_accepting == 0;
2701 if (peer_ni->ibp_retries > *kiblnd_tunables.kib_retry_count) {
2702 reason = "retry count exceeded due to no listener";
2711 case IBLND_REJECT_RDMA_FRAGS: {
2712 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2715 reason = "can't negotiate max frags";
2718 tunables = &peer_ni->ibp_ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2719 #ifdef HAVE_IB_GET_DMA_MR
2721 * This check only makes sense if the kernel supports global
2722 * memory registration. Otherwise, map_on_demand will never == 0
2724 if (!tunables->lnd_map_on_demand) {
2725 reason = "map_on_demand must be enabled";
2729 if (conn->ibc_max_frags <= frag_num) {
2730 reason = "unsupported max frags";
2734 peer_ni->ibp_max_frags = frag_num;
2735 reason = "rdma fragments";
2738 case IBLND_REJECT_MSG_QUEUE_SIZE:
2740 reason = "can't negotiate queue depth";
2743 if (conn->ibc_queue_depth <= queue_dep) {
2744 reason = "unsupported queue depth";
2748 peer_ni->ibp_queue_depth = queue_dep;
2749 reason = "queue depth";
2752 case IBLND_REJECT_CONN_STALE:
2756 case IBLND_REJECT_CONN_RACE:
2757 reason = "conn race";
2760 case IBLND_REJECT_CONN_UNCOMPAT:
2761 reason = "version negotiation";
2764 case IBLND_REJECT_INVALID_SRV_ID:
2765 reason = "invalid service id";
2769 conn->ibc_reconnect = 1;
2770 peer_ni->ibp_reconnecting++;
2771 peer_ni->ibp_version = version;
2772 if (incarnation != 0)
2773 peer_ni->ibp_incarnation = incarnation;
2775 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2777 CNETERR("%s: %s (%s), %x, %x, msg_size: %d, queue_depth: %d/%d, max_frags: %d/%d\n",
2778 libcfs_nid2str(peer_ni->ibp_nid),
2779 reconnect ? "reconnect" : "don't reconnect",
2780 reason, IBLND_MSG_VERSION, version, msg_size,
2781 conn->ibc_queue_depth, queue_dep,
2782 conn->ibc_max_frags, frag_num);
2784 * if conn::ibc_reconnect is TRUE, connd will reconnect to the peer_ni
2785 * while destroying the zombie
2790 kiblnd_rejected(struct kib_conn *conn, int reason, void *priv, int priv_nob)
2792 struct kib_peer_ni *peer_ni = conn->ibc_peer;
2794 LASSERT (!in_interrupt());
2795 LASSERT (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2798 case IB_CM_REJ_STALE_CONN:
2799 kiblnd_check_reconnect(conn, IBLND_MSG_VERSION, 0,
2800 IBLND_REJECT_CONN_STALE, NULL);
2803 case IB_CM_REJ_INVALID_SERVICE_ID:
2804 peer_ni->ibp_retries++;
2805 kiblnd_check_reconnect(conn, IBLND_MSG_VERSION, 0,
2806 IBLND_REJECT_INVALID_SRV_ID, NULL);
2807 CNETERR("%s rejected: no listener at %d\n",
2808 libcfs_nid2str(peer_ni->ibp_nid),
2809 *kiblnd_tunables.kib_service);
2812 case IB_CM_REJ_CONSUMER_DEFINED:
2813 if (priv_nob >= offsetof(struct kib_rej, ibr_padding)) {
2814 struct kib_rej *rej = priv;
2815 struct kib_connparams *cp = NULL;
2817 __u64 incarnation = -1;
2819 /* NB. default incarnation is -1 because:
2820 * a) V1 will ignore dst incarnation in connreq.
2821 * b) V2 will provide incarnation while rejecting me,
2822 * -1 will be overwrote.
2824 * if I try to connect to a V1 peer_ni with V2 protocol,
2825 * it rejected me then upgrade to V2, I have no idea
2826 * about the upgrading and try to reconnect with V1,
2827 * in this case upgraded V2 can find out I'm trying to
2828 * talk to the old guy and reject me(incarnation is -1).
2831 if (rej->ibr_magic == __swab32(IBLND_MSG_MAGIC) ||
2832 rej->ibr_magic == __swab32(LNET_PROTO_MAGIC)) {
2833 __swab32s(&rej->ibr_magic);
2834 __swab16s(&rej->ibr_version);
2838 if (priv_nob >= sizeof(struct kib_rej) &&
2839 rej->ibr_version > IBLND_MSG_VERSION_1) {
2840 /* priv_nob is always 148 in current version
2841 * of OFED, so we still need to check version.
2842 * (define of IB_CM_REJ_PRIVATE_DATA_SIZE) */
2846 __swab64s(&rej->ibr_incarnation);
2847 __swab16s(&cp->ibcp_queue_depth);
2848 __swab16s(&cp->ibcp_max_frags);
2849 __swab32s(&cp->ibcp_max_msg_size);
2852 incarnation = rej->ibr_incarnation;
2855 if (rej->ibr_magic != IBLND_MSG_MAGIC &&
2856 rej->ibr_magic != LNET_PROTO_MAGIC) {
2857 CERROR("%s rejected: consumer defined fatal error\n",
2858 libcfs_nid2str(peer_ni->ibp_nid));
2862 if (rej->ibr_version != IBLND_MSG_VERSION &&
2863 rej->ibr_version != IBLND_MSG_VERSION_1) {
2864 CERROR("%s rejected: o2iblnd version %x error\n",
2865 libcfs_nid2str(peer_ni->ibp_nid),
2870 if (rej->ibr_why == IBLND_REJECT_FATAL &&
2871 rej->ibr_version == IBLND_MSG_VERSION_1) {
2872 CDEBUG(D_NET, "rejected by old version peer_ni %s: %x\n",
2873 libcfs_nid2str(peer_ni->ibp_nid), rej->ibr_version);
2875 if (conn->ibc_version != IBLND_MSG_VERSION_1)
2876 rej->ibr_why = IBLND_REJECT_CONN_UNCOMPAT;
2879 switch (rej->ibr_why) {
2880 case IBLND_REJECT_CONN_RACE:
2881 case IBLND_REJECT_CONN_STALE:
2882 case IBLND_REJECT_CONN_UNCOMPAT:
2883 case IBLND_REJECT_MSG_QUEUE_SIZE:
2884 case IBLND_REJECT_RDMA_FRAGS:
2885 kiblnd_check_reconnect(conn, rej->ibr_version,
2886 incarnation, rej->ibr_why, cp);
2889 case IBLND_REJECT_NO_RESOURCES:
2890 CERROR("%s rejected: o2iblnd no resources\n",
2891 libcfs_nid2str(peer_ni->ibp_nid));
2894 case IBLND_REJECT_FATAL:
2895 CERROR("%s rejected: o2iblnd fatal error\n",
2896 libcfs_nid2str(peer_ni->ibp_nid));
2900 CERROR("%s rejected: o2iblnd reason %d\n",
2901 libcfs_nid2str(peer_ni->ibp_nid),
2909 CNETERR("%s rejected: reason %d, size %d\n",
2910 libcfs_nid2str(peer_ni->ibp_nid), reason, priv_nob);
2914 kiblnd_connreq_done(conn, -ECONNREFUSED);
2918 kiblnd_check_connreply(struct kib_conn *conn, void *priv, int priv_nob)
2920 struct kib_peer_ni *peer_ni = conn->ibc_peer;
2921 struct lnet_ni *ni = peer_ni->ibp_ni;
2922 struct kib_net *net = ni->ni_data;
2923 struct kib_msg *msg = priv;
2924 int ver = conn->ibc_version;
2925 int rc = kiblnd_unpack_msg(msg, priv_nob);
2926 unsigned long flags;
2928 LASSERT (net != NULL);
2931 CERROR("Can't unpack connack from %s: %d\n",
2932 libcfs_nid2str(peer_ni->ibp_nid), rc);
2936 if (msg->ibm_type != IBLND_MSG_CONNACK) {
2937 CERROR("Unexpected message %d from %s\n",
2938 msg->ibm_type, libcfs_nid2str(peer_ni->ibp_nid));
2943 if (ver != msg->ibm_version) {
2944 CERROR("%s replied version %x is different with "
2945 "requested version %x\n",
2946 libcfs_nid2str(peer_ni->ibp_nid), msg->ibm_version, ver);
2951 if (msg->ibm_u.connparams.ibcp_queue_depth >
2952 conn->ibc_queue_depth) {
2953 CERROR("%s has incompatible queue depth %d (<=%d wanted)\n",
2954 libcfs_nid2str(peer_ni->ibp_nid),
2955 msg->ibm_u.connparams.ibcp_queue_depth,
2956 conn->ibc_queue_depth);
2961 if (msg->ibm_u.connparams.ibcp_max_frags >
2962 conn->ibc_max_frags) {
2963 CERROR("%s has incompatible max_frags %d (<=%d wanted)\n",
2964 libcfs_nid2str(peer_ni->ibp_nid),
2965 msg->ibm_u.connparams.ibcp_max_frags,
2966 conn->ibc_max_frags);
2971 if (msg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2972 CERROR("%s max message size %d too big (%d max)\n",
2973 libcfs_nid2str(peer_ni->ibp_nid),
2974 msg->ibm_u.connparams.ibcp_max_msg_size,
2980 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2981 if (msg->ibm_dstnid == ni->ni_nid &&
2982 msg->ibm_dststamp == net->ibn_incarnation)
2986 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2989 CERROR("Bad connection reply from %s, rc = %d, "
2990 "version: %x max_frags: %d\n",
2991 libcfs_nid2str(peer_ni->ibp_nid), rc,
2992 msg->ibm_version, msg->ibm_u.connparams.ibcp_max_frags);
2996 conn->ibc_incarnation = msg->ibm_srcstamp;
2997 conn->ibc_credits = msg->ibm_u.connparams.ibcp_queue_depth;
2998 conn->ibc_reserved_credits = msg->ibm_u.connparams.ibcp_queue_depth;
2999 conn->ibc_queue_depth = msg->ibm_u.connparams.ibcp_queue_depth;
3000 conn->ibc_max_frags = msg->ibm_u.connparams.ibcp_max_frags;
3001 LASSERT(conn->ibc_credits + conn->ibc_reserved_credits +
3002 IBLND_OOB_MSGS(ver) <= IBLND_RX_MSGS(conn));
3004 kiblnd_connreq_done(conn, 0);
3008 /* NB My QP has already established itself, so I handle anything going
3009 * wrong here by setting ibc_comms_error.
3010 * kiblnd_connreq_done(0) moves the conn state to ESTABLISHED, but then
3011 * immediately tears it down. */
3014 conn->ibc_comms_error = rc;
3015 kiblnd_connreq_done(conn, 0);
3019 kiblnd_active_connect(struct rdma_cm_id *cmid)
3021 struct kib_peer_ni *peer_ni = cmid->context;
3022 struct kib_conn *conn;
3023 struct kib_msg *msg;
3024 struct rdma_conn_param cp;
3027 unsigned long flags;
3030 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3032 incarnation = peer_ni->ibp_incarnation;
3033 version = (peer_ni->ibp_version == 0) ? IBLND_MSG_VERSION :
3034 peer_ni->ibp_version;
3036 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3038 conn = kiblnd_create_conn(peer_ni, cmid, IBLND_CONN_ACTIVE_CONNECT,
3041 kiblnd_peer_connect_failed(peer_ni, 1, -ENOMEM);
3042 kiblnd_peer_decref(peer_ni); /* lose cmid's ref */
3046 /* conn "owns" cmid now, so I return success from here on to ensure the
3047 * CM callback doesn't destroy cmid. conn also takes over cmid's ref
3050 msg = &conn->ibc_connvars->cv_msg;
3052 memset(msg, 0, sizeof(*msg));
3053 kiblnd_init_msg(msg, IBLND_MSG_CONNREQ, sizeof(msg->ibm_u.connparams));
3054 msg->ibm_u.connparams.ibcp_queue_depth = conn->ibc_queue_depth;
3055 msg->ibm_u.connparams.ibcp_max_frags = conn->ibc_max_frags;
3056 msg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
3058 kiblnd_pack_msg(peer_ni->ibp_ni, msg, version,
3059 0, peer_ni->ibp_nid, incarnation);
3061 memset(&cp, 0, sizeof(cp));
3062 cp.private_data = msg;
3063 cp.private_data_len = msg->ibm_nob;
3064 cp.responder_resources = 0; /* No atomic ops or RDMA reads */
3065 cp.initiator_depth = 0;
3066 cp.flow_control = 1;
3067 cp.retry_count = *kiblnd_tunables.kib_retry_count;
3068 cp.rnr_retry_count = *kiblnd_tunables.kib_rnr_retry_count;
3070 LASSERT(cmid->context == (void *)conn);
3071 LASSERT(conn->ibc_cmid == cmid);
3073 rc = rdma_connect(cmid, &cp);
3075 CERROR("Can't connect to %s: %d\n",
3076 libcfs_nid2str(peer_ni->ibp_nid), rc);
3077 kiblnd_connreq_done(conn, rc);
3078 kiblnd_conn_decref(conn);
3085 kiblnd_cm_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
3087 struct kib_peer_ni *peer_ni;
3088 struct kib_conn *conn;
3091 switch (event->event) {
3093 CERROR("Unexpected event: %d, status: %d\n",
3094 event->event, event->status);
3097 case RDMA_CM_EVENT_CONNECT_REQUEST:
3098 /* destroy cmid on failure */
3099 rc = kiblnd_passive_connect(cmid,
3100 (void *)KIBLND_CONN_PARAM(event),
3101 KIBLND_CONN_PARAM_LEN(event));
3102 CDEBUG(D_NET, "connreq: %d\n", rc);
3105 case RDMA_CM_EVENT_ADDR_ERROR:
3106 peer_ni = cmid->context;
3107 CNETERR("%s: ADDR ERROR %d\n",
3108 libcfs_nid2str(peer_ni->ibp_nid), event->status);
3109 kiblnd_peer_connect_failed(peer_ni, 1, -EHOSTUNREACH);
3110 kiblnd_peer_decref(peer_ni);
3111 return -EHOSTUNREACH; /* rc != 0 destroys cmid */
3113 case RDMA_CM_EVENT_ADDR_RESOLVED:
3114 peer_ni = cmid->context;
3116 CDEBUG(D_NET,"%s Addr resolved: %d\n",
3117 libcfs_nid2str(peer_ni->ibp_nid), event->status);
3119 if (event->status != 0) {
3120 CNETERR("Can't resolve address for %s: %d\n",
3121 libcfs_nid2str(peer_ni->ibp_nid), event->status);
3124 rc = rdma_resolve_route(
3125 cmid, kiblnd_timeout() * 1000);
3127 struct kib_net *net = peer_ni->ibp_ni->ni_data;
3128 struct kib_dev *dev = net->ibn_dev;
3130 CDEBUG(D_NET, "%s: connection bound to "\
3132 libcfs_nid2str(peer_ni->ibp_nid),
3134 &dev->ibd_ifip, cmid->device->name);
3139 /* Can't initiate route resolution */
3140 CERROR("Can't resolve route for %s: %d\n",
3141 libcfs_nid2str(peer_ni->ibp_nid), rc);
3143 kiblnd_peer_connect_failed(peer_ni, 1, rc);
3144 kiblnd_peer_decref(peer_ni);
3145 return rc; /* rc != 0 destroys cmid */
3147 case RDMA_CM_EVENT_ROUTE_ERROR:
3148 peer_ni = cmid->context;
3149 CNETERR("%s: ROUTE ERROR %d\n",
3150 libcfs_nid2str(peer_ni->ibp_nid), event->status);
3151 kiblnd_peer_connect_failed(peer_ni, 1, -EHOSTUNREACH);
3152 kiblnd_peer_decref(peer_ni);
3153 return -EHOSTUNREACH; /* rc != 0 destroys cmid */
3155 case RDMA_CM_EVENT_ROUTE_RESOLVED:
3156 peer_ni = cmid->context;
3157 CDEBUG(D_NET,"%s Route resolved: %d\n",
3158 libcfs_nid2str(peer_ni->ibp_nid), event->status);
3160 if (event->status == 0)
3161 return kiblnd_active_connect(cmid);
3163 CNETERR("Can't resolve route for %s: %d\n",
3164 libcfs_nid2str(peer_ni->ibp_nid), event->status);
3165 kiblnd_peer_connect_failed(peer_ni, 1, event->status);
3166 kiblnd_peer_decref(peer_ni);
3167 return event->status; /* rc != 0 destroys cmid */
3169 case RDMA_CM_EVENT_UNREACHABLE:
3170 conn = cmid->context;
3171 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
3172 conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
3173 CNETERR("%s: UNREACHABLE %d\n",
3174 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
3175 kiblnd_connreq_done(conn, -ENETDOWN);
3176 kiblnd_conn_decref(conn);
3179 case RDMA_CM_EVENT_CONNECT_ERROR:
3180 conn = cmid->context;
3181 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
3182 conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
3183 CNETERR("%s: CONNECT ERROR %d\n",
3184 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
3185 kiblnd_connreq_done(conn, -ENOTCONN);
3186 kiblnd_conn_decref(conn);
3189 case RDMA_CM_EVENT_REJECTED:
3190 conn = cmid->context;
3191 switch (conn->ibc_state) {
3195 case IBLND_CONN_PASSIVE_WAIT:
3196 CERROR ("%s: REJECTED %d\n",
3197 libcfs_nid2str(conn->ibc_peer->ibp_nid),
3199 kiblnd_connreq_done(conn, -ECONNRESET);
3202 case IBLND_CONN_ACTIVE_CONNECT:
3203 kiblnd_rejected(conn, event->status,
3204 (void *)KIBLND_CONN_PARAM(event),
3205 KIBLND_CONN_PARAM_LEN(event));
3208 kiblnd_conn_decref(conn);
3211 case RDMA_CM_EVENT_ESTABLISHED:
3212 conn = cmid->context;
3213 switch (conn->ibc_state) {
3217 case IBLND_CONN_PASSIVE_WAIT:
3218 CDEBUG(D_NET, "ESTABLISHED (passive): %s\n",
3219 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3220 kiblnd_connreq_done(conn, 0);
3223 case IBLND_CONN_ACTIVE_CONNECT:
3224 CDEBUG(D_NET, "ESTABLISHED(active): %s\n",
3225 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3226 kiblnd_check_connreply(conn,
3227 (void *)KIBLND_CONN_PARAM(event),
3228 KIBLND_CONN_PARAM_LEN(event));
3231 /* net keeps its ref on conn! */
3234 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
3235 CDEBUG(D_NET, "Ignore TIMEWAIT_EXIT event\n");
3238 case RDMA_CM_EVENT_DISCONNECTED:
3239 conn = cmid->context;
3240 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
3241 CERROR("%s DISCONNECTED\n",
3242 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3243 kiblnd_connreq_done(conn, -ECONNRESET);
3245 kiblnd_close_conn(conn, 0);
3247 kiblnd_conn_decref(conn);
3248 cmid->context = NULL;
3251 case RDMA_CM_EVENT_DEVICE_REMOVAL:
3252 LCONSOLE_ERROR_MSG(0x131,
3253 "Received notification of device removal\n"
3254 "Please shutdown LNET to allow this to proceed\n");
3255 /* Can't remove network from underneath LNET for now, so I have
3259 case RDMA_CM_EVENT_ADDR_CHANGE:
3260 LCONSOLE_INFO("Physical link changed (eg hca/port)\n");
3266 kiblnd_check_txs_locked(struct kib_conn *conn, struct list_head *txs)
3269 struct list_head *ttmp;
3271 list_for_each(ttmp, txs) {
3272 tx = list_entry(ttmp, struct kib_tx, tx_list);
3274 if (txs != &conn->ibc_active_txs) {
3275 LASSERT(tx->tx_queued);
3277 LASSERT(!tx->tx_queued);
3278 LASSERT(tx->tx_waiting || tx->tx_sending != 0);
3281 if (ktime_compare(ktime_get(), tx->tx_deadline) >= 0) {
3282 CERROR("Timed out tx: %s, %lld seconds\n",
3283 kiblnd_queue2str(conn, txs),
3285 ktime_ms_delta(ktime_get(),
3286 tx->tx_deadline) / MSEC_PER_SEC);
3295 kiblnd_conn_timed_out_locked(struct kib_conn *conn)
3297 return kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue) ||
3298 kiblnd_check_txs_locked(conn, &conn->ibc_tx_noops) ||
3299 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_rsrvd) ||
3300 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_nocred) ||
3301 kiblnd_check_txs_locked(conn, &conn->ibc_active_txs);
3305 kiblnd_check_conns (int idx)
3308 LIST_HEAD(checksends);
3309 LIST_HEAD(timedout_txs);
3310 struct list_head *peers = &kiblnd_data.kib_peers[idx];
3311 struct list_head *ptmp;
3312 struct kib_peer_ni *peer_ni;
3313 struct kib_conn *conn;
3314 struct kib_tx *tx, *tx_tmp;
3315 struct list_head *ctmp;
3316 unsigned long flags;
3318 /* NB. We expect to have a look at all the peers and not find any
3319 * RDMAs to time out, so we just use a shared lock while we
3321 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3323 list_for_each(ptmp, peers) {
3324 peer_ni = list_entry(ptmp, struct kib_peer_ni, ibp_list);
3326 /* Check tx_deadline */
3327 list_for_each_entry_safe(tx, tx_tmp, &peer_ni->ibp_tx_queue, tx_list) {
3328 if (ktime_compare(ktime_get(), tx->tx_deadline) >= 0) {
3329 CWARN("Timed out tx for %s: %lld seconds\n",
3330 libcfs_nid2str(peer_ni->ibp_nid),
3331 ktime_ms_delta(ktime_get(),
3332 tx->tx_deadline) / MSEC_PER_SEC);
3333 list_move(&tx->tx_list, &timedout_txs);
3337 list_for_each(ctmp, &peer_ni->ibp_conns) {
3341 conn = list_entry(ctmp, struct kib_conn, ibc_list);
3343 LASSERT(conn->ibc_state == IBLND_CONN_ESTABLISHED);
3345 spin_lock(&conn->ibc_lock);
3347 sendnoop = kiblnd_need_noop(conn);
3348 timedout = kiblnd_conn_timed_out_locked(conn);
3349 if (!sendnoop && !timedout) {
3350 spin_unlock(&conn->ibc_lock);
3355 CERROR("Timed out RDMA with %s (%lld): "
3356 "c: %u, oc: %u, rc: %u\n",
3357 libcfs_nid2str(peer_ni->ibp_nid),
3358 ktime_get_seconds() - peer_ni->ibp_last_alive,
3360 conn->ibc_outstanding_credits,
3361 conn->ibc_reserved_credits);
3362 list_add(&conn->ibc_connd_list, &closes);
3364 list_add(&conn->ibc_connd_list, &checksends);
3366 /* +ref for 'closes' or 'checksends' */
3367 kiblnd_conn_addref(conn);
3369 spin_unlock(&conn->ibc_lock);
3373 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3375 if (!list_empty(&timedout_txs))
3376 kiblnd_txlist_done(&timedout_txs, -ETIMEDOUT,
3377 LNET_MSG_STATUS_LOCAL_TIMEOUT);
3379 /* Handle timeout by closing the whole
3380 * connection. We can only be sure RDMA activity
3381 * has ceased once the QP has been modified. */
3382 while (!list_empty(&closes)) {
3383 conn = list_entry(closes.next,
3384 struct kib_conn, ibc_connd_list);
3385 list_del(&conn->ibc_connd_list);
3386 kiblnd_close_conn(conn, -ETIMEDOUT);
3387 kiblnd_conn_decref(conn);
3390 /* In case we have enough credits to return via a
3391 * NOOP, but there were no non-blocking tx descs
3392 * free to do it last time... */
3393 while (!list_empty(&checksends)) {
3394 conn = list_entry(checksends.next,
3395 struct kib_conn, ibc_connd_list);
3396 list_del(&conn->ibc_connd_list);
3398 spin_lock(&conn->ibc_lock);
3399 kiblnd_check_sends_locked(conn);
3400 spin_unlock(&conn->ibc_lock);
3402 kiblnd_conn_decref(conn);
3407 kiblnd_disconnect_conn(struct kib_conn *conn)
3409 LASSERT (!in_interrupt());
3410 LASSERT (current == kiblnd_data.kib_connd);
3411 LASSERT (conn->ibc_state == IBLND_CONN_CLOSING);
3413 rdma_disconnect(conn->ibc_cmid);
3414 kiblnd_finalise_conn(conn);
3416 kiblnd_peer_notify(conn->ibc_peer);
3420 * High-water for reconnection to the same peer_ni, reconnection attempt should
3421 * be delayed after trying more than KIB_RECONN_HIGH_RACE.
3423 #define KIB_RECONN_HIGH_RACE 10
3425 * Allow connd to take a break and handle other things after consecutive
3426 * reconnection attemps.
3428 #define KIB_RECONN_BREAK 100
3431 kiblnd_connd (void *arg)
3433 spinlock_t *lock= &kiblnd_data.kib_connd_lock;
3434 wait_queue_entry_t wait;
3435 unsigned long flags;
3436 struct kib_conn *conn;
3441 unsigned long deadline = jiffies;
3443 init_waitqueue_entry(&wait, current);
3444 kiblnd_data.kib_connd = current;
3446 spin_lock_irqsave(lock, flags);
3448 while (!kiblnd_data.kib_shutdown) {
3453 if (!list_empty(&kiblnd_data.kib_connd_zombies)) {
3454 struct kib_peer_ni *peer_ni = NULL;
3456 conn = list_entry(kiblnd_data.kib_connd_zombies.next,
3457 struct kib_conn, ibc_list);
3458 list_del(&conn->ibc_list);
3459 if (conn->ibc_reconnect) {
3460 peer_ni = conn->ibc_peer;
3461 kiblnd_peer_addref(peer_ni);
3464 spin_unlock_irqrestore(lock, flags);
3467 kiblnd_destroy_conn(conn);
3469 spin_lock_irqsave(lock, flags);
3471 LIBCFS_FREE(conn, sizeof(*conn));
3475 conn->ibc_peer = peer_ni;
3476 if (peer_ni->ibp_reconnected < KIB_RECONN_HIGH_RACE)
3477 list_add_tail(&conn->ibc_list,
3478 &kiblnd_data.kib_reconn_list);
3480 list_add_tail(&conn->ibc_list,
3481 &kiblnd_data.kib_reconn_wait);
3484 if (!list_empty(&kiblnd_data.kib_connd_conns)) {
3485 conn = list_entry(kiblnd_data.kib_connd_conns.next,
3486 struct kib_conn, ibc_list);
3487 list_del(&conn->ibc_list);
3489 spin_unlock_irqrestore(lock, flags);
3492 kiblnd_disconnect_conn(conn);
3493 kiblnd_conn_decref(conn);
3495 spin_lock_irqsave(lock, flags);
3498 while (reconn < KIB_RECONN_BREAK) {
3499 if (kiblnd_data.kib_reconn_sec !=
3500 ktime_get_real_seconds()) {
3501 kiblnd_data.kib_reconn_sec = ktime_get_real_seconds();
3502 list_splice_init(&kiblnd_data.kib_reconn_wait,
3503 &kiblnd_data.kib_reconn_list);
3506 if (list_empty(&kiblnd_data.kib_reconn_list))
3509 conn = list_entry(kiblnd_data.kib_reconn_list.next,
3510 struct kib_conn, ibc_list);
3511 list_del(&conn->ibc_list);
3513 spin_unlock_irqrestore(lock, flags);
3516 reconn += kiblnd_reconnect_peer(conn->ibc_peer);
3517 kiblnd_peer_decref(conn->ibc_peer);
3518 LIBCFS_FREE(conn, sizeof(*conn));
3520 spin_lock_irqsave(lock, flags);
3523 /* careful with the jiffy wrap... */
3524 timeout = (int)(deadline - jiffies);
3528 int chunk = kiblnd_data.kib_peer_hash_size;
3529 unsigned int lnd_timeout;
3531 spin_unlock_irqrestore(lock, flags);
3534 /* Time to check for RDMA timeouts on a few more
3535 * peers: I do checks every 'p' seconds on a
3536 * proportion of the peer_ni table and I need to check
3537 * every connection 'n' times within a timeout
3538 * interval, to ensure I detect a timeout on any
3539 * connection within (n+1)/n times the timeout
3542 lnd_timeout = kiblnd_timeout();
3543 if (lnd_timeout > n * p)
3544 chunk = (chunk * n * p) / lnd_timeout;
3548 for (i = 0; i < chunk; i++) {
3549 kiblnd_check_conns(peer_index);
3550 peer_index = (peer_index + 1) %
3551 kiblnd_data.kib_peer_hash_size;
3554 deadline += cfs_time_seconds(p);
3555 spin_lock_irqsave(lock, flags);
3561 /* Nothing to do for 'timeout' */
3562 set_current_state(TASK_INTERRUPTIBLE);
3563 add_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3564 spin_unlock_irqrestore(lock, flags);
3566 schedule_timeout(timeout);
3568 remove_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3569 spin_lock_irqsave(lock, flags);
3572 spin_unlock_irqrestore(lock, flags);
3574 kiblnd_thread_fini();
3579 kiblnd_qp_event(struct ib_event *event, void *arg)
3581 struct kib_conn *conn = arg;
3583 switch (event->event) {
3584 case IB_EVENT_COMM_EST:
3585 CDEBUG(D_NET, "%s established\n",
3586 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3587 /* We received a packet but connection isn't established
3588 * probably handshake packet was lost, so free to
3589 * force make connection established */
3590 rdma_notify(conn->ibc_cmid, IB_EVENT_COMM_EST);
3593 case IB_EVENT_PORT_ERR:
3594 case IB_EVENT_DEVICE_FATAL:
3595 CERROR("Fatal device error for NI %s\n",
3596 libcfs_nid2str(conn->ibc_peer->ibp_ni->ni_nid));
3597 atomic_set(&conn->ibc_peer->ibp_ni->ni_fatal_error_on, 1);
3600 case IB_EVENT_PORT_ACTIVE:
3601 CERROR("Port reactivated for NI %s\n",
3602 libcfs_nid2str(conn->ibc_peer->ibp_ni->ni_nid));
3603 atomic_set(&conn->ibc_peer->ibp_ni->ni_fatal_error_on, 0);
3607 CERROR("%s: Async QP event type %d\n",
3608 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3614 kiblnd_complete (struct ib_wc *wc)
3616 switch (kiblnd_wreqid2type(wc->wr_id)) {
3621 if (wc->status != IB_WC_SUCCESS &&
3622 wc->status != IB_WC_WR_FLUSH_ERR)
3623 CNETERR("FastReg failed: %d\n", wc->status);
3626 case IBLND_WID_RDMA:
3627 /* We only get RDMA completion notification if it fails. All
3628 * subsequent work items, including the final SEND will fail
3629 * too. However we can't print out any more info about the
3630 * failing RDMA because 'tx' might be back on the idle list or
3631 * even reused already if we didn't manage to post all our work
3633 CNETERR("RDMA (tx: %p) failed: %d\n",
3634 kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3638 kiblnd_tx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3642 kiblnd_rx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status,
3649 kiblnd_cq_completion(struct ib_cq *cq, void *arg)
3651 /* NB I'm not allowed to schedule this conn once its refcount has
3652 * reached 0. Since fundamentally I'm racing with scheduler threads
3653 * consuming my CQ I could be called after all completions have
3654 * occurred. But in this case, ibc_nrx == 0 && ibc_nsends_posted == 0
3655 * and this CQ is about to be destroyed so I NOOP. */
3656 struct kib_conn *conn = arg;
3657 struct kib_sched_info *sched = conn->ibc_sched;
3658 unsigned long flags;
3660 LASSERT(cq == conn->ibc_cq);
3662 spin_lock_irqsave(&sched->ibs_lock, flags);
3664 conn->ibc_ready = 1;
3666 if (!conn->ibc_scheduled &&
3667 (conn->ibc_nrx > 0 ||
3668 conn->ibc_nsends_posted > 0)) {
3669 kiblnd_conn_addref(conn); /* +1 ref for sched_conns */
3670 conn->ibc_scheduled = 1;
3671 list_add_tail(&conn->ibc_sched_list, &sched->ibs_conns);
3673 if (waitqueue_active(&sched->ibs_waitq))
3674 wake_up(&sched->ibs_waitq);
3677 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3681 kiblnd_cq_event(struct ib_event *event, void *arg)
3683 struct kib_conn *conn = arg;
3685 CERROR("%s: async CQ event type %d\n",
3686 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3690 kiblnd_scheduler(void *arg)
3692 long id = (long)arg;
3693 struct kib_sched_info *sched;
3694 struct kib_conn *conn;
3695 wait_queue_entry_t wait;
3696 unsigned long flags;
3701 init_waitqueue_entry(&wait, current);
3703 sched = kiblnd_data.kib_scheds[KIB_THREAD_CPT(id)];
3705 rc = cfs_cpt_bind(lnet_cpt_table(), sched->ibs_cpt);
3707 CWARN("Unable to bind on CPU partition %d, please verify "
3708 "whether all CPUs are healthy and reload modules if "
3709 "necessary, otherwise your system might under risk of "
3710 "low performance\n", sched->ibs_cpt);
3713 spin_lock_irqsave(&sched->ibs_lock, flags);
3715 while (!kiblnd_data.kib_shutdown) {
3716 if (need_resched()) {
3717 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3721 spin_lock_irqsave(&sched->ibs_lock, flags);
3726 if (!list_empty(&sched->ibs_conns)) {
3727 conn = list_entry(sched->ibs_conns.next,
3728 struct kib_conn, ibc_sched_list);
3729 /* take over kib_sched_conns' ref on conn... */
3730 LASSERT(conn->ibc_scheduled);
3731 list_del(&conn->ibc_sched_list);
3732 conn->ibc_ready = 0;
3734 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3736 wc.wr_id = IBLND_WID_INVAL;
3738 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3740 rc = ib_req_notify_cq(conn->ibc_cq,
3743 CWARN("%s: ib_req_notify_cq failed: %d, "
3744 "closing connection\n",
3745 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
3746 kiblnd_close_conn(conn, -EIO);
3747 kiblnd_conn_decref(conn);
3748 spin_lock_irqsave(&sched->ibs_lock,
3753 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3756 if (unlikely(rc > 0 && wc.wr_id == IBLND_WID_INVAL)) {
3758 "ib_poll_cq (rc: %d) returned invalid "
3759 "wr_id, opcode %d, status: %d, "
3760 "vendor_err: %d, conn: %s status: %d\n"
3761 "please upgrade firmware and OFED or "
3762 "contact vendor.\n", rc,
3763 wc.opcode, wc.status, wc.vendor_err,
3764 libcfs_nid2str(conn->ibc_peer->ibp_nid),
3770 CWARN("%s: ib_poll_cq failed: %d, "
3771 "closing connection\n",
3772 libcfs_nid2str(conn->ibc_peer->ibp_nid),
3774 kiblnd_close_conn(conn, -EIO);
3775 kiblnd_conn_decref(conn);
3776 spin_lock_irqsave(&sched->ibs_lock, flags);
3780 spin_lock_irqsave(&sched->ibs_lock, flags);
3782 if (rc != 0 || conn->ibc_ready) {
3783 /* There may be another completion waiting; get
3784 * another scheduler to check while I handle
3786 /* +1 ref for sched_conns */
3787 kiblnd_conn_addref(conn);
3788 list_add_tail(&conn->ibc_sched_list,
3790 if (waitqueue_active(&sched->ibs_waitq))
3791 wake_up(&sched->ibs_waitq);
3793 conn->ibc_scheduled = 0;
3797 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3798 kiblnd_complete(&wc);
3800 spin_lock_irqsave(&sched->ibs_lock, flags);
3803 kiblnd_conn_decref(conn); /* ...drop my ref from above */
3810 set_current_state(TASK_INTERRUPTIBLE);
3811 add_wait_queue_exclusive(&sched->ibs_waitq, &wait);
3812 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3816 remove_wait_queue(&sched->ibs_waitq, &wait);
3817 set_current_state(TASK_RUNNING);
3818 spin_lock_irqsave(&sched->ibs_lock, flags);
3821 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3823 kiblnd_thread_fini();
3828 kiblnd_failover_thread(void *arg)
3830 rwlock_t *glock = &kiblnd_data.kib_global_lock;
3831 struct kib_dev *dev;
3832 struct net *ns = arg;
3833 wait_queue_entry_t wait;
3834 unsigned long flags;
3837 LASSERT(*kiblnd_tunables.kib_dev_failover != 0);
3839 init_waitqueue_entry(&wait, current);
3840 write_lock_irqsave(glock, flags);
3842 while (!kiblnd_data.kib_shutdown) {
3843 int do_failover = 0;
3846 list_for_each_entry(dev, &kiblnd_data.kib_failed_devs,
3848 if (ktime_get_seconds() < dev->ibd_next_failover)
3855 list_del_init(&dev->ibd_fail_list);
3856 dev->ibd_failover = 1;
3857 write_unlock_irqrestore(glock, flags);
3859 rc = kiblnd_dev_failover(dev, ns);
3861 write_lock_irqsave(glock, flags);
3863 LASSERT (dev->ibd_failover);
3864 dev->ibd_failover = 0;
3865 if (rc >= 0) { /* Device is OK or failover succeed */
3866 dev->ibd_next_failover = ktime_get_seconds() + 3;
3870 /* failed to failover, retry later */
3871 dev->ibd_next_failover = ktime_get_seconds() +
3872 min(dev->ibd_failed_failover, 10);
3873 if (kiblnd_dev_can_failover(dev)) {
3874 list_add_tail(&dev->ibd_fail_list,
3875 &kiblnd_data.kib_failed_devs);
3881 /* long sleep if no more pending failover */
3882 long_sleep = list_empty(&kiblnd_data.kib_failed_devs);
3884 set_current_state(TASK_INTERRUPTIBLE);
3885 add_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3886 write_unlock_irqrestore(glock, flags);
3888 rc = schedule_timeout(long_sleep ? cfs_time_seconds(10) :
3889 cfs_time_seconds(1));
3890 set_current_state(TASK_RUNNING);
3891 remove_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3892 write_lock_irqsave(glock, flags);
3894 if (!long_sleep || rc != 0)
3897 /* have a long sleep, routine check all active devices,
3898 * we need checking like this because if there is not active
3899 * connection on the dev and no SEND from local, we may listen
3900 * on wrong HCA for ever while there is a bonding failover */
3901 list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3902 if (kiblnd_dev_can_failover(dev)) {
3903 list_add_tail(&dev->ibd_fail_list,
3904 &kiblnd_data.kib_failed_devs);
3909 write_unlock_irqrestore(glock, flags);
3911 kiblnd_thread_fini();