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/
31 * lnet/klnds/o2iblnd/o2iblnd_cb.c
33 * Author: Eric Barton <eric@bartonsoftware.com>
38 #define MAX_CONN_RACES_BEFORE_ABORT 20
40 static void kiblnd_peer_alive(struct kib_peer_ni *peer_ni);
41 static void kiblnd_peer_connect_failed(struct kib_peer_ni *peer_ni, int active,
43 static struct ib_rdma_wr *
44 kiblnd_init_tx_msg_payload(struct lnet_ni *ni, struct kib_tx *tx,
45 int type, int body_nob, int payload_nob);
46 #define kiblnd_init_tx_msg(ni, tx, type, body) \
47 kiblnd_init_tx_msg_payload(ni, tx, type, body, 0)
48 static int kiblnd_init_rdma(struct kib_conn *conn, struct kib_tx *tx, int type,
49 int resid, struct kib_rdma_desc *dstrd, u64 dstcookie);
50 static void kiblnd_queue_tx_locked(struct kib_tx *tx, struct kib_conn *conn);
51 static void kiblnd_queue_tx(struct kib_tx *tx, struct kib_conn *conn);
53 static void kiblnd_unmap_tx(struct kib_tx *tx);
54 static void kiblnd_check_sends_locked(struct kib_conn *conn);
57 kiblnd_tx_done(struct kib_tx *tx)
59 struct lnet_msg *lntmsg[2];
63 LASSERT (!in_interrupt());
64 LASSERT (!tx->tx_queued); /* mustn't be queued for sending */
65 LASSERT (tx->tx_sending == 0); /* mustn't be awaiting sent callback */
66 LASSERT (!tx->tx_waiting); /* mustn't be awaiting peer_ni response */
67 LASSERT (tx->tx_pool != NULL);
71 /* tx may have up to 2 lnet msgs to finalise */
72 lntmsg[0] = tx->tx_lntmsg[0]; tx->tx_lntmsg[0] = NULL;
73 lntmsg[1] = tx->tx_lntmsg[1]; tx->tx_lntmsg[1] = NULL;
76 if (tx->tx_conn != NULL) {
77 kiblnd_conn_decref(tx->tx_conn);
81 tx->tx_nwrq = tx->tx_nsge = 0;
84 kiblnd_pool_free_node(&tx->tx_pool->tpo_pool, &tx->tx_list);
86 /* delay finalize until my descs have been freed */
87 for (i = 0; i < 2; i++) {
88 if (lntmsg[i] == NULL)
91 /* propagate health status to LNet for requests */
92 if (i == 0 && lntmsg[i])
93 lntmsg[i]->msg_health_status = tx->tx_hstatus;
95 lnet_finalize(lntmsg[i], rc);
100 kiblnd_txlist_done(struct list_head *txlist, int status,
101 enum lnet_msg_hstatus hstatus)
105 while ((tx = list_first_entry_or_null(txlist,
108 list_del(&tx->tx_list);
111 tx->tx_status = status;
112 if (hstatus != LNET_MSG_STATUS_OK)
113 tx->tx_hstatus = hstatus;
118 static struct kib_tx *
119 kiblnd_get_idle_tx(struct lnet_ni *ni, lnet_nid_t target)
121 struct kib_net *net = ni->ni_data;
122 struct list_head *node;
124 struct kib_tx_poolset *tps;
126 tps = net->ibn_tx_ps[lnet_cpt_of_nid(target, ni)];
127 node = kiblnd_pool_alloc_node(&tps->tps_poolset);
130 tx = container_of(node, struct kib_tx, tx_list);
132 LASSERT (tx->tx_nwrq == 0);
133 LASSERT (!tx->tx_queued);
134 LASSERT (tx->tx_sending == 0);
135 LASSERT (!tx->tx_waiting);
136 LASSERT (tx->tx_status == 0);
137 LASSERT (tx->tx_conn == NULL);
138 LASSERT (tx->tx_lntmsg[0] == NULL);
139 LASSERT (tx->tx_lntmsg[1] == NULL);
140 LASSERT (tx->tx_nfrags == 0);
143 tx->tx_hstatus = LNET_MSG_STATUS_OK;
149 kiblnd_drop_rx(struct kib_rx *rx)
151 struct kib_conn *conn = rx->rx_conn;
152 struct kib_sched_info *sched = conn->ibc_sched;
155 spin_lock_irqsave(&sched->ibs_lock, flags);
156 LASSERT(conn->ibc_nrx > 0);
158 spin_unlock_irqrestore(&sched->ibs_lock, flags);
160 kiblnd_conn_decref(conn);
164 kiblnd_post_rx(struct kib_rx *rx, int credit)
166 struct kib_conn *conn = rx->rx_conn;
167 struct kib_net *net = conn->ibc_peer->ibp_ni->ni_data;
168 struct ib_recv_wr *bad_wrq = NULL;
169 #ifdef HAVE_IB_GET_DMA_MR
170 struct ib_mr *mr = conn->ibc_hdev->ibh_mrs;
174 LASSERT (net != NULL);
175 LASSERT (!in_interrupt());
176 LASSERT (credit == IBLND_POSTRX_NO_CREDIT ||
177 credit == IBLND_POSTRX_PEER_CREDIT ||
178 credit == IBLND_POSTRX_RSRVD_CREDIT);
179 #ifdef HAVE_IB_GET_DMA_MR
182 rx->rx_sge.lkey = mr->lkey;
184 rx->rx_sge.lkey = conn->ibc_hdev->ibh_pd->local_dma_lkey;
186 rx->rx_sge.addr = rx->rx_msgaddr;
187 rx->rx_sge.length = IBLND_MSG_SIZE;
189 rx->rx_wrq.next = NULL;
190 rx->rx_wrq.sg_list = &rx->rx_sge;
191 rx->rx_wrq.num_sge = 1;
192 rx->rx_wrq.wr_id = kiblnd_ptr2wreqid(rx, IBLND_WID_RX);
194 LASSERT (conn->ibc_state >= IBLND_CONN_INIT);
195 LASSERT (rx->rx_nob >= 0); /* not posted */
197 if (conn->ibc_state > IBLND_CONN_ESTABLISHED) {
198 kiblnd_drop_rx(rx); /* No more posts for this rx */
202 rx->rx_nob = -1; /* flag posted */
204 /* NB: need an extra reference after ib_post_recv because we don't
205 * own this rx (and rx::rx_conn) anymore, LU-5678.
207 kiblnd_conn_addref(conn);
208 #ifdef HAVE_IB_POST_SEND_RECV_CONST
209 rc = ib_post_recv(conn->ibc_cmid->qp, &rx->rx_wrq,
210 (const struct ib_recv_wr **)&bad_wrq);
212 rc = ib_post_recv(conn->ibc_cmid->qp, &rx->rx_wrq, &bad_wrq);
214 if (unlikely(rc != 0)) {
215 CERROR("Can't post rx for %s: %d, bad_wrq: %p\n",
216 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc, bad_wrq);
220 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) /* Initial post */
223 if (unlikely(rc != 0)) {
224 kiblnd_close_conn(conn, rc);
225 kiblnd_drop_rx(rx); /* No more posts for this rx */
229 if (credit == IBLND_POSTRX_NO_CREDIT)
232 spin_lock(&conn->ibc_lock);
233 if (credit == IBLND_POSTRX_PEER_CREDIT)
234 conn->ibc_outstanding_credits++;
236 conn->ibc_reserved_credits++;
237 kiblnd_check_sends_locked(conn);
238 spin_unlock(&conn->ibc_lock);
241 kiblnd_conn_decref(conn);
245 static struct kib_tx *
246 kiblnd_find_waiting_tx_locked(struct kib_conn *conn, int txtype, u64 cookie)
250 list_for_each_entry(tx, &conn->ibc_active_txs, 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 struct lnet_nid srcnid;
341 LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
343 CDEBUG (D_NET, "Received %x[%d] from %s\n",
344 msg->ibm_type, credits,
345 libcfs_nid2str(conn->ibc_peer->ibp_nid));
348 /* Have I received credits that will let me send? */
349 spin_lock(&conn->ibc_lock);
351 if (conn->ibc_credits + credits >
352 conn->ibc_queue_depth) {
353 rc2 = conn->ibc_credits;
354 spin_unlock(&conn->ibc_lock);
356 CERROR("Bad credits from %s: %d + %d > %d\n",
357 libcfs_nid2str(conn->ibc_peer->ibp_nid),
359 conn->ibc_queue_depth);
361 kiblnd_close_conn(conn, -EPROTO);
362 kiblnd_post_rx(rx, IBLND_POSTRX_NO_CREDIT);
366 conn->ibc_credits += credits;
368 /* This ensures the credit taken by NOOP can be returned */
369 if (msg->ibm_type == IBLND_MSG_NOOP &&
370 !IBLND_OOB_CAPABLE(conn->ibc_version)) /* v1 only */
371 conn->ibc_outstanding_credits++;
373 kiblnd_check_sends_locked(conn);
374 spin_unlock(&conn->ibc_lock);
377 switch (msg->ibm_type) {
379 CERROR("Bad IBLND message type %x from %s\n",
380 msg->ibm_type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
381 post_credit = IBLND_POSTRX_NO_CREDIT;
386 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
387 post_credit = IBLND_POSTRX_NO_CREDIT;
391 if (credits != 0) /* credit already posted */
392 post_credit = IBLND_POSTRX_NO_CREDIT;
393 else /* a keepalive NOOP */
394 post_credit = IBLND_POSTRX_PEER_CREDIT;
397 case IBLND_MSG_IMMEDIATE:
398 post_credit = IBLND_POSTRX_DONT_POST;
399 lnet_hdr_from_nid4(&hdr, &msg->ibm_u.immediate.ibim_hdr);
400 lnet_nid4_to_nid(msg->ibm_srcnid, &srcnid);
401 rc = lnet_parse(ni, &hdr, &srcnid, rx, 0);
402 if (rc < 0) /* repost on error */
403 post_credit = IBLND_POSTRX_PEER_CREDIT;
406 case IBLND_MSG_PUT_REQ:
407 post_credit = IBLND_POSTRX_DONT_POST;
408 lnet_hdr_from_nid4(&hdr, &msg->ibm_u.putreq.ibprm_hdr);
409 lnet_nid4_to_nid(msg->ibm_srcnid, &srcnid);
410 rc = lnet_parse(ni, &hdr, &srcnid, rx, 1);
411 if (rc < 0) /* repost on error */
412 post_credit = IBLND_POSTRX_PEER_CREDIT;
415 case IBLND_MSG_PUT_NAK:
416 CWARN ("PUT_NACK from %s\n",
417 libcfs_nid2str(conn->ibc_peer->ibp_nid));
418 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
419 kiblnd_handle_completion(conn, IBLND_MSG_PUT_REQ,
420 msg->ibm_u.completion.ibcm_status,
421 msg->ibm_u.completion.ibcm_cookie);
424 case IBLND_MSG_PUT_ACK:
425 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
427 spin_lock(&conn->ibc_lock);
428 tx = kiblnd_find_waiting_tx_locked(conn, IBLND_MSG_PUT_REQ,
429 msg->ibm_u.putack.ibpam_src_cookie);
431 list_del(&tx->tx_list);
432 spin_unlock(&conn->ibc_lock);
435 CERROR("Unmatched PUT_ACK from %s\n",
436 libcfs_nid2str(conn->ibc_peer->ibp_nid));
441 LASSERT (tx->tx_waiting);
442 /* CAVEAT EMPTOR: I could be racing with tx_complete, but...
443 * (a) I can overwrite tx_msg since my peer_ni has received it!
444 * (b) tx_waiting set tells tx_complete() it's not done. */
446 tx->tx_nwrq = tx->tx_nsge = 0; /* overwrite PUT_REQ */
448 rc2 = kiblnd_init_rdma(conn, tx, IBLND_MSG_PUT_DONE,
449 kiblnd_rd_size(&msg->ibm_u.putack.ibpam_rd),
450 &msg->ibm_u.putack.ibpam_rd,
451 msg->ibm_u.putack.ibpam_dst_cookie);
453 CERROR("Can't setup rdma for PUT to %s: %d\n",
454 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc2);
456 spin_lock(&conn->ibc_lock);
457 tx->tx_waiting = 0; /* clear waiting and queue atomically */
458 kiblnd_queue_tx_locked(tx, conn);
459 spin_unlock(&conn->ibc_lock);
462 case IBLND_MSG_PUT_DONE:
463 post_credit = IBLND_POSTRX_PEER_CREDIT;
464 kiblnd_handle_completion(conn, IBLND_MSG_PUT_ACK,
465 msg->ibm_u.completion.ibcm_status,
466 msg->ibm_u.completion.ibcm_cookie);
469 case IBLND_MSG_GET_REQ:
470 post_credit = IBLND_POSTRX_DONT_POST;
471 lnet_hdr_from_nid4(&hdr, &msg->ibm_u.get.ibgm_hdr);
472 lnet_nid4_to_nid(msg->ibm_srcnid, &srcnid);
473 rc = lnet_parse(ni, &hdr, &srcnid, rx, 1);
474 if (rc < 0) /* repost on error */
475 post_credit = IBLND_POSTRX_PEER_CREDIT;
478 case IBLND_MSG_GET_DONE:
479 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
480 kiblnd_handle_completion(conn, IBLND_MSG_GET_REQ,
481 msg->ibm_u.completion.ibcm_status,
482 msg->ibm_u.completion.ibcm_cookie);
486 if (rc < 0) /* protocol error */
487 kiblnd_close_conn(conn, rc);
489 if (post_credit != IBLND_POSTRX_DONT_POST)
490 kiblnd_post_rx(rx, post_credit);
494 kiblnd_rx_complete(struct kib_rx *rx, int status, int nob)
496 struct kib_msg *msg = rx->rx_msg;
497 struct kib_conn *conn = rx->rx_conn;
498 struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
499 struct kib_net *net = ni->ni_data;
504 LASSERT(rx->rx_nob < 0); /* was posted */
505 rx->rx_nob = 0; /* isn't now */
507 if (conn->ibc_state > IBLND_CONN_ESTABLISHED)
510 if (status != IB_WC_SUCCESS) {
511 CNETERR("Rx from %s failed: %d\n",
512 libcfs_nid2str(conn->ibc_peer->ibp_nid), status);
519 rc = kiblnd_unpack_msg(msg, rx->rx_nob);
521 CERROR("Error %d unpacking rx from %s\n",
522 rc, libcfs_nid2str(conn->ibc_peer->ibp_nid));
526 if (msg->ibm_srcnid != conn->ibc_peer->ibp_nid ||
527 msg->ibm_dstnid != lnet_nid_to_nid4(&ni->ni_nid) ||
528 msg->ibm_srcstamp != conn->ibc_incarnation ||
529 msg->ibm_dststamp != net->ibn_incarnation) {
530 CERROR("Stale rx from %s\n",
531 libcfs_nid2str(conn->ibc_peer->ibp_nid));
536 /* set time last known alive */
537 kiblnd_peer_alive(conn->ibc_peer);
539 /* racing with connection establishment/teardown! */
541 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
542 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
545 write_lock_irqsave(g_lock, flags);
546 /* must check holding global lock to eliminate race */
547 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
548 list_add_tail(&rx->rx_list, &conn->ibc_early_rxs);
549 write_unlock_irqrestore(g_lock, flags);
552 write_unlock_irqrestore(g_lock, flags);
554 kiblnd_handle_rx(rx);
558 CDEBUG(D_NET, "rx %p conn %p\n", rx, conn);
559 kiblnd_close_conn(conn, err);
561 kiblnd_drop_rx(rx); /* Don't re-post rx. */
565 kiblnd_fmr_map_tx(struct kib_net *net, struct kib_tx *tx,
566 struct kib_rdma_desc *rd, u32 nob)
568 struct kib_hca_dev *hdev;
570 struct kib_fmr_poolset *fps;
575 LASSERT(tx->tx_pool != NULL);
576 LASSERT(tx->tx_pool->tpo_pool.po_owner != NULL);
579 hdev = tx->tx_pool->tpo_hdev;
580 cpt = tx->tx_pool->tpo_pool.po_owner->ps_cpt;
583 * If we're dealing with FastReg, but the device doesn't
584 * support GAPS and the tx has GAPS, then there is no real point
585 * in trying to map the memory, because it'll just fail. So
586 * preemptively fail with an appropriate message
588 if (IS_FAST_REG_DEV(dev) &&
589 !(dev->ibd_dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT) &&
591 CERROR("Using FastReg with no GAPS support, but tx has gaps. "
592 "Try setting use_fastreg_gaps to 1\n");
593 return -EPROTONOSUPPORT;
596 #ifdef HAVE_FMR_POOL_API
598 * FMR does not support gaps but the tx has gaps then
599 * we should make sure that the number of fragments we'll be sending
600 * over fits within the number of fragments negotiated on the
601 * connection, otherwise, we won't be able to RDMA the data.
602 * We need to maintain the number of fragments negotiation on the
603 * connection for backwards compatibility.
605 if (tx->tx_gaps && (dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED)) {
607 tx->tx_conn->ibc_max_frags <= rd->rd_nfrags) {
608 CERROR("TX number of frags (%d) is <= than connection"
609 " number of frags (%d). Consider setting peer's"
610 " map_on_demand to 256\n", tx->tx_nfrags,
611 tx->tx_conn->ibc_max_frags);
617 fps = net->ibn_fmr_ps[cpt];
618 rc = kiblnd_fmr_pool_map(fps, tx, rd, nob, 0, &tx->tx_fmr);
620 CERROR("Can't map %u bytes (%u/%u)s: %d\n", nob,
621 tx->tx_nfrags, rd->rd_nfrags, rc);
626 * If rd is not tx_rd, it's going to get sent to a peer_ni, who will
629 rd->rd_key = tx->tx_fmr.fmr_key;
631 * for FastReg or FMR with no gaps we can accumulate all
632 * the fragments in one FastReg or FMR fragment.
635 #ifdef HAVE_FMR_POOL_API
636 ((dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED)
639 IS_FAST_REG_DEV(dev)) {
640 /* FMR requires zero based address */
641 #ifdef HAVE_FMR_POOL_API
642 if (dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED)
643 rd->rd_frags[0].rf_addr &= ~hdev->ibh_page_mask;
645 rd->rd_frags[0].rf_nob = nob;
649 * We're transmitting with gaps using FMR.
650 * We'll need to use multiple fragments and identify the
651 * zero based address of each fragment.
653 for (i = 0; i < rd->rd_nfrags; i++) {
654 rd->rd_frags[i].rf_addr &= ~hdev->ibh_page_mask;
655 rd->rd_frags[i].rf_addr += i << hdev->ibh_page_shift;
663 kiblnd_unmap_tx(struct kib_tx *tx)
666 #ifdef HAVE_FMR_POOL_API
667 tx->tx_fmr.fmr_pfmr ||
670 kiblnd_fmr_pool_unmap(&tx->tx_fmr, tx->tx_status);
672 if (tx->tx_nfrags != 0) {
673 kiblnd_dma_unmap_sg(tx->tx_pool->tpo_hdev, tx);
678 #ifdef HAVE_IB_GET_DMA_MR
679 static struct ib_mr *
680 kiblnd_find_rd_dma_mr(struct lnet_ni *ni, struct kib_rdma_desc *rd)
682 struct kib_net *net = ni->ni_data;
683 struct kib_hca_dev *hdev = net->ibn_dev->ibd_hdev;
684 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
686 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
689 * if map-on-demand is turned on and the device supports
690 * either FMR or FastReg then use that. Otherwise use global
691 * memory regions. If that's not available either, then you're
692 * dead in the water and fail the operation.
694 if (tunables->lnd_map_on_demand && (IS_FAST_REG_DEV(net->ibn_dev)
695 #ifdef HAVE_FMR_POOL_API
696 || net->ibn_dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED
702 * hdev->ibh_mrs can be NULL. This case is dealt with gracefully
703 * in the call chain. The mapping will fail with appropriate error
706 return hdev->ibh_mrs;
710 static int kiblnd_map_tx(struct lnet_ni *ni, struct kib_tx *tx,
711 struct kib_rdma_desc *rd, int nfrags)
713 struct kib_net *net = ni->ni_data;
714 struct kib_hca_dev *hdev = net->ibn_dev->ibd_hdev;
715 #ifdef HAVE_IB_GET_DMA_MR
716 struct ib_mr *mr = NULL;
721 /* If rd is not tx_rd, it's going to get sent to a peer_ni and I'm the
723 tx->tx_dmadir = (rd != tx->tx_rd) ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
724 tx->tx_nfrags = nfrags;
726 rd->rd_nfrags = kiblnd_dma_map_sg(hdev, tx);
727 for (i = 0, nob = 0; i < rd->rd_nfrags; i++) {
728 rd->rd_frags[i].rf_nob = kiblnd_sg_dma_len(
729 hdev->ibh_ibdev, &tx->tx_frags[i]);
730 rd->rd_frags[i].rf_addr = kiblnd_sg_dma_address(
731 hdev->ibh_ibdev, &tx->tx_frags[i]);
732 nob += rd->rd_frags[i].rf_nob;
735 #ifdef HAVE_IB_GET_DMA_MR
736 mr = kiblnd_find_rd_dma_mr(ni, rd);
738 /* found pre-mapping MR */
739 rd->rd_key = (rd != tx->tx_rd) ? mr->rkey : mr->lkey;
744 if (net->ibn_fmr_ps != NULL)
745 return kiblnd_fmr_map_tx(net, tx, rd, nob);
750 static int kiblnd_setup_rd_kiov(struct lnet_ni *ni, struct kib_tx *tx,
751 struct kib_rdma_desc *rd, int nkiov,
752 struct bio_vec *kiov, int offset, int nob)
754 struct kib_net *net = ni->ni_data;
755 struct scatterlist *sg;
760 CDEBUG(D_NET, "niov %d offset %d nob %d\n", nkiov, offset, nob);
764 LASSERT(net != NULL);
766 while (offset >= kiov->bv_len) {
767 offset -= kiov->bv_len;
780 CERROR("lacking enough sg entries to map tx\n");
785 fragnob = min((int)(kiov->bv_len - offset), nob);
788 * We're allowed to start at a non-aligned page offset in
789 * the first fragment and end at a non-aligned page offset
790 * in the last fragment.
792 if ((fragnob < (int)(kiov->bv_len - offset)) &&
793 nkiov < max_nkiov && nob > fragnob) {
794 CDEBUG(D_NET, "fragnob %d < available page %d: with"
795 " remaining %d kiovs with %d nob left\n",
796 fragnob, (int)(kiov->bv_len - offset),
801 sg_set_page(sg, kiov->bv_page, fragnob,
802 kiov->bv_offset + offset);
811 return kiblnd_map_tx(ni, tx, rd, sg_count);
815 kiblnd_post_tx_locked(struct kib_conn *conn, struct kib_tx *tx, int credit)
816 __must_hold(&conn->ibc_lock)
818 struct kib_msg *msg = tx->tx_msg;
819 struct kib_peer_ni *peer_ni = conn->ibc_peer;
820 struct lnet_ni *ni = peer_ni->ibp_ni;
821 struct kib_fast_reg_descriptor *frd = tx->tx_fmr.fmr_frd;
822 int ver = conn->ibc_version;
826 LASSERT(tx->tx_queued);
827 /* We rely on this for QP sizing */
828 LASSERT(tx->tx_nwrq > 0 && tx->tx_nsge >= 0);
829 LASSERT(tx->tx_nwrq <= 1 + conn->ibc_max_frags);
831 LASSERT(credit == 0 || credit == 1);
832 LASSERT(conn->ibc_outstanding_credits >= 0);
833 LASSERT(conn->ibc_outstanding_credits <= conn->ibc_queue_depth);
834 LASSERT(conn->ibc_credits >= 0);
835 LASSERT(conn->ibc_credits <= conn->ibc_queue_depth);
837 if (conn->ibc_nsends_posted ==
838 kiblnd_concurrent_sends(ver, ni)) {
839 /* tx completions outstanding... */
840 CDEBUG(D_NET, "%s: posted enough\n",
841 libcfs_nid2str(peer_ni->ibp_nid));
845 if (credit != 0 && conn->ibc_credits == 0) { /* no credits */
846 CDEBUG(D_NET, "%s: no credits\n",
847 libcfs_nid2str(peer_ni->ibp_nid));
851 if (credit != 0 && !IBLND_OOB_CAPABLE(ver) &&
852 conn->ibc_credits == 1 && /* last credit reserved */
853 msg->ibm_type != IBLND_MSG_NOOP) { /* for NOOP */
854 CDEBUG(D_NET, "%s: not using last credit\n",
855 libcfs_nid2str(peer_ni->ibp_nid));
859 /* NB don't drop ibc_lock before bumping tx_sending */
860 list_del(&tx->tx_list);
863 if (msg->ibm_type == IBLND_MSG_NOOP &&
864 (!kiblnd_need_noop(conn) || /* redundant NOOP */
865 (IBLND_OOB_CAPABLE(ver) && /* posted enough NOOP */
866 conn->ibc_noops_posted == IBLND_OOB_MSGS(ver)))) {
867 /* OK to drop when posted enough NOOPs, since
868 * kiblnd_check_sends_locked will queue NOOP again when
869 * posted NOOPs complete */
870 spin_unlock(&conn->ibc_lock);
871 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
873 spin_lock(&conn->ibc_lock);
874 CDEBUG(D_NET, "%s(%d): redundant or enough NOOP\n",
875 libcfs_nid2str(peer_ni->ibp_nid),
876 conn->ibc_noops_posted);
880 kiblnd_pack_msg(peer_ni->ibp_ni, msg, ver, conn->ibc_outstanding_credits,
881 peer_ni->ibp_nid, conn->ibc_incarnation);
883 conn->ibc_credits -= credit;
884 conn->ibc_outstanding_credits = 0;
885 conn->ibc_nsends_posted++;
886 if (msg->ibm_type == IBLND_MSG_NOOP)
887 conn->ibc_noops_posted++;
889 /* CAVEAT EMPTOR! This tx could be the PUT_DONE of an RDMA
890 * PUT. If so, it was first queued here as a PUT_REQ, sent and
891 * stashed on ibc_active_txs, matched by an incoming PUT_ACK,
892 * and then re-queued here. It's (just) possible that
893 * tx_sending is non-zero if we've not done the tx_complete()
894 * from the first send; hence the ++ rather than = below. */
896 list_add(&tx->tx_list, &conn->ibc_active_txs);
898 /* I'm still holding ibc_lock! */
899 if (conn->ibc_state != IBLND_CONN_ESTABLISHED) {
901 } else if (tx->tx_pool->tpo_pool.po_failed ||
902 conn->ibc_hdev != tx->tx_pool->tpo_hdev) {
903 /* close_conn will launch failover */
906 struct ib_send_wr *bad = &tx->tx_wrq[tx->tx_nwrq - 1].wr;
907 struct ib_send_wr *wr = &tx->tx_wrq[0].wr;
909 if (frd != NULL && !frd->frd_posted) {
910 if (!frd->frd_valid) {
911 wr = &frd->frd_inv_wr.wr;
912 wr->next = &frd->frd_fastreg_wr.wr;
914 wr = &frd->frd_fastreg_wr.wr;
916 frd->frd_fastreg_wr.wr.next = &tx->tx_wrq[0].wr;
919 LASSERTF(bad->wr_id == kiblnd_ptr2wreqid(tx, IBLND_WID_TX),
920 "bad wr_id %#llx, opc %d, flags %d, peer_ni: %s\n",
921 bad->wr_id, bad->opcode, bad->send_flags,
922 libcfs_nid2str(conn->ibc_peer->ibp_nid));
925 if (lnet_send_error_simulation(tx->tx_lntmsg[0], &tx->tx_hstatus))
928 #ifdef HAVE_IB_POST_SEND_RECV_CONST
929 rc = ib_post_send(conn->ibc_cmid->qp, wr,
930 (const struct ib_send_wr **)&bad);
932 rc = ib_post_send(conn->ibc_cmid->qp, wr, &bad);
936 conn->ibc_last_send = ktime_get();
940 frd->frd_posted = true;
944 /* NB credits are transferred in the actual
945 * message, which can only be the last work item */
946 conn->ibc_credits += credit;
947 conn->ibc_outstanding_credits += msg->ibm_credits;
948 conn->ibc_nsends_posted--;
949 if (msg->ibm_type == IBLND_MSG_NOOP)
950 conn->ibc_noops_posted--;
956 done = (tx->tx_sending == 0);
958 list_del(&tx->tx_list);
960 spin_unlock(&conn->ibc_lock);
962 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
963 CERROR("Error %d posting transmit to %s\n",
964 rc, libcfs_nid2str(peer_ni->ibp_nid));
966 CDEBUG(D_NET, "Error %d posting transmit to %s\n",
967 rc, libcfs_nid2str(peer_ni->ibp_nid));
969 kiblnd_close_conn(conn, rc);
974 spin_lock(&conn->ibc_lock);
980 kiblnd_check_sends_locked(struct kib_conn *conn)
982 int ver = conn->ibc_version;
983 struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
986 /* Don't send anything until after the connection is established */
987 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
988 CDEBUG(D_NET, "%s too soon\n",
989 libcfs_nid2str(conn->ibc_peer->ibp_nid));
993 LASSERT(conn->ibc_nsends_posted <=
994 kiblnd_concurrent_sends(ver, ni));
995 LASSERT (!IBLND_OOB_CAPABLE(ver) ||
996 conn->ibc_noops_posted <= IBLND_OOB_MSGS(ver));
997 LASSERT (conn->ibc_reserved_credits >= 0);
999 while (conn->ibc_reserved_credits > 0 &&
1000 (tx = list_first_entry_or_null(&conn->ibc_tx_queue_rsrvd,
1001 struct kib_tx, tx_list)) != NULL) {
1002 list_move_tail(&tx->tx_list, &conn->ibc_tx_queue);
1003 conn->ibc_reserved_credits--;
1006 if (kiblnd_need_noop(conn)) {
1007 spin_unlock(&conn->ibc_lock);
1009 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
1011 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_NOOP, 0);
1013 spin_lock(&conn->ibc_lock);
1015 kiblnd_queue_tx_locked(tx, conn);
1021 if (!list_empty(&conn->ibc_tx_queue_nocred)) {
1023 tx = list_first_entry(&conn->ibc_tx_queue_nocred,
1024 struct kib_tx, tx_list);
1025 } else if (!list_empty(&conn->ibc_tx_noops)) {
1026 LASSERT (!IBLND_OOB_CAPABLE(ver));
1028 tx = list_first_entry(&conn->ibc_tx_noops,
1029 struct kib_tx, tx_list);
1030 } else if (!list_empty(&conn->ibc_tx_queue)) {
1032 tx = list_first_entry(&conn->ibc_tx_queue,
1033 struct kib_tx, tx_list);
1037 if (kiblnd_post_tx_locked(conn, tx, credit) != 0)
1043 kiblnd_tx_complete(struct kib_tx *tx, int status)
1045 int failed = (status != IB_WC_SUCCESS);
1046 struct kib_conn *conn = tx->tx_conn;
1049 if (tx->tx_sending <= 0) {
1050 CERROR("Received an event on a freed tx: %p status %d\n",
1056 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
1057 CNETERR("Tx -> %s cookie %#llx"
1058 " sending %d waiting %d: failed %d\n",
1059 libcfs_nid2str(conn->ibc_peer->ibp_nid),
1060 tx->tx_cookie, tx->tx_sending, tx->tx_waiting,
1063 kiblnd_close_conn(conn, -EIO);
1065 kiblnd_peer_alive(conn->ibc_peer);
1068 spin_lock(&conn->ibc_lock);
1070 /* I could be racing with rdma completion. Whoever makes 'tx' idle
1071 * gets to free it, which also drops its ref on 'conn'. */
1074 conn->ibc_nsends_posted--;
1075 if (tx->tx_msg->ibm_type == IBLND_MSG_NOOP)
1076 conn->ibc_noops_posted--;
1079 tx->tx_hstatus = LNET_MSG_STATUS_REMOTE_DROPPED;
1080 tx->tx_waiting = 0; /* don't wait for peer_ni */
1081 tx->tx_status = -EIO;
1084 idle = (tx->tx_sending == 0) && /* This is the final callback */
1085 !tx->tx_waiting && /* Not waiting for peer_ni */
1086 !tx->tx_queued; /* Not re-queued (PUT_DONE) */
1088 list_del(&tx->tx_list);
1090 kiblnd_check_sends_locked(conn);
1091 spin_unlock(&conn->ibc_lock);
1099 kiblnd_init_tx_sge(struct kib_tx *tx, u64 addr, unsigned int len)
1101 struct ib_sge *sge = &tx->tx_sge[tx->tx_nsge];
1102 struct kib_hca_dev *hdev = tx->tx_pool->tpo_hdev;
1103 #ifdef HAVE_IB_GET_DMA_MR
1104 struct ib_mr *mr = hdev->ibh_mrs;
1107 *sge = (struct ib_sge) {
1108 #ifdef HAVE_IB_GET_DMA_MR
1111 .lkey = hdev->ibh_pd->local_dma_lkey,
1120 static struct ib_rdma_wr *
1121 kiblnd_init_tx_msg_payload(struct lnet_ni *ni, struct kib_tx *tx, int type,
1122 int body_nob, int payload)
1124 struct ib_rdma_wr *wrq;
1125 int nob = offsetof(struct kib_msg, ibm_u) + body_nob;
1127 LASSERT(tx->tx_nwrq >= 0);
1128 LASSERT(tx->tx_nwrq < IBLND_MAX_RDMA_FRAGS + 1);
1129 LASSERT(nob <= IBLND_MSG_SIZE);
1131 kiblnd_init_msg(tx->tx_msg, type, body_nob + payload);
1133 wrq = &tx->tx_wrq[tx->tx_nwrq];
1135 *wrq = (struct ib_rdma_wr) {
1137 .wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_TX),
1139 .sg_list = &tx->tx_sge[tx->tx_nsge],
1140 .opcode = IB_WR_SEND,
1141 .send_flags = IB_SEND_SIGNALED,
1145 kiblnd_init_tx_sge(tx, tx->tx_msgaddr, nob);
1152 kiblnd_init_rdma(struct kib_conn *conn, struct kib_tx *tx, int type,
1153 int resid, struct kib_rdma_desc *dstrd, u64 dstcookie)
1155 struct kib_msg *ibmsg = tx->tx_msg;
1156 struct kib_rdma_desc *srcrd = tx->tx_rd;
1157 struct ib_rdma_wr *wrq = NULL;
1165 LASSERT(!in_interrupt());
1166 LASSERT(tx->tx_nwrq == 0 && tx->tx_nsge == 0);
1167 LASSERT(type == IBLND_MSG_GET_DONE || type == IBLND_MSG_PUT_DONE);
1169 for (srcidx = dstidx = wrq_sge = sge_nob = 0;
1170 resid > 0; resid -= sge_nob) {
1173 if (srcidx >= srcrd->rd_nfrags) {
1174 CERROR("Src buffer exhausted: %d frags %px\n",
1180 if (dstidx >= dstrd->rd_nfrags) {
1181 CERROR("Dst buffer exhausted: %d frags\n", dstidx);
1186 if (tx->tx_nwrq >= conn->ibc_max_frags) {
1187 CERROR("RDMA has too many fragments for peer_ni %s (%d), "
1188 "src idx/frags: %d/%d dst idx/frags: %d/%d\n",
1189 libcfs_nid2str(conn->ibc_peer->ibp_nid),
1190 conn->ibc_max_frags,
1191 srcidx, srcrd->rd_nfrags,
1192 dstidx, dstrd->rd_nfrags);
1197 sge_nob = min3(kiblnd_rd_frag_size(srcrd, srcidx),
1198 kiblnd_rd_frag_size(dstrd, dstidx),
1201 sge = &tx->tx_sge[tx->tx_nsge];
1202 sge->addr = kiblnd_rd_frag_addr(srcrd, srcidx);
1203 sge->lkey = kiblnd_rd_frag_key(srcrd, srcidx);
1204 sge->length = sge_nob;
1207 wrq = &tx->tx_wrq[tx->tx_nwrq];
1209 wrq->wr.next = &(wrq + 1)->wr;
1210 wrq->wr.wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_RDMA);
1211 wrq->wr.sg_list = sge;
1212 wrq->wr.opcode = IB_WR_RDMA_WRITE;
1213 wrq->wr.send_flags = 0;
1215 #ifdef HAVE_IB_RDMA_WR
1216 wrq->remote_addr = kiblnd_rd_frag_addr(dstrd,
1218 wrq->rkey = kiblnd_rd_frag_key(dstrd,
1221 wrq->wr.wr.rdma.remote_addr = kiblnd_rd_frag_addr(dstrd,
1223 wrq->wr.wr.rdma.rkey = kiblnd_rd_frag_key(dstrd,
1228 srcidx = kiblnd_rd_consume_frag(srcrd, srcidx, sge_nob);
1229 dstidx = kiblnd_rd_consume_frag(dstrd, dstidx, sge_nob);
1232 if (wrq_sge == *kiblnd_tunables.kib_wrq_sge || dstidx != prev) {
1234 wrq->wr.num_sge = wrq_sge;
1240 if (rc < 0) /* no RDMA if completing with failure */
1241 tx->tx_nwrq = tx->tx_nsge = 0;
1243 ibmsg->ibm_u.completion.ibcm_status = rc;
1244 ibmsg->ibm_u.completion.ibcm_cookie = dstcookie;
1245 kiblnd_init_tx_msg(conn->ibc_peer->ibp_ni, tx,
1246 type, sizeof(struct kib_completion_msg));
1252 kiblnd_queue_tx_locked(struct kib_tx *tx, struct kib_conn *conn)
1254 struct list_head *q;
1257 LASSERT(tx->tx_nwrq > 0); /* work items set up */
1258 LASSERT(!tx->tx_queued); /* not queued for sending already */
1259 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1261 if (conn->ibc_state >= IBLND_CONN_DISCONNECTED) {
1262 tx->tx_status = -ECONNABORTED;
1264 if (tx->tx_conn != NULL) {
1265 /* PUT_DONE first attached to conn as a PUT_REQ */
1266 LASSERT(tx->tx_conn == conn);
1267 LASSERT(tx->tx_msg->ibm_type == IBLND_MSG_PUT_DONE);
1269 kiblnd_conn_decref(conn);
1271 list_add(&tx->tx_list, &conn->ibc_zombie_txs);
1276 timeout_ns = kiblnd_timeout() * NSEC_PER_SEC;
1278 tx->tx_deadline = ktime_add_ns(ktime_get(), timeout_ns);
1280 if (tx->tx_conn == NULL) {
1281 kiblnd_conn_addref(conn);
1283 LASSERT (tx->tx_msg->ibm_type != IBLND_MSG_PUT_DONE);
1285 /* PUT_DONE first attached to conn as a PUT_REQ */
1286 LASSERT (tx->tx_conn == conn);
1287 LASSERT (tx->tx_msg->ibm_type == IBLND_MSG_PUT_DONE);
1290 switch (tx->tx_msg->ibm_type) {
1294 case IBLND_MSG_PUT_REQ:
1295 case IBLND_MSG_GET_REQ:
1296 q = &conn->ibc_tx_queue_rsrvd;
1299 case IBLND_MSG_PUT_NAK:
1300 case IBLND_MSG_PUT_ACK:
1301 case IBLND_MSG_PUT_DONE:
1302 case IBLND_MSG_GET_DONE:
1303 q = &conn->ibc_tx_queue_nocred;
1306 case IBLND_MSG_NOOP:
1307 if (IBLND_OOB_CAPABLE(conn->ibc_version))
1308 q = &conn->ibc_tx_queue_nocred;
1310 q = &conn->ibc_tx_noops;
1313 case IBLND_MSG_IMMEDIATE:
1314 q = &conn->ibc_tx_queue;
1318 list_add_tail(&tx->tx_list, q);
1322 kiblnd_queue_tx(struct kib_tx *tx, struct kib_conn *conn)
1324 spin_lock(&conn->ibc_lock);
1325 kiblnd_queue_tx_locked(tx, conn);
1326 kiblnd_check_sends_locked(conn);
1327 spin_unlock(&conn->ibc_lock);
1331 kiblnd_resolve_addr_cap(struct rdma_cm_id *cmid,
1332 struct sockaddr_in *srcaddr,
1333 struct sockaddr_in *dstaddr,
1336 unsigned short port;
1339 /* allow the port to be reused */
1340 rc = rdma_set_reuseaddr(cmid, 1);
1342 CERROR("Unable to set reuse on cmid: %d\n", rc);
1346 /* look for a free privileged port */
1347 for (port = PROT_SOCK-1; port > 0; port--) {
1348 srcaddr->sin_port = htons(port);
1349 rc = rdma_resolve_addr(cmid,
1350 (struct sockaddr *)srcaddr,
1351 (struct sockaddr *)dstaddr,
1354 CDEBUG(D_NET, "bound to port %hu\n", port);
1356 } else if (rc == -EADDRINUSE || rc == -EADDRNOTAVAIL) {
1357 CDEBUG(D_NET, "bind to port %hu failed: %d\n",
1364 CERROR("cannot bind to a free privileged port: rc = %d\n", rc);
1370 kiblnd_resolve_addr(struct rdma_cm_id *cmid,
1371 struct sockaddr_in *srcaddr,
1372 struct sockaddr_in *dstaddr,
1375 const struct cred *old_creds = NULL;
1376 struct cred *new_creds;
1379 if (!capable(CAP_NET_BIND_SERVICE)) {
1380 new_creds = prepare_kernel_cred(NULL);
1384 cap_raise(new_creds->cap_effective, CAP_NET_BIND_SERVICE);
1385 old_creds = override_creds(new_creds);
1388 rc = kiblnd_resolve_addr_cap(cmid, srcaddr, dstaddr, timeout_ms);
1391 revert_creds(old_creds);
1397 kiblnd_connect_peer(struct kib_peer_ni *peer_ni)
1399 struct rdma_cm_id *cmid;
1400 struct kib_dev *dev;
1401 struct kib_net *net = peer_ni->ibp_ni->ni_data;
1402 struct sockaddr_in srcaddr;
1403 struct sockaddr_in dstaddr;
1406 LASSERT (net != NULL);
1407 LASSERT (peer_ni->ibp_connecting > 0);
1409 cmid = kiblnd_rdma_create_id(peer_ni->ibp_ni->ni_net_ns,
1410 kiblnd_cm_callback, peer_ni,
1411 RDMA_PS_TCP, IB_QPT_RC);
1414 CERROR("Can't create CMID for %s: %ld\n",
1415 libcfs_nid2str(peer_ni->ibp_nid), PTR_ERR(cmid));
1421 memset(&srcaddr, 0, sizeof(srcaddr));
1422 srcaddr.sin_family = AF_INET;
1423 srcaddr.sin_addr.s_addr = htonl(dev->ibd_ifip);
1425 memset(&dstaddr, 0, sizeof(dstaddr));
1426 dstaddr.sin_family = AF_INET;
1427 dstaddr.sin_port = htons(*kiblnd_tunables.kib_service);
1428 dstaddr.sin_addr.s_addr = htonl(LNET_NIDADDR(peer_ni->ibp_nid));
1430 kiblnd_peer_addref(peer_ni); /* cmid's ref */
1432 if (*kiblnd_tunables.kib_use_priv_port) {
1433 rc = kiblnd_resolve_addr(cmid, &srcaddr, &dstaddr,
1434 kiblnd_timeout() * 1000);
1436 rc = rdma_resolve_addr(cmid,
1437 (struct sockaddr *)&srcaddr,
1438 (struct sockaddr *)&dstaddr,
1439 kiblnd_timeout() * 1000);
1442 /* Can't initiate address resolution: */
1443 CERROR("Can't resolve addr for %s: %d\n",
1444 libcfs_nid2str(peer_ni->ibp_nid), rc);
1451 kiblnd_peer_connect_failed(peer_ni, 1, rc);
1452 kiblnd_peer_decref(peer_ni); /* cmid's ref */
1453 rdma_destroy_id(cmid);
1456 kiblnd_peer_connect_failed(peer_ni, 1, rc);
1460 kiblnd_reconnect_peer(struct kib_peer_ni *peer_ni)
1462 rwlock_t *glock = &kiblnd_data.kib_global_lock;
1463 char *reason = NULL;
1465 unsigned long flags;
1467 write_lock_irqsave(glock, flags);
1468 if (peer_ni->ibp_reconnecting == 0) {
1469 if (peer_ni->ibp_accepting)
1470 reason = "accepting";
1471 else if (peer_ni->ibp_connecting)
1472 reason = "connecting";
1473 else if (!list_empty(&peer_ni->ibp_conns))
1474 reason = "connected";
1475 else /* connected then closed */
1481 if (peer_ni->ibp_accepting)
1482 CNETERR("Detecting race between accepting and reconnecting\n");
1483 peer_ni->ibp_reconnecting--;
1485 if (!kiblnd_peer_active(peer_ni)) {
1486 list_splice_init(&peer_ni->ibp_tx_queue, &txs);
1487 reason = "unlinked";
1491 peer_ni->ibp_connecting++;
1492 peer_ni->ibp_reconnected++;
1494 write_unlock_irqrestore(glock, flags);
1496 kiblnd_connect_peer(peer_ni);
1500 write_unlock_irqrestore(glock, flags);
1502 CWARN("Abort reconnection of %s: %s\n",
1503 libcfs_nid2str(peer_ni->ibp_nid), reason);
1504 kiblnd_txlist_done(&txs, -ECONNABORTED,
1505 LNET_MSG_STATUS_LOCAL_ABORTED);
1510 kiblnd_launch_tx(struct lnet_ni *ni, struct kib_tx *tx, lnet_nid_t nid)
1512 struct kib_peer_ni *peer_ni;
1513 struct kib_peer_ni *peer2;
1514 struct kib_conn *conn;
1515 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
1516 unsigned long flags;
1519 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
1521 /* If I get here, I've committed to send, so I complete the tx with
1522 * failure on any problems
1525 LASSERT(!tx || !tx->tx_conn); /* only set when assigned a conn */
1526 LASSERT(!tx || tx->tx_nwrq > 0); /* work items have been set up */
1528 /* First time, just use a read lock since I expect to find my peer_ni
1531 read_lock_irqsave(g_lock, flags);
1533 peer_ni = kiblnd_find_peer_locked(ni, nid);
1534 if (peer_ni != NULL && !list_empty(&peer_ni->ibp_conns)) {
1535 /* Found a peer_ni with an established connection */
1536 conn = kiblnd_get_conn_locked(peer_ni);
1537 kiblnd_conn_addref(conn); /* 1 ref for me... */
1539 read_unlock_irqrestore(g_lock, flags);
1542 kiblnd_queue_tx(tx, conn);
1543 kiblnd_conn_decref(conn); /* ...to here */
1547 read_unlock(g_lock);
1548 /* Re-try with a write lock */
1551 peer_ni = kiblnd_find_peer_locked(ni, nid);
1552 if (peer_ni != NULL) {
1553 if (list_empty(&peer_ni->ibp_conns)) {
1554 /* found a peer_ni, but it's still connecting... */
1555 LASSERT(kiblnd_peer_connecting(peer_ni));
1557 list_add_tail(&tx->tx_list,
1558 &peer_ni->ibp_tx_queue);
1559 write_unlock_irqrestore(g_lock, flags);
1561 conn = kiblnd_get_conn_locked(peer_ni);
1562 kiblnd_conn_addref(conn); /* 1 ref for me... */
1564 write_unlock_irqrestore(g_lock, flags);
1567 kiblnd_queue_tx(tx, conn);
1568 kiblnd_conn_decref(conn); /* ...to here */
1573 write_unlock_irqrestore(g_lock, flags);
1575 /* Allocate a peer_ni ready to add to the peer_ni table and retry */
1576 rc = kiblnd_create_peer(ni, &peer_ni, nid);
1578 CERROR("Can't create peer_ni %s\n", libcfs_nid2str(nid));
1580 tx->tx_status = -EHOSTUNREACH;
1582 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
1588 write_lock_irqsave(g_lock, flags);
1590 peer2 = kiblnd_find_peer_locked(ni, nid);
1591 if (peer2 != NULL) {
1592 if (list_empty(&peer2->ibp_conns)) {
1593 /* found a peer_ni, but it's still connecting... */
1594 LASSERT(kiblnd_peer_connecting(peer2));
1596 list_add_tail(&tx->tx_list,
1597 &peer2->ibp_tx_queue);
1598 write_unlock_irqrestore(g_lock, flags);
1600 conn = kiblnd_get_conn_locked(peer2);
1601 kiblnd_conn_addref(conn); /* 1 ref for me... */
1603 write_unlock_irqrestore(g_lock, flags);
1606 kiblnd_queue_tx(tx, conn);
1607 kiblnd_conn_decref(conn); /* ...to here */
1610 kiblnd_peer_decref(peer_ni);
1614 /* Brand new peer_ni */
1615 LASSERT(peer_ni->ibp_connecting == 0);
1616 tunables = &peer_ni->ibp_ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
1617 peer_ni->ibp_connecting = tunables->lnd_conns_per_peer;
1619 /* always called with a ref on ni, which prevents ni being shutdown */
1620 LASSERT(((struct kib_net *)ni->ni_data)->ibn_shutdown == 0);
1623 list_add_tail(&tx->tx_list, &peer_ni->ibp_tx_queue);
1625 kiblnd_peer_addref(peer_ni);
1626 hash_add(kiblnd_data.kib_peers, &peer_ni->ibp_list, nid);
1628 write_unlock_irqrestore(g_lock, flags);
1630 for (i = 0; i < tunables->lnd_conns_per_peer; i++)
1631 kiblnd_connect_peer(peer_ni);
1632 kiblnd_peer_decref(peer_ni);
1636 kiblnd_send(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg)
1638 struct kib_dev *dev = ((struct kib_net *)ni->ni_data)->ibn_dev;
1639 struct lnet_hdr *hdr = &lntmsg->msg_hdr;
1640 int type = lntmsg->msg_type;
1641 struct lnet_processid *target = &lntmsg->msg_target;
1642 int target_is_router = lntmsg->msg_target_is_router;
1643 int routing = lntmsg->msg_routing;
1644 unsigned int payload_niov = lntmsg->msg_niov;
1645 struct bio_vec *payload_kiov = lntmsg->msg_kiov;
1646 unsigned int payload_offset = lntmsg->msg_offset;
1647 unsigned int payload_nob = lntmsg->msg_len;
1649 struct kib_msg *ibmsg;
1650 struct kib_rdma_desc *rd;
1655 /* NB 'private' is different depending on what we're sending.... */
1657 CDEBUG(D_NET, "sending %d bytes in %d frags to %s\n",
1658 payload_nob, payload_niov, libcfs_idstr(target));
1660 LASSERT (payload_nob == 0 || payload_niov > 0);
1661 LASSERT (payload_niov <= LNET_MAX_IOV);
1663 /* Thread context */
1664 LASSERT (!in_interrupt());
1666 tx = kiblnd_get_idle_tx(ni, lnet_nid_to_nid4(&target->nid));
1668 CERROR("Can't allocate %s txd for %s\n",
1669 lnet_msgtyp2str(type),
1670 libcfs_nidstr(&target->nid));
1674 gpu = (lntmsg->msg_md->md_flags & LNET_MD_FLAG_GPU);
1682 LASSERT(payload_nob == 0);
1686 if (routing || target_is_router)
1687 break; /* send IMMEDIATE */
1689 /* is the REPLY message too small for RDMA? */
1690 nob = offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[lntmsg->msg_md->md_length]);
1691 if (nob <= IBLND_MSG_SIZE && !gpu)
1692 break; /* send IMMEDIATE */
1694 rd = &ibmsg->ibm_u.get.ibgm_rd;
1696 rc = kiblnd_setup_rd_kiov(ni, tx, rd,
1697 lntmsg->msg_md->md_niov,
1698 lntmsg->msg_md->md_kiov,
1699 0, lntmsg->msg_md->md_length);
1701 CERROR("Can't setup GET sink for %s: %d\n",
1702 libcfs_nidstr(&target->nid), rc);
1703 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
1708 nob = offsetof(struct kib_get_msg, ibgm_rd.rd_frags[rd->rd_nfrags]);
1709 ibmsg->ibm_u.get.ibgm_cookie = tx->tx_cookie;
1710 lnet_hdr_to_nid4(hdr, &ibmsg->ibm_u.get.ibgm_hdr);
1712 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_GET_REQ, nob);
1714 tx->tx_lntmsg[1] = lnet_create_reply_msg(ni, lntmsg);
1715 if (tx->tx_lntmsg[1] == NULL) {
1716 CERROR("Can't create reply for GET -> %s\n",
1717 libcfs_nidstr(&target->nid));
1722 /* finalise lntmsg[0,1] on completion */
1723 tx->tx_lntmsg[0] = lntmsg;
1724 tx->tx_waiting = 1; /* waiting for GET_DONE */
1725 kiblnd_launch_tx(ni, tx, lnet_nid_to_nid4(&target->nid));
1728 case LNET_MSG_REPLY:
1730 /* Is the payload small enough not to need RDMA? */
1731 nob = offsetof(struct kib_msg,
1732 ibm_u.immediate.ibim_payload[payload_nob]);
1733 if (nob <= IBLND_MSG_SIZE && !gpu)
1734 break; /* send IMMEDIATE */
1738 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1739 payload_niov, payload_kiov,
1740 payload_offset, payload_nob);
1742 CERROR("Can't setup PUT src for %s: %d\n",
1743 libcfs_nidstr(&target->nid), rc);
1748 lnet_hdr_to_nid4(hdr, &ibmsg->ibm_u.putreq.ibprm_hdr);
1749 ibmsg->ibm_u.putreq.ibprm_cookie = tx->tx_cookie;
1750 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_REQ,
1751 sizeof(struct kib_putreq_msg));
1753 /* finalise lntmsg[0,1] on completion */
1754 tx->tx_lntmsg[0] = lntmsg;
1755 tx->tx_waiting = 1; /* waiting for PUT_{ACK,NAK} */
1756 kiblnd_launch_tx(ni, tx, lnet_nid_to_nid4(&target->nid));
1760 /* send IMMEDIATE */
1761 LASSERT(offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[payload_nob])
1765 lnet_hdr_to_nid4(hdr, &ibmsg->ibm_u.immediate.ibim_hdr);
1767 if (IS_FAST_REG_DEV(dev) && payload_nob) {
1768 struct ib_rdma_wr *wrq;
1771 nob = offsetof(struct kib_immediate_msg, ibim_payload[0]);
1772 wrq = kiblnd_init_tx_msg_payload(ni, tx, IBLND_MSG_IMMEDIATE,
1776 rc = kiblnd_setup_rd_kiov(ni, tx, rd,
1777 payload_niov, payload_kiov,
1778 payload_offset, payload_nob);
1780 CERROR("Can't setup IMMEDIATE src for %s: %d\n",
1781 libcfs_nidstr(&target->nid), rc);
1786 /* lets generate a SGE chain */
1787 for (i = 0; i < rd->rd_nfrags; i++) {
1788 kiblnd_init_tx_sge(tx, rd->rd_frags[i].rf_addr,
1789 rd->rd_frags[i].rf_nob);
1793 lnet_copy_kiov2flat(IBLND_MSG_SIZE, ibmsg,
1794 offsetof(struct kib_msg,
1795 ibm_u.immediate.ibim_payload),
1796 payload_niov, payload_kiov,
1797 payload_offset, payload_nob);
1799 nob = offsetof(struct kib_immediate_msg,
1800 ibim_payload[payload_nob]);
1802 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_IMMEDIATE, nob);
1805 /* finalise lntmsg on completion */
1806 tx->tx_lntmsg[0] = lntmsg;
1808 kiblnd_launch_tx(ni, tx, lnet_nid_to_nid4(&target->nid));
1813 kiblnd_reply(struct lnet_ni *ni, struct kib_rx *rx, struct lnet_msg *lntmsg)
1815 struct lnet_processid *target = &lntmsg->msg_target;
1816 unsigned int niov = lntmsg->msg_niov;
1817 struct bio_vec *kiov = lntmsg->msg_kiov;
1818 unsigned int offset = lntmsg->msg_offset;
1819 unsigned int nob = lntmsg->msg_len;
1820 struct lnet_libmd *payload_md = lntmsg->msg_md;
1824 tx = kiblnd_get_idle_tx(ni, rx->rx_conn->ibc_peer->ibp_nid);
1826 CERROR("Can't get tx for REPLY to %s\n",
1827 libcfs_nidstr(&target->nid));
1831 tx->tx_gpu = !!(payload_md->md_flags & LNET_MD_FLAG_GPU);
1835 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1836 niov, kiov, offset, nob);
1839 CERROR("Can't setup GET src for %s: %d\n",
1840 libcfs_nidstr(&target->nid), rc);
1844 rc = kiblnd_init_rdma(rx->rx_conn, tx,
1845 IBLND_MSG_GET_DONE, nob,
1846 &rx->rx_msg->ibm_u.get.ibgm_rd,
1847 rx->rx_msg->ibm_u.get.ibgm_cookie);
1849 CERROR("Can't setup rdma for GET from %s: %d\n",
1850 libcfs_nidstr(&target->nid), rc);
1855 /* No RDMA: local completion may happen now! */
1856 lnet_finalize(lntmsg, 0);
1858 /* RDMA: lnet_finalize(lntmsg) when it
1860 tx->tx_lntmsg[0] = lntmsg;
1863 kiblnd_queue_tx(tx, rx->rx_conn);
1868 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
1871 lnet_finalize(lntmsg, -EIO);
1875 kiblnd_get_dev_prio(struct lnet_ni *ni, unsigned int dev_idx)
1877 struct kib_net *net = ni->ni_data;
1878 struct device *dev = NULL;
1881 dev = net->ibn_dev->ibd_hdev->ibh_ibdev->dma_device;
1883 return lnet_get_dev_prio(dev, dev_idx);
1888 kiblnd_recv(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg,
1889 int delayed, unsigned int niov, struct bio_vec *kiov,
1890 unsigned int offset, unsigned int mlen, unsigned int rlen)
1892 struct kib_rx *rx = private;
1893 struct kib_msg *rxmsg = rx->rx_msg;
1894 struct kib_conn *conn = rx->rx_conn;
1898 int post_credit = IBLND_POSTRX_PEER_CREDIT;
1901 LASSERT (mlen <= rlen);
1902 LASSERT (!in_interrupt());
1904 switch (rxmsg->ibm_type) {
1908 case IBLND_MSG_IMMEDIATE:
1909 nob = offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[rlen]);
1910 if (nob > rx->rx_nob) {
1911 CERROR("Immediate message from %s too big: %d(%d)\n",
1912 libcfs_nidstr(&lntmsg->msg_hdr.src_nid),
1918 lnet_copy_flat2kiov(niov, kiov, offset,
1919 IBLND_MSG_SIZE, rxmsg,
1920 offsetof(struct kib_msg,
1921 ibm_u.immediate.ibim_payload),
1923 lnet_finalize(lntmsg, 0);
1926 case IBLND_MSG_PUT_REQ: {
1927 struct kib_msg *txmsg;
1928 struct kib_rdma_desc *rd;
1929 struct lnet_libmd *payload_md = lntmsg->msg_md;
1931 ibprm_cookie = rxmsg->ibm_u.putreq.ibprm_cookie;
1933 lnet_finalize(lntmsg, 0);
1934 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK,
1939 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
1941 CERROR("Can't allocate tx for %s\n",
1942 libcfs_nid2str(conn->ibc_peer->ibp_nid));
1943 /* Not replying will break the connection */
1948 tx->tx_gpu = !!(payload_md->md_flags & LNET_MD_FLAG_GPU);
1950 rd = &txmsg->ibm_u.putack.ibpam_rd;
1951 rc = kiblnd_setup_rd_kiov(ni, tx, rd,
1952 niov, kiov, offset, mlen);
1954 CERROR("Can't setup PUT sink for %s: %d\n",
1955 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
1956 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
1958 /* tell peer_ni it's over */
1959 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK,
1964 nob = offsetof(struct kib_putack_msg, ibpam_rd.rd_frags[rd->rd_nfrags]);
1965 txmsg->ibm_u.putack.ibpam_src_cookie = ibprm_cookie;
1966 txmsg->ibm_u.putack.ibpam_dst_cookie = tx->tx_cookie;
1968 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_ACK, nob);
1970 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1971 tx->tx_waiting = 1; /* waiting for PUT_DONE */
1972 kiblnd_queue_tx(tx, conn);
1974 /* reposted buffer reserved for PUT_DONE */
1975 post_credit = IBLND_POSTRX_NO_CREDIT;
1979 case IBLND_MSG_GET_REQ:
1980 if (lntmsg != NULL) {
1981 /* Optimized GET; RDMA lntmsg's payload */
1982 kiblnd_reply(ni, rx, lntmsg);
1984 /* GET didn't match anything */
1985 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_GET_DONE,
1987 rxmsg->ibm_u.get.ibgm_cookie);
1992 kiblnd_post_rx(rx, post_credit);
1997 kiblnd_thread_fini (void)
1999 atomic_dec (&kiblnd_data.kib_nthreads);
2003 kiblnd_peer_alive(struct kib_peer_ni *peer_ni)
2005 /* This is racy, but everyone's only writing ktime_get_seconds() */
2006 peer_ni->ibp_last_alive = ktime_get_seconds();
2011 kiblnd_peer_notify(struct kib_peer_ni *peer_ni)
2014 time64_t last_alive = 0;
2015 unsigned long flags;
2017 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2019 if (kiblnd_peer_idle(peer_ni) && peer_ni->ibp_error != 0) {
2020 error = peer_ni->ibp_error;
2021 peer_ni->ibp_error = 0;
2023 last_alive = peer_ni->ibp_last_alive;
2026 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2029 lnet_notify(peer_ni->ibp_ni,
2030 peer_ni->ibp_nid, false, false, last_alive);
2034 kiblnd_close_conn_locked(struct kib_conn *conn, int error)
2036 /* This just does the immediate housekeeping. 'error' is zero for a
2037 * normal shutdown which can happen only after the connection has been
2038 * established. If the connection is established, schedule the
2039 * connection to be finished off by the connd. Otherwise the connd is
2040 * already dealing with it (either to set it up or tear it down).
2041 * Caller holds kib_global_lock exclusively in irq context */
2042 struct kib_peer_ni *peer_ni = conn->ibc_peer;
2043 struct kib_dev *dev;
2044 unsigned long flags;
2046 LASSERT (error != 0 || conn->ibc_state >= IBLND_CONN_ESTABLISHED);
2048 if (error != 0 && conn->ibc_comms_error == 0)
2049 conn->ibc_comms_error = error;
2051 if (conn->ibc_state != IBLND_CONN_ESTABLISHED)
2052 return; /* already being handled */
2055 list_empty(&conn->ibc_tx_noops) &&
2056 list_empty(&conn->ibc_tx_queue) &&
2057 list_empty(&conn->ibc_tx_queue_rsrvd) &&
2058 list_empty(&conn->ibc_tx_queue_nocred) &&
2059 list_empty(&conn->ibc_active_txs)) {
2060 CDEBUG(D_NET, "closing conn to %s\n",
2061 libcfs_nid2str(peer_ni->ibp_nid));
2063 CNETERR("Closing conn to %s: error %d%s%s%s%s%s\n",
2064 libcfs_nid2str(peer_ni->ibp_nid), error,
2065 list_empty(&conn->ibc_tx_queue) ? "" : "(sending)",
2066 list_empty(&conn->ibc_tx_noops) ? "" : "(sending_noops)",
2067 list_empty(&conn->ibc_tx_queue_rsrvd) ?
2068 "" : "(sending_rsrvd)",
2069 list_empty(&conn->ibc_tx_queue_nocred) ?
2070 "" : "(sending_nocred)",
2071 list_empty(&conn->ibc_active_txs) ? "" : "(waiting)");
2074 dev = ((struct kib_net *)peer_ni->ibp_ni->ni_data)->ibn_dev;
2075 if (peer_ni->ibp_next_conn == conn)
2076 /* clear next_conn so it won't be used */
2077 peer_ni->ibp_next_conn = NULL;
2078 list_del(&conn->ibc_list);
2079 /* connd (see below) takes over ibc_list's ref */
2081 if (list_empty(&peer_ni->ibp_conns) && /* no more conns */
2082 kiblnd_peer_active(peer_ni)) { /* still in peer_ni table */
2083 kiblnd_unlink_peer_locked(peer_ni);
2085 /* set/clear error on last conn */
2086 peer_ni->ibp_error = conn->ibc_comms_error;
2089 kiblnd_set_conn_state(conn, IBLND_CONN_CLOSING);
2092 kiblnd_dev_can_failover(dev)) {
2093 list_add_tail(&dev->ibd_fail_list,
2094 &kiblnd_data.kib_failed_devs);
2095 wake_up(&kiblnd_data.kib_failover_waitq);
2098 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
2100 list_add_tail(&conn->ibc_list, &kiblnd_data.kib_connd_conns);
2101 wake_up(&kiblnd_data.kib_connd_waitq);
2103 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
2107 kiblnd_close_conn(struct kib_conn *conn, int error)
2109 unsigned long flags;
2111 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2113 kiblnd_close_conn_locked(conn, error);
2115 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2119 kiblnd_handle_early_rxs(struct kib_conn *conn)
2121 unsigned long flags;
2124 LASSERT(!in_interrupt());
2125 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
2127 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2128 while ((rx = list_first_entry_or_null(&conn->ibc_early_rxs,
2130 rx_list)) != NULL) {
2131 list_del(&rx->rx_list);
2132 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2134 kiblnd_handle_rx(rx);
2136 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2138 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2142 kiblnd_abort_txs(struct kib_conn *conn, struct list_head *txs)
2148 spin_lock(&conn->ibc_lock);
2150 list_for_each_entry_safe(tx, nxt, txs, tx_list) {
2151 if (txs == &conn->ibc_active_txs) {
2152 LASSERT(!tx->tx_queued);
2153 LASSERT(tx->tx_waiting ||
2154 tx->tx_sending != 0);
2155 if (conn->ibc_comms_error == -ETIMEDOUT) {
2156 if (tx->tx_waiting && !tx->tx_sending)
2158 LNET_MSG_STATUS_REMOTE_TIMEOUT;
2159 else if (tx->tx_sending)
2161 LNET_MSG_STATUS_NETWORK_TIMEOUT;
2164 LASSERT(tx->tx_queued);
2165 if (conn->ibc_comms_error == -ETIMEDOUT)
2166 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_TIMEOUT;
2168 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
2171 tx->tx_status = -ECONNABORTED;
2175 * TODO: This makes an assumption that
2176 * kiblnd_tx_complete() will be called for each tx. If
2177 * that event is dropped we could end up with stale
2178 * connections floating around. We'd like to deal with
2179 * that in a better way.
2181 * Also that means we can exceed the timeout by many
2184 if (tx->tx_sending == 0) {
2186 list_move(&tx->tx_list, &zombies);
2188 /* keep tx until cq destroy */
2189 list_move(&tx->tx_list, &conn->ibc_zombie_txs);
2194 spin_unlock(&conn->ibc_lock);
2197 * aborting transmits occurs when finalizing the connection.
2198 * The connection is finalized on error.
2199 * Passing LNET_MSG_STATUS_OK to txlist_done() will not
2200 * override the value already set in tx->tx_hstatus above.
2202 kiblnd_txlist_done(&zombies, -ECONNABORTED, LNET_MSG_STATUS_OK);
2206 kiblnd_tx_may_discard(struct kib_conn *conn)
2212 spin_lock(&conn->ibc_lock);
2214 list_for_each_entry_safe(tx, nxt, &conn->ibc_zombie_txs, tx_list) {
2215 if (tx->tx_sending > 0 && tx->tx_lntmsg[0] &&
2216 lnet_md_discarded(tx->tx_lntmsg[0]->msg_md)) {
2218 if (tx->tx_sending == 0) {
2219 kiblnd_conn_decref(tx->tx_conn);
2226 spin_unlock(&conn->ibc_lock);
2231 kiblnd_finalise_conn(struct kib_conn *conn)
2233 LASSERT (!in_interrupt());
2234 LASSERT (conn->ibc_state > IBLND_CONN_INIT);
2236 /* abort_receives moves QP state to IB_QPS_ERR. This is only required
2237 * for connections that didn't get as far as being connected, because
2238 * rdma_disconnect() does this for free. */
2239 kiblnd_abort_receives(conn);
2241 kiblnd_set_conn_state(conn, IBLND_CONN_DISCONNECTED);
2243 /* Complete all tx descs not waiting for sends to complete.
2244 * NB we should be safe from RDMA now that the QP has changed state */
2246 kiblnd_abort_txs(conn, &conn->ibc_tx_noops);
2247 kiblnd_abort_txs(conn, &conn->ibc_tx_queue);
2248 kiblnd_abort_txs(conn, &conn->ibc_tx_queue_rsrvd);
2249 kiblnd_abort_txs(conn, &conn->ibc_tx_queue_nocred);
2250 kiblnd_abort_txs(conn, &conn->ibc_active_txs);
2252 kiblnd_handle_early_rxs(conn);
2256 kiblnd_peer_connect_failed(struct kib_peer_ni *peer_ni, int active,
2260 unsigned long flags;
2261 enum lnet_msg_hstatus hstatus;
2263 LASSERT(error != 0);
2264 LASSERT(!in_interrupt());
2266 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2269 LASSERT(peer_ni->ibp_connecting > 0);
2270 peer_ni->ibp_connecting--;
2272 LASSERT (peer_ni->ibp_accepting > 0);
2273 peer_ni->ibp_accepting--;
2276 if (kiblnd_peer_connecting(peer_ni)) {
2277 /* another connection attempt under way... */
2278 write_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2283 peer_ni->ibp_reconnected = 0;
2284 if (list_empty(&peer_ni->ibp_conns)) {
2285 /* Take peer_ni's blocked transmits to complete with error */
2286 list_splice_init(&peer_ni->ibp_tx_queue, &zombies);
2288 if (kiblnd_peer_active(peer_ni))
2289 kiblnd_unlink_peer_locked(peer_ni);
2291 peer_ni->ibp_error = error;
2293 /* Can't have blocked transmits if there are connections */
2294 LASSERT(list_empty(&peer_ni->ibp_tx_queue));
2297 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2299 kiblnd_peer_notify(peer_ni);
2301 if (list_empty(&zombies))
2304 CNETERR("Deleting messages for %s: connection failed\n",
2305 libcfs_nid2str(peer_ni->ibp_nid));
2310 hstatus = LNET_MSG_STATUS_NETWORK_TIMEOUT;
2313 hstatus = LNET_MSG_STATUS_REMOTE_DROPPED;
2316 hstatus = LNET_MSG_STATUS_LOCAL_DROPPED;
2320 kiblnd_txlist_done(&zombies, error, hstatus);
2324 kiblnd_connreq_done(struct kib_conn *conn, int status)
2326 struct kib_peer_ni *peer_ni = conn->ibc_peer;
2329 unsigned long flags;
2332 active = (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2334 CDEBUG(D_NET,"%s: active(%d), version(%x), status(%d)\n",
2335 libcfs_nid2str(peer_ni->ibp_nid), active,
2336 conn->ibc_version, status);
2338 LASSERT (!in_interrupt());
2339 LASSERT ((conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT &&
2340 peer_ni->ibp_connecting > 0) ||
2341 (conn->ibc_state == IBLND_CONN_PASSIVE_WAIT &&
2342 peer_ni->ibp_accepting > 0));
2344 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
2345 conn->ibc_connvars = NULL;
2348 /* failed to establish connection */
2349 kiblnd_peer_connect_failed(peer_ni, active, status);
2350 kiblnd_finalise_conn(conn);
2354 /* connection established */
2355 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2357 conn->ibc_last_send = ktime_get();
2358 kiblnd_set_conn_state(conn, IBLND_CONN_ESTABLISHED);
2359 kiblnd_peer_alive(peer_ni);
2361 /* Add conn to peer_ni's list and nuke any dangling conns from a different
2362 * peer_ni instance... */
2363 kiblnd_conn_addref(conn); /* +1 ref for ibc_list */
2364 list_add(&conn->ibc_list, &peer_ni->ibp_conns);
2365 peer_ni->ibp_reconnected = 0;
2367 peer_ni->ibp_connecting--;
2369 peer_ni->ibp_accepting--;
2371 if (peer_ni->ibp_version == 0) {
2372 peer_ni->ibp_version = conn->ibc_version;
2373 peer_ni->ibp_incarnation = conn->ibc_incarnation;
2376 if (peer_ni->ibp_version != conn->ibc_version ||
2377 peer_ni->ibp_incarnation != conn->ibc_incarnation) {
2378 kiblnd_close_stale_conns_locked(peer_ni, conn->ibc_version,
2379 conn->ibc_incarnation);
2380 peer_ni->ibp_version = conn->ibc_version;
2381 peer_ni->ibp_incarnation = conn->ibc_incarnation;
2384 /* grab pending txs while I have the lock */
2385 list_splice_init(&peer_ni->ibp_tx_queue, &txs);
2387 if (!kiblnd_peer_active(peer_ni) || /* peer_ni has been deleted */
2388 conn->ibc_comms_error != 0) { /* error has happened already */
2390 /* start to shut down connection */
2391 kiblnd_close_conn_locked(conn, -ECONNABORTED);
2392 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2394 kiblnd_txlist_done(&txs, -ECONNABORTED,
2395 LNET_MSG_STATUS_LOCAL_ERROR);
2400 /* +1 ref for myself, this connection is visible to other threads
2401 * now, refcount of peer:ibp_conns can be released by connection
2402 * close from either a different thread, or the calling of
2403 * kiblnd_check_sends_locked() below. See bz21911 for details.
2405 kiblnd_conn_addref(conn);
2406 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2408 /* Schedule blocked txs
2409 * Note: if we are running with conns_per_peer > 1, these blocked
2410 * txs will all get scheduled to the first connection which gets
2411 * scheduled. We won't be using round robin on this first batch.
2413 spin_lock(&conn->ibc_lock);
2414 while ((tx = list_first_entry_or_null(&txs, struct kib_tx,
2415 tx_list)) != NULL) {
2416 list_del(&tx->tx_list);
2418 kiblnd_queue_tx_locked(tx, conn);
2420 kiblnd_check_sends_locked(conn);
2421 spin_unlock(&conn->ibc_lock);
2423 /* schedule blocked rxs */
2424 kiblnd_handle_early_rxs(conn);
2425 kiblnd_conn_decref(conn);
2429 kiblnd_reject(struct rdma_cm_id *cmid, struct kib_rej *rej)
2433 #ifdef HAVE_RDMA_REJECT_4ARGS
2434 rc = rdma_reject(cmid, rej, sizeof(*rej), IB_CM_REJ_CONSUMER_DEFINED);
2436 rc = rdma_reject(cmid, rej, sizeof(*rej));
2440 CWARN("Error %d sending reject\n", rc);
2444 kiblnd_passive_connect(struct rdma_cm_id *cmid, void *priv, int priv_nob)
2446 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2447 struct kib_msg *reqmsg = priv;
2448 struct kib_msg *ackmsg;
2449 struct kib_dev *ibdev;
2450 struct kib_peer_ni *peer_ni;
2451 struct kib_peer_ni *peer2;
2452 struct kib_conn *conn;
2453 struct lnet_ni *ni = NULL;
2454 struct kib_net *net = NULL;
2456 struct rdma_conn_param cp;
2458 int version = IBLND_MSG_VERSION;
2459 unsigned long flags;
2461 struct sockaddr_in *peer_addr;
2463 LASSERT(!in_interrupt());
2464 /* cmid inherits 'context' from the corresponding listener id */
2465 ibdev = cmid->context;
2468 memset(&rej, 0, sizeof(rej));
2469 rej.ibr_magic = IBLND_MSG_MAGIC;
2470 rej.ibr_why = IBLND_REJECT_FATAL;
2471 rej.ibr_cp.ibcp_max_msg_size = IBLND_MSG_SIZE;
2473 peer_addr = (struct sockaddr_in *)&(cmid->route.addr.dst_addr);
2474 if (*kiblnd_tunables.kib_require_priv_port &&
2475 ntohs(peer_addr->sin_port) >= PROT_SOCK) {
2476 __u32 ip = ntohl(peer_addr->sin_addr.s_addr);
2477 CERROR("peer_ni's port (%pI4h:%hu) is not privileged\n",
2478 &ip, ntohs(peer_addr->sin_port));
2482 if (priv_nob < offsetof(struct kib_msg, ibm_type)) {
2483 CERROR("Short connection request\n");
2487 /* Future protocol version compatibility support! If the
2488 * o2iblnd-specific protocol changes, or when LNET unifies
2489 * protocols over all LNDs, the initial connection will
2490 * negotiate a protocol version. I trap this here to avoid
2491 * console errors; the reject tells the peer_ni which protocol I
2493 if (reqmsg->ibm_magic == LNET_PROTO_MAGIC ||
2494 reqmsg->ibm_magic == __swab32(LNET_PROTO_MAGIC))
2496 if (reqmsg->ibm_magic == IBLND_MSG_MAGIC &&
2497 reqmsg->ibm_version != IBLND_MSG_VERSION &&
2498 reqmsg->ibm_version != IBLND_MSG_VERSION_1)
2500 if (reqmsg->ibm_magic == __swab32(IBLND_MSG_MAGIC) &&
2501 reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION) &&
2502 reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION_1))
2505 rc = kiblnd_unpack_msg(reqmsg, priv_nob);
2507 CERROR("Can't parse connection request: %d\n", rc);
2511 nid = reqmsg->ibm_srcnid;
2512 ni = lnet_nid2ni_addref(reqmsg->ibm_dstnid);
2515 net = (struct kib_net *)ni->ni_data;
2516 rej.ibr_incarnation = net->ibn_incarnation;
2519 if (ni == NULL || /* no matching net */
2520 lnet_nid_to_nid4(&ni->ni_nid) !=
2521 reqmsg->ibm_dstnid || /* right NET, wrong NID! */
2522 net->ibn_dev != ibdev) { /* wrong device */
2523 CERROR("Can't accept conn from %s on %s (%s:%d:%pI4h): bad dst nid %s\n", libcfs_nid2str(nid),
2524 ni ? libcfs_nidstr(&ni->ni_nid) : "NA",
2525 ibdev->ibd_ifname, ibdev->ibd_nnets,
2527 libcfs_nid2str(reqmsg->ibm_dstnid));
2532 /* check time stamp as soon as possible */
2533 if (reqmsg->ibm_dststamp != 0 &&
2534 reqmsg->ibm_dststamp != net->ibn_incarnation) {
2535 CWARN("Stale connection request\n");
2536 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2540 /* I can accept peer_ni's version */
2541 version = reqmsg->ibm_version;
2543 if (reqmsg->ibm_type != IBLND_MSG_CONNREQ) {
2544 CERROR("Unexpected connreq msg type: %x from %s\n",
2545 reqmsg->ibm_type, libcfs_nid2str(nid));
2549 if (reqmsg->ibm_u.connparams.ibcp_queue_depth >
2550 kiblnd_msg_queue_size(version, ni)) {
2551 CERROR("Can't accept conn from %s, queue depth too large: %d (<=%d wanted)\n",
2552 libcfs_nid2str(nid),
2553 reqmsg->ibm_u.connparams.ibcp_queue_depth,
2554 kiblnd_msg_queue_size(version, ni));
2556 if (version == IBLND_MSG_VERSION)
2557 rej.ibr_why = IBLND_REJECT_MSG_QUEUE_SIZE;
2562 if (reqmsg->ibm_u.connparams.ibcp_max_frags >
2563 IBLND_MAX_RDMA_FRAGS) {
2564 CWARN("Can't accept conn from %s (version %x): max_frags %d too large (%d wanted)\n",
2565 libcfs_nid2str(nid), version,
2566 reqmsg->ibm_u.connparams.ibcp_max_frags,
2567 IBLND_MAX_RDMA_FRAGS);
2569 if (version >= IBLND_MSG_VERSION)
2570 rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2573 } else if (reqmsg->ibm_u.connparams.ibcp_max_frags <
2574 IBLND_MAX_RDMA_FRAGS &&
2575 net->ibn_fmr_ps == NULL) {
2576 CWARN("Can't accept conn from %s (version %x): max_frags %d incompatible without FMR pool (%d wanted)\n",
2577 libcfs_nid2str(nid), version,
2578 reqmsg->ibm_u.connparams.ibcp_max_frags,
2579 IBLND_MAX_RDMA_FRAGS);
2581 if (version == IBLND_MSG_VERSION)
2582 rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2587 if (reqmsg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2588 CERROR("Can't accept %s: message size %d too big (%d max)\n",
2589 libcfs_nid2str(nid),
2590 reqmsg->ibm_u.connparams.ibcp_max_msg_size,
2595 /* assume 'nid' is a new peer_ni; create */
2596 rc = kiblnd_create_peer(ni, &peer_ni, nid);
2598 CERROR("Can't create peer_ni for %s\n", libcfs_nid2str(nid));
2599 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2603 /* We have validated the peer's parameters so use those */
2604 peer_ni->ibp_max_frags = reqmsg->ibm_u.connparams.ibcp_max_frags;
2605 peer_ni->ibp_queue_depth = reqmsg->ibm_u.connparams.ibcp_queue_depth;
2607 write_lock_irqsave(g_lock, flags);
2609 peer2 = kiblnd_find_peer_locked(ni, nid);
2610 if (peer2 != NULL) {
2611 if (peer2->ibp_version == 0) {
2612 peer2->ibp_version = version;
2613 peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2616 /* not the guy I've talked with */
2617 if (peer2->ibp_incarnation != reqmsg->ibm_srcstamp ||
2618 peer2->ibp_version != version) {
2619 kiblnd_close_peer_conns_locked(peer2, -ESTALE);
2621 if (kiblnd_peer_active(peer2)) {
2622 peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2623 peer2->ibp_version = version;
2625 write_unlock_irqrestore(g_lock, flags);
2627 CWARN("Conn stale %s version %x/%x incarnation %llu/%llu\n",
2628 libcfs_nid2str(nid), peer2->ibp_version, version,
2629 peer2->ibp_incarnation, reqmsg->ibm_srcstamp);
2631 kiblnd_peer_decref(peer_ni);
2632 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2636 /* Tie-break connection race in favour of the higher NID.
2637 * If we keep running into a race condition multiple times,
2638 * we have to assume that the connection attempt with the
2639 * higher NID is stuck in a connecting state and will never
2640 * recover. As such, we pass through this if-block and let
2641 * the lower NID connection win so we can move forward.
2643 if (peer2->ibp_connecting != 0 &&
2644 nid < lnet_nid_to_nid4(&ni->ni_nid) &&
2645 peer2->ibp_races < MAX_CONN_RACES_BEFORE_ABORT) {
2647 write_unlock_irqrestore(g_lock, flags);
2649 CDEBUG(D_NET, "Conn race %s\n",
2650 libcfs_nid2str(peer2->ibp_nid));
2652 kiblnd_peer_decref(peer_ni);
2653 rej.ibr_why = IBLND_REJECT_CONN_RACE;
2656 if (peer2->ibp_races >= MAX_CONN_RACES_BEFORE_ABORT)
2657 CNETERR("Conn race %s: unresolved after %d attempts, letting lower NID win\n",
2658 libcfs_nid2str(peer2->ibp_nid),
2659 MAX_CONN_RACES_BEFORE_ABORT);
2661 * passive connection is allowed even this peer_ni is waiting for
2664 peer2->ibp_reconnecting = 0;
2665 peer2->ibp_races = 0;
2666 peer2->ibp_accepting++;
2667 kiblnd_peer_addref(peer2);
2669 /* Race with kiblnd_launch_tx (active connect) to create peer_ni
2670 * so copy validated parameters since we now know what the
2671 * peer_ni's limits are */
2672 peer2->ibp_max_frags = peer_ni->ibp_max_frags;
2673 peer2->ibp_queue_depth = peer_ni->ibp_queue_depth;
2675 write_unlock_irqrestore(g_lock, flags);
2676 kiblnd_peer_decref(peer_ni);
2679 /* Brand new peer_ni */
2680 LASSERT(peer_ni->ibp_accepting == 0);
2681 LASSERT(peer_ni->ibp_version == 0 &&
2682 peer_ni->ibp_incarnation == 0);
2684 peer_ni->ibp_accepting = 1;
2685 peer_ni->ibp_version = version;
2686 peer_ni->ibp_incarnation = reqmsg->ibm_srcstamp;
2688 /* I have a ref on ni that prevents it being shutdown */
2689 LASSERT(net->ibn_shutdown == 0);
2691 kiblnd_peer_addref(peer_ni);
2692 hash_add(kiblnd_data.kib_peers, &peer_ni->ibp_list, nid);
2694 write_unlock_irqrestore(g_lock, flags);
2697 conn = kiblnd_create_conn(peer_ni, cmid, IBLND_CONN_PASSIVE_WAIT,
2700 kiblnd_peer_connect_failed(peer_ni, 0, -ENOMEM);
2701 kiblnd_peer_decref(peer_ni);
2702 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2706 /* conn now "owns" cmid, so I return success from here on to ensure the
2707 * CM callback doesn't destroy cmid.
2709 conn->ibc_incarnation = reqmsg->ibm_srcstamp;
2710 conn->ibc_credits = conn->ibc_queue_depth;
2711 conn->ibc_reserved_credits = conn->ibc_queue_depth;
2712 LASSERT(conn->ibc_credits + conn->ibc_reserved_credits +
2713 IBLND_OOB_MSGS(version) <= IBLND_RX_MSGS(conn));
2715 ackmsg = &conn->ibc_connvars->cv_msg;
2716 memset(ackmsg, 0, sizeof(*ackmsg));
2718 kiblnd_init_msg(ackmsg, IBLND_MSG_CONNACK,
2719 sizeof(ackmsg->ibm_u.connparams));
2720 ackmsg->ibm_u.connparams.ibcp_queue_depth = conn->ibc_queue_depth;
2721 ackmsg->ibm_u.connparams.ibcp_max_frags = conn->ibc_max_frags;
2722 ackmsg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2724 kiblnd_pack_msg(ni, ackmsg, version, 0, nid, reqmsg->ibm_srcstamp);
2726 memset(&cp, 0, sizeof(cp));
2727 cp.private_data = ackmsg;
2728 cp.private_data_len = ackmsg->ibm_nob;
2729 cp.responder_resources = 0; /* No atomic ops or RDMA reads */
2730 cp.initiator_depth = 0;
2731 cp.flow_control = 1;
2732 cp.retry_count = *kiblnd_tunables.kib_retry_count;
2733 cp.rnr_retry_count = *kiblnd_tunables.kib_rnr_retry_count;
2735 CDEBUG(D_NET, "Accept %s\n", libcfs_nid2str(nid));
2737 rc = rdma_accept(cmid, &cp);
2739 CERROR("Can't accept %s: %d\n", libcfs_nid2str(nid), rc);
2740 rej.ibr_version = version;
2741 rej.ibr_why = IBLND_REJECT_FATAL;
2743 kiblnd_reject(cmid, &rej);
2744 kiblnd_connreq_done(conn, rc);
2745 kiblnd_conn_decref(conn);
2753 rej.ibr_cp.ibcp_queue_depth =
2754 kiblnd_msg_queue_size(version, ni);
2755 rej.ibr_cp.ibcp_max_frags = IBLND_MAX_RDMA_FRAGS;
2759 rej.ibr_version = version;
2760 kiblnd_reject(cmid, &rej);
2762 return -ECONNREFUSED;
2766 kiblnd_check_reconnect(struct kib_conn *conn, int version,
2767 u64 incarnation, int why, struct kib_connparams *cp)
2769 rwlock_t *glock = &kiblnd_data.kib_global_lock;
2770 struct kib_peer_ni *peer_ni = conn->ibc_peer;
2772 int msg_size = IBLND_MSG_SIZE;
2776 unsigned long flags;
2778 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2779 LASSERT(peer_ni->ibp_connecting > 0); /* 'conn' at least */
2782 msg_size = cp->ibcp_max_msg_size;
2783 frag_num = cp->ibcp_max_frags;
2784 queue_dep = cp->ibcp_queue_depth;
2787 write_lock_irqsave(glock, flags);
2788 /* retry connection if it's still needed and no other connection
2789 * attempts (active or passive) are in progress
2790 * NB: reconnect is still needed even when ibp_tx_queue is
2791 * empty if ibp_version != version because reconnect may be
2794 reconnect = (!list_empty(&peer_ni->ibp_tx_queue) ||
2795 peer_ni->ibp_version != version) &&
2796 peer_ni->ibp_connecting &&
2797 peer_ni->ibp_accepting == 0;
2808 case IBLND_REJECT_RDMA_FRAGS: {
2810 reason = "can't negotiate max frags";
2814 if (conn->ibc_max_frags <= frag_num) {
2815 reason = "unsupported max frags";
2819 peer_ni->ibp_max_frags = frag_num;
2820 reason = "rdma fragments";
2823 case IBLND_REJECT_MSG_QUEUE_SIZE:
2825 reason = "can't negotiate queue depth";
2828 if (conn->ibc_queue_depth <= queue_dep) {
2829 reason = "unsupported queue depth";
2833 peer_ni->ibp_queue_depth = queue_dep;
2834 reason = "queue depth";
2837 case IBLND_REJECT_CONN_STALE:
2841 case IBLND_REJECT_CONN_RACE:
2842 reason = "conn race";
2845 case IBLND_REJECT_CONN_UNCOMPAT:
2846 reason = "version negotiation";
2850 conn->ibc_reconnect = 1;
2851 peer_ni->ibp_reconnecting++;
2852 peer_ni->ibp_version = version;
2853 if (incarnation != 0)
2854 peer_ni->ibp_incarnation = incarnation;
2856 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2858 CNETERR("%s: %s (%s), %x, %x, msg_size: %d, queue_depth: %d/%d, max_frags: %d/%d\n",
2859 libcfs_nid2str(peer_ni->ibp_nid),
2860 reconnect ? "reconnect" : "don't reconnect",
2861 reason, IBLND_MSG_VERSION, version, msg_size,
2862 conn->ibc_queue_depth, queue_dep,
2863 conn->ibc_max_frags, frag_num);
2865 * if conn::ibc_reconnect is TRUE, connd will reconnect to the peer_ni
2866 * while destroying the zombie
2871 kiblnd_rejected(struct kib_conn *conn, int reason, void *priv, int priv_nob)
2873 struct kib_peer_ni *peer_ni = conn->ibc_peer;
2874 int status = -ECONNREFUSED;
2876 LASSERT (!in_interrupt());
2877 LASSERT (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2880 case IB_CM_REJ_STALE_CONN:
2881 kiblnd_check_reconnect(conn, IBLND_MSG_VERSION, 0,
2882 IBLND_REJECT_CONN_STALE, NULL);
2885 case IB_CM_REJ_INVALID_SERVICE_ID:
2886 status = -EHOSTUNREACH;
2887 CNETERR("%s rejected: no listener at %d\n",
2888 libcfs_nid2str(peer_ni->ibp_nid),
2889 *kiblnd_tunables.kib_service);
2892 case IB_CM_REJ_CONSUMER_DEFINED:
2893 if (priv_nob >= offsetof(struct kib_rej, ibr_padding)) {
2894 struct kib_rej *rej = priv;
2895 struct kib_connparams *cp = NULL;
2897 __u64 incarnation = -1;
2899 /* NB. default incarnation is -1 because:
2900 * a) V1 will ignore dst incarnation in connreq.
2901 * b) V2 will provide incarnation while rejecting me,
2902 * -1 will be overwrote.
2904 * if I try to connect to a V1 peer_ni with V2 protocol,
2905 * it rejected me then upgrade to V2, I have no idea
2906 * about the upgrading and try to reconnect with V1,
2907 * in this case upgraded V2 can find out I'm trying to
2908 * talk to the old guy and reject me(incarnation is -1).
2911 if (rej->ibr_magic == __swab32(IBLND_MSG_MAGIC) ||
2912 rej->ibr_magic == __swab32(LNET_PROTO_MAGIC)) {
2913 __swab32s(&rej->ibr_magic);
2914 __swab16s(&rej->ibr_version);
2918 if (priv_nob >= sizeof(struct kib_rej) &&
2919 rej->ibr_version > IBLND_MSG_VERSION_1) {
2920 /* priv_nob is always 148 in current version
2921 * of OFED, so we still need to check version.
2922 * (define of IB_CM_REJ_PRIVATE_DATA_SIZE)
2927 __swab64s(&rej->ibr_incarnation);
2928 __swab16s(&cp->ibcp_queue_depth);
2929 __swab16s(&cp->ibcp_max_frags);
2930 __swab32s(&cp->ibcp_max_msg_size);
2933 incarnation = rej->ibr_incarnation;
2936 if (rej->ibr_magic != IBLND_MSG_MAGIC &&
2937 rej->ibr_magic != LNET_PROTO_MAGIC) {
2938 CERROR("%s rejected: consumer defined fatal error\n",
2939 libcfs_nid2str(peer_ni->ibp_nid));
2943 if (rej->ibr_version != IBLND_MSG_VERSION &&
2944 rej->ibr_version != IBLND_MSG_VERSION_1) {
2945 CERROR("%s rejected: o2iblnd version %x error\n",
2946 libcfs_nid2str(peer_ni->ibp_nid),
2951 if (rej->ibr_why == IBLND_REJECT_FATAL &&
2952 rej->ibr_version == IBLND_MSG_VERSION_1) {
2953 CDEBUG(D_NET, "rejected by old version peer_ni %s: %x\n",
2954 libcfs_nid2str(peer_ni->ibp_nid),
2957 if (conn->ibc_version != IBLND_MSG_VERSION_1)
2958 rej->ibr_why = IBLND_REJECT_CONN_UNCOMPAT;
2961 switch (rej->ibr_why) {
2962 case IBLND_REJECT_CONN_RACE:
2963 case IBLND_REJECT_CONN_STALE:
2964 case IBLND_REJECT_CONN_UNCOMPAT:
2965 case IBLND_REJECT_MSG_QUEUE_SIZE:
2966 case IBLND_REJECT_RDMA_FRAGS:
2967 kiblnd_check_reconnect(conn, rej->ibr_version,
2972 case IBLND_REJECT_NO_RESOURCES:
2973 CERROR("%s rejected: o2iblnd no resources\n",
2974 libcfs_nid2str(peer_ni->ibp_nid));
2977 case IBLND_REJECT_FATAL:
2978 CERROR("%s rejected: o2iblnd fatal error\n",
2979 libcfs_nid2str(peer_ni->ibp_nid));
2983 CERROR("%s rejected: o2iblnd reason %d\n",
2984 libcfs_nid2str(peer_ni->ibp_nid),
2992 CNETERR("%s rejected: reason %d, size %d\n",
2993 libcfs_nid2str(peer_ni->ibp_nid), reason, priv_nob);
2997 kiblnd_connreq_done(conn, status);
3001 kiblnd_check_connreply(struct kib_conn *conn, void *priv, int priv_nob)
3003 struct kib_peer_ni *peer_ni = conn->ibc_peer;
3004 struct lnet_ni *ni = peer_ni->ibp_ni;
3005 struct kib_net *net = ni->ni_data;
3006 struct kib_msg *msg = priv;
3007 int ver = conn->ibc_version;
3008 int rc = kiblnd_unpack_msg(msg, priv_nob);
3009 unsigned long flags;
3011 LASSERT (net != NULL);
3014 CERROR("Can't unpack connack from %s: %d\n",
3015 libcfs_nid2str(peer_ni->ibp_nid), rc);
3019 if (msg->ibm_type != IBLND_MSG_CONNACK) {
3020 CERROR("Unexpected message %d from %s\n",
3021 msg->ibm_type, libcfs_nid2str(peer_ni->ibp_nid));
3026 if (ver != msg->ibm_version) {
3027 CERROR("%s replied version %x is different with "
3028 "requested version %x\n",
3029 libcfs_nid2str(peer_ni->ibp_nid), msg->ibm_version, ver);
3034 if (msg->ibm_u.connparams.ibcp_queue_depth >
3035 conn->ibc_queue_depth) {
3036 CERROR("%s has incompatible queue depth %d (<=%d wanted)\n",
3037 libcfs_nid2str(peer_ni->ibp_nid),
3038 msg->ibm_u.connparams.ibcp_queue_depth,
3039 conn->ibc_queue_depth);
3044 if (msg->ibm_u.connparams.ibcp_max_frags >
3045 conn->ibc_max_frags) {
3046 CERROR("%s has incompatible max_frags %d (<=%d wanted)\n",
3047 libcfs_nid2str(peer_ni->ibp_nid),
3048 msg->ibm_u.connparams.ibcp_max_frags,
3049 conn->ibc_max_frags);
3054 if (msg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
3055 CERROR("%s max message size %d too big (%d max)\n",
3056 libcfs_nid2str(peer_ni->ibp_nid),
3057 msg->ibm_u.connparams.ibcp_max_msg_size,
3063 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3064 if (msg->ibm_dstnid == lnet_nid_to_nid4(&ni->ni_nid) &&
3065 msg->ibm_dststamp == net->ibn_incarnation)
3069 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3072 CERROR("Bad connection reply from %s, rc = %d, "
3073 "version: %x max_frags: %d\n",
3074 libcfs_nid2str(peer_ni->ibp_nid), rc,
3075 msg->ibm_version, msg->ibm_u.connparams.ibcp_max_frags);
3079 conn->ibc_incarnation = msg->ibm_srcstamp;
3080 conn->ibc_credits = msg->ibm_u.connparams.ibcp_queue_depth;
3081 conn->ibc_reserved_credits = msg->ibm_u.connparams.ibcp_queue_depth;
3082 conn->ibc_queue_depth = msg->ibm_u.connparams.ibcp_queue_depth;
3083 conn->ibc_max_frags = msg->ibm_u.connparams.ibcp_max_frags;
3084 LASSERT(conn->ibc_credits + conn->ibc_reserved_credits +
3085 IBLND_OOB_MSGS(ver) <= IBLND_RX_MSGS(conn));
3087 kiblnd_connreq_done(conn, 0);
3091 /* NB My QP has already established itself, so I handle anything going
3092 * wrong here by setting ibc_comms_error.
3093 * kiblnd_connreq_done(0) moves the conn state to ESTABLISHED, but then
3094 * immediately tears it down. */
3097 conn->ibc_comms_error = rc;
3098 kiblnd_connreq_done(conn, 0);
3102 kiblnd_active_connect(struct rdma_cm_id *cmid)
3104 struct kib_peer_ni *peer_ni = cmid->context;
3105 struct kib_conn *conn;
3106 struct kib_msg *msg;
3107 struct rdma_conn_param cp;
3110 unsigned long flags;
3113 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3115 incarnation = peer_ni->ibp_incarnation;
3116 version = (peer_ni->ibp_version == 0) ? IBLND_MSG_VERSION :
3117 peer_ni->ibp_version;
3119 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3121 conn = kiblnd_create_conn(peer_ni, cmid, IBLND_CONN_ACTIVE_CONNECT,
3124 kiblnd_peer_connect_failed(peer_ni, 1, -ENOMEM);
3125 kiblnd_peer_decref(peer_ni); /* lose cmid's ref */
3129 /* conn "owns" cmid now, so I return success from here on to ensure the
3130 * CM callback doesn't destroy cmid. conn also takes over cmid's ref
3133 msg = &conn->ibc_connvars->cv_msg;
3135 memset(msg, 0, sizeof(*msg));
3136 kiblnd_init_msg(msg, IBLND_MSG_CONNREQ, sizeof(msg->ibm_u.connparams));
3137 msg->ibm_u.connparams.ibcp_queue_depth = conn->ibc_queue_depth;
3138 msg->ibm_u.connparams.ibcp_max_frags = conn->ibc_max_frags;
3139 msg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
3141 kiblnd_pack_msg(peer_ni->ibp_ni, msg, version,
3142 0, peer_ni->ibp_nid, incarnation);
3144 memset(&cp, 0, sizeof(cp));
3145 cp.private_data = msg;
3146 cp.private_data_len = msg->ibm_nob;
3147 cp.responder_resources = 0; /* No atomic ops or RDMA reads */
3148 cp.initiator_depth = 0;
3149 cp.flow_control = 1;
3150 cp.retry_count = *kiblnd_tunables.kib_retry_count;
3151 cp.rnr_retry_count = *kiblnd_tunables.kib_rnr_retry_count;
3153 LASSERT(cmid->context == (void *)conn);
3154 LASSERT(conn->ibc_cmid == cmid);
3155 rc = rdma_connect_locked(cmid, &cp);
3157 CERROR("Can't connect to %s: %d\n",
3158 libcfs_nid2str(peer_ni->ibp_nid), rc);
3159 kiblnd_connreq_done(conn, rc);
3160 kiblnd_conn_decref(conn);
3167 kiblnd_cm_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
3169 struct kib_peer_ni *peer_ni;
3170 struct kib_conn *conn;
3173 switch (event->event) {
3175 CERROR("Unexpected event: %d, status: %d\n",
3176 event->event, event->status);
3179 case RDMA_CM_EVENT_CONNECT_REQUEST:
3180 /* destroy cmid on failure */
3181 rc = kiblnd_passive_connect(cmid,
3182 (void *)KIBLND_CONN_PARAM(event),
3183 KIBLND_CONN_PARAM_LEN(event));
3184 CDEBUG(D_NET, "connreq: %d\n", rc);
3187 case RDMA_CM_EVENT_ADDR_ERROR:
3188 peer_ni = cmid->context;
3189 CNETERR("%s: ADDR ERROR %d\n",
3190 libcfs_nid2str(peer_ni->ibp_nid), event->status);
3191 kiblnd_peer_connect_failed(peer_ni, 1, -EHOSTUNREACH);
3192 kiblnd_peer_decref(peer_ni);
3193 return -EHOSTUNREACH; /* rc != 0 destroys cmid */
3195 case RDMA_CM_EVENT_ADDR_RESOLVED:
3196 peer_ni = cmid->context;
3198 CDEBUG(D_NET,"%s Addr resolved: %d\n",
3199 libcfs_nid2str(peer_ni->ibp_nid), event->status);
3201 if (event->status != 0) {
3202 CNETERR("Can't resolve address for %s: %d\n",
3203 libcfs_nid2str(peer_ni->ibp_nid), event->status);
3206 rc = rdma_resolve_route(
3207 cmid, kiblnd_timeout() * 1000);
3209 struct kib_net *net = peer_ni->ibp_ni->ni_data;
3210 struct kib_dev *dev = net->ibn_dev;
3212 CDEBUG(D_NET, "%s: connection bound to "\
3214 libcfs_nid2str(peer_ni->ibp_nid),
3216 &dev->ibd_ifip, cmid->device->name);
3221 /* Can't initiate route resolution */
3222 CERROR("Can't resolve route for %s: %d\n",
3223 libcfs_nid2str(peer_ni->ibp_nid), rc);
3225 kiblnd_peer_connect_failed(peer_ni, 1, rc);
3226 kiblnd_peer_decref(peer_ni);
3227 return rc; /* rc != 0 destroys cmid */
3229 case RDMA_CM_EVENT_ROUTE_ERROR:
3230 peer_ni = cmid->context;
3231 CNETERR("%s: ROUTE ERROR %d\n",
3232 libcfs_nid2str(peer_ni->ibp_nid), event->status);
3233 kiblnd_peer_connect_failed(peer_ni, 1, -EHOSTUNREACH);
3234 kiblnd_peer_decref(peer_ni);
3235 return -EHOSTUNREACH; /* rc != 0 destroys cmid */
3237 case RDMA_CM_EVENT_ROUTE_RESOLVED:
3238 peer_ni = cmid->context;
3239 CDEBUG(D_NET,"%s Route resolved: %d\n",
3240 libcfs_nid2str(peer_ni->ibp_nid), event->status);
3242 if (event->status == 0)
3243 return kiblnd_active_connect(cmid);
3245 CNETERR("Can't resolve route for %s: %d\n",
3246 libcfs_nid2str(peer_ni->ibp_nid), event->status);
3247 kiblnd_peer_connect_failed(peer_ni, 1, event->status);
3248 kiblnd_peer_decref(peer_ni);
3249 return event->status; /* rc != 0 destroys cmid */
3251 case RDMA_CM_EVENT_UNREACHABLE:
3252 conn = cmid->context;
3253 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
3254 conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
3255 CNETERR("%s: UNREACHABLE %d\n",
3256 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
3257 kiblnd_connreq_done(conn, -ENETDOWN);
3258 kiblnd_conn_decref(conn);
3261 case RDMA_CM_EVENT_CONNECT_ERROR:
3262 conn = cmid->context;
3263 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
3264 conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
3265 CNETERR("%s: CONNECT ERROR %d\n",
3266 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
3267 kiblnd_connreq_done(conn, -ENOTCONN);
3268 kiblnd_conn_decref(conn);
3271 case RDMA_CM_EVENT_REJECTED:
3272 conn = cmid->context;
3273 switch (conn->ibc_state) {
3277 case IBLND_CONN_PASSIVE_WAIT:
3278 CERROR ("%s: REJECTED %d\n",
3279 libcfs_nid2str(conn->ibc_peer->ibp_nid),
3281 kiblnd_connreq_done(conn, -ECONNRESET);
3284 case IBLND_CONN_ACTIVE_CONNECT:
3285 kiblnd_rejected(conn, event->status,
3286 (void *)KIBLND_CONN_PARAM(event),
3287 KIBLND_CONN_PARAM_LEN(event));
3290 kiblnd_conn_decref(conn);
3293 case RDMA_CM_EVENT_ESTABLISHED:
3294 conn = cmid->context;
3295 switch (conn->ibc_state) {
3299 case IBLND_CONN_PASSIVE_WAIT:
3300 CDEBUG(D_NET, "ESTABLISHED (passive): %s\n",
3301 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3302 kiblnd_connreq_done(conn, 0);
3305 case IBLND_CONN_ACTIVE_CONNECT:
3306 CDEBUG(D_NET, "ESTABLISHED(active): %s\n",
3307 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3308 kiblnd_check_connreply(conn,
3309 (void *)KIBLND_CONN_PARAM(event),
3310 KIBLND_CONN_PARAM_LEN(event));
3313 /* net keeps its ref on conn! */
3316 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
3317 CDEBUG(D_NET, "Ignore TIMEWAIT_EXIT event\n");
3320 case RDMA_CM_EVENT_DISCONNECTED:
3321 conn = cmid->context;
3322 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
3323 CERROR("%s DISCONNECTED\n",
3324 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3325 kiblnd_connreq_done(conn, -ECONNRESET);
3327 kiblnd_close_conn(conn, 0);
3329 kiblnd_conn_decref(conn);
3330 cmid->context = NULL;
3333 case RDMA_CM_EVENT_DEVICE_REMOVAL:
3334 LCONSOLE_ERROR_MSG(0x131,
3335 "Received notification of device removal\n"
3336 "Please shutdown LNET to allow this to proceed\n");
3337 /* Can't remove network from underneath LNET for now, so I have
3341 case RDMA_CM_EVENT_ADDR_CHANGE:
3342 LCONSOLE_INFO("Physical link changed (eg hca/port)\n");
3348 kiblnd_check_txs_locked(struct kib_conn *conn, struct list_head *txs)
3352 list_for_each_entry(tx, txs, tx_list) {
3353 if (txs != &conn->ibc_active_txs) {
3354 LASSERT(tx->tx_queued);
3356 LASSERT(!tx->tx_queued);
3357 LASSERT(tx->tx_waiting || tx->tx_sending != 0);
3360 if (ktime_compare(ktime_get(), tx->tx_deadline) >= 0) {
3361 CERROR("Timed out tx: %s(WSQ:%d%d%d), %lld seconds\n",
3362 kiblnd_queue2str(conn, txs),
3363 tx->tx_waiting, tx->tx_sending, tx->tx_queued,
3365 ktime_ms_delta(ktime_get(),
3366 tx->tx_deadline) / MSEC_PER_SEC);
3375 kiblnd_conn_timed_out_locked(struct kib_conn *conn)
3377 return kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue) ||
3378 kiblnd_check_txs_locked(conn, &conn->ibc_tx_noops) ||
3379 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_rsrvd) ||
3380 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_nocred) ||
3381 kiblnd_check_txs_locked(conn, &conn->ibc_active_txs);
3385 kiblnd_check_conns (int idx)
3388 LIST_HEAD(checksends);
3389 LIST_HEAD(timedout_txs);
3390 struct hlist_head *peers = &kiblnd_data.kib_peers[idx];
3391 struct kib_peer_ni *peer_ni;
3392 struct kib_conn *conn;
3393 struct kib_tx *tx, *tx_tmp;
3394 unsigned long flags;
3396 /* NB. We expect to have a look at all the peers and not find any
3397 * RDMAs to time out, so we just use a shared lock while we
3400 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3402 hlist_for_each_entry(peer_ni, peers, ibp_list) {
3403 /* Check tx_deadline */
3404 list_for_each_entry_safe(tx, tx_tmp, &peer_ni->ibp_tx_queue, tx_list) {
3405 if (ktime_compare(ktime_get(), tx->tx_deadline) >= 0) {
3406 CWARN("Timed out tx for %s: %lld seconds\n",
3407 libcfs_nid2str(peer_ni->ibp_nid),
3408 ktime_ms_delta(ktime_get(),
3409 tx->tx_deadline) / MSEC_PER_SEC);
3410 list_move(&tx->tx_list, &timedout_txs);
3414 list_for_each_entry(conn, &peer_ni->ibp_conns, ibc_list) {
3418 LASSERT(conn->ibc_state == IBLND_CONN_ESTABLISHED);
3420 spin_lock(&conn->ibc_lock);
3422 sendnoop = kiblnd_need_noop(conn);
3423 timedout = kiblnd_conn_timed_out_locked(conn);
3424 if (!sendnoop && !timedout) {
3425 spin_unlock(&conn->ibc_lock);
3430 CERROR("Timed out RDMA with %s (%lld): c: %u, oc: %u, rc: %u\n",
3431 libcfs_nid2str(peer_ni->ibp_nid),
3433 - peer_ni->ibp_last_alive,
3435 conn->ibc_outstanding_credits,
3436 conn->ibc_reserved_credits);
3437 list_add(&conn->ibc_connd_list, &closes);
3439 list_add(&conn->ibc_connd_list, &checksends);
3441 /* +ref for 'closes' or 'checksends' */
3442 kiblnd_conn_addref(conn);
3444 spin_unlock(&conn->ibc_lock);
3448 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3450 if (!list_empty(&timedout_txs))
3451 kiblnd_txlist_done(&timedout_txs, -ETIMEDOUT,
3452 LNET_MSG_STATUS_NETWORK_TIMEOUT);
3454 /* Handle timeout by closing the whole
3455 * connection. We can only be sure RDMA activity
3456 * has ceased once the QP has been modified.
3458 while ((conn = list_first_entry_or_null(&closes,
3460 ibc_connd_list)) != NULL) {
3461 list_del(&conn->ibc_connd_list);
3462 kiblnd_close_conn(conn, -ETIMEDOUT);
3463 kiblnd_conn_decref(conn);
3466 /* In case we have enough credits to return via a
3467 * NOOP, but there were no non-blocking tx descs
3468 * free to do it last time...
3470 while ((conn = list_first_entry_or_null(&checksends,
3472 ibc_connd_list)) != NULL) {
3473 list_del(&conn->ibc_connd_list);
3475 spin_lock(&conn->ibc_lock);
3476 kiblnd_check_sends_locked(conn);
3477 spin_unlock(&conn->ibc_lock);
3479 kiblnd_conn_decref(conn);
3484 kiblnd_disconnect_conn(struct kib_conn *conn)
3486 LASSERT (!in_interrupt());
3487 LASSERT (current == kiblnd_data.kib_connd);
3488 LASSERT (conn->ibc_state == IBLND_CONN_CLOSING);
3490 rdma_disconnect(conn->ibc_cmid);
3491 kiblnd_finalise_conn(conn);
3493 kiblnd_peer_notify(conn->ibc_peer);
3497 * High-water for reconnection to the same peer_ni, reconnection attempt should
3498 * be delayed after trying more than KIB_RECONN_HIGH_RACE.
3500 #define KIB_RECONN_HIGH_RACE 10
3502 * Allow connd to take a break and handle other things after consecutive
3503 * reconnection attemps.
3505 #define KIB_RECONN_BREAK 100
3508 kiblnd_connd (void *arg)
3510 spinlock_t *lock = &kiblnd_data.kib_connd_lock;
3511 wait_queue_entry_t wait;
3512 unsigned long flags;
3513 struct kib_conn *conn;
3518 unsigned long deadline = jiffies;
3521 kiblnd_data.kib_connd = current;
3523 spin_lock_irqsave(lock, flags);
3525 while (!kiblnd_data.kib_shutdown) {
3528 dropped_lock = false;
3530 conn = list_first_entry_or_null(&kiblnd_data.kib_connd_zombies,
3531 struct kib_conn, ibc_list);
3533 struct kib_peer_ni *peer_ni = NULL;
3535 list_del(&conn->ibc_list);
3536 if (conn->ibc_reconnect) {
3537 peer_ni = conn->ibc_peer;
3538 kiblnd_peer_addref(peer_ni);
3541 spin_unlock_irqrestore(lock, flags);
3542 dropped_lock = true;
3544 kiblnd_destroy_conn(conn);
3546 spin_lock_irqsave(lock, flags);
3548 LIBCFS_FREE(conn, sizeof(*conn));
3552 conn->ibc_peer = peer_ni;
3553 if (peer_ni->ibp_reconnected < KIB_RECONN_HIGH_RACE)
3554 list_add_tail(&conn->ibc_list,
3555 &kiblnd_data.kib_reconn_list);
3557 list_add_tail(&conn->ibc_list,
3558 &kiblnd_data.kib_reconn_wait);
3561 conn = list_first_entry_or_null(&kiblnd_data.kib_connd_conns,
3562 struct kib_conn, ibc_list);
3566 list_del(&conn->ibc_list);
3568 spin_unlock_irqrestore(lock, flags);
3569 dropped_lock = true;
3571 kiblnd_disconnect_conn(conn);
3572 wait = conn->ibc_waits;
3573 if (wait == 0) /* keep ref for connd_wait, see below */
3574 kiblnd_conn_decref(conn);
3576 spin_lock_irqsave(lock, flags);
3579 list_add_tail(&conn->ibc_list,
3580 &kiblnd_data.kib_connd_waits);
3583 while (reconn < KIB_RECONN_BREAK) {
3584 if (kiblnd_data.kib_reconn_sec !=
3585 ktime_get_real_seconds()) {
3586 kiblnd_data.kib_reconn_sec = ktime_get_real_seconds();
3587 list_splice_init(&kiblnd_data.kib_reconn_wait,
3588 &kiblnd_data.kib_reconn_list);
3591 conn = list_first_entry_or_null(&kiblnd_data.kib_reconn_list,
3592 struct kib_conn, ibc_list);
3596 list_del(&conn->ibc_list);
3598 spin_unlock_irqrestore(lock, flags);
3599 dropped_lock = true;
3601 reconn += kiblnd_reconnect_peer(conn->ibc_peer);
3602 kiblnd_peer_decref(conn->ibc_peer);
3603 LIBCFS_FREE(conn, sizeof(*conn));
3605 spin_lock_irqsave(lock, flags);
3608 conn = list_first_entry_or_null(&kiblnd_data.kib_connd_waits,
3609 struct kib_conn, ibc_list);
3611 list_del(&conn->ibc_list);
3612 spin_unlock_irqrestore(lock, flags);
3614 dropped_lock = kiblnd_tx_may_discard(conn);
3616 kiblnd_conn_decref(conn);
3618 spin_lock_irqsave(lock, flags);
3620 list_add_tail(&conn->ibc_list,
3621 &kiblnd_data.kib_connd_waits);
3624 /* careful with the jiffy wrap... */
3625 timeout = (int)(deadline - jiffies);
3629 int chunk = HASH_SIZE(kiblnd_data.kib_peers);
3630 unsigned int lnd_timeout;
3632 spin_unlock_irqrestore(lock, flags);
3633 dropped_lock = true;
3635 /* Time to check for RDMA timeouts on a few more
3636 * peers: I do checks every 'p' seconds on a
3637 * proportion of the peer_ni table and I need to check
3638 * every connection 'n' times within a timeout
3639 * interval, to ensure I detect a timeout on any
3640 * connection within (n+1)/n times the timeout
3644 lnd_timeout = kiblnd_timeout();
3645 if (lnd_timeout > n * p)
3646 chunk = (chunk * n * p) / lnd_timeout;
3650 for (i = 0; i < chunk; i++) {
3651 kiblnd_check_conns(peer_index);
3652 peer_index = (peer_index + 1) %
3653 HASH_SIZE(kiblnd_data.kib_peers);
3656 deadline += cfs_time_seconds(p);
3657 spin_lock_irqsave(lock, flags);
3663 /* Nothing to do for 'timeout' */
3664 set_current_state(TASK_INTERRUPTIBLE);
3665 add_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3666 spin_unlock_irqrestore(lock, flags);
3668 schedule_timeout(timeout);
3670 remove_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3671 spin_lock_irqsave(lock, flags);
3674 spin_unlock_irqrestore(lock, flags);
3676 kiblnd_thread_fini();
3681 kiblnd_qp_event(struct ib_event *event, void *arg)
3683 struct kib_conn *conn = arg;
3685 switch (event->event) {
3686 case IB_EVENT_COMM_EST:
3687 CDEBUG(D_NET, "%s established\n",
3688 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3689 /* We received a packet but connection isn't established
3690 * probably handshake packet was lost, so free to
3691 * force make connection established */
3692 rdma_notify(conn->ibc_cmid, IB_EVENT_COMM_EST);
3695 case IB_EVENT_PORT_ERR:
3696 case IB_EVENT_DEVICE_FATAL:
3697 CERROR("Fatal device error for NI %s\n",
3698 libcfs_nidstr(&conn->ibc_peer->ibp_ni->ni_nid));
3699 atomic_set(&conn->ibc_peer->ibp_ni->ni_fatal_error_on, 1);
3702 case IB_EVENT_PORT_ACTIVE:
3703 CERROR("Port reactivated for NI %s\n",
3704 libcfs_nidstr(&conn->ibc_peer->ibp_ni->ni_nid));
3705 atomic_set(&conn->ibc_peer->ibp_ni->ni_fatal_error_on, 0);
3709 CERROR("%s: Async QP event type %d\n",
3710 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3716 kiblnd_complete (struct ib_wc *wc)
3718 switch (kiblnd_wreqid2type(wc->wr_id)) {
3723 if (wc->status != IB_WC_SUCCESS &&
3724 wc->status != IB_WC_WR_FLUSH_ERR)
3725 CNETERR("FastReg failed: %d\n", wc->status);
3728 case IBLND_WID_RDMA:
3729 /* We only get RDMA completion notification if it fails. All
3730 * subsequent work items, including the final SEND will fail
3731 * too. However we can't print out any more info about the
3732 * failing RDMA because 'tx' might be back on the idle list or
3733 * even reused already if we didn't manage to post all our work
3735 CNETERR("RDMA (tx: %p) failed: %d\n",
3736 kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3740 kiblnd_tx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3744 kiblnd_rx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status,
3751 kiblnd_cq_completion(struct ib_cq *cq, void *arg)
3753 /* NB I'm not allowed to schedule this conn once its refcount has
3754 * reached 0. Since fundamentally I'm racing with scheduler threads
3755 * consuming my CQ I could be called after all completions have
3756 * occurred. But in this case, ibc_nrx == 0 && ibc_nsends_posted == 0
3757 * and this CQ is about to be destroyed so I NOOP. */
3758 struct kib_conn *conn = arg;
3759 struct kib_sched_info *sched = conn->ibc_sched;
3760 unsigned long flags;
3762 LASSERT(cq == conn->ibc_cq);
3764 spin_lock_irqsave(&sched->ibs_lock, flags);
3766 conn->ibc_ready = 1;
3768 if (!conn->ibc_scheduled &&
3769 (conn->ibc_nrx > 0 ||
3770 conn->ibc_nsends_posted > 0)) {
3771 kiblnd_conn_addref(conn); /* +1 ref for sched_conns */
3772 conn->ibc_scheduled = 1;
3773 list_add_tail(&conn->ibc_sched_list, &sched->ibs_conns);
3775 if (waitqueue_active(&sched->ibs_waitq))
3776 wake_up(&sched->ibs_waitq);
3779 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3783 kiblnd_cq_event(struct ib_event *event, void *arg)
3785 struct kib_conn *conn = arg;
3787 CERROR("%s: async CQ event type %d\n",
3788 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3792 kiblnd_scheduler(void *arg)
3794 long id = (long)arg;
3795 struct kib_sched_info *sched;
3796 struct kib_conn *conn;
3797 wait_queue_entry_t wait;
3798 unsigned long flags;
3805 sched = kiblnd_data.kib_scheds[KIB_THREAD_CPT(id)];
3807 rc = cfs_cpt_bind(lnet_cpt_table(), sched->ibs_cpt);
3809 CWARN("Unable to bind on CPU partition %d, please verify whether all CPUs are healthy and reload modules if necessary, otherwise your system might under risk of low performance\n", sched->ibs_cpt);
3812 spin_lock_irqsave(&sched->ibs_lock, flags);
3814 while (!kiblnd_data.kib_shutdown) {
3815 if (need_resched()) {
3816 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3820 spin_lock_irqsave(&sched->ibs_lock, flags);
3823 did_something = false;
3825 conn = list_first_entry_or_null(&sched->ibs_conns,
3829 /* take over kib_sched_conns' ref on conn... */
3830 LASSERT(conn->ibc_scheduled);
3831 list_del(&conn->ibc_sched_list);
3832 conn->ibc_ready = 0;
3834 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3836 wc.wr_id = IBLND_WID_INVAL;
3838 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3840 rc = ib_req_notify_cq(conn->ibc_cq,
3843 CWARN("%s: ib_req_notify_cq failed: %d, closing connection\n",
3844 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
3845 kiblnd_close_conn(conn, -EIO);
3846 kiblnd_conn_decref(conn);
3847 spin_lock_irqsave(&sched->ibs_lock,
3852 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3855 if (unlikely(rc > 0 && wc.wr_id == IBLND_WID_INVAL)) {
3857 "ib_poll_cq (rc: %d) returned invalid "
3858 "wr_id, opcode %d, status: %d, "
3859 "vendor_err: %d, conn: %s status: %d\n"
3860 "please upgrade firmware and OFED or "
3861 "contact vendor.\n", rc,
3862 wc.opcode, wc.status, wc.vendor_err,
3863 libcfs_nid2str(conn->ibc_peer->ibp_nid),
3869 CWARN("%s: ib_poll_cq failed: %d, closing connection\n",
3870 libcfs_nid2str(conn->ibc_peer->ibp_nid),
3872 kiblnd_close_conn(conn, -EIO);
3873 kiblnd_conn_decref(conn);
3874 spin_lock_irqsave(&sched->ibs_lock, flags);
3878 spin_lock_irqsave(&sched->ibs_lock, flags);
3880 if (rc != 0 || conn->ibc_ready) {
3881 /* There may be another completion waiting; get
3882 * another scheduler to check while I handle
3884 /* +1 ref for sched_conns */
3885 kiblnd_conn_addref(conn);
3886 list_add_tail(&conn->ibc_sched_list,
3888 if (waitqueue_active(&sched->ibs_waitq))
3889 wake_up(&sched->ibs_waitq);
3891 conn->ibc_scheduled = 0;
3895 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3896 kiblnd_complete(&wc);
3898 spin_lock_irqsave(&sched->ibs_lock, flags);
3901 kiblnd_conn_decref(conn); /* ..drop my ref from above */
3902 did_something = true;
3908 set_current_state(TASK_INTERRUPTIBLE);
3909 add_wait_queue_exclusive(&sched->ibs_waitq, &wait);
3910 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3914 remove_wait_queue(&sched->ibs_waitq, &wait);
3915 set_current_state(TASK_RUNNING);
3916 spin_lock_irqsave(&sched->ibs_lock, flags);
3919 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3921 kiblnd_thread_fini();
3926 kiblnd_failover_thread(void *arg)
3928 rwlock_t *glock = &kiblnd_data.kib_global_lock;
3929 struct kib_dev *dev;
3930 struct net *ns = arg;
3931 wait_queue_entry_t wait;
3932 unsigned long flags;
3935 LASSERT(*kiblnd_tunables.kib_dev_failover != 0);
3938 write_lock_irqsave(glock, flags);
3940 while (!kiblnd_data.kib_shutdown) {
3941 bool do_failover = false;
3944 list_for_each_entry(dev, &kiblnd_data.kib_failed_devs,
3946 if (ktime_get_seconds() < dev->ibd_next_failover)
3953 list_del_init(&dev->ibd_fail_list);
3954 dev->ibd_failover = 1;
3955 write_unlock_irqrestore(glock, flags);
3957 rc = kiblnd_dev_failover(dev, ns);
3959 write_lock_irqsave(glock, flags);
3961 LASSERT(dev->ibd_failover);
3962 dev->ibd_failover = 0;
3963 if (rc >= 0) { /* Device is OK or failover succeed */
3964 dev->ibd_next_failover = ktime_get_seconds() + 3;
3968 /* failed to failover, retry later */
3969 dev->ibd_next_failover = ktime_get_seconds() +
3970 min(dev->ibd_failed_failover, 10);
3971 if (kiblnd_dev_can_failover(dev)) {
3972 list_add_tail(&dev->ibd_fail_list,
3973 &kiblnd_data.kib_failed_devs);
3979 /* long sleep if no more pending failover */
3980 long_sleep = list_empty(&kiblnd_data.kib_failed_devs);
3982 set_current_state(TASK_INTERRUPTIBLE);
3983 add_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3984 write_unlock_irqrestore(glock, flags);
3986 rc = schedule_timeout(long_sleep ? cfs_time_seconds(10) :
3987 cfs_time_seconds(1));
3988 set_current_state(TASK_RUNNING);
3989 remove_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3990 write_lock_irqsave(glock, flags);
3992 if (!long_sleep || rc != 0)
3995 /* have a long sleep, routine check all active devices,
3996 * we need checking like this because if there is not active
3997 * connection on the dev and no SEND from local, we may listen
3998 * on wrong HCA for ever while there is a bonding failover
4000 list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
4001 if (kiblnd_dev_can_failover(dev)) {
4002 list_add_tail(&dev->ibd_fail_list,
4003 &kiblnd_data.kib_failed_devs);
4008 write_unlock_irqrestore(glock, flags);
4010 kiblnd_thread_fini();