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 void kiblnd_init_tx_msg(struct lnet_ni *ni, struct kib_tx *tx,
44 int type, int body_nob);
45 static int kiblnd_init_rdma(struct kib_conn *conn, struct kib_tx *tx, int type,
46 int resid, struct kib_rdma_desc *dstrd, u64 dstcookie);
47 static void kiblnd_queue_tx_locked(struct kib_tx *tx, struct kib_conn *conn);
48 static void kiblnd_queue_tx(struct kib_tx *tx, struct kib_conn *conn);
50 static void kiblnd_unmap_tx(struct kib_tx *tx);
51 static void kiblnd_check_sends_locked(struct kib_conn *conn);
54 kiblnd_tx_done(struct kib_tx *tx)
56 struct lnet_msg *lntmsg[2];
60 LASSERT (!in_interrupt());
61 LASSERT (!tx->tx_queued); /* mustn't be queued for sending */
62 LASSERT (tx->tx_sending == 0); /* mustn't be awaiting sent callback */
63 LASSERT (!tx->tx_waiting); /* mustn't be awaiting peer_ni response */
64 LASSERT (tx->tx_pool != NULL);
68 /* tx may have up to 2 lnet msgs to finalise */
69 lntmsg[0] = tx->tx_lntmsg[0]; tx->tx_lntmsg[0] = NULL;
70 lntmsg[1] = tx->tx_lntmsg[1]; tx->tx_lntmsg[1] = NULL;
73 if (tx->tx_conn != NULL) {
74 kiblnd_conn_decref(tx->tx_conn);
78 tx->tx_nwrq = tx->tx_nsge = 0;
81 kiblnd_pool_free_node(&tx->tx_pool->tpo_pool, &tx->tx_list);
83 /* delay finalize until my descs have been freed */
84 for (i = 0; i < 2; i++) {
85 if (lntmsg[i] == NULL)
88 /* propagate health status to LNet for requests */
89 if (i == 0 && lntmsg[i])
90 lntmsg[i]->msg_health_status = tx->tx_hstatus;
92 lnet_finalize(lntmsg[i], rc);
97 kiblnd_txlist_done(struct list_head *txlist, int status,
98 enum lnet_msg_hstatus hstatus)
102 while ((tx = list_first_entry_or_null(txlist,
105 list_del(&tx->tx_list);
108 tx->tx_status = status;
109 if (hstatus != LNET_MSG_STATUS_OK)
110 tx->tx_hstatus = hstatus;
115 static struct kib_tx *
116 kiblnd_get_idle_tx(struct lnet_ni *ni, lnet_nid_t target)
118 struct kib_net *net = ni->ni_data;
119 struct list_head *node;
121 struct kib_tx_poolset *tps;
123 tps = net->ibn_tx_ps[lnet_cpt_of_nid(target, ni)];
124 node = kiblnd_pool_alloc_node(&tps->tps_poolset);
127 tx = container_of(node, struct kib_tx, tx_list);
129 LASSERT (tx->tx_nwrq == 0);
130 LASSERT (!tx->tx_queued);
131 LASSERT (tx->tx_sending == 0);
132 LASSERT (!tx->tx_waiting);
133 LASSERT (tx->tx_status == 0);
134 LASSERT (tx->tx_conn == NULL);
135 LASSERT (tx->tx_lntmsg[0] == NULL);
136 LASSERT (tx->tx_lntmsg[1] == NULL);
137 LASSERT (tx->tx_nfrags == 0);
140 tx->tx_hstatus = LNET_MSG_STATUS_OK;
146 kiblnd_drop_rx(struct kib_rx *rx)
148 struct kib_conn *conn = rx->rx_conn;
149 struct kib_sched_info *sched = conn->ibc_sched;
152 spin_lock_irqsave(&sched->ibs_lock, flags);
153 LASSERT(conn->ibc_nrx > 0);
155 spin_unlock_irqrestore(&sched->ibs_lock, flags);
157 kiblnd_conn_decref(conn);
161 kiblnd_post_rx(struct kib_rx *rx, int credit)
163 struct kib_conn *conn = rx->rx_conn;
164 struct kib_net *net = conn->ibc_peer->ibp_ni->ni_data;
165 struct ib_recv_wr *bad_wrq = NULL;
166 #ifdef HAVE_IB_GET_DMA_MR
167 struct ib_mr *mr = conn->ibc_hdev->ibh_mrs;
171 LASSERT (net != NULL);
172 LASSERT (!in_interrupt());
173 LASSERT (credit == IBLND_POSTRX_NO_CREDIT ||
174 credit == IBLND_POSTRX_PEER_CREDIT ||
175 credit == IBLND_POSTRX_RSRVD_CREDIT);
176 #ifdef HAVE_IB_GET_DMA_MR
179 rx->rx_sge.lkey = mr->lkey;
181 rx->rx_sge.lkey = conn->ibc_hdev->ibh_pd->local_dma_lkey;
183 rx->rx_sge.addr = rx->rx_msgaddr;
184 rx->rx_sge.length = IBLND_MSG_SIZE;
186 rx->rx_wrq.next = NULL;
187 rx->rx_wrq.sg_list = &rx->rx_sge;
188 rx->rx_wrq.num_sge = 1;
189 rx->rx_wrq.wr_id = kiblnd_ptr2wreqid(rx, IBLND_WID_RX);
191 LASSERT (conn->ibc_state >= IBLND_CONN_INIT);
192 LASSERT (rx->rx_nob >= 0); /* not posted */
194 if (conn->ibc_state > IBLND_CONN_ESTABLISHED) {
195 kiblnd_drop_rx(rx); /* No more posts for this rx */
199 rx->rx_nob = -1; /* flag posted */
201 /* NB: need an extra reference after ib_post_recv because we don't
202 * own this rx (and rx::rx_conn) anymore, LU-5678.
204 kiblnd_conn_addref(conn);
205 #ifdef HAVE_IB_POST_SEND_RECV_CONST
206 rc = ib_post_recv(conn->ibc_cmid->qp, &rx->rx_wrq,
207 (const struct ib_recv_wr **)&bad_wrq);
209 rc = ib_post_recv(conn->ibc_cmid->qp, &rx->rx_wrq, &bad_wrq);
211 if (unlikely(rc != 0)) {
212 CERROR("Can't post rx for %s: %d, bad_wrq: %p\n",
213 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc, bad_wrq);
217 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) /* Initial post */
220 if (unlikely(rc != 0)) {
221 kiblnd_close_conn(conn, rc);
222 kiblnd_drop_rx(rx); /* No more posts for this rx */
226 if (credit == IBLND_POSTRX_NO_CREDIT)
229 spin_lock(&conn->ibc_lock);
230 if (credit == IBLND_POSTRX_PEER_CREDIT)
231 conn->ibc_outstanding_credits++;
233 conn->ibc_reserved_credits++;
234 kiblnd_check_sends_locked(conn);
235 spin_unlock(&conn->ibc_lock);
238 kiblnd_conn_decref(conn);
242 static struct kib_tx *
243 kiblnd_find_waiting_tx_locked(struct kib_conn *conn, int txtype, u64 cookie)
247 list_for_each_entry(tx, &conn->ibc_active_txs, tx_list) {
248 LASSERT(!tx->tx_queued);
249 LASSERT(tx->tx_sending != 0 || tx->tx_waiting);
251 if (tx->tx_cookie != cookie)
254 if (tx->tx_waiting &&
255 tx->tx_msg->ibm_type == txtype)
258 CWARN("Bad completion: %swaiting, type %x (wanted %x)\n",
259 tx->tx_waiting ? "" : "NOT ",
260 tx->tx_msg->ibm_type, txtype);
266 kiblnd_handle_completion(struct kib_conn *conn, int txtype, int status, u64 cookie)
269 struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
272 spin_lock(&conn->ibc_lock);
274 tx = kiblnd_find_waiting_tx_locked(conn, txtype, cookie);
276 spin_unlock(&conn->ibc_lock);
278 CWARN("Unmatched completion type %x cookie %#llx from %s\n",
279 txtype, cookie, libcfs_nid2str(conn->ibc_peer->ibp_nid));
280 kiblnd_close_conn(conn, -EPROTO);
284 if (tx->tx_status == 0) { /* success so far */
285 if (status < 0) { /* failed? */
286 tx->tx_status = status;
287 tx->tx_hstatus = LNET_MSG_STATUS_REMOTE_ERROR;
288 } else if (txtype == IBLND_MSG_GET_REQ) {
289 lnet_set_reply_msg_len(ni, tx->tx_lntmsg[1], status);
295 idle = !tx->tx_queued && (tx->tx_sending == 0);
297 list_del(&tx->tx_list);
299 spin_unlock(&conn->ibc_lock);
306 kiblnd_send_completion(struct kib_conn *conn, int type, int status, u64 cookie)
308 struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
309 struct kib_tx *tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
312 CERROR("Can't get tx for completion %x for %s\n",
313 type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
317 tx->tx_msg->ibm_u.completion.ibcm_status = status;
318 tx->tx_msg->ibm_u.completion.ibcm_cookie = cookie;
319 kiblnd_init_tx_msg(ni, tx, type, sizeof(struct kib_completion_msg));
321 kiblnd_queue_tx(tx, conn);
325 kiblnd_handle_rx(struct kib_rx *rx)
327 struct kib_msg *msg = rx->rx_msg;
328 struct kib_conn *conn = rx->rx_conn;
329 struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
330 int credits = msg->ibm_credits;
336 LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
338 CDEBUG (D_NET, "Received %x[%d] from %s\n",
339 msg->ibm_type, credits,
340 libcfs_nid2str(conn->ibc_peer->ibp_nid));
343 /* Have I received credits that will let me send? */
344 spin_lock(&conn->ibc_lock);
346 if (conn->ibc_credits + credits >
347 conn->ibc_queue_depth) {
348 rc2 = conn->ibc_credits;
349 spin_unlock(&conn->ibc_lock);
351 CERROR("Bad credits from %s: %d + %d > %d\n",
352 libcfs_nid2str(conn->ibc_peer->ibp_nid),
354 conn->ibc_queue_depth);
356 kiblnd_close_conn(conn, -EPROTO);
357 kiblnd_post_rx(rx, IBLND_POSTRX_NO_CREDIT);
361 conn->ibc_credits += credits;
363 /* This ensures the credit taken by NOOP can be returned */
364 if (msg->ibm_type == IBLND_MSG_NOOP &&
365 !IBLND_OOB_CAPABLE(conn->ibc_version)) /* v1 only */
366 conn->ibc_outstanding_credits++;
368 kiblnd_check_sends_locked(conn);
369 spin_unlock(&conn->ibc_lock);
372 switch (msg->ibm_type) {
374 CERROR("Bad IBLND message type %x from %s\n",
375 msg->ibm_type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
376 post_credit = IBLND_POSTRX_NO_CREDIT;
381 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
382 post_credit = IBLND_POSTRX_NO_CREDIT;
386 if (credits != 0) /* credit already posted */
387 post_credit = IBLND_POSTRX_NO_CREDIT;
388 else /* a keepalive NOOP */
389 post_credit = IBLND_POSTRX_PEER_CREDIT;
392 case IBLND_MSG_IMMEDIATE:
393 post_credit = IBLND_POSTRX_DONT_POST;
394 rc = lnet_parse(ni, &msg->ibm_u.immediate.ibim_hdr,
395 msg->ibm_srcnid, rx, 0);
396 if (rc < 0) /* repost on error */
397 post_credit = IBLND_POSTRX_PEER_CREDIT;
400 case IBLND_MSG_PUT_REQ:
401 post_credit = IBLND_POSTRX_DONT_POST;
402 rc = lnet_parse(ni, &msg->ibm_u.putreq.ibprm_hdr,
403 msg->ibm_srcnid, rx, 1);
404 if (rc < 0) /* repost on error */
405 post_credit = IBLND_POSTRX_PEER_CREDIT;
408 case IBLND_MSG_PUT_NAK:
409 CWARN ("PUT_NACK from %s\n",
410 libcfs_nid2str(conn->ibc_peer->ibp_nid));
411 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
412 kiblnd_handle_completion(conn, IBLND_MSG_PUT_REQ,
413 msg->ibm_u.completion.ibcm_status,
414 msg->ibm_u.completion.ibcm_cookie);
417 case IBLND_MSG_PUT_ACK:
418 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
420 spin_lock(&conn->ibc_lock);
421 tx = kiblnd_find_waiting_tx_locked(conn, IBLND_MSG_PUT_REQ,
422 msg->ibm_u.putack.ibpam_src_cookie);
424 list_del(&tx->tx_list);
425 spin_unlock(&conn->ibc_lock);
428 CERROR("Unmatched PUT_ACK from %s\n",
429 libcfs_nid2str(conn->ibc_peer->ibp_nid));
434 LASSERT (tx->tx_waiting);
435 /* CAVEAT EMPTOR: I could be racing with tx_complete, but...
436 * (a) I can overwrite tx_msg since my peer_ni has received it!
437 * (b) tx_waiting set tells tx_complete() it's not done. */
439 tx->tx_nwrq = tx->tx_nsge = 0; /* overwrite PUT_REQ */
441 rc2 = kiblnd_init_rdma(conn, tx, IBLND_MSG_PUT_DONE,
442 kiblnd_rd_size(&msg->ibm_u.putack.ibpam_rd),
443 &msg->ibm_u.putack.ibpam_rd,
444 msg->ibm_u.putack.ibpam_dst_cookie);
446 CERROR("Can't setup rdma for PUT to %s: %d\n",
447 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc2);
449 spin_lock(&conn->ibc_lock);
450 tx->tx_waiting = 0; /* clear waiting and queue atomically */
451 kiblnd_queue_tx_locked(tx, conn);
452 spin_unlock(&conn->ibc_lock);
455 case IBLND_MSG_PUT_DONE:
456 post_credit = IBLND_POSTRX_PEER_CREDIT;
457 kiblnd_handle_completion(conn, IBLND_MSG_PUT_ACK,
458 msg->ibm_u.completion.ibcm_status,
459 msg->ibm_u.completion.ibcm_cookie);
462 case IBLND_MSG_GET_REQ:
463 post_credit = IBLND_POSTRX_DONT_POST;
464 rc = lnet_parse(ni, &msg->ibm_u.get.ibgm_hdr,
465 msg->ibm_srcnid, rx, 1);
466 if (rc < 0) /* repost on error */
467 post_credit = IBLND_POSTRX_PEER_CREDIT;
470 case IBLND_MSG_GET_DONE:
471 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
472 kiblnd_handle_completion(conn, IBLND_MSG_GET_REQ,
473 msg->ibm_u.completion.ibcm_status,
474 msg->ibm_u.completion.ibcm_cookie);
478 if (rc < 0) /* protocol error */
479 kiblnd_close_conn(conn, rc);
481 if (post_credit != IBLND_POSTRX_DONT_POST)
482 kiblnd_post_rx(rx, post_credit);
486 kiblnd_rx_complete(struct kib_rx *rx, int status, int nob)
488 struct kib_msg *msg = rx->rx_msg;
489 struct kib_conn *conn = rx->rx_conn;
490 struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
491 struct kib_net *net = ni->ni_data;
496 LASSERT(rx->rx_nob < 0); /* was posted */
497 rx->rx_nob = 0; /* isn't now */
499 if (conn->ibc_state > IBLND_CONN_ESTABLISHED)
502 if (status != IB_WC_SUCCESS) {
503 CNETERR("Rx from %s failed: %d\n",
504 libcfs_nid2str(conn->ibc_peer->ibp_nid), status);
511 rc = kiblnd_unpack_msg(msg, rx->rx_nob);
513 CERROR("Error %d unpacking rx from %s\n",
514 rc, libcfs_nid2str(conn->ibc_peer->ibp_nid));
518 if (msg->ibm_srcnid != conn->ibc_peer->ibp_nid ||
519 msg->ibm_dstnid != lnet_nid_to_nid4(&ni->ni_nid) ||
520 msg->ibm_srcstamp != conn->ibc_incarnation ||
521 msg->ibm_dststamp != net->ibn_incarnation) {
522 CERROR("Stale rx from %s\n",
523 libcfs_nid2str(conn->ibc_peer->ibp_nid));
528 /* set time last known alive */
529 kiblnd_peer_alive(conn->ibc_peer);
531 /* racing with connection establishment/teardown! */
533 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
534 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
537 write_lock_irqsave(g_lock, flags);
538 /* must check holding global lock to eliminate race */
539 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
540 list_add_tail(&rx->rx_list, &conn->ibc_early_rxs);
541 write_unlock_irqrestore(g_lock, flags);
544 write_unlock_irqrestore(g_lock, flags);
546 kiblnd_handle_rx(rx);
550 CDEBUG(D_NET, "rx %p conn %p\n", rx, conn);
551 kiblnd_close_conn(conn, err);
553 kiblnd_drop_rx(rx); /* Don't re-post rx. */
557 kiblnd_fmr_map_tx(struct kib_net *net, struct kib_tx *tx,
558 struct kib_rdma_desc *rd, u32 nob)
560 struct kib_hca_dev *hdev;
562 struct kib_fmr_poolset *fps;
567 LASSERT(tx->tx_pool != NULL);
568 LASSERT(tx->tx_pool->tpo_pool.po_owner != NULL);
571 hdev = tx->tx_pool->tpo_hdev;
572 cpt = tx->tx_pool->tpo_pool.po_owner->ps_cpt;
575 * If we're dealing with FastReg, but the device doesn't
576 * support GAPS and the tx has GAPS, then there is no real point
577 * in trying to map the memory, because it'll just fail. So
578 * preemptively fail with an appropriate message
580 if ((dev->ibd_dev_caps & IBLND_DEV_CAPS_FASTREG_ENABLED) &&
581 !(dev->ibd_dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT) &&
583 CERROR("Using FastReg with no GAPS support, but tx has gaps. "
584 "Try setting use_fastreg_gaps to 1\n");
585 return -EPROTONOSUPPORT;
588 #ifdef HAVE_FMR_POOL_API
590 * FMR does not support gaps but the tx has gaps then
591 * we should make sure that the number of fragments we'll be sending
592 * over fits within the number of fragments negotiated on the
593 * connection, otherwise, we won't be able to RDMA the data.
594 * We need to maintain the number of fragments negotiation on the
595 * connection for backwards compatibility.
597 if (tx->tx_gaps && (dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED)) {
599 tx->tx_conn->ibc_max_frags <= rd->rd_nfrags) {
600 CERROR("TX number of frags (%d) is <= than connection"
601 " number of frags (%d). Consider setting peer's"
602 " map_on_demand to 256\n", tx->tx_nfrags,
603 tx->tx_conn->ibc_max_frags);
609 fps = net->ibn_fmr_ps[cpt];
610 rc = kiblnd_fmr_pool_map(fps, tx, rd, nob, 0, &tx->tx_fmr);
612 CERROR("Can't map %u bytes (%u/%u)s: %d\n", nob,
613 tx->tx_nfrags, rd->rd_nfrags, rc);
618 * If rd is not tx_rd, it's going to get sent to a peer_ni, who will
621 rd->rd_key = tx->tx_fmr.fmr_key;
623 * for FastReg or FMR with no gaps we can accumulate all
624 * the fragments in one FastReg or FMR fragment.
627 #ifdef HAVE_FMR_POOL_API
628 ((dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED)
631 (dev->ibd_dev_caps & IBLND_DEV_CAPS_FASTREG_ENABLED)) {
632 /* FMR requires zero based address */
633 #ifdef HAVE_FMR_POOL_API
634 if (dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED)
635 rd->rd_frags[0].rf_addr &= ~hdev->ibh_page_mask;
637 rd->rd_frags[0].rf_nob = nob;
641 * We're transmitting with gaps using FMR.
642 * We'll need to use multiple fragments and identify the
643 * zero based address of each fragment.
645 for (i = 0; i < rd->rd_nfrags; i++) {
646 rd->rd_frags[i].rf_addr &= ~hdev->ibh_page_mask;
647 rd->rd_frags[i].rf_addr += i << hdev->ibh_page_shift;
655 kiblnd_unmap_tx(struct kib_tx *tx)
658 #ifdef HAVE_FMR_POOL_API
659 tx->tx_fmr.fmr_pfmr ||
662 kiblnd_fmr_pool_unmap(&tx->tx_fmr, tx->tx_status);
664 if (tx->tx_nfrags != 0) {
665 kiblnd_dma_unmap_sg(tx->tx_pool->tpo_hdev,
666 tx->tx_frags, tx->tx_nfrags, tx->tx_dmadir);
671 #ifdef HAVE_IB_GET_DMA_MR
672 static struct ib_mr *
673 kiblnd_find_rd_dma_mr(struct lnet_ni *ni, struct kib_rdma_desc *rd)
675 struct kib_net *net = ni->ni_data;
676 struct kib_hca_dev *hdev = net->ibn_dev->ibd_hdev;
677 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
679 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
682 * if map-on-demand is turned on and the device supports
683 * either FMR or FastReg then use that. Otherwise use global
684 * memory regions. If that's not available either, then you're
685 * dead in the water and fail the operation.
687 if (tunables->lnd_map_on_demand &&
688 (net->ibn_dev->ibd_dev_caps & IBLND_DEV_CAPS_FASTREG_ENABLED
689 #ifdef HAVE_FMR_POOL_API
690 || net->ibn_dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED
696 * hdev->ibh_mrs can be NULL. This case is dealt with gracefully
697 * in the call chain. The mapping will fail with appropriate error
700 return hdev->ibh_mrs;
704 static int kiblnd_map_tx(struct lnet_ni *ni, struct kib_tx *tx,
705 struct kib_rdma_desc *rd, int nfrags)
707 struct kib_net *net = ni->ni_data;
708 struct kib_hca_dev *hdev = net->ibn_dev->ibd_hdev;
709 #ifdef HAVE_IB_GET_DMA_MR
710 struct ib_mr *mr = NULL;
715 /* If rd is not tx_rd, it's going to get sent to a peer_ni and I'm the
717 tx->tx_dmadir = (rd != tx->tx_rd) ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
718 tx->tx_nfrags = nfrags;
720 rd->rd_nfrags = kiblnd_dma_map_sg(hdev, tx->tx_frags,
721 tx->tx_nfrags, tx->tx_dmadir);
723 for (i = 0, nob = 0; i < rd->rd_nfrags; i++) {
724 rd->rd_frags[i].rf_nob = kiblnd_sg_dma_len(
725 hdev->ibh_ibdev, &tx->tx_frags[i]);
726 rd->rd_frags[i].rf_addr = kiblnd_sg_dma_address(
727 hdev->ibh_ibdev, &tx->tx_frags[i]);
728 nob += rd->rd_frags[i].rf_nob;
731 #ifdef HAVE_IB_GET_DMA_MR
732 mr = kiblnd_find_rd_dma_mr(ni, rd);
734 /* found pre-mapping MR */
735 rd->rd_key = (rd != tx->tx_rd) ? mr->rkey : mr->lkey;
740 if (net->ibn_fmr_ps != NULL)
741 return kiblnd_fmr_map_tx(net, tx, rd, nob);
746 static int kiblnd_setup_rd_kiov(struct lnet_ni *ni, struct kib_tx *tx,
747 struct kib_rdma_desc *rd, int nkiov,
748 struct bio_vec *kiov, int offset, int nob)
750 struct kib_net *net = ni->ni_data;
751 struct scatterlist *sg;
756 CDEBUG(D_NET, "niov %d offset %d nob %d\n", nkiov, offset, nob);
760 LASSERT(net != NULL);
762 while (offset >= kiov->bv_len) {
763 offset -= kiov->bv_len;
776 CERROR("lacking enough sg entries to map tx\n");
781 fragnob = min((int)(kiov->bv_len - offset), nob);
784 * We're allowed to start at a non-aligned page offset in
785 * the first fragment and end at a non-aligned page offset
786 * in the last fragment.
788 if ((fragnob < (int)(kiov->bv_len - offset)) &&
789 nkiov < max_nkiov && nob > fragnob) {
790 CDEBUG(D_NET, "fragnob %d < available page %d: with"
791 " remaining %d kiovs with %d nob left\n",
792 fragnob, (int)(kiov->bv_len - offset),
797 sg_set_page(sg, kiov->bv_page, fragnob,
798 kiov->bv_offset + offset);
807 return kiblnd_map_tx(ni, tx, rd, sg_count);
811 kiblnd_post_tx_locked(struct kib_conn *conn, struct kib_tx *tx, int credit)
812 __must_hold(&conn->ibc_lock)
814 struct kib_msg *msg = tx->tx_msg;
815 struct kib_peer_ni *peer_ni = conn->ibc_peer;
816 struct lnet_ni *ni = peer_ni->ibp_ni;
817 struct kib_fast_reg_descriptor *frd = tx->tx_fmr.fmr_frd;
818 int ver = conn->ibc_version;
822 LASSERT(tx->tx_queued);
823 /* We rely on this for QP sizing */
824 LASSERT(tx->tx_nwrq > 0 && tx->tx_nsge >= 0);
825 LASSERT(tx->tx_nwrq <= 1 + conn->ibc_max_frags);
827 LASSERT(credit == 0 || credit == 1);
828 LASSERT(conn->ibc_outstanding_credits >= 0);
829 LASSERT(conn->ibc_outstanding_credits <= conn->ibc_queue_depth);
830 LASSERT(conn->ibc_credits >= 0);
831 LASSERT(conn->ibc_credits <= conn->ibc_queue_depth);
833 if (conn->ibc_nsends_posted ==
834 kiblnd_concurrent_sends(ver, ni)) {
835 /* tx completions outstanding... */
836 CDEBUG(D_NET, "%s: posted enough\n",
837 libcfs_nid2str(peer_ni->ibp_nid));
841 if (credit != 0 && conn->ibc_credits == 0) { /* no credits */
842 CDEBUG(D_NET, "%s: no credits\n",
843 libcfs_nid2str(peer_ni->ibp_nid));
847 if (credit != 0 && !IBLND_OOB_CAPABLE(ver) &&
848 conn->ibc_credits == 1 && /* last credit reserved */
849 msg->ibm_type != IBLND_MSG_NOOP) { /* for NOOP */
850 CDEBUG(D_NET, "%s: not using last credit\n",
851 libcfs_nid2str(peer_ni->ibp_nid));
855 /* NB don't drop ibc_lock before bumping tx_sending */
856 list_del(&tx->tx_list);
859 if (msg->ibm_type == IBLND_MSG_NOOP &&
860 (!kiblnd_need_noop(conn) || /* redundant NOOP */
861 (IBLND_OOB_CAPABLE(ver) && /* posted enough NOOP */
862 conn->ibc_noops_posted == IBLND_OOB_MSGS(ver)))) {
863 /* OK to drop when posted enough NOOPs, since
864 * kiblnd_check_sends_locked will queue NOOP again when
865 * posted NOOPs complete */
866 spin_unlock(&conn->ibc_lock);
867 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
869 spin_lock(&conn->ibc_lock);
870 CDEBUG(D_NET, "%s(%d): redundant or enough NOOP\n",
871 libcfs_nid2str(peer_ni->ibp_nid),
872 conn->ibc_noops_posted);
876 kiblnd_pack_msg(peer_ni->ibp_ni, msg, ver, conn->ibc_outstanding_credits,
877 peer_ni->ibp_nid, conn->ibc_incarnation);
879 conn->ibc_credits -= credit;
880 conn->ibc_outstanding_credits = 0;
881 conn->ibc_nsends_posted++;
882 if (msg->ibm_type == IBLND_MSG_NOOP)
883 conn->ibc_noops_posted++;
885 /* CAVEAT EMPTOR! This tx could be the PUT_DONE of an RDMA
886 * PUT. If so, it was first queued here as a PUT_REQ, sent and
887 * stashed on ibc_active_txs, matched by an incoming PUT_ACK,
888 * and then re-queued here. It's (just) possible that
889 * tx_sending is non-zero if we've not done the tx_complete()
890 * from the first send; hence the ++ rather than = below. */
892 list_add(&tx->tx_list, &conn->ibc_active_txs);
894 /* I'm still holding ibc_lock! */
895 if (conn->ibc_state != IBLND_CONN_ESTABLISHED) {
897 } else if (tx->tx_pool->tpo_pool.po_failed ||
898 conn->ibc_hdev != tx->tx_pool->tpo_hdev) {
899 /* close_conn will launch failover */
902 struct ib_send_wr *bad = &tx->tx_wrq[tx->tx_nwrq - 1].wr;
903 struct ib_send_wr *wr = &tx->tx_wrq[0].wr;
905 if (frd != NULL && !frd->frd_posted) {
906 if (!frd->frd_valid) {
907 wr = &frd->frd_inv_wr.wr;
908 wr->next = &frd->frd_fastreg_wr.wr;
910 wr = &frd->frd_fastreg_wr.wr;
912 frd->frd_fastreg_wr.wr.next = &tx->tx_wrq[0].wr;
915 LASSERTF(bad->wr_id == kiblnd_ptr2wreqid(tx, IBLND_WID_TX),
916 "bad wr_id %#llx, opc %d, flags %d, peer_ni: %s\n",
917 bad->wr_id, bad->opcode, bad->send_flags,
918 libcfs_nid2str(conn->ibc_peer->ibp_nid));
921 if (lnet_send_error_simulation(tx->tx_lntmsg[0], &tx->tx_hstatus))
924 #ifdef HAVE_IB_POST_SEND_RECV_CONST
925 rc = ib_post_send(conn->ibc_cmid->qp, wr,
926 (const struct ib_send_wr **)&bad);
928 rc = ib_post_send(conn->ibc_cmid->qp, wr, &bad);
932 conn->ibc_last_send = ktime_get();
936 frd->frd_posted = true;
940 /* NB credits are transferred in the actual
941 * message, which can only be the last work item */
942 conn->ibc_credits += credit;
943 conn->ibc_outstanding_credits += msg->ibm_credits;
944 conn->ibc_nsends_posted--;
945 if (msg->ibm_type == IBLND_MSG_NOOP)
946 conn->ibc_noops_posted--;
952 done = (tx->tx_sending == 0);
954 list_del(&tx->tx_list);
956 spin_unlock(&conn->ibc_lock);
958 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
959 CERROR("Error %d posting transmit to %s\n",
960 rc, libcfs_nid2str(peer_ni->ibp_nid));
962 CDEBUG(D_NET, "Error %d posting transmit to %s\n",
963 rc, libcfs_nid2str(peer_ni->ibp_nid));
965 kiblnd_close_conn(conn, rc);
970 spin_lock(&conn->ibc_lock);
976 kiblnd_check_sends_locked(struct kib_conn *conn)
978 int ver = conn->ibc_version;
979 struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
982 /* Don't send anything until after the connection is established */
983 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
984 CDEBUG(D_NET, "%s too soon\n",
985 libcfs_nid2str(conn->ibc_peer->ibp_nid));
989 LASSERT(conn->ibc_nsends_posted <=
990 kiblnd_concurrent_sends(ver, ni));
991 LASSERT (!IBLND_OOB_CAPABLE(ver) ||
992 conn->ibc_noops_posted <= IBLND_OOB_MSGS(ver));
993 LASSERT (conn->ibc_reserved_credits >= 0);
995 while (conn->ibc_reserved_credits > 0 &&
996 (tx = list_first_entry_or_null(&conn->ibc_tx_queue_rsrvd,
997 struct kib_tx, tx_list)) != NULL) {
998 list_move_tail(&tx->tx_list, &conn->ibc_tx_queue);
999 conn->ibc_reserved_credits--;
1002 if (kiblnd_need_noop(conn)) {
1003 spin_unlock(&conn->ibc_lock);
1005 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
1007 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_NOOP, 0);
1009 spin_lock(&conn->ibc_lock);
1011 kiblnd_queue_tx_locked(tx, conn);
1017 if (!list_empty(&conn->ibc_tx_queue_nocred)) {
1019 tx = list_first_entry(&conn->ibc_tx_queue_nocred,
1020 struct kib_tx, tx_list);
1021 } else if (!list_empty(&conn->ibc_tx_noops)) {
1022 LASSERT (!IBLND_OOB_CAPABLE(ver));
1024 tx = list_first_entry(&conn->ibc_tx_noops,
1025 struct kib_tx, tx_list);
1026 } else if (!list_empty(&conn->ibc_tx_queue)) {
1028 tx = list_first_entry(&conn->ibc_tx_queue,
1029 struct kib_tx, tx_list);
1033 if (kiblnd_post_tx_locked(conn, tx, credit) != 0)
1039 kiblnd_tx_complete(struct kib_tx *tx, int status)
1041 int failed = (status != IB_WC_SUCCESS);
1042 struct kib_conn *conn = tx->tx_conn;
1045 if (tx->tx_sending <= 0) {
1046 CERROR("Received an event on a freed tx: %p status %d\n",
1052 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
1053 CNETERR("Tx -> %s cookie %#llx"
1054 " sending %d waiting %d: failed %d\n",
1055 libcfs_nid2str(conn->ibc_peer->ibp_nid),
1056 tx->tx_cookie, tx->tx_sending, tx->tx_waiting,
1059 kiblnd_close_conn(conn, -EIO);
1061 kiblnd_peer_alive(conn->ibc_peer);
1064 spin_lock(&conn->ibc_lock);
1066 /* I could be racing with rdma completion. Whoever makes 'tx' idle
1067 * gets to free it, which also drops its ref on 'conn'. */
1070 conn->ibc_nsends_posted--;
1071 if (tx->tx_msg->ibm_type == IBLND_MSG_NOOP)
1072 conn->ibc_noops_posted--;
1075 tx->tx_hstatus = LNET_MSG_STATUS_REMOTE_DROPPED;
1076 tx->tx_waiting = 0; /* don't wait for peer_ni */
1077 tx->tx_status = -EIO;
1080 idle = (tx->tx_sending == 0) && /* This is the final callback */
1081 !tx->tx_waiting && /* Not waiting for peer_ni */
1082 !tx->tx_queued; /* Not re-queued (PUT_DONE) */
1084 list_del(&tx->tx_list);
1086 kiblnd_check_sends_locked(conn);
1087 spin_unlock(&conn->ibc_lock);
1094 kiblnd_init_tx_msg(struct lnet_ni *ni, struct kib_tx *tx, int type,
1097 struct kib_hca_dev *hdev = tx->tx_pool->tpo_hdev;
1098 struct ib_sge *sge = &tx->tx_msgsge;
1099 struct ib_rdma_wr *wrq;
1100 int nob = offsetof(struct kib_msg, ibm_u) + body_nob;
1101 #ifdef HAVE_IB_GET_DMA_MR
1102 struct ib_mr *mr = hdev->ibh_mrs;
1105 LASSERT(tx->tx_nwrq >= 0);
1106 LASSERT(tx->tx_nwrq <= IBLND_MAX_RDMA_FRAGS);
1107 LASSERT(nob <= IBLND_MSG_SIZE);
1108 #ifdef HAVE_IB_GET_DMA_MR
1109 LASSERT(mr != NULL);
1112 kiblnd_init_msg(tx->tx_msg, type, body_nob);
1114 #ifdef HAVE_IB_GET_DMA_MR
1115 sge->lkey = mr->lkey;
1117 sge->lkey = hdev->ibh_pd->local_dma_lkey;
1119 sge->addr = tx->tx_msgaddr;
1122 wrq = &tx->tx_wrq[tx->tx_nwrq];
1123 memset(wrq, 0, sizeof(*wrq));
1125 wrq->wr.next = NULL;
1126 wrq->wr.wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_TX);
1127 wrq->wr.sg_list = sge;
1128 wrq->wr.num_sge = 1;
1129 wrq->wr.opcode = IB_WR_SEND;
1130 wrq->wr.send_flags = IB_SEND_SIGNALED;
1136 kiblnd_init_rdma(struct kib_conn *conn, struct kib_tx *tx, int type,
1137 int resid, struct kib_rdma_desc *dstrd, u64 dstcookie)
1139 struct kib_msg *ibmsg = tx->tx_msg;
1140 struct kib_rdma_desc *srcrd = tx->tx_rd;
1141 struct ib_rdma_wr *wrq = NULL;
1149 LASSERT(!in_interrupt());
1150 LASSERT(tx->tx_nwrq == 0 && tx->tx_nsge == 0);
1151 LASSERT(type == IBLND_MSG_GET_DONE || type == IBLND_MSG_PUT_DONE);
1153 for (srcidx = dstidx = wrq_sge = sge_nob = 0;
1154 resid > 0; resid -= sge_nob) {
1157 if (srcidx >= srcrd->rd_nfrags) {
1158 CERROR("Src buffer exhausted: %d frags\n", srcidx);
1163 if (dstidx >= dstrd->rd_nfrags) {
1164 CERROR("Dst buffer exhausted: %d frags\n", dstidx);
1169 if (tx->tx_nwrq >= conn->ibc_max_frags) {
1170 CERROR("RDMA has too many fragments for peer_ni %s (%d), "
1171 "src idx/frags: %d/%d dst idx/frags: %d/%d\n",
1172 libcfs_nid2str(conn->ibc_peer->ibp_nid),
1173 conn->ibc_max_frags,
1174 srcidx, srcrd->rd_nfrags,
1175 dstidx, dstrd->rd_nfrags);
1180 sge_nob = min3(kiblnd_rd_frag_size(srcrd, srcidx),
1181 kiblnd_rd_frag_size(dstrd, dstidx),
1184 sge = &tx->tx_sge[tx->tx_nsge];
1185 sge->addr = kiblnd_rd_frag_addr(srcrd, srcidx);
1186 sge->lkey = kiblnd_rd_frag_key(srcrd, srcidx);
1187 sge->length = sge_nob;
1190 wrq = &tx->tx_wrq[tx->tx_nwrq];
1192 wrq->wr.next = &(wrq + 1)->wr;
1193 wrq->wr.wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_RDMA);
1194 wrq->wr.sg_list = sge;
1195 wrq->wr.opcode = IB_WR_RDMA_WRITE;
1196 wrq->wr.send_flags = 0;
1198 #ifdef HAVE_IB_RDMA_WR
1199 wrq->remote_addr = kiblnd_rd_frag_addr(dstrd,
1201 wrq->rkey = kiblnd_rd_frag_key(dstrd,
1204 wrq->wr.wr.rdma.remote_addr = kiblnd_rd_frag_addr(dstrd,
1206 wrq->wr.wr.rdma.rkey = kiblnd_rd_frag_key(dstrd,
1211 srcidx = kiblnd_rd_consume_frag(srcrd, srcidx, sge_nob);
1212 dstidx = kiblnd_rd_consume_frag(dstrd, dstidx, sge_nob);
1215 if (wrq_sge == *kiblnd_tunables.kib_wrq_sge || dstidx != prev) {
1217 wrq->wr.num_sge = wrq_sge;
1223 if (rc < 0) /* no RDMA if completing with failure */
1224 tx->tx_nwrq = tx->tx_nsge = 0;
1226 ibmsg->ibm_u.completion.ibcm_status = rc;
1227 ibmsg->ibm_u.completion.ibcm_cookie = dstcookie;
1228 kiblnd_init_tx_msg(conn->ibc_peer->ibp_ni, tx,
1229 type, sizeof(struct kib_completion_msg));
1235 kiblnd_queue_tx_locked(struct kib_tx *tx, struct kib_conn *conn)
1237 struct list_head *q;
1240 LASSERT(tx->tx_nwrq > 0); /* work items set up */
1241 LASSERT(!tx->tx_queued); /* not queued for sending already */
1242 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1244 if (conn->ibc_state >= IBLND_CONN_DISCONNECTED) {
1245 tx->tx_status = -ECONNABORTED;
1247 if (tx->tx_conn != NULL) {
1248 /* PUT_DONE first attached to conn as a PUT_REQ */
1249 LASSERT(tx->tx_conn == conn);
1250 LASSERT(tx->tx_msg->ibm_type == IBLND_MSG_PUT_DONE);
1252 kiblnd_conn_decref(conn);
1254 list_add(&tx->tx_list, &conn->ibc_zombie_txs);
1259 timeout_ns = kiblnd_timeout() * NSEC_PER_SEC;
1261 tx->tx_deadline = ktime_add_ns(ktime_get(), timeout_ns);
1263 if (tx->tx_conn == NULL) {
1264 kiblnd_conn_addref(conn);
1266 LASSERT (tx->tx_msg->ibm_type != IBLND_MSG_PUT_DONE);
1268 /* PUT_DONE first attached to conn as a PUT_REQ */
1269 LASSERT (tx->tx_conn == conn);
1270 LASSERT (tx->tx_msg->ibm_type == IBLND_MSG_PUT_DONE);
1273 switch (tx->tx_msg->ibm_type) {
1277 case IBLND_MSG_PUT_REQ:
1278 case IBLND_MSG_GET_REQ:
1279 q = &conn->ibc_tx_queue_rsrvd;
1282 case IBLND_MSG_PUT_NAK:
1283 case IBLND_MSG_PUT_ACK:
1284 case IBLND_MSG_PUT_DONE:
1285 case IBLND_MSG_GET_DONE:
1286 q = &conn->ibc_tx_queue_nocred;
1289 case IBLND_MSG_NOOP:
1290 if (IBLND_OOB_CAPABLE(conn->ibc_version))
1291 q = &conn->ibc_tx_queue_nocred;
1293 q = &conn->ibc_tx_noops;
1296 case IBLND_MSG_IMMEDIATE:
1297 q = &conn->ibc_tx_queue;
1301 list_add_tail(&tx->tx_list, q);
1305 kiblnd_queue_tx(struct kib_tx *tx, struct kib_conn *conn)
1307 spin_lock(&conn->ibc_lock);
1308 kiblnd_queue_tx_locked(tx, conn);
1309 kiblnd_check_sends_locked(conn);
1310 spin_unlock(&conn->ibc_lock);
1314 kiblnd_resolve_addr_cap(struct rdma_cm_id *cmid,
1315 struct sockaddr_in *srcaddr,
1316 struct sockaddr_in *dstaddr,
1319 unsigned short port;
1322 /* allow the port to be reused */
1323 rc = rdma_set_reuseaddr(cmid, 1);
1325 CERROR("Unable to set reuse on cmid: %d\n", rc);
1329 /* look for a free privileged port */
1330 for (port = PROT_SOCK-1; port > 0; port--) {
1331 srcaddr->sin_port = htons(port);
1332 rc = rdma_resolve_addr(cmid,
1333 (struct sockaddr *)srcaddr,
1334 (struct sockaddr *)dstaddr,
1337 CDEBUG(D_NET, "bound to port %hu\n", port);
1339 } else if (rc == -EADDRINUSE || rc == -EADDRNOTAVAIL) {
1340 CDEBUG(D_NET, "bind to port %hu failed: %d\n",
1347 CERROR("cannot bind to a free privileged port: rc = %d\n", rc);
1353 kiblnd_resolve_addr(struct rdma_cm_id *cmid,
1354 struct sockaddr_in *srcaddr,
1355 struct sockaddr_in *dstaddr,
1358 const struct cred *old_creds = NULL;
1359 struct cred *new_creds;
1362 if (!capable(CAP_NET_BIND_SERVICE)) {
1363 new_creds = prepare_kernel_cred(NULL);
1367 cap_raise(new_creds->cap_effective, CAP_NET_BIND_SERVICE);
1368 old_creds = override_creds(new_creds);
1371 rc = kiblnd_resolve_addr_cap(cmid, srcaddr, dstaddr, timeout_ms);
1374 revert_creds(old_creds);
1380 kiblnd_connect_peer(struct kib_peer_ni *peer_ni)
1382 struct rdma_cm_id *cmid;
1383 struct kib_dev *dev;
1384 struct kib_net *net = peer_ni->ibp_ni->ni_data;
1385 struct sockaddr_in srcaddr;
1386 struct sockaddr_in dstaddr;
1389 LASSERT (net != NULL);
1390 LASSERT (peer_ni->ibp_connecting > 0);
1392 cmid = kiblnd_rdma_create_id(peer_ni->ibp_ni->ni_net_ns,
1393 kiblnd_cm_callback, peer_ni,
1394 RDMA_PS_TCP, IB_QPT_RC);
1397 CERROR("Can't create CMID for %s: %ld\n",
1398 libcfs_nid2str(peer_ni->ibp_nid), PTR_ERR(cmid));
1404 memset(&srcaddr, 0, sizeof(srcaddr));
1405 srcaddr.sin_family = AF_INET;
1406 srcaddr.sin_addr.s_addr = htonl(dev->ibd_ifip);
1408 memset(&dstaddr, 0, sizeof(dstaddr));
1409 dstaddr.sin_family = AF_INET;
1410 dstaddr.sin_port = htons(*kiblnd_tunables.kib_service);
1411 dstaddr.sin_addr.s_addr = htonl(LNET_NIDADDR(peer_ni->ibp_nid));
1413 kiblnd_peer_addref(peer_ni); /* cmid's ref */
1415 if (*kiblnd_tunables.kib_use_priv_port) {
1416 rc = kiblnd_resolve_addr(cmid, &srcaddr, &dstaddr,
1417 kiblnd_timeout() * 1000);
1419 rc = rdma_resolve_addr(cmid,
1420 (struct sockaddr *)&srcaddr,
1421 (struct sockaddr *)&dstaddr,
1422 kiblnd_timeout() * 1000);
1425 /* Can't initiate address resolution: */
1426 CERROR("Can't resolve addr for %s: %d\n",
1427 libcfs_nid2str(peer_ni->ibp_nid), rc);
1434 kiblnd_peer_connect_failed(peer_ni, 1, rc);
1435 kiblnd_peer_decref(peer_ni); /* cmid's ref */
1436 rdma_destroy_id(cmid);
1439 kiblnd_peer_connect_failed(peer_ni, 1, rc);
1443 kiblnd_reconnect_peer(struct kib_peer_ni *peer_ni)
1445 rwlock_t *glock = &kiblnd_data.kib_global_lock;
1446 char *reason = NULL;
1448 unsigned long flags;
1450 write_lock_irqsave(glock, flags);
1451 if (peer_ni->ibp_reconnecting == 0) {
1452 if (peer_ni->ibp_accepting)
1453 reason = "accepting";
1454 else if (peer_ni->ibp_connecting)
1455 reason = "connecting";
1456 else if (!list_empty(&peer_ni->ibp_conns))
1457 reason = "connected";
1458 else /* connected then closed */
1464 if (peer_ni->ibp_accepting)
1465 CNETERR("Detecting race between accepting and reconnecting\n");
1466 peer_ni->ibp_reconnecting--;
1468 if (!kiblnd_peer_active(peer_ni)) {
1469 list_splice_init(&peer_ni->ibp_tx_queue, &txs);
1470 reason = "unlinked";
1474 peer_ni->ibp_connecting++;
1475 peer_ni->ibp_reconnected++;
1477 write_unlock_irqrestore(glock, flags);
1479 kiblnd_connect_peer(peer_ni);
1483 write_unlock_irqrestore(glock, flags);
1485 CWARN("Abort reconnection of %s: %s\n",
1486 libcfs_nid2str(peer_ni->ibp_nid), reason);
1487 kiblnd_txlist_done(&txs, -ECONNABORTED,
1488 LNET_MSG_STATUS_LOCAL_ABORTED);
1493 kiblnd_launch_tx(struct lnet_ni *ni, struct kib_tx *tx, lnet_nid_t nid)
1495 struct kib_peer_ni *peer_ni;
1496 struct kib_peer_ni *peer2;
1497 struct kib_conn *conn;
1498 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
1499 unsigned long flags;
1502 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
1504 /* If I get here, I've committed to send, so I complete the tx with
1505 * failure on any problems
1508 LASSERT(!tx || !tx->tx_conn); /* only set when assigned a conn */
1509 LASSERT(!tx || tx->tx_nwrq > 0); /* work items have been set up */
1511 /* First time, just use a read lock since I expect to find my peer_ni
1514 read_lock_irqsave(g_lock, flags);
1516 peer_ni = kiblnd_find_peer_locked(ni, nid);
1517 if (peer_ni != NULL && !list_empty(&peer_ni->ibp_conns)) {
1518 /* Found a peer_ni with an established connection */
1519 conn = kiblnd_get_conn_locked(peer_ni);
1520 kiblnd_conn_addref(conn); /* 1 ref for me... */
1522 read_unlock_irqrestore(g_lock, flags);
1525 kiblnd_queue_tx(tx, conn);
1526 kiblnd_conn_decref(conn); /* ...to here */
1530 read_unlock(g_lock);
1531 /* Re-try with a write lock */
1534 peer_ni = kiblnd_find_peer_locked(ni, nid);
1535 if (peer_ni != NULL) {
1536 if (list_empty(&peer_ni->ibp_conns)) {
1537 /* found a peer_ni, but it's still connecting... */
1538 LASSERT(kiblnd_peer_connecting(peer_ni));
1540 list_add_tail(&tx->tx_list,
1541 &peer_ni->ibp_tx_queue);
1542 write_unlock_irqrestore(g_lock, flags);
1544 conn = kiblnd_get_conn_locked(peer_ni);
1545 kiblnd_conn_addref(conn); /* 1 ref for me... */
1547 write_unlock_irqrestore(g_lock, flags);
1550 kiblnd_queue_tx(tx, conn);
1551 kiblnd_conn_decref(conn); /* ...to here */
1556 write_unlock_irqrestore(g_lock, flags);
1558 /* Allocate a peer_ni ready to add to the peer_ni table and retry */
1559 rc = kiblnd_create_peer(ni, &peer_ni, nid);
1561 CERROR("Can't create peer_ni %s\n", libcfs_nid2str(nid));
1563 tx->tx_status = -EHOSTUNREACH;
1565 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
1571 write_lock_irqsave(g_lock, flags);
1573 peer2 = kiblnd_find_peer_locked(ni, nid);
1574 if (peer2 != NULL) {
1575 if (list_empty(&peer2->ibp_conns)) {
1576 /* found a peer_ni, but it's still connecting... */
1577 LASSERT(kiblnd_peer_connecting(peer2));
1579 list_add_tail(&tx->tx_list,
1580 &peer2->ibp_tx_queue);
1581 write_unlock_irqrestore(g_lock, flags);
1583 conn = kiblnd_get_conn_locked(peer2);
1584 kiblnd_conn_addref(conn); /* 1 ref for me... */
1586 write_unlock_irqrestore(g_lock, flags);
1589 kiblnd_queue_tx(tx, conn);
1590 kiblnd_conn_decref(conn); /* ...to here */
1593 kiblnd_peer_decref(peer_ni);
1597 /* Brand new peer_ni */
1598 LASSERT(peer_ni->ibp_connecting == 0);
1599 tunables = &peer_ni->ibp_ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
1600 peer_ni->ibp_connecting = tunables->lnd_conns_per_peer;
1602 /* always called with a ref on ni, which prevents ni being shutdown */
1603 LASSERT(((struct kib_net *)ni->ni_data)->ibn_shutdown == 0);
1606 list_add_tail(&tx->tx_list, &peer_ni->ibp_tx_queue);
1608 kiblnd_peer_addref(peer_ni);
1609 hash_add(kiblnd_data.kib_peers, &peer_ni->ibp_list, nid);
1611 write_unlock_irqrestore(g_lock, flags);
1613 for (i = 0; i < tunables->lnd_conns_per_peer; i++)
1614 kiblnd_connect_peer(peer_ni);
1615 kiblnd_peer_decref(peer_ni);
1619 kiblnd_send(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg)
1621 struct lnet_hdr *hdr = &lntmsg->msg_hdr;
1622 int type = lntmsg->msg_type;
1623 struct lnet_process_id target = lntmsg->msg_target;
1624 int target_is_router = lntmsg->msg_target_is_router;
1625 int routing = lntmsg->msg_routing;
1626 unsigned int payload_niov = lntmsg->msg_niov;
1627 struct bio_vec *payload_kiov = lntmsg->msg_kiov;
1628 unsigned int payload_offset = lntmsg->msg_offset;
1629 unsigned int payload_nob = lntmsg->msg_len;
1630 struct kib_msg *ibmsg;
1631 struct kib_rdma_desc *rd;
1636 /* NB 'private' is different depending on what we're sending.... */
1638 CDEBUG(D_NET, "sending %d bytes in %d frags to %s\n",
1639 payload_nob, payload_niov, libcfs_id2str(target));
1641 LASSERT (payload_nob == 0 || payload_niov > 0);
1642 LASSERT (payload_niov <= LNET_MAX_IOV);
1644 /* Thread context */
1645 LASSERT (!in_interrupt());
1647 tx = kiblnd_get_idle_tx(ni, target.nid);
1649 CERROR("Can't allocate %s txd for %s\n",
1650 lnet_msgtyp2str(type),
1651 libcfs_nid2str(target.nid));
1662 LASSERT (payload_nob == 0);
1666 if (routing || target_is_router)
1667 break; /* send IMMEDIATE */
1669 /* is the REPLY message too small for RDMA? */
1670 nob = offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[lntmsg->msg_md->md_length]);
1671 if (nob <= IBLND_MSG_SIZE && !lntmsg->msg_rdma_force)
1672 break; /* send IMMEDIATE */
1674 rd = &ibmsg->ibm_u.get.ibgm_rd;
1675 rc = kiblnd_setup_rd_kiov(ni, tx, rd,
1676 lntmsg->msg_md->md_niov,
1677 lntmsg->msg_md->md_kiov,
1678 0, lntmsg->msg_md->md_length);
1680 CERROR("Can't setup GET sink for %s: %d\n",
1681 libcfs_nid2str(target.nid), rc);
1682 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
1687 nob = offsetof(struct kib_get_msg, ibgm_rd.rd_frags[rd->rd_nfrags]);
1688 ibmsg->ibm_u.get.ibgm_cookie = tx->tx_cookie;
1689 ibmsg->ibm_u.get.ibgm_hdr = *hdr;
1691 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_GET_REQ, nob);
1693 tx->tx_lntmsg[1] = lnet_create_reply_msg(ni, lntmsg);
1694 if (tx->tx_lntmsg[1] == NULL) {
1695 CERROR("Can't create reply for GET -> %s\n",
1696 libcfs_nid2str(target.nid));
1701 /* finalise lntmsg[0,1] on completion */
1702 tx->tx_lntmsg[0] = lntmsg;
1703 tx->tx_waiting = 1; /* waiting for GET_DONE */
1704 kiblnd_launch_tx(ni, tx, target.nid);
1707 case LNET_MSG_REPLY:
1709 /* Is the payload small enough not to need RDMA? */
1710 nob = offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[payload_nob]);
1711 if (nob <= IBLND_MSG_SIZE && !lntmsg->msg_rdma_force)
1712 break; /* send IMMEDIATE */
1714 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1715 payload_niov, payload_kiov,
1716 payload_offset, payload_nob);
1718 CERROR("Can't setup PUT src for %s: %d\n",
1719 libcfs_nid2str(target.nid), rc);
1724 ibmsg->ibm_u.putreq.ibprm_hdr = *hdr;
1725 ibmsg->ibm_u.putreq.ibprm_cookie = tx->tx_cookie;
1726 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_REQ,
1727 sizeof(struct kib_putreq_msg));
1729 /* finalise lntmsg[0,1] on completion */
1730 tx->tx_lntmsg[0] = lntmsg;
1731 tx->tx_waiting = 1; /* waiting for PUT_{ACK,NAK} */
1732 kiblnd_launch_tx(ni, tx, target.nid);
1736 /* send IMMEDIATE */
1737 LASSERT(offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[payload_nob])
1741 ibmsg->ibm_u.immediate.ibim_hdr = *hdr;
1743 lnet_copy_kiov2flat(IBLND_MSG_SIZE, ibmsg,
1744 offsetof(struct kib_msg,
1745 ibm_u.immediate.ibim_payload),
1746 payload_niov, payload_kiov,
1747 payload_offset, payload_nob);
1749 nob = offsetof(struct kib_immediate_msg, ibim_payload[payload_nob]);
1751 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_IMMEDIATE, nob);
1753 /* finalise lntmsg on completion */
1754 tx->tx_lntmsg[0] = lntmsg;
1756 kiblnd_launch_tx(ni, tx, target.nid);
1761 kiblnd_reply(struct lnet_ni *ni, struct kib_rx *rx, struct lnet_msg *lntmsg)
1763 struct lnet_process_id target = lntmsg->msg_target;
1764 unsigned int niov = lntmsg->msg_niov;
1765 struct bio_vec *kiov = lntmsg->msg_kiov;
1766 unsigned int offset = lntmsg->msg_offset;
1767 unsigned int nob = lntmsg->msg_len;
1771 tx = kiblnd_get_idle_tx(ni, rx->rx_conn->ibc_peer->ibp_nid);
1773 CERROR("Can't get tx for REPLY to %s\n",
1774 libcfs_nid2str(target.nid));
1781 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1782 niov, kiov, offset, nob);
1785 CERROR("Can't setup GET src for %s: %d\n",
1786 libcfs_nid2str(target.nid), rc);
1790 rc = kiblnd_init_rdma(rx->rx_conn, tx,
1791 IBLND_MSG_GET_DONE, nob,
1792 &rx->rx_msg->ibm_u.get.ibgm_rd,
1793 rx->rx_msg->ibm_u.get.ibgm_cookie);
1795 CERROR("Can't setup rdma for GET from %s: %d\n",
1796 libcfs_nid2str(target.nid), rc);
1801 /* No RDMA: local completion may happen now! */
1802 lnet_finalize(lntmsg, 0);
1804 /* RDMA: lnet_finalize(lntmsg) when it
1806 tx->tx_lntmsg[0] = lntmsg;
1809 kiblnd_queue_tx(tx, rx->rx_conn);
1814 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
1817 lnet_finalize(lntmsg, -EIO);
1821 kiblnd_get_dev_prio(struct lnet_ni *ni, unsigned int dev_idx)
1823 struct kib_net *net = ni->ni_data;
1824 struct device *dev = NULL;
1827 dev = net->ibn_dev->ibd_hdev->ibh_ibdev->dma_device;
1829 return lnet_get_dev_prio(dev, dev_idx);
1834 kiblnd_recv(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg,
1835 int delayed, unsigned int niov, struct bio_vec *kiov,
1836 unsigned int offset, unsigned int mlen, unsigned int rlen)
1838 struct kib_rx *rx = private;
1839 struct kib_msg *rxmsg = rx->rx_msg;
1840 struct kib_conn *conn = rx->rx_conn;
1844 int post_credit = IBLND_POSTRX_PEER_CREDIT;
1847 LASSERT (mlen <= rlen);
1848 LASSERT (!in_interrupt());
1850 switch (rxmsg->ibm_type) {
1854 case IBLND_MSG_IMMEDIATE:
1855 nob = offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[rlen]);
1856 if (nob > rx->rx_nob) {
1857 CERROR ("Immediate message from %s too big: %d(%d)\n",
1858 libcfs_nid2str(rxmsg->ibm_u.immediate.ibim_hdr.src_nid),
1864 lnet_copy_flat2kiov(niov, kiov, offset,
1865 IBLND_MSG_SIZE, rxmsg,
1866 offsetof(struct kib_msg,
1867 ibm_u.immediate.ibim_payload),
1869 lnet_finalize(lntmsg, 0);
1872 case IBLND_MSG_PUT_REQ: {
1873 struct kib_msg *txmsg;
1874 struct kib_rdma_desc *rd;
1875 ibprm_cookie = rxmsg->ibm_u.putreq.ibprm_cookie;
1878 lnet_finalize(lntmsg, 0);
1879 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK,
1884 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
1886 CERROR("Can't allocate tx for %s\n",
1887 libcfs_nid2str(conn->ibc_peer->ibp_nid));
1888 /* Not replying will break the connection */
1894 rd = &txmsg->ibm_u.putack.ibpam_rd;
1895 rc = kiblnd_setup_rd_kiov(ni, tx, rd,
1896 niov, kiov, offset, mlen);
1898 CERROR("Can't setup PUT sink for %s: %d\n",
1899 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
1900 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
1902 /* tell peer_ni it's over */
1903 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK,
1908 nob = offsetof(struct kib_putack_msg, ibpam_rd.rd_frags[rd->rd_nfrags]);
1909 txmsg->ibm_u.putack.ibpam_src_cookie = ibprm_cookie;
1910 txmsg->ibm_u.putack.ibpam_dst_cookie = tx->tx_cookie;
1912 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_ACK, nob);
1914 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1915 tx->tx_waiting = 1; /* waiting for PUT_DONE */
1916 kiblnd_queue_tx(tx, conn);
1918 /* reposted buffer reserved for PUT_DONE */
1919 post_credit = IBLND_POSTRX_NO_CREDIT;
1923 case IBLND_MSG_GET_REQ:
1924 if (lntmsg != NULL) {
1925 /* Optimized GET; RDMA lntmsg's payload */
1926 kiblnd_reply(ni, rx, lntmsg);
1928 /* GET didn't match anything */
1929 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_GET_DONE,
1931 rxmsg->ibm_u.get.ibgm_cookie);
1936 kiblnd_post_rx(rx, post_credit);
1941 kiblnd_thread_fini (void)
1943 atomic_dec (&kiblnd_data.kib_nthreads);
1947 kiblnd_peer_alive(struct kib_peer_ni *peer_ni)
1949 /* This is racy, but everyone's only writing ktime_get_seconds() */
1950 peer_ni->ibp_last_alive = ktime_get_seconds();
1955 kiblnd_peer_notify(struct kib_peer_ni *peer_ni)
1958 time64_t last_alive = 0;
1959 unsigned long flags;
1961 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1963 if (kiblnd_peer_idle(peer_ni) && peer_ni->ibp_error != 0) {
1964 error = peer_ni->ibp_error;
1965 peer_ni->ibp_error = 0;
1967 last_alive = peer_ni->ibp_last_alive;
1970 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1973 lnet_notify(peer_ni->ibp_ni,
1974 peer_ni->ibp_nid, false, false, last_alive);
1978 kiblnd_close_conn_locked(struct kib_conn *conn, int error)
1980 /* This just does the immediate housekeeping. 'error' is zero for a
1981 * normal shutdown which can happen only after the connection has been
1982 * established. If the connection is established, schedule the
1983 * connection to be finished off by the connd. Otherwise the connd is
1984 * already dealing with it (either to set it up or tear it down).
1985 * Caller holds kib_global_lock exclusively in irq context */
1986 struct kib_peer_ni *peer_ni = conn->ibc_peer;
1987 struct kib_dev *dev;
1988 unsigned long flags;
1990 LASSERT (error != 0 || conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1992 if (error != 0 && conn->ibc_comms_error == 0)
1993 conn->ibc_comms_error = error;
1995 if (conn->ibc_state != IBLND_CONN_ESTABLISHED)
1996 return; /* already being handled */
1999 list_empty(&conn->ibc_tx_noops) &&
2000 list_empty(&conn->ibc_tx_queue) &&
2001 list_empty(&conn->ibc_tx_queue_rsrvd) &&
2002 list_empty(&conn->ibc_tx_queue_nocred) &&
2003 list_empty(&conn->ibc_active_txs)) {
2004 CDEBUG(D_NET, "closing conn to %s\n",
2005 libcfs_nid2str(peer_ni->ibp_nid));
2007 CNETERR("Closing conn to %s: error %d%s%s%s%s%s\n",
2008 libcfs_nid2str(peer_ni->ibp_nid), error,
2009 list_empty(&conn->ibc_tx_queue) ? "" : "(sending)",
2010 list_empty(&conn->ibc_tx_noops) ? "" : "(sending_noops)",
2011 list_empty(&conn->ibc_tx_queue_rsrvd) ?
2012 "" : "(sending_rsrvd)",
2013 list_empty(&conn->ibc_tx_queue_nocred) ?
2014 "" : "(sending_nocred)",
2015 list_empty(&conn->ibc_active_txs) ? "" : "(waiting)");
2018 dev = ((struct kib_net *)peer_ni->ibp_ni->ni_data)->ibn_dev;
2019 if (peer_ni->ibp_next_conn == conn)
2020 /* clear next_conn so it won't be used */
2021 peer_ni->ibp_next_conn = NULL;
2022 list_del(&conn->ibc_list);
2023 /* connd (see below) takes over ibc_list's ref */
2025 if (list_empty(&peer_ni->ibp_conns) && /* no more conns */
2026 kiblnd_peer_active(peer_ni)) { /* still in peer_ni table */
2027 kiblnd_unlink_peer_locked(peer_ni);
2029 /* set/clear error on last conn */
2030 peer_ni->ibp_error = conn->ibc_comms_error;
2033 kiblnd_set_conn_state(conn, IBLND_CONN_CLOSING);
2036 kiblnd_dev_can_failover(dev)) {
2037 list_add_tail(&dev->ibd_fail_list,
2038 &kiblnd_data.kib_failed_devs);
2039 wake_up(&kiblnd_data.kib_failover_waitq);
2042 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
2044 list_add_tail(&conn->ibc_list, &kiblnd_data.kib_connd_conns);
2045 wake_up(&kiblnd_data.kib_connd_waitq);
2047 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
2051 kiblnd_close_conn(struct kib_conn *conn, int error)
2053 unsigned long flags;
2055 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2057 kiblnd_close_conn_locked(conn, error);
2059 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2063 kiblnd_handle_early_rxs(struct kib_conn *conn)
2065 unsigned long flags;
2068 LASSERT(!in_interrupt());
2069 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
2071 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2072 while ((rx = list_first_entry_or_null(&conn->ibc_early_rxs,
2074 rx_list)) != NULL) {
2075 list_del(&rx->rx_list);
2076 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2078 kiblnd_handle_rx(rx);
2080 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2082 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2086 kiblnd_abort_txs(struct kib_conn *conn, struct list_head *txs)
2092 spin_lock(&conn->ibc_lock);
2094 list_for_each_entry_safe(tx, nxt, txs, tx_list) {
2095 if (txs == &conn->ibc_active_txs) {
2096 LASSERT(!tx->tx_queued);
2097 LASSERT(tx->tx_waiting ||
2098 tx->tx_sending != 0);
2099 if (conn->ibc_comms_error == -ETIMEDOUT) {
2100 if (tx->tx_waiting && !tx->tx_sending)
2102 LNET_MSG_STATUS_REMOTE_TIMEOUT;
2103 else if (tx->tx_sending)
2105 LNET_MSG_STATUS_NETWORK_TIMEOUT;
2108 LASSERT(tx->tx_queued);
2109 if (conn->ibc_comms_error == -ETIMEDOUT)
2110 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_TIMEOUT;
2112 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
2115 tx->tx_status = -ECONNABORTED;
2119 * TODO: This makes an assumption that
2120 * kiblnd_tx_complete() will be called for each tx. If
2121 * that event is dropped we could end up with stale
2122 * connections floating around. We'd like to deal with
2123 * that in a better way.
2125 * Also that means we can exceed the timeout by many
2128 if (tx->tx_sending == 0) {
2130 list_move(&tx->tx_list, &zombies);
2132 /* keep tx until cq destroy */
2133 list_move(&tx->tx_list, &conn->ibc_zombie_txs);
2138 spin_unlock(&conn->ibc_lock);
2141 * aborting transmits occurs when finalizing the connection.
2142 * The connection is finalized on error.
2143 * Passing LNET_MSG_STATUS_OK to txlist_done() will not
2144 * override the value already set in tx->tx_hstatus above.
2146 kiblnd_txlist_done(&zombies, -ECONNABORTED, LNET_MSG_STATUS_OK);
2150 kiblnd_tx_may_discard(struct kib_conn *conn)
2156 spin_lock(&conn->ibc_lock);
2158 list_for_each_entry_safe(tx, nxt, &conn->ibc_zombie_txs, tx_list) {
2159 if (tx->tx_sending > 0 && tx->tx_lntmsg[0] &&
2160 lnet_md_discarded(tx->tx_lntmsg[0]->msg_md)) {
2162 if (tx->tx_sending == 0) {
2163 kiblnd_conn_decref(tx->tx_conn);
2170 spin_unlock(&conn->ibc_lock);
2175 kiblnd_finalise_conn(struct kib_conn *conn)
2177 LASSERT (!in_interrupt());
2178 LASSERT (conn->ibc_state > IBLND_CONN_INIT);
2180 /* abort_receives moves QP state to IB_QPS_ERR. This is only required
2181 * for connections that didn't get as far as being connected, because
2182 * rdma_disconnect() does this for free. */
2183 kiblnd_abort_receives(conn);
2185 kiblnd_set_conn_state(conn, IBLND_CONN_DISCONNECTED);
2187 /* Complete all tx descs not waiting for sends to complete.
2188 * NB we should be safe from RDMA now that the QP has changed state */
2190 kiblnd_abort_txs(conn, &conn->ibc_tx_noops);
2191 kiblnd_abort_txs(conn, &conn->ibc_tx_queue);
2192 kiblnd_abort_txs(conn, &conn->ibc_tx_queue_rsrvd);
2193 kiblnd_abort_txs(conn, &conn->ibc_tx_queue_nocred);
2194 kiblnd_abort_txs(conn, &conn->ibc_active_txs);
2196 kiblnd_handle_early_rxs(conn);
2200 kiblnd_peer_connect_failed(struct kib_peer_ni *peer_ni, int active,
2204 unsigned long flags;
2205 enum lnet_msg_hstatus hstatus;
2207 LASSERT(error != 0);
2208 LASSERT(!in_interrupt());
2210 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2213 LASSERT(peer_ni->ibp_connecting > 0);
2214 peer_ni->ibp_connecting--;
2216 LASSERT (peer_ni->ibp_accepting > 0);
2217 peer_ni->ibp_accepting--;
2220 if (kiblnd_peer_connecting(peer_ni)) {
2221 /* another connection attempt under way... */
2222 write_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2227 peer_ni->ibp_reconnected = 0;
2228 if (list_empty(&peer_ni->ibp_conns)) {
2229 /* Take peer_ni's blocked transmits to complete with error */
2230 list_splice_init(&peer_ni->ibp_tx_queue, &zombies);
2232 if (kiblnd_peer_active(peer_ni))
2233 kiblnd_unlink_peer_locked(peer_ni);
2235 peer_ni->ibp_error = error;
2237 /* Can't have blocked transmits if there are connections */
2238 LASSERT(list_empty(&peer_ni->ibp_tx_queue));
2241 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2243 kiblnd_peer_notify(peer_ni);
2245 if (list_empty(&zombies))
2248 CNETERR("Deleting messages for %s: connection failed\n",
2249 libcfs_nid2str(peer_ni->ibp_nid));
2254 hstatus = LNET_MSG_STATUS_NETWORK_TIMEOUT;
2257 hstatus = LNET_MSG_STATUS_REMOTE_DROPPED;
2260 hstatus = LNET_MSG_STATUS_LOCAL_DROPPED;
2264 kiblnd_txlist_done(&zombies, error, hstatus);
2268 kiblnd_connreq_done(struct kib_conn *conn, int status)
2270 struct kib_peer_ni *peer_ni = conn->ibc_peer;
2273 unsigned long flags;
2276 active = (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2278 CDEBUG(D_NET,"%s: active(%d), version(%x), status(%d)\n",
2279 libcfs_nid2str(peer_ni->ibp_nid), active,
2280 conn->ibc_version, status);
2282 LASSERT (!in_interrupt());
2283 LASSERT ((conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT &&
2284 peer_ni->ibp_connecting > 0) ||
2285 (conn->ibc_state == IBLND_CONN_PASSIVE_WAIT &&
2286 peer_ni->ibp_accepting > 0));
2288 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
2289 conn->ibc_connvars = NULL;
2292 /* failed to establish connection */
2293 kiblnd_peer_connect_failed(peer_ni, active, status);
2294 kiblnd_finalise_conn(conn);
2298 /* connection established */
2299 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2301 conn->ibc_last_send = ktime_get();
2302 kiblnd_set_conn_state(conn, IBLND_CONN_ESTABLISHED);
2303 kiblnd_peer_alive(peer_ni);
2305 /* Add conn to peer_ni's list and nuke any dangling conns from a different
2306 * peer_ni instance... */
2307 kiblnd_conn_addref(conn); /* +1 ref for ibc_list */
2308 list_add(&conn->ibc_list, &peer_ni->ibp_conns);
2309 peer_ni->ibp_reconnected = 0;
2311 peer_ni->ibp_connecting--;
2313 peer_ni->ibp_accepting--;
2315 if (peer_ni->ibp_version == 0) {
2316 peer_ni->ibp_version = conn->ibc_version;
2317 peer_ni->ibp_incarnation = conn->ibc_incarnation;
2320 if (peer_ni->ibp_version != conn->ibc_version ||
2321 peer_ni->ibp_incarnation != conn->ibc_incarnation) {
2322 kiblnd_close_stale_conns_locked(peer_ni, conn->ibc_version,
2323 conn->ibc_incarnation);
2324 peer_ni->ibp_version = conn->ibc_version;
2325 peer_ni->ibp_incarnation = conn->ibc_incarnation;
2328 /* grab pending txs while I have the lock */
2329 list_splice_init(&peer_ni->ibp_tx_queue, &txs);
2331 if (!kiblnd_peer_active(peer_ni) || /* peer_ni has been deleted */
2332 conn->ibc_comms_error != 0) { /* error has happened already */
2334 /* start to shut down connection */
2335 kiblnd_close_conn_locked(conn, -ECONNABORTED);
2336 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2338 kiblnd_txlist_done(&txs, -ECONNABORTED,
2339 LNET_MSG_STATUS_LOCAL_ERROR);
2344 /* +1 ref for myself, this connection is visible to other threads
2345 * now, refcount of peer:ibp_conns can be released by connection
2346 * close from either a different thread, or the calling of
2347 * kiblnd_check_sends_locked() below. See bz21911 for details.
2349 kiblnd_conn_addref(conn);
2350 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2352 /* Schedule blocked txs
2353 * Note: if we are running with conns_per_peer > 1, these blocked
2354 * txs will all get scheduled to the first connection which gets
2355 * scheduled. We won't be using round robin on this first batch.
2357 spin_lock(&conn->ibc_lock);
2358 while ((tx = list_first_entry_or_null(&txs, struct kib_tx,
2359 tx_list)) != NULL) {
2360 list_del(&tx->tx_list);
2362 kiblnd_queue_tx_locked(tx, conn);
2364 kiblnd_check_sends_locked(conn);
2365 spin_unlock(&conn->ibc_lock);
2367 /* schedule blocked rxs */
2368 kiblnd_handle_early_rxs(conn);
2369 kiblnd_conn_decref(conn);
2373 kiblnd_reject(struct rdma_cm_id *cmid, struct kib_rej *rej)
2377 #ifdef HAVE_RDMA_REJECT_4ARGS
2378 rc = rdma_reject(cmid, rej, sizeof(*rej), IB_CM_REJ_CONSUMER_DEFINED);
2380 rc = rdma_reject(cmid, rej, sizeof(*rej));
2384 CWARN("Error %d sending reject\n", rc);
2388 kiblnd_passive_connect(struct rdma_cm_id *cmid, void *priv, int priv_nob)
2390 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2391 struct kib_msg *reqmsg = priv;
2392 struct kib_msg *ackmsg;
2393 struct kib_dev *ibdev;
2394 struct kib_peer_ni *peer_ni;
2395 struct kib_peer_ni *peer2;
2396 struct kib_conn *conn;
2397 struct lnet_ni *ni = NULL;
2398 struct kib_net *net = NULL;
2400 struct rdma_conn_param cp;
2402 int version = IBLND_MSG_VERSION;
2403 unsigned long flags;
2405 struct sockaddr_in *peer_addr;
2407 LASSERT(!in_interrupt());
2408 /* cmid inherits 'context' from the corresponding listener id */
2409 ibdev = cmid->context;
2412 memset(&rej, 0, sizeof(rej));
2413 rej.ibr_magic = IBLND_MSG_MAGIC;
2414 rej.ibr_why = IBLND_REJECT_FATAL;
2415 rej.ibr_cp.ibcp_max_msg_size = IBLND_MSG_SIZE;
2417 peer_addr = (struct sockaddr_in *)&(cmid->route.addr.dst_addr);
2418 if (*kiblnd_tunables.kib_require_priv_port &&
2419 ntohs(peer_addr->sin_port) >= PROT_SOCK) {
2420 __u32 ip = ntohl(peer_addr->sin_addr.s_addr);
2421 CERROR("peer_ni's port (%pI4h:%hu) is not privileged\n",
2422 &ip, ntohs(peer_addr->sin_port));
2426 if (priv_nob < offsetof(struct kib_msg, ibm_type)) {
2427 CERROR("Short connection request\n");
2431 /* Future protocol version compatibility support! If the
2432 * o2iblnd-specific protocol changes, or when LNET unifies
2433 * protocols over all LNDs, the initial connection will
2434 * negotiate a protocol version. I trap this here to avoid
2435 * console errors; the reject tells the peer_ni which protocol I
2437 if (reqmsg->ibm_magic == LNET_PROTO_MAGIC ||
2438 reqmsg->ibm_magic == __swab32(LNET_PROTO_MAGIC))
2440 if (reqmsg->ibm_magic == IBLND_MSG_MAGIC &&
2441 reqmsg->ibm_version != IBLND_MSG_VERSION &&
2442 reqmsg->ibm_version != IBLND_MSG_VERSION_1)
2444 if (reqmsg->ibm_magic == __swab32(IBLND_MSG_MAGIC) &&
2445 reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION) &&
2446 reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION_1))
2449 rc = kiblnd_unpack_msg(reqmsg, priv_nob);
2451 CERROR("Can't parse connection request: %d\n", rc);
2455 nid = reqmsg->ibm_srcnid;
2456 ni = lnet_nid2ni_addref(reqmsg->ibm_dstnid);
2459 net = (struct kib_net *)ni->ni_data;
2460 rej.ibr_incarnation = net->ibn_incarnation;
2463 if (ni == NULL || /* no matching net */
2464 lnet_nid_to_nid4(&ni->ni_nid) !=
2465 reqmsg->ibm_dstnid || /* right NET, wrong NID! */
2466 net->ibn_dev != ibdev) { /* wrong device */
2467 CERROR("Can't accept conn from %s on %s (%s:%d:%pI4h): bad dst nid %s\n", libcfs_nid2str(nid),
2468 ni ? libcfs_nidstr(&ni->ni_nid) : "NA",
2469 ibdev->ibd_ifname, ibdev->ibd_nnets,
2471 libcfs_nid2str(reqmsg->ibm_dstnid));
2476 /* check time stamp as soon as possible */
2477 if (reqmsg->ibm_dststamp != 0 &&
2478 reqmsg->ibm_dststamp != net->ibn_incarnation) {
2479 CWARN("Stale connection request\n");
2480 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2484 /* I can accept peer_ni's version */
2485 version = reqmsg->ibm_version;
2487 if (reqmsg->ibm_type != IBLND_MSG_CONNREQ) {
2488 CERROR("Unexpected connreq msg type: %x from %s\n",
2489 reqmsg->ibm_type, libcfs_nid2str(nid));
2493 if (reqmsg->ibm_u.connparams.ibcp_queue_depth >
2494 kiblnd_msg_queue_size(version, ni)) {
2495 CERROR("Can't accept conn from %s, queue depth too large: %d (<=%d wanted)\n",
2496 libcfs_nid2str(nid),
2497 reqmsg->ibm_u.connparams.ibcp_queue_depth,
2498 kiblnd_msg_queue_size(version, ni));
2500 if (version == IBLND_MSG_VERSION)
2501 rej.ibr_why = IBLND_REJECT_MSG_QUEUE_SIZE;
2506 if (reqmsg->ibm_u.connparams.ibcp_max_frags >
2507 IBLND_MAX_RDMA_FRAGS) {
2508 CWARN("Can't accept conn from %s (version %x): max_frags %d too large (%d wanted)\n",
2509 libcfs_nid2str(nid), version,
2510 reqmsg->ibm_u.connparams.ibcp_max_frags,
2511 IBLND_MAX_RDMA_FRAGS);
2513 if (version >= IBLND_MSG_VERSION)
2514 rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2517 } else if (reqmsg->ibm_u.connparams.ibcp_max_frags <
2518 IBLND_MAX_RDMA_FRAGS &&
2519 net->ibn_fmr_ps == NULL) {
2520 CWARN("Can't accept conn from %s (version %x): max_frags %d incompatible without FMR pool (%d wanted)\n",
2521 libcfs_nid2str(nid), version,
2522 reqmsg->ibm_u.connparams.ibcp_max_frags,
2523 IBLND_MAX_RDMA_FRAGS);
2525 if (version == IBLND_MSG_VERSION)
2526 rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2531 if (reqmsg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2532 CERROR("Can't accept %s: message size %d too big (%d max)\n",
2533 libcfs_nid2str(nid),
2534 reqmsg->ibm_u.connparams.ibcp_max_msg_size,
2539 /* assume 'nid' is a new peer_ni; create */
2540 rc = kiblnd_create_peer(ni, &peer_ni, nid);
2542 CERROR("Can't create peer_ni for %s\n", libcfs_nid2str(nid));
2543 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2547 /* We have validated the peer's parameters so use those */
2548 peer_ni->ibp_max_frags = reqmsg->ibm_u.connparams.ibcp_max_frags;
2549 peer_ni->ibp_queue_depth = reqmsg->ibm_u.connparams.ibcp_queue_depth;
2551 write_lock_irqsave(g_lock, flags);
2553 peer2 = kiblnd_find_peer_locked(ni, nid);
2554 if (peer2 != NULL) {
2555 if (peer2->ibp_version == 0) {
2556 peer2->ibp_version = version;
2557 peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2560 /* not the guy I've talked with */
2561 if (peer2->ibp_incarnation != reqmsg->ibm_srcstamp ||
2562 peer2->ibp_version != version) {
2563 kiblnd_close_peer_conns_locked(peer2, -ESTALE);
2565 if (kiblnd_peer_active(peer2)) {
2566 peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2567 peer2->ibp_version = version;
2569 write_unlock_irqrestore(g_lock, flags);
2571 CWARN("Conn stale %s version %x/%x incarnation %llu/%llu\n",
2572 libcfs_nid2str(nid), peer2->ibp_version, version,
2573 peer2->ibp_incarnation, reqmsg->ibm_srcstamp);
2575 kiblnd_peer_decref(peer_ni);
2576 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2580 /* Tie-break connection race in favour of the higher NID.
2581 * If we keep running into a race condition multiple times,
2582 * we have to assume that the connection attempt with the
2583 * higher NID is stuck in a connecting state and will never
2584 * recover. As such, we pass through this if-block and let
2585 * the lower NID connection win so we can move forward.
2587 if (peer2->ibp_connecting != 0 &&
2588 nid < lnet_nid_to_nid4(&ni->ni_nid) &&
2589 peer2->ibp_races < MAX_CONN_RACES_BEFORE_ABORT) {
2591 write_unlock_irqrestore(g_lock, flags);
2593 CDEBUG(D_NET, "Conn race %s\n",
2594 libcfs_nid2str(peer2->ibp_nid));
2596 kiblnd_peer_decref(peer_ni);
2597 rej.ibr_why = IBLND_REJECT_CONN_RACE;
2600 if (peer2->ibp_races >= MAX_CONN_RACES_BEFORE_ABORT)
2601 CNETERR("Conn race %s: unresolved after %d attempts, letting lower NID win\n",
2602 libcfs_nid2str(peer2->ibp_nid),
2603 MAX_CONN_RACES_BEFORE_ABORT);
2605 * passive connection is allowed even this peer_ni is waiting for
2608 peer2->ibp_reconnecting = 0;
2609 peer2->ibp_races = 0;
2610 peer2->ibp_accepting++;
2611 kiblnd_peer_addref(peer2);
2613 /* Race with kiblnd_launch_tx (active connect) to create peer_ni
2614 * so copy validated parameters since we now know what the
2615 * peer_ni's limits are */
2616 peer2->ibp_max_frags = peer_ni->ibp_max_frags;
2617 peer2->ibp_queue_depth = peer_ni->ibp_queue_depth;
2619 write_unlock_irqrestore(g_lock, flags);
2620 kiblnd_peer_decref(peer_ni);
2623 /* Brand new peer_ni */
2624 LASSERT(peer_ni->ibp_accepting == 0);
2625 LASSERT(peer_ni->ibp_version == 0 &&
2626 peer_ni->ibp_incarnation == 0);
2628 peer_ni->ibp_accepting = 1;
2629 peer_ni->ibp_version = version;
2630 peer_ni->ibp_incarnation = reqmsg->ibm_srcstamp;
2632 /* I have a ref on ni that prevents it being shutdown */
2633 LASSERT(net->ibn_shutdown == 0);
2635 kiblnd_peer_addref(peer_ni);
2636 hash_add(kiblnd_data.kib_peers, &peer_ni->ibp_list, nid);
2638 write_unlock_irqrestore(g_lock, flags);
2641 conn = kiblnd_create_conn(peer_ni, cmid, IBLND_CONN_PASSIVE_WAIT,
2644 kiblnd_peer_connect_failed(peer_ni, 0, -ENOMEM);
2645 kiblnd_peer_decref(peer_ni);
2646 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2650 /* conn now "owns" cmid, so I return success from here on to ensure the
2651 * CM callback doesn't destroy cmid.
2653 conn->ibc_incarnation = reqmsg->ibm_srcstamp;
2654 conn->ibc_credits = conn->ibc_queue_depth;
2655 conn->ibc_reserved_credits = conn->ibc_queue_depth;
2656 LASSERT(conn->ibc_credits + conn->ibc_reserved_credits +
2657 IBLND_OOB_MSGS(version) <= IBLND_RX_MSGS(conn));
2659 ackmsg = &conn->ibc_connvars->cv_msg;
2660 memset(ackmsg, 0, sizeof(*ackmsg));
2662 kiblnd_init_msg(ackmsg, IBLND_MSG_CONNACK,
2663 sizeof(ackmsg->ibm_u.connparams));
2664 ackmsg->ibm_u.connparams.ibcp_queue_depth = conn->ibc_queue_depth;
2665 ackmsg->ibm_u.connparams.ibcp_max_frags = conn->ibc_max_frags;
2666 ackmsg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2668 kiblnd_pack_msg(ni, ackmsg, version, 0, nid, reqmsg->ibm_srcstamp);
2670 memset(&cp, 0, sizeof(cp));
2671 cp.private_data = ackmsg;
2672 cp.private_data_len = ackmsg->ibm_nob;
2673 cp.responder_resources = 0; /* No atomic ops or RDMA reads */
2674 cp.initiator_depth = 0;
2675 cp.flow_control = 1;
2676 cp.retry_count = *kiblnd_tunables.kib_retry_count;
2677 cp.rnr_retry_count = *kiblnd_tunables.kib_rnr_retry_count;
2679 CDEBUG(D_NET, "Accept %s\n", libcfs_nid2str(nid));
2681 rc = rdma_accept(cmid, &cp);
2683 CERROR("Can't accept %s: %d\n", libcfs_nid2str(nid), rc);
2684 rej.ibr_version = version;
2685 rej.ibr_why = IBLND_REJECT_FATAL;
2687 kiblnd_reject(cmid, &rej);
2688 kiblnd_connreq_done(conn, rc);
2689 kiblnd_conn_decref(conn);
2697 rej.ibr_cp.ibcp_queue_depth =
2698 kiblnd_msg_queue_size(version, ni);
2699 rej.ibr_cp.ibcp_max_frags = IBLND_MAX_RDMA_FRAGS;
2703 rej.ibr_version = version;
2704 kiblnd_reject(cmid, &rej);
2706 return -ECONNREFUSED;
2710 kiblnd_check_reconnect(struct kib_conn *conn, int version,
2711 u64 incarnation, int why, struct kib_connparams *cp)
2713 rwlock_t *glock = &kiblnd_data.kib_global_lock;
2714 struct kib_peer_ni *peer_ni = conn->ibc_peer;
2716 int msg_size = IBLND_MSG_SIZE;
2720 unsigned long flags;
2722 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2723 LASSERT(peer_ni->ibp_connecting > 0); /* 'conn' at least */
2726 msg_size = cp->ibcp_max_msg_size;
2727 frag_num = cp->ibcp_max_frags;
2728 queue_dep = cp->ibcp_queue_depth;
2731 write_lock_irqsave(glock, flags);
2732 /* retry connection if it's still needed and no other connection
2733 * attempts (active or passive) are in progress
2734 * NB: reconnect is still needed even when ibp_tx_queue is
2735 * empty if ibp_version != version because reconnect may be
2738 reconnect = (!list_empty(&peer_ni->ibp_tx_queue) ||
2739 peer_ni->ibp_version != version) &&
2740 peer_ni->ibp_connecting &&
2741 peer_ni->ibp_accepting == 0;
2752 case IBLND_REJECT_RDMA_FRAGS: {
2754 reason = "can't negotiate max frags";
2758 if (conn->ibc_max_frags <= frag_num) {
2759 reason = "unsupported max frags";
2763 peer_ni->ibp_max_frags = frag_num;
2764 reason = "rdma fragments";
2767 case IBLND_REJECT_MSG_QUEUE_SIZE:
2769 reason = "can't negotiate queue depth";
2772 if (conn->ibc_queue_depth <= queue_dep) {
2773 reason = "unsupported queue depth";
2777 peer_ni->ibp_queue_depth = queue_dep;
2778 reason = "queue depth";
2781 case IBLND_REJECT_CONN_STALE:
2785 case IBLND_REJECT_CONN_RACE:
2786 reason = "conn race";
2789 case IBLND_REJECT_CONN_UNCOMPAT:
2790 reason = "version negotiation";
2794 conn->ibc_reconnect = 1;
2795 peer_ni->ibp_reconnecting++;
2796 peer_ni->ibp_version = version;
2797 if (incarnation != 0)
2798 peer_ni->ibp_incarnation = incarnation;
2800 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2802 CNETERR("%s: %s (%s), %x, %x, msg_size: %d, queue_depth: %d/%d, max_frags: %d/%d\n",
2803 libcfs_nid2str(peer_ni->ibp_nid),
2804 reconnect ? "reconnect" : "don't reconnect",
2805 reason, IBLND_MSG_VERSION, version, msg_size,
2806 conn->ibc_queue_depth, queue_dep,
2807 conn->ibc_max_frags, frag_num);
2809 * if conn::ibc_reconnect is TRUE, connd will reconnect to the peer_ni
2810 * while destroying the zombie
2815 kiblnd_rejected(struct kib_conn *conn, int reason, void *priv, int priv_nob)
2817 struct kib_peer_ni *peer_ni = conn->ibc_peer;
2818 int status = -ECONNREFUSED;
2820 LASSERT (!in_interrupt());
2821 LASSERT (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2824 case IB_CM_REJ_STALE_CONN:
2825 kiblnd_check_reconnect(conn, IBLND_MSG_VERSION, 0,
2826 IBLND_REJECT_CONN_STALE, NULL);
2829 case IB_CM_REJ_INVALID_SERVICE_ID:
2830 status = -EHOSTUNREACH;
2831 CNETERR("%s rejected: no listener at %d\n",
2832 libcfs_nid2str(peer_ni->ibp_nid),
2833 *kiblnd_tunables.kib_service);
2836 case IB_CM_REJ_CONSUMER_DEFINED:
2837 if (priv_nob >= offsetof(struct kib_rej, ibr_padding)) {
2838 struct kib_rej *rej = priv;
2839 struct kib_connparams *cp = NULL;
2841 __u64 incarnation = -1;
2843 /* NB. default incarnation is -1 because:
2844 * a) V1 will ignore dst incarnation in connreq.
2845 * b) V2 will provide incarnation while rejecting me,
2846 * -1 will be overwrote.
2848 * if I try to connect to a V1 peer_ni with V2 protocol,
2849 * it rejected me then upgrade to V2, I have no idea
2850 * about the upgrading and try to reconnect with V1,
2851 * in this case upgraded V2 can find out I'm trying to
2852 * talk to the old guy and reject me(incarnation is -1).
2855 if (rej->ibr_magic == __swab32(IBLND_MSG_MAGIC) ||
2856 rej->ibr_magic == __swab32(LNET_PROTO_MAGIC)) {
2857 __swab32s(&rej->ibr_magic);
2858 __swab16s(&rej->ibr_version);
2862 if (priv_nob >= sizeof(struct kib_rej) &&
2863 rej->ibr_version > IBLND_MSG_VERSION_1) {
2864 /* priv_nob is always 148 in current version
2865 * of OFED, so we still need to check version.
2866 * (define of IB_CM_REJ_PRIVATE_DATA_SIZE)
2871 __swab64s(&rej->ibr_incarnation);
2872 __swab16s(&cp->ibcp_queue_depth);
2873 __swab16s(&cp->ibcp_max_frags);
2874 __swab32s(&cp->ibcp_max_msg_size);
2877 incarnation = rej->ibr_incarnation;
2880 if (rej->ibr_magic != IBLND_MSG_MAGIC &&
2881 rej->ibr_magic != LNET_PROTO_MAGIC) {
2882 CERROR("%s rejected: consumer defined fatal error\n",
2883 libcfs_nid2str(peer_ni->ibp_nid));
2887 if (rej->ibr_version != IBLND_MSG_VERSION &&
2888 rej->ibr_version != IBLND_MSG_VERSION_1) {
2889 CERROR("%s rejected: o2iblnd version %x error\n",
2890 libcfs_nid2str(peer_ni->ibp_nid),
2895 if (rej->ibr_why == IBLND_REJECT_FATAL &&
2896 rej->ibr_version == IBLND_MSG_VERSION_1) {
2897 CDEBUG(D_NET, "rejected by old version peer_ni %s: %x\n",
2898 libcfs_nid2str(peer_ni->ibp_nid),
2901 if (conn->ibc_version != IBLND_MSG_VERSION_1)
2902 rej->ibr_why = IBLND_REJECT_CONN_UNCOMPAT;
2905 switch (rej->ibr_why) {
2906 case IBLND_REJECT_CONN_RACE:
2907 case IBLND_REJECT_CONN_STALE:
2908 case IBLND_REJECT_CONN_UNCOMPAT:
2909 case IBLND_REJECT_MSG_QUEUE_SIZE:
2910 case IBLND_REJECT_RDMA_FRAGS:
2911 kiblnd_check_reconnect(conn, rej->ibr_version,
2916 case IBLND_REJECT_NO_RESOURCES:
2917 CERROR("%s rejected: o2iblnd no resources\n",
2918 libcfs_nid2str(peer_ni->ibp_nid));
2921 case IBLND_REJECT_FATAL:
2922 CERROR("%s rejected: o2iblnd fatal error\n",
2923 libcfs_nid2str(peer_ni->ibp_nid));
2927 CERROR("%s rejected: o2iblnd reason %d\n",
2928 libcfs_nid2str(peer_ni->ibp_nid),
2936 CNETERR("%s rejected: reason %d, size %d\n",
2937 libcfs_nid2str(peer_ni->ibp_nid), reason, priv_nob);
2941 kiblnd_connreq_done(conn, status);
2945 kiblnd_check_connreply(struct kib_conn *conn, void *priv, int priv_nob)
2947 struct kib_peer_ni *peer_ni = conn->ibc_peer;
2948 struct lnet_ni *ni = peer_ni->ibp_ni;
2949 struct kib_net *net = ni->ni_data;
2950 struct kib_msg *msg = priv;
2951 int ver = conn->ibc_version;
2952 int rc = kiblnd_unpack_msg(msg, priv_nob);
2953 unsigned long flags;
2955 LASSERT (net != NULL);
2958 CERROR("Can't unpack connack from %s: %d\n",
2959 libcfs_nid2str(peer_ni->ibp_nid), rc);
2963 if (msg->ibm_type != IBLND_MSG_CONNACK) {
2964 CERROR("Unexpected message %d from %s\n",
2965 msg->ibm_type, libcfs_nid2str(peer_ni->ibp_nid));
2970 if (ver != msg->ibm_version) {
2971 CERROR("%s replied version %x is different with "
2972 "requested version %x\n",
2973 libcfs_nid2str(peer_ni->ibp_nid), msg->ibm_version, ver);
2978 if (msg->ibm_u.connparams.ibcp_queue_depth >
2979 conn->ibc_queue_depth) {
2980 CERROR("%s has incompatible queue depth %d (<=%d wanted)\n",
2981 libcfs_nid2str(peer_ni->ibp_nid),
2982 msg->ibm_u.connparams.ibcp_queue_depth,
2983 conn->ibc_queue_depth);
2988 if (msg->ibm_u.connparams.ibcp_max_frags >
2989 conn->ibc_max_frags) {
2990 CERROR("%s has incompatible max_frags %d (<=%d wanted)\n",
2991 libcfs_nid2str(peer_ni->ibp_nid),
2992 msg->ibm_u.connparams.ibcp_max_frags,
2993 conn->ibc_max_frags);
2998 if (msg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2999 CERROR("%s max message size %d too big (%d max)\n",
3000 libcfs_nid2str(peer_ni->ibp_nid),
3001 msg->ibm_u.connparams.ibcp_max_msg_size,
3007 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3008 if (msg->ibm_dstnid == lnet_nid_to_nid4(&ni->ni_nid) &&
3009 msg->ibm_dststamp == net->ibn_incarnation)
3013 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3016 CERROR("Bad connection reply from %s, rc = %d, "
3017 "version: %x max_frags: %d\n",
3018 libcfs_nid2str(peer_ni->ibp_nid), rc,
3019 msg->ibm_version, msg->ibm_u.connparams.ibcp_max_frags);
3023 conn->ibc_incarnation = msg->ibm_srcstamp;
3024 conn->ibc_credits = msg->ibm_u.connparams.ibcp_queue_depth;
3025 conn->ibc_reserved_credits = msg->ibm_u.connparams.ibcp_queue_depth;
3026 conn->ibc_queue_depth = msg->ibm_u.connparams.ibcp_queue_depth;
3027 conn->ibc_max_frags = msg->ibm_u.connparams.ibcp_max_frags;
3028 LASSERT(conn->ibc_credits + conn->ibc_reserved_credits +
3029 IBLND_OOB_MSGS(ver) <= IBLND_RX_MSGS(conn));
3031 kiblnd_connreq_done(conn, 0);
3035 /* NB My QP has already established itself, so I handle anything going
3036 * wrong here by setting ibc_comms_error.
3037 * kiblnd_connreq_done(0) moves the conn state to ESTABLISHED, but then
3038 * immediately tears it down. */
3041 conn->ibc_comms_error = rc;
3042 kiblnd_connreq_done(conn, 0);
3046 kiblnd_active_connect(struct rdma_cm_id *cmid)
3048 struct kib_peer_ni *peer_ni = cmid->context;
3049 struct kib_conn *conn;
3050 struct kib_msg *msg;
3051 struct rdma_conn_param cp;
3054 unsigned long flags;
3057 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3059 incarnation = peer_ni->ibp_incarnation;
3060 version = (peer_ni->ibp_version == 0) ? IBLND_MSG_VERSION :
3061 peer_ni->ibp_version;
3063 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3065 conn = kiblnd_create_conn(peer_ni, cmid, IBLND_CONN_ACTIVE_CONNECT,
3068 kiblnd_peer_connect_failed(peer_ni, 1, -ENOMEM);
3069 kiblnd_peer_decref(peer_ni); /* lose cmid's ref */
3073 /* conn "owns" cmid now, so I return success from here on to ensure the
3074 * CM callback doesn't destroy cmid. conn also takes over cmid's ref
3077 msg = &conn->ibc_connvars->cv_msg;
3079 memset(msg, 0, sizeof(*msg));
3080 kiblnd_init_msg(msg, IBLND_MSG_CONNREQ, sizeof(msg->ibm_u.connparams));
3081 msg->ibm_u.connparams.ibcp_queue_depth = conn->ibc_queue_depth;
3082 msg->ibm_u.connparams.ibcp_max_frags = conn->ibc_max_frags;
3083 msg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
3085 kiblnd_pack_msg(peer_ni->ibp_ni, msg, version,
3086 0, peer_ni->ibp_nid, incarnation);
3088 memset(&cp, 0, sizeof(cp));
3089 cp.private_data = msg;
3090 cp.private_data_len = msg->ibm_nob;
3091 cp.responder_resources = 0; /* No atomic ops or RDMA reads */
3092 cp.initiator_depth = 0;
3093 cp.flow_control = 1;
3094 cp.retry_count = *kiblnd_tunables.kib_retry_count;
3095 cp.rnr_retry_count = *kiblnd_tunables.kib_rnr_retry_count;
3097 LASSERT(cmid->context == (void *)conn);
3098 LASSERT(conn->ibc_cmid == cmid);
3099 rc = rdma_connect_locked(cmid, &cp);
3101 CERROR("Can't connect to %s: %d\n",
3102 libcfs_nid2str(peer_ni->ibp_nid), rc);
3103 kiblnd_connreq_done(conn, rc);
3104 kiblnd_conn_decref(conn);
3111 kiblnd_cm_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
3113 struct kib_peer_ni *peer_ni;
3114 struct kib_conn *conn;
3117 switch (event->event) {
3119 CERROR("Unexpected event: %d, status: %d\n",
3120 event->event, event->status);
3123 case RDMA_CM_EVENT_CONNECT_REQUEST:
3124 /* destroy cmid on failure */
3125 rc = kiblnd_passive_connect(cmid,
3126 (void *)KIBLND_CONN_PARAM(event),
3127 KIBLND_CONN_PARAM_LEN(event));
3128 CDEBUG(D_NET, "connreq: %d\n", rc);
3131 case RDMA_CM_EVENT_ADDR_ERROR:
3132 peer_ni = cmid->context;
3133 CNETERR("%s: ADDR ERROR %d\n",
3134 libcfs_nid2str(peer_ni->ibp_nid), event->status);
3135 kiblnd_peer_connect_failed(peer_ni, 1, -EHOSTUNREACH);
3136 kiblnd_peer_decref(peer_ni);
3137 return -EHOSTUNREACH; /* rc != 0 destroys cmid */
3139 case RDMA_CM_EVENT_ADDR_RESOLVED:
3140 peer_ni = cmid->context;
3142 CDEBUG(D_NET,"%s Addr resolved: %d\n",
3143 libcfs_nid2str(peer_ni->ibp_nid), event->status);
3145 if (event->status != 0) {
3146 CNETERR("Can't resolve address for %s: %d\n",
3147 libcfs_nid2str(peer_ni->ibp_nid), event->status);
3150 rc = rdma_resolve_route(
3151 cmid, kiblnd_timeout() * 1000);
3153 struct kib_net *net = peer_ni->ibp_ni->ni_data;
3154 struct kib_dev *dev = net->ibn_dev;
3156 CDEBUG(D_NET, "%s: connection bound to "\
3158 libcfs_nid2str(peer_ni->ibp_nid),
3160 &dev->ibd_ifip, cmid->device->name);
3165 /* Can't initiate route resolution */
3166 CERROR("Can't resolve route for %s: %d\n",
3167 libcfs_nid2str(peer_ni->ibp_nid), rc);
3169 kiblnd_peer_connect_failed(peer_ni, 1, rc);
3170 kiblnd_peer_decref(peer_ni);
3171 return rc; /* rc != 0 destroys cmid */
3173 case RDMA_CM_EVENT_ROUTE_ERROR:
3174 peer_ni = cmid->context;
3175 CNETERR("%s: ROUTE ERROR %d\n",
3176 libcfs_nid2str(peer_ni->ibp_nid), event->status);
3177 kiblnd_peer_connect_failed(peer_ni, 1, -EHOSTUNREACH);
3178 kiblnd_peer_decref(peer_ni);
3179 return -EHOSTUNREACH; /* rc != 0 destroys cmid */
3181 case RDMA_CM_EVENT_ROUTE_RESOLVED:
3182 peer_ni = cmid->context;
3183 CDEBUG(D_NET,"%s Route resolved: %d\n",
3184 libcfs_nid2str(peer_ni->ibp_nid), event->status);
3186 if (event->status == 0)
3187 return kiblnd_active_connect(cmid);
3189 CNETERR("Can't resolve route for %s: %d\n",
3190 libcfs_nid2str(peer_ni->ibp_nid), event->status);
3191 kiblnd_peer_connect_failed(peer_ni, 1, event->status);
3192 kiblnd_peer_decref(peer_ni);
3193 return event->status; /* rc != 0 destroys cmid */
3195 case RDMA_CM_EVENT_UNREACHABLE:
3196 conn = cmid->context;
3197 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
3198 conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
3199 CNETERR("%s: UNREACHABLE %d\n",
3200 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
3201 kiblnd_connreq_done(conn, -ENETDOWN);
3202 kiblnd_conn_decref(conn);
3205 case RDMA_CM_EVENT_CONNECT_ERROR:
3206 conn = cmid->context;
3207 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
3208 conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
3209 CNETERR("%s: CONNECT ERROR %d\n",
3210 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
3211 kiblnd_connreq_done(conn, -ENOTCONN);
3212 kiblnd_conn_decref(conn);
3215 case RDMA_CM_EVENT_REJECTED:
3216 conn = cmid->context;
3217 switch (conn->ibc_state) {
3221 case IBLND_CONN_PASSIVE_WAIT:
3222 CERROR ("%s: REJECTED %d\n",
3223 libcfs_nid2str(conn->ibc_peer->ibp_nid),
3225 kiblnd_connreq_done(conn, -ECONNRESET);
3228 case IBLND_CONN_ACTIVE_CONNECT:
3229 kiblnd_rejected(conn, event->status,
3230 (void *)KIBLND_CONN_PARAM(event),
3231 KIBLND_CONN_PARAM_LEN(event));
3234 kiblnd_conn_decref(conn);
3237 case RDMA_CM_EVENT_ESTABLISHED:
3238 conn = cmid->context;
3239 switch (conn->ibc_state) {
3243 case IBLND_CONN_PASSIVE_WAIT:
3244 CDEBUG(D_NET, "ESTABLISHED (passive): %s\n",
3245 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3246 kiblnd_connreq_done(conn, 0);
3249 case IBLND_CONN_ACTIVE_CONNECT:
3250 CDEBUG(D_NET, "ESTABLISHED(active): %s\n",
3251 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3252 kiblnd_check_connreply(conn,
3253 (void *)KIBLND_CONN_PARAM(event),
3254 KIBLND_CONN_PARAM_LEN(event));
3257 /* net keeps its ref on conn! */
3260 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
3261 CDEBUG(D_NET, "Ignore TIMEWAIT_EXIT event\n");
3264 case RDMA_CM_EVENT_DISCONNECTED:
3265 conn = cmid->context;
3266 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
3267 CERROR("%s DISCONNECTED\n",
3268 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3269 kiblnd_connreq_done(conn, -ECONNRESET);
3271 kiblnd_close_conn(conn, 0);
3273 kiblnd_conn_decref(conn);
3274 cmid->context = NULL;
3277 case RDMA_CM_EVENT_DEVICE_REMOVAL:
3278 LCONSOLE_ERROR_MSG(0x131,
3279 "Received notification of device removal\n"
3280 "Please shutdown LNET to allow this to proceed\n");
3281 /* Can't remove network from underneath LNET for now, so I have
3285 case RDMA_CM_EVENT_ADDR_CHANGE:
3286 LCONSOLE_INFO("Physical link changed (eg hca/port)\n");
3292 kiblnd_check_txs_locked(struct kib_conn *conn, struct list_head *txs)
3296 list_for_each_entry(tx, txs, tx_list) {
3297 if (txs != &conn->ibc_active_txs) {
3298 LASSERT(tx->tx_queued);
3300 LASSERT(!tx->tx_queued);
3301 LASSERT(tx->tx_waiting || tx->tx_sending != 0);
3304 if (ktime_compare(ktime_get(), tx->tx_deadline) >= 0) {
3305 CERROR("Timed out tx: %s(WSQ:%d%d%d), %lld seconds\n",
3306 kiblnd_queue2str(conn, txs),
3307 tx->tx_waiting, tx->tx_sending, tx->tx_queued,
3309 ktime_ms_delta(ktime_get(),
3310 tx->tx_deadline) / MSEC_PER_SEC);
3319 kiblnd_conn_timed_out_locked(struct kib_conn *conn)
3321 return kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue) ||
3322 kiblnd_check_txs_locked(conn, &conn->ibc_tx_noops) ||
3323 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_rsrvd) ||
3324 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_nocred) ||
3325 kiblnd_check_txs_locked(conn, &conn->ibc_active_txs);
3329 kiblnd_check_conns (int idx)
3332 LIST_HEAD(checksends);
3333 LIST_HEAD(timedout_txs);
3334 struct hlist_head *peers = &kiblnd_data.kib_peers[idx];
3335 struct kib_peer_ni *peer_ni;
3336 struct kib_conn *conn;
3337 struct kib_tx *tx, *tx_tmp;
3338 unsigned long flags;
3340 /* NB. We expect to have a look at all the peers and not find any
3341 * RDMAs to time out, so we just use a shared lock while we
3344 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3346 hlist_for_each_entry(peer_ni, peers, ibp_list) {
3347 /* Check tx_deadline */
3348 list_for_each_entry_safe(tx, tx_tmp, &peer_ni->ibp_tx_queue, tx_list) {
3349 if (ktime_compare(ktime_get(), tx->tx_deadline) >= 0) {
3350 CWARN("Timed out tx for %s: %lld seconds\n",
3351 libcfs_nid2str(peer_ni->ibp_nid),
3352 ktime_ms_delta(ktime_get(),
3353 tx->tx_deadline) / MSEC_PER_SEC);
3354 list_move(&tx->tx_list, &timedout_txs);
3358 list_for_each_entry(conn, &peer_ni->ibp_conns, ibc_list) {
3362 LASSERT(conn->ibc_state == IBLND_CONN_ESTABLISHED);
3364 spin_lock(&conn->ibc_lock);
3366 sendnoop = kiblnd_need_noop(conn);
3367 timedout = kiblnd_conn_timed_out_locked(conn);
3368 if (!sendnoop && !timedout) {
3369 spin_unlock(&conn->ibc_lock);
3374 CERROR("Timed out RDMA with %s (%lld): c: %u, oc: %u, rc: %u\n",
3375 libcfs_nid2str(peer_ni->ibp_nid),
3377 - peer_ni->ibp_last_alive,
3379 conn->ibc_outstanding_credits,
3380 conn->ibc_reserved_credits);
3381 list_add(&conn->ibc_connd_list, &closes);
3383 list_add(&conn->ibc_connd_list, &checksends);
3385 /* +ref for 'closes' or 'checksends' */
3386 kiblnd_conn_addref(conn);
3388 spin_unlock(&conn->ibc_lock);
3392 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3394 if (!list_empty(&timedout_txs))
3395 kiblnd_txlist_done(&timedout_txs, -ETIMEDOUT,
3396 LNET_MSG_STATUS_NETWORK_TIMEOUT);
3398 /* Handle timeout by closing the whole
3399 * connection. We can only be sure RDMA activity
3400 * has ceased once the QP has been modified.
3402 while ((conn = list_first_entry_or_null(&closes,
3404 ibc_connd_list)) != NULL) {
3405 list_del(&conn->ibc_connd_list);
3406 kiblnd_close_conn(conn, -ETIMEDOUT);
3407 kiblnd_conn_decref(conn);
3410 /* In case we have enough credits to return via a
3411 * NOOP, but there were no non-blocking tx descs
3412 * free to do it last time...
3414 while ((conn = list_first_entry_or_null(&checksends,
3416 ibc_connd_list)) != NULL) {
3417 list_del(&conn->ibc_connd_list);
3419 spin_lock(&conn->ibc_lock);
3420 kiblnd_check_sends_locked(conn);
3421 spin_unlock(&conn->ibc_lock);
3423 kiblnd_conn_decref(conn);
3428 kiblnd_disconnect_conn(struct kib_conn *conn)
3430 LASSERT (!in_interrupt());
3431 LASSERT (current == kiblnd_data.kib_connd);
3432 LASSERT (conn->ibc_state == IBLND_CONN_CLOSING);
3434 rdma_disconnect(conn->ibc_cmid);
3435 kiblnd_finalise_conn(conn);
3437 kiblnd_peer_notify(conn->ibc_peer);
3441 * High-water for reconnection to the same peer_ni, reconnection attempt should
3442 * be delayed after trying more than KIB_RECONN_HIGH_RACE.
3444 #define KIB_RECONN_HIGH_RACE 10
3446 * Allow connd to take a break and handle other things after consecutive
3447 * reconnection attemps.
3449 #define KIB_RECONN_BREAK 100
3452 kiblnd_connd (void *arg)
3454 spinlock_t *lock = &kiblnd_data.kib_connd_lock;
3455 wait_queue_entry_t wait;
3456 unsigned long flags;
3457 struct kib_conn *conn;
3462 unsigned long deadline = jiffies;
3465 kiblnd_data.kib_connd = current;
3467 spin_lock_irqsave(lock, flags);
3469 while (!kiblnd_data.kib_shutdown) {
3472 dropped_lock = false;
3474 conn = list_first_entry_or_null(&kiblnd_data.kib_connd_zombies,
3475 struct kib_conn, ibc_list);
3477 struct kib_peer_ni *peer_ni = NULL;
3479 list_del(&conn->ibc_list);
3480 if (conn->ibc_reconnect) {
3481 peer_ni = conn->ibc_peer;
3482 kiblnd_peer_addref(peer_ni);
3485 spin_unlock_irqrestore(lock, flags);
3486 dropped_lock = true;
3488 kiblnd_destroy_conn(conn);
3490 spin_lock_irqsave(lock, flags);
3492 LIBCFS_FREE(conn, sizeof(*conn));
3496 conn->ibc_peer = peer_ni;
3497 if (peer_ni->ibp_reconnected < KIB_RECONN_HIGH_RACE)
3498 list_add_tail(&conn->ibc_list,
3499 &kiblnd_data.kib_reconn_list);
3501 list_add_tail(&conn->ibc_list,
3502 &kiblnd_data.kib_reconn_wait);
3505 conn = list_first_entry_or_null(&kiblnd_data.kib_connd_conns,
3506 struct kib_conn, ibc_list);
3510 list_del(&conn->ibc_list);
3512 spin_unlock_irqrestore(lock, flags);
3513 dropped_lock = true;
3515 kiblnd_disconnect_conn(conn);
3516 wait = conn->ibc_waits;
3517 if (wait == 0) /* keep ref for connd_wait, see below */
3518 kiblnd_conn_decref(conn);
3520 spin_lock_irqsave(lock, flags);
3523 list_add_tail(&conn->ibc_list,
3524 &kiblnd_data.kib_connd_waits);
3527 while (reconn < KIB_RECONN_BREAK) {
3528 if (kiblnd_data.kib_reconn_sec !=
3529 ktime_get_real_seconds()) {
3530 kiblnd_data.kib_reconn_sec = ktime_get_real_seconds();
3531 list_splice_init(&kiblnd_data.kib_reconn_wait,
3532 &kiblnd_data.kib_reconn_list);
3535 conn = list_first_entry_or_null(&kiblnd_data.kib_reconn_list,
3536 struct kib_conn, ibc_list);
3540 list_del(&conn->ibc_list);
3542 spin_unlock_irqrestore(lock, flags);
3543 dropped_lock = true;
3545 reconn += kiblnd_reconnect_peer(conn->ibc_peer);
3546 kiblnd_peer_decref(conn->ibc_peer);
3547 LIBCFS_FREE(conn, sizeof(*conn));
3549 spin_lock_irqsave(lock, flags);
3552 conn = list_first_entry_or_null(&kiblnd_data.kib_connd_waits,
3553 struct kib_conn, ibc_list);
3555 list_del(&conn->ibc_list);
3556 spin_unlock_irqrestore(lock, flags);
3558 dropped_lock = kiblnd_tx_may_discard(conn);
3560 kiblnd_conn_decref(conn);
3562 spin_lock_irqsave(lock, flags);
3564 list_add_tail(&conn->ibc_list,
3565 &kiblnd_data.kib_connd_waits);
3568 /* careful with the jiffy wrap... */
3569 timeout = (int)(deadline - jiffies);
3573 int chunk = HASH_SIZE(kiblnd_data.kib_peers);
3574 unsigned int lnd_timeout;
3576 spin_unlock_irqrestore(lock, flags);
3577 dropped_lock = true;
3579 /* Time to check for RDMA timeouts on a few more
3580 * peers: I do checks every 'p' seconds on a
3581 * proportion of the peer_ni table and I need to check
3582 * every connection 'n' times within a timeout
3583 * interval, to ensure I detect a timeout on any
3584 * connection within (n+1)/n times the timeout
3588 lnd_timeout = kiblnd_timeout();
3589 if (lnd_timeout > n * p)
3590 chunk = (chunk * n * p) / lnd_timeout;
3594 for (i = 0; i < chunk; i++) {
3595 kiblnd_check_conns(peer_index);
3596 peer_index = (peer_index + 1) %
3597 HASH_SIZE(kiblnd_data.kib_peers);
3600 deadline += cfs_time_seconds(p);
3601 spin_lock_irqsave(lock, flags);
3607 /* Nothing to do for 'timeout' */
3608 set_current_state(TASK_INTERRUPTIBLE);
3609 add_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3610 spin_unlock_irqrestore(lock, flags);
3612 schedule_timeout(timeout);
3614 remove_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3615 spin_lock_irqsave(lock, flags);
3618 spin_unlock_irqrestore(lock, flags);
3620 kiblnd_thread_fini();
3625 kiblnd_qp_event(struct ib_event *event, void *arg)
3627 struct kib_conn *conn = arg;
3629 switch (event->event) {
3630 case IB_EVENT_COMM_EST:
3631 CDEBUG(D_NET, "%s established\n",
3632 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3633 /* We received a packet but connection isn't established
3634 * probably handshake packet was lost, so free to
3635 * force make connection established */
3636 rdma_notify(conn->ibc_cmid, IB_EVENT_COMM_EST);
3639 case IB_EVENT_PORT_ERR:
3640 case IB_EVENT_DEVICE_FATAL:
3641 CERROR("Fatal device error for NI %s\n",
3642 libcfs_nidstr(&conn->ibc_peer->ibp_ni->ni_nid));
3643 atomic_set(&conn->ibc_peer->ibp_ni->ni_fatal_error_on, 1);
3646 case IB_EVENT_PORT_ACTIVE:
3647 CERROR("Port reactivated for NI %s\n",
3648 libcfs_nidstr(&conn->ibc_peer->ibp_ni->ni_nid));
3649 atomic_set(&conn->ibc_peer->ibp_ni->ni_fatal_error_on, 0);
3653 CERROR("%s: Async QP event type %d\n",
3654 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3660 kiblnd_complete (struct ib_wc *wc)
3662 switch (kiblnd_wreqid2type(wc->wr_id)) {
3667 if (wc->status != IB_WC_SUCCESS &&
3668 wc->status != IB_WC_WR_FLUSH_ERR)
3669 CNETERR("FastReg failed: %d\n", wc->status);
3672 case IBLND_WID_RDMA:
3673 /* We only get RDMA completion notification if it fails. All
3674 * subsequent work items, including the final SEND will fail
3675 * too. However we can't print out any more info about the
3676 * failing RDMA because 'tx' might be back on the idle list or
3677 * even reused already if we didn't manage to post all our work
3679 CNETERR("RDMA (tx: %p) failed: %d\n",
3680 kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3684 kiblnd_tx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3688 kiblnd_rx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status,
3695 kiblnd_cq_completion(struct ib_cq *cq, void *arg)
3697 /* NB I'm not allowed to schedule this conn once its refcount has
3698 * reached 0. Since fundamentally I'm racing with scheduler threads
3699 * consuming my CQ I could be called after all completions have
3700 * occurred. But in this case, ibc_nrx == 0 && ibc_nsends_posted == 0
3701 * and this CQ is about to be destroyed so I NOOP. */
3702 struct kib_conn *conn = arg;
3703 struct kib_sched_info *sched = conn->ibc_sched;
3704 unsigned long flags;
3706 LASSERT(cq == conn->ibc_cq);
3708 spin_lock_irqsave(&sched->ibs_lock, flags);
3710 conn->ibc_ready = 1;
3712 if (!conn->ibc_scheduled &&
3713 (conn->ibc_nrx > 0 ||
3714 conn->ibc_nsends_posted > 0)) {
3715 kiblnd_conn_addref(conn); /* +1 ref for sched_conns */
3716 conn->ibc_scheduled = 1;
3717 list_add_tail(&conn->ibc_sched_list, &sched->ibs_conns);
3719 if (waitqueue_active(&sched->ibs_waitq))
3720 wake_up(&sched->ibs_waitq);
3723 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3727 kiblnd_cq_event(struct ib_event *event, void *arg)
3729 struct kib_conn *conn = arg;
3731 CERROR("%s: async CQ event type %d\n",
3732 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3736 kiblnd_scheduler(void *arg)
3738 long id = (long)arg;
3739 struct kib_sched_info *sched;
3740 struct kib_conn *conn;
3741 wait_queue_entry_t wait;
3742 unsigned long flags;
3749 sched = kiblnd_data.kib_scheds[KIB_THREAD_CPT(id)];
3751 rc = cfs_cpt_bind(lnet_cpt_table(), sched->ibs_cpt);
3753 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);
3756 spin_lock_irqsave(&sched->ibs_lock, flags);
3758 while (!kiblnd_data.kib_shutdown) {
3759 if (need_resched()) {
3760 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3764 spin_lock_irqsave(&sched->ibs_lock, flags);
3767 did_something = false;
3769 conn = list_first_entry_or_null(&sched->ibs_conns,
3773 /* take over kib_sched_conns' ref on conn... */
3774 LASSERT(conn->ibc_scheduled);
3775 list_del(&conn->ibc_sched_list);
3776 conn->ibc_ready = 0;
3778 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3780 wc.wr_id = IBLND_WID_INVAL;
3782 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3784 rc = ib_req_notify_cq(conn->ibc_cq,
3787 CWARN("%s: ib_req_notify_cq failed: %d, closing connection\n",
3788 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
3789 kiblnd_close_conn(conn, -EIO);
3790 kiblnd_conn_decref(conn);
3791 spin_lock_irqsave(&sched->ibs_lock,
3796 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3799 if (unlikely(rc > 0 && wc.wr_id == IBLND_WID_INVAL)) {
3801 "ib_poll_cq (rc: %d) returned invalid "
3802 "wr_id, opcode %d, status: %d, "
3803 "vendor_err: %d, conn: %s status: %d\n"
3804 "please upgrade firmware and OFED or "
3805 "contact vendor.\n", rc,
3806 wc.opcode, wc.status, wc.vendor_err,
3807 libcfs_nid2str(conn->ibc_peer->ibp_nid),
3813 CWARN("%s: ib_poll_cq failed: %d, closing connection\n",
3814 libcfs_nid2str(conn->ibc_peer->ibp_nid),
3816 kiblnd_close_conn(conn, -EIO);
3817 kiblnd_conn_decref(conn);
3818 spin_lock_irqsave(&sched->ibs_lock, flags);
3822 spin_lock_irqsave(&sched->ibs_lock, flags);
3824 if (rc != 0 || conn->ibc_ready) {
3825 /* There may be another completion waiting; get
3826 * another scheduler to check while I handle
3828 /* +1 ref for sched_conns */
3829 kiblnd_conn_addref(conn);
3830 list_add_tail(&conn->ibc_sched_list,
3832 if (waitqueue_active(&sched->ibs_waitq))
3833 wake_up(&sched->ibs_waitq);
3835 conn->ibc_scheduled = 0;
3839 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3840 kiblnd_complete(&wc);
3842 spin_lock_irqsave(&sched->ibs_lock, flags);
3845 kiblnd_conn_decref(conn); /* ..drop my ref from above */
3846 did_something = true;
3852 set_current_state(TASK_INTERRUPTIBLE);
3853 add_wait_queue_exclusive(&sched->ibs_waitq, &wait);
3854 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3858 remove_wait_queue(&sched->ibs_waitq, &wait);
3859 set_current_state(TASK_RUNNING);
3860 spin_lock_irqsave(&sched->ibs_lock, flags);
3863 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3865 kiblnd_thread_fini();
3870 kiblnd_failover_thread(void *arg)
3872 rwlock_t *glock = &kiblnd_data.kib_global_lock;
3873 struct kib_dev *dev;
3874 struct net *ns = arg;
3875 wait_queue_entry_t wait;
3876 unsigned long flags;
3879 LASSERT(*kiblnd_tunables.kib_dev_failover != 0);
3882 write_lock_irqsave(glock, flags);
3884 while (!kiblnd_data.kib_shutdown) {
3885 bool do_failover = false;
3888 list_for_each_entry(dev, &kiblnd_data.kib_failed_devs,
3890 if (ktime_get_seconds() < dev->ibd_next_failover)
3897 list_del_init(&dev->ibd_fail_list);
3898 dev->ibd_failover = 1;
3899 write_unlock_irqrestore(glock, flags);
3901 rc = kiblnd_dev_failover(dev, ns);
3903 write_lock_irqsave(glock, flags);
3905 LASSERT(dev->ibd_failover);
3906 dev->ibd_failover = 0;
3907 if (rc >= 0) { /* Device is OK or failover succeed */
3908 dev->ibd_next_failover = ktime_get_seconds() + 3;
3912 /* failed to failover, retry later */
3913 dev->ibd_next_failover = ktime_get_seconds() +
3914 min(dev->ibd_failed_failover, 10);
3915 if (kiblnd_dev_can_failover(dev)) {
3916 list_add_tail(&dev->ibd_fail_list,
3917 &kiblnd_data.kib_failed_devs);
3923 /* long sleep if no more pending failover */
3924 long_sleep = list_empty(&kiblnd_data.kib_failed_devs);
3926 set_current_state(TASK_INTERRUPTIBLE);
3927 add_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3928 write_unlock_irqrestore(glock, flags);
3930 rc = schedule_timeout(long_sleep ? cfs_time_seconds(10) :
3931 cfs_time_seconds(1));
3932 set_current_state(TASK_RUNNING);
3933 remove_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3934 write_lock_irqsave(glock, flags);
3936 if (!long_sleep || rc != 0)
3939 /* have a long sleep, routine check all active devices,
3940 * we need checking like this because if there is not active
3941 * connection on the dev and no SEND from local, we may listen
3942 * on wrong HCA for ever while there is a bonding failover
3944 list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3945 if (kiblnd_dev_can_failover(dev)) {
3946 list_add_tail(&dev->ibd_fail_list,
3947 &kiblnd_data.kib_failed_devs);
3952 write_unlock_irqrestore(glock, flags);
3954 kiblnd_thread_fini();