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;
495 LASSERT (net != NULL);
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 != 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->ibh_ibdev,
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->ibh_ibdev, 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;
755 CDEBUG(D_NET, "niov %d offset %d nob %d\n", nkiov, offset, nob);
759 LASSERT(net != NULL);
761 while (offset >= kiov->bv_len) {
762 offset -= kiov->bv_len;
774 fragnob = min((int)(kiov->bv_len - offset), nob);
777 * We're allowed to start at a non-aligned page offset in
778 * the first fragment and end at a non-aligned page offset
779 * in the last fragment.
781 if ((fragnob < (int)(kiov->bv_len - offset)) &&
782 nkiov < max_nkiov && nob > fragnob) {
783 CDEBUG(D_NET, "fragnob %d < available page %d: with"
784 " remaining %d kiovs with %d nob left\n",
785 fragnob, (int)(kiov->bv_len - offset),
790 sg_set_page(sg, kiov->bv_page, fragnob,
791 kiov->bv_offset + offset);
794 CERROR("lacking enough sg entries to map tx\n");
804 return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags);
808 kiblnd_post_tx_locked(struct kib_conn *conn, struct kib_tx *tx, int credit)
809 __must_hold(&conn->ibc_lock)
811 struct kib_msg *msg = tx->tx_msg;
812 struct kib_peer_ni *peer_ni = conn->ibc_peer;
813 struct lnet_ni *ni = peer_ni->ibp_ni;
814 struct kib_fast_reg_descriptor *frd = tx->tx_fmr.fmr_frd;
815 int ver = conn->ibc_version;
819 LASSERT(tx->tx_queued);
820 /* We rely on this for QP sizing */
821 LASSERT(tx->tx_nwrq > 0 && tx->tx_nsge >= 0);
822 LASSERT(tx->tx_nwrq <= 1 + conn->ibc_max_frags);
824 LASSERT(credit == 0 || credit == 1);
825 LASSERT(conn->ibc_outstanding_credits >= 0);
826 LASSERT(conn->ibc_outstanding_credits <= conn->ibc_queue_depth);
827 LASSERT(conn->ibc_credits >= 0);
828 LASSERT(conn->ibc_credits <= conn->ibc_queue_depth);
830 if (conn->ibc_nsends_posted ==
831 kiblnd_concurrent_sends(ver, ni)) {
832 /* tx completions outstanding... */
833 CDEBUG(D_NET, "%s: posted enough\n",
834 libcfs_nid2str(peer_ni->ibp_nid));
838 if (credit != 0 && conn->ibc_credits == 0) { /* no credits */
839 CDEBUG(D_NET, "%s: no credits\n",
840 libcfs_nid2str(peer_ni->ibp_nid));
844 if (credit != 0 && !IBLND_OOB_CAPABLE(ver) &&
845 conn->ibc_credits == 1 && /* last credit reserved */
846 msg->ibm_type != IBLND_MSG_NOOP) { /* for NOOP */
847 CDEBUG(D_NET, "%s: not using last credit\n",
848 libcfs_nid2str(peer_ni->ibp_nid));
852 /* NB don't drop ibc_lock before bumping tx_sending */
853 list_del(&tx->tx_list);
856 if (msg->ibm_type == IBLND_MSG_NOOP &&
857 (!kiblnd_need_noop(conn) || /* redundant NOOP */
858 (IBLND_OOB_CAPABLE(ver) && /* posted enough NOOP */
859 conn->ibc_noops_posted == IBLND_OOB_MSGS(ver)))) {
860 /* OK to drop when posted enough NOOPs, since
861 * kiblnd_check_sends_locked will queue NOOP again when
862 * posted NOOPs complete */
863 spin_unlock(&conn->ibc_lock);
864 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
866 spin_lock(&conn->ibc_lock);
867 CDEBUG(D_NET, "%s(%d): redundant or enough NOOP\n",
868 libcfs_nid2str(peer_ni->ibp_nid),
869 conn->ibc_noops_posted);
873 kiblnd_pack_msg(peer_ni->ibp_ni, msg, ver, conn->ibc_outstanding_credits,
874 peer_ni->ibp_nid, conn->ibc_incarnation);
876 conn->ibc_credits -= credit;
877 conn->ibc_outstanding_credits = 0;
878 conn->ibc_nsends_posted++;
879 if (msg->ibm_type == IBLND_MSG_NOOP)
880 conn->ibc_noops_posted++;
882 /* CAVEAT EMPTOR! This tx could be the PUT_DONE of an RDMA
883 * PUT. If so, it was first queued here as a PUT_REQ, sent and
884 * stashed on ibc_active_txs, matched by an incoming PUT_ACK,
885 * and then re-queued here. It's (just) possible that
886 * tx_sending is non-zero if we've not done the tx_complete()
887 * from the first send; hence the ++ rather than = below. */
889 list_add(&tx->tx_list, &conn->ibc_active_txs);
891 /* I'm still holding ibc_lock! */
892 if (conn->ibc_state != IBLND_CONN_ESTABLISHED) {
894 } else if (tx->tx_pool->tpo_pool.po_failed ||
895 conn->ibc_hdev != tx->tx_pool->tpo_hdev) {
896 /* close_conn will launch failover */
899 struct ib_send_wr *bad = &tx->tx_wrq[tx->tx_nwrq - 1].wr;
900 struct ib_send_wr *wr = &tx->tx_wrq[0].wr;
902 if (frd != NULL && !frd->frd_posted) {
903 if (!frd->frd_valid) {
904 wr = &frd->frd_inv_wr.wr;
905 wr->next = &frd->frd_fastreg_wr.wr;
907 wr = &frd->frd_fastreg_wr.wr;
909 frd->frd_fastreg_wr.wr.next = &tx->tx_wrq[0].wr;
912 LASSERTF(bad->wr_id == kiblnd_ptr2wreqid(tx, IBLND_WID_TX),
913 "bad wr_id %#llx, opc %d, flags %d, peer_ni: %s\n",
914 bad->wr_id, bad->opcode, bad->send_flags,
915 libcfs_nid2str(conn->ibc_peer->ibp_nid));
918 if (lnet_send_error_simulation(tx->tx_lntmsg[0], &tx->tx_hstatus))
921 #ifdef HAVE_IB_POST_SEND_RECV_CONST
922 rc = ib_post_send(conn->ibc_cmid->qp, wr,
923 (const struct ib_send_wr **)&bad);
925 rc = ib_post_send(conn->ibc_cmid->qp, wr, &bad);
929 conn->ibc_last_send = ktime_get();
933 frd->frd_posted = true;
937 /* NB credits are transferred in the actual
938 * message, which can only be the last work item */
939 conn->ibc_credits += credit;
940 conn->ibc_outstanding_credits += msg->ibm_credits;
941 conn->ibc_nsends_posted--;
942 if (msg->ibm_type == IBLND_MSG_NOOP)
943 conn->ibc_noops_posted--;
949 done = (tx->tx_sending == 0);
951 list_del(&tx->tx_list);
953 spin_unlock(&conn->ibc_lock);
955 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
956 CERROR("Error %d posting transmit to %s\n",
957 rc, libcfs_nid2str(peer_ni->ibp_nid));
959 CDEBUG(D_NET, "Error %d posting transmit to %s\n",
960 rc, libcfs_nid2str(peer_ni->ibp_nid));
962 kiblnd_close_conn(conn, rc);
967 spin_lock(&conn->ibc_lock);
973 kiblnd_check_sends_locked(struct kib_conn *conn)
975 int ver = conn->ibc_version;
976 struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
979 /* Don't send anything until after the connection is established */
980 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
981 CDEBUG(D_NET, "%s too soon\n",
982 libcfs_nid2str(conn->ibc_peer->ibp_nid));
986 LASSERT(conn->ibc_nsends_posted <=
987 kiblnd_concurrent_sends(ver, ni));
988 LASSERT (!IBLND_OOB_CAPABLE(ver) ||
989 conn->ibc_noops_posted <= IBLND_OOB_MSGS(ver));
990 LASSERT (conn->ibc_reserved_credits >= 0);
992 while (conn->ibc_reserved_credits > 0 &&
993 (tx = list_first_entry_or_null(&conn->ibc_tx_queue_rsrvd,
994 struct kib_tx, tx_list)) != NULL) {
995 list_move_tail(&tx->tx_list, &conn->ibc_tx_queue);
996 conn->ibc_reserved_credits--;
999 if (kiblnd_need_noop(conn)) {
1000 spin_unlock(&conn->ibc_lock);
1002 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
1004 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_NOOP, 0);
1006 spin_lock(&conn->ibc_lock);
1008 kiblnd_queue_tx_locked(tx, conn);
1014 if (!list_empty(&conn->ibc_tx_queue_nocred)) {
1016 tx = list_first_entry(&conn->ibc_tx_queue_nocred,
1017 struct kib_tx, tx_list);
1018 } else if (!list_empty(&conn->ibc_tx_noops)) {
1019 LASSERT (!IBLND_OOB_CAPABLE(ver));
1021 tx = list_first_entry(&conn->ibc_tx_noops,
1022 struct kib_tx, tx_list);
1023 } else if (!list_empty(&conn->ibc_tx_queue)) {
1025 tx = list_first_entry(&conn->ibc_tx_queue,
1026 struct kib_tx, tx_list);
1030 if (kiblnd_post_tx_locked(conn, tx, credit) != 0)
1036 kiblnd_tx_complete(struct kib_tx *tx, int status)
1038 int failed = (status != IB_WC_SUCCESS);
1039 struct kib_conn *conn = tx->tx_conn;
1042 if (tx->tx_sending <= 0) {
1043 CERROR("Received an event on a freed tx: %p status %d\n",
1049 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
1050 CNETERR("Tx -> %s cookie %#llx"
1051 " sending %d waiting %d: failed %d\n",
1052 libcfs_nid2str(conn->ibc_peer->ibp_nid),
1053 tx->tx_cookie, tx->tx_sending, tx->tx_waiting,
1056 kiblnd_close_conn(conn, -EIO);
1058 kiblnd_peer_alive(conn->ibc_peer);
1061 spin_lock(&conn->ibc_lock);
1063 /* I could be racing with rdma completion. Whoever makes 'tx' idle
1064 * gets to free it, which also drops its ref on 'conn'. */
1067 conn->ibc_nsends_posted--;
1068 if (tx->tx_msg->ibm_type == IBLND_MSG_NOOP)
1069 conn->ibc_noops_posted--;
1072 tx->tx_hstatus = LNET_MSG_STATUS_REMOTE_DROPPED;
1073 tx->tx_waiting = 0; /* don't wait for peer_ni */
1074 tx->tx_status = -EIO;
1077 idle = (tx->tx_sending == 0) && /* This is the final callback */
1078 !tx->tx_waiting && /* Not waiting for peer_ni */
1079 !tx->tx_queued; /* Not re-queued (PUT_DONE) */
1081 list_del(&tx->tx_list);
1083 kiblnd_check_sends_locked(conn);
1084 spin_unlock(&conn->ibc_lock);
1091 kiblnd_init_tx_msg(struct lnet_ni *ni, struct kib_tx *tx, int type,
1094 struct kib_hca_dev *hdev = tx->tx_pool->tpo_hdev;
1095 struct ib_sge *sge = &tx->tx_msgsge;
1096 struct ib_rdma_wr *wrq;
1097 int nob = offsetof(struct kib_msg, ibm_u) + body_nob;
1098 #ifdef HAVE_IB_GET_DMA_MR
1099 struct ib_mr *mr = hdev->ibh_mrs;
1102 LASSERT(tx->tx_nwrq >= 0);
1103 LASSERT(tx->tx_nwrq < IBLND_MAX_RDMA_FRAGS + 1);
1104 LASSERT(nob <= IBLND_MSG_SIZE);
1105 #ifdef HAVE_IB_GET_DMA_MR
1106 LASSERT(mr != NULL);
1109 kiblnd_init_msg(tx->tx_msg, type, body_nob);
1111 #ifdef HAVE_IB_GET_DMA_MR
1112 sge->lkey = mr->lkey;
1114 sge->lkey = hdev->ibh_pd->local_dma_lkey;
1116 sge->addr = tx->tx_msgaddr;
1119 wrq = &tx->tx_wrq[tx->tx_nwrq];
1120 memset(wrq, 0, sizeof(*wrq));
1122 wrq->wr.next = NULL;
1123 wrq->wr.wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_TX);
1124 wrq->wr.sg_list = sge;
1125 wrq->wr.num_sge = 1;
1126 wrq->wr.opcode = IB_WR_SEND;
1127 wrq->wr.send_flags = IB_SEND_SIGNALED;
1133 kiblnd_init_rdma(struct kib_conn *conn, struct kib_tx *tx, int type,
1134 int resid, struct kib_rdma_desc *dstrd, u64 dstcookie)
1136 struct kib_msg *ibmsg = tx->tx_msg;
1137 struct kib_rdma_desc *srcrd = tx->tx_rd;
1138 struct ib_rdma_wr *wrq = NULL;
1146 LASSERT(!in_interrupt());
1147 LASSERT(tx->tx_nwrq == 0 && tx->tx_nsge == 0);
1148 LASSERT(type == IBLND_MSG_GET_DONE || type == IBLND_MSG_PUT_DONE);
1150 for (srcidx = dstidx = wrq_sge = sge_nob = 0;
1151 resid > 0; resid -= sge_nob) {
1154 if (srcidx >= srcrd->rd_nfrags) {
1155 CERROR("Src buffer exhausted: %d frags\n", srcidx);
1160 if (dstidx >= dstrd->rd_nfrags) {
1161 CERROR("Dst buffer exhausted: %d frags\n", dstidx);
1166 if (tx->tx_nwrq >= conn->ibc_max_frags) {
1167 CERROR("RDMA has too many fragments for peer_ni %s (%d), "
1168 "src idx/frags: %d/%d dst idx/frags: %d/%d\n",
1169 libcfs_nid2str(conn->ibc_peer->ibp_nid),
1170 conn->ibc_max_frags,
1171 srcidx, srcrd->rd_nfrags,
1172 dstidx, dstrd->rd_nfrags);
1177 sge_nob = min3(kiblnd_rd_frag_size(srcrd, srcidx),
1178 kiblnd_rd_frag_size(dstrd, dstidx),
1181 sge = &tx->tx_sge[tx->tx_nsge];
1182 sge->addr = kiblnd_rd_frag_addr(srcrd, srcidx);
1183 sge->lkey = kiblnd_rd_frag_key(srcrd, srcidx);
1184 sge->length = sge_nob;
1187 wrq = &tx->tx_wrq[tx->tx_nwrq];
1189 wrq->wr.next = &(wrq + 1)->wr;
1190 wrq->wr.wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_RDMA);
1191 wrq->wr.sg_list = sge;
1192 wrq->wr.opcode = IB_WR_RDMA_WRITE;
1193 wrq->wr.send_flags = 0;
1195 #ifdef HAVE_IB_RDMA_WR
1196 wrq->remote_addr = kiblnd_rd_frag_addr(dstrd,
1198 wrq->rkey = kiblnd_rd_frag_key(dstrd,
1201 wrq->wr.wr.rdma.remote_addr = kiblnd_rd_frag_addr(dstrd,
1203 wrq->wr.wr.rdma.rkey = kiblnd_rd_frag_key(dstrd,
1208 srcidx = kiblnd_rd_consume_frag(srcrd, srcidx, sge_nob);
1209 dstidx = kiblnd_rd_consume_frag(dstrd, dstidx, sge_nob);
1212 if (wrq_sge == *kiblnd_tunables.kib_wrq_sge || dstidx != prev) {
1214 wrq->wr.num_sge = wrq_sge;
1220 if (rc < 0) /* no RDMA if completing with failure */
1221 tx->tx_nwrq = tx->tx_nsge = 0;
1223 ibmsg->ibm_u.completion.ibcm_status = rc;
1224 ibmsg->ibm_u.completion.ibcm_cookie = dstcookie;
1225 kiblnd_init_tx_msg(conn->ibc_peer->ibp_ni, tx,
1226 type, sizeof(struct kib_completion_msg));
1232 kiblnd_queue_tx_locked(struct kib_tx *tx, struct kib_conn *conn)
1234 struct list_head *q;
1237 LASSERT(tx->tx_nwrq > 0); /* work items set up */
1238 LASSERT(!tx->tx_queued); /* not queued for sending already */
1239 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1241 if (conn->ibc_state >= IBLND_CONN_DISCONNECTED) {
1242 tx->tx_status = -ECONNABORTED;
1244 if (tx->tx_conn != NULL) {
1245 /* PUT_DONE first attached to conn as a PUT_REQ */
1246 LASSERT(tx->tx_conn == conn);
1247 LASSERT(tx->tx_msg->ibm_type == IBLND_MSG_PUT_DONE);
1249 kiblnd_conn_decref(conn);
1251 list_add(&tx->tx_list, &conn->ibc_zombie_txs);
1256 timeout_ns = kiblnd_timeout() * NSEC_PER_SEC;
1258 tx->tx_deadline = ktime_add_ns(ktime_get(), timeout_ns);
1260 if (tx->tx_conn == NULL) {
1261 kiblnd_conn_addref(conn);
1263 LASSERT (tx->tx_msg->ibm_type != IBLND_MSG_PUT_DONE);
1265 /* PUT_DONE first attached to conn as a PUT_REQ */
1266 LASSERT (tx->tx_conn == conn);
1267 LASSERT (tx->tx_msg->ibm_type == IBLND_MSG_PUT_DONE);
1270 switch (tx->tx_msg->ibm_type) {
1274 case IBLND_MSG_PUT_REQ:
1275 case IBLND_MSG_GET_REQ:
1276 q = &conn->ibc_tx_queue_rsrvd;
1279 case IBLND_MSG_PUT_NAK:
1280 case IBLND_MSG_PUT_ACK:
1281 case IBLND_MSG_PUT_DONE:
1282 case IBLND_MSG_GET_DONE:
1283 q = &conn->ibc_tx_queue_nocred;
1286 case IBLND_MSG_NOOP:
1287 if (IBLND_OOB_CAPABLE(conn->ibc_version))
1288 q = &conn->ibc_tx_queue_nocred;
1290 q = &conn->ibc_tx_noops;
1293 case IBLND_MSG_IMMEDIATE:
1294 q = &conn->ibc_tx_queue;
1298 list_add_tail(&tx->tx_list, q);
1302 kiblnd_queue_tx(struct kib_tx *tx, struct kib_conn *conn)
1304 spin_lock(&conn->ibc_lock);
1305 kiblnd_queue_tx_locked(tx, conn);
1306 kiblnd_check_sends_locked(conn);
1307 spin_unlock(&conn->ibc_lock);
1311 kiblnd_resolve_addr_cap(struct rdma_cm_id *cmid,
1312 struct sockaddr_in *srcaddr,
1313 struct sockaddr_in *dstaddr,
1316 unsigned short port;
1319 /* allow the port to be reused */
1320 rc = rdma_set_reuseaddr(cmid, 1);
1322 CERROR("Unable to set reuse on cmid: %d\n", rc);
1326 /* look for a free privileged port */
1327 for (port = PROT_SOCK-1; port > 0; port--) {
1328 srcaddr->sin_port = htons(port);
1329 rc = rdma_resolve_addr(cmid,
1330 (struct sockaddr *)srcaddr,
1331 (struct sockaddr *)dstaddr,
1334 CDEBUG(D_NET, "bound to port %hu\n", port);
1336 } else if (rc == -EADDRINUSE || rc == -EADDRNOTAVAIL) {
1337 CDEBUG(D_NET, "bind to port %hu failed: %d\n",
1344 CERROR("cannot bind to a free privileged port: rc = %d\n", rc);
1350 kiblnd_resolve_addr(struct rdma_cm_id *cmid,
1351 struct sockaddr_in *srcaddr,
1352 struct sockaddr_in *dstaddr,
1355 const struct cred *old_creds = NULL;
1356 struct cred *new_creds;
1359 if (!capable(CAP_NET_BIND_SERVICE)) {
1360 new_creds = prepare_kernel_cred(NULL);
1364 cap_raise(new_creds->cap_effective, CAP_NET_BIND_SERVICE);
1365 old_creds = override_creds(new_creds);
1368 rc = kiblnd_resolve_addr_cap(cmid, srcaddr, dstaddr, timeout_ms);
1371 revert_creds(old_creds);
1377 kiblnd_connect_peer(struct kib_peer_ni *peer_ni)
1379 struct rdma_cm_id *cmid;
1380 struct kib_dev *dev;
1381 struct kib_net *net = peer_ni->ibp_ni->ni_data;
1382 struct sockaddr_in srcaddr;
1383 struct sockaddr_in dstaddr;
1386 LASSERT (net != NULL);
1387 LASSERT (peer_ni->ibp_connecting > 0);
1389 cmid = kiblnd_rdma_create_id(peer_ni->ibp_ni->ni_net_ns,
1390 kiblnd_cm_callback, peer_ni,
1391 RDMA_PS_TCP, IB_QPT_RC);
1394 CERROR("Can't create CMID for %s: %ld\n",
1395 libcfs_nid2str(peer_ni->ibp_nid), PTR_ERR(cmid));
1401 memset(&srcaddr, 0, sizeof(srcaddr));
1402 srcaddr.sin_family = AF_INET;
1403 srcaddr.sin_addr.s_addr = htonl(dev->ibd_ifip);
1405 memset(&dstaddr, 0, sizeof(dstaddr));
1406 dstaddr.sin_family = AF_INET;
1407 dstaddr.sin_port = htons(*kiblnd_tunables.kib_service);
1408 dstaddr.sin_addr.s_addr = htonl(LNET_NIDADDR(peer_ni->ibp_nid));
1410 kiblnd_peer_addref(peer_ni); /* cmid's ref */
1412 if (*kiblnd_tunables.kib_use_priv_port) {
1413 rc = kiblnd_resolve_addr(cmid, &srcaddr, &dstaddr,
1414 kiblnd_timeout() * 1000);
1416 rc = rdma_resolve_addr(cmid,
1417 (struct sockaddr *)&srcaddr,
1418 (struct sockaddr *)&dstaddr,
1419 kiblnd_timeout() * 1000);
1422 /* Can't initiate address resolution: */
1423 CERROR("Can't resolve addr for %s: %d\n",
1424 libcfs_nid2str(peer_ni->ibp_nid), rc);
1431 kiblnd_peer_connect_failed(peer_ni, 1, rc);
1432 kiblnd_peer_decref(peer_ni); /* cmid's ref */
1433 rdma_destroy_id(cmid);
1436 kiblnd_peer_connect_failed(peer_ni, 1, rc);
1440 kiblnd_reconnect_peer(struct kib_peer_ni *peer_ni)
1442 rwlock_t *glock = &kiblnd_data.kib_global_lock;
1443 char *reason = NULL;
1445 unsigned long flags;
1447 write_lock_irqsave(glock, flags);
1448 if (peer_ni->ibp_reconnecting == 0) {
1449 if (peer_ni->ibp_accepting)
1450 reason = "accepting";
1451 else if (peer_ni->ibp_connecting)
1452 reason = "connecting";
1453 else if (!list_empty(&peer_ni->ibp_conns))
1454 reason = "connected";
1455 else /* connected then closed */
1461 if (peer_ni->ibp_accepting)
1462 CNETERR("Detecting race between accepting and reconnecting\n");
1463 peer_ni->ibp_reconnecting--;
1465 if (!kiblnd_peer_active(peer_ni)) {
1466 list_splice_init(&peer_ni->ibp_tx_queue, &txs);
1467 reason = "unlinked";
1471 peer_ni->ibp_connecting++;
1472 peer_ni->ibp_reconnected++;
1474 write_unlock_irqrestore(glock, flags);
1476 kiblnd_connect_peer(peer_ni);
1480 write_unlock_irqrestore(glock, flags);
1482 CWARN("Abort reconnection of %s: %s\n",
1483 libcfs_nid2str(peer_ni->ibp_nid), reason);
1484 kiblnd_txlist_done(&txs, -ECONNABORTED,
1485 LNET_MSG_STATUS_LOCAL_ABORTED);
1490 kiblnd_launch_tx(struct lnet_ni *ni, struct kib_tx *tx, lnet_nid_t nid)
1492 struct kib_peer_ni *peer_ni;
1493 struct kib_peer_ni *peer2;
1494 struct kib_conn *conn;
1495 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
1496 unsigned long flags;
1499 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
1501 /* If I get here, I've committed to send, so I complete the tx with
1502 * failure on any problems
1505 LASSERT(!tx || !tx->tx_conn); /* only set when assigned a conn */
1506 LASSERT(!tx || tx->tx_nwrq > 0); /* work items have been set up */
1508 /* First time, just use a read lock since I expect to find my peer_ni
1511 read_lock_irqsave(g_lock, flags);
1513 peer_ni = kiblnd_find_peer_locked(ni, nid);
1514 if (peer_ni != NULL && !list_empty(&peer_ni->ibp_conns)) {
1515 /* Found a peer_ni with an established connection */
1516 conn = kiblnd_get_conn_locked(peer_ni);
1517 kiblnd_conn_addref(conn); /* 1 ref for me... */
1519 read_unlock_irqrestore(g_lock, flags);
1522 kiblnd_queue_tx(tx, conn);
1523 kiblnd_conn_decref(conn); /* ...to here */
1527 read_unlock(g_lock);
1528 /* Re-try with a write lock */
1531 peer_ni = kiblnd_find_peer_locked(ni, nid);
1532 if (peer_ni != NULL) {
1533 if (list_empty(&peer_ni->ibp_conns)) {
1534 /* found a peer_ni, but it's still connecting... */
1535 LASSERT(kiblnd_peer_connecting(peer_ni));
1537 list_add_tail(&tx->tx_list,
1538 &peer_ni->ibp_tx_queue);
1539 write_unlock_irqrestore(g_lock, flags);
1541 conn = kiblnd_get_conn_locked(peer_ni);
1542 kiblnd_conn_addref(conn); /* 1 ref for me... */
1544 write_unlock_irqrestore(g_lock, flags);
1547 kiblnd_queue_tx(tx, conn);
1548 kiblnd_conn_decref(conn); /* ...to here */
1553 write_unlock_irqrestore(g_lock, flags);
1555 /* Allocate a peer_ni ready to add to the peer_ni table and retry */
1556 rc = kiblnd_create_peer(ni, &peer_ni, nid);
1558 CERROR("Can't create peer_ni %s\n", libcfs_nid2str(nid));
1560 tx->tx_status = -EHOSTUNREACH;
1562 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
1568 write_lock_irqsave(g_lock, flags);
1570 peer2 = kiblnd_find_peer_locked(ni, nid);
1571 if (peer2 != NULL) {
1572 if (list_empty(&peer2->ibp_conns)) {
1573 /* found a peer_ni, but it's still connecting... */
1574 LASSERT(kiblnd_peer_connecting(peer2));
1576 list_add_tail(&tx->tx_list,
1577 &peer2->ibp_tx_queue);
1578 write_unlock_irqrestore(g_lock, flags);
1580 conn = kiblnd_get_conn_locked(peer2);
1581 kiblnd_conn_addref(conn); /* 1 ref for me... */
1583 write_unlock_irqrestore(g_lock, flags);
1586 kiblnd_queue_tx(tx, conn);
1587 kiblnd_conn_decref(conn); /* ...to here */
1590 kiblnd_peer_decref(peer_ni);
1594 /* Brand new peer_ni */
1595 LASSERT(peer_ni->ibp_connecting == 0);
1596 tunables = &peer_ni->ibp_ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
1597 peer_ni->ibp_connecting = tunables->lnd_conns_per_peer;
1599 /* always called with a ref on ni, which prevents ni being shutdown */
1600 LASSERT(((struct kib_net *)ni->ni_data)->ibn_shutdown == 0);
1603 list_add_tail(&tx->tx_list, &peer_ni->ibp_tx_queue);
1605 kiblnd_peer_addref(peer_ni);
1606 hash_add(kiblnd_data.kib_peers, &peer_ni->ibp_list, nid);
1608 write_unlock_irqrestore(g_lock, flags);
1610 for (i = 0; i < tunables->lnd_conns_per_peer; i++)
1611 kiblnd_connect_peer(peer_ni);
1612 kiblnd_peer_decref(peer_ni);
1616 kiblnd_send(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg)
1618 struct lnet_hdr *hdr = &lntmsg->msg_hdr;
1619 int type = lntmsg->msg_type;
1620 struct lnet_process_id target = lntmsg->msg_target;
1621 int target_is_router = lntmsg->msg_target_is_router;
1622 int routing = lntmsg->msg_routing;
1623 unsigned int payload_niov = lntmsg->msg_niov;
1624 struct bio_vec *payload_kiov = lntmsg->msg_kiov;
1625 unsigned int payload_offset = lntmsg->msg_offset;
1626 unsigned int payload_nob = lntmsg->msg_len;
1627 struct kib_msg *ibmsg;
1628 struct kib_rdma_desc *rd;
1633 /* NB 'private' is different depending on what we're sending.... */
1635 CDEBUG(D_NET, "sending %d bytes in %d frags to %s\n",
1636 payload_nob, payload_niov, libcfs_id2str(target));
1638 LASSERT (payload_nob == 0 || payload_niov > 0);
1639 LASSERT (payload_niov <= LNET_MAX_IOV);
1641 /* Thread context */
1642 LASSERT (!in_interrupt());
1650 LASSERT (payload_nob == 0);
1654 if (routing || target_is_router)
1655 break; /* send IMMEDIATE */
1657 /* is the REPLY message too small for RDMA? */
1658 nob = offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[lntmsg->msg_md->md_length]);
1659 if (nob <= IBLND_MSG_SIZE && !lntmsg->msg_rdma_force)
1660 break; /* send IMMEDIATE */
1662 tx = kiblnd_get_idle_tx(ni, target.nid);
1664 CERROR("Can't allocate txd for GET to %s\n",
1665 libcfs_nid2str(target.nid));
1670 rd = &ibmsg->ibm_u.get.ibgm_rd;
1671 rc = kiblnd_setup_rd_kiov(ni, tx, rd,
1672 lntmsg->msg_md->md_niov,
1673 lntmsg->msg_md->md_kiov,
1674 0, lntmsg->msg_md->md_length);
1676 CERROR("Can't setup GET sink for %s: %d\n",
1677 libcfs_nid2str(target.nid), rc);
1678 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
1683 nob = offsetof(struct kib_get_msg, ibgm_rd.rd_frags[rd->rd_nfrags]);
1684 ibmsg->ibm_u.get.ibgm_cookie = tx->tx_cookie;
1685 ibmsg->ibm_u.get.ibgm_hdr = *hdr;
1687 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_GET_REQ, nob);
1689 tx->tx_lntmsg[1] = lnet_create_reply_msg(ni, lntmsg);
1690 if (tx->tx_lntmsg[1] == NULL) {
1691 CERROR("Can't create reply for GET -> %s\n",
1692 libcfs_nid2str(target.nid));
1697 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg[0,1] on completion */
1698 tx->tx_waiting = 1; /* waiting for GET_DONE */
1699 kiblnd_launch_tx(ni, tx, target.nid);
1702 case LNET_MSG_REPLY:
1704 /* Is the payload small enough not to need RDMA? */
1705 nob = offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[payload_nob]);
1706 if (nob <= IBLND_MSG_SIZE && !lntmsg->msg_rdma_force)
1707 break; /* send IMMEDIATE */
1709 tx = kiblnd_get_idle_tx(ni, target.nid);
1711 CERROR("Can't allocate %s txd for %s\n",
1712 type == LNET_MSG_PUT ? "PUT" : "REPLY",
1713 libcfs_nid2str(target.nid));
1717 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1718 payload_niov, payload_kiov,
1719 payload_offset, payload_nob);
1721 CERROR("Can't setup PUT src for %s: %d\n",
1722 libcfs_nid2str(target.nid), rc);
1728 ibmsg->ibm_u.putreq.ibprm_hdr = *hdr;
1729 ibmsg->ibm_u.putreq.ibprm_cookie = tx->tx_cookie;
1730 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_REQ,
1731 sizeof(struct kib_putreq_msg));
1733 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1734 tx->tx_waiting = 1; /* waiting for PUT_{ACK,NAK} */
1735 kiblnd_launch_tx(ni, tx, target.nid);
1739 /* send IMMEDIATE */
1740 LASSERT(offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[payload_nob])
1743 tx = kiblnd_get_idle_tx(ni, target.nid);
1745 CERROR ("Can't send %d to %s: tx descs exhausted\n",
1746 type, libcfs_nid2str(target.nid));
1751 ibmsg->ibm_u.immediate.ibim_hdr = *hdr;
1753 lnet_copy_kiov2flat(IBLND_MSG_SIZE, ibmsg,
1754 offsetof(struct kib_msg,
1755 ibm_u.immediate.ibim_payload),
1756 payload_niov, payload_kiov,
1757 payload_offset, payload_nob);
1759 nob = offsetof(struct kib_immediate_msg, ibim_payload[payload_nob]);
1760 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_IMMEDIATE, nob);
1762 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1763 kiblnd_launch_tx(ni, tx, target.nid);
1768 kiblnd_reply(struct lnet_ni *ni, struct kib_rx *rx, struct lnet_msg *lntmsg)
1770 struct lnet_process_id target = lntmsg->msg_target;
1771 unsigned int niov = lntmsg->msg_niov;
1772 struct bio_vec *kiov = lntmsg->msg_kiov;
1773 unsigned int offset = lntmsg->msg_offset;
1774 unsigned int nob = lntmsg->msg_len;
1778 tx = kiblnd_get_idle_tx(ni, rx->rx_conn->ibc_peer->ibp_nid);
1780 CERROR("Can't get tx for REPLY to %s\n",
1781 libcfs_nid2str(target.nid));
1788 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1789 niov, kiov, offset, nob);
1792 CERROR("Can't setup GET src for %s: %d\n",
1793 libcfs_nid2str(target.nid), rc);
1797 rc = kiblnd_init_rdma(rx->rx_conn, tx,
1798 IBLND_MSG_GET_DONE, nob,
1799 &rx->rx_msg->ibm_u.get.ibgm_rd,
1800 rx->rx_msg->ibm_u.get.ibgm_cookie);
1802 CERROR("Can't setup rdma for GET from %s: %d\n",
1803 libcfs_nid2str(target.nid), rc);
1808 /* No RDMA: local completion may happen now! */
1809 lnet_finalize(lntmsg, 0);
1811 /* RDMA: lnet_finalize(lntmsg) when it
1813 tx->tx_lntmsg[0] = lntmsg;
1816 kiblnd_queue_tx(tx, rx->rx_conn);
1821 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
1824 lnet_finalize(lntmsg, -EIO);
1828 kiblnd_recv(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg,
1829 int delayed, unsigned int niov, struct bio_vec *kiov,
1830 unsigned int offset, unsigned int mlen, unsigned int rlen)
1832 struct kib_rx *rx = private;
1833 struct kib_msg *rxmsg = rx->rx_msg;
1834 struct kib_conn *conn = rx->rx_conn;
1838 int post_credit = IBLND_POSTRX_PEER_CREDIT;
1841 LASSERT (mlen <= rlen);
1842 LASSERT (!in_interrupt());
1844 switch (rxmsg->ibm_type) {
1848 case IBLND_MSG_IMMEDIATE:
1849 nob = offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[rlen]);
1850 if (nob > rx->rx_nob) {
1851 CERROR ("Immediate message from %s too big: %d(%d)\n",
1852 libcfs_nid2str(rxmsg->ibm_u.immediate.ibim_hdr.src_nid),
1858 lnet_copy_flat2kiov(niov, kiov, offset,
1859 IBLND_MSG_SIZE, rxmsg,
1860 offsetof(struct kib_msg,
1861 ibm_u.immediate.ibim_payload),
1863 lnet_finalize(lntmsg, 0);
1866 case IBLND_MSG_PUT_REQ: {
1867 struct kib_msg *txmsg;
1868 struct kib_rdma_desc *rd;
1869 ibprm_cookie = rxmsg->ibm_u.putreq.ibprm_cookie;
1872 lnet_finalize(lntmsg, 0);
1873 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK,
1878 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
1880 CERROR("Can't allocate tx for %s\n",
1881 libcfs_nid2str(conn->ibc_peer->ibp_nid));
1882 /* Not replying will break the connection */
1888 rd = &txmsg->ibm_u.putack.ibpam_rd;
1889 rc = kiblnd_setup_rd_kiov(ni, tx, rd,
1890 niov, kiov, offset, mlen);
1892 CERROR("Can't setup PUT sink for %s: %d\n",
1893 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
1894 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
1896 /* tell peer_ni it's over */
1897 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK,
1902 nob = offsetof(struct kib_putack_msg, ibpam_rd.rd_frags[rd->rd_nfrags]);
1903 txmsg->ibm_u.putack.ibpam_src_cookie = ibprm_cookie;
1904 txmsg->ibm_u.putack.ibpam_dst_cookie = tx->tx_cookie;
1906 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_ACK, nob);
1908 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1909 tx->tx_waiting = 1; /* waiting for PUT_DONE */
1910 kiblnd_queue_tx(tx, conn);
1912 /* reposted buffer reserved for PUT_DONE */
1913 post_credit = IBLND_POSTRX_NO_CREDIT;
1917 case IBLND_MSG_GET_REQ:
1918 if (lntmsg != NULL) {
1919 /* Optimized GET; RDMA lntmsg's payload */
1920 kiblnd_reply(ni, rx, lntmsg);
1922 /* GET didn't match anything */
1923 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_GET_DONE,
1925 rxmsg->ibm_u.get.ibgm_cookie);
1930 kiblnd_post_rx(rx, post_credit);
1935 kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name)
1937 struct task_struct *task = kthread_run(fn, arg, "%s", name);
1940 return PTR_ERR(task);
1942 atomic_inc(&kiblnd_data.kib_nthreads);
1947 kiblnd_thread_fini (void)
1949 atomic_dec (&kiblnd_data.kib_nthreads);
1953 kiblnd_peer_alive(struct kib_peer_ni *peer_ni)
1955 /* This is racy, but everyone's only writing ktime_get_seconds() */
1956 peer_ni->ibp_last_alive = ktime_get_seconds();
1961 kiblnd_peer_notify(struct kib_peer_ni *peer_ni)
1964 time64_t last_alive = 0;
1965 unsigned long flags;
1967 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1969 if (kiblnd_peer_idle(peer_ni) && peer_ni->ibp_error != 0) {
1970 error = peer_ni->ibp_error;
1971 peer_ni->ibp_error = 0;
1973 last_alive = peer_ni->ibp_last_alive;
1976 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1979 lnet_notify(peer_ni->ibp_ni,
1980 peer_ni->ibp_nid, false, false, last_alive);
1984 kiblnd_close_conn_locked(struct kib_conn *conn, int error)
1986 /* This just does the immediate housekeeping. 'error' is zero for a
1987 * normal shutdown which can happen only after the connection has been
1988 * established. If the connection is established, schedule the
1989 * connection to be finished off by the connd. Otherwise the connd is
1990 * already dealing with it (either to set it up or tear it down).
1991 * Caller holds kib_global_lock exclusively in irq context */
1992 struct kib_peer_ni *peer_ni = conn->ibc_peer;
1993 struct kib_dev *dev;
1994 unsigned long flags;
1996 LASSERT (error != 0 || conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1998 if (error != 0 && conn->ibc_comms_error == 0)
1999 conn->ibc_comms_error = error;
2001 if (conn->ibc_state != IBLND_CONN_ESTABLISHED)
2002 return; /* already being handled */
2005 list_empty(&conn->ibc_tx_noops) &&
2006 list_empty(&conn->ibc_tx_queue) &&
2007 list_empty(&conn->ibc_tx_queue_rsrvd) &&
2008 list_empty(&conn->ibc_tx_queue_nocred) &&
2009 list_empty(&conn->ibc_active_txs)) {
2010 CDEBUG(D_NET, "closing conn to %s\n",
2011 libcfs_nid2str(peer_ni->ibp_nid));
2013 CNETERR("Closing conn to %s: error %d%s%s%s%s%s\n",
2014 libcfs_nid2str(peer_ni->ibp_nid), error,
2015 list_empty(&conn->ibc_tx_queue) ? "" : "(sending)",
2016 list_empty(&conn->ibc_tx_noops) ? "" : "(sending_noops)",
2017 list_empty(&conn->ibc_tx_queue_rsrvd) ?
2018 "" : "(sending_rsrvd)",
2019 list_empty(&conn->ibc_tx_queue_nocred) ?
2020 "" : "(sending_nocred)",
2021 list_empty(&conn->ibc_active_txs) ? "" : "(waiting)");
2024 dev = ((struct kib_net *)peer_ni->ibp_ni->ni_data)->ibn_dev;
2025 if (peer_ni->ibp_next_conn == conn)
2026 /* clear next_conn so it won't be used */
2027 peer_ni->ibp_next_conn = NULL;
2028 list_del(&conn->ibc_list);
2029 /* connd (see below) takes over ibc_list's ref */
2031 if (list_empty(&peer_ni->ibp_conns) && /* no more conns */
2032 kiblnd_peer_active(peer_ni)) { /* still in peer_ni table */
2033 kiblnd_unlink_peer_locked(peer_ni);
2035 /* set/clear error on last conn */
2036 peer_ni->ibp_error = conn->ibc_comms_error;
2039 kiblnd_set_conn_state(conn, IBLND_CONN_CLOSING);
2042 kiblnd_dev_can_failover(dev)) {
2043 list_add_tail(&dev->ibd_fail_list,
2044 &kiblnd_data.kib_failed_devs);
2045 wake_up(&kiblnd_data.kib_failover_waitq);
2048 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
2050 list_add_tail(&conn->ibc_list, &kiblnd_data.kib_connd_conns);
2051 wake_up(&kiblnd_data.kib_connd_waitq);
2053 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
2057 kiblnd_close_conn(struct kib_conn *conn, int error)
2059 unsigned long flags;
2061 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2063 kiblnd_close_conn_locked(conn, error);
2065 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2069 kiblnd_handle_early_rxs(struct kib_conn *conn)
2071 unsigned long flags;
2074 LASSERT(!in_interrupt());
2075 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
2077 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2078 while ((rx = list_first_entry_or_null(&conn->ibc_early_rxs,
2080 rx_list)) != NULL) {
2081 list_del(&rx->rx_list);
2082 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2084 kiblnd_handle_rx(rx);
2086 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2088 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2092 kiblnd_abort_txs(struct kib_conn *conn, struct list_head *txs)
2098 spin_lock(&conn->ibc_lock);
2100 list_for_each_entry_safe(tx, nxt, txs, tx_list) {
2101 if (txs == &conn->ibc_active_txs) {
2102 LASSERT(!tx->tx_queued);
2103 LASSERT(tx->tx_waiting ||
2104 tx->tx_sending != 0);
2105 if (conn->ibc_comms_error == -ETIMEDOUT) {
2106 if (tx->tx_waiting && !tx->tx_sending)
2108 LNET_MSG_STATUS_REMOTE_TIMEOUT;
2109 else if (tx->tx_sending)
2111 LNET_MSG_STATUS_NETWORK_TIMEOUT;
2114 LASSERT(tx->tx_queued);
2115 if (conn->ibc_comms_error == -ETIMEDOUT)
2116 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_TIMEOUT;
2118 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
2121 tx->tx_status = -ECONNABORTED;
2125 * TODO: This makes an assumption that
2126 * kiblnd_tx_complete() will be called for each tx. If
2127 * that event is dropped we could end up with stale
2128 * connections floating around. We'd like to deal with
2129 * that in a better way.
2131 * Also that means we can exceed the timeout by many
2134 if (tx->tx_sending == 0) {
2136 list_move(&tx->tx_list, &zombies);
2138 /* keep tx until cq destroy */
2139 list_move(&tx->tx_list, &conn->ibc_zombie_txs);
2144 spin_unlock(&conn->ibc_lock);
2147 * aborting transmits occurs when finalizing the connection.
2148 * The connection is finalized on error.
2149 * Passing LNET_MSG_STATUS_OK to txlist_done() will not
2150 * override the value already set in tx->tx_hstatus above.
2152 kiblnd_txlist_done(&zombies, -ECONNABORTED, LNET_MSG_STATUS_OK);
2156 kiblnd_tx_may_discard(struct kib_conn *conn)
2162 spin_lock(&conn->ibc_lock);
2164 list_for_each_entry_safe(tx, nxt, &conn->ibc_zombie_txs, tx_list) {
2165 if (tx->tx_sending > 0 && tx->tx_lntmsg[0] &&
2166 lnet_md_discarded(tx->tx_lntmsg[0]->msg_md)) {
2168 if (tx->tx_sending == 0) {
2169 kiblnd_conn_decref(tx->tx_conn);
2176 spin_unlock(&conn->ibc_lock);
2181 kiblnd_finalise_conn(struct kib_conn *conn)
2183 LASSERT (!in_interrupt());
2184 LASSERT (conn->ibc_state > IBLND_CONN_INIT);
2186 /* abort_receives moves QP state to IB_QPS_ERR. This is only required
2187 * for connections that didn't get as far as being connected, because
2188 * rdma_disconnect() does this for free. */
2189 kiblnd_abort_receives(conn);
2191 kiblnd_set_conn_state(conn, IBLND_CONN_DISCONNECTED);
2193 /* Complete all tx descs not waiting for sends to complete.
2194 * NB we should be safe from RDMA now that the QP has changed state */
2196 kiblnd_abort_txs(conn, &conn->ibc_tx_noops);
2197 kiblnd_abort_txs(conn, &conn->ibc_tx_queue);
2198 kiblnd_abort_txs(conn, &conn->ibc_tx_queue_rsrvd);
2199 kiblnd_abort_txs(conn, &conn->ibc_tx_queue_nocred);
2200 kiblnd_abort_txs(conn, &conn->ibc_active_txs);
2202 kiblnd_handle_early_rxs(conn);
2206 kiblnd_peer_connect_failed(struct kib_peer_ni *peer_ni, int active,
2210 unsigned long flags;
2211 enum lnet_msg_hstatus hstatus;
2213 LASSERT(error != 0);
2214 LASSERT(!in_interrupt());
2216 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2219 LASSERT(peer_ni->ibp_connecting > 0);
2220 peer_ni->ibp_connecting--;
2222 LASSERT (peer_ni->ibp_accepting > 0);
2223 peer_ni->ibp_accepting--;
2226 if (kiblnd_peer_connecting(peer_ni)) {
2227 /* another connection attempt under way... */
2228 write_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2233 peer_ni->ibp_reconnected = 0;
2234 if (list_empty(&peer_ni->ibp_conns)) {
2235 /* Take peer_ni's blocked transmits to complete with error */
2236 list_splice_init(&peer_ni->ibp_tx_queue, &zombies);
2238 if (kiblnd_peer_active(peer_ni))
2239 kiblnd_unlink_peer_locked(peer_ni);
2241 peer_ni->ibp_error = error;
2243 /* Can't have blocked transmits if there are connections */
2244 LASSERT(list_empty(&peer_ni->ibp_tx_queue));
2247 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2249 kiblnd_peer_notify(peer_ni);
2251 if (list_empty(&zombies))
2254 CNETERR("Deleting messages for %s: connection failed\n",
2255 libcfs_nid2str(peer_ni->ibp_nid));
2260 hstatus = LNET_MSG_STATUS_NETWORK_TIMEOUT;
2263 hstatus = LNET_MSG_STATUS_REMOTE_DROPPED;
2266 hstatus = LNET_MSG_STATUS_LOCAL_DROPPED;
2270 kiblnd_txlist_done(&zombies, error, hstatus);
2274 kiblnd_connreq_done(struct kib_conn *conn, int status)
2276 struct kib_peer_ni *peer_ni = conn->ibc_peer;
2279 unsigned long flags;
2282 active = (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2284 CDEBUG(D_NET,"%s: active(%d), version(%x), status(%d)\n",
2285 libcfs_nid2str(peer_ni->ibp_nid), active,
2286 conn->ibc_version, status);
2288 LASSERT (!in_interrupt());
2289 LASSERT ((conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT &&
2290 peer_ni->ibp_connecting > 0) ||
2291 (conn->ibc_state == IBLND_CONN_PASSIVE_WAIT &&
2292 peer_ni->ibp_accepting > 0));
2294 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
2295 conn->ibc_connvars = NULL;
2298 /* failed to establish connection */
2299 kiblnd_peer_connect_failed(peer_ni, active, status);
2300 kiblnd_finalise_conn(conn);
2304 /* connection established */
2305 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2307 conn->ibc_last_send = ktime_get();
2308 kiblnd_set_conn_state(conn, IBLND_CONN_ESTABLISHED);
2309 kiblnd_peer_alive(peer_ni);
2311 /* Add conn to peer_ni's list and nuke any dangling conns from a different
2312 * peer_ni instance... */
2313 kiblnd_conn_addref(conn); /* +1 ref for ibc_list */
2314 list_add(&conn->ibc_list, &peer_ni->ibp_conns);
2315 peer_ni->ibp_reconnected = 0;
2317 peer_ni->ibp_connecting--;
2319 peer_ni->ibp_accepting--;
2321 if (peer_ni->ibp_version == 0) {
2322 peer_ni->ibp_version = conn->ibc_version;
2323 peer_ni->ibp_incarnation = conn->ibc_incarnation;
2326 if (peer_ni->ibp_version != conn->ibc_version ||
2327 peer_ni->ibp_incarnation != conn->ibc_incarnation) {
2328 kiblnd_close_stale_conns_locked(peer_ni, conn->ibc_version,
2329 conn->ibc_incarnation);
2330 peer_ni->ibp_version = conn->ibc_version;
2331 peer_ni->ibp_incarnation = conn->ibc_incarnation;
2334 /* grab pending txs while I have the lock */
2335 list_splice_init(&peer_ni->ibp_tx_queue, &txs);
2337 if (!kiblnd_peer_active(peer_ni) || /* peer_ni has been deleted */
2338 conn->ibc_comms_error != 0) { /* error has happened already */
2340 /* start to shut down connection */
2341 kiblnd_close_conn_locked(conn, -ECONNABORTED);
2342 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2344 kiblnd_txlist_done(&txs, -ECONNABORTED,
2345 LNET_MSG_STATUS_LOCAL_ERROR);
2350 /* +1 ref for myself, this connection is visible to other threads
2351 * now, refcount of peer:ibp_conns can be released by connection
2352 * close from either a different thread, or the calling of
2353 * kiblnd_check_sends_locked() below. See bz21911 for details.
2355 kiblnd_conn_addref(conn);
2356 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2358 /* Schedule blocked txs
2359 * Note: if we are running with conns_per_peer > 1, these blocked
2360 * txs will all get scheduled to the first connection which gets
2361 * scheduled. We won't be using round robin on this first batch.
2363 spin_lock(&conn->ibc_lock);
2364 while ((tx = list_first_entry_or_null(&txs, struct kib_tx,
2365 tx_list)) != NULL) {
2366 list_del(&tx->tx_list);
2368 kiblnd_queue_tx_locked(tx, conn);
2370 kiblnd_check_sends_locked(conn);
2371 spin_unlock(&conn->ibc_lock);
2373 /* schedule blocked rxs */
2374 kiblnd_handle_early_rxs(conn);
2375 kiblnd_conn_decref(conn);
2379 kiblnd_reject(struct rdma_cm_id *cmid, struct kib_rej *rej)
2383 #ifdef HAVE_RDMA_REJECT_4ARGS
2384 rc = rdma_reject(cmid, rej, sizeof(*rej), IB_CM_REJ_CONSUMER_DEFINED);
2386 rc = rdma_reject(cmid, rej, sizeof(*rej));
2390 CWARN("Error %d sending reject\n", rc);
2394 kiblnd_passive_connect(struct rdma_cm_id *cmid, void *priv, int priv_nob)
2396 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2397 struct kib_msg *reqmsg = priv;
2398 struct kib_msg *ackmsg;
2399 struct kib_dev *ibdev;
2400 struct kib_peer_ni *peer_ni;
2401 struct kib_peer_ni *peer2;
2402 struct kib_conn *conn;
2403 struct lnet_ni *ni = NULL;
2404 struct kib_net *net = NULL;
2406 struct rdma_conn_param cp;
2408 int version = IBLND_MSG_VERSION;
2409 unsigned long flags;
2411 struct sockaddr_in *peer_addr;
2413 LASSERT(!in_interrupt());
2414 /* cmid inherits 'context' from the corresponding listener id */
2415 ibdev = cmid->context;
2418 memset(&rej, 0, sizeof(rej));
2419 rej.ibr_magic = IBLND_MSG_MAGIC;
2420 rej.ibr_why = IBLND_REJECT_FATAL;
2421 rej.ibr_cp.ibcp_max_msg_size = IBLND_MSG_SIZE;
2423 peer_addr = (struct sockaddr_in *)&(cmid->route.addr.dst_addr);
2424 if (*kiblnd_tunables.kib_require_priv_port &&
2425 ntohs(peer_addr->sin_port) >= PROT_SOCK) {
2426 __u32 ip = ntohl(peer_addr->sin_addr.s_addr);
2427 CERROR("peer_ni's port (%pI4h:%hu) is not privileged\n",
2428 &ip, ntohs(peer_addr->sin_port));
2432 if (priv_nob < offsetof(struct kib_msg, ibm_type)) {
2433 CERROR("Short connection request\n");
2437 /* Future protocol version compatibility support! If the
2438 * o2iblnd-specific protocol changes, or when LNET unifies
2439 * protocols over all LNDs, the initial connection will
2440 * negotiate a protocol version. I trap this here to avoid
2441 * console errors; the reject tells the peer_ni which protocol I
2443 if (reqmsg->ibm_magic == LNET_PROTO_MAGIC ||
2444 reqmsg->ibm_magic == __swab32(LNET_PROTO_MAGIC))
2446 if (reqmsg->ibm_magic == IBLND_MSG_MAGIC &&
2447 reqmsg->ibm_version != IBLND_MSG_VERSION &&
2448 reqmsg->ibm_version != IBLND_MSG_VERSION_1)
2450 if (reqmsg->ibm_magic == __swab32(IBLND_MSG_MAGIC) &&
2451 reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION) &&
2452 reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION_1))
2455 rc = kiblnd_unpack_msg(reqmsg, priv_nob);
2457 CERROR("Can't parse connection request: %d\n", rc);
2461 nid = reqmsg->ibm_srcnid;
2462 ni = lnet_nid2ni_addref(reqmsg->ibm_dstnid);
2465 net = (struct kib_net *)ni->ni_data;
2466 rej.ibr_incarnation = net->ibn_incarnation;
2469 if (ni == NULL || /* no matching net */
2470 ni->ni_nid != reqmsg->ibm_dstnid || /* right NET, wrong NID! */
2471 net->ibn_dev != ibdev) { /* wrong device */
2472 CERROR("Can't accept conn from %s on %s (%s:%d:%pI4h): bad dst nid %s\n", libcfs_nid2str(nid),
2473 ni ? libcfs_nid2str(ni->ni_nid) : "NA",
2474 ibdev->ibd_ifname, ibdev->ibd_nnets,
2476 libcfs_nid2str(reqmsg->ibm_dstnid));
2481 /* check time stamp as soon as possible */
2482 if (reqmsg->ibm_dststamp != 0 &&
2483 reqmsg->ibm_dststamp != net->ibn_incarnation) {
2484 CWARN("Stale connection request\n");
2485 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2489 /* I can accept peer_ni's version */
2490 version = reqmsg->ibm_version;
2492 if (reqmsg->ibm_type != IBLND_MSG_CONNREQ) {
2493 CERROR("Unexpected connreq msg type: %x from %s\n",
2494 reqmsg->ibm_type, libcfs_nid2str(nid));
2498 if (reqmsg->ibm_u.connparams.ibcp_queue_depth >
2499 kiblnd_msg_queue_size(version, ni)) {
2500 CERROR("Can't accept conn from %s, queue depth too large: %d (<=%d wanted)\n",
2501 libcfs_nid2str(nid),
2502 reqmsg->ibm_u.connparams.ibcp_queue_depth,
2503 kiblnd_msg_queue_size(version, ni));
2505 if (version == IBLND_MSG_VERSION)
2506 rej.ibr_why = IBLND_REJECT_MSG_QUEUE_SIZE;
2511 if (reqmsg->ibm_u.connparams.ibcp_max_frags >
2512 IBLND_MAX_RDMA_FRAGS) {
2513 CWARN("Can't accept conn from %s (version %x): max_frags %d too large (%d wanted)\n",
2514 libcfs_nid2str(nid), version,
2515 reqmsg->ibm_u.connparams.ibcp_max_frags,
2516 IBLND_MAX_RDMA_FRAGS);
2518 if (version >= IBLND_MSG_VERSION)
2519 rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2522 } else if (reqmsg->ibm_u.connparams.ibcp_max_frags <
2523 IBLND_MAX_RDMA_FRAGS &&
2524 net->ibn_fmr_ps == NULL) {
2525 CWARN("Can't accept conn from %s (version %x): max_frags %d incompatible without FMR pool (%d wanted)\n",
2526 libcfs_nid2str(nid), version,
2527 reqmsg->ibm_u.connparams.ibcp_max_frags,
2528 IBLND_MAX_RDMA_FRAGS);
2530 if (version == IBLND_MSG_VERSION)
2531 rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2536 if (reqmsg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2537 CERROR("Can't accept %s: message size %d too big (%d max)\n",
2538 libcfs_nid2str(nid),
2539 reqmsg->ibm_u.connparams.ibcp_max_msg_size,
2544 /* assume 'nid' is a new peer_ni; create */
2545 rc = kiblnd_create_peer(ni, &peer_ni, nid);
2547 CERROR("Can't create peer_ni for %s\n", libcfs_nid2str(nid));
2548 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2552 /* We have validated the peer's parameters so use those */
2553 peer_ni->ibp_max_frags = reqmsg->ibm_u.connparams.ibcp_max_frags;
2554 peer_ni->ibp_queue_depth = reqmsg->ibm_u.connparams.ibcp_queue_depth;
2556 write_lock_irqsave(g_lock, flags);
2558 peer2 = kiblnd_find_peer_locked(ni, nid);
2559 if (peer2 != NULL) {
2560 if (peer2->ibp_version == 0) {
2561 peer2->ibp_version = version;
2562 peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2565 /* not the guy I've talked with */
2566 if (peer2->ibp_incarnation != reqmsg->ibm_srcstamp ||
2567 peer2->ibp_version != version) {
2568 kiblnd_close_peer_conns_locked(peer2, -ESTALE);
2570 if (kiblnd_peer_active(peer2)) {
2571 peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2572 peer2->ibp_version = version;
2574 write_unlock_irqrestore(g_lock, flags);
2576 CWARN("Conn stale %s version %x/%x incarnation %llu/%llu\n",
2577 libcfs_nid2str(nid), peer2->ibp_version, version,
2578 peer2->ibp_incarnation, reqmsg->ibm_srcstamp);
2580 kiblnd_peer_decref(peer_ni);
2581 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2585 /* Tie-break connection race in favour of the higher NID.
2586 * If we keep running into a race condition multiple times,
2587 * we have to assume that the connection attempt with the
2588 * higher NID is stuck in a connecting state and will never
2589 * recover. As such, we pass through this if-block and let
2590 * the lower NID connection win so we can move forward.
2592 if (peer2->ibp_connecting != 0 &&
2593 nid < ni->ni_nid && peer2->ibp_races <
2594 MAX_CONN_RACES_BEFORE_ABORT) {
2596 write_unlock_irqrestore(g_lock, flags);
2598 CDEBUG(D_NET, "Conn race %s\n",
2599 libcfs_nid2str(peer2->ibp_nid));
2601 kiblnd_peer_decref(peer_ni);
2602 rej.ibr_why = IBLND_REJECT_CONN_RACE;
2605 if (peer2->ibp_races >= MAX_CONN_RACES_BEFORE_ABORT)
2606 CNETERR("Conn race %s: unresolved after %d attempts, letting lower NID win\n",
2607 libcfs_nid2str(peer2->ibp_nid),
2608 MAX_CONN_RACES_BEFORE_ABORT);
2610 * passive connection is allowed even this peer_ni is waiting for
2613 peer2->ibp_reconnecting = 0;
2614 peer2->ibp_races = 0;
2615 peer2->ibp_accepting++;
2616 kiblnd_peer_addref(peer2);
2618 /* Race with kiblnd_launch_tx (active connect) to create peer_ni
2619 * so copy validated parameters since we now know what the
2620 * peer_ni's limits are */
2621 peer2->ibp_max_frags = peer_ni->ibp_max_frags;
2622 peer2->ibp_queue_depth = peer_ni->ibp_queue_depth;
2624 write_unlock_irqrestore(g_lock, flags);
2625 kiblnd_peer_decref(peer_ni);
2628 /* Brand new peer_ni */
2629 LASSERT(peer_ni->ibp_accepting == 0);
2630 LASSERT(peer_ni->ibp_version == 0 &&
2631 peer_ni->ibp_incarnation == 0);
2633 peer_ni->ibp_accepting = 1;
2634 peer_ni->ibp_version = version;
2635 peer_ni->ibp_incarnation = reqmsg->ibm_srcstamp;
2637 /* I have a ref on ni that prevents it being shutdown */
2638 LASSERT(net->ibn_shutdown == 0);
2640 kiblnd_peer_addref(peer_ni);
2641 hash_add(kiblnd_data.kib_peers, &peer_ni->ibp_list, nid);
2643 write_unlock_irqrestore(g_lock, flags);
2646 conn = kiblnd_create_conn(peer_ni, cmid, IBLND_CONN_PASSIVE_WAIT,
2649 kiblnd_peer_connect_failed(peer_ni, 0, -ENOMEM);
2650 kiblnd_peer_decref(peer_ni);
2651 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2655 /* conn now "owns" cmid, so I return success from here on to ensure the
2656 * CM callback doesn't destroy cmid.
2658 conn->ibc_incarnation = reqmsg->ibm_srcstamp;
2659 conn->ibc_credits = conn->ibc_queue_depth;
2660 conn->ibc_reserved_credits = conn->ibc_queue_depth;
2661 LASSERT(conn->ibc_credits + conn->ibc_reserved_credits +
2662 IBLND_OOB_MSGS(version) <= IBLND_RX_MSGS(conn));
2664 ackmsg = &conn->ibc_connvars->cv_msg;
2665 memset(ackmsg, 0, sizeof(*ackmsg));
2667 kiblnd_init_msg(ackmsg, IBLND_MSG_CONNACK,
2668 sizeof(ackmsg->ibm_u.connparams));
2669 ackmsg->ibm_u.connparams.ibcp_queue_depth = conn->ibc_queue_depth;
2670 ackmsg->ibm_u.connparams.ibcp_max_frags = conn->ibc_max_frags;
2671 ackmsg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2673 kiblnd_pack_msg(ni, ackmsg, version, 0, nid, reqmsg->ibm_srcstamp);
2675 memset(&cp, 0, sizeof(cp));
2676 cp.private_data = ackmsg;
2677 cp.private_data_len = ackmsg->ibm_nob;
2678 cp.responder_resources = 0; /* No atomic ops or RDMA reads */
2679 cp.initiator_depth = 0;
2680 cp.flow_control = 1;
2681 cp.retry_count = *kiblnd_tunables.kib_retry_count;
2682 cp.rnr_retry_count = *kiblnd_tunables.kib_rnr_retry_count;
2684 CDEBUG(D_NET, "Accept %s\n", libcfs_nid2str(nid));
2686 rc = rdma_accept(cmid, &cp);
2688 CERROR("Can't accept %s: %d\n", libcfs_nid2str(nid), rc);
2689 rej.ibr_version = version;
2690 rej.ibr_why = IBLND_REJECT_FATAL;
2692 kiblnd_reject(cmid, &rej);
2693 kiblnd_connreq_done(conn, rc);
2694 kiblnd_conn_decref(conn);
2702 rej.ibr_cp.ibcp_queue_depth =
2703 kiblnd_msg_queue_size(version, ni);
2704 rej.ibr_cp.ibcp_max_frags = IBLND_MAX_RDMA_FRAGS;
2708 rej.ibr_version = version;
2709 kiblnd_reject(cmid, &rej);
2711 return -ECONNREFUSED;
2715 kiblnd_check_reconnect(struct kib_conn *conn, int version,
2716 u64 incarnation, int why, struct kib_connparams *cp)
2718 rwlock_t *glock = &kiblnd_data.kib_global_lock;
2719 struct kib_peer_ni *peer_ni = conn->ibc_peer;
2721 int msg_size = IBLND_MSG_SIZE;
2725 unsigned long flags;
2727 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2728 LASSERT(peer_ni->ibp_connecting > 0); /* 'conn' at least */
2731 msg_size = cp->ibcp_max_msg_size;
2732 frag_num = cp->ibcp_max_frags;
2733 queue_dep = cp->ibcp_queue_depth;
2736 write_lock_irqsave(glock, flags);
2737 /* retry connection if it's still needed and no other connection
2738 * attempts (active or passive) are in progress
2739 * NB: reconnect is still needed even when ibp_tx_queue is
2740 * empty if ibp_version != version because reconnect may be
2743 reconnect = (!list_empty(&peer_ni->ibp_tx_queue) ||
2744 peer_ni->ibp_version != version) &&
2745 peer_ni->ibp_connecting &&
2746 peer_ni->ibp_accepting == 0;
2757 case IBLND_REJECT_RDMA_FRAGS: {
2758 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2761 reason = "can't negotiate max frags";
2764 tunables = &peer_ni->ibp_ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2765 #ifdef HAVE_IB_GET_DMA_MR
2767 * This check only makes sense if the kernel supports global
2768 * memory registration. Otherwise, map_on_demand will never == 0
2770 if (!tunables->lnd_map_on_demand) {
2771 reason = "map_on_demand must be enabled";
2775 if (conn->ibc_max_frags <= frag_num) {
2776 reason = "unsupported max frags";
2780 peer_ni->ibp_max_frags = frag_num;
2781 reason = "rdma fragments";
2784 case IBLND_REJECT_MSG_QUEUE_SIZE:
2786 reason = "can't negotiate queue depth";
2789 if (conn->ibc_queue_depth <= queue_dep) {
2790 reason = "unsupported queue depth";
2794 peer_ni->ibp_queue_depth = queue_dep;
2795 reason = "queue depth";
2798 case IBLND_REJECT_CONN_STALE:
2802 case IBLND_REJECT_CONN_RACE:
2803 reason = "conn race";
2806 case IBLND_REJECT_CONN_UNCOMPAT:
2807 reason = "version negotiation";
2811 conn->ibc_reconnect = 1;
2812 peer_ni->ibp_reconnecting++;
2813 peer_ni->ibp_version = version;
2814 if (incarnation != 0)
2815 peer_ni->ibp_incarnation = incarnation;
2817 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2819 CNETERR("%s: %s (%s), %x, %x, msg_size: %d, queue_depth: %d/%d, max_frags: %d/%d\n",
2820 libcfs_nid2str(peer_ni->ibp_nid),
2821 reconnect ? "reconnect" : "don't reconnect",
2822 reason, IBLND_MSG_VERSION, version, msg_size,
2823 conn->ibc_queue_depth, queue_dep,
2824 conn->ibc_max_frags, frag_num);
2826 * if conn::ibc_reconnect is TRUE, connd will reconnect to the peer_ni
2827 * while destroying the zombie
2832 kiblnd_rejected(struct kib_conn *conn, int reason, void *priv, int priv_nob)
2834 struct kib_peer_ni *peer_ni = conn->ibc_peer;
2835 int status = -ECONNREFUSED;
2837 LASSERT (!in_interrupt());
2838 LASSERT (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2841 case IB_CM_REJ_STALE_CONN:
2842 kiblnd_check_reconnect(conn, IBLND_MSG_VERSION, 0,
2843 IBLND_REJECT_CONN_STALE, NULL);
2846 case IB_CM_REJ_INVALID_SERVICE_ID:
2847 status = -EHOSTUNREACH;
2848 CNETERR("%s rejected: no listener at %d\n",
2849 libcfs_nid2str(peer_ni->ibp_nid),
2850 *kiblnd_tunables.kib_service);
2853 case IB_CM_REJ_CONSUMER_DEFINED:
2854 if (priv_nob >= offsetof(struct kib_rej, ibr_padding)) {
2855 struct kib_rej *rej = priv;
2856 struct kib_connparams *cp = NULL;
2858 __u64 incarnation = -1;
2860 /* NB. default incarnation is -1 because:
2861 * a) V1 will ignore dst incarnation in connreq.
2862 * b) V2 will provide incarnation while rejecting me,
2863 * -1 will be overwrote.
2865 * if I try to connect to a V1 peer_ni with V2 protocol,
2866 * it rejected me then upgrade to V2, I have no idea
2867 * about the upgrading and try to reconnect with V1,
2868 * in this case upgraded V2 can find out I'm trying to
2869 * talk to the old guy and reject me(incarnation is -1).
2872 if (rej->ibr_magic == __swab32(IBLND_MSG_MAGIC) ||
2873 rej->ibr_magic == __swab32(LNET_PROTO_MAGIC)) {
2874 __swab32s(&rej->ibr_magic);
2875 __swab16s(&rej->ibr_version);
2879 if (priv_nob >= sizeof(struct kib_rej) &&
2880 rej->ibr_version > IBLND_MSG_VERSION_1) {
2881 /* priv_nob is always 148 in current version
2882 * of OFED, so we still need to check version.
2883 * (define of IB_CM_REJ_PRIVATE_DATA_SIZE)
2888 __swab64s(&rej->ibr_incarnation);
2889 __swab16s(&cp->ibcp_queue_depth);
2890 __swab16s(&cp->ibcp_max_frags);
2891 __swab32s(&cp->ibcp_max_msg_size);
2894 incarnation = rej->ibr_incarnation;
2897 if (rej->ibr_magic != IBLND_MSG_MAGIC &&
2898 rej->ibr_magic != LNET_PROTO_MAGIC) {
2899 CERROR("%s rejected: consumer defined fatal error\n",
2900 libcfs_nid2str(peer_ni->ibp_nid));
2904 if (rej->ibr_version != IBLND_MSG_VERSION &&
2905 rej->ibr_version != IBLND_MSG_VERSION_1) {
2906 CERROR("%s rejected: o2iblnd version %x error\n",
2907 libcfs_nid2str(peer_ni->ibp_nid),
2912 if (rej->ibr_why == IBLND_REJECT_FATAL &&
2913 rej->ibr_version == IBLND_MSG_VERSION_1) {
2914 CDEBUG(D_NET, "rejected by old version peer_ni %s: %x\n",
2915 libcfs_nid2str(peer_ni->ibp_nid),
2918 if (conn->ibc_version != IBLND_MSG_VERSION_1)
2919 rej->ibr_why = IBLND_REJECT_CONN_UNCOMPAT;
2922 switch (rej->ibr_why) {
2923 case IBLND_REJECT_CONN_RACE:
2924 case IBLND_REJECT_CONN_STALE:
2925 case IBLND_REJECT_CONN_UNCOMPAT:
2926 case IBLND_REJECT_MSG_QUEUE_SIZE:
2927 case IBLND_REJECT_RDMA_FRAGS:
2928 kiblnd_check_reconnect(conn, rej->ibr_version,
2933 case IBLND_REJECT_NO_RESOURCES:
2934 CERROR("%s rejected: o2iblnd no resources\n",
2935 libcfs_nid2str(peer_ni->ibp_nid));
2938 case IBLND_REJECT_FATAL:
2939 CERROR("%s rejected: o2iblnd fatal error\n",
2940 libcfs_nid2str(peer_ni->ibp_nid));
2944 CERROR("%s rejected: o2iblnd reason %d\n",
2945 libcfs_nid2str(peer_ni->ibp_nid),
2953 CNETERR("%s rejected: reason %d, size %d\n",
2954 libcfs_nid2str(peer_ni->ibp_nid), reason, priv_nob);
2958 kiblnd_connreq_done(conn, status);
2962 kiblnd_check_connreply(struct kib_conn *conn, void *priv, int priv_nob)
2964 struct kib_peer_ni *peer_ni = conn->ibc_peer;
2965 struct lnet_ni *ni = peer_ni->ibp_ni;
2966 struct kib_net *net = ni->ni_data;
2967 struct kib_msg *msg = priv;
2968 int ver = conn->ibc_version;
2969 int rc = kiblnd_unpack_msg(msg, priv_nob);
2970 unsigned long flags;
2972 LASSERT (net != NULL);
2975 CERROR("Can't unpack connack from %s: %d\n",
2976 libcfs_nid2str(peer_ni->ibp_nid), rc);
2980 if (msg->ibm_type != IBLND_MSG_CONNACK) {
2981 CERROR("Unexpected message %d from %s\n",
2982 msg->ibm_type, libcfs_nid2str(peer_ni->ibp_nid));
2987 if (ver != msg->ibm_version) {
2988 CERROR("%s replied version %x is different with "
2989 "requested version %x\n",
2990 libcfs_nid2str(peer_ni->ibp_nid), msg->ibm_version, ver);
2995 if (msg->ibm_u.connparams.ibcp_queue_depth >
2996 conn->ibc_queue_depth) {
2997 CERROR("%s has incompatible queue depth %d (<=%d wanted)\n",
2998 libcfs_nid2str(peer_ni->ibp_nid),
2999 msg->ibm_u.connparams.ibcp_queue_depth,
3000 conn->ibc_queue_depth);
3005 if (msg->ibm_u.connparams.ibcp_max_frags >
3006 conn->ibc_max_frags) {
3007 CERROR("%s has incompatible max_frags %d (<=%d wanted)\n",
3008 libcfs_nid2str(peer_ni->ibp_nid),
3009 msg->ibm_u.connparams.ibcp_max_frags,
3010 conn->ibc_max_frags);
3015 if (msg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
3016 CERROR("%s max message size %d too big (%d max)\n",
3017 libcfs_nid2str(peer_ni->ibp_nid),
3018 msg->ibm_u.connparams.ibcp_max_msg_size,
3024 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3025 if (msg->ibm_dstnid == ni->ni_nid &&
3026 msg->ibm_dststamp == net->ibn_incarnation)
3030 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3033 CERROR("Bad connection reply from %s, rc = %d, "
3034 "version: %x max_frags: %d\n",
3035 libcfs_nid2str(peer_ni->ibp_nid), rc,
3036 msg->ibm_version, msg->ibm_u.connparams.ibcp_max_frags);
3040 conn->ibc_incarnation = msg->ibm_srcstamp;
3041 conn->ibc_credits = msg->ibm_u.connparams.ibcp_queue_depth;
3042 conn->ibc_reserved_credits = msg->ibm_u.connparams.ibcp_queue_depth;
3043 conn->ibc_queue_depth = msg->ibm_u.connparams.ibcp_queue_depth;
3044 conn->ibc_max_frags = msg->ibm_u.connparams.ibcp_max_frags;
3045 LASSERT(conn->ibc_credits + conn->ibc_reserved_credits +
3046 IBLND_OOB_MSGS(ver) <= IBLND_RX_MSGS(conn));
3048 kiblnd_connreq_done(conn, 0);
3052 /* NB My QP has already established itself, so I handle anything going
3053 * wrong here by setting ibc_comms_error.
3054 * kiblnd_connreq_done(0) moves the conn state to ESTABLISHED, but then
3055 * immediately tears it down. */
3058 conn->ibc_comms_error = rc;
3059 kiblnd_connreq_done(conn, 0);
3063 kiblnd_active_connect(struct rdma_cm_id *cmid)
3065 struct kib_peer_ni *peer_ni = cmid->context;
3066 struct kib_conn *conn;
3067 struct kib_msg *msg;
3068 struct rdma_conn_param cp;
3071 unsigned long flags;
3074 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3076 incarnation = peer_ni->ibp_incarnation;
3077 version = (peer_ni->ibp_version == 0) ? IBLND_MSG_VERSION :
3078 peer_ni->ibp_version;
3080 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3082 conn = kiblnd_create_conn(peer_ni, cmid, IBLND_CONN_ACTIVE_CONNECT,
3085 kiblnd_peer_connect_failed(peer_ni, 1, -ENOMEM);
3086 kiblnd_peer_decref(peer_ni); /* lose cmid's ref */
3090 /* conn "owns" cmid now, so I return success from here on to ensure the
3091 * CM callback doesn't destroy cmid. conn also takes over cmid's ref
3094 msg = &conn->ibc_connvars->cv_msg;
3096 memset(msg, 0, sizeof(*msg));
3097 kiblnd_init_msg(msg, IBLND_MSG_CONNREQ, sizeof(msg->ibm_u.connparams));
3098 msg->ibm_u.connparams.ibcp_queue_depth = conn->ibc_queue_depth;
3099 msg->ibm_u.connparams.ibcp_max_frags = conn->ibc_max_frags;
3100 msg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
3102 kiblnd_pack_msg(peer_ni->ibp_ni, msg, version,
3103 0, peer_ni->ibp_nid, incarnation);
3105 memset(&cp, 0, sizeof(cp));
3106 cp.private_data = msg;
3107 cp.private_data_len = msg->ibm_nob;
3108 cp.responder_resources = 0; /* No atomic ops or RDMA reads */
3109 cp.initiator_depth = 0;
3110 cp.flow_control = 1;
3111 cp.retry_count = *kiblnd_tunables.kib_retry_count;
3112 cp.rnr_retry_count = *kiblnd_tunables.kib_rnr_retry_count;
3114 LASSERT(cmid->context == (void *)conn);
3115 LASSERT(conn->ibc_cmid == cmid);
3116 rc = rdma_connect_locked(cmid, &cp);
3118 CERROR("Can't connect to %s: %d\n",
3119 libcfs_nid2str(peer_ni->ibp_nid), rc);
3120 kiblnd_connreq_done(conn, rc);
3121 kiblnd_conn_decref(conn);
3128 kiblnd_cm_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
3130 struct kib_peer_ni *peer_ni;
3131 struct kib_conn *conn;
3134 switch (event->event) {
3136 CERROR("Unexpected event: %d, status: %d\n",
3137 event->event, event->status);
3140 case RDMA_CM_EVENT_CONNECT_REQUEST:
3141 /* destroy cmid on failure */
3142 rc = kiblnd_passive_connect(cmid,
3143 (void *)KIBLND_CONN_PARAM(event),
3144 KIBLND_CONN_PARAM_LEN(event));
3145 CDEBUG(D_NET, "connreq: %d\n", rc);
3148 case RDMA_CM_EVENT_ADDR_ERROR:
3149 peer_ni = cmid->context;
3150 CNETERR("%s: ADDR ERROR %d\n",
3151 libcfs_nid2str(peer_ni->ibp_nid), event->status);
3152 kiblnd_peer_connect_failed(peer_ni, 1, -EHOSTUNREACH);
3153 kiblnd_peer_decref(peer_ni);
3154 return -EHOSTUNREACH; /* rc != 0 destroys cmid */
3156 case RDMA_CM_EVENT_ADDR_RESOLVED:
3157 peer_ni = cmid->context;
3159 CDEBUG(D_NET,"%s Addr resolved: %d\n",
3160 libcfs_nid2str(peer_ni->ibp_nid), event->status);
3162 if (event->status != 0) {
3163 CNETERR("Can't resolve address for %s: %d\n",
3164 libcfs_nid2str(peer_ni->ibp_nid), event->status);
3167 rc = rdma_resolve_route(
3168 cmid, kiblnd_timeout() * 1000);
3170 struct kib_net *net = peer_ni->ibp_ni->ni_data;
3171 struct kib_dev *dev = net->ibn_dev;
3173 CDEBUG(D_NET, "%s: connection bound to "\
3175 libcfs_nid2str(peer_ni->ibp_nid),
3177 &dev->ibd_ifip, cmid->device->name);
3182 /* Can't initiate route resolution */
3183 CERROR("Can't resolve route for %s: %d\n",
3184 libcfs_nid2str(peer_ni->ibp_nid), rc);
3186 kiblnd_peer_connect_failed(peer_ni, 1, rc);
3187 kiblnd_peer_decref(peer_ni);
3188 return rc; /* rc != 0 destroys cmid */
3190 case RDMA_CM_EVENT_ROUTE_ERROR:
3191 peer_ni = cmid->context;
3192 CNETERR("%s: ROUTE ERROR %d\n",
3193 libcfs_nid2str(peer_ni->ibp_nid), event->status);
3194 kiblnd_peer_connect_failed(peer_ni, 1, -EHOSTUNREACH);
3195 kiblnd_peer_decref(peer_ni);
3196 return -EHOSTUNREACH; /* rc != 0 destroys cmid */
3198 case RDMA_CM_EVENT_ROUTE_RESOLVED:
3199 peer_ni = cmid->context;
3200 CDEBUG(D_NET,"%s Route resolved: %d\n",
3201 libcfs_nid2str(peer_ni->ibp_nid), event->status);
3203 if (event->status == 0)
3204 return kiblnd_active_connect(cmid);
3206 CNETERR("Can't resolve route for %s: %d\n",
3207 libcfs_nid2str(peer_ni->ibp_nid), event->status);
3208 kiblnd_peer_connect_failed(peer_ni, 1, event->status);
3209 kiblnd_peer_decref(peer_ni);
3210 return event->status; /* rc != 0 destroys cmid */
3212 case RDMA_CM_EVENT_UNREACHABLE:
3213 conn = cmid->context;
3214 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
3215 conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
3216 CNETERR("%s: UNREACHABLE %d\n",
3217 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
3218 kiblnd_connreq_done(conn, -ENETDOWN);
3219 kiblnd_conn_decref(conn);
3222 case RDMA_CM_EVENT_CONNECT_ERROR:
3223 conn = cmid->context;
3224 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
3225 conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
3226 CNETERR("%s: CONNECT ERROR %d\n",
3227 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
3228 kiblnd_connreq_done(conn, -ENOTCONN);
3229 kiblnd_conn_decref(conn);
3232 case RDMA_CM_EVENT_REJECTED:
3233 conn = cmid->context;
3234 switch (conn->ibc_state) {
3238 case IBLND_CONN_PASSIVE_WAIT:
3239 CERROR ("%s: REJECTED %d\n",
3240 libcfs_nid2str(conn->ibc_peer->ibp_nid),
3242 kiblnd_connreq_done(conn, -ECONNRESET);
3245 case IBLND_CONN_ACTIVE_CONNECT:
3246 kiblnd_rejected(conn, event->status,
3247 (void *)KIBLND_CONN_PARAM(event),
3248 KIBLND_CONN_PARAM_LEN(event));
3251 kiblnd_conn_decref(conn);
3254 case RDMA_CM_EVENT_ESTABLISHED:
3255 conn = cmid->context;
3256 switch (conn->ibc_state) {
3260 case IBLND_CONN_PASSIVE_WAIT:
3261 CDEBUG(D_NET, "ESTABLISHED (passive): %s\n",
3262 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3263 kiblnd_connreq_done(conn, 0);
3266 case IBLND_CONN_ACTIVE_CONNECT:
3267 CDEBUG(D_NET, "ESTABLISHED(active): %s\n",
3268 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3269 kiblnd_check_connreply(conn,
3270 (void *)KIBLND_CONN_PARAM(event),
3271 KIBLND_CONN_PARAM_LEN(event));
3274 /* net keeps its ref on conn! */
3277 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
3278 CDEBUG(D_NET, "Ignore TIMEWAIT_EXIT event\n");
3281 case RDMA_CM_EVENT_DISCONNECTED:
3282 conn = cmid->context;
3283 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
3284 CERROR("%s DISCONNECTED\n",
3285 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3286 kiblnd_connreq_done(conn, -ECONNRESET);
3288 kiblnd_close_conn(conn, 0);
3290 kiblnd_conn_decref(conn);
3291 cmid->context = NULL;
3294 case RDMA_CM_EVENT_DEVICE_REMOVAL:
3295 LCONSOLE_ERROR_MSG(0x131,
3296 "Received notification of device removal\n"
3297 "Please shutdown LNET to allow this to proceed\n");
3298 /* Can't remove network from underneath LNET for now, so I have
3302 case RDMA_CM_EVENT_ADDR_CHANGE:
3303 LCONSOLE_INFO("Physical link changed (eg hca/port)\n");
3309 kiblnd_check_txs_locked(struct kib_conn *conn, struct list_head *txs)
3313 list_for_each_entry(tx, txs, tx_list) {
3314 if (txs != &conn->ibc_active_txs) {
3315 LASSERT(tx->tx_queued);
3317 LASSERT(!tx->tx_queued);
3318 LASSERT(tx->tx_waiting || tx->tx_sending != 0);
3321 if (ktime_compare(ktime_get(), tx->tx_deadline) >= 0) {
3322 CERROR("Timed out tx: %s(WSQ:%d%d%d), %lld seconds\n",
3323 kiblnd_queue2str(conn, txs),
3324 tx->tx_waiting, tx->tx_sending, tx->tx_queued,
3326 ktime_ms_delta(ktime_get(),
3327 tx->tx_deadline) / MSEC_PER_SEC);
3336 kiblnd_conn_timed_out_locked(struct kib_conn *conn)
3338 return kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue) ||
3339 kiblnd_check_txs_locked(conn, &conn->ibc_tx_noops) ||
3340 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_rsrvd) ||
3341 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_nocred) ||
3342 kiblnd_check_txs_locked(conn, &conn->ibc_active_txs);
3346 kiblnd_check_conns (int idx)
3349 LIST_HEAD(checksends);
3350 LIST_HEAD(timedout_txs);
3351 struct hlist_head *peers = &kiblnd_data.kib_peers[idx];
3352 struct kib_peer_ni *peer_ni;
3353 struct kib_conn *conn;
3354 struct kib_tx *tx, *tx_tmp;
3355 unsigned long flags;
3357 /* NB. We expect to have a look at all the peers and not find any
3358 * RDMAs to time out, so we just use a shared lock while we
3361 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3363 hlist_for_each_entry(peer_ni, peers, ibp_list) {
3364 /* Check tx_deadline */
3365 list_for_each_entry_safe(tx, tx_tmp, &peer_ni->ibp_tx_queue, tx_list) {
3366 if (ktime_compare(ktime_get(), tx->tx_deadline) >= 0) {
3367 CWARN("Timed out tx for %s: %lld seconds\n",
3368 libcfs_nid2str(peer_ni->ibp_nid),
3369 ktime_ms_delta(ktime_get(),
3370 tx->tx_deadline) / MSEC_PER_SEC);
3371 list_move(&tx->tx_list, &timedout_txs);
3375 list_for_each_entry(conn, &peer_ni->ibp_conns, ibc_list) {
3379 LASSERT(conn->ibc_state == IBLND_CONN_ESTABLISHED);
3381 spin_lock(&conn->ibc_lock);
3383 sendnoop = kiblnd_need_noop(conn);
3384 timedout = kiblnd_conn_timed_out_locked(conn);
3385 if (!sendnoop && !timedout) {
3386 spin_unlock(&conn->ibc_lock);
3391 CERROR("Timed out RDMA with %s (%lld): c: %u, oc: %u, rc: %u\n",
3392 libcfs_nid2str(peer_ni->ibp_nid),
3394 - peer_ni->ibp_last_alive,
3396 conn->ibc_outstanding_credits,
3397 conn->ibc_reserved_credits);
3398 list_add(&conn->ibc_connd_list, &closes);
3400 list_add(&conn->ibc_connd_list, &checksends);
3402 /* +ref for 'closes' or 'checksends' */
3403 kiblnd_conn_addref(conn);
3405 spin_unlock(&conn->ibc_lock);
3409 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3411 if (!list_empty(&timedout_txs))
3412 kiblnd_txlist_done(&timedout_txs, -ETIMEDOUT,
3413 LNET_MSG_STATUS_NETWORK_TIMEOUT);
3415 /* Handle timeout by closing the whole
3416 * connection. We can only be sure RDMA activity
3417 * has ceased once the QP has been modified.
3419 while ((conn = list_first_entry_or_null(&closes,
3421 ibc_connd_list)) != NULL) {
3422 list_del(&conn->ibc_connd_list);
3423 kiblnd_close_conn(conn, -ETIMEDOUT);
3424 kiblnd_conn_decref(conn);
3427 /* In case we have enough credits to return via a
3428 * NOOP, but there were no non-blocking tx descs
3429 * free to do it last time...
3431 while ((conn = list_first_entry_or_null(&checksends,
3433 ibc_connd_list)) != NULL) {
3434 list_del(&conn->ibc_connd_list);
3436 spin_lock(&conn->ibc_lock);
3437 kiblnd_check_sends_locked(conn);
3438 spin_unlock(&conn->ibc_lock);
3440 kiblnd_conn_decref(conn);
3445 kiblnd_disconnect_conn(struct kib_conn *conn)
3447 LASSERT (!in_interrupt());
3448 LASSERT (current == kiblnd_data.kib_connd);
3449 LASSERT (conn->ibc_state == IBLND_CONN_CLOSING);
3451 rdma_disconnect(conn->ibc_cmid);
3452 kiblnd_finalise_conn(conn);
3454 kiblnd_peer_notify(conn->ibc_peer);
3458 * High-water for reconnection to the same peer_ni, reconnection attempt should
3459 * be delayed after trying more than KIB_RECONN_HIGH_RACE.
3461 #define KIB_RECONN_HIGH_RACE 10
3463 * Allow connd to take a break and handle other things after consecutive
3464 * reconnection attemps.
3466 #define KIB_RECONN_BREAK 100
3469 kiblnd_connd (void *arg)
3471 spinlock_t *lock = &kiblnd_data.kib_connd_lock;
3472 wait_queue_entry_t wait;
3473 unsigned long flags;
3474 struct kib_conn *conn;
3479 unsigned long deadline = jiffies;
3482 kiblnd_data.kib_connd = current;
3484 spin_lock_irqsave(lock, flags);
3486 while (!kiblnd_data.kib_shutdown) {
3489 dropped_lock = false;
3491 conn = list_first_entry_or_null(&kiblnd_data.kib_connd_zombies,
3492 struct kib_conn, ibc_list);
3494 struct kib_peer_ni *peer_ni = NULL;
3496 list_del(&conn->ibc_list);
3497 if (conn->ibc_reconnect) {
3498 peer_ni = conn->ibc_peer;
3499 kiblnd_peer_addref(peer_ni);
3502 spin_unlock_irqrestore(lock, flags);
3503 dropped_lock = true;
3505 kiblnd_destroy_conn(conn);
3507 spin_lock_irqsave(lock, flags);
3509 LIBCFS_FREE(conn, sizeof(*conn));
3513 conn->ibc_peer = peer_ni;
3514 if (peer_ni->ibp_reconnected < KIB_RECONN_HIGH_RACE)
3515 list_add_tail(&conn->ibc_list,
3516 &kiblnd_data.kib_reconn_list);
3518 list_add_tail(&conn->ibc_list,
3519 &kiblnd_data.kib_reconn_wait);
3522 conn = list_first_entry_or_null(&kiblnd_data.kib_connd_conns,
3523 struct kib_conn, ibc_list);
3527 list_del(&conn->ibc_list);
3529 spin_unlock_irqrestore(lock, flags);
3530 dropped_lock = true;
3532 kiblnd_disconnect_conn(conn);
3533 wait = conn->ibc_waits;
3534 if (wait == 0) /* keep ref for connd_wait, see below */
3535 kiblnd_conn_decref(conn);
3537 spin_lock_irqsave(lock, flags);
3540 list_add_tail(&conn->ibc_list,
3541 &kiblnd_data.kib_connd_waits);
3544 while (reconn < KIB_RECONN_BREAK) {
3545 if (kiblnd_data.kib_reconn_sec !=
3546 ktime_get_real_seconds()) {
3547 kiblnd_data.kib_reconn_sec = ktime_get_real_seconds();
3548 list_splice_init(&kiblnd_data.kib_reconn_wait,
3549 &kiblnd_data.kib_reconn_list);
3552 conn = list_first_entry_or_null(&kiblnd_data.kib_reconn_list,
3553 struct kib_conn, ibc_list);
3557 list_del(&conn->ibc_list);
3559 spin_unlock_irqrestore(lock, flags);
3560 dropped_lock = true;
3562 reconn += kiblnd_reconnect_peer(conn->ibc_peer);
3563 kiblnd_peer_decref(conn->ibc_peer);
3564 LIBCFS_FREE(conn, sizeof(*conn));
3566 spin_lock_irqsave(lock, flags);
3569 conn = list_first_entry_or_null(&kiblnd_data.kib_connd_waits,
3573 list_del(&conn->ibc_list);
3574 spin_unlock_irqrestore(lock, flags);
3576 dropped_lock = kiblnd_tx_may_discard(conn);
3578 kiblnd_conn_decref(conn);
3580 spin_lock_irqsave(lock, flags);
3582 list_add_tail(&conn->ibc_list,
3583 &kiblnd_data.kib_connd_waits);
3586 /* careful with the jiffy wrap... */
3587 timeout = (int)(deadline - jiffies);
3591 int chunk = HASH_SIZE(kiblnd_data.kib_peers);
3592 unsigned int lnd_timeout;
3594 spin_unlock_irqrestore(lock, flags);
3595 dropped_lock = true;
3597 /* Time to check for RDMA timeouts on a few more
3598 * peers: I do checks every 'p' seconds on a
3599 * proportion of the peer_ni table and I need to check
3600 * every connection 'n' times within a timeout
3601 * interval, to ensure I detect a timeout on any
3602 * connection within (n+1)/n times the timeout
3606 lnd_timeout = kiblnd_timeout();
3607 if (lnd_timeout > n * p)
3608 chunk = (chunk * n * p) / lnd_timeout;
3612 for (i = 0; i < chunk; i++) {
3613 kiblnd_check_conns(peer_index);
3614 peer_index = (peer_index + 1) %
3615 HASH_SIZE(kiblnd_data.kib_peers);
3618 deadline += cfs_time_seconds(p);
3619 spin_lock_irqsave(lock, flags);
3625 /* Nothing to do for 'timeout' */
3626 set_current_state(TASK_INTERRUPTIBLE);
3627 add_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3628 spin_unlock_irqrestore(lock, flags);
3630 schedule_timeout(timeout);
3632 remove_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3633 spin_lock_irqsave(lock, flags);
3636 spin_unlock_irqrestore(lock, flags);
3638 kiblnd_thread_fini();
3643 kiblnd_qp_event(struct ib_event *event, void *arg)
3645 struct kib_conn *conn = arg;
3647 switch (event->event) {
3648 case IB_EVENT_COMM_EST:
3649 CDEBUG(D_NET, "%s established\n",
3650 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3651 /* We received a packet but connection isn't established
3652 * probably handshake packet was lost, so free to
3653 * force make connection established */
3654 rdma_notify(conn->ibc_cmid, IB_EVENT_COMM_EST);
3657 case IB_EVENT_PORT_ERR:
3658 case IB_EVENT_DEVICE_FATAL:
3659 CERROR("Fatal device error for NI %s\n",
3660 libcfs_nid2str(conn->ibc_peer->ibp_ni->ni_nid));
3661 atomic_set(&conn->ibc_peer->ibp_ni->ni_fatal_error_on, 1);
3664 case IB_EVENT_PORT_ACTIVE:
3665 CERROR("Port reactivated for NI %s\n",
3666 libcfs_nid2str(conn->ibc_peer->ibp_ni->ni_nid));
3667 atomic_set(&conn->ibc_peer->ibp_ni->ni_fatal_error_on, 0);
3671 CERROR("%s: Async QP event type %d\n",
3672 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3678 kiblnd_complete (struct ib_wc *wc)
3680 switch (kiblnd_wreqid2type(wc->wr_id)) {
3685 if (wc->status != IB_WC_SUCCESS &&
3686 wc->status != IB_WC_WR_FLUSH_ERR)
3687 CNETERR("FastReg failed: %d\n", wc->status);
3690 case IBLND_WID_RDMA:
3691 /* We only get RDMA completion notification if it fails. All
3692 * subsequent work items, including the final SEND will fail
3693 * too. However we can't print out any more info about the
3694 * failing RDMA because 'tx' might be back on the idle list or
3695 * even reused already if we didn't manage to post all our work
3697 CNETERR("RDMA (tx: %p) failed: %d\n",
3698 kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3702 kiblnd_tx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3706 kiblnd_rx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status,
3713 kiblnd_cq_completion(struct ib_cq *cq, void *arg)
3715 /* NB I'm not allowed to schedule this conn once its refcount has
3716 * reached 0. Since fundamentally I'm racing with scheduler threads
3717 * consuming my CQ I could be called after all completions have
3718 * occurred. But in this case, ibc_nrx == 0 && ibc_nsends_posted == 0
3719 * and this CQ is about to be destroyed so I NOOP. */
3720 struct kib_conn *conn = arg;
3721 struct kib_sched_info *sched = conn->ibc_sched;
3722 unsigned long flags;
3724 LASSERT(cq == conn->ibc_cq);
3726 spin_lock_irqsave(&sched->ibs_lock, flags);
3728 conn->ibc_ready = 1;
3730 if (!conn->ibc_scheduled &&
3731 (conn->ibc_nrx > 0 ||
3732 conn->ibc_nsends_posted > 0)) {
3733 kiblnd_conn_addref(conn); /* +1 ref for sched_conns */
3734 conn->ibc_scheduled = 1;
3735 list_add_tail(&conn->ibc_sched_list, &sched->ibs_conns);
3737 if (waitqueue_active(&sched->ibs_waitq))
3738 wake_up(&sched->ibs_waitq);
3741 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3745 kiblnd_cq_event(struct ib_event *event, void *arg)
3747 struct kib_conn *conn = arg;
3749 CERROR("%s: async CQ event type %d\n",
3750 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3754 kiblnd_scheduler(void *arg)
3756 long id = (long)arg;
3757 struct kib_sched_info *sched;
3758 struct kib_conn *conn;
3759 wait_queue_entry_t wait;
3760 unsigned long flags;
3767 sched = kiblnd_data.kib_scheds[KIB_THREAD_CPT(id)];
3769 rc = cfs_cpt_bind(lnet_cpt_table(), sched->ibs_cpt);
3771 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);
3774 spin_lock_irqsave(&sched->ibs_lock, flags);
3776 while (!kiblnd_data.kib_shutdown) {
3777 if (need_resched()) {
3778 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3782 spin_lock_irqsave(&sched->ibs_lock, flags);
3785 did_something = false;
3787 conn = list_first_entry_or_null(&sched->ibs_conns,
3791 /* take over kib_sched_conns' ref on conn... */
3792 LASSERT(conn->ibc_scheduled);
3793 list_del(&conn->ibc_sched_list);
3794 conn->ibc_ready = 0;
3796 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3798 wc.wr_id = IBLND_WID_INVAL;
3800 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3802 rc = ib_req_notify_cq(conn->ibc_cq,
3805 CWARN("%s: ib_req_notify_cq failed: %d, closing connection\n",
3806 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
3807 kiblnd_close_conn(conn, -EIO);
3808 kiblnd_conn_decref(conn);
3809 spin_lock_irqsave(&sched->ibs_lock,
3814 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3817 if (unlikely(rc > 0 && wc.wr_id == IBLND_WID_INVAL)) {
3819 "ib_poll_cq (rc: %d) returned invalid "
3820 "wr_id, opcode %d, status: %d, "
3821 "vendor_err: %d, conn: %s status: %d\n"
3822 "please upgrade firmware and OFED or "
3823 "contact vendor.\n", rc,
3824 wc.opcode, wc.status, wc.vendor_err,
3825 libcfs_nid2str(conn->ibc_peer->ibp_nid),
3831 CWARN("%s: ib_poll_cq failed: %d, closing connection\n",
3832 libcfs_nid2str(conn->ibc_peer->ibp_nid),
3834 kiblnd_close_conn(conn, -EIO);
3835 kiblnd_conn_decref(conn);
3836 spin_lock_irqsave(&sched->ibs_lock, flags);
3840 spin_lock_irqsave(&sched->ibs_lock, flags);
3842 if (rc != 0 || conn->ibc_ready) {
3843 /* There may be another completion waiting; get
3844 * another scheduler to check while I handle
3846 /* +1 ref for sched_conns */
3847 kiblnd_conn_addref(conn);
3848 list_add_tail(&conn->ibc_sched_list,
3850 if (waitqueue_active(&sched->ibs_waitq))
3851 wake_up(&sched->ibs_waitq);
3853 conn->ibc_scheduled = 0;
3857 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3858 kiblnd_complete(&wc);
3860 spin_lock_irqsave(&sched->ibs_lock, flags);
3863 kiblnd_conn_decref(conn); /* ..drop my ref from above */
3864 did_something = true;
3870 set_current_state(TASK_INTERRUPTIBLE);
3871 add_wait_queue_exclusive(&sched->ibs_waitq, &wait);
3872 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3876 remove_wait_queue(&sched->ibs_waitq, &wait);
3877 set_current_state(TASK_RUNNING);
3878 spin_lock_irqsave(&sched->ibs_lock, flags);
3881 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3883 kiblnd_thread_fini();
3888 kiblnd_failover_thread(void *arg)
3890 rwlock_t *glock = &kiblnd_data.kib_global_lock;
3891 struct kib_dev *dev;
3892 struct net *ns = arg;
3893 wait_queue_entry_t wait;
3894 unsigned long flags;
3897 LASSERT(*kiblnd_tunables.kib_dev_failover != 0);
3900 write_lock_irqsave(glock, flags);
3902 while (!kiblnd_data.kib_shutdown) {
3903 bool do_failover = false;
3906 list_for_each_entry(dev, &kiblnd_data.kib_failed_devs,
3908 if (ktime_get_seconds() < dev->ibd_next_failover)
3915 list_del_init(&dev->ibd_fail_list);
3916 dev->ibd_failover = 1;
3917 write_unlock_irqrestore(glock, flags);
3919 rc = kiblnd_dev_failover(dev, ns);
3921 write_lock_irqsave(glock, flags);
3923 LASSERT(dev->ibd_failover);
3924 dev->ibd_failover = 0;
3925 if (rc >= 0) { /* Device is OK or failover succeed */
3926 dev->ibd_next_failover = ktime_get_seconds() + 3;
3930 /* failed to failover, retry later */
3931 dev->ibd_next_failover = ktime_get_seconds() +
3932 min(dev->ibd_failed_failover, 10);
3933 if (kiblnd_dev_can_failover(dev)) {
3934 list_add_tail(&dev->ibd_fail_list,
3935 &kiblnd_data.kib_failed_devs);
3941 /* long sleep if no more pending failover */
3942 long_sleep = list_empty(&kiblnd_data.kib_failed_devs);
3944 set_current_state(TASK_INTERRUPTIBLE);
3945 add_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3946 write_unlock_irqrestore(glock, flags);
3948 rc = schedule_timeout(long_sleep ? cfs_time_seconds(10) :
3949 cfs_time_seconds(1));
3950 set_current_state(TASK_RUNNING);
3951 remove_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3952 write_lock_irqsave(glock, flags);
3954 if (!long_sleep || rc != 0)
3957 /* have a long sleep, routine check all active devices,
3958 * we need checking like this because if there is not active
3959 * connection on the dev and no SEND from local, we may listen
3960 * on wrong HCA for ever while there is a bonding failover
3962 list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3963 if (kiblnd_dev_can_failover(dev)) {
3964 list_add_tail(&dev->ibd_fail_list,
3965 &kiblnd_data.kib_failed_devs);
3970 write_unlock_irqrestore(glock, flags);
3972 kiblnd_thread_fini();