4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lnet/klnds/o2iblnd/o2iblnd_cb.c
34 * Author: Eric Barton <eric@bartonsoftware.com>
39 #define MAX_CONN_RACES_BEFORE_ABORT 20
41 static void kiblnd_peer_alive(struct kib_peer_ni *peer_ni);
42 static void kiblnd_peer_connect_failed(struct kib_peer_ni *peer_ni, int active,
44 static void kiblnd_init_tx_msg(struct lnet_ni *ni, struct kib_tx *tx,
45 int type, int body_nob);
46 static int kiblnd_init_rdma(struct kib_conn *conn, struct kib_tx *tx, int type,
47 int resid, struct kib_rdma_desc *dstrd, u64 dstcookie);
48 static void kiblnd_queue_tx_locked(struct kib_tx *tx, struct kib_conn *conn);
49 static void kiblnd_queue_tx(struct kib_tx *tx, struct kib_conn *conn);
51 static void kiblnd_unmap_tx(struct kib_tx *tx);
52 static void kiblnd_check_sends_locked(struct kib_conn *conn);
55 kiblnd_tx_done(struct kib_tx *tx)
57 struct lnet_msg *lntmsg[2];
61 LASSERT (!in_interrupt());
62 LASSERT (!tx->tx_queued); /* mustn't be queued for sending */
63 LASSERT (tx->tx_sending == 0); /* mustn't be awaiting sent callback */
64 LASSERT (!tx->tx_waiting); /* mustn't be awaiting peer_ni response */
65 LASSERT (tx->tx_pool != NULL);
69 /* tx may have up to 2 lnet msgs to finalise */
70 lntmsg[0] = tx->tx_lntmsg[0]; tx->tx_lntmsg[0] = NULL;
71 lntmsg[1] = tx->tx_lntmsg[1]; tx->tx_lntmsg[1] = NULL;
74 if (tx->tx_conn != NULL) {
75 kiblnd_conn_decref(tx->tx_conn);
79 tx->tx_nwrq = tx->tx_nsge = 0;
82 kiblnd_pool_free_node(&tx->tx_pool->tpo_pool, &tx->tx_list);
84 /* delay finalize until my descs have been freed */
85 for (i = 0; i < 2; i++) {
86 if (lntmsg[i] == NULL)
89 /* propagate health status to LNet for requests */
90 if (i == 0 && lntmsg[i])
91 lntmsg[i]->msg_health_status = tx->tx_hstatus;
93 lnet_finalize(lntmsg[i], rc);
98 kiblnd_txlist_done(struct list_head *txlist, int status,
99 enum lnet_msg_hstatus hstatus)
103 while (!list_empty(txlist)) {
104 tx = list_entry(txlist->next, struct kib_tx, tx_list);
106 list_del(&tx->tx_list);
109 tx->tx_status = status;
110 if (hstatus != LNET_MSG_STATUS_OK)
111 tx->tx_hstatus = hstatus;
116 static struct kib_tx *
117 kiblnd_get_idle_tx(struct lnet_ni *ni, lnet_nid_t target)
119 struct kib_net *net = ni->ni_data;
120 struct list_head *node;
122 struct kib_tx_poolset *tps;
124 tps = net->ibn_tx_ps[lnet_cpt_of_nid(target, ni)];
125 node = kiblnd_pool_alloc_node(&tps->tps_poolset);
128 tx = container_of(node, struct kib_tx, tx_list);
130 LASSERT (tx->tx_nwrq == 0);
131 LASSERT (!tx->tx_queued);
132 LASSERT (tx->tx_sending == 0);
133 LASSERT (!tx->tx_waiting);
134 LASSERT (tx->tx_status == 0);
135 LASSERT (tx->tx_conn == NULL);
136 LASSERT (tx->tx_lntmsg[0] == NULL);
137 LASSERT (tx->tx_lntmsg[1] == NULL);
138 LASSERT (tx->tx_nfrags == 0);
141 tx->tx_hstatus = LNET_MSG_STATUS_OK;
147 kiblnd_drop_rx(struct kib_rx *rx)
149 struct kib_conn *conn = rx->rx_conn;
150 struct kib_sched_info *sched = conn->ibc_sched;
153 spin_lock_irqsave(&sched->ibs_lock, flags);
154 LASSERT(conn->ibc_nrx > 0);
156 spin_unlock_irqrestore(&sched->ibs_lock, flags);
158 kiblnd_conn_decref(conn);
162 kiblnd_post_rx(struct kib_rx *rx, int credit)
164 struct kib_conn *conn = rx->rx_conn;
165 struct kib_net *net = conn->ibc_peer->ibp_ni->ni_data;
166 struct ib_recv_wr *bad_wrq = NULL;
167 #ifdef HAVE_IB_GET_DMA_MR
168 struct ib_mr *mr = conn->ibc_hdev->ibh_mrs;
172 LASSERT (net != NULL);
173 LASSERT (!in_interrupt());
174 LASSERT (credit == IBLND_POSTRX_NO_CREDIT ||
175 credit == IBLND_POSTRX_PEER_CREDIT ||
176 credit == IBLND_POSTRX_RSRVD_CREDIT);
177 #ifdef HAVE_IB_GET_DMA_MR
180 rx->rx_sge.lkey = mr->lkey;
182 rx->rx_sge.lkey = conn->ibc_hdev->ibh_pd->local_dma_lkey;
184 rx->rx_sge.addr = rx->rx_msgaddr;
185 rx->rx_sge.length = IBLND_MSG_SIZE;
187 rx->rx_wrq.next = NULL;
188 rx->rx_wrq.sg_list = &rx->rx_sge;
189 rx->rx_wrq.num_sge = 1;
190 rx->rx_wrq.wr_id = kiblnd_ptr2wreqid(rx, IBLND_WID_RX);
192 LASSERT (conn->ibc_state >= IBLND_CONN_INIT);
193 LASSERT (rx->rx_nob >= 0); /* not posted */
195 if (conn->ibc_state > IBLND_CONN_ESTABLISHED) {
196 kiblnd_drop_rx(rx); /* No more posts for this rx */
200 rx->rx_nob = -1; /* flag posted */
202 /* NB: need an extra reference after ib_post_recv because we don't
203 * own this rx (and rx::rx_conn) anymore, LU-5678.
205 kiblnd_conn_addref(conn);
206 #ifdef HAVE_IB_POST_SEND_RECV_CONST
207 rc = ib_post_recv(conn->ibc_cmid->qp, &rx->rx_wrq,
208 (const struct ib_recv_wr **)&bad_wrq);
210 rc = ib_post_recv(conn->ibc_cmid->qp, &rx->rx_wrq, &bad_wrq);
212 if (unlikely(rc != 0)) {
213 CERROR("Can't post rx for %s: %d, bad_wrq: %p\n",
214 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc, bad_wrq);
218 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) /* Initial post */
221 if (unlikely(rc != 0)) {
222 kiblnd_close_conn(conn, rc);
223 kiblnd_drop_rx(rx); /* No more posts for this rx */
227 if (credit == IBLND_POSTRX_NO_CREDIT)
230 spin_lock(&conn->ibc_lock);
231 if (credit == IBLND_POSTRX_PEER_CREDIT)
232 conn->ibc_outstanding_credits++;
234 conn->ibc_reserved_credits++;
235 kiblnd_check_sends_locked(conn);
236 spin_unlock(&conn->ibc_lock);
239 kiblnd_conn_decref(conn);
243 static struct kib_tx *
244 kiblnd_find_waiting_tx_locked(struct kib_conn *conn, int txtype, u64 cookie)
246 struct list_head *tmp;
248 list_for_each(tmp, &conn->ibc_active_txs) {
249 struct kib_tx *tx = list_entry(tmp, struct kib_tx, tx_list);
251 LASSERT(!tx->tx_queued);
252 LASSERT(tx->tx_sending != 0 || tx->tx_waiting);
254 if (tx->tx_cookie != cookie)
257 if (tx->tx_waiting &&
258 tx->tx_msg->ibm_type == txtype)
261 CWARN("Bad completion: %swaiting, type %x (wanted %x)\n",
262 tx->tx_waiting ? "" : "NOT ",
263 tx->tx_msg->ibm_type, txtype);
269 kiblnd_handle_completion(struct kib_conn *conn, int txtype, int status, u64 cookie)
272 struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
275 spin_lock(&conn->ibc_lock);
277 tx = kiblnd_find_waiting_tx_locked(conn, txtype, cookie);
279 spin_unlock(&conn->ibc_lock);
281 CWARN("Unmatched completion type %x cookie %#llx from %s\n",
282 txtype, cookie, libcfs_nid2str(conn->ibc_peer->ibp_nid));
283 kiblnd_close_conn(conn, -EPROTO);
287 if (tx->tx_status == 0) { /* success so far */
288 if (status < 0) { /* failed? */
289 tx->tx_status = status;
290 tx->tx_hstatus = LNET_MSG_STATUS_REMOTE_ERROR;
291 } else if (txtype == IBLND_MSG_GET_REQ) {
292 lnet_set_reply_msg_len(ni, tx->tx_lntmsg[1], status);
298 idle = !tx->tx_queued && (tx->tx_sending == 0);
300 list_del(&tx->tx_list);
302 spin_unlock(&conn->ibc_lock);
309 kiblnd_send_completion(struct kib_conn *conn, int type, int status, u64 cookie)
311 struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
312 struct kib_tx *tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
315 CERROR("Can't get tx for completion %x for %s\n",
316 type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
320 tx->tx_msg->ibm_u.completion.ibcm_status = status;
321 tx->tx_msg->ibm_u.completion.ibcm_cookie = cookie;
322 kiblnd_init_tx_msg(ni, tx, type, sizeof(struct kib_completion_msg));
324 kiblnd_queue_tx(tx, conn);
328 kiblnd_handle_rx(struct kib_rx *rx)
330 struct kib_msg *msg = rx->rx_msg;
331 struct kib_conn *conn = rx->rx_conn;
332 struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
333 int credits = msg->ibm_credits;
339 LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
341 CDEBUG (D_NET, "Received %x[%d] from %s\n",
342 msg->ibm_type, credits,
343 libcfs_nid2str(conn->ibc_peer->ibp_nid));
346 /* Have I received credits that will let me send? */
347 spin_lock(&conn->ibc_lock);
349 if (conn->ibc_credits + credits >
350 conn->ibc_queue_depth) {
351 rc2 = conn->ibc_credits;
352 spin_unlock(&conn->ibc_lock);
354 CERROR("Bad credits from %s: %d + %d > %d\n",
355 libcfs_nid2str(conn->ibc_peer->ibp_nid),
357 conn->ibc_queue_depth);
359 kiblnd_close_conn(conn, -EPROTO);
360 kiblnd_post_rx(rx, IBLND_POSTRX_NO_CREDIT);
364 conn->ibc_credits += credits;
366 /* This ensures the credit taken by NOOP can be returned */
367 if (msg->ibm_type == IBLND_MSG_NOOP &&
368 !IBLND_OOB_CAPABLE(conn->ibc_version)) /* v1 only */
369 conn->ibc_outstanding_credits++;
371 kiblnd_check_sends_locked(conn);
372 spin_unlock(&conn->ibc_lock);
375 switch (msg->ibm_type) {
377 CERROR("Bad IBLND message type %x from %s\n",
378 msg->ibm_type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
379 post_credit = IBLND_POSTRX_NO_CREDIT;
384 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
385 post_credit = IBLND_POSTRX_NO_CREDIT;
389 if (credits != 0) /* credit already posted */
390 post_credit = IBLND_POSTRX_NO_CREDIT;
391 else /* a keepalive NOOP */
392 post_credit = IBLND_POSTRX_PEER_CREDIT;
395 case IBLND_MSG_IMMEDIATE:
396 post_credit = IBLND_POSTRX_DONT_POST;
397 rc = lnet_parse(ni, &msg->ibm_u.immediate.ibim_hdr,
398 msg->ibm_srcnid, rx, 0);
399 if (rc < 0) /* repost on error */
400 post_credit = IBLND_POSTRX_PEER_CREDIT;
403 case IBLND_MSG_PUT_REQ:
404 post_credit = IBLND_POSTRX_DONT_POST;
405 rc = lnet_parse(ni, &msg->ibm_u.putreq.ibprm_hdr,
406 msg->ibm_srcnid, rx, 1);
407 if (rc < 0) /* repost on error */
408 post_credit = IBLND_POSTRX_PEER_CREDIT;
411 case IBLND_MSG_PUT_NAK:
412 CWARN ("PUT_NACK from %s\n",
413 libcfs_nid2str(conn->ibc_peer->ibp_nid));
414 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
415 kiblnd_handle_completion(conn, IBLND_MSG_PUT_REQ,
416 msg->ibm_u.completion.ibcm_status,
417 msg->ibm_u.completion.ibcm_cookie);
420 case IBLND_MSG_PUT_ACK:
421 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
423 spin_lock(&conn->ibc_lock);
424 tx = kiblnd_find_waiting_tx_locked(conn, IBLND_MSG_PUT_REQ,
425 msg->ibm_u.putack.ibpam_src_cookie);
427 list_del(&tx->tx_list);
428 spin_unlock(&conn->ibc_lock);
431 CERROR("Unmatched PUT_ACK from %s\n",
432 libcfs_nid2str(conn->ibc_peer->ibp_nid));
437 LASSERT (tx->tx_waiting);
438 /* CAVEAT EMPTOR: I could be racing with tx_complete, but...
439 * (a) I can overwrite tx_msg since my peer_ni has received it!
440 * (b) tx_waiting set tells tx_complete() it's not done. */
442 tx->tx_nwrq = tx->tx_nsge = 0; /* overwrite PUT_REQ */
444 rc2 = kiblnd_init_rdma(conn, tx, IBLND_MSG_PUT_DONE,
445 kiblnd_rd_size(&msg->ibm_u.putack.ibpam_rd),
446 &msg->ibm_u.putack.ibpam_rd,
447 msg->ibm_u.putack.ibpam_dst_cookie);
449 CERROR("Can't setup rdma for PUT to %s: %d\n",
450 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc2);
452 spin_lock(&conn->ibc_lock);
453 tx->tx_waiting = 0; /* clear waiting and queue atomically */
454 kiblnd_queue_tx_locked(tx, conn);
455 spin_unlock(&conn->ibc_lock);
458 case IBLND_MSG_PUT_DONE:
459 post_credit = IBLND_POSTRX_PEER_CREDIT;
460 kiblnd_handle_completion(conn, IBLND_MSG_PUT_ACK,
461 msg->ibm_u.completion.ibcm_status,
462 msg->ibm_u.completion.ibcm_cookie);
465 case IBLND_MSG_GET_REQ:
466 post_credit = IBLND_POSTRX_DONT_POST;
467 rc = lnet_parse(ni, &msg->ibm_u.get.ibgm_hdr,
468 msg->ibm_srcnid, rx, 1);
469 if (rc < 0) /* repost on error */
470 post_credit = IBLND_POSTRX_PEER_CREDIT;
473 case IBLND_MSG_GET_DONE:
474 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
475 kiblnd_handle_completion(conn, IBLND_MSG_GET_REQ,
476 msg->ibm_u.completion.ibcm_status,
477 msg->ibm_u.completion.ibcm_cookie);
481 if (rc < 0) /* protocol error */
482 kiblnd_close_conn(conn, rc);
484 if (post_credit != IBLND_POSTRX_DONT_POST)
485 kiblnd_post_rx(rx, post_credit);
489 kiblnd_rx_complete(struct kib_rx *rx, int status, int nob)
491 struct kib_msg *msg = rx->rx_msg;
492 struct kib_conn *conn = rx->rx_conn;
493 struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
494 struct kib_net *net = ni->ni_data;
498 LASSERT (net != NULL);
499 LASSERT (rx->rx_nob < 0); /* was posted */
500 rx->rx_nob = 0; /* isn't now */
502 if (conn->ibc_state > IBLND_CONN_ESTABLISHED)
505 if (status != IB_WC_SUCCESS) {
506 CNETERR("Rx from %s failed: %d\n",
507 libcfs_nid2str(conn->ibc_peer->ibp_nid), status);
514 rc = kiblnd_unpack_msg(msg, rx->rx_nob);
516 CERROR ("Error %d unpacking rx from %s\n",
517 rc, libcfs_nid2str(conn->ibc_peer->ibp_nid));
521 if (msg->ibm_srcnid != conn->ibc_peer->ibp_nid ||
522 msg->ibm_dstnid != ni->ni_nid ||
523 msg->ibm_srcstamp != conn->ibc_incarnation ||
524 msg->ibm_dststamp != net->ibn_incarnation) {
525 CERROR ("Stale rx from %s\n",
526 libcfs_nid2str(conn->ibc_peer->ibp_nid));
531 /* set time last known alive */
532 kiblnd_peer_alive(conn->ibc_peer);
534 /* racing with connection establishment/teardown! */
536 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
537 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
540 write_lock_irqsave(g_lock, flags);
541 /* must check holding global lock to eliminate race */
542 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
543 list_add_tail(&rx->rx_list, &conn->ibc_early_rxs);
544 write_unlock_irqrestore(g_lock, flags);
547 write_unlock_irqrestore(g_lock, flags);
549 kiblnd_handle_rx(rx);
553 CDEBUG(D_NET, "rx %p conn %p\n", rx, conn);
554 kiblnd_close_conn(conn, err);
556 kiblnd_drop_rx(rx); /* Don't re-post rx. */
560 kiblnd_fmr_map_tx(struct kib_net *net, struct kib_tx *tx,
561 struct kib_rdma_desc *rd, u32 nob)
563 struct kib_hca_dev *hdev;
565 struct kib_fmr_poolset *fps;
570 LASSERT(tx->tx_pool != NULL);
571 LASSERT(tx->tx_pool->tpo_pool.po_owner != NULL);
574 hdev = tx->tx_pool->tpo_hdev;
575 cpt = tx->tx_pool->tpo_pool.po_owner->ps_cpt;
578 * If we're dealing with FastReg, but the device doesn't
579 * support GAPS and the tx has GAPS, then there is no real point
580 * in trying to map the memory, because it'll just fail. So
581 * preemptively fail with an appropriate message
583 if ((dev->ibd_dev_caps & IBLND_DEV_CAPS_FASTREG_ENABLED) &&
584 !(dev->ibd_dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT) &&
586 CERROR("Using FastReg with no GAPS support, but tx has gaps. "
587 "Try setting use_fastreg_gaps to 1\n");
588 return -EPROTONOSUPPORT;
592 * FMR does not support gaps but the tx has gaps then
593 * we should make sure that the number of fragments we'll be sending
594 * over fits within the number of fragments negotiated on the
595 * connection, otherwise, we won't be able to RDMA the data.
596 * We need to maintain the number of fragments negotiation on the
597 * connection for backwards compatibility.
599 if (tx->tx_gaps && (dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED)) {
601 tx->tx_conn->ibc_max_frags <= rd->rd_nfrags) {
602 CERROR("TX number of frags (%d) is <= than connection"
603 " number of frags (%d). Consider setting peer's"
604 " map_on_demand to 256\n", tx->tx_nfrags,
605 tx->tx_conn->ibc_max_frags);
610 fps = net->ibn_fmr_ps[cpt];
611 rc = kiblnd_fmr_pool_map(fps, tx, rd, nob, 0, &tx->tx_fmr);
613 CERROR("Can't map %u pages: %d\n", nob, 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.
626 if (((dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED) && !tx->tx_gaps) ||
627 (dev->ibd_dev_caps & IBLND_DEV_CAPS_FASTREG_ENABLED)) {
628 /* FMR requires zero based address */
629 if (dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED)
630 rd->rd_frags[0].rf_addr &= ~hdev->ibh_page_mask;
631 rd->rd_frags[0].rf_nob = nob;
635 * We're transmitting with gaps using FMR.
636 * We'll need to use multiple fragments and identify the
637 * zero based address of each fragment.
639 for (i = 0; i < rd->rd_nfrags; i++) {
640 rd->rd_frags[i].rf_addr &= ~hdev->ibh_page_mask;
641 rd->rd_frags[i].rf_addr += i << hdev->ibh_page_shift;
649 kiblnd_unmap_tx(struct kib_tx *tx)
651 if (tx->tx_fmr.fmr_pfmr || tx->tx_fmr.fmr_frd)
652 kiblnd_fmr_pool_unmap(&tx->tx_fmr, tx->tx_status);
654 if (tx->tx_nfrags != 0) {
655 kiblnd_dma_unmap_sg(tx->tx_pool->tpo_hdev->ibh_ibdev,
656 tx->tx_frags, tx->tx_nfrags, tx->tx_dmadir);
661 #ifdef HAVE_IB_GET_DMA_MR
662 static struct ib_mr *
663 kiblnd_find_rd_dma_mr(struct lnet_ni *ni, struct kib_rdma_desc *rd)
665 struct kib_net *net = ni->ni_data;
666 struct kib_hca_dev *hdev = net->ibn_dev->ibd_hdev;
667 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
669 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
672 * if map-on-demand is turned on and the device supports
673 * either FMR or FastReg then use that. Otherwise use global
674 * memory regions. If that's not available either, then you're
675 * dead in the water and fail the operation.
677 if (tunables->lnd_map_on_demand &&
678 (net->ibn_dev->ibd_dev_caps & IBLND_DEV_CAPS_FASTREG_ENABLED ||
679 net->ibn_dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED))
683 * hdev->ibh_mrs can be NULL. This case is dealt with gracefully
684 * in the call chain. The mapping will fail with appropriate error
687 return hdev->ibh_mrs;
691 static int kiblnd_map_tx(struct lnet_ni *ni, struct kib_tx *tx,
692 struct kib_rdma_desc *rd, int nfrags)
694 struct kib_net *net = ni->ni_data;
695 struct kib_hca_dev *hdev = net->ibn_dev->ibd_hdev;
696 #ifdef HAVE_IB_GET_DMA_MR
697 struct ib_mr *mr = NULL;
702 /* If rd is not tx_rd, it's going to get sent to a peer_ni and I'm the
704 tx->tx_dmadir = (rd != tx->tx_rd) ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
705 tx->tx_nfrags = nfrags;
707 rd->rd_nfrags = kiblnd_dma_map_sg(hdev->ibh_ibdev, tx->tx_frags,
708 tx->tx_nfrags, tx->tx_dmadir);
710 for (i = 0, nob = 0; i < rd->rd_nfrags; i++) {
711 rd->rd_frags[i].rf_nob = kiblnd_sg_dma_len(
712 hdev->ibh_ibdev, &tx->tx_frags[i]);
713 rd->rd_frags[i].rf_addr = kiblnd_sg_dma_address(
714 hdev->ibh_ibdev, &tx->tx_frags[i]);
715 nob += rd->rd_frags[i].rf_nob;
718 #ifdef HAVE_IB_GET_DMA_MR
719 mr = kiblnd_find_rd_dma_mr(ni, rd);
721 /* found pre-mapping MR */
722 rd->rd_key = (rd != tx->tx_rd) ? mr->rkey : mr->lkey;
727 if (net->ibn_fmr_ps != NULL)
728 return kiblnd_fmr_map_tx(net, tx, rd, nob);
733 static int kiblnd_setup_rd_kiov(struct lnet_ni *ni, struct kib_tx *tx,
734 struct kib_rdma_desc *rd, int nkiov,
735 struct bio_vec *kiov, int offset, int nob)
737 struct kib_net *net = ni->ni_data;
738 struct scatterlist *sg;
742 CDEBUG(D_NET, "niov %d offset %d nob %d\n", nkiov, offset, nob);
746 LASSERT(net != NULL);
748 while (offset >= kiov->bv_len) {
749 offset -= kiov->bv_len;
761 fragnob = min((int)(kiov->bv_len - offset), nob);
764 * We're allowed to start at a non-aligned page offset in
765 * the first fragment and end at a non-aligned page offset
766 * in the last fragment.
768 if ((fragnob < (int)(kiov->bv_len - offset)) &&
769 nkiov < max_nkiov && nob > fragnob) {
770 CDEBUG(D_NET, "fragnob %d < available page %d: with"
771 " remaining %d kiovs with %d nob left\n",
772 fragnob, (int)(kiov->bv_len - offset),
777 sg_set_page(sg, kiov->bv_page, fragnob,
778 kiov->bv_offset + offset);
781 CERROR("lacking enough sg entries to map tx\n");
791 return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags);
795 kiblnd_post_tx_locked(struct kib_conn *conn, struct kib_tx *tx, int credit)
796 __must_hold(&conn->ibc_lock)
798 struct kib_msg *msg = tx->tx_msg;
799 struct kib_peer_ni *peer_ni = conn->ibc_peer;
800 struct lnet_ni *ni = peer_ni->ibp_ni;
801 int ver = conn->ibc_version;
805 LASSERT(tx->tx_queued);
806 /* We rely on this for QP sizing */
807 LASSERT(tx->tx_nwrq > 0 && tx->tx_nsge >= 0);
808 LASSERT(tx->tx_nwrq <= 1 + conn->ibc_max_frags);
810 LASSERT(credit == 0 || credit == 1);
811 LASSERT(conn->ibc_outstanding_credits >= 0);
812 LASSERT(conn->ibc_outstanding_credits <= conn->ibc_queue_depth);
813 LASSERT(conn->ibc_credits >= 0);
814 LASSERT(conn->ibc_credits <= conn->ibc_queue_depth);
816 if (conn->ibc_nsends_posted ==
817 kiblnd_concurrent_sends(ver, ni)) {
818 /* tx completions outstanding... */
819 CDEBUG(D_NET, "%s: posted enough\n",
820 libcfs_nid2str(peer_ni->ibp_nid));
824 if (credit != 0 && conn->ibc_credits == 0) { /* no credits */
825 CDEBUG(D_NET, "%s: no credits\n",
826 libcfs_nid2str(peer_ni->ibp_nid));
830 if (credit != 0 && !IBLND_OOB_CAPABLE(ver) &&
831 conn->ibc_credits == 1 && /* last credit reserved */
832 msg->ibm_type != IBLND_MSG_NOOP) { /* for NOOP */
833 CDEBUG(D_NET, "%s: not using last credit\n",
834 libcfs_nid2str(peer_ni->ibp_nid));
838 /* NB don't drop ibc_lock before bumping tx_sending */
839 list_del(&tx->tx_list);
842 if (msg->ibm_type == IBLND_MSG_NOOP &&
843 (!kiblnd_need_noop(conn) || /* redundant NOOP */
844 (IBLND_OOB_CAPABLE(ver) && /* posted enough NOOP */
845 conn->ibc_noops_posted == IBLND_OOB_MSGS(ver)))) {
846 /* OK to drop when posted enough NOOPs, since
847 * kiblnd_check_sends_locked will queue NOOP again when
848 * posted NOOPs complete */
849 spin_unlock(&conn->ibc_lock);
850 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
852 spin_lock(&conn->ibc_lock);
853 CDEBUG(D_NET, "%s(%d): redundant or enough NOOP\n",
854 libcfs_nid2str(peer_ni->ibp_nid),
855 conn->ibc_noops_posted);
859 kiblnd_pack_msg(peer_ni->ibp_ni, msg, ver, conn->ibc_outstanding_credits,
860 peer_ni->ibp_nid, conn->ibc_incarnation);
862 conn->ibc_credits -= credit;
863 conn->ibc_outstanding_credits = 0;
864 conn->ibc_nsends_posted++;
865 if (msg->ibm_type == IBLND_MSG_NOOP)
866 conn->ibc_noops_posted++;
868 /* CAVEAT EMPTOR! This tx could be the PUT_DONE of an RDMA
869 * PUT. If so, it was first queued here as a PUT_REQ, sent and
870 * stashed on ibc_active_txs, matched by an incoming PUT_ACK,
871 * and then re-queued here. It's (just) possible that
872 * tx_sending is non-zero if we've not done the tx_complete()
873 * from the first send; hence the ++ rather than = below. */
875 list_add(&tx->tx_list, &conn->ibc_active_txs);
877 /* I'm still holding ibc_lock! */
878 if (conn->ibc_state != IBLND_CONN_ESTABLISHED) {
880 } else if (tx->tx_pool->tpo_pool.po_failed ||
881 conn->ibc_hdev != tx->tx_pool->tpo_hdev) {
882 /* close_conn will launch failover */
885 struct kib_fast_reg_descriptor *frd = tx->tx_fmr.fmr_frd;
886 struct ib_send_wr *bad = &tx->tx_wrq[tx->tx_nwrq - 1].wr;
887 struct ib_send_wr *wr = &tx->tx_wrq[0].wr;
890 if (!frd->frd_valid) {
891 wr = &frd->frd_inv_wr.wr;
892 wr->next = &frd->frd_fastreg_wr.wr;
894 wr = &frd->frd_fastreg_wr.wr;
896 frd->frd_fastreg_wr.wr.next = &tx->tx_wrq[0].wr;
899 LASSERTF(bad->wr_id == kiblnd_ptr2wreqid(tx, IBLND_WID_TX),
900 "bad wr_id %#llx, opc %d, flags %d, peer_ni: %s\n",
901 bad->wr_id, bad->opcode, bad->send_flags,
902 libcfs_nid2str(conn->ibc_peer->ibp_nid));
905 if (lnet_send_error_simulation(tx->tx_lntmsg[0], &tx->tx_hstatus))
908 #ifdef HAVE_IB_POST_SEND_RECV_CONST
909 rc = ib_post_send(conn->ibc_cmid->qp, wr,
910 (const struct ib_send_wr **)&bad);
912 rc = ib_post_send(conn->ibc_cmid->qp, wr, &bad);
916 conn->ibc_last_send = ktime_get();
921 /* NB credits are transferred in the actual
922 * message, which can only be the last work item */
923 conn->ibc_credits += credit;
924 conn->ibc_outstanding_credits += msg->ibm_credits;
925 conn->ibc_nsends_posted--;
926 if (msg->ibm_type == IBLND_MSG_NOOP)
927 conn->ibc_noops_posted--;
933 done = (tx->tx_sending == 0);
935 list_del(&tx->tx_list);
937 spin_unlock(&conn->ibc_lock);
939 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
940 CERROR("Error %d posting transmit to %s\n",
941 rc, libcfs_nid2str(peer_ni->ibp_nid));
943 CDEBUG(D_NET, "Error %d posting transmit to %s\n",
944 rc, libcfs_nid2str(peer_ni->ibp_nid));
946 kiblnd_close_conn(conn, rc);
951 spin_lock(&conn->ibc_lock);
957 kiblnd_check_sends_locked(struct kib_conn *conn)
959 int ver = conn->ibc_version;
960 struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
963 /* Don't send anything until after the connection is established */
964 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
965 CDEBUG(D_NET, "%s too soon\n",
966 libcfs_nid2str(conn->ibc_peer->ibp_nid));
970 LASSERT(conn->ibc_nsends_posted <=
971 kiblnd_concurrent_sends(ver, ni));
972 LASSERT (!IBLND_OOB_CAPABLE(ver) ||
973 conn->ibc_noops_posted <= IBLND_OOB_MSGS(ver));
974 LASSERT (conn->ibc_reserved_credits >= 0);
976 while (conn->ibc_reserved_credits > 0 &&
977 !list_empty(&conn->ibc_tx_queue_rsrvd)) {
978 tx = list_entry(conn->ibc_tx_queue_rsrvd.next,
979 struct kib_tx, tx_list);
980 list_move_tail(&tx->tx_list, &conn->ibc_tx_queue);
981 conn->ibc_reserved_credits--;
984 if (kiblnd_need_noop(conn)) {
985 spin_unlock(&conn->ibc_lock);
987 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
989 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_NOOP, 0);
991 spin_lock(&conn->ibc_lock);
993 kiblnd_queue_tx_locked(tx, conn);
999 if (!list_empty(&conn->ibc_tx_queue_nocred)) {
1001 tx = list_entry(conn->ibc_tx_queue_nocred.next,
1002 struct kib_tx, tx_list);
1003 } else if (!list_empty(&conn->ibc_tx_noops)) {
1004 LASSERT (!IBLND_OOB_CAPABLE(ver));
1006 tx = list_entry(conn->ibc_tx_noops.next,
1007 struct kib_tx, tx_list);
1008 } else if (!list_empty(&conn->ibc_tx_queue)) {
1010 tx = list_entry(conn->ibc_tx_queue.next,
1011 struct kib_tx, tx_list);
1015 if (kiblnd_post_tx_locked(conn, tx, credit) != 0)
1021 kiblnd_tx_complete(struct kib_tx *tx, int status)
1023 int failed = (status != IB_WC_SUCCESS);
1024 struct kib_conn *conn = tx->tx_conn;
1027 LASSERT (tx->tx_sending > 0);
1030 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
1031 CNETERR("Tx -> %s cookie %#llx"
1032 " sending %d waiting %d: failed %d\n",
1033 libcfs_nid2str(conn->ibc_peer->ibp_nid),
1034 tx->tx_cookie, tx->tx_sending, tx->tx_waiting,
1037 kiblnd_close_conn(conn, -EIO);
1039 kiblnd_peer_alive(conn->ibc_peer);
1042 spin_lock(&conn->ibc_lock);
1044 /* I could be racing with rdma completion. Whoever makes 'tx' idle
1045 * gets to free it, which also drops its ref on 'conn'. */
1048 conn->ibc_nsends_posted--;
1049 if (tx->tx_msg->ibm_type == IBLND_MSG_NOOP)
1050 conn->ibc_noops_posted--;
1053 tx->tx_hstatus = LNET_MSG_STATUS_REMOTE_DROPPED;
1054 tx->tx_waiting = 0; /* don't wait for peer_ni */
1055 tx->tx_status = -EIO;
1058 idle = (tx->tx_sending == 0) && /* This is the final callback */
1059 !tx->tx_waiting && /* Not waiting for peer_ni */
1060 !tx->tx_queued; /* Not re-queued (PUT_DONE) */
1062 list_del(&tx->tx_list);
1064 kiblnd_check_sends_locked(conn);
1065 spin_unlock(&conn->ibc_lock);
1072 kiblnd_init_tx_msg(struct lnet_ni *ni, struct kib_tx *tx, int type,
1075 struct kib_hca_dev *hdev = tx->tx_pool->tpo_hdev;
1076 struct ib_sge *sge = &tx->tx_msgsge;
1077 struct ib_rdma_wr *wrq;
1078 int nob = offsetof(struct kib_msg, ibm_u) + body_nob;
1079 #ifdef HAVE_IB_GET_DMA_MR
1080 struct ib_mr *mr = hdev->ibh_mrs;
1083 LASSERT(tx->tx_nwrq >= 0);
1084 LASSERT(tx->tx_nwrq < IBLND_MAX_RDMA_FRAGS + 1);
1085 LASSERT(nob <= IBLND_MSG_SIZE);
1086 #ifdef HAVE_IB_GET_DMA_MR
1087 LASSERT(mr != NULL);
1090 kiblnd_init_msg(tx->tx_msg, type, body_nob);
1092 #ifdef HAVE_IB_GET_DMA_MR
1093 sge->lkey = mr->lkey;
1095 sge->lkey = hdev->ibh_pd->local_dma_lkey;
1097 sge->addr = tx->tx_msgaddr;
1100 wrq = &tx->tx_wrq[tx->tx_nwrq];
1101 memset(wrq, 0, sizeof(*wrq));
1103 wrq->wr.next = NULL;
1104 wrq->wr.wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_TX);
1105 wrq->wr.sg_list = sge;
1106 wrq->wr.num_sge = 1;
1107 wrq->wr.opcode = IB_WR_SEND;
1108 wrq->wr.send_flags = IB_SEND_SIGNALED;
1114 kiblnd_init_rdma(struct kib_conn *conn, struct kib_tx *tx, int type,
1115 int resid, struct kib_rdma_desc *dstrd, u64 dstcookie)
1117 struct kib_msg *ibmsg = tx->tx_msg;
1118 struct kib_rdma_desc *srcrd = tx->tx_rd;
1119 struct ib_rdma_wr *wrq = NULL;
1127 LASSERT(!in_interrupt());
1128 LASSERT(tx->tx_nwrq == 0 && tx->tx_nsge == 0);
1129 LASSERT(type == IBLND_MSG_GET_DONE || type == IBLND_MSG_PUT_DONE);
1131 for (srcidx = dstidx = wrq_sge = sge_nob = 0;
1132 resid > 0; resid -= sge_nob) {
1135 if (srcidx >= srcrd->rd_nfrags) {
1136 CERROR("Src buffer exhausted: %d frags\n", srcidx);
1141 if (dstidx >= dstrd->rd_nfrags) {
1142 CERROR("Dst buffer exhausted: %d frags\n", dstidx);
1147 if (tx->tx_nwrq >= conn->ibc_max_frags) {
1148 CERROR("RDMA has too many fragments for peer_ni %s (%d), "
1149 "src idx/frags: %d/%d dst idx/frags: %d/%d\n",
1150 libcfs_nid2str(conn->ibc_peer->ibp_nid),
1151 conn->ibc_max_frags,
1152 srcidx, srcrd->rd_nfrags,
1153 dstidx, dstrd->rd_nfrags);
1158 sge_nob = min3(kiblnd_rd_frag_size(srcrd, srcidx),
1159 kiblnd_rd_frag_size(dstrd, dstidx),
1162 sge = &tx->tx_sge[tx->tx_nsge];
1163 sge->addr = kiblnd_rd_frag_addr(srcrd, srcidx);
1164 sge->lkey = kiblnd_rd_frag_key(srcrd, srcidx);
1165 sge->length = sge_nob;
1168 wrq = &tx->tx_wrq[tx->tx_nwrq];
1170 wrq->wr.next = &(wrq + 1)->wr;
1171 wrq->wr.wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_RDMA);
1172 wrq->wr.sg_list = sge;
1173 wrq->wr.opcode = IB_WR_RDMA_WRITE;
1174 wrq->wr.send_flags = 0;
1176 #ifdef HAVE_IB_RDMA_WR
1177 wrq->remote_addr = kiblnd_rd_frag_addr(dstrd,
1179 wrq->rkey = kiblnd_rd_frag_key(dstrd,
1182 wrq->wr.wr.rdma.remote_addr = kiblnd_rd_frag_addr(dstrd,
1184 wrq->wr.wr.rdma.rkey = kiblnd_rd_frag_key(dstrd,
1189 srcidx = kiblnd_rd_consume_frag(srcrd, srcidx, sge_nob);
1190 dstidx = kiblnd_rd_consume_frag(dstrd, dstidx, sge_nob);
1193 if (wrq_sge == *kiblnd_tunables.kib_wrq_sge || dstidx != prev) {
1195 wrq->wr.num_sge = wrq_sge;
1201 if (rc < 0) /* no RDMA if completing with failure */
1202 tx->tx_nwrq = tx->tx_nsge = 0;
1204 ibmsg->ibm_u.completion.ibcm_status = rc;
1205 ibmsg->ibm_u.completion.ibcm_cookie = dstcookie;
1206 kiblnd_init_tx_msg(conn->ibc_peer->ibp_ni, tx,
1207 type, sizeof(struct kib_completion_msg));
1213 kiblnd_queue_tx_locked(struct kib_tx *tx, struct kib_conn *conn)
1215 struct list_head *q;
1218 LASSERT(tx->tx_nwrq > 0); /* work items set up */
1219 LASSERT(!tx->tx_queued); /* not queued for sending already */
1220 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1222 if (conn->ibc_state >= IBLND_CONN_DISCONNECTED) {
1223 tx->tx_status = -ECONNABORTED;
1225 if (tx->tx_conn != NULL) {
1226 /* PUT_DONE first attached to conn as a PUT_REQ */
1227 LASSERT(tx->tx_conn == conn);
1228 LASSERT(tx->tx_msg->ibm_type == IBLND_MSG_PUT_DONE);
1230 kiblnd_conn_decref(conn);
1232 list_add(&tx->tx_list, &conn->ibc_zombie_txs);
1237 timeout_ns = kiblnd_timeout() * NSEC_PER_SEC;
1239 tx->tx_deadline = ktime_add_ns(ktime_get(), timeout_ns);
1241 if (tx->tx_conn == NULL) {
1242 kiblnd_conn_addref(conn);
1244 LASSERT (tx->tx_msg->ibm_type != IBLND_MSG_PUT_DONE);
1246 /* PUT_DONE first attached to conn as a PUT_REQ */
1247 LASSERT (tx->tx_conn == conn);
1248 LASSERT (tx->tx_msg->ibm_type == IBLND_MSG_PUT_DONE);
1251 switch (tx->tx_msg->ibm_type) {
1255 case IBLND_MSG_PUT_REQ:
1256 case IBLND_MSG_GET_REQ:
1257 q = &conn->ibc_tx_queue_rsrvd;
1260 case IBLND_MSG_PUT_NAK:
1261 case IBLND_MSG_PUT_ACK:
1262 case IBLND_MSG_PUT_DONE:
1263 case IBLND_MSG_GET_DONE:
1264 q = &conn->ibc_tx_queue_nocred;
1267 case IBLND_MSG_NOOP:
1268 if (IBLND_OOB_CAPABLE(conn->ibc_version))
1269 q = &conn->ibc_tx_queue_nocred;
1271 q = &conn->ibc_tx_noops;
1274 case IBLND_MSG_IMMEDIATE:
1275 q = &conn->ibc_tx_queue;
1279 list_add_tail(&tx->tx_list, q);
1283 kiblnd_queue_tx(struct kib_tx *tx, struct kib_conn *conn)
1285 spin_lock(&conn->ibc_lock);
1286 kiblnd_queue_tx_locked(tx, conn);
1287 kiblnd_check_sends_locked(conn);
1288 spin_unlock(&conn->ibc_lock);
1291 static int kiblnd_resolve_addr(struct rdma_cm_id *cmid,
1292 struct sockaddr_in *srcaddr,
1293 struct sockaddr_in *dstaddr,
1296 unsigned short port;
1299 /* allow the port to be reused */
1300 rc = rdma_set_reuseaddr(cmid, 1);
1302 CERROR("Unable to set reuse on cmid: %d\n", rc);
1306 /* look for a free privileged port */
1307 for (port = PROT_SOCK-1; port > 0; port--) {
1308 srcaddr->sin_port = htons(port);
1309 rc = rdma_resolve_addr(cmid,
1310 (struct sockaddr *)srcaddr,
1311 (struct sockaddr *)dstaddr,
1314 CDEBUG(D_NET, "bound to port %hu\n", port);
1316 } else if (rc == -EADDRINUSE || rc == -EADDRNOTAVAIL) {
1317 CDEBUG(D_NET, "bind to port %hu failed: %d\n",
1324 CERROR("Failed to bind to a free privileged port\n");
1329 kiblnd_connect_peer(struct kib_peer_ni *peer_ni)
1331 struct rdma_cm_id *cmid;
1332 struct kib_dev *dev;
1333 struct kib_net *net = peer_ni->ibp_ni->ni_data;
1334 struct sockaddr_in srcaddr;
1335 struct sockaddr_in dstaddr;
1338 LASSERT (net != NULL);
1339 LASSERT (peer_ni->ibp_connecting > 0);
1341 cmid = kiblnd_rdma_create_id(peer_ni->ibp_ni->ni_net_ns,
1342 kiblnd_cm_callback, peer_ni,
1343 RDMA_PS_TCP, IB_QPT_RC);
1346 CERROR("Can't create CMID for %s: %ld\n",
1347 libcfs_nid2str(peer_ni->ibp_nid), PTR_ERR(cmid));
1353 memset(&srcaddr, 0, sizeof(srcaddr));
1354 srcaddr.sin_family = AF_INET;
1355 srcaddr.sin_addr.s_addr = htonl(dev->ibd_ifip);
1357 memset(&dstaddr, 0, sizeof(dstaddr));
1358 dstaddr.sin_family = AF_INET;
1359 dstaddr.sin_port = htons(*kiblnd_tunables.kib_service);
1360 dstaddr.sin_addr.s_addr = htonl(LNET_NIDADDR(peer_ni->ibp_nid));
1362 kiblnd_peer_addref(peer_ni); /* cmid's ref */
1364 if (*kiblnd_tunables.kib_use_priv_port) {
1365 rc = kiblnd_resolve_addr(cmid, &srcaddr, &dstaddr,
1366 kiblnd_timeout() * 1000);
1368 rc = rdma_resolve_addr(cmid,
1369 (struct sockaddr *)&srcaddr,
1370 (struct sockaddr *)&dstaddr,
1371 kiblnd_timeout() * 1000);
1374 /* Can't initiate address resolution: */
1375 CERROR("Can't resolve addr for %s: %d\n",
1376 libcfs_nid2str(peer_ni->ibp_nid), rc);
1383 kiblnd_peer_connect_failed(peer_ni, 1, rc);
1384 kiblnd_peer_decref(peer_ni); /* cmid's ref */
1385 rdma_destroy_id(cmid);
1388 kiblnd_peer_connect_failed(peer_ni, 1, rc);
1392 kiblnd_reconnect_peer(struct kib_peer_ni *peer_ni)
1394 rwlock_t *glock = &kiblnd_data.kib_global_lock;
1395 char *reason = NULL;
1397 unsigned long flags;
1399 write_lock_irqsave(glock, flags);
1400 if (peer_ni->ibp_reconnecting == 0) {
1401 if (peer_ni->ibp_accepting)
1402 reason = "accepting";
1403 else if (peer_ni->ibp_connecting)
1404 reason = "connecting";
1405 else if (!list_empty(&peer_ni->ibp_conns))
1406 reason = "connected";
1407 else /* connected then closed */
1413 if (peer_ni->ibp_accepting)
1414 CNETERR("Detecting race between accepting and reconnecting\n");
1415 peer_ni->ibp_reconnecting--;
1417 if (!kiblnd_peer_active(peer_ni)) {
1418 list_splice_init(&peer_ni->ibp_tx_queue, &txs);
1419 reason = "unlinked";
1423 peer_ni->ibp_connecting++;
1424 peer_ni->ibp_reconnected++;
1426 write_unlock_irqrestore(glock, flags);
1428 kiblnd_connect_peer(peer_ni);
1432 write_unlock_irqrestore(glock, flags);
1434 CWARN("Abort reconnection of %s: %s\n",
1435 libcfs_nid2str(peer_ni->ibp_nid), reason);
1436 kiblnd_txlist_done(&txs, -ECONNABORTED,
1437 LNET_MSG_STATUS_LOCAL_ABORTED);
1442 kiblnd_launch_tx(struct lnet_ni *ni, struct kib_tx *tx, lnet_nid_t nid)
1444 struct kib_peer_ni *peer_ni;
1445 struct kib_peer_ni *peer2;
1446 struct kib_conn *conn;
1447 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
1448 unsigned long flags;
1451 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
1453 /* If I get here, I've committed to send, so I complete the tx with
1454 * failure on any problems */
1456 LASSERT (tx == NULL || tx->tx_conn == NULL); /* only set when assigned a conn */
1457 LASSERT (tx == NULL || tx->tx_nwrq > 0); /* work items have been set up */
1459 /* First time, just use a read lock since I expect to find my peer_ni
1461 read_lock_irqsave(g_lock, flags);
1463 peer_ni = kiblnd_find_peer_locked(ni, nid);
1464 if (peer_ni != NULL && !list_empty(&peer_ni->ibp_conns)) {
1465 /* Found a peer_ni with an established connection */
1466 conn = kiblnd_get_conn_locked(peer_ni);
1467 kiblnd_conn_addref(conn); /* 1 ref for me... */
1469 read_unlock_irqrestore(g_lock, flags);
1472 kiblnd_queue_tx(tx, conn);
1473 kiblnd_conn_decref(conn); /* ...to here */
1477 read_unlock(g_lock);
1478 /* Re-try with a write lock */
1481 peer_ni = kiblnd_find_peer_locked(ni, nid);
1482 if (peer_ni != NULL) {
1483 if (list_empty(&peer_ni->ibp_conns)) {
1484 /* found a peer_ni, but it's still connecting... */
1485 LASSERT(kiblnd_peer_connecting(peer_ni));
1487 list_add_tail(&tx->tx_list,
1488 &peer_ni->ibp_tx_queue);
1489 write_unlock_irqrestore(g_lock, flags);
1491 conn = kiblnd_get_conn_locked(peer_ni);
1492 kiblnd_conn_addref(conn); /* 1 ref for me... */
1494 write_unlock_irqrestore(g_lock, flags);
1497 kiblnd_queue_tx(tx, conn);
1498 kiblnd_conn_decref(conn); /* ...to here */
1503 write_unlock_irqrestore(g_lock, flags);
1505 /* Allocate a peer_ni ready to add to the peer_ni table and retry */
1506 rc = kiblnd_create_peer(ni, &peer_ni, nid);
1508 CERROR("Can't create peer_ni %s\n", libcfs_nid2str(nid));
1510 tx->tx_status = -EHOSTUNREACH;
1512 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
1518 write_lock_irqsave(g_lock, flags);
1520 peer2 = kiblnd_find_peer_locked(ni, nid);
1521 if (peer2 != NULL) {
1522 if (list_empty(&peer2->ibp_conns)) {
1523 /* found a peer_ni, but it's still connecting... */
1524 LASSERT(kiblnd_peer_connecting(peer2));
1526 list_add_tail(&tx->tx_list,
1527 &peer2->ibp_tx_queue);
1528 write_unlock_irqrestore(g_lock, flags);
1530 conn = kiblnd_get_conn_locked(peer2);
1531 kiblnd_conn_addref(conn); /* 1 ref for me... */
1533 write_unlock_irqrestore(g_lock, flags);
1536 kiblnd_queue_tx(tx, conn);
1537 kiblnd_conn_decref(conn); /* ...to here */
1540 kiblnd_peer_decref(peer_ni);
1544 /* Brand new peer_ni */
1545 LASSERT(peer_ni->ibp_connecting == 0);
1546 tunables = &peer_ni->ibp_ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
1547 peer_ni->ibp_connecting = tunables->lnd_conns_per_peer;
1549 /* always called with a ref on ni, which prevents ni being shutdown */
1550 LASSERT(((struct kib_net *)ni->ni_data)->ibn_shutdown == 0);
1553 list_add_tail(&tx->tx_list, &peer_ni->ibp_tx_queue);
1555 kiblnd_peer_addref(peer_ni);
1556 list_add_tail(&peer_ni->ibp_list, kiblnd_nid2peerlist(nid));
1558 write_unlock_irqrestore(g_lock, flags);
1560 for (i = 0; i < tunables->lnd_conns_per_peer; i++)
1561 kiblnd_connect_peer(peer_ni);
1562 kiblnd_peer_decref(peer_ni);
1566 kiblnd_send(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg)
1568 struct lnet_hdr *hdr = &lntmsg->msg_hdr;
1569 int type = lntmsg->msg_type;
1570 struct lnet_process_id target = lntmsg->msg_target;
1571 int target_is_router = lntmsg->msg_target_is_router;
1572 int routing = lntmsg->msg_routing;
1573 unsigned int payload_niov = lntmsg->msg_niov;
1574 struct bio_vec *payload_kiov = lntmsg->msg_kiov;
1575 unsigned int payload_offset = lntmsg->msg_offset;
1576 unsigned int payload_nob = lntmsg->msg_len;
1577 struct kib_msg *ibmsg;
1578 struct kib_rdma_desc *rd;
1583 /* NB 'private' is different depending on what we're sending.... */
1585 CDEBUG(D_NET, "sending %d bytes in %d frags to %s\n",
1586 payload_nob, payload_niov, libcfs_id2str(target));
1588 LASSERT (payload_nob == 0 || payload_niov > 0);
1589 LASSERT (payload_niov <= LNET_MAX_IOV);
1591 /* Thread context */
1592 LASSERT (!in_interrupt());
1600 LASSERT (payload_nob == 0);
1604 if (routing || target_is_router)
1605 break; /* send IMMEDIATE */
1607 /* is the REPLY message too small for RDMA? */
1608 nob = offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[lntmsg->msg_md->md_length]);
1609 if (nob <= IBLND_MSG_SIZE)
1610 break; /* send IMMEDIATE */
1612 tx = kiblnd_get_idle_tx(ni, target.nid);
1614 CERROR("Can't allocate txd for GET to %s\n",
1615 libcfs_nid2str(target.nid));
1620 rd = &ibmsg->ibm_u.get.ibgm_rd;
1621 rc = kiblnd_setup_rd_kiov(ni, tx, rd,
1622 lntmsg->msg_md->md_niov,
1623 lntmsg->msg_md->md_kiov,
1624 0, lntmsg->msg_md->md_length);
1626 CERROR("Can't setup GET sink for %s: %d\n",
1627 libcfs_nid2str(target.nid), rc);
1628 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
1633 nob = offsetof(struct kib_get_msg, ibgm_rd.rd_frags[rd->rd_nfrags]);
1634 ibmsg->ibm_u.get.ibgm_cookie = tx->tx_cookie;
1635 ibmsg->ibm_u.get.ibgm_hdr = *hdr;
1637 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_GET_REQ, nob);
1639 tx->tx_lntmsg[1] = lnet_create_reply_msg(ni, lntmsg);
1640 if (tx->tx_lntmsg[1] == NULL) {
1641 CERROR("Can't create reply for GET -> %s\n",
1642 libcfs_nid2str(target.nid));
1647 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg[0,1] on completion */
1648 tx->tx_waiting = 1; /* waiting for GET_DONE */
1649 kiblnd_launch_tx(ni, tx, target.nid);
1652 case LNET_MSG_REPLY:
1654 /* Is the payload small enough not to need RDMA? */
1655 nob = offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[payload_nob]);
1656 if (nob <= IBLND_MSG_SIZE)
1657 break; /* send IMMEDIATE */
1659 tx = kiblnd_get_idle_tx(ni, target.nid);
1661 CERROR("Can't allocate %s txd for %s\n",
1662 type == LNET_MSG_PUT ? "PUT" : "REPLY",
1663 libcfs_nid2str(target.nid));
1667 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1668 payload_niov, payload_kiov,
1669 payload_offset, payload_nob);
1671 CERROR("Can't setup PUT src for %s: %d\n",
1672 libcfs_nid2str(target.nid), rc);
1678 ibmsg->ibm_u.putreq.ibprm_hdr = *hdr;
1679 ibmsg->ibm_u.putreq.ibprm_cookie = tx->tx_cookie;
1680 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_REQ,
1681 sizeof(struct kib_putreq_msg));
1683 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1684 tx->tx_waiting = 1; /* waiting for PUT_{ACK,NAK} */
1685 kiblnd_launch_tx(ni, tx, target.nid);
1689 /* send IMMEDIATE */
1690 LASSERT(offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[payload_nob])
1693 tx = kiblnd_get_idle_tx(ni, target.nid);
1695 CERROR ("Can't send %d to %s: tx descs exhausted\n",
1696 type, libcfs_nid2str(target.nid));
1701 ibmsg->ibm_u.immediate.ibim_hdr = *hdr;
1703 lnet_copy_kiov2flat(IBLND_MSG_SIZE, ibmsg,
1704 offsetof(struct kib_msg,
1705 ibm_u.immediate.ibim_payload),
1706 payload_niov, payload_kiov,
1707 payload_offset, payload_nob);
1709 nob = offsetof(struct kib_immediate_msg, ibim_payload[payload_nob]);
1710 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_IMMEDIATE, nob);
1712 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1713 kiblnd_launch_tx(ni, tx, target.nid);
1718 kiblnd_reply(struct lnet_ni *ni, struct kib_rx *rx, struct lnet_msg *lntmsg)
1720 struct lnet_process_id target = lntmsg->msg_target;
1721 unsigned int niov = lntmsg->msg_niov;
1722 struct bio_vec *kiov = lntmsg->msg_kiov;
1723 unsigned int offset = lntmsg->msg_offset;
1724 unsigned int nob = lntmsg->msg_len;
1728 tx = kiblnd_get_idle_tx(ni, rx->rx_conn->ibc_peer->ibp_nid);
1730 CERROR("Can't get tx for REPLY to %s\n",
1731 libcfs_nid2str(target.nid));
1738 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1739 niov, kiov, offset, nob);
1742 CERROR("Can't setup GET src for %s: %d\n",
1743 libcfs_nid2str(target.nid), rc);
1747 rc = kiblnd_init_rdma(rx->rx_conn, tx,
1748 IBLND_MSG_GET_DONE, nob,
1749 &rx->rx_msg->ibm_u.get.ibgm_rd,
1750 rx->rx_msg->ibm_u.get.ibgm_cookie);
1752 CERROR("Can't setup rdma for GET from %s: %d\n",
1753 libcfs_nid2str(target.nid), rc);
1758 /* No RDMA: local completion may happen now! */
1759 lnet_finalize(lntmsg, 0);
1761 /* RDMA: lnet_finalize(lntmsg) when it
1763 tx->tx_lntmsg[0] = lntmsg;
1766 kiblnd_queue_tx(tx, rx->rx_conn);
1771 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
1774 lnet_finalize(lntmsg, -EIO);
1778 kiblnd_recv(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg,
1779 int delayed, unsigned int niov, struct bio_vec *kiov,
1780 unsigned int offset, unsigned int mlen, unsigned int rlen)
1782 struct kib_rx *rx = private;
1783 struct kib_msg *rxmsg = rx->rx_msg;
1784 struct kib_conn *conn = rx->rx_conn;
1788 int post_credit = IBLND_POSTRX_PEER_CREDIT;
1791 LASSERT (mlen <= rlen);
1792 LASSERT (!in_interrupt());
1794 switch (rxmsg->ibm_type) {
1798 case IBLND_MSG_IMMEDIATE:
1799 nob = offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[rlen]);
1800 if (nob > rx->rx_nob) {
1801 CERROR ("Immediate message from %s too big: %d(%d)\n",
1802 libcfs_nid2str(rxmsg->ibm_u.immediate.ibim_hdr.src_nid),
1808 lnet_copy_flat2kiov(niov, kiov, offset,
1809 IBLND_MSG_SIZE, rxmsg,
1810 offsetof(struct kib_msg,
1811 ibm_u.immediate.ibim_payload),
1813 lnet_finalize(lntmsg, 0);
1816 case IBLND_MSG_PUT_REQ: {
1817 struct kib_msg *txmsg;
1818 struct kib_rdma_desc *rd;
1819 ibprm_cookie = rxmsg->ibm_u.putreq.ibprm_cookie;
1822 lnet_finalize(lntmsg, 0);
1823 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK,
1828 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
1830 CERROR("Can't allocate tx for %s\n",
1831 libcfs_nid2str(conn->ibc_peer->ibp_nid));
1832 /* Not replying will break the connection */
1838 rd = &txmsg->ibm_u.putack.ibpam_rd;
1839 rc = kiblnd_setup_rd_kiov(ni, tx, rd,
1840 niov, kiov, offset, mlen);
1842 CERROR("Can't setup PUT sink for %s: %d\n",
1843 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
1844 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
1846 /* tell peer_ni it's over */
1847 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK,
1852 nob = offsetof(struct kib_putack_msg, ibpam_rd.rd_frags[rd->rd_nfrags]);
1853 txmsg->ibm_u.putack.ibpam_src_cookie = ibprm_cookie;
1854 txmsg->ibm_u.putack.ibpam_dst_cookie = tx->tx_cookie;
1856 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_ACK, nob);
1858 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1859 tx->tx_waiting = 1; /* waiting for PUT_DONE */
1860 kiblnd_queue_tx(tx, conn);
1862 /* reposted buffer reserved for PUT_DONE */
1863 post_credit = IBLND_POSTRX_NO_CREDIT;
1867 case IBLND_MSG_GET_REQ:
1868 if (lntmsg != NULL) {
1869 /* Optimized GET; RDMA lntmsg's payload */
1870 kiblnd_reply(ni, rx, lntmsg);
1872 /* GET didn't match anything */
1873 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_GET_DONE,
1875 rxmsg->ibm_u.get.ibgm_cookie);
1880 kiblnd_post_rx(rx, post_credit);
1885 kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name)
1887 struct task_struct *task = kthread_run(fn, arg, name);
1890 return PTR_ERR(task);
1892 atomic_inc(&kiblnd_data.kib_nthreads);
1897 kiblnd_thread_fini (void)
1899 atomic_dec (&kiblnd_data.kib_nthreads);
1903 kiblnd_peer_alive(struct kib_peer_ni *peer_ni)
1905 /* This is racy, but everyone's only writing ktime_get_seconds() */
1906 peer_ni->ibp_last_alive = ktime_get_seconds();
1911 kiblnd_peer_notify(struct kib_peer_ni *peer_ni)
1914 time64_t last_alive = 0;
1915 unsigned long flags;
1917 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1919 if (kiblnd_peer_idle(peer_ni) && peer_ni->ibp_error != 0) {
1920 error = peer_ni->ibp_error;
1921 peer_ni->ibp_error = 0;
1923 last_alive = peer_ni->ibp_last_alive;
1926 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1929 lnet_notify(peer_ni->ibp_ni,
1930 peer_ni->ibp_nid, false, false, last_alive);
1934 kiblnd_close_conn_locked(struct kib_conn *conn, int error)
1936 /* This just does the immediate housekeeping. 'error' is zero for a
1937 * normal shutdown which can happen only after the connection has been
1938 * established. If the connection is established, schedule the
1939 * connection to be finished off by the connd. Otherwise the connd is
1940 * already dealing with it (either to set it up or tear it down).
1941 * Caller holds kib_global_lock exclusively in irq context */
1942 struct kib_peer_ni *peer_ni = conn->ibc_peer;
1943 struct kib_dev *dev;
1944 unsigned long flags;
1946 LASSERT (error != 0 || conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1948 if (error != 0 && conn->ibc_comms_error == 0)
1949 conn->ibc_comms_error = error;
1951 if (conn->ibc_state != IBLND_CONN_ESTABLISHED)
1952 return; /* already being handled */
1955 list_empty(&conn->ibc_tx_noops) &&
1956 list_empty(&conn->ibc_tx_queue) &&
1957 list_empty(&conn->ibc_tx_queue_rsrvd) &&
1958 list_empty(&conn->ibc_tx_queue_nocred) &&
1959 list_empty(&conn->ibc_active_txs)) {
1960 CDEBUG(D_NET, "closing conn to %s\n",
1961 libcfs_nid2str(peer_ni->ibp_nid));
1963 CNETERR("Closing conn to %s: error %d%s%s%s%s%s\n",
1964 libcfs_nid2str(peer_ni->ibp_nid), error,
1965 list_empty(&conn->ibc_tx_queue) ? "" : "(sending)",
1966 list_empty(&conn->ibc_tx_noops) ? "" : "(sending_noops)",
1967 list_empty(&conn->ibc_tx_queue_rsrvd) ?
1968 "" : "(sending_rsrvd)",
1969 list_empty(&conn->ibc_tx_queue_nocred) ?
1970 "" : "(sending_nocred)",
1971 list_empty(&conn->ibc_active_txs) ? "" : "(waiting)");
1974 dev = ((struct kib_net *)peer_ni->ibp_ni->ni_data)->ibn_dev;
1975 if (peer_ni->ibp_next_conn == conn)
1976 /* clear next_conn so it won't be used */
1977 peer_ni->ibp_next_conn = NULL;
1978 list_del(&conn->ibc_list);
1979 /* connd (see below) takes over ibc_list's ref */
1981 if (list_empty(&peer_ni->ibp_conns) && /* no more conns */
1982 kiblnd_peer_active(peer_ni)) { /* still in peer_ni table */
1983 kiblnd_unlink_peer_locked(peer_ni);
1985 /* set/clear error on last conn */
1986 peer_ni->ibp_error = conn->ibc_comms_error;
1989 kiblnd_set_conn_state(conn, IBLND_CONN_CLOSING);
1992 kiblnd_dev_can_failover(dev)) {
1993 list_add_tail(&dev->ibd_fail_list,
1994 &kiblnd_data.kib_failed_devs);
1995 wake_up(&kiblnd_data.kib_failover_waitq);
1998 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
2000 list_add_tail(&conn->ibc_list, &kiblnd_data.kib_connd_conns);
2001 wake_up(&kiblnd_data.kib_connd_waitq);
2003 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
2007 kiblnd_close_conn(struct kib_conn *conn, int error)
2009 unsigned long flags;
2011 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2013 kiblnd_close_conn_locked(conn, error);
2015 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2019 kiblnd_handle_early_rxs(struct kib_conn *conn)
2021 unsigned long flags;
2024 LASSERT(!in_interrupt());
2025 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
2027 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2028 while (!list_empty(&conn->ibc_early_rxs)) {
2029 rx = list_entry(conn->ibc_early_rxs.next,
2030 struct kib_rx, rx_list);
2031 list_del(&rx->rx_list);
2032 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2034 kiblnd_handle_rx(rx);
2036 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2038 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2042 kiblnd_abort_txs(struct kib_conn *conn, struct list_head *txs)
2045 struct list_head *tmp;
2046 struct list_head *nxt;
2049 spin_lock(&conn->ibc_lock);
2051 list_for_each_safe(tmp, nxt, txs) {
2052 tx = list_entry(tmp, struct kib_tx, tx_list);
2054 if (txs == &conn->ibc_active_txs) {
2055 LASSERT(!tx->tx_queued);
2056 LASSERT(tx->tx_waiting ||
2057 tx->tx_sending != 0);
2058 if (conn->ibc_comms_error == -ETIMEDOUT) {
2059 if (tx->tx_waiting && !tx->tx_sending)
2061 LNET_MSG_STATUS_REMOTE_TIMEOUT;
2062 else if (tx->tx_sending)
2064 LNET_MSG_STATUS_NETWORK_TIMEOUT;
2067 LASSERT(tx->tx_queued);
2068 if (conn->ibc_comms_error == -ETIMEDOUT)
2069 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_TIMEOUT;
2071 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
2074 tx->tx_status = -ECONNABORTED;
2078 * TODO: This makes an assumption that
2079 * kiblnd_tx_complete() will be called for each tx. If
2080 * that event is dropped we could end up with stale
2081 * connections floating around. We'd like to deal with
2082 * that in a better way.
2084 * Also that means we can exceed the timeout by many
2087 if (tx->tx_sending == 0) {
2089 list_move(&tx->tx_list, &zombies);
2093 spin_unlock(&conn->ibc_lock);
2096 * aborting transmits occurs when finalizing the connection.
2097 * The connection is finalized on error.
2098 * Passing LNET_MSG_STATUS_OK to txlist_done() will not
2099 * override the value already set in tx->tx_hstatus above.
2101 kiblnd_txlist_done(&zombies, -ECONNABORTED, LNET_MSG_STATUS_OK);
2105 kiblnd_finalise_conn(struct kib_conn *conn)
2107 LASSERT (!in_interrupt());
2108 LASSERT (conn->ibc_state > IBLND_CONN_INIT);
2110 /* abort_receives moves QP state to IB_QPS_ERR. This is only required
2111 * for connections that didn't get as far as being connected, because
2112 * rdma_disconnect() does this for free. */
2113 kiblnd_abort_receives(conn);
2115 kiblnd_set_conn_state(conn, IBLND_CONN_DISCONNECTED);
2117 /* Complete all tx descs not waiting for sends to complete.
2118 * NB we should be safe from RDMA now that the QP has changed state */
2120 kiblnd_abort_txs(conn, &conn->ibc_tx_noops);
2121 kiblnd_abort_txs(conn, &conn->ibc_tx_queue);
2122 kiblnd_abort_txs(conn, &conn->ibc_tx_queue_rsrvd);
2123 kiblnd_abort_txs(conn, &conn->ibc_tx_queue_nocred);
2124 kiblnd_abort_txs(conn, &conn->ibc_active_txs);
2126 kiblnd_handle_early_rxs(conn);
2130 kiblnd_peer_connect_failed(struct kib_peer_ni *peer_ni, int active,
2134 unsigned long flags;
2136 LASSERT (error != 0);
2137 LASSERT (!in_interrupt());
2139 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2142 LASSERT(peer_ni->ibp_connecting > 0);
2143 peer_ni->ibp_connecting--;
2145 LASSERT (peer_ni->ibp_accepting > 0);
2146 peer_ni->ibp_accepting--;
2149 if (kiblnd_peer_connecting(peer_ni)) {
2150 /* another connection attempt under way... */
2151 write_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2156 peer_ni->ibp_reconnected = 0;
2157 if (list_empty(&peer_ni->ibp_conns)) {
2158 /* Take peer_ni's blocked transmits to complete with error */
2159 list_splice_init(&peer_ni->ibp_tx_queue, &zombies);
2161 if (kiblnd_peer_active(peer_ni))
2162 kiblnd_unlink_peer_locked(peer_ni);
2164 peer_ni->ibp_error = error;
2166 /* Can't have blocked transmits if there are connections */
2167 LASSERT(list_empty(&peer_ni->ibp_tx_queue));
2170 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2172 kiblnd_peer_notify(peer_ni);
2174 if (list_empty(&zombies))
2177 CNETERR("Deleting messages for %s: connection failed\n",
2178 libcfs_nid2str(peer_ni->ibp_nid));
2180 kiblnd_txlist_done(&zombies, error,
2181 LNET_MSG_STATUS_LOCAL_DROPPED);
2185 kiblnd_connreq_done(struct kib_conn *conn, int status)
2187 struct kib_peer_ni *peer_ni = conn->ibc_peer;
2190 unsigned long flags;
2193 active = (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2195 CDEBUG(D_NET,"%s: active(%d), version(%x), status(%d)\n",
2196 libcfs_nid2str(peer_ni->ibp_nid), active,
2197 conn->ibc_version, status);
2199 LASSERT (!in_interrupt());
2200 LASSERT ((conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT &&
2201 peer_ni->ibp_connecting > 0) ||
2202 (conn->ibc_state == IBLND_CONN_PASSIVE_WAIT &&
2203 peer_ni->ibp_accepting > 0));
2205 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
2206 conn->ibc_connvars = NULL;
2209 /* failed to establish connection */
2210 kiblnd_peer_connect_failed(peer_ni, active, status);
2211 kiblnd_finalise_conn(conn);
2215 /* connection established */
2216 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2218 conn->ibc_last_send = ktime_get();
2219 kiblnd_set_conn_state(conn, IBLND_CONN_ESTABLISHED);
2220 kiblnd_peer_alive(peer_ni);
2222 /* Add conn to peer_ni's list and nuke any dangling conns from a different
2223 * peer_ni instance... */
2224 kiblnd_conn_addref(conn); /* +1 ref for ibc_list */
2225 list_add(&conn->ibc_list, &peer_ni->ibp_conns);
2226 peer_ni->ibp_reconnected = 0;
2228 peer_ni->ibp_connecting--;
2230 peer_ni->ibp_accepting--;
2232 if (peer_ni->ibp_version == 0) {
2233 peer_ni->ibp_version = conn->ibc_version;
2234 peer_ni->ibp_incarnation = conn->ibc_incarnation;
2237 if (peer_ni->ibp_version != conn->ibc_version ||
2238 peer_ni->ibp_incarnation != conn->ibc_incarnation) {
2239 kiblnd_close_stale_conns_locked(peer_ni, conn->ibc_version,
2240 conn->ibc_incarnation);
2241 peer_ni->ibp_version = conn->ibc_version;
2242 peer_ni->ibp_incarnation = conn->ibc_incarnation;
2245 /* grab pending txs while I have the lock */
2246 list_splice_init(&peer_ni->ibp_tx_queue, &txs);
2248 if (!kiblnd_peer_active(peer_ni) || /* peer_ni has been deleted */
2249 conn->ibc_comms_error != 0) { /* error has happened already */
2251 /* start to shut down connection */
2252 kiblnd_close_conn_locked(conn, -ECONNABORTED);
2253 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2255 kiblnd_txlist_done(&txs, -ECONNABORTED,
2256 LNET_MSG_STATUS_LOCAL_ERROR);
2261 /* +1 ref for myself, this connection is visible to other threads
2262 * now, refcount of peer:ibp_conns can be released by connection
2263 * close from either a different thread, or the calling of
2264 * kiblnd_check_sends_locked() below. See bz21911 for details.
2266 kiblnd_conn_addref(conn);
2267 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2269 /* Schedule blocked txs
2270 * Note: if we are running with conns_per_peer > 1, these blocked
2271 * txs will all get scheduled to the first connection which gets
2272 * scheduled. We won't be using round robin on this first batch.
2274 spin_lock(&conn->ibc_lock);
2275 while (!list_empty(&txs)) {
2276 tx = list_entry(txs.next, struct kib_tx, tx_list);
2277 list_del(&tx->tx_list);
2279 kiblnd_queue_tx_locked(tx, conn);
2281 kiblnd_check_sends_locked(conn);
2282 spin_unlock(&conn->ibc_lock);
2284 /* schedule blocked rxs */
2285 kiblnd_handle_early_rxs(conn);
2286 kiblnd_conn_decref(conn);
2290 kiblnd_reject(struct rdma_cm_id *cmid, struct kib_rej *rej)
2294 rc = rdma_reject(cmid, rej, sizeof(*rej));
2297 CWARN("Error %d sending reject\n", rc);
2301 kiblnd_passive_connect(struct rdma_cm_id *cmid, void *priv, int priv_nob)
2303 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2304 struct kib_msg *reqmsg = priv;
2305 struct kib_msg *ackmsg;
2306 struct kib_dev *ibdev;
2307 struct kib_peer_ni *peer_ni;
2308 struct kib_peer_ni *peer2;
2309 struct kib_conn *conn;
2310 struct lnet_ni *ni = NULL;
2311 struct kib_net *net = NULL;
2313 struct rdma_conn_param cp;
2315 int version = IBLND_MSG_VERSION;
2316 unsigned long flags;
2318 struct sockaddr_in *peer_addr;
2319 LASSERT (!in_interrupt());
2321 /* cmid inherits 'context' from the corresponding listener id */
2322 ibdev = cmid->context;
2325 memset(&rej, 0, sizeof(rej));
2326 rej.ibr_magic = IBLND_MSG_MAGIC;
2327 rej.ibr_why = IBLND_REJECT_FATAL;
2328 rej.ibr_cp.ibcp_max_msg_size = IBLND_MSG_SIZE;
2330 peer_addr = (struct sockaddr_in *)&(cmid->route.addr.dst_addr);
2331 if (*kiblnd_tunables.kib_require_priv_port &&
2332 ntohs(peer_addr->sin_port) >= PROT_SOCK) {
2333 __u32 ip = ntohl(peer_addr->sin_addr.s_addr);
2334 CERROR("peer_ni's port (%pI4h:%hu) is not privileged\n",
2335 &ip, ntohs(peer_addr->sin_port));
2339 if (priv_nob < offsetof(struct kib_msg, ibm_type)) {
2340 CERROR("Short connection request\n");
2344 /* Future protocol version compatibility support! If the
2345 * o2iblnd-specific protocol changes, or when LNET unifies
2346 * protocols over all LNDs, the initial connection will
2347 * negotiate a protocol version. I trap this here to avoid
2348 * console errors; the reject tells the peer_ni which protocol I
2350 if (reqmsg->ibm_magic == LNET_PROTO_MAGIC ||
2351 reqmsg->ibm_magic == __swab32(LNET_PROTO_MAGIC))
2353 if (reqmsg->ibm_magic == IBLND_MSG_MAGIC &&
2354 reqmsg->ibm_version != IBLND_MSG_VERSION &&
2355 reqmsg->ibm_version != IBLND_MSG_VERSION_1)
2357 if (reqmsg->ibm_magic == __swab32(IBLND_MSG_MAGIC) &&
2358 reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION) &&
2359 reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION_1))
2362 rc = kiblnd_unpack_msg(reqmsg, priv_nob);
2364 CERROR("Can't parse connection request: %d\n", rc);
2368 nid = reqmsg->ibm_srcnid;
2369 ni = lnet_nid2ni_addref(reqmsg->ibm_dstnid);
2372 net = (struct kib_net *)ni->ni_data;
2373 rej.ibr_incarnation = net->ibn_incarnation;
2376 if (ni == NULL || /* no matching net */
2377 ni->ni_nid != reqmsg->ibm_dstnid || /* right NET, wrong NID! */
2378 net->ibn_dev != ibdev) { /* wrong device */
2379 CERROR("Can't accept conn from %s on %s (%s:%d:%pI4h): "
2380 "bad dst nid %s\n", libcfs_nid2str(nid),
2381 ni == NULL ? "NA" : libcfs_nid2str(ni->ni_nid),
2382 ibdev->ibd_ifname, ibdev->ibd_nnets,
2384 libcfs_nid2str(reqmsg->ibm_dstnid));
2389 /* check time stamp as soon as possible */
2390 if (reqmsg->ibm_dststamp != 0 &&
2391 reqmsg->ibm_dststamp != net->ibn_incarnation) {
2392 CWARN("Stale connection request\n");
2393 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2397 /* I can accept peer_ni's version */
2398 version = reqmsg->ibm_version;
2400 if (reqmsg->ibm_type != IBLND_MSG_CONNREQ) {
2401 CERROR("Unexpected connreq msg type: %x from %s\n",
2402 reqmsg->ibm_type, libcfs_nid2str(nid));
2406 if (reqmsg->ibm_u.connparams.ibcp_queue_depth >
2407 kiblnd_msg_queue_size(version, ni)) {
2408 CERROR("Can't accept conn from %s, queue depth too large: "
2409 " %d (<=%d wanted)\n",
2410 libcfs_nid2str(nid),
2411 reqmsg->ibm_u.connparams.ibcp_queue_depth,
2412 kiblnd_msg_queue_size(version, ni));
2414 if (version == IBLND_MSG_VERSION)
2415 rej.ibr_why = IBLND_REJECT_MSG_QUEUE_SIZE;
2420 if (reqmsg->ibm_u.connparams.ibcp_max_frags >
2421 IBLND_MAX_RDMA_FRAGS) {
2422 CWARN("Can't accept conn from %s (version %x): "
2423 "max_frags %d too large (%d wanted)\n",
2424 libcfs_nid2str(nid), version,
2425 reqmsg->ibm_u.connparams.ibcp_max_frags,
2426 IBLND_MAX_RDMA_FRAGS);
2428 if (version >= IBLND_MSG_VERSION)
2429 rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2432 } else if (reqmsg->ibm_u.connparams.ibcp_max_frags <
2433 IBLND_MAX_RDMA_FRAGS &&
2434 net->ibn_fmr_ps == NULL) {
2435 CWARN("Can't accept conn from %s (version %x): "
2436 "max_frags %d incompatible without FMR pool "
2438 libcfs_nid2str(nid), version,
2439 reqmsg->ibm_u.connparams.ibcp_max_frags,
2440 IBLND_MAX_RDMA_FRAGS);
2442 if (version == IBLND_MSG_VERSION)
2443 rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2448 if (reqmsg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2449 CERROR("Can't accept %s: message size %d too big (%d max)\n",
2450 libcfs_nid2str(nid),
2451 reqmsg->ibm_u.connparams.ibcp_max_msg_size,
2456 /* assume 'nid' is a new peer_ni; create */
2457 rc = kiblnd_create_peer(ni, &peer_ni, nid);
2459 CERROR("Can't create peer_ni for %s\n", libcfs_nid2str(nid));
2460 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2464 /* We have validated the peer's parameters so use those */
2465 peer_ni->ibp_max_frags = reqmsg->ibm_u.connparams.ibcp_max_frags;
2466 peer_ni->ibp_queue_depth = reqmsg->ibm_u.connparams.ibcp_queue_depth;
2468 write_lock_irqsave(g_lock, flags);
2470 peer2 = kiblnd_find_peer_locked(ni, nid);
2471 if (peer2 != NULL) {
2472 if (peer2->ibp_version == 0) {
2473 peer2->ibp_version = version;
2474 peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2477 /* not the guy I've talked with */
2478 if (peer2->ibp_incarnation != reqmsg->ibm_srcstamp ||
2479 peer2->ibp_version != version) {
2480 kiblnd_close_peer_conns_locked(peer2, -ESTALE);
2482 if (kiblnd_peer_active(peer2)) {
2483 peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2484 peer2->ibp_version = version;
2486 write_unlock_irqrestore(g_lock, flags);
2488 CWARN("Conn stale %s version %x/%x incarnation %llu/%llu\n",
2489 libcfs_nid2str(nid), peer2->ibp_version, version,
2490 peer2->ibp_incarnation, reqmsg->ibm_srcstamp);
2492 kiblnd_peer_decref(peer_ni);
2493 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2497 /* Tie-break connection race in favour of the higher NID.
2498 * If we keep running into a race condition multiple times,
2499 * we have to assume that the connection attempt with the
2500 * higher NID is stuck in a connecting state and will never
2501 * recover. As such, we pass through this if-block and let
2502 * the lower NID connection win so we can move forward.
2504 if (peer2->ibp_connecting != 0 &&
2505 nid < ni->ni_nid && peer2->ibp_races <
2506 MAX_CONN_RACES_BEFORE_ABORT) {
2508 write_unlock_irqrestore(g_lock, flags);
2510 CDEBUG(D_NET, "Conn race %s\n",
2511 libcfs_nid2str(peer2->ibp_nid));
2513 kiblnd_peer_decref(peer_ni);
2514 rej.ibr_why = IBLND_REJECT_CONN_RACE;
2517 if (peer2->ibp_races >= MAX_CONN_RACES_BEFORE_ABORT)
2518 CNETERR("Conn race %s: unresolved after %d attempts, letting lower NID win\n",
2519 libcfs_nid2str(peer2->ibp_nid),
2520 MAX_CONN_RACES_BEFORE_ABORT);
2522 * passive connection is allowed even this peer_ni is waiting for
2525 peer2->ibp_reconnecting = 0;
2526 peer2->ibp_races = 0;
2527 peer2->ibp_accepting++;
2528 kiblnd_peer_addref(peer2);
2530 /* Race with kiblnd_launch_tx (active connect) to create peer_ni
2531 * so copy validated parameters since we now know what the
2532 * peer_ni's limits are */
2533 peer2->ibp_max_frags = peer_ni->ibp_max_frags;
2534 peer2->ibp_queue_depth = peer_ni->ibp_queue_depth;
2536 write_unlock_irqrestore(g_lock, flags);
2537 kiblnd_peer_decref(peer_ni);
2540 /* Brand new peer_ni */
2541 LASSERT (peer_ni->ibp_accepting == 0);
2542 LASSERT (peer_ni->ibp_version == 0 &&
2543 peer_ni->ibp_incarnation == 0);
2545 peer_ni->ibp_accepting = 1;
2546 peer_ni->ibp_version = version;
2547 peer_ni->ibp_incarnation = reqmsg->ibm_srcstamp;
2549 /* I have a ref on ni that prevents it being shutdown */
2550 LASSERT (net->ibn_shutdown == 0);
2552 kiblnd_peer_addref(peer_ni);
2553 list_add_tail(&peer_ni->ibp_list, kiblnd_nid2peerlist(nid));
2555 write_unlock_irqrestore(g_lock, flags);
2558 conn = kiblnd_create_conn(peer_ni, cmid, IBLND_CONN_PASSIVE_WAIT, version);
2560 kiblnd_peer_connect_failed(peer_ni, 0, -ENOMEM);
2561 kiblnd_peer_decref(peer_ni);
2562 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2566 /* conn now "owns" cmid, so I return success from here on to ensure the
2567 * CM callback doesn't destroy cmid. */
2568 conn->ibc_incarnation = reqmsg->ibm_srcstamp;
2569 conn->ibc_credits = conn->ibc_queue_depth;
2570 conn->ibc_reserved_credits = conn->ibc_queue_depth;
2571 LASSERT(conn->ibc_credits + conn->ibc_reserved_credits +
2572 IBLND_OOB_MSGS(version) <= IBLND_RX_MSGS(conn));
2574 ackmsg = &conn->ibc_connvars->cv_msg;
2575 memset(ackmsg, 0, sizeof(*ackmsg));
2577 kiblnd_init_msg(ackmsg, IBLND_MSG_CONNACK,
2578 sizeof(ackmsg->ibm_u.connparams));
2579 ackmsg->ibm_u.connparams.ibcp_queue_depth = conn->ibc_queue_depth;
2580 ackmsg->ibm_u.connparams.ibcp_max_frags = conn->ibc_max_frags;
2581 ackmsg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2583 kiblnd_pack_msg(ni, ackmsg, version, 0, nid, reqmsg->ibm_srcstamp);
2585 memset(&cp, 0, sizeof(cp));
2586 cp.private_data = ackmsg;
2587 cp.private_data_len = ackmsg->ibm_nob;
2588 cp.responder_resources = 0; /* No atomic ops or RDMA reads */
2589 cp.initiator_depth = 0;
2590 cp.flow_control = 1;
2591 cp.retry_count = *kiblnd_tunables.kib_retry_count;
2592 cp.rnr_retry_count = *kiblnd_tunables.kib_rnr_retry_count;
2594 CDEBUG(D_NET, "Accept %s\n", libcfs_nid2str(nid));
2596 rc = rdma_accept(cmid, &cp);
2598 CERROR("Can't accept %s: %d\n", libcfs_nid2str(nid), rc);
2599 rej.ibr_version = version;
2600 rej.ibr_why = IBLND_REJECT_FATAL;
2602 kiblnd_reject(cmid, &rej);
2603 kiblnd_connreq_done(conn, rc);
2604 kiblnd_conn_decref(conn);
2612 rej.ibr_cp.ibcp_queue_depth =
2613 kiblnd_msg_queue_size(version, ni);
2614 rej.ibr_cp.ibcp_max_frags = IBLND_MAX_RDMA_FRAGS;
2618 rej.ibr_version = version;
2619 kiblnd_reject(cmid, &rej);
2621 return -ECONNREFUSED;
2625 kiblnd_check_reconnect(struct kib_conn *conn, int version,
2626 u64 incarnation, int why, struct kib_connparams *cp)
2628 rwlock_t *glock = &kiblnd_data.kib_global_lock;
2629 struct kib_peer_ni *peer_ni = conn->ibc_peer;
2631 int msg_size = IBLND_MSG_SIZE;
2635 unsigned long flags;
2637 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2638 LASSERT(peer_ni->ibp_connecting > 0); /* 'conn' at least */
2641 msg_size = cp->ibcp_max_msg_size;
2642 frag_num = cp->ibcp_max_frags;
2643 queue_dep = cp->ibcp_queue_depth;
2646 write_lock_irqsave(glock, flags);
2647 /* retry connection if it's still needed and no other connection
2648 * attempts (active or passive) are in progress
2649 * NB: reconnect is still needed even when ibp_tx_queue is
2650 * empty if ibp_version != version because reconnect may be
2653 reconnect = (!list_empty(&peer_ni->ibp_tx_queue) ||
2654 peer_ni->ibp_version != version) &&
2655 peer_ni->ibp_connecting &&
2656 peer_ni->ibp_accepting == 0;
2667 case IBLND_REJECT_RDMA_FRAGS: {
2668 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2671 reason = "can't negotiate max frags";
2674 tunables = &peer_ni->ibp_ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2675 #ifdef HAVE_IB_GET_DMA_MR
2677 * This check only makes sense if the kernel supports global
2678 * memory registration. Otherwise, map_on_demand will never == 0
2680 if (!tunables->lnd_map_on_demand) {
2681 reason = "map_on_demand must be enabled";
2685 if (conn->ibc_max_frags <= frag_num) {
2686 reason = "unsupported max frags";
2690 peer_ni->ibp_max_frags = frag_num;
2691 reason = "rdma fragments";
2694 case IBLND_REJECT_MSG_QUEUE_SIZE:
2696 reason = "can't negotiate queue depth";
2699 if (conn->ibc_queue_depth <= queue_dep) {
2700 reason = "unsupported queue depth";
2704 peer_ni->ibp_queue_depth = queue_dep;
2705 reason = "queue depth";
2708 case IBLND_REJECT_CONN_STALE:
2712 case IBLND_REJECT_CONN_RACE:
2713 reason = "conn race";
2716 case IBLND_REJECT_CONN_UNCOMPAT:
2717 reason = "version negotiation";
2720 case IBLND_REJECT_INVALID_SRV_ID:
2721 reason = "invalid service id";
2725 conn->ibc_reconnect = 1;
2726 peer_ni->ibp_reconnecting++;
2727 peer_ni->ibp_version = version;
2728 if (incarnation != 0)
2729 peer_ni->ibp_incarnation = incarnation;
2731 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2733 CNETERR("%s: %s (%s), %x, %x, msg_size: %d, queue_depth: %d/%d, max_frags: %d/%d\n",
2734 libcfs_nid2str(peer_ni->ibp_nid),
2735 reconnect ? "reconnect" : "don't reconnect",
2736 reason, IBLND_MSG_VERSION, version, msg_size,
2737 conn->ibc_queue_depth, queue_dep,
2738 conn->ibc_max_frags, frag_num);
2740 * if conn::ibc_reconnect is TRUE, connd will reconnect to the peer_ni
2741 * while destroying the zombie
2746 kiblnd_rejected(struct kib_conn *conn, int reason, void *priv, int priv_nob)
2748 struct kib_peer_ni *peer_ni = conn->ibc_peer;
2750 LASSERT (!in_interrupt());
2751 LASSERT (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2754 case IB_CM_REJ_STALE_CONN:
2755 kiblnd_check_reconnect(conn, IBLND_MSG_VERSION, 0,
2756 IBLND_REJECT_CONN_STALE, NULL);
2759 case IB_CM_REJ_INVALID_SERVICE_ID:
2760 kiblnd_check_reconnect(conn, IBLND_MSG_VERSION, 0,
2761 IBLND_REJECT_INVALID_SRV_ID, NULL);
2762 CNETERR("%s rejected: no listener at %d\n",
2763 libcfs_nid2str(peer_ni->ibp_nid),
2764 *kiblnd_tunables.kib_service);
2767 case IB_CM_REJ_CONSUMER_DEFINED:
2768 if (priv_nob >= offsetof(struct kib_rej, ibr_padding)) {
2769 struct kib_rej *rej = priv;
2770 struct kib_connparams *cp = NULL;
2772 __u64 incarnation = -1;
2774 /* NB. default incarnation is -1 because:
2775 * a) V1 will ignore dst incarnation in connreq.
2776 * b) V2 will provide incarnation while rejecting me,
2777 * -1 will be overwrote.
2779 * if I try to connect to a V1 peer_ni with V2 protocol,
2780 * it rejected me then upgrade to V2, I have no idea
2781 * about the upgrading and try to reconnect with V1,
2782 * in this case upgraded V2 can find out I'm trying to
2783 * talk to the old guy and reject me(incarnation is -1).
2786 if (rej->ibr_magic == __swab32(IBLND_MSG_MAGIC) ||
2787 rej->ibr_magic == __swab32(LNET_PROTO_MAGIC)) {
2788 __swab32s(&rej->ibr_magic);
2789 __swab16s(&rej->ibr_version);
2793 if (priv_nob >= sizeof(struct kib_rej) &&
2794 rej->ibr_version > IBLND_MSG_VERSION_1) {
2795 /* priv_nob is always 148 in current version
2796 * of OFED, so we still need to check version.
2797 * (define of IB_CM_REJ_PRIVATE_DATA_SIZE) */
2801 __swab64s(&rej->ibr_incarnation);
2802 __swab16s(&cp->ibcp_queue_depth);
2803 __swab16s(&cp->ibcp_max_frags);
2804 __swab32s(&cp->ibcp_max_msg_size);
2807 incarnation = rej->ibr_incarnation;
2810 if (rej->ibr_magic != IBLND_MSG_MAGIC &&
2811 rej->ibr_magic != LNET_PROTO_MAGIC) {
2812 CERROR("%s rejected: consumer defined fatal error\n",
2813 libcfs_nid2str(peer_ni->ibp_nid));
2817 if (rej->ibr_version != IBLND_MSG_VERSION &&
2818 rej->ibr_version != IBLND_MSG_VERSION_1) {
2819 CERROR("%s rejected: o2iblnd version %x error\n",
2820 libcfs_nid2str(peer_ni->ibp_nid),
2825 if (rej->ibr_why == IBLND_REJECT_FATAL &&
2826 rej->ibr_version == IBLND_MSG_VERSION_1) {
2827 CDEBUG(D_NET, "rejected by old version peer_ni %s: %x\n",
2828 libcfs_nid2str(peer_ni->ibp_nid), rej->ibr_version);
2830 if (conn->ibc_version != IBLND_MSG_VERSION_1)
2831 rej->ibr_why = IBLND_REJECT_CONN_UNCOMPAT;
2834 switch (rej->ibr_why) {
2835 case IBLND_REJECT_CONN_RACE:
2836 case IBLND_REJECT_CONN_STALE:
2837 case IBLND_REJECT_CONN_UNCOMPAT:
2838 case IBLND_REJECT_MSG_QUEUE_SIZE:
2839 case IBLND_REJECT_RDMA_FRAGS:
2840 kiblnd_check_reconnect(conn, rej->ibr_version,
2841 incarnation, rej->ibr_why, cp);
2844 case IBLND_REJECT_NO_RESOURCES:
2845 CERROR("%s rejected: o2iblnd no resources\n",
2846 libcfs_nid2str(peer_ni->ibp_nid));
2849 case IBLND_REJECT_FATAL:
2850 CERROR("%s rejected: o2iblnd fatal error\n",
2851 libcfs_nid2str(peer_ni->ibp_nid));
2855 CERROR("%s rejected: o2iblnd reason %d\n",
2856 libcfs_nid2str(peer_ni->ibp_nid),
2864 CNETERR("%s rejected: reason %d, size %d\n",
2865 libcfs_nid2str(peer_ni->ibp_nid), reason, priv_nob);
2869 kiblnd_connreq_done(conn, -ECONNREFUSED);
2873 kiblnd_check_connreply(struct kib_conn *conn, void *priv, int priv_nob)
2875 struct kib_peer_ni *peer_ni = conn->ibc_peer;
2876 struct lnet_ni *ni = peer_ni->ibp_ni;
2877 struct kib_net *net = ni->ni_data;
2878 struct kib_msg *msg = priv;
2879 int ver = conn->ibc_version;
2880 int rc = kiblnd_unpack_msg(msg, priv_nob);
2881 unsigned long flags;
2883 LASSERT (net != NULL);
2886 CERROR("Can't unpack connack from %s: %d\n",
2887 libcfs_nid2str(peer_ni->ibp_nid), rc);
2891 if (msg->ibm_type != IBLND_MSG_CONNACK) {
2892 CERROR("Unexpected message %d from %s\n",
2893 msg->ibm_type, libcfs_nid2str(peer_ni->ibp_nid));
2898 if (ver != msg->ibm_version) {
2899 CERROR("%s replied version %x is different with "
2900 "requested version %x\n",
2901 libcfs_nid2str(peer_ni->ibp_nid), msg->ibm_version, ver);
2906 if (msg->ibm_u.connparams.ibcp_queue_depth >
2907 conn->ibc_queue_depth) {
2908 CERROR("%s has incompatible queue depth %d (<=%d wanted)\n",
2909 libcfs_nid2str(peer_ni->ibp_nid),
2910 msg->ibm_u.connparams.ibcp_queue_depth,
2911 conn->ibc_queue_depth);
2916 if (msg->ibm_u.connparams.ibcp_max_frags >
2917 conn->ibc_max_frags) {
2918 CERROR("%s has incompatible max_frags %d (<=%d wanted)\n",
2919 libcfs_nid2str(peer_ni->ibp_nid),
2920 msg->ibm_u.connparams.ibcp_max_frags,
2921 conn->ibc_max_frags);
2926 if (msg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2927 CERROR("%s max message size %d too big (%d max)\n",
2928 libcfs_nid2str(peer_ni->ibp_nid),
2929 msg->ibm_u.connparams.ibcp_max_msg_size,
2935 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2936 if (msg->ibm_dstnid == ni->ni_nid &&
2937 msg->ibm_dststamp == net->ibn_incarnation)
2941 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2944 CERROR("Bad connection reply from %s, rc = %d, "
2945 "version: %x max_frags: %d\n",
2946 libcfs_nid2str(peer_ni->ibp_nid), rc,
2947 msg->ibm_version, msg->ibm_u.connparams.ibcp_max_frags);
2951 conn->ibc_incarnation = msg->ibm_srcstamp;
2952 conn->ibc_credits = msg->ibm_u.connparams.ibcp_queue_depth;
2953 conn->ibc_reserved_credits = msg->ibm_u.connparams.ibcp_queue_depth;
2954 conn->ibc_queue_depth = msg->ibm_u.connparams.ibcp_queue_depth;
2955 conn->ibc_max_frags = msg->ibm_u.connparams.ibcp_max_frags;
2956 LASSERT(conn->ibc_credits + conn->ibc_reserved_credits +
2957 IBLND_OOB_MSGS(ver) <= IBLND_RX_MSGS(conn));
2959 kiblnd_connreq_done(conn, 0);
2963 /* NB My QP has already established itself, so I handle anything going
2964 * wrong here by setting ibc_comms_error.
2965 * kiblnd_connreq_done(0) moves the conn state to ESTABLISHED, but then
2966 * immediately tears it down. */
2969 conn->ibc_comms_error = rc;
2970 kiblnd_connreq_done(conn, 0);
2974 kiblnd_active_connect(struct rdma_cm_id *cmid)
2976 struct kib_peer_ni *peer_ni = cmid->context;
2977 struct kib_conn *conn;
2978 struct kib_msg *msg;
2979 struct rdma_conn_param cp;
2982 unsigned long flags;
2985 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2987 incarnation = peer_ni->ibp_incarnation;
2988 version = (peer_ni->ibp_version == 0) ? IBLND_MSG_VERSION :
2989 peer_ni->ibp_version;
2991 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2993 conn = kiblnd_create_conn(peer_ni, cmid, IBLND_CONN_ACTIVE_CONNECT,
2996 kiblnd_peer_connect_failed(peer_ni, 1, -ENOMEM);
2997 kiblnd_peer_decref(peer_ni); /* lose cmid's ref */
3001 /* conn "owns" cmid now, so I return success from here on to ensure the
3002 * CM callback doesn't destroy cmid. conn also takes over cmid's ref
3005 msg = &conn->ibc_connvars->cv_msg;
3007 memset(msg, 0, sizeof(*msg));
3008 kiblnd_init_msg(msg, IBLND_MSG_CONNREQ, sizeof(msg->ibm_u.connparams));
3009 msg->ibm_u.connparams.ibcp_queue_depth = conn->ibc_queue_depth;
3010 msg->ibm_u.connparams.ibcp_max_frags = conn->ibc_max_frags;
3011 msg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
3013 kiblnd_pack_msg(peer_ni->ibp_ni, msg, version,
3014 0, peer_ni->ibp_nid, incarnation);
3016 memset(&cp, 0, sizeof(cp));
3017 cp.private_data = msg;
3018 cp.private_data_len = msg->ibm_nob;
3019 cp.responder_resources = 0; /* No atomic ops or RDMA reads */
3020 cp.initiator_depth = 0;
3021 cp.flow_control = 1;
3022 cp.retry_count = *kiblnd_tunables.kib_retry_count;
3023 cp.rnr_retry_count = *kiblnd_tunables.kib_rnr_retry_count;
3025 LASSERT(cmid->context == (void *)conn);
3026 LASSERT(conn->ibc_cmid == cmid);
3028 rc = rdma_connect(cmid, &cp);
3030 CERROR("Can't connect to %s: %d\n",
3031 libcfs_nid2str(peer_ni->ibp_nid), rc);
3032 kiblnd_connreq_done(conn, rc);
3033 kiblnd_conn_decref(conn);
3040 kiblnd_cm_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
3042 struct kib_peer_ni *peer_ni;
3043 struct kib_conn *conn;
3046 switch (event->event) {
3048 CERROR("Unexpected event: %d, status: %d\n",
3049 event->event, event->status);
3052 case RDMA_CM_EVENT_CONNECT_REQUEST:
3053 /* destroy cmid on failure */
3054 rc = kiblnd_passive_connect(cmid,
3055 (void *)KIBLND_CONN_PARAM(event),
3056 KIBLND_CONN_PARAM_LEN(event));
3057 CDEBUG(D_NET, "connreq: %d\n", rc);
3060 case RDMA_CM_EVENT_ADDR_ERROR:
3061 peer_ni = cmid->context;
3062 CNETERR("%s: ADDR ERROR %d\n",
3063 libcfs_nid2str(peer_ni->ibp_nid), event->status);
3064 kiblnd_peer_connect_failed(peer_ni, 1, -EHOSTUNREACH);
3065 kiblnd_peer_decref(peer_ni);
3066 return -EHOSTUNREACH; /* rc != 0 destroys cmid */
3068 case RDMA_CM_EVENT_ADDR_RESOLVED:
3069 peer_ni = cmid->context;
3071 CDEBUG(D_NET,"%s Addr resolved: %d\n",
3072 libcfs_nid2str(peer_ni->ibp_nid), event->status);
3074 if (event->status != 0) {
3075 CNETERR("Can't resolve address for %s: %d\n",
3076 libcfs_nid2str(peer_ni->ibp_nid), event->status);
3079 rc = rdma_resolve_route(
3080 cmid, kiblnd_timeout() * 1000);
3082 struct kib_net *net = peer_ni->ibp_ni->ni_data;
3083 struct kib_dev *dev = net->ibn_dev;
3085 CDEBUG(D_NET, "%s: connection bound to "\
3087 libcfs_nid2str(peer_ni->ibp_nid),
3089 &dev->ibd_ifip, cmid->device->name);
3094 /* Can't initiate route resolution */
3095 CERROR("Can't resolve route for %s: %d\n",
3096 libcfs_nid2str(peer_ni->ibp_nid), rc);
3098 kiblnd_peer_connect_failed(peer_ni, 1, rc);
3099 kiblnd_peer_decref(peer_ni);
3100 return rc; /* rc != 0 destroys cmid */
3102 case RDMA_CM_EVENT_ROUTE_ERROR:
3103 peer_ni = cmid->context;
3104 CNETERR("%s: ROUTE ERROR %d\n",
3105 libcfs_nid2str(peer_ni->ibp_nid), event->status);
3106 kiblnd_peer_connect_failed(peer_ni, 1, -EHOSTUNREACH);
3107 kiblnd_peer_decref(peer_ni);
3108 return -EHOSTUNREACH; /* rc != 0 destroys cmid */
3110 case RDMA_CM_EVENT_ROUTE_RESOLVED:
3111 peer_ni = cmid->context;
3112 CDEBUG(D_NET,"%s Route resolved: %d\n",
3113 libcfs_nid2str(peer_ni->ibp_nid), event->status);
3115 if (event->status == 0)
3116 return kiblnd_active_connect(cmid);
3118 CNETERR("Can't resolve route for %s: %d\n",
3119 libcfs_nid2str(peer_ni->ibp_nid), event->status);
3120 kiblnd_peer_connect_failed(peer_ni, 1, event->status);
3121 kiblnd_peer_decref(peer_ni);
3122 return event->status; /* rc != 0 destroys cmid */
3124 case RDMA_CM_EVENT_UNREACHABLE:
3125 conn = cmid->context;
3126 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
3127 conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
3128 CNETERR("%s: UNREACHABLE %d\n",
3129 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
3130 kiblnd_connreq_done(conn, -ENETDOWN);
3131 kiblnd_conn_decref(conn);
3134 case RDMA_CM_EVENT_CONNECT_ERROR:
3135 conn = cmid->context;
3136 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
3137 conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
3138 CNETERR("%s: CONNECT ERROR %d\n",
3139 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
3140 kiblnd_connreq_done(conn, -ENOTCONN);
3141 kiblnd_conn_decref(conn);
3144 case RDMA_CM_EVENT_REJECTED:
3145 conn = cmid->context;
3146 switch (conn->ibc_state) {
3150 case IBLND_CONN_PASSIVE_WAIT:
3151 CERROR ("%s: REJECTED %d\n",
3152 libcfs_nid2str(conn->ibc_peer->ibp_nid),
3154 kiblnd_connreq_done(conn, -ECONNRESET);
3157 case IBLND_CONN_ACTIVE_CONNECT:
3158 kiblnd_rejected(conn, event->status,
3159 (void *)KIBLND_CONN_PARAM(event),
3160 KIBLND_CONN_PARAM_LEN(event));
3163 kiblnd_conn_decref(conn);
3166 case RDMA_CM_EVENT_ESTABLISHED:
3167 conn = cmid->context;
3168 switch (conn->ibc_state) {
3172 case IBLND_CONN_PASSIVE_WAIT:
3173 CDEBUG(D_NET, "ESTABLISHED (passive): %s\n",
3174 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3175 kiblnd_connreq_done(conn, 0);
3178 case IBLND_CONN_ACTIVE_CONNECT:
3179 CDEBUG(D_NET, "ESTABLISHED(active): %s\n",
3180 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3181 kiblnd_check_connreply(conn,
3182 (void *)KIBLND_CONN_PARAM(event),
3183 KIBLND_CONN_PARAM_LEN(event));
3186 /* net keeps its ref on conn! */
3189 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
3190 CDEBUG(D_NET, "Ignore TIMEWAIT_EXIT event\n");
3193 case RDMA_CM_EVENT_DISCONNECTED:
3194 conn = cmid->context;
3195 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
3196 CERROR("%s DISCONNECTED\n",
3197 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3198 kiblnd_connreq_done(conn, -ECONNRESET);
3200 kiblnd_close_conn(conn, 0);
3202 kiblnd_conn_decref(conn);
3203 cmid->context = NULL;
3206 case RDMA_CM_EVENT_DEVICE_REMOVAL:
3207 LCONSOLE_ERROR_MSG(0x131,
3208 "Received notification of device removal\n"
3209 "Please shutdown LNET to allow this to proceed\n");
3210 /* Can't remove network from underneath LNET for now, so I have
3214 case RDMA_CM_EVENT_ADDR_CHANGE:
3215 LCONSOLE_INFO("Physical link changed (eg hca/port)\n");
3221 kiblnd_check_txs_locked(struct kib_conn *conn, struct list_head *txs)
3224 struct list_head *ttmp;
3226 list_for_each(ttmp, txs) {
3227 tx = list_entry(ttmp, struct kib_tx, tx_list);
3229 if (txs != &conn->ibc_active_txs) {
3230 LASSERT(tx->tx_queued);
3232 LASSERT(!tx->tx_queued);
3233 LASSERT(tx->tx_waiting || tx->tx_sending != 0);
3236 if (ktime_compare(ktime_get(), tx->tx_deadline) >= 0) {
3237 CERROR("Timed out tx: %s, %lld seconds\n",
3238 kiblnd_queue2str(conn, txs),
3239 ktime_ms_delta(ktime_get(),
3240 tx->tx_deadline) / MSEC_PER_SEC);
3249 kiblnd_conn_timed_out_locked(struct kib_conn *conn)
3251 return kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue) ||
3252 kiblnd_check_txs_locked(conn, &conn->ibc_tx_noops) ||
3253 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_rsrvd) ||
3254 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_nocred) ||
3255 kiblnd_check_txs_locked(conn, &conn->ibc_active_txs);
3259 kiblnd_check_conns (int idx)
3262 LIST_HEAD(checksends);
3263 LIST_HEAD(timedout_txs);
3264 struct list_head *peers = &kiblnd_data.kib_peers[idx];
3265 struct list_head *ptmp;
3266 struct kib_peer_ni *peer_ni;
3267 struct kib_conn *conn;
3268 struct kib_tx *tx, *tx_tmp;
3269 struct list_head *ctmp;
3270 unsigned long flags;
3272 /* NB. We expect to have a look at all the peers and not find any
3273 * RDMAs to time out, so we just use a shared lock while we
3275 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3277 list_for_each(ptmp, peers) {
3278 peer_ni = list_entry(ptmp, struct kib_peer_ni, ibp_list);
3280 /* Check tx_deadline */
3281 list_for_each_entry_safe(tx, tx_tmp, &peer_ni->ibp_tx_queue, tx_list) {
3282 if (ktime_compare(ktime_get(), tx->tx_deadline) >= 0) {
3283 CWARN("Timed out tx for %s: %lld seconds\n",
3284 libcfs_nid2str(peer_ni->ibp_nid),
3285 ktime_ms_delta(ktime_get(),
3286 tx->tx_deadline) / MSEC_PER_SEC);
3287 list_move(&tx->tx_list, &timedout_txs);
3291 list_for_each(ctmp, &peer_ni->ibp_conns) {
3295 conn = list_entry(ctmp, struct kib_conn, ibc_list);
3297 LASSERT(conn->ibc_state == IBLND_CONN_ESTABLISHED);
3299 spin_lock(&conn->ibc_lock);
3301 sendnoop = kiblnd_need_noop(conn);
3302 timedout = kiblnd_conn_timed_out_locked(conn);
3303 if (!sendnoop && !timedout) {
3304 spin_unlock(&conn->ibc_lock);
3309 CERROR("Timed out RDMA with %s (%lld): "
3310 "c: %u, oc: %u, rc: %u\n",
3311 libcfs_nid2str(peer_ni->ibp_nid),
3312 ktime_get_seconds() - peer_ni->ibp_last_alive,
3314 conn->ibc_outstanding_credits,
3315 conn->ibc_reserved_credits);
3316 list_add(&conn->ibc_connd_list, &closes);
3318 list_add(&conn->ibc_connd_list, &checksends);
3320 /* +ref for 'closes' or 'checksends' */
3321 kiblnd_conn_addref(conn);
3323 spin_unlock(&conn->ibc_lock);
3327 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3329 if (!list_empty(&timedout_txs))
3330 kiblnd_txlist_done(&timedout_txs, -ETIMEDOUT,
3331 LNET_MSG_STATUS_LOCAL_TIMEOUT);
3333 /* Handle timeout by closing the whole
3334 * connection. We can only be sure RDMA activity
3335 * has ceased once the QP has been modified. */
3336 while (!list_empty(&closes)) {
3337 conn = list_entry(closes.next,
3338 struct kib_conn, ibc_connd_list);
3339 list_del(&conn->ibc_connd_list);
3340 kiblnd_close_conn(conn, -ETIMEDOUT);
3341 kiblnd_conn_decref(conn);
3344 /* In case we have enough credits to return via a
3345 * NOOP, but there were no non-blocking tx descs
3346 * free to do it last time... */
3347 while (!list_empty(&checksends)) {
3348 conn = list_entry(checksends.next,
3349 struct kib_conn, ibc_connd_list);
3350 list_del(&conn->ibc_connd_list);
3352 spin_lock(&conn->ibc_lock);
3353 kiblnd_check_sends_locked(conn);
3354 spin_unlock(&conn->ibc_lock);
3356 kiblnd_conn_decref(conn);
3361 kiblnd_disconnect_conn(struct kib_conn *conn)
3363 LASSERT (!in_interrupt());
3364 LASSERT (current == kiblnd_data.kib_connd);
3365 LASSERT (conn->ibc_state == IBLND_CONN_CLOSING);
3367 rdma_disconnect(conn->ibc_cmid);
3368 kiblnd_finalise_conn(conn);
3370 kiblnd_peer_notify(conn->ibc_peer);
3374 * High-water for reconnection to the same peer_ni, reconnection attempt should
3375 * be delayed after trying more than KIB_RECONN_HIGH_RACE.
3377 #define KIB_RECONN_HIGH_RACE 10
3379 * Allow connd to take a break and handle other things after consecutive
3380 * reconnection attemps.
3382 #define KIB_RECONN_BREAK 100
3385 kiblnd_connd (void *arg)
3387 spinlock_t *lock= &kiblnd_data.kib_connd_lock;
3388 wait_queue_entry_t wait;
3389 unsigned long flags;
3390 struct kib_conn *conn;
3395 unsigned long deadline = jiffies;
3397 init_waitqueue_entry(&wait, current);
3398 kiblnd_data.kib_connd = current;
3400 spin_lock_irqsave(lock, flags);
3402 while (!kiblnd_data.kib_shutdown) {
3407 if (!list_empty(&kiblnd_data.kib_connd_zombies)) {
3408 struct kib_peer_ni *peer_ni = NULL;
3410 conn = list_entry(kiblnd_data.kib_connd_zombies.next,
3411 struct kib_conn, ibc_list);
3412 list_del(&conn->ibc_list);
3413 if (conn->ibc_reconnect) {
3414 peer_ni = conn->ibc_peer;
3415 kiblnd_peer_addref(peer_ni);
3418 spin_unlock_irqrestore(lock, flags);
3421 kiblnd_destroy_conn(conn);
3423 spin_lock_irqsave(lock, flags);
3425 LIBCFS_FREE(conn, sizeof(*conn));
3429 conn->ibc_peer = peer_ni;
3430 if (peer_ni->ibp_reconnected < KIB_RECONN_HIGH_RACE)
3431 list_add_tail(&conn->ibc_list,
3432 &kiblnd_data.kib_reconn_list);
3434 list_add_tail(&conn->ibc_list,
3435 &kiblnd_data.kib_reconn_wait);
3438 if (!list_empty(&kiblnd_data.kib_connd_conns)) {
3439 conn = list_entry(kiblnd_data.kib_connd_conns.next,
3440 struct kib_conn, ibc_list);
3441 list_del(&conn->ibc_list);
3443 spin_unlock_irqrestore(lock, flags);
3446 kiblnd_disconnect_conn(conn);
3447 kiblnd_conn_decref(conn);
3449 spin_lock_irqsave(lock, flags);
3452 while (reconn < KIB_RECONN_BREAK) {
3453 if (kiblnd_data.kib_reconn_sec !=
3454 ktime_get_real_seconds()) {
3455 kiblnd_data.kib_reconn_sec = ktime_get_real_seconds();
3456 list_splice_init(&kiblnd_data.kib_reconn_wait,
3457 &kiblnd_data.kib_reconn_list);
3460 if (list_empty(&kiblnd_data.kib_reconn_list))
3463 conn = list_entry(kiblnd_data.kib_reconn_list.next,
3464 struct kib_conn, ibc_list);
3465 list_del(&conn->ibc_list);
3467 spin_unlock_irqrestore(lock, flags);
3470 reconn += kiblnd_reconnect_peer(conn->ibc_peer);
3471 kiblnd_peer_decref(conn->ibc_peer);
3472 LIBCFS_FREE(conn, sizeof(*conn));
3474 spin_lock_irqsave(lock, flags);
3477 /* careful with the jiffy wrap... */
3478 timeout = (int)(deadline - jiffies);
3482 int chunk = kiblnd_data.kib_peer_hash_size;
3483 unsigned int lnd_timeout;
3485 spin_unlock_irqrestore(lock, flags);
3488 /* Time to check for RDMA timeouts on a few more
3489 * peers: I do checks every 'p' seconds on a
3490 * proportion of the peer_ni table and I need to check
3491 * every connection 'n' times within a timeout
3492 * interval, to ensure I detect a timeout on any
3493 * connection within (n+1)/n times the timeout
3496 lnd_timeout = kiblnd_timeout();
3497 if (lnd_timeout > n * p)
3498 chunk = (chunk * n * p) / lnd_timeout;
3502 for (i = 0; i < chunk; i++) {
3503 kiblnd_check_conns(peer_index);
3504 peer_index = (peer_index + 1) %
3505 kiblnd_data.kib_peer_hash_size;
3508 deadline += cfs_time_seconds(p);
3509 spin_lock_irqsave(lock, flags);
3515 /* Nothing to do for 'timeout' */
3516 set_current_state(TASK_INTERRUPTIBLE);
3517 add_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3518 spin_unlock_irqrestore(lock, flags);
3520 schedule_timeout(timeout);
3522 remove_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3523 spin_lock_irqsave(lock, flags);
3526 spin_unlock_irqrestore(lock, flags);
3528 kiblnd_thread_fini();
3533 kiblnd_qp_event(struct ib_event *event, void *arg)
3535 struct kib_conn *conn = arg;
3537 switch (event->event) {
3538 case IB_EVENT_COMM_EST:
3539 CDEBUG(D_NET, "%s established\n",
3540 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3541 /* We received a packet but connection isn't established
3542 * probably handshake packet was lost, so free to
3543 * force make connection established */
3544 rdma_notify(conn->ibc_cmid, IB_EVENT_COMM_EST);
3547 case IB_EVENT_PORT_ERR:
3548 case IB_EVENT_DEVICE_FATAL:
3549 CERROR("Fatal device error for NI %s\n",
3550 libcfs_nid2str(conn->ibc_peer->ibp_ni->ni_nid));
3551 atomic_set(&conn->ibc_peer->ibp_ni->ni_fatal_error_on, 1);
3554 case IB_EVENT_PORT_ACTIVE:
3555 CERROR("Port reactivated for NI %s\n",
3556 libcfs_nid2str(conn->ibc_peer->ibp_ni->ni_nid));
3557 atomic_set(&conn->ibc_peer->ibp_ni->ni_fatal_error_on, 0);
3561 CERROR("%s: Async QP event type %d\n",
3562 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3568 kiblnd_complete (struct ib_wc *wc)
3570 switch (kiblnd_wreqid2type(wc->wr_id)) {
3575 if (wc->status != IB_WC_SUCCESS &&
3576 wc->status != IB_WC_WR_FLUSH_ERR)
3577 CNETERR("FastReg failed: %d\n", wc->status);
3580 case IBLND_WID_RDMA:
3581 /* We only get RDMA completion notification if it fails. All
3582 * subsequent work items, including the final SEND will fail
3583 * too. However we can't print out any more info about the
3584 * failing RDMA because 'tx' might be back on the idle list or
3585 * even reused already if we didn't manage to post all our work
3587 CNETERR("RDMA (tx: %p) failed: %d\n",
3588 kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3592 kiblnd_tx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3596 kiblnd_rx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status,
3603 kiblnd_cq_completion(struct ib_cq *cq, void *arg)
3605 /* NB I'm not allowed to schedule this conn once its refcount has
3606 * reached 0. Since fundamentally I'm racing with scheduler threads
3607 * consuming my CQ I could be called after all completions have
3608 * occurred. But in this case, ibc_nrx == 0 && ibc_nsends_posted == 0
3609 * and this CQ is about to be destroyed so I NOOP. */
3610 struct kib_conn *conn = arg;
3611 struct kib_sched_info *sched = conn->ibc_sched;
3612 unsigned long flags;
3614 LASSERT(cq == conn->ibc_cq);
3616 spin_lock_irqsave(&sched->ibs_lock, flags);
3618 conn->ibc_ready = 1;
3620 if (!conn->ibc_scheduled &&
3621 (conn->ibc_nrx > 0 ||
3622 conn->ibc_nsends_posted > 0)) {
3623 kiblnd_conn_addref(conn); /* +1 ref for sched_conns */
3624 conn->ibc_scheduled = 1;
3625 list_add_tail(&conn->ibc_sched_list, &sched->ibs_conns);
3627 if (waitqueue_active(&sched->ibs_waitq))
3628 wake_up(&sched->ibs_waitq);
3631 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3635 kiblnd_cq_event(struct ib_event *event, void *arg)
3637 struct kib_conn *conn = arg;
3639 CERROR("%s: async CQ event type %d\n",
3640 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3644 kiblnd_scheduler(void *arg)
3646 long id = (long)arg;
3647 struct kib_sched_info *sched;
3648 struct kib_conn *conn;
3649 wait_queue_entry_t wait;
3650 unsigned long flags;
3656 init_waitqueue_entry(&wait, current);
3658 sched = kiblnd_data.kib_scheds[KIB_THREAD_CPT(id)];
3660 rc = cfs_cpt_bind(lnet_cpt_table(), sched->ibs_cpt);
3662 CWARN("Unable to bind on CPU partition %d, please verify "
3663 "whether all CPUs are healthy and reload modules if "
3664 "necessary, otherwise your system might under risk of "
3665 "low performance\n", sched->ibs_cpt);
3668 spin_lock_irqsave(&sched->ibs_lock, flags);
3670 while (!kiblnd_data.kib_shutdown) {
3671 if (busy_loops++ >= IBLND_RESCHED) {
3672 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3677 spin_lock_irqsave(&sched->ibs_lock, flags);
3682 if (!list_empty(&sched->ibs_conns)) {
3683 conn = list_entry(sched->ibs_conns.next,
3684 struct kib_conn, ibc_sched_list);
3685 /* take over kib_sched_conns' ref on conn... */
3686 LASSERT(conn->ibc_scheduled);
3687 list_del(&conn->ibc_sched_list);
3688 conn->ibc_ready = 0;
3690 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3692 wc.wr_id = IBLND_WID_INVAL;
3694 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3696 rc = ib_req_notify_cq(conn->ibc_cq,
3699 CWARN("%s: ib_req_notify_cq failed: %d, "
3700 "closing connection\n",
3701 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
3702 kiblnd_close_conn(conn, -EIO);
3703 kiblnd_conn_decref(conn);
3704 spin_lock_irqsave(&sched->ibs_lock,
3709 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3712 if (unlikely(rc > 0 && wc.wr_id == IBLND_WID_INVAL)) {
3714 "ib_poll_cq (rc: %d) returned invalid "
3715 "wr_id, opcode %d, status: %d, "
3716 "vendor_err: %d, conn: %s status: %d\n"
3717 "please upgrade firmware and OFED or "
3718 "contact vendor.\n", rc,
3719 wc.opcode, wc.status, wc.vendor_err,
3720 libcfs_nid2str(conn->ibc_peer->ibp_nid),
3726 CWARN("%s: ib_poll_cq failed: %d, "
3727 "closing connection\n",
3728 libcfs_nid2str(conn->ibc_peer->ibp_nid),
3730 kiblnd_close_conn(conn, -EIO);
3731 kiblnd_conn_decref(conn);
3732 spin_lock_irqsave(&sched->ibs_lock, flags);
3736 spin_lock_irqsave(&sched->ibs_lock, flags);
3738 if (rc != 0 || conn->ibc_ready) {
3739 /* There may be another completion waiting; get
3740 * another scheduler to check while I handle
3742 /* +1 ref for sched_conns */
3743 kiblnd_conn_addref(conn);
3744 list_add_tail(&conn->ibc_sched_list,
3746 if (waitqueue_active(&sched->ibs_waitq))
3747 wake_up(&sched->ibs_waitq);
3749 conn->ibc_scheduled = 0;
3753 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3754 kiblnd_complete(&wc);
3756 spin_lock_irqsave(&sched->ibs_lock, flags);
3759 kiblnd_conn_decref(conn); /* ...drop my ref from above */
3766 set_current_state(TASK_INTERRUPTIBLE);
3767 add_wait_queue_exclusive(&sched->ibs_waitq, &wait);
3768 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3773 remove_wait_queue(&sched->ibs_waitq, &wait);
3774 set_current_state(TASK_RUNNING);
3775 spin_lock_irqsave(&sched->ibs_lock, flags);
3778 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3780 kiblnd_thread_fini();
3785 kiblnd_failover_thread(void *arg)
3787 rwlock_t *glock = &kiblnd_data.kib_global_lock;
3788 struct kib_dev *dev;
3789 struct net *ns = arg;
3790 wait_queue_entry_t wait;
3791 unsigned long flags;
3794 LASSERT(*kiblnd_tunables.kib_dev_failover != 0);
3796 init_waitqueue_entry(&wait, current);
3797 write_lock_irqsave(glock, flags);
3799 while (!kiblnd_data.kib_shutdown) {
3800 int do_failover = 0;
3803 list_for_each_entry(dev, &kiblnd_data.kib_failed_devs,
3805 if (ktime_get_seconds() < dev->ibd_next_failover)
3812 list_del_init(&dev->ibd_fail_list);
3813 dev->ibd_failover = 1;
3814 write_unlock_irqrestore(glock, flags);
3816 rc = kiblnd_dev_failover(dev, ns);
3818 write_lock_irqsave(glock, flags);
3820 LASSERT (dev->ibd_failover);
3821 dev->ibd_failover = 0;
3822 if (rc >= 0) { /* Device is OK or failover succeed */
3823 dev->ibd_next_failover = ktime_get_seconds() + 3;
3827 /* failed to failover, retry later */
3828 dev->ibd_next_failover = ktime_get_seconds() +
3829 min(dev->ibd_failed_failover, 10);
3830 if (kiblnd_dev_can_failover(dev)) {
3831 list_add_tail(&dev->ibd_fail_list,
3832 &kiblnd_data.kib_failed_devs);
3838 /* long sleep if no more pending failover */
3839 long_sleep = list_empty(&kiblnd_data.kib_failed_devs);
3841 set_current_state(TASK_INTERRUPTIBLE);
3842 add_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3843 write_unlock_irqrestore(glock, flags);
3845 rc = schedule_timeout(long_sleep ? cfs_time_seconds(10) :
3846 cfs_time_seconds(1));
3847 set_current_state(TASK_RUNNING);
3848 remove_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3849 write_lock_irqsave(glock, flags);
3851 if (!long_sleep || rc != 0)
3854 /* have a long sleep, routine check all active devices,
3855 * we need checking like this because if there is not active
3856 * connection on the dev and no SEND from local, we may listen
3857 * on wrong HCA for ever while there is a bonding failover */
3858 list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3859 if (kiblnd_dev_can_failover(dev)) {
3860 list_add_tail(&dev->ibd_fail_list,
3861 &kiblnd_data.kib_failed_devs);
3866 write_unlock_irqrestore(glock, flags);
3868 kiblnd_thread_fini();