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(kib_peer_ni_t *peer_ni);
42 static void kiblnd_peer_connect_failed(kib_peer_ni_t *peer_ni, int active, int error);
43 static void kiblnd_init_tx_msg(struct lnet_ni *ni, kib_tx_t *tx,
44 int type, int body_nob);
45 static int kiblnd_init_rdma(kib_conn_t *conn, kib_tx_t *tx, int type,
46 int resid, kib_rdma_desc_t *dstrd, __u64 dstcookie);
47 static void kiblnd_queue_tx_locked(kib_tx_t *tx, kib_conn_t *conn);
48 static void kiblnd_queue_tx(kib_tx_t *tx, kib_conn_t *conn);
50 static void kiblnd_unmap_tx(kib_tx_t *tx);
51 static void kiblnd_check_sends_locked(kib_conn_t *conn);
54 kiblnd_tx_done(kib_tx_t *tx)
56 struct lnet_msg *lntmsg[2];
60 LASSERT (!in_interrupt());
61 LASSERT (!tx->tx_queued); /* mustn't be queued for sending */
62 LASSERT (tx->tx_sending == 0); /* mustn't be awaiting sent callback */
63 LASSERT (!tx->tx_waiting); /* mustn't be awaiting peer_ni response */
64 LASSERT (tx->tx_pool != NULL);
68 /* tx may have up to 2 lnet msgs to finalise */
69 lntmsg[0] = tx->tx_lntmsg[0]; tx->tx_lntmsg[0] = NULL;
70 lntmsg[1] = tx->tx_lntmsg[1]; tx->tx_lntmsg[1] = NULL;
73 if (tx->tx_conn != NULL) {
74 kiblnd_conn_decref(tx->tx_conn);
78 tx->tx_nwrq = tx->tx_nsge = 0;
81 kiblnd_pool_free_node(&tx->tx_pool->tpo_pool, &tx->tx_list);
83 /* delay finalize until my descs have been freed */
84 for (i = 0; i < 2; i++) {
85 if (lntmsg[i] == NULL)
88 lnet_finalize(lntmsg[i], rc);
93 kiblnd_txlist_done(struct list_head *txlist, int status)
97 while (!list_empty(txlist)) {
98 tx = list_entry(txlist->next, kib_tx_t, tx_list);
100 list_del(&tx->tx_list);
103 tx->tx_status = status;
109 kiblnd_get_idle_tx(struct lnet_ni *ni, lnet_nid_t target)
111 kib_net_t *net = (kib_net_t *)ni->ni_data;
112 struct list_head *node;
114 kib_tx_poolset_t *tps;
116 tps = net->ibn_tx_ps[lnet_cpt_of_nid(target, ni)];
117 node = kiblnd_pool_alloc_node(&tps->tps_poolset);
120 tx = container_of(node, kib_tx_t, tx_list);
122 LASSERT (tx->tx_nwrq == 0);
123 LASSERT (!tx->tx_queued);
124 LASSERT (tx->tx_sending == 0);
125 LASSERT (!tx->tx_waiting);
126 LASSERT (tx->tx_status == 0);
127 LASSERT (tx->tx_conn == NULL);
128 LASSERT (tx->tx_lntmsg[0] == NULL);
129 LASSERT (tx->tx_lntmsg[1] == NULL);
130 LASSERT (tx->tx_nfrags == 0);
138 kiblnd_drop_rx(kib_rx_t *rx)
140 kib_conn_t *conn = rx->rx_conn;
141 struct kib_sched_info *sched = conn->ibc_sched;
144 spin_lock_irqsave(&sched->ibs_lock, flags);
145 LASSERT(conn->ibc_nrx > 0);
147 spin_unlock_irqrestore(&sched->ibs_lock, flags);
149 kiblnd_conn_decref(conn);
153 kiblnd_post_rx (kib_rx_t *rx, int credit)
155 kib_conn_t *conn = rx->rx_conn;
156 kib_net_t *net = conn->ibc_peer->ibp_ni->ni_data;
157 struct ib_recv_wr *bad_wrq = NULL;
158 #ifdef HAVE_IB_GET_DMA_MR
159 struct ib_mr *mr = conn->ibc_hdev->ibh_mrs;
163 LASSERT (net != NULL);
164 LASSERT (!in_interrupt());
165 LASSERT (credit == IBLND_POSTRX_NO_CREDIT ||
166 credit == IBLND_POSTRX_PEER_CREDIT ||
167 credit == IBLND_POSTRX_RSRVD_CREDIT);
168 #ifdef HAVE_IB_GET_DMA_MR
171 rx->rx_sge.lkey = mr->lkey;
173 rx->rx_sge.lkey = conn->ibc_hdev->ibh_pd->local_dma_lkey;
175 rx->rx_sge.addr = rx->rx_msgaddr;
176 rx->rx_sge.length = IBLND_MSG_SIZE;
178 rx->rx_wrq.next = NULL;
179 rx->rx_wrq.sg_list = &rx->rx_sge;
180 rx->rx_wrq.num_sge = 1;
181 rx->rx_wrq.wr_id = kiblnd_ptr2wreqid(rx, IBLND_WID_RX);
183 LASSERT (conn->ibc_state >= IBLND_CONN_INIT);
184 LASSERT (rx->rx_nob >= 0); /* not posted */
186 if (conn->ibc_state > IBLND_CONN_ESTABLISHED) {
187 kiblnd_drop_rx(rx); /* No more posts for this rx */
191 rx->rx_nob = -1; /* flag posted */
193 /* NB: need an extra reference after ib_post_recv because we don't
194 * own this rx (and rx::rx_conn) anymore, LU-5678.
196 kiblnd_conn_addref(conn);
197 rc = ib_post_recv(conn->ibc_cmid->qp, &rx->rx_wrq, &bad_wrq);
198 if (unlikely(rc != 0)) {
199 CERROR("Can't post rx for %s: %d, bad_wrq: %p\n",
200 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc, bad_wrq);
204 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) /* Initial post */
207 if (unlikely(rc != 0)) {
208 kiblnd_close_conn(conn, rc);
209 kiblnd_drop_rx(rx); /* No more posts for this rx */
213 if (credit == IBLND_POSTRX_NO_CREDIT)
216 spin_lock(&conn->ibc_lock);
217 if (credit == IBLND_POSTRX_PEER_CREDIT)
218 conn->ibc_outstanding_credits++;
220 conn->ibc_reserved_credits++;
221 kiblnd_check_sends_locked(conn);
222 spin_unlock(&conn->ibc_lock);
225 kiblnd_conn_decref(conn);
230 kiblnd_find_waiting_tx_locked(kib_conn_t *conn, int txtype, __u64 cookie)
232 struct list_head *tmp;
234 list_for_each(tmp, &conn->ibc_active_txs) {
235 kib_tx_t *tx = list_entry(tmp, kib_tx_t, tx_list);
237 LASSERT(!tx->tx_queued);
238 LASSERT(tx->tx_sending != 0 || tx->tx_waiting);
240 if (tx->tx_cookie != cookie)
243 if (tx->tx_waiting &&
244 tx->tx_msg->ibm_type == txtype)
247 CWARN("Bad completion: %swaiting, type %x (wanted %x)\n",
248 tx->tx_waiting ? "" : "NOT ",
249 tx->tx_msg->ibm_type, txtype);
255 kiblnd_handle_completion(kib_conn_t *conn, int txtype, int status, __u64 cookie)
258 struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
261 spin_lock(&conn->ibc_lock);
263 tx = kiblnd_find_waiting_tx_locked(conn, txtype, cookie);
265 spin_unlock(&conn->ibc_lock);
267 CWARN("Unmatched completion type %x cookie %#llx from %s\n",
268 txtype, cookie, libcfs_nid2str(conn->ibc_peer->ibp_nid));
269 kiblnd_close_conn(conn, -EPROTO);
273 if (tx->tx_status == 0) { /* success so far */
274 if (status < 0) { /* failed? */
275 tx->tx_status = status;
276 } else if (txtype == IBLND_MSG_GET_REQ) {
277 lnet_set_reply_msg_len(ni, tx->tx_lntmsg[1], status);
283 idle = !tx->tx_queued && (tx->tx_sending == 0);
285 list_del(&tx->tx_list);
287 spin_unlock(&conn->ibc_lock);
294 kiblnd_send_completion(kib_conn_t *conn, int type, int status, __u64 cookie)
296 struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
297 kib_tx_t *tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
300 CERROR("Can't get tx for completion %x for %s\n",
301 type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
305 tx->tx_msg->ibm_u.completion.ibcm_status = status;
306 tx->tx_msg->ibm_u.completion.ibcm_cookie = cookie;
307 kiblnd_init_tx_msg(ni, tx, type, sizeof(kib_completion_msg_t));
309 kiblnd_queue_tx(tx, conn);
313 kiblnd_handle_rx (kib_rx_t *rx)
315 kib_msg_t *msg = rx->rx_msg;
316 kib_conn_t *conn = rx->rx_conn;
317 struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
318 int credits = msg->ibm_credits;
324 LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
326 CDEBUG (D_NET, "Received %x[%d] from %s\n",
327 msg->ibm_type, credits,
328 libcfs_nid2str(conn->ibc_peer->ibp_nid));
331 /* Have I received credits that will let me send? */
332 spin_lock(&conn->ibc_lock);
334 if (conn->ibc_credits + credits >
335 conn->ibc_queue_depth) {
336 rc2 = conn->ibc_credits;
337 spin_unlock(&conn->ibc_lock);
339 CERROR("Bad credits from %s: %d + %d > %d\n",
340 libcfs_nid2str(conn->ibc_peer->ibp_nid),
342 conn->ibc_queue_depth);
344 kiblnd_close_conn(conn, -EPROTO);
345 kiblnd_post_rx(rx, IBLND_POSTRX_NO_CREDIT);
349 conn->ibc_credits += credits;
351 /* This ensures the credit taken by NOOP can be returned */
352 if (msg->ibm_type == IBLND_MSG_NOOP &&
353 !IBLND_OOB_CAPABLE(conn->ibc_version)) /* v1 only */
354 conn->ibc_outstanding_credits++;
356 kiblnd_check_sends_locked(conn);
357 spin_unlock(&conn->ibc_lock);
360 switch (msg->ibm_type) {
362 CERROR("Bad IBLND message type %x from %s\n",
363 msg->ibm_type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
364 post_credit = IBLND_POSTRX_NO_CREDIT;
369 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
370 post_credit = IBLND_POSTRX_NO_CREDIT;
374 if (credits != 0) /* credit already posted */
375 post_credit = IBLND_POSTRX_NO_CREDIT;
376 else /* a keepalive NOOP */
377 post_credit = IBLND_POSTRX_PEER_CREDIT;
380 case IBLND_MSG_IMMEDIATE:
381 post_credit = IBLND_POSTRX_DONT_POST;
382 rc = lnet_parse(ni, &msg->ibm_u.immediate.ibim_hdr,
383 msg->ibm_srcnid, rx, 0);
384 if (rc < 0) /* repost on error */
385 post_credit = IBLND_POSTRX_PEER_CREDIT;
388 case IBLND_MSG_PUT_REQ:
389 post_credit = IBLND_POSTRX_DONT_POST;
390 rc = lnet_parse(ni, &msg->ibm_u.putreq.ibprm_hdr,
391 msg->ibm_srcnid, rx, 1);
392 if (rc < 0) /* repost on error */
393 post_credit = IBLND_POSTRX_PEER_CREDIT;
396 case IBLND_MSG_PUT_NAK:
397 CWARN ("PUT_NACK from %s\n",
398 libcfs_nid2str(conn->ibc_peer->ibp_nid));
399 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
400 kiblnd_handle_completion(conn, IBLND_MSG_PUT_REQ,
401 msg->ibm_u.completion.ibcm_status,
402 msg->ibm_u.completion.ibcm_cookie);
405 case IBLND_MSG_PUT_ACK:
406 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
408 spin_lock(&conn->ibc_lock);
409 tx = kiblnd_find_waiting_tx_locked(conn, IBLND_MSG_PUT_REQ,
410 msg->ibm_u.putack.ibpam_src_cookie);
412 list_del(&tx->tx_list);
413 spin_unlock(&conn->ibc_lock);
416 CERROR("Unmatched PUT_ACK from %s\n",
417 libcfs_nid2str(conn->ibc_peer->ibp_nid));
422 LASSERT (tx->tx_waiting);
423 /* CAVEAT EMPTOR: I could be racing with tx_complete, but...
424 * (a) I can overwrite tx_msg since my peer_ni has received it!
425 * (b) tx_waiting set tells tx_complete() it's not done. */
427 tx->tx_nwrq = tx->tx_nsge = 0; /* overwrite PUT_REQ */
429 rc2 = kiblnd_init_rdma(conn, tx, IBLND_MSG_PUT_DONE,
430 kiblnd_rd_size(&msg->ibm_u.putack.ibpam_rd),
431 &msg->ibm_u.putack.ibpam_rd,
432 msg->ibm_u.putack.ibpam_dst_cookie);
434 CERROR("Can't setup rdma for PUT to %s: %d\n",
435 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc2);
437 spin_lock(&conn->ibc_lock);
438 tx->tx_waiting = 0; /* clear waiting and queue atomically */
439 kiblnd_queue_tx_locked(tx, conn);
440 spin_unlock(&conn->ibc_lock);
443 case IBLND_MSG_PUT_DONE:
444 post_credit = IBLND_POSTRX_PEER_CREDIT;
445 kiblnd_handle_completion(conn, IBLND_MSG_PUT_ACK,
446 msg->ibm_u.completion.ibcm_status,
447 msg->ibm_u.completion.ibcm_cookie);
450 case IBLND_MSG_GET_REQ:
451 post_credit = IBLND_POSTRX_DONT_POST;
452 rc = lnet_parse(ni, &msg->ibm_u.get.ibgm_hdr,
453 msg->ibm_srcnid, rx, 1);
454 if (rc < 0) /* repost on error */
455 post_credit = IBLND_POSTRX_PEER_CREDIT;
458 case IBLND_MSG_GET_DONE:
459 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
460 kiblnd_handle_completion(conn, IBLND_MSG_GET_REQ,
461 msg->ibm_u.completion.ibcm_status,
462 msg->ibm_u.completion.ibcm_cookie);
466 if (rc < 0) /* protocol error */
467 kiblnd_close_conn(conn, rc);
469 if (post_credit != IBLND_POSTRX_DONT_POST)
470 kiblnd_post_rx(rx, post_credit);
474 kiblnd_rx_complete (kib_rx_t *rx, int status, int nob)
476 kib_msg_t *msg = rx->rx_msg;
477 kib_conn_t *conn = rx->rx_conn;
478 struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
479 kib_net_t *net = ni->ni_data;
483 LASSERT (net != NULL);
484 LASSERT (rx->rx_nob < 0); /* was posted */
485 rx->rx_nob = 0; /* isn't now */
487 if (conn->ibc_state > IBLND_CONN_ESTABLISHED)
490 if (status != IB_WC_SUCCESS) {
491 CNETERR("Rx from %s failed: %d\n",
492 libcfs_nid2str(conn->ibc_peer->ibp_nid), status);
499 rc = kiblnd_unpack_msg(msg, rx->rx_nob);
501 CERROR ("Error %d unpacking rx from %s\n",
502 rc, libcfs_nid2str(conn->ibc_peer->ibp_nid));
506 if (msg->ibm_srcnid != conn->ibc_peer->ibp_nid ||
507 msg->ibm_dstnid != ni->ni_nid ||
508 msg->ibm_srcstamp != conn->ibc_incarnation ||
509 msg->ibm_dststamp != net->ibn_incarnation) {
510 CERROR ("Stale rx from %s\n",
511 libcfs_nid2str(conn->ibc_peer->ibp_nid));
516 /* set time last known alive */
517 kiblnd_peer_alive(conn->ibc_peer);
519 /* racing with connection establishment/teardown! */
521 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
522 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
525 write_lock_irqsave(g_lock, flags);
526 /* must check holding global lock to eliminate race */
527 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
528 list_add_tail(&rx->rx_list, &conn->ibc_early_rxs);
529 write_unlock_irqrestore(g_lock, flags);
532 write_unlock_irqrestore(g_lock, flags);
534 kiblnd_handle_rx(rx);
538 CDEBUG(D_NET, "rx %p conn %p\n", rx, conn);
539 kiblnd_close_conn(conn, err);
541 kiblnd_drop_rx(rx); /* Don't re-post rx. */
545 kiblnd_fmr_map_tx(kib_net_t *net, kib_tx_t *tx, kib_rdma_desc_t *rd, __u32 nob)
549 kib_fmr_poolset_t *fps;
554 LASSERT(tx->tx_pool != NULL);
555 LASSERT(tx->tx_pool->tpo_pool.po_owner != NULL);
558 hdev = tx->tx_pool->tpo_hdev;
559 cpt = tx->tx_pool->tpo_pool.po_owner->ps_cpt;
562 * If we're dealing with FastReg, but the device doesn't
563 * support GAPS and the tx has GAPS, then there is no real point
564 * in trying to map the memory, because it'll just fail. So
565 * preemptively fail with an appropriate message
567 if ((dev->ibd_dev_caps & IBLND_DEV_CAPS_FASTREG_ENABLED) &&
568 !(dev->ibd_dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT) &&
570 CERROR("Using FastReg with no GAPS support, but tx has gaps\n");
571 return -EPROTONOSUPPORT;
575 * FMR does not support gaps but the tx has gaps then
576 * we should make sure that the number of fragments we'll be sending
577 * over fits within the number of fragments negotiated on the
578 * connection, otherwise, we won't be able to RDMA the data.
579 * We need to maintain the number of fragments negotiation on the
580 * connection for backwards compatibility.
582 if (tx->tx_gaps && (dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED)) {
584 tx->tx_conn->ibc_max_frags <= rd->rd_nfrags) {
585 CERROR("TX number of frags (%d) is <= than connection"
586 " number of frags (%d). Consider setting peer's"
587 " map_on_demand to 256\n", tx->tx_nfrags,
588 tx->tx_conn->ibc_max_frags);
593 fps = net->ibn_fmr_ps[cpt];
594 rc = kiblnd_fmr_pool_map(fps, tx, rd, nob, 0, &tx->tx_fmr);
596 CERROR("Can't map %u pages: %d\n", nob, rc);
601 * If rd is not tx_rd, it's going to get sent to a peer_ni, who will
604 rd->rd_key = tx->tx_fmr.fmr_key;
606 * for FastReg or FMR with no gaps we can accumulate all
607 * the fragments in one FastReg or FMR fragment.
609 if (((dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED) && !tx->tx_gaps) ||
610 (dev->ibd_dev_caps & IBLND_DEV_CAPS_FASTREG_ENABLED)) {
611 /* FMR requires zero based address */
612 if (dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED)
613 rd->rd_frags[0].rf_addr &= ~hdev->ibh_page_mask;
614 rd->rd_frags[0].rf_nob = nob;
618 * We're transmitting with gaps using FMR.
619 * We'll need to use multiple fragments and identify the
620 * zero based address of each fragment.
622 for (i = 0; i < rd->rd_nfrags; i++) {
623 rd->rd_frags[i].rf_addr &= ~hdev->ibh_page_mask;
624 rd->rd_frags[i].rf_addr += i << hdev->ibh_page_shift;
632 kiblnd_unmap_tx(kib_tx_t *tx)
634 if (tx->tx_fmr.fmr_pfmr || tx->tx_fmr.fmr_frd)
635 kiblnd_fmr_pool_unmap(&tx->tx_fmr, tx->tx_status);
637 if (tx->tx_nfrags != 0) {
638 kiblnd_dma_unmap_sg(tx->tx_pool->tpo_hdev->ibh_ibdev,
639 tx->tx_frags, tx->tx_nfrags, tx->tx_dmadir);
644 #ifdef HAVE_IB_GET_DMA_MR
645 static struct ib_mr *
646 kiblnd_find_rd_dma_mr(struct lnet_ni *ni, kib_rdma_desc_t *rd)
648 kib_net_t *net = ni->ni_data;
649 kib_hca_dev_t *hdev = net->ibn_dev->ibd_hdev;
650 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
652 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
655 * if map-on-demand is turned on and the device supports
656 * either FMR or FastReg then use that. Otherwise use global
657 * memory regions. If that's not available either, then you're
658 * dead in the water and fail the operation.
660 if (tunables->lnd_map_on_demand &&
661 (net->ibn_dev->ibd_dev_caps & IBLND_DEV_CAPS_FASTREG_ENABLED ||
662 net->ibn_dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED))
666 * hdev->ibh_mrs can be NULL. This case is dealt with gracefully
667 * in the call chain. The mapping will fail with appropriate error
670 return hdev->ibh_mrs;
675 kiblnd_map_tx(struct lnet_ni *ni, kib_tx_t *tx, kib_rdma_desc_t *rd, int nfrags)
677 kib_net_t *net = ni->ni_data;
678 kib_hca_dev_t *hdev = net->ibn_dev->ibd_hdev;
679 #ifdef HAVE_IB_GET_DMA_MR
680 struct ib_mr *mr = NULL;
685 /* If rd is not tx_rd, it's going to get sent to a peer_ni and I'm the
687 tx->tx_dmadir = (rd != tx->tx_rd) ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
688 tx->tx_nfrags = nfrags;
690 rd->rd_nfrags = kiblnd_dma_map_sg(hdev->ibh_ibdev, tx->tx_frags,
691 tx->tx_nfrags, tx->tx_dmadir);
693 for (i = 0, nob = 0; i < rd->rd_nfrags; i++) {
694 rd->rd_frags[i].rf_nob = kiblnd_sg_dma_len(
695 hdev->ibh_ibdev, &tx->tx_frags[i]);
696 rd->rd_frags[i].rf_addr = kiblnd_sg_dma_address(
697 hdev->ibh_ibdev, &tx->tx_frags[i]);
698 nob += rd->rd_frags[i].rf_nob;
701 #ifdef HAVE_IB_GET_DMA_MR
702 mr = kiblnd_find_rd_dma_mr(ni, rd);
704 /* found pre-mapping MR */
705 rd->rd_key = (rd != tx->tx_rd) ? mr->rkey : mr->lkey;
710 if (net->ibn_fmr_ps != NULL)
711 return kiblnd_fmr_map_tx(net, tx, rd, nob);
717 kiblnd_setup_rd_iov(struct lnet_ni *ni, kib_tx_t *tx, kib_rdma_desc_t *rd,
718 unsigned int niov, struct kvec *iov, int offset, int nob)
720 kib_net_t *net = ni->ni_data;
722 struct scatterlist *sg;
729 LASSERT (net != NULL);
731 while (offset >= iov->iov_len) {
732 offset -= iov->iov_len;
742 vaddr = ((unsigned long)iov->iov_base) + offset;
743 page_offset = vaddr & (PAGE_SIZE - 1);
744 page = lnet_kvaddr_to_page(vaddr);
746 CERROR("Can't find page\n");
750 fragnob = min((int)(iov->iov_len - offset), nob);
751 fragnob = min(fragnob, (int)PAGE_SIZE - page_offset);
753 if ((fragnob < (int)PAGE_SIZE - page_offset) && (niov > 1)) {
754 CDEBUG(D_NET, "fragnob %d < available page %d: with"
755 " remaining %d iovs\n",
756 fragnob, (int)PAGE_SIZE - page_offset, niov);
760 sg_set_page(sg, page, fragnob, page_offset);
763 CERROR("lacking enough sg entries to map tx\n");
767 if (offset + fragnob < iov->iov_len) {
777 return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags);
781 kiblnd_setup_rd_kiov(struct lnet_ni *ni, kib_tx_t *tx, kib_rdma_desc_t *rd,
782 int nkiov, lnet_kiov_t *kiov, int offset, int nob)
784 kib_net_t *net = ni->ni_data;
785 struct scatterlist *sg;
788 CDEBUG(D_NET, "niov %d offset %d nob %d\n", nkiov, offset, nob);
792 LASSERT(net != NULL);
794 while (offset >= kiov->kiov_len) {
795 offset -= kiov->kiov_len;
805 fragnob = min((int)(kiov->kiov_len - offset), nob);
807 if ((fragnob < (int)(kiov->kiov_len - offset)) && nkiov > 1) {
808 CDEBUG(D_NET, "fragnob %d < available page %d: with"
809 " remaining %d kiovs\n",
810 fragnob, (int)(kiov->kiov_len - offset), nkiov);
814 sg_set_page(sg, kiov->kiov_page, fragnob,
815 kiov->kiov_offset + offset);
818 CERROR("lacking enough sg entries to map tx\n");
828 return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags);
832 kiblnd_post_tx_locked (kib_conn_t *conn, kib_tx_t *tx, int credit)
833 __must_hold(&conn->ibc_lock)
835 kib_msg_t *msg = tx->tx_msg;
836 kib_peer_ni_t *peer_ni = conn->ibc_peer;
837 int ver = conn->ibc_version;
841 LASSERT(tx->tx_queued);
842 /* We rely on this for QP sizing */
843 LASSERT(tx->tx_nwrq > 0 && tx->tx_nsge >= 0);
844 LASSERT(tx->tx_nwrq <= 1 + conn->ibc_max_frags);
846 LASSERT(credit == 0 || credit == 1);
847 LASSERT(conn->ibc_outstanding_credits >= 0);
848 LASSERT(conn->ibc_outstanding_credits <= conn->ibc_queue_depth);
849 LASSERT(conn->ibc_credits >= 0);
850 LASSERT(conn->ibc_credits <= conn->ibc_queue_depth);
852 if (conn->ibc_nsends_posted ==
853 conn->ibc_queue_depth) {
854 /* tx completions outstanding... */
855 CDEBUG(D_NET, "%s: posted enough\n",
856 libcfs_nid2str(peer_ni->ibp_nid));
860 if (credit != 0 && conn->ibc_credits == 0) { /* no credits */
861 CDEBUG(D_NET, "%s: no credits\n",
862 libcfs_nid2str(peer_ni->ibp_nid));
866 if (credit != 0 && !IBLND_OOB_CAPABLE(ver) &&
867 conn->ibc_credits == 1 && /* last credit reserved */
868 msg->ibm_type != IBLND_MSG_NOOP) { /* for NOOP */
869 CDEBUG(D_NET, "%s: not using last credit\n",
870 libcfs_nid2str(peer_ni->ibp_nid));
874 /* NB don't drop ibc_lock before bumping tx_sending */
875 list_del(&tx->tx_list);
878 if (msg->ibm_type == IBLND_MSG_NOOP &&
879 (!kiblnd_need_noop(conn) || /* redundant NOOP */
880 (IBLND_OOB_CAPABLE(ver) && /* posted enough NOOP */
881 conn->ibc_noops_posted == IBLND_OOB_MSGS(ver)))) {
882 /* OK to drop when posted enough NOOPs, since
883 * kiblnd_check_sends_locked will queue NOOP again when
884 * posted NOOPs complete */
885 spin_unlock(&conn->ibc_lock);
887 spin_lock(&conn->ibc_lock);
888 CDEBUG(D_NET, "%s(%d): redundant or enough NOOP\n",
889 libcfs_nid2str(peer_ni->ibp_nid),
890 conn->ibc_noops_posted);
894 kiblnd_pack_msg(peer_ni->ibp_ni, msg, ver, conn->ibc_outstanding_credits,
895 peer_ni->ibp_nid, conn->ibc_incarnation);
897 conn->ibc_credits -= credit;
898 conn->ibc_outstanding_credits = 0;
899 conn->ibc_nsends_posted++;
900 if (msg->ibm_type == IBLND_MSG_NOOP)
901 conn->ibc_noops_posted++;
903 /* CAVEAT EMPTOR! This tx could be the PUT_DONE of an RDMA
904 * PUT. If so, it was first queued here as a PUT_REQ, sent and
905 * stashed on ibc_active_txs, matched by an incoming PUT_ACK,
906 * and then re-queued here. It's (just) possible that
907 * tx_sending is non-zero if we've not done the tx_complete()
908 * from the first send; hence the ++ rather than = below. */
910 list_add(&tx->tx_list, &conn->ibc_active_txs);
912 /* I'm still holding ibc_lock! */
913 if (conn->ibc_state != IBLND_CONN_ESTABLISHED) {
915 } else if (tx->tx_pool->tpo_pool.po_failed ||
916 conn->ibc_hdev != tx->tx_pool->tpo_hdev) {
917 /* close_conn will launch failover */
920 struct kib_fast_reg_descriptor *frd = tx->tx_fmr.fmr_frd;
921 struct ib_send_wr *bad = &tx->tx_wrq[tx->tx_nwrq - 1].wr;
922 struct ib_send_wr *wr = &tx->tx_wrq[0].wr;
925 if (!frd->frd_valid) {
926 wr = &frd->frd_inv_wr.wr;
927 wr->next = &frd->frd_fastreg_wr.wr;
929 wr = &frd->frd_fastreg_wr.wr;
931 frd->frd_fastreg_wr.wr.next = &tx->tx_wrq[0].wr;
934 LASSERTF(bad->wr_id == kiblnd_ptr2wreqid(tx, IBLND_WID_TX),
935 "bad wr_id %#llx, opc %d, flags %d, peer_ni: %s\n",
936 bad->wr_id, bad->opcode, bad->send_flags,
937 libcfs_nid2str(conn->ibc_peer->ibp_nid));
940 rc = ib_post_send(conn->ibc_cmid->qp, wr, &bad);
943 conn->ibc_last_send = jiffies;
948 /* NB credits are transferred in the actual
949 * message, which can only be the last work item */
950 conn->ibc_credits += credit;
951 conn->ibc_outstanding_credits += msg->ibm_credits;
952 conn->ibc_nsends_posted--;
953 if (msg->ibm_type == IBLND_MSG_NOOP)
954 conn->ibc_noops_posted--;
960 done = (tx->tx_sending == 0);
962 list_del(&tx->tx_list);
964 spin_unlock(&conn->ibc_lock);
966 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
967 CERROR("Error %d posting transmit to %s\n",
968 rc, libcfs_nid2str(peer_ni->ibp_nid));
970 CDEBUG(D_NET, "Error %d posting transmit to %s\n",
971 rc, libcfs_nid2str(peer_ni->ibp_nid));
973 kiblnd_close_conn(conn, rc);
978 spin_lock(&conn->ibc_lock);
984 kiblnd_check_sends_locked(kib_conn_t *conn)
986 int ver = conn->ibc_version;
987 struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
990 /* Don't send anything until after the connection is established */
991 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
992 CDEBUG(D_NET, "%s too soon\n",
993 libcfs_nid2str(conn->ibc_peer->ibp_nid));
997 LASSERT(conn->ibc_nsends_posted <=
998 conn->ibc_queue_depth);
999 LASSERT (!IBLND_OOB_CAPABLE(ver) ||
1000 conn->ibc_noops_posted <= IBLND_OOB_MSGS(ver));
1001 LASSERT (conn->ibc_reserved_credits >= 0);
1003 while (conn->ibc_reserved_credits > 0 &&
1004 !list_empty(&conn->ibc_tx_queue_rsrvd)) {
1005 tx = list_entry(conn->ibc_tx_queue_rsrvd.next,
1007 list_del(&tx->tx_list);
1008 list_add_tail(&tx->tx_list, &conn->ibc_tx_queue);
1009 conn->ibc_reserved_credits--;
1012 if (kiblnd_need_noop(conn)) {
1013 spin_unlock(&conn->ibc_lock);
1015 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
1017 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_NOOP, 0);
1019 spin_lock(&conn->ibc_lock);
1021 kiblnd_queue_tx_locked(tx, conn);
1027 if (!list_empty(&conn->ibc_tx_queue_nocred)) {
1029 tx = list_entry(conn->ibc_tx_queue_nocred.next,
1031 } else if (!list_empty(&conn->ibc_tx_noops)) {
1032 LASSERT (!IBLND_OOB_CAPABLE(ver));
1034 tx = list_entry(conn->ibc_tx_noops.next,
1036 } else if (!list_empty(&conn->ibc_tx_queue)) {
1038 tx = list_entry(conn->ibc_tx_queue.next,
1043 if (kiblnd_post_tx_locked(conn, tx, credit) != 0)
1049 kiblnd_tx_complete (kib_tx_t *tx, int status)
1051 int failed = (status != IB_WC_SUCCESS);
1052 kib_conn_t *conn = tx->tx_conn;
1055 LASSERT (tx->tx_sending > 0);
1058 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
1059 CNETERR("Tx -> %s cookie %#llx"
1060 " sending %d waiting %d: failed %d\n",
1061 libcfs_nid2str(conn->ibc_peer->ibp_nid),
1062 tx->tx_cookie, tx->tx_sending, tx->tx_waiting,
1065 kiblnd_close_conn(conn, -EIO);
1067 kiblnd_peer_alive(conn->ibc_peer);
1070 spin_lock(&conn->ibc_lock);
1072 /* I could be racing with rdma completion. Whoever makes 'tx' idle
1073 * gets to free it, which also drops its ref on 'conn'. */
1076 conn->ibc_nsends_posted--;
1077 if (tx->tx_msg->ibm_type == IBLND_MSG_NOOP)
1078 conn->ibc_noops_posted--;
1081 tx->tx_waiting = 0; /* don't wait for peer_ni */
1082 tx->tx_status = -EIO;
1085 idle = (tx->tx_sending == 0) && /* This is the final callback */
1086 !tx->tx_waiting && /* Not waiting for peer_ni */
1087 !tx->tx_queued; /* Not re-queued (PUT_DONE) */
1089 list_del(&tx->tx_list);
1091 kiblnd_check_sends_locked(conn);
1092 spin_unlock(&conn->ibc_lock);
1099 kiblnd_init_tx_msg(struct lnet_ni *ni, kib_tx_t *tx, int type, int body_nob)
1101 kib_hca_dev_t *hdev = tx->tx_pool->tpo_hdev;
1102 struct ib_sge *sge = &tx->tx_msgsge;
1103 struct ib_rdma_wr *wrq;
1104 int nob = offsetof(kib_msg_t, ibm_u) + body_nob;
1105 #ifdef HAVE_IB_GET_DMA_MR
1106 struct ib_mr *mr = hdev->ibh_mrs;
1109 LASSERT(tx->tx_nwrq >= 0);
1110 LASSERT(tx->tx_nwrq < IBLND_MAX_RDMA_FRAGS + 1);
1111 LASSERT(nob <= IBLND_MSG_SIZE);
1112 #ifdef HAVE_IB_GET_DMA_MR
1113 LASSERT(mr != NULL);
1116 kiblnd_init_msg(tx->tx_msg, type, body_nob);
1118 #ifdef HAVE_IB_GET_DMA_MR
1119 sge->lkey = mr->lkey;
1121 sge->lkey = hdev->ibh_pd->local_dma_lkey;
1123 sge->addr = tx->tx_msgaddr;
1126 wrq = &tx->tx_wrq[tx->tx_nwrq];
1127 memset(wrq, 0, sizeof(*wrq));
1129 wrq->wr.next = NULL;
1130 wrq->wr.wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_TX);
1131 wrq->wr.sg_list = sge;
1132 wrq->wr.num_sge = 1;
1133 wrq->wr.opcode = IB_WR_SEND;
1134 wrq->wr.send_flags = IB_SEND_SIGNALED;
1140 kiblnd_init_rdma(kib_conn_t *conn, kib_tx_t *tx, int type,
1141 int resid, kib_rdma_desc_t *dstrd, __u64 dstcookie)
1143 kib_msg_t *ibmsg = tx->tx_msg;
1144 kib_rdma_desc_t *srcrd = tx->tx_rd;
1145 struct ib_rdma_wr *wrq = NULL;
1153 LASSERT(!in_interrupt());
1154 LASSERT(tx->tx_nwrq == 0 && tx->tx_nsge == 0);
1155 LASSERT(type == IBLND_MSG_GET_DONE || type == IBLND_MSG_PUT_DONE);
1157 for (srcidx = dstidx = wrq_sge = sge_nob = 0;
1158 resid > 0; resid -= sge_nob) {
1161 if (srcidx >= srcrd->rd_nfrags) {
1162 CERROR("Src buffer exhausted: %d frags\n", srcidx);
1167 if (dstidx >= dstrd->rd_nfrags) {
1168 CERROR("Dst buffer exhausted: %d frags\n", dstidx);
1173 if (tx->tx_nwrq >= conn->ibc_max_frags) {
1174 CERROR("RDMA has too many fragments for peer_ni %s (%d), "
1175 "src idx/frags: %d/%d dst idx/frags: %d/%d\n",
1176 libcfs_nid2str(conn->ibc_peer->ibp_nid),
1177 conn->ibc_max_frags,
1178 srcidx, srcrd->rd_nfrags,
1179 dstidx, dstrd->rd_nfrags);
1184 sge_nob = MIN(MIN(kiblnd_rd_frag_size(srcrd, srcidx),
1185 kiblnd_rd_frag_size(dstrd, dstidx)), resid);
1187 sge = &tx->tx_sge[tx->tx_nsge];
1188 sge->addr = kiblnd_rd_frag_addr(srcrd, srcidx);
1189 sge->lkey = kiblnd_rd_frag_key(srcrd, srcidx);
1190 sge->length = sge_nob;
1193 wrq = &tx->tx_wrq[tx->tx_nwrq];
1195 wrq->wr.next = &(wrq + 1)->wr;
1196 wrq->wr.wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_RDMA);
1197 wrq->wr.sg_list = sge;
1198 wrq->wr.opcode = IB_WR_RDMA_WRITE;
1199 wrq->wr.send_flags = 0;
1201 #ifdef HAVE_IB_RDMA_WR
1202 wrq->remote_addr = kiblnd_rd_frag_addr(dstrd,
1204 wrq->rkey = kiblnd_rd_frag_key(dstrd,
1207 wrq->wr.wr.rdma.remote_addr = kiblnd_rd_frag_addr(dstrd,
1209 wrq->wr.wr.rdma.rkey = kiblnd_rd_frag_key(dstrd,
1214 srcidx = kiblnd_rd_consume_frag(srcrd, srcidx, sge_nob);
1215 dstidx = kiblnd_rd_consume_frag(dstrd, dstidx, sge_nob);
1218 if (wrq_sge == *kiblnd_tunables.kib_wrq_sge || dstidx != prev) {
1220 wrq->wr.num_sge = wrq_sge;
1226 if (rc < 0) /* no RDMA if completing with failure */
1227 tx->tx_nwrq = tx->tx_nsge = 0;
1229 ibmsg->ibm_u.completion.ibcm_status = rc;
1230 ibmsg->ibm_u.completion.ibcm_cookie = dstcookie;
1231 kiblnd_init_tx_msg(conn->ibc_peer->ibp_ni, tx,
1232 type, sizeof (kib_completion_msg_t));
1238 kiblnd_queue_tx_locked(kib_tx_t *tx, kib_conn_t *conn)
1240 struct list_head *q;
1242 LASSERT(tx->tx_nwrq > 0); /* work items set up */
1243 LASSERT(!tx->tx_queued); /* not queued for sending already */
1244 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1247 tx->tx_deadline = jiffies +
1248 msecs_to_jiffies(*kiblnd_tunables.kib_timeout *
1251 if (tx->tx_conn == NULL) {
1252 kiblnd_conn_addref(conn);
1254 LASSERT (tx->tx_msg->ibm_type != IBLND_MSG_PUT_DONE);
1256 /* PUT_DONE first attached to conn as a PUT_REQ */
1257 LASSERT (tx->tx_conn == conn);
1258 LASSERT (tx->tx_msg->ibm_type == IBLND_MSG_PUT_DONE);
1261 switch (tx->tx_msg->ibm_type) {
1265 case IBLND_MSG_PUT_REQ:
1266 case IBLND_MSG_GET_REQ:
1267 q = &conn->ibc_tx_queue_rsrvd;
1270 case IBLND_MSG_PUT_NAK:
1271 case IBLND_MSG_PUT_ACK:
1272 case IBLND_MSG_PUT_DONE:
1273 case IBLND_MSG_GET_DONE:
1274 q = &conn->ibc_tx_queue_nocred;
1277 case IBLND_MSG_NOOP:
1278 if (IBLND_OOB_CAPABLE(conn->ibc_version))
1279 q = &conn->ibc_tx_queue_nocred;
1281 q = &conn->ibc_tx_noops;
1284 case IBLND_MSG_IMMEDIATE:
1285 q = &conn->ibc_tx_queue;
1289 list_add_tail(&tx->tx_list, q);
1293 kiblnd_queue_tx (kib_tx_t *tx, kib_conn_t *conn)
1295 spin_lock(&conn->ibc_lock);
1296 kiblnd_queue_tx_locked(tx, conn);
1297 kiblnd_check_sends_locked(conn);
1298 spin_unlock(&conn->ibc_lock);
1301 static int kiblnd_resolve_addr(struct rdma_cm_id *cmid,
1302 struct sockaddr_in *srcaddr,
1303 struct sockaddr_in *dstaddr,
1306 unsigned short port;
1309 /* allow the port to be reused */
1310 rc = rdma_set_reuseaddr(cmid, 1);
1312 CERROR("Unable to set reuse on cmid: %d\n", rc);
1316 /* look for a free privileged port */
1317 for (port = PROT_SOCK-1; port > 0; port--) {
1318 srcaddr->sin_port = htons(port);
1319 rc = rdma_resolve_addr(cmid,
1320 (struct sockaddr *)srcaddr,
1321 (struct sockaddr *)dstaddr,
1324 CDEBUG(D_NET, "bound to port %hu\n", port);
1326 } else if (rc == -EADDRINUSE || rc == -EADDRNOTAVAIL) {
1327 CDEBUG(D_NET, "bind to port %hu failed: %d\n",
1334 CERROR("Failed to bind to a free privileged port\n");
1339 kiblnd_connect_peer (kib_peer_ni_t *peer_ni)
1341 struct rdma_cm_id *cmid;
1343 kib_net_t *net = peer_ni->ibp_ni->ni_data;
1344 struct sockaddr_in srcaddr;
1345 struct sockaddr_in dstaddr;
1348 LASSERT (net != NULL);
1349 LASSERT (peer_ni->ibp_connecting > 0);
1351 cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, peer_ni, RDMA_PS_TCP,
1355 CERROR("Can't create CMID for %s: %ld\n",
1356 libcfs_nid2str(peer_ni->ibp_nid), PTR_ERR(cmid));
1362 memset(&srcaddr, 0, sizeof(srcaddr));
1363 srcaddr.sin_family = AF_INET;
1364 srcaddr.sin_addr.s_addr = htonl(dev->ibd_ifip);
1366 memset(&dstaddr, 0, sizeof(dstaddr));
1367 dstaddr.sin_family = AF_INET;
1368 dstaddr.sin_port = htons(*kiblnd_tunables.kib_service);
1369 dstaddr.sin_addr.s_addr = htonl(LNET_NIDADDR(peer_ni->ibp_nid));
1371 kiblnd_peer_addref(peer_ni); /* cmid's ref */
1373 if (*kiblnd_tunables.kib_use_priv_port) {
1374 rc = kiblnd_resolve_addr(cmid, &srcaddr, &dstaddr,
1375 *kiblnd_tunables.kib_timeout * 1000);
1377 rc = rdma_resolve_addr(cmid,
1378 (struct sockaddr *)&srcaddr,
1379 (struct sockaddr *)&dstaddr,
1380 *kiblnd_tunables.kib_timeout * 1000);
1383 /* Can't initiate address resolution: */
1384 CERROR("Can't resolve addr for %s: %d\n",
1385 libcfs_nid2str(peer_ni->ibp_nid), rc);
1392 kiblnd_peer_connect_failed(peer_ni, 1, rc);
1393 kiblnd_peer_decref(peer_ni); /* cmid's ref */
1394 rdma_destroy_id(cmid);
1397 kiblnd_peer_connect_failed(peer_ni, 1, rc);
1401 kiblnd_reconnect_peer(kib_peer_ni_t *peer_ni)
1403 rwlock_t *glock = &kiblnd_data.kib_global_lock;
1404 char *reason = NULL;
1405 struct list_head txs;
1406 unsigned long flags;
1408 INIT_LIST_HEAD(&txs);
1410 write_lock_irqsave(glock, flags);
1411 if (peer_ni->ibp_reconnecting == 0) {
1412 if (peer_ni->ibp_accepting)
1413 reason = "accepting";
1414 else if (peer_ni->ibp_connecting)
1415 reason = "connecting";
1416 else if (!list_empty(&peer_ni->ibp_conns))
1417 reason = "connected";
1418 else /* connected then closed */
1424 if (peer_ni->ibp_accepting)
1425 CNETERR("Detecting race between accepting and reconnecting\n");
1426 peer_ni->ibp_reconnecting--;
1428 if (!kiblnd_peer_active(peer_ni)) {
1429 list_splice_init(&peer_ni->ibp_tx_queue, &txs);
1430 reason = "unlinked";
1434 peer_ni->ibp_connecting++;
1435 peer_ni->ibp_reconnected++;
1437 write_unlock_irqrestore(glock, flags);
1439 kiblnd_connect_peer(peer_ni);
1443 write_unlock_irqrestore(glock, flags);
1445 CWARN("Abort reconnection of %s: %s\n",
1446 libcfs_nid2str(peer_ni->ibp_nid), reason);
1447 kiblnd_txlist_done(&txs, -ECONNABORTED);
1452 kiblnd_launch_tx(struct lnet_ni *ni, kib_tx_t *tx, lnet_nid_t nid)
1454 kib_peer_ni_t *peer_ni;
1455 kib_peer_ni_t *peer2;
1457 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
1458 unsigned long flags;
1461 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
1463 /* If I get here, I've committed to send, so I complete the tx with
1464 * failure on any problems */
1466 LASSERT (tx == NULL || tx->tx_conn == NULL); /* only set when assigned a conn */
1467 LASSERT (tx == NULL || tx->tx_nwrq > 0); /* work items have been set up */
1469 /* First time, just use a read lock since I expect to find my peer_ni
1471 read_lock_irqsave(g_lock, flags);
1473 peer_ni = kiblnd_find_peer_locked(ni, nid);
1474 if (peer_ni != NULL && !list_empty(&peer_ni->ibp_conns)) {
1475 /* Found a peer_ni with an established connection */
1476 conn = kiblnd_get_conn_locked(peer_ni);
1477 kiblnd_conn_addref(conn); /* 1 ref for me... */
1479 read_unlock_irqrestore(g_lock, flags);
1482 kiblnd_queue_tx(tx, conn);
1483 kiblnd_conn_decref(conn); /* ...to here */
1487 read_unlock(g_lock);
1488 /* Re-try with a write lock */
1491 peer_ni = kiblnd_find_peer_locked(ni, nid);
1492 if (peer_ni != NULL) {
1493 if (list_empty(&peer_ni->ibp_conns)) {
1494 /* found a peer_ni, but it's still connecting... */
1495 LASSERT(kiblnd_peer_connecting(peer_ni));
1497 list_add_tail(&tx->tx_list,
1498 &peer_ni->ibp_tx_queue);
1499 write_unlock_irqrestore(g_lock, flags);
1501 conn = kiblnd_get_conn_locked(peer_ni);
1502 kiblnd_conn_addref(conn); /* 1 ref for me... */
1504 write_unlock_irqrestore(g_lock, flags);
1507 kiblnd_queue_tx(tx, conn);
1508 kiblnd_conn_decref(conn); /* ...to here */
1513 write_unlock_irqrestore(g_lock, flags);
1515 /* Allocate a peer_ni ready to add to the peer_ni table and retry */
1516 rc = kiblnd_create_peer(ni, &peer_ni, nid);
1518 CERROR("Can't create peer_ni %s\n", libcfs_nid2str(nid));
1520 tx->tx_status = -EHOSTUNREACH;
1527 write_lock_irqsave(g_lock, flags);
1529 peer2 = kiblnd_find_peer_locked(ni, nid);
1530 if (peer2 != NULL) {
1531 if (list_empty(&peer2->ibp_conns)) {
1532 /* found a peer_ni, but it's still connecting... */
1533 LASSERT(kiblnd_peer_connecting(peer2));
1535 list_add_tail(&tx->tx_list,
1536 &peer2->ibp_tx_queue);
1537 write_unlock_irqrestore(g_lock, flags);
1539 conn = kiblnd_get_conn_locked(peer2);
1540 kiblnd_conn_addref(conn); /* 1 ref for me... */
1542 write_unlock_irqrestore(g_lock, flags);
1545 kiblnd_queue_tx(tx, conn);
1546 kiblnd_conn_decref(conn); /* ...to here */
1549 kiblnd_peer_decref(peer_ni);
1553 /* Brand new peer_ni */
1554 LASSERT(peer_ni->ibp_connecting == 0);
1555 tunables = &peer_ni->ibp_ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
1556 peer_ni->ibp_connecting = tunables->lnd_conns_per_peer;
1558 /* always called with a ref on ni, which prevents ni being shutdown */
1559 LASSERT(((kib_net_t *)ni->ni_data)->ibn_shutdown == 0);
1562 list_add_tail(&tx->tx_list, &peer_ni->ibp_tx_queue);
1564 kiblnd_peer_addref(peer_ni);
1565 list_add_tail(&peer_ni->ibp_list, kiblnd_nid2peerlist(nid));
1567 write_unlock_irqrestore(g_lock, flags);
1569 for (i = 0; i < tunables->lnd_conns_per_peer; i++)
1570 kiblnd_connect_peer(peer_ni);
1571 kiblnd_peer_decref(peer_ni);
1575 kiblnd_send(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg)
1577 struct lnet_hdr *hdr = &lntmsg->msg_hdr;
1578 int type = lntmsg->msg_type;
1579 struct lnet_process_id target = lntmsg->msg_target;
1580 int target_is_router = lntmsg->msg_target_is_router;
1581 int routing = lntmsg->msg_routing;
1582 unsigned int payload_niov = lntmsg->msg_niov;
1583 struct kvec *payload_iov = lntmsg->msg_iov;
1584 lnet_kiov_t *payload_kiov = lntmsg->msg_kiov;
1585 unsigned int payload_offset = lntmsg->msg_offset;
1586 unsigned int payload_nob = lntmsg->msg_len;
1588 kib_rdma_desc_t *rd;
1593 /* NB 'private' is different depending on what we're sending.... */
1595 CDEBUG(D_NET, "sending %d bytes in %d frags to %s\n",
1596 payload_nob, payload_niov, libcfs_id2str(target));
1598 LASSERT (payload_nob == 0 || payload_niov > 0);
1599 LASSERT (payload_niov <= LNET_MAX_IOV);
1601 /* Thread context */
1602 LASSERT (!in_interrupt());
1603 /* payload is either all vaddrs or all pages */
1604 LASSERT (!(payload_kiov != NULL && payload_iov != NULL));
1612 LASSERT (payload_nob == 0);
1616 if (routing || target_is_router)
1617 break; /* send IMMEDIATE */
1619 /* is the REPLY message too small for RDMA? */
1620 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[lntmsg->msg_md->md_length]);
1621 if (nob <= IBLND_MSG_SIZE)
1622 break; /* send IMMEDIATE */
1624 tx = kiblnd_get_idle_tx(ni, target.nid);
1626 CERROR("Can't allocate txd for GET to %s\n",
1627 libcfs_nid2str(target.nid));
1632 rd = &ibmsg->ibm_u.get.ibgm_rd;
1633 if ((lntmsg->msg_md->md_options & LNET_MD_KIOV) == 0)
1634 rc = kiblnd_setup_rd_iov(ni, tx, rd,
1635 lntmsg->msg_md->md_niov,
1636 lntmsg->msg_md->md_iov.iov,
1637 0, lntmsg->msg_md->md_length);
1639 rc = kiblnd_setup_rd_kiov(ni, tx, rd,
1640 lntmsg->msg_md->md_niov,
1641 lntmsg->msg_md->md_iov.kiov,
1642 0, lntmsg->msg_md->md_length);
1644 CERROR("Can't setup GET sink for %s: %d\n",
1645 libcfs_nid2str(target.nid), rc);
1650 nob = offsetof(kib_get_msg_t, ibgm_rd.rd_frags[rd->rd_nfrags]);
1651 ibmsg->ibm_u.get.ibgm_cookie = tx->tx_cookie;
1652 ibmsg->ibm_u.get.ibgm_hdr = *hdr;
1654 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_GET_REQ, nob);
1656 tx->tx_lntmsg[1] = lnet_create_reply_msg(ni, lntmsg);
1657 if (tx->tx_lntmsg[1] == NULL) {
1658 CERROR("Can't create reply for GET -> %s\n",
1659 libcfs_nid2str(target.nid));
1664 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg[0,1] on completion */
1665 tx->tx_waiting = 1; /* waiting for GET_DONE */
1666 kiblnd_launch_tx(ni, tx, target.nid);
1669 case LNET_MSG_REPLY:
1671 /* Is the payload small enough not to need RDMA? */
1672 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob]);
1673 if (nob <= IBLND_MSG_SIZE)
1674 break; /* send IMMEDIATE */
1676 tx = kiblnd_get_idle_tx(ni, target.nid);
1678 CERROR("Can't allocate %s txd for %s\n",
1679 type == LNET_MSG_PUT ? "PUT" : "REPLY",
1680 libcfs_nid2str(target.nid));
1684 if (payload_kiov == NULL)
1685 rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1686 payload_niov, payload_iov,
1687 payload_offset, payload_nob);
1689 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1690 payload_niov, payload_kiov,
1691 payload_offset, payload_nob);
1693 CERROR("Can't setup PUT src for %s: %d\n",
1694 libcfs_nid2str(target.nid), rc);
1700 ibmsg->ibm_u.putreq.ibprm_hdr = *hdr;
1701 ibmsg->ibm_u.putreq.ibprm_cookie = tx->tx_cookie;
1702 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_REQ, sizeof(kib_putreq_msg_t));
1704 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1705 tx->tx_waiting = 1; /* waiting for PUT_{ACK,NAK} */
1706 kiblnd_launch_tx(ni, tx, target.nid);
1710 /* send IMMEDIATE */
1712 LASSERT (offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob])
1715 tx = kiblnd_get_idle_tx(ni, target.nid);
1717 CERROR ("Can't send %d to %s: tx descs exhausted\n",
1718 type, libcfs_nid2str(target.nid));
1723 ibmsg->ibm_u.immediate.ibim_hdr = *hdr;
1725 if (payload_kiov != NULL)
1726 lnet_copy_kiov2flat(IBLND_MSG_SIZE, ibmsg,
1727 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1728 payload_niov, payload_kiov,
1729 payload_offset, payload_nob);
1731 lnet_copy_iov2flat(IBLND_MSG_SIZE, ibmsg,
1732 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1733 payload_niov, payload_iov,
1734 payload_offset, payload_nob);
1736 nob = offsetof(kib_immediate_msg_t, ibim_payload[payload_nob]);
1737 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_IMMEDIATE, nob);
1739 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1740 kiblnd_launch_tx(ni, tx, target.nid);
1745 kiblnd_reply(struct lnet_ni *ni, kib_rx_t *rx, struct lnet_msg *lntmsg)
1747 struct lnet_process_id target = lntmsg->msg_target;
1748 unsigned int niov = lntmsg->msg_niov;
1749 struct kvec *iov = lntmsg->msg_iov;
1750 lnet_kiov_t *kiov = lntmsg->msg_kiov;
1751 unsigned int offset = lntmsg->msg_offset;
1752 unsigned int nob = lntmsg->msg_len;
1756 tx = kiblnd_get_idle_tx(ni, rx->rx_conn->ibc_peer->ibp_nid);
1758 CERROR("Can't get tx for REPLY to %s\n",
1759 libcfs_nid2str(target.nid));
1765 else if (kiov == NULL)
1766 rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1767 niov, iov, offset, nob);
1769 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1770 niov, kiov, offset, nob);
1773 CERROR("Can't setup GET src for %s: %d\n",
1774 libcfs_nid2str(target.nid), rc);
1778 rc = kiblnd_init_rdma(rx->rx_conn, tx,
1779 IBLND_MSG_GET_DONE, nob,
1780 &rx->rx_msg->ibm_u.get.ibgm_rd,
1781 rx->rx_msg->ibm_u.get.ibgm_cookie);
1783 CERROR("Can't setup rdma for GET from %s: %d\n",
1784 libcfs_nid2str(target.nid), rc);
1789 /* No RDMA: local completion may happen now! */
1790 lnet_finalize(lntmsg, 0);
1792 /* RDMA: lnet_finalize(lntmsg) when it
1794 tx->tx_lntmsg[0] = lntmsg;
1797 kiblnd_queue_tx(tx, rx->rx_conn);
1803 lnet_finalize(lntmsg, -EIO);
1807 kiblnd_recv(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg,
1808 int delayed, unsigned int niov, struct kvec *iov, lnet_kiov_t *kiov,
1809 unsigned int offset, unsigned int mlen, unsigned int rlen)
1811 kib_rx_t *rx = private;
1812 kib_msg_t *rxmsg = rx->rx_msg;
1813 kib_conn_t *conn = rx->rx_conn;
1817 int post_credit = IBLND_POSTRX_PEER_CREDIT;
1820 LASSERT (mlen <= rlen);
1821 LASSERT (!in_interrupt());
1822 /* Either all pages or all vaddrs */
1823 LASSERT (!(kiov != NULL && iov != NULL));
1825 switch (rxmsg->ibm_type) {
1829 case IBLND_MSG_IMMEDIATE:
1830 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[rlen]);
1831 if (nob > rx->rx_nob) {
1832 CERROR ("Immediate message from %s too big: %d(%d)\n",
1833 libcfs_nid2str(rxmsg->ibm_u.immediate.ibim_hdr.src_nid),
1840 lnet_copy_flat2kiov(niov, kiov, offset,
1841 IBLND_MSG_SIZE, rxmsg,
1842 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1845 lnet_copy_flat2iov(niov, iov, offset,
1846 IBLND_MSG_SIZE, rxmsg,
1847 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1849 lnet_finalize(lntmsg, 0);
1852 case IBLND_MSG_PUT_REQ: {
1854 kib_rdma_desc_t *rd;
1855 ibprm_cookie = rxmsg->ibm_u.putreq.ibprm_cookie;
1858 lnet_finalize(lntmsg, 0);
1859 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK,
1864 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
1866 CERROR("Can't allocate tx for %s\n",
1867 libcfs_nid2str(conn->ibc_peer->ibp_nid));
1868 /* Not replying will break the connection */
1874 rd = &txmsg->ibm_u.putack.ibpam_rd;
1876 rc = kiblnd_setup_rd_iov(ni, tx, rd,
1877 niov, iov, offset, mlen);
1879 rc = kiblnd_setup_rd_kiov(ni, tx, rd,
1880 niov, kiov, offset, mlen);
1882 CERROR("Can't setup PUT sink for %s: %d\n",
1883 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
1885 /* tell peer_ni it's over */
1886 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK,
1891 nob = offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[rd->rd_nfrags]);
1892 txmsg->ibm_u.putack.ibpam_src_cookie = ibprm_cookie;
1893 txmsg->ibm_u.putack.ibpam_dst_cookie = tx->tx_cookie;
1895 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_ACK, nob);
1897 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1898 tx->tx_waiting = 1; /* waiting for PUT_DONE */
1899 kiblnd_queue_tx(tx, conn);
1901 /* reposted buffer reserved for PUT_DONE */
1902 post_credit = IBLND_POSTRX_NO_CREDIT;
1906 case IBLND_MSG_GET_REQ:
1907 if (lntmsg != NULL) {
1908 /* Optimized GET; RDMA lntmsg's payload */
1909 kiblnd_reply(ni, rx, lntmsg);
1911 /* GET didn't match anything */
1912 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_GET_DONE,
1914 rxmsg->ibm_u.get.ibgm_cookie);
1919 kiblnd_post_rx(rx, post_credit);
1924 kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name)
1926 struct task_struct *task = kthread_run(fn, arg, name);
1929 return PTR_ERR(task);
1931 atomic_inc(&kiblnd_data.kib_nthreads);
1936 kiblnd_thread_fini (void)
1938 atomic_dec (&kiblnd_data.kib_nthreads);
1942 kiblnd_peer_alive (kib_peer_ni_t *peer_ni)
1944 /* This is racy, but everyone's only writing cfs_time_current() */
1945 peer_ni->ibp_last_alive = cfs_time_current();
1950 kiblnd_peer_notify (kib_peer_ni_t *peer_ni)
1953 cfs_time_t last_alive = 0;
1954 unsigned long flags;
1956 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1958 if (kiblnd_peer_idle(peer_ni) && peer_ni->ibp_error != 0) {
1959 error = peer_ni->ibp_error;
1960 peer_ni->ibp_error = 0;
1962 last_alive = peer_ni->ibp_last_alive;
1965 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1968 lnet_notify(peer_ni->ibp_ni,
1969 peer_ni->ibp_nid, 0, last_alive);
1973 kiblnd_close_conn_locked (kib_conn_t *conn, int error)
1975 /* This just does the immediate housekeeping. 'error' is zero for a
1976 * normal shutdown which can happen only after the connection has been
1977 * established. If the connection is established, schedule the
1978 * connection to be finished off by the connd. Otherwise the connd is
1979 * already dealing with it (either to set it up or tear it down).
1980 * Caller holds kib_global_lock exclusively in irq context */
1981 kib_peer_ni_t *peer_ni = conn->ibc_peer;
1983 unsigned long flags;
1985 LASSERT (error != 0 || conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1987 if (error != 0 && conn->ibc_comms_error == 0)
1988 conn->ibc_comms_error = error;
1990 if (conn->ibc_state != IBLND_CONN_ESTABLISHED)
1991 return; /* already being handled */
1994 list_empty(&conn->ibc_tx_noops) &&
1995 list_empty(&conn->ibc_tx_queue) &&
1996 list_empty(&conn->ibc_tx_queue_rsrvd) &&
1997 list_empty(&conn->ibc_tx_queue_nocred) &&
1998 list_empty(&conn->ibc_active_txs)) {
1999 CDEBUG(D_NET, "closing conn to %s\n",
2000 libcfs_nid2str(peer_ni->ibp_nid));
2002 CNETERR("Closing conn to %s: error %d%s%s%s%s%s\n",
2003 libcfs_nid2str(peer_ni->ibp_nid), error,
2004 list_empty(&conn->ibc_tx_queue) ? "" : "(sending)",
2005 list_empty(&conn->ibc_tx_noops) ? "" : "(sending_noops)",
2006 list_empty(&conn->ibc_tx_queue_rsrvd) ?
2007 "" : "(sending_rsrvd)",
2008 list_empty(&conn->ibc_tx_queue_nocred) ?
2009 "" : "(sending_nocred)",
2010 list_empty(&conn->ibc_active_txs) ? "" : "(waiting)");
2013 dev = ((kib_net_t *)peer_ni->ibp_ni->ni_data)->ibn_dev;
2014 if (peer_ni->ibp_next_conn == conn)
2015 /* clear next_conn so it won't be used */
2016 peer_ni->ibp_next_conn = NULL;
2017 list_del(&conn->ibc_list);
2018 /* connd (see below) takes over ibc_list's ref */
2020 if (list_empty(&peer_ni->ibp_conns) && /* no more conns */
2021 kiblnd_peer_active(peer_ni)) { /* still in peer_ni table */
2022 kiblnd_unlink_peer_locked(peer_ni);
2024 /* set/clear error on last conn */
2025 peer_ni->ibp_error = conn->ibc_comms_error;
2028 kiblnd_set_conn_state(conn, IBLND_CONN_CLOSING);
2031 kiblnd_dev_can_failover(dev)) {
2032 list_add_tail(&dev->ibd_fail_list,
2033 &kiblnd_data.kib_failed_devs);
2034 wake_up(&kiblnd_data.kib_failover_waitq);
2037 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
2039 list_add_tail(&conn->ibc_list, &kiblnd_data.kib_connd_conns);
2040 wake_up(&kiblnd_data.kib_connd_waitq);
2042 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
2046 kiblnd_close_conn(kib_conn_t *conn, int error)
2048 unsigned long flags;
2050 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2052 kiblnd_close_conn_locked(conn, error);
2054 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2058 kiblnd_handle_early_rxs(kib_conn_t *conn)
2060 unsigned long flags;
2063 LASSERT(!in_interrupt());
2064 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
2066 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2067 while (!list_empty(&conn->ibc_early_rxs)) {
2068 rx = list_entry(conn->ibc_early_rxs.next,
2070 list_del(&rx->rx_list);
2071 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2073 kiblnd_handle_rx(rx);
2075 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2077 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2081 kiblnd_abort_txs(kib_conn_t *conn, struct list_head *txs)
2083 struct list_head zombies = LIST_HEAD_INIT(zombies);
2084 struct list_head *tmp;
2085 struct list_head *nxt;
2088 spin_lock(&conn->ibc_lock);
2090 list_for_each_safe(tmp, nxt, txs) {
2091 tx = list_entry(tmp, kib_tx_t, tx_list);
2093 if (txs == &conn->ibc_active_txs) {
2094 LASSERT(!tx->tx_queued);
2095 LASSERT(tx->tx_waiting ||
2096 tx->tx_sending != 0);
2098 LASSERT(tx->tx_queued);
2101 tx->tx_status = -ECONNABORTED;
2104 if (tx->tx_sending == 0) {
2106 list_del(&tx->tx_list);
2107 list_add(&tx->tx_list, &zombies);
2111 spin_unlock(&conn->ibc_lock);
2113 kiblnd_txlist_done(&zombies, -ECONNABORTED);
2117 kiblnd_finalise_conn (kib_conn_t *conn)
2119 LASSERT (!in_interrupt());
2120 LASSERT (conn->ibc_state > IBLND_CONN_INIT);
2122 kiblnd_set_conn_state(conn, IBLND_CONN_DISCONNECTED);
2124 /* abort_receives moves QP state to IB_QPS_ERR. This is only required
2125 * for connections that didn't get as far as being connected, because
2126 * rdma_disconnect() does this for free. */
2127 kiblnd_abort_receives(conn);
2129 /* Complete all tx descs not waiting for sends to complete.
2130 * NB we should be safe from RDMA now that the QP has changed state */
2132 kiblnd_abort_txs(conn, &conn->ibc_tx_noops);
2133 kiblnd_abort_txs(conn, &conn->ibc_tx_queue);
2134 kiblnd_abort_txs(conn, &conn->ibc_tx_queue_rsrvd);
2135 kiblnd_abort_txs(conn, &conn->ibc_tx_queue_nocred);
2136 kiblnd_abort_txs(conn, &conn->ibc_active_txs);
2138 kiblnd_handle_early_rxs(conn);
2142 kiblnd_peer_connect_failed(kib_peer_ni_t *peer_ni, int active, int error)
2144 struct list_head zombies = LIST_HEAD_INIT(zombies);
2145 unsigned long flags;
2147 LASSERT (error != 0);
2148 LASSERT (!in_interrupt());
2150 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2153 LASSERT(peer_ni->ibp_connecting > 0);
2154 peer_ni->ibp_connecting--;
2156 LASSERT (peer_ni->ibp_accepting > 0);
2157 peer_ni->ibp_accepting--;
2160 if (kiblnd_peer_connecting(peer_ni)) {
2161 /* another connection attempt under way... */
2162 write_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2167 peer_ni->ibp_reconnected = 0;
2168 if (list_empty(&peer_ni->ibp_conns)) {
2169 /* Take peer_ni's blocked transmits to complete with error */
2170 list_add(&zombies, &peer_ni->ibp_tx_queue);
2171 list_del_init(&peer_ni->ibp_tx_queue);
2173 if (kiblnd_peer_active(peer_ni))
2174 kiblnd_unlink_peer_locked(peer_ni);
2176 peer_ni->ibp_error = error;
2178 /* Can't have blocked transmits if there are connections */
2179 LASSERT(list_empty(&peer_ni->ibp_tx_queue));
2182 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2184 kiblnd_peer_notify(peer_ni);
2186 if (list_empty(&zombies))
2189 CNETERR("Deleting messages for %s: connection failed\n",
2190 libcfs_nid2str(peer_ni->ibp_nid));
2192 kiblnd_txlist_done(&zombies, -EHOSTUNREACH);
2196 kiblnd_connreq_done(kib_conn_t *conn, int status)
2198 kib_peer_ni_t *peer_ni = conn->ibc_peer;
2200 struct list_head txs;
2201 unsigned long flags;
2204 active = (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2206 CDEBUG(D_NET,"%s: active(%d), version(%x), status(%d)\n",
2207 libcfs_nid2str(peer_ni->ibp_nid), active,
2208 conn->ibc_version, status);
2210 LASSERT (!in_interrupt());
2211 LASSERT ((conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT &&
2212 peer_ni->ibp_connecting > 0) ||
2213 (conn->ibc_state == IBLND_CONN_PASSIVE_WAIT &&
2214 peer_ni->ibp_accepting > 0));
2216 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
2217 conn->ibc_connvars = NULL;
2220 /* failed to establish connection */
2221 kiblnd_peer_connect_failed(peer_ni, active, status);
2222 kiblnd_finalise_conn(conn);
2226 /* connection established */
2227 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2229 conn->ibc_last_send = jiffies;
2230 kiblnd_set_conn_state(conn, IBLND_CONN_ESTABLISHED);
2231 kiblnd_peer_alive(peer_ni);
2233 /* Add conn to peer_ni's list and nuke any dangling conns from a different
2234 * peer_ni instance... */
2235 kiblnd_conn_addref(conn); /* +1 ref for ibc_list */
2236 list_add(&conn->ibc_list, &peer_ni->ibp_conns);
2237 peer_ni->ibp_reconnected = 0;
2239 peer_ni->ibp_connecting--;
2241 peer_ni->ibp_accepting--;
2243 if (peer_ni->ibp_version == 0) {
2244 peer_ni->ibp_version = conn->ibc_version;
2245 peer_ni->ibp_incarnation = conn->ibc_incarnation;
2248 if (peer_ni->ibp_version != conn->ibc_version ||
2249 peer_ni->ibp_incarnation != conn->ibc_incarnation) {
2250 kiblnd_close_stale_conns_locked(peer_ni, conn->ibc_version,
2251 conn->ibc_incarnation);
2252 peer_ni->ibp_version = conn->ibc_version;
2253 peer_ni->ibp_incarnation = conn->ibc_incarnation;
2256 /* grab pending txs while I have the lock */
2257 list_add(&txs, &peer_ni->ibp_tx_queue);
2258 list_del_init(&peer_ni->ibp_tx_queue);
2260 if (!kiblnd_peer_active(peer_ni) || /* peer_ni has been deleted */
2261 conn->ibc_comms_error != 0) { /* error has happened already */
2263 /* start to shut down connection */
2264 kiblnd_close_conn_locked(conn, -ECONNABORTED);
2265 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2267 kiblnd_txlist_done(&txs, -ECONNABORTED);
2272 /* +1 ref for myself, this connection is visible to other threads
2273 * now, refcount of peer:ibp_conns can be released by connection
2274 * close from either a different thread, or the calling of
2275 * kiblnd_check_sends_locked() below. See bz21911 for details.
2277 kiblnd_conn_addref(conn);
2278 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2280 /* Schedule blocked txs
2281 * Note: if we are running with conns_per_peer > 1, these blocked
2282 * txs will all get scheduled to the first connection which gets
2283 * scheduled. We won't be using round robin on this first batch.
2285 spin_lock(&conn->ibc_lock);
2286 while (!list_empty(&txs)) {
2287 tx = list_entry(txs.next, kib_tx_t, tx_list);
2288 list_del(&tx->tx_list);
2290 kiblnd_queue_tx_locked(tx, conn);
2292 kiblnd_check_sends_locked(conn);
2293 spin_unlock(&conn->ibc_lock);
2295 /* schedule blocked rxs */
2296 kiblnd_handle_early_rxs(conn);
2297 kiblnd_conn_decref(conn);
2301 kiblnd_reject(struct rdma_cm_id *cmid, kib_rej_t *rej)
2305 rc = rdma_reject(cmid, rej, sizeof(*rej));
2308 CWARN("Error %d sending reject\n", rc);
2312 kiblnd_passive_connect(struct rdma_cm_id *cmid, void *priv, int priv_nob)
2314 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2315 kib_msg_t *reqmsg = priv;
2318 kib_peer_ni_t *peer_ni;
2319 kib_peer_ni_t *peer2;
2321 struct lnet_ni *ni = NULL;
2322 kib_net_t *net = NULL;
2324 struct rdma_conn_param cp;
2326 int version = IBLND_MSG_VERSION;
2327 unsigned long flags;
2329 struct sockaddr_in *peer_addr;
2330 LASSERT (!in_interrupt());
2332 /* cmid inherits 'context' from the corresponding listener id */
2333 ibdev = (kib_dev_t *)cmid->context;
2334 LASSERT (ibdev != NULL);
2336 memset(&rej, 0, sizeof(rej));
2337 rej.ibr_magic = IBLND_MSG_MAGIC;
2338 rej.ibr_why = IBLND_REJECT_FATAL;
2339 rej.ibr_cp.ibcp_max_msg_size = IBLND_MSG_SIZE;
2341 peer_addr = (struct sockaddr_in *)&(cmid->route.addr.dst_addr);
2342 if (*kiblnd_tunables.kib_require_priv_port &&
2343 ntohs(peer_addr->sin_port) >= PROT_SOCK) {
2344 __u32 ip = ntohl(peer_addr->sin_addr.s_addr);
2345 CERROR("peer_ni's port (%pI4h:%hu) is not privileged\n",
2346 &ip, ntohs(peer_addr->sin_port));
2350 if (priv_nob < offsetof(kib_msg_t, ibm_type)) {
2351 CERROR("Short connection request\n");
2355 /* Future protocol version compatibility support! If the
2356 * o2iblnd-specific protocol changes, or when LNET unifies
2357 * protocols over all LNDs, the initial connection will
2358 * negotiate a protocol version. I trap this here to avoid
2359 * console errors; the reject tells the peer_ni which protocol I
2361 if (reqmsg->ibm_magic == LNET_PROTO_MAGIC ||
2362 reqmsg->ibm_magic == __swab32(LNET_PROTO_MAGIC))
2364 if (reqmsg->ibm_magic == IBLND_MSG_MAGIC &&
2365 reqmsg->ibm_version != IBLND_MSG_VERSION &&
2366 reqmsg->ibm_version != IBLND_MSG_VERSION_1)
2368 if (reqmsg->ibm_magic == __swab32(IBLND_MSG_MAGIC) &&
2369 reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION) &&
2370 reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION_1))
2373 rc = kiblnd_unpack_msg(reqmsg, priv_nob);
2375 CERROR("Can't parse connection request: %d\n", rc);
2379 nid = reqmsg->ibm_srcnid;
2380 ni = lnet_nid2ni_addref(reqmsg->ibm_dstnid);
2383 net = (kib_net_t *)ni->ni_data;
2384 rej.ibr_incarnation = net->ibn_incarnation;
2387 if (ni == NULL || /* no matching net */
2388 ni->ni_nid != reqmsg->ibm_dstnid || /* right NET, wrong NID! */
2389 net->ibn_dev != ibdev) { /* wrong device */
2390 CERROR("Can't accept conn from %s on %s (%s:%d:%pI4h): "
2391 "bad dst nid %s\n", libcfs_nid2str(nid),
2392 ni == NULL ? "NA" : libcfs_nid2str(ni->ni_nid),
2393 ibdev->ibd_ifname, ibdev->ibd_nnets,
2395 libcfs_nid2str(reqmsg->ibm_dstnid));
2400 /* check time stamp as soon as possible */
2401 if (reqmsg->ibm_dststamp != 0 &&
2402 reqmsg->ibm_dststamp != net->ibn_incarnation) {
2403 CWARN("Stale connection request\n");
2404 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2408 /* I can accept peer_ni's version */
2409 version = reqmsg->ibm_version;
2411 if (reqmsg->ibm_type != IBLND_MSG_CONNREQ) {
2412 CERROR("Unexpected connreq msg type: %x from %s\n",
2413 reqmsg->ibm_type, libcfs_nid2str(nid));
2417 if (reqmsg->ibm_u.connparams.ibcp_queue_depth >
2418 kiblnd_msg_queue_size(version, ni)) {
2419 CERROR("Can't accept conn from %s, queue depth too large: "
2420 " %d (<=%d wanted)\n",
2421 libcfs_nid2str(nid),
2422 reqmsg->ibm_u.connparams.ibcp_queue_depth,
2423 kiblnd_msg_queue_size(version, ni));
2425 if (version == IBLND_MSG_VERSION)
2426 rej.ibr_why = IBLND_REJECT_MSG_QUEUE_SIZE;
2431 if (reqmsg->ibm_u.connparams.ibcp_max_frags >
2432 IBLND_MAX_RDMA_FRAGS) {
2433 CWARN("Can't accept conn from %s (version %x): "
2434 "max_frags %d too large (%d wanted)\n",
2435 libcfs_nid2str(nid), version,
2436 reqmsg->ibm_u.connparams.ibcp_max_frags,
2437 IBLND_MAX_RDMA_FRAGS);
2439 if (version >= IBLND_MSG_VERSION)
2440 rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2443 } else if (reqmsg->ibm_u.connparams.ibcp_max_frags <
2444 IBLND_MAX_RDMA_FRAGS &&
2445 net->ibn_fmr_ps == NULL) {
2446 CWARN("Can't accept conn from %s (version %x): "
2447 "max_frags %d incompatible without FMR pool "
2449 libcfs_nid2str(nid), version,
2450 reqmsg->ibm_u.connparams.ibcp_max_frags,
2451 IBLND_MAX_RDMA_FRAGS);
2453 if (version == IBLND_MSG_VERSION)
2454 rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2459 if (reqmsg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2460 CERROR("Can't accept %s: message size %d too big (%d max)\n",
2461 libcfs_nid2str(nid),
2462 reqmsg->ibm_u.connparams.ibcp_max_msg_size,
2467 /* assume 'nid' is a new peer_ni; create */
2468 rc = kiblnd_create_peer(ni, &peer_ni, nid);
2470 CERROR("Can't create peer_ni for %s\n", libcfs_nid2str(nid));
2471 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2475 /* We have validated the peer's parameters so use those */
2476 peer_ni->ibp_max_frags = reqmsg->ibm_u.connparams.ibcp_max_frags;
2477 peer_ni->ibp_queue_depth = reqmsg->ibm_u.connparams.ibcp_queue_depth;
2479 write_lock_irqsave(g_lock, flags);
2481 peer2 = kiblnd_find_peer_locked(ni, nid);
2482 if (peer2 != NULL) {
2483 if (peer2->ibp_version == 0) {
2484 peer2->ibp_version = version;
2485 peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2488 /* not the guy I've talked with */
2489 if (peer2->ibp_incarnation != reqmsg->ibm_srcstamp ||
2490 peer2->ibp_version != version) {
2491 kiblnd_close_peer_conns_locked(peer2, -ESTALE);
2493 if (kiblnd_peer_active(peer2)) {
2494 peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2495 peer2->ibp_version = version;
2497 write_unlock_irqrestore(g_lock, flags);
2499 CWARN("Conn stale %s version %x/%x incarnation %llu/%llu\n",
2500 libcfs_nid2str(nid), peer2->ibp_version, version,
2501 peer2->ibp_incarnation, reqmsg->ibm_srcstamp);
2503 kiblnd_peer_decref(peer_ni);
2504 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2508 /* Tie-break connection race in favour of the higher NID.
2509 * If we keep running into a race condition multiple times,
2510 * we have to assume that the connection attempt with the
2511 * higher NID is stuck in a connecting state and will never
2512 * recover. As such, we pass through this if-block and let
2513 * the lower NID connection win so we can move forward.
2515 if (peer2->ibp_connecting != 0 &&
2516 nid < ni->ni_nid && peer2->ibp_races <
2517 MAX_CONN_RACES_BEFORE_ABORT) {
2519 write_unlock_irqrestore(g_lock, flags);
2521 CDEBUG(D_NET, "Conn race %s\n",
2522 libcfs_nid2str(peer2->ibp_nid));
2524 kiblnd_peer_decref(peer_ni);
2525 rej.ibr_why = IBLND_REJECT_CONN_RACE;
2528 if (peer2->ibp_races >= MAX_CONN_RACES_BEFORE_ABORT)
2529 CNETERR("Conn race %s: unresolved after %d attempts, letting lower NID win\n",
2530 libcfs_nid2str(peer2->ibp_nid),
2531 MAX_CONN_RACES_BEFORE_ABORT);
2533 * passive connection is allowed even this peer_ni is waiting for
2536 peer2->ibp_reconnecting = 0;
2537 peer2->ibp_races = 0;
2538 peer2->ibp_accepting++;
2539 kiblnd_peer_addref(peer2);
2541 /* Race with kiblnd_launch_tx (active connect) to create peer_ni
2542 * so copy validated parameters since we now know what the
2543 * peer_ni's limits are */
2544 peer2->ibp_max_frags = peer_ni->ibp_max_frags;
2545 peer2->ibp_queue_depth = peer_ni->ibp_queue_depth;
2547 write_unlock_irqrestore(g_lock, flags);
2548 kiblnd_peer_decref(peer_ni);
2551 /* Brand new peer_ni */
2552 LASSERT (peer_ni->ibp_accepting == 0);
2553 LASSERT (peer_ni->ibp_version == 0 &&
2554 peer_ni->ibp_incarnation == 0);
2556 peer_ni->ibp_accepting = 1;
2557 peer_ni->ibp_version = version;
2558 peer_ni->ibp_incarnation = reqmsg->ibm_srcstamp;
2560 /* I have a ref on ni that prevents it being shutdown */
2561 LASSERT (net->ibn_shutdown == 0);
2563 kiblnd_peer_addref(peer_ni);
2564 list_add_tail(&peer_ni->ibp_list, kiblnd_nid2peerlist(nid));
2566 write_unlock_irqrestore(g_lock, flags);
2569 conn = kiblnd_create_conn(peer_ni, cmid, IBLND_CONN_PASSIVE_WAIT, version);
2571 kiblnd_peer_connect_failed(peer_ni, 0, -ENOMEM);
2572 kiblnd_peer_decref(peer_ni);
2573 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2577 /* conn now "owns" cmid, so I return success from here on to ensure the
2578 * CM callback doesn't destroy cmid. */
2579 conn->ibc_incarnation = reqmsg->ibm_srcstamp;
2580 conn->ibc_credits = conn->ibc_queue_depth;
2581 conn->ibc_reserved_credits = conn->ibc_queue_depth;
2582 LASSERT(conn->ibc_credits + conn->ibc_reserved_credits +
2583 IBLND_OOB_MSGS(version) <= IBLND_RX_MSGS(conn));
2585 ackmsg = &conn->ibc_connvars->cv_msg;
2586 memset(ackmsg, 0, sizeof(*ackmsg));
2588 kiblnd_init_msg(ackmsg, IBLND_MSG_CONNACK,
2589 sizeof(ackmsg->ibm_u.connparams));
2590 ackmsg->ibm_u.connparams.ibcp_queue_depth = conn->ibc_queue_depth;
2591 ackmsg->ibm_u.connparams.ibcp_max_frags = conn->ibc_max_frags;
2592 ackmsg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2594 kiblnd_pack_msg(ni, ackmsg, version, 0, nid, reqmsg->ibm_srcstamp);
2596 memset(&cp, 0, sizeof(cp));
2597 cp.private_data = ackmsg;
2598 cp.private_data_len = ackmsg->ibm_nob;
2599 cp.responder_resources = 0; /* No atomic ops or RDMA reads */
2600 cp.initiator_depth = 0;
2601 cp.flow_control = 1;
2602 cp.retry_count = *kiblnd_tunables.kib_retry_count;
2603 cp.rnr_retry_count = *kiblnd_tunables.kib_rnr_retry_count;
2605 CDEBUG(D_NET, "Accept %s\n", libcfs_nid2str(nid));
2607 rc = rdma_accept(cmid, &cp);
2609 CERROR("Can't accept %s: %d\n", libcfs_nid2str(nid), rc);
2610 rej.ibr_version = version;
2611 rej.ibr_why = IBLND_REJECT_FATAL;
2613 kiblnd_reject(cmid, &rej);
2614 kiblnd_connreq_done(conn, rc);
2615 kiblnd_conn_decref(conn);
2623 rej.ibr_cp.ibcp_queue_depth =
2624 kiblnd_msg_queue_size(version, ni);
2625 rej.ibr_cp.ibcp_max_frags = IBLND_MAX_RDMA_FRAGS;
2629 rej.ibr_version = version;
2630 kiblnd_reject(cmid, &rej);
2632 return -ECONNREFUSED;
2636 kiblnd_check_reconnect(kib_conn_t *conn, int version,
2637 __u64 incarnation, int why, kib_connparams_t *cp)
2639 rwlock_t *glock = &kiblnd_data.kib_global_lock;
2640 kib_peer_ni_t *peer_ni = conn->ibc_peer;
2642 int msg_size = IBLND_MSG_SIZE;
2646 unsigned long flags;
2648 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2649 LASSERT(peer_ni->ibp_connecting > 0); /* 'conn' at least */
2652 msg_size = cp->ibcp_max_msg_size;
2653 frag_num = cp->ibcp_max_frags;
2654 queue_dep = cp->ibcp_queue_depth;
2657 write_lock_irqsave(glock, flags);
2658 /* retry connection if it's still needed and no other connection
2659 * attempts (active or passive) are in progress
2660 * NB: reconnect is still needed even when ibp_tx_queue is
2661 * empty if ibp_version != version because reconnect may be
2662 * initiated by kiblnd_query() */
2663 reconnect = (!list_empty(&peer_ni->ibp_tx_queue) ||
2664 peer_ni->ibp_version != version) &&
2665 peer_ni->ibp_connecting &&
2666 peer_ni->ibp_accepting == 0;
2677 case IBLND_REJECT_RDMA_FRAGS: {
2678 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2681 reason = "can't negotiate max frags";
2684 tunables = &peer_ni->ibp_ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2685 #ifdef HAVE_IB_GET_DMA_MR
2687 * This check only makes sense if the kernel supports global
2688 * memory registration. Otherwise, map_on_demand will never == 0
2690 if (!tunables->lnd_map_on_demand) {
2691 reason = "map_on_demand must be enabled";
2695 if (conn->ibc_max_frags <= frag_num) {
2696 reason = "unsupported max frags";
2700 peer_ni->ibp_max_frags = frag_num;
2701 reason = "rdma fragments";
2704 case IBLND_REJECT_MSG_QUEUE_SIZE:
2706 reason = "can't negotiate queue depth";
2709 if (conn->ibc_queue_depth <= queue_dep) {
2710 reason = "unsupported queue depth";
2714 peer_ni->ibp_queue_depth = queue_dep;
2715 reason = "queue depth";
2718 case IBLND_REJECT_CONN_STALE:
2722 case IBLND_REJECT_CONN_RACE:
2723 reason = "conn race";
2726 case IBLND_REJECT_CONN_UNCOMPAT:
2727 reason = "version negotiation";
2730 case IBLND_REJECT_INVALID_SRV_ID:
2731 reason = "invalid service id";
2735 conn->ibc_reconnect = 1;
2736 peer_ni->ibp_reconnecting++;
2737 peer_ni->ibp_version = version;
2738 if (incarnation != 0)
2739 peer_ni->ibp_incarnation = incarnation;
2741 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2743 CNETERR("%s: %s (%s), %x, %x, msg_size: %d, queue_depth: %d/%d, max_frags: %d/%d\n",
2744 libcfs_nid2str(peer_ni->ibp_nid),
2745 reconnect ? "reconnect" : "don't reconnect",
2746 reason, IBLND_MSG_VERSION, version, msg_size,
2747 conn->ibc_queue_depth, queue_dep,
2748 conn->ibc_max_frags, frag_num);
2750 * if conn::ibc_reconnect is TRUE, connd will reconnect to the peer_ni
2751 * while destroying the zombie
2756 kiblnd_rejected (kib_conn_t *conn, int reason, void *priv, int priv_nob)
2758 kib_peer_ni_t *peer_ni = conn->ibc_peer;
2760 LASSERT (!in_interrupt());
2761 LASSERT (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2764 case IB_CM_REJ_STALE_CONN:
2765 kiblnd_check_reconnect(conn, IBLND_MSG_VERSION, 0,
2766 IBLND_REJECT_CONN_STALE, NULL);
2769 case IB_CM_REJ_INVALID_SERVICE_ID:
2770 kiblnd_check_reconnect(conn, IBLND_MSG_VERSION, 0,
2771 IBLND_REJECT_INVALID_SRV_ID, NULL);
2772 CNETERR("%s rejected: no listener at %d\n",
2773 libcfs_nid2str(peer_ni->ibp_nid),
2774 *kiblnd_tunables.kib_service);
2777 case IB_CM_REJ_CONSUMER_DEFINED:
2778 if (priv_nob >= offsetof(kib_rej_t, ibr_padding)) {
2779 kib_rej_t *rej = priv;
2780 kib_connparams_t *cp = NULL;
2782 __u64 incarnation = -1;
2784 /* NB. default incarnation is -1 because:
2785 * a) V1 will ignore dst incarnation in connreq.
2786 * b) V2 will provide incarnation while rejecting me,
2787 * -1 will be overwrote.
2789 * if I try to connect to a V1 peer_ni with V2 protocol,
2790 * it rejected me then upgrade to V2, I have no idea
2791 * about the upgrading and try to reconnect with V1,
2792 * in this case upgraded V2 can find out I'm trying to
2793 * talk to the old guy and reject me(incarnation is -1).
2796 if (rej->ibr_magic == __swab32(IBLND_MSG_MAGIC) ||
2797 rej->ibr_magic == __swab32(LNET_PROTO_MAGIC)) {
2798 __swab32s(&rej->ibr_magic);
2799 __swab16s(&rej->ibr_version);
2803 if (priv_nob >= sizeof(kib_rej_t) &&
2804 rej->ibr_version > IBLND_MSG_VERSION_1) {
2805 /* priv_nob is always 148 in current version
2806 * of OFED, so we still need to check version.
2807 * (define of IB_CM_REJ_PRIVATE_DATA_SIZE) */
2811 __swab64s(&rej->ibr_incarnation);
2812 __swab16s(&cp->ibcp_queue_depth);
2813 __swab16s(&cp->ibcp_max_frags);
2814 __swab32s(&cp->ibcp_max_msg_size);
2817 incarnation = rej->ibr_incarnation;
2820 if (rej->ibr_magic != IBLND_MSG_MAGIC &&
2821 rej->ibr_magic != LNET_PROTO_MAGIC) {
2822 CERROR("%s rejected: consumer defined fatal error\n",
2823 libcfs_nid2str(peer_ni->ibp_nid));
2827 if (rej->ibr_version != IBLND_MSG_VERSION &&
2828 rej->ibr_version != IBLND_MSG_VERSION_1) {
2829 CERROR("%s rejected: o2iblnd version %x error\n",
2830 libcfs_nid2str(peer_ni->ibp_nid),
2835 if (rej->ibr_why == IBLND_REJECT_FATAL &&
2836 rej->ibr_version == IBLND_MSG_VERSION_1) {
2837 CDEBUG(D_NET, "rejected by old version peer_ni %s: %x\n",
2838 libcfs_nid2str(peer_ni->ibp_nid), rej->ibr_version);
2840 if (conn->ibc_version != IBLND_MSG_VERSION_1)
2841 rej->ibr_why = IBLND_REJECT_CONN_UNCOMPAT;
2844 switch (rej->ibr_why) {
2845 case IBLND_REJECT_CONN_RACE:
2846 case IBLND_REJECT_CONN_STALE:
2847 case IBLND_REJECT_CONN_UNCOMPAT:
2848 case IBLND_REJECT_MSG_QUEUE_SIZE:
2849 case IBLND_REJECT_RDMA_FRAGS:
2850 kiblnd_check_reconnect(conn, rej->ibr_version,
2851 incarnation, rej->ibr_why, cp);
2854 case IBLND_REJECT_NO_RESOURCES:
2855 CERROR("%s rejected: o2iblnd no resources\n",
2856 libcfs_nid2str(peer_ni->ibp_nid));
2859 case IBLND_REJECT_FATAL:
2860 CERROR("%s rejected: o2iblnd fatal error\n",
2861 libcfs_nid2str(peer_ni->ibp_nid));
2865 CERROR("%s rejected: o2iblnd reason %d\n",
2866 libcfs_nid2str(peer_ni->ibp_nid),
2874 CNETERR("%s rejected: reason %d, size %d\n",
2875 libcfs_nid2str(peer_ni->ibp_nid), reason, priv_nob);
2879 kiblnd_connreq_done(conn, -ECONNREFUSED);
2883 kiblnd_check_connreply (kib_conn_t *conn, void *priv, int priv_nob)
2885 kib_peer_ni_t *peer_ni = conn->ibc_peer;
2886 struct lnet_ni *ni = peer_ni->ibp_ni;
2887 kib_net_t *net = ni->ni_data;
2888 kib_msg_t *msg = priv;
2889 int ver = conn->ibc_version;
2890 int rc = kiblnd_unpack_msg(msg, priv_nob);
2891 unsigned long flags;
2893 LASSERT (net != NULL);
2896 CERROR("Can't unpack connack from %s: %d\n",
2897 libcfs_nid2str(peer_ni->ibp_nid), rc);
2901 if (msg->ibm_type != IBLND_MSG_CONNACK) {
2902 CERROR("Unexpected message %d from %s\n",
2903 msg->ibm_type, libcfs_nid2str(peer_ni->ibp_nid));
2908 if (ver != msg->ibm_version) {
2909 CERROR("%s replied version %x is different with "
2910 "requested version %x\n",
2911 libcfs_nid2str(peer_ni->ibp_nid), msg->ibm_version, ver);
2916 if (msg->ibm_u.connparams.ibcp_queue_depth >
2917 conn->ibc_queue_depth) {
2918 CERROR("%s has incompatible queue depth %d (<=%d wanted)\n",
2919 libcfs_nid2str(peer_ni->ibp_nid),
2920 msg->ibm_u.connparams.ibcp_queue_depth,
2921 conn->ibc_queue_depth);
2926 if (msg->ibm_u.connparams.ibcp_max_frags >
2927 conn->ibc_max_frags) {
2928 CERROR("%s has incompatible max_frags %d (<=%d wanted)\n",
2929 libcfs_nid2str(peer_ni->ibp_nid),
2930 msg->ibm_u.connparams.ibcp_max_frags,
2931 conn->ibc_max_frags);
2936 if (msg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2937 CERROR("%s max message size %d too big (%d max)\n",
2938 libcfs_nid2str(peer_ni->ibp_nid),
2939 msg->ibm_u.connparams.ibcp_max_msg_size,
2945 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2946 if (msg->ibm_dstnid == ni->ni_nid &&
2947 msg->ibm_dststamp == net->ibn_incarnation)
2951 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2954 CERROR("Bad connection reply from %s, rc = %d, "
2955 "version: %x max_frags: %d\n",
2956 libcfs_nid2str(peer_ni->ibp_nid), rc,
2957 msg->ibm_version, msg->ibm_u.connparams.ibcp_max_frags);
2961 conn->ibc_incarnation = msg->ibm_srcstamp;
2962 conn->ibc_credits = msg->ibm_u.connparams.ibcp_queue_depth;
2963 conn->ibc_reserved_credits = msg->ibm_u.connparams.ibcp_queue_depth;
2964 conn->ibc_queue_depth = msg->ibm_u.connparams.ibcp_queue_depth;
2965 conn->ibc_max_frags = msg->ibm_u.connparams.ibcp_max_frags;
2966 LASSERT(conn->ibc_credits + conn->ibc_reserved_credits +
2967 IBLND_OOB_MSGS(ver) <= IBLND_RX_MSGS(conn));
2969 kiblnd_connreq_done(conn, 0);
2973 /* NB My QP has already established itself, so I handle anything going
2974 * wrong here by setting ibc_comms_error.
2975 * kiblnd_connreq_done(0) moves the conn state to ESTABLISHED, but then
2976 * immediately tears it down. */
2979 conn->ibc_comms_error = rc;
2980 kiblnd_connreq_done(conn, 0);
2984 kiblnd_active_connect (struct rdma_cm_id *cmid)
2986 kib_peer_ni_t *peer_ni = (kib_peer_ni_t *)cmid->context;
2989 struct rdma_conn_param cp;
2992 unsigned long flags;
2995 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2997 incarnation = peer_ni->ibp_incarnation;
2998 version = (peer_ni->ibp_version == 0) ? IBLND_MSG_VERSION :
2999 peer_ni->ibp_version;
3001 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3003 conn = kiblnd_create_conn(peer_ni, cmid, IBLND_CONN_ACTIVE_CONNECT,
3006 kiblnd_peer_connect_failed(peer_ni, 1, -ENOMEM);
3007 kiblnd_peer_decref(peer_ni); /* lose cmid's ref */
3011 /* conn "owns" cmid now, so I return success from here on to ensure the
3012 * CM callback doesn't destroy cmid. conn also takes over cmid's ref
3015 msg = &conn->ibc_connvars->cv_msg;
3017 memset(msg, 0, sizeof(*msg));
3018 kiblnd_init_msg(msg, IBLND_MSG_CONNREQ, sizeof(msg->ibm_u.connparams));
3019 msg->ibm_u.connparams.ibcp_queue_depth = conn->ibc_queue_depth;
3020 msg->ibm_u.connparams.ibcp_max_frags = conn->ibc_max_frags;
3021 msg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
3023 kiblnd_pack_msg(peer_ni->ibp_ni, msg, version,
3024 0, peer_ni->ibp_nid, incarnation);
3026 memset(&cp, 0, sizeof(cp));
3027 cp.private_data = msg;
3028 cp.private_data_len = msg->ibm_nob;
3029 cp.responder_resources = 0; /* No atomic ops or RDMA reads */
3030 cp.initiator_depth = 0;
3031 cp.flow_control = 1;
3032 cp.retry_count = *kiblnd_tunables.kib_retry_count;
3033 cp.rnr_retry_count = *kiblnd_tunables.kib_rnr_retry_count;
3035 LASSERT(cmid->context == (void *)conn);
3036 LASSERT(conn->ibc_cmid == cmid);
3038 rc = rdma_connect(cmid, &cp);
3040 CERROR("Can't connect to %s: %d\n",
3041 libcfs_nid2str(peer_ni->ibp_nid), rc);
3042 kiblnd_connreq_done(conn, rc);
3043 kiblnd_conn_decref(conn);
3050 kiblnd_cm_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
3052 kib_peer_ni_t *peer_ni;
3056 switch (event->event) {
3058 CERROR("Unexpected event: %d, status: %d\n",
3059 event->event, event->status);
3062 case RDMA_CM_EVENT_CONNECT_REQUEST:
3063 /* destroy cmid on failure */
3064 rc = kiblnd_passive_connect(cmid,
3065 (void *)KIBLND_CONN_PARAM(event),
3066 KIBLND_CONN_PARAM_LEN(event));
3067 CDEBUG(D_NET, "connreq: %d\n", rc);
3070 case RDMA_CM_EVENT_ADDR_ERROR:
3071 peer_ni = (kib_peer_ni_t *)cmid->context;
3072 CNETERR("%s: ADDR ERROR %d\n",
3073 libcfs_nid2str(peer_ni->ibp_nid), event->status);
3074 kiblnd_peer_connect_failed(peer_ni, 1, -EHOSTUNREACH);
3075 kiblnd_peer_decref(peer_ni);
3076 return -EHOSTUNREACH; /* rc != 0 destroys cmid */
3078 case RDMA_CM_EVENT_ADDR_RESOLVED:
3079 peer_ni = (kib_peer_ni_t *)cmid->context;
3081 CDEBUG(D_NET,"%s Addr resolved: %d\n",
3082 libcfs_nid2str(peer_ni->ibp_nid), event->status);
3084 if (event->status != 0) {
3085 CNETERR("Can't resolve address for %s: %d\n",
3086 libcfs_nid2str(peer_ni->ibp_nid), event->status);
3089 rc = rdma_resolve_route(
3090 cmid, *kiblnd_tunables.kib_timeout * 1000);
3092 kib_net_t *net = peer_ni->ibp_ni->ni_data;
3093 kib_dev_t *dev = net->ibn_dev;
3095 CDEBUG(D_NET, "%s: connection bound to "\
3097 libcfs_nid2str(peer_ni->ibp_nid),
3099 &dev->ibd_ifip, cmid->device->name);
3104 /* Can't initiate route resolution */
3105 CERROR("Can't resolve route for %s: %d\n",
3106 libcfs_nid2str(peer_ni->ibp_nid), rc);
3108 kiblnd_peer_connect_failed(peer_ni, 1, rc);
3109 kiblnd_peer_decref(peer_ni);
3110 return rc; /* rc != 0 destroys cmid */
3112 case RDMA_CM_EVENT_ROUTE_ERROR:
3113 peer_ni = (kib_peer_ni_t *)cmid->context;
3114 CNETERR("%s: ROUTE ERROR %d\n",
3115 libcfs_nid2str(peer_ni->ibp_nid), event->status);
3116 kiblnd_peer_connect_failed(peer_ni, 1, -EHOSTUNREACH);
3117 kiblnd_peer_decref(peer_ni);
3118 return -EHOSTUNREACH; /* rc != 0 destroys cmid */
3120 case RDMA_CM_EVENT_ROUTE_RESOLVED:
3121 peer_ni = (kib_peer_ni_t *)cmid->context;
3122 CDEBUG(D_NET,"%s Route resolved: %d\n",
3123 libcfs_nid2str(peer_ni->ibp_nid), event->status);
3125 if (event->status == 0)
3126 return kiblnd_active_connect(cmid);
3128 CNETERR("Can't resolve route for %s: %d\n",
3129 libcfs_nid2str(peer_ni->ibp_nid), event->status);
3130 kiblnd_peer_connect_failed(peer_ni, 1, event->status);
3131 kiblnd_peer_decref(peer_ni);
3132 return event->status; /* rc != 0 destroys cmid */
3134 case RDMA_CM_EVENT_UNREACHABLE:
3135 conn = (kib_conn_t *)cmid->context;
3136 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
3137 conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
3138 CNETERR("%s: UNREACHABLE %d\n",
3139 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
3140 kiblnd_connreq_done(conn, -ENETDOWN);
3141 kiblnd_conn_decref(conn);
3144 case RDMA_CM_EVENT_CONNECT_ERROR:
3145 conn = (kib_conn_t *)cmid->context;
3146 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
3147 conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
3148 CNETERR("%s: CONNECT ERROR %d\n",
3149 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
3150 kiblnd_connreq_done(conn, -ENOTCONN);
3151 kiblnd_conn_decref(conn);
3154 case RDMA_CM_EVENT_REJECTED:
3155 conn = (kib_conn_t *)cmid->context;
3156 switch (conn->ibc_state) {
3160 case IBLND_CONN_PASSIVE_WAIT:
3161 CERROR ("%s: REJECTED %d\n",
3162 libcfs_nid2str(conn->ibc_peer->ibp_nid),
3164 kiblnd_connreq_done(conn, -ECONNRESET);
3167 case IBLND_CONN_ACTIVE_CONNECT:
3168 kiblnd_rejected(conn, event->status,
3169 (void *)KIBLND_CONN_PARAM(event),
3170 KIBLND_CONN_PARAM_LEN(event));
3173 kiblnd_conn_decref(conn);
3176 case RDMA_CM_EVENT_ESTABLISHED:
3177 conn = (kib_conn_t *)cmid->context;
3178 switch (conn->ibc_state) {
3182 case IBLND_CONN_PASSIVE_WAIT:
3183 CDEBUG(D_NET, "ESTABLISHED (passive): %s\n",
3184 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3185 kiblnd_connreq_done(conn, 0);
3188 case IBLND_CONN_ACTIVE_CONNECT:
3189 CDEBUG(D_NET, "ESTABLISHED(active): %s\n",
3190 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3191 kiblnd_check_connreply(conn,
3192 (void *)KIBLND_CONN_PARAM(event),
3193 KIBLND_CONN_PARAM_LEN(event));
3196 /* net keeps its ref on conn! */
3199 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
3200 CDEBUG(D_NET, "Ignore TIMEWAIT_EXIT event\n");
3203 case RDMA_CM_EVENT_DISCONNECTED:
3204 conn = (kib_conn_t *)cmid->context;
3205 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
3206 CERROR("%s DISCONNECTED\n",
3207 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3208 kiblnd_connreq_done(conn, -ECONNRESET);
3210 kiblnd_close_conn(conn, 0);
3212 kiblnd_conn_decref(conn);
3213 cmid->context = NULL;
3216 case RDMA_CM_EVENT_DEVICE_REMOVAL:
3217 LCONSOLE_ERROR_MSG(0x131,
3218 "Received notification of device removal\n"
3219 "Please shutdown LNET to allow this to proceed\n");
3220 /* Can't remove network from underneath LNET for now, so I have
3224 case RDMA_CM_EVENT_ADDR_CHANGE:
3225 LCONSOLE_INFO("Physical link changed (eg hca/port)\n");
3231 kiblnd_check_txs_locked(kib_conn_t *conn, struct list_head *txs)
3234 struct list_head *ttmp;
3236 list_for_each(ttmp, txs) {
3237 tx = list_entry(ttmp, kib_tx_t, tx_list);
3239 if (txs != &conn->ibc_active_txs) {
3240 LASSERT(tx->tx_queued);
3242 LASSERT(!tx->tx_queued);
3243 LASSERT(tx->tx_waiting || tx->tx_sending != 0);
3246 if (cfs_time_aftereq(jiffies, tx->tx_deadline)) {
3247 CERROR("Timed out tx: %s, %lu seconds\n",
3248 kiblnd_queue2str(conn, txs),
3249 cfs_duration_sec(jiffies - tx->tx_deadline));
3258 kiblnd_conn_timed_out_locked(kib_conn_t *conn)
3260 return kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue) ||
3261 kiblnd_check_txs_locked(conn, &conn->ibc_tx_noops) ||
3262 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_rsrvd) ||
3263 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_nocred) ||
3264 kiblnd_check_txs_locked(conn, &conn->ibc_active_txs);
3268 kiblnd_check_conns (int idx)
3270 struct list_head closes = LIST_HEAD_INIT(closes);
3271 struct list_head checksends = LIST_HEAD_INIT(checksends);
3272 struct list_head timedout_txs = LIST_HEAD_INIT(timedout_txs);
3273 struct list_head *peers = &kiblnd_data.kib_peers[idx];
3274 struct list_head *ptmp;
3275 kib_peer_ni_t *peer_ni;
3277 kib_tx_t *tx, *tx_tmp;
3278 struct list_head *ctmp;
3279 unsigned long flags;
3281 /* NB. We expect to have a look at all the peers and not find any
3282 * RDMAs to time out, so we just use a shared lock while we
3284 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3286 list_for_each(ptmp, peers) {
3287 peer_ni = list_entry(ptmp, kib_peer_ni_t, ibp_list);
3289 /* Check tx_deadline */
3290 list_for_each_entry_safe(tx, tx_tmp, &peer_ni->ibp_tx_queue, tx_list) {
3291 if (cfs_time_aftereq(jiffies, tx->tx_deadline)) {
3292 CWARN("Timed out tx for %s: %lu seconds\n",
3293 libcfs_nid2str(peer_ni->ibp_nid),
3294 cfs_duration_sec(jiffies - tx->tx_deadline));
3295 list_move(&tx->tx_list, &timedout_txs);
3299 list_for_each(ctmp, &peer_ni->ibp_conns) {
3303 conn = list_entry(ctmp, kib_conn_t, ibc_list);
3305 LASSERT(conn->ibc_state == IBLND_CONN_ESTABLISHED);
3307 spin_lock(&conn->ibc_lock);
3309 sendnoop = kiblnd_need_noop(conn);
3310 timedout = kiblnd_conn_timed_out_locked(conn);
3311 if (!sendnoop && !timedout) {
3312 spin_unlock(&conn->ibc_lock);
3317 CERROR("Timed out RDMA with %s (%lu): "
3318 "c: %u, oc: %u, rc: %u\n",
3319 libcfs_nid2str(peer_ni->ibp_nid),
3320 cfs_duration_sec(cfs_time_current() -
3321 peer_ni->ibp_last_alive),
3323 conn->ibc_outstanding_credits,
3324 conn->ibc_reserved_credits);
3325 list_add(&conn->ibc_connd_list, &closes);
3327 list_add(&conn->ibc_connd_list, &checksends);
3329 /* +ref for 'closes' or 'checksends' */
3330 kiblnd_conn_addref(conn);
3332 spin_unlock(&conn->ibc_lock);
3336 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3338 if (!list_empty(&timedout_txs))
3339 kiblnd_txlist_done(&timedout_txs, -ETIMEDOUT);
3341 /* Handle timeout by closing the whole
3342 * connection. We can only be sure RDMA activity
3343 * has ceased once the QP has been modified. */
3344 while (!list_empty(&closes)) {
3345 conn = list_entry(closes.next,
3346 kib_conn_t, ibc_connd_list);
3347 list_del(&conn->ibc_connd_list);
3348 kiblnd_close_conn(conn, -ETIMEDOUT);
3349 kiblnd_conn_decref(conn);
3352 /* In case we have enough credits to return via a
3353 * NOOP, but there were no non-blocking tx descs
3354 * free to do it last time... */
3355 while (!list_empty(&checksends)) {
3356 conn = list_entry(checksends.next,
3357 kib_conn_t, ibc_connd_list);
3358 list_del(&conn->ibc_connd_list);
3360 spin_lock(&conn->ibc_lock);
3361 kiblnd_check_sends_locked(conn);
3362 spin_unlock(&conn->ibc_lock);
3364 kiblnd_conn_decref(conn);
3369 kiblnd_disconnect_conn (kib_conn_t *conn)
3371 LASSERT (!in_interrupt());
3372 LASSERT (current == kiblnd_data.kib_connd);
3373 LASSERT (conn->ibc_state == IBLND_CONN_CLOSING);
3375 rdma_disconnect(conn->ibc_cmid);
3376 kiblnd_finalise_conn(conn);
3378 kiblnd_peer_notify(conn->ibc_peer);
3382 * High-water for reconnection to the same peer_ni, reconnection attempt should
3383 * be delayed after trying more than KIB_RECONN_HIGH_RACE.
3385 #define KIB_RECONN_HIGH_RACE 10
3387 * Allow connd to take a break and handle other things after consecutive
3388 * reconnection attemps.
3390 #define KIB_RECONN_BREAK 100
3393 kiblnd_connd (void *arg)
3395 spinlock_t *lock= &kiblnd_data.kib_connd_lock;
3397 unsigned long flags;
3403 unsigned long deadline = jiffies;
3405 cfs_block_allsigs();
3407 init_waitqueue_entry(&wait, current);
3408 kiblnd_data.kib_connd = current;
3410 spin_lock_irqsave(lock, flags);
3412 while (!kiblnd_data.kib_shutdown) {
3417 if (!list_empty(&kiblnd_data.kib_connd_zombies)) {
3418 kib_peer_ni_t *peer_ni = NULL;
3420 conn = list_entry(kiblnd_data.kib_connd_zombies.next,
3421 kib_conn_t, ibc_list);
3422 list_del(&conn->ibc_list);
3423 if (conn->ibc_reconnect) {
3424 peer_ni = conn->ibc_peer;
3425 kiblnd_peer_addref(peer_ni);
3428 spin_unlock_irqrestore(lock, flags);
3431 kiblnd_destroy_conn(conn, !peer_ni);
3433 spin_lock_irqsave(lock, flags);
3437 conn->ibc_peer = peer_ni;
3438 if (peer_ni->ibp_reconnected < KIB_RECONN_HIGH_RACE)
3439 list_add_tail(&conn->ibc_list,
3440 &kiblnd_data.kib_reconn_list);
3442 list_add_tail(&conn->ibc_list,
3443 &kiblnd_data.kib_reconn_wait);
3446 if (!list_empty(&kiblnd_data.kib_connd_conns)) {
3447 conn = list_entry(kiblnd_data.kib_connd_conns.next,
3448 kib_conn_t, ibc_list);
3449 list_del(&conn->ibc_list);
3451 spin_unlock_irqrestore(lock, flags);
3454 kiblnd_disconnect_conn(conn);
3455 kiblnd_conn_decref(conn);
3457 spin_lock_irqsave(lock, flags);
3460 while (reconn < KIB_RECONN_BREAK) {
3461 if (kiblnd_data.kib_reconn_sec !=
3462 ktime_get_real_seconds()) {
3463 kiblnd_data.kib_reconn_sec = ktime_get_real_seconds();
3464 list_splice_init(&kiblnd_data.kib_reconn_wait,
3465 &kiblnd_data.kib_reconn_list);
3468 if (list_empty(&kiblnd_data.kib_reconn_list))
3471 conn = list_entry(kiblnd_data.kib_reconn_list.next,
3472 kib_conn_t, ibc_list);
3473 list_del(&conn->ibc_list);
3475 spin_unlock_irqrestore(lock, flags);
3478 reconn += kiblnd_reconnect_peer(conn->ibc_peer);
3479 kiblnd_peer_decref(conn->ibc_peer);
3480 LIBCFS_FREE(conn, sizeof(*conn));
3482 spin_lock_irqsave(lock, flags);
3485 /* careful with the jiffy wrap... */
3486 timeout = (int)(deadline - jiffies);
3490 int chunk = kiblnd_data.kib_peer_hash_size;
3492 spin_unlock_irqrestore(lock, flags);
3495 /* Time to check for RDMA timeouts on a few more
3496 * peers: I do checks every 'p' seconds on a
3497 * proportion of the peer_ni table and I need to check
3498 * every connection 'n' times within a timeout
3499 * interval, to ensure I detect a timeout on any
3500 * connection within (n+1)/n times the timeout
3503 if (*kiblnd_tunables.kib_timeout > n * p)
3504 chunk = (chunk * n * p) /
3505 *kiblnd_tunables.kib_timeout;
3509 for (i = 0; i < chunk; i++) {
3510 kiblnd_check_conns(peer_index);
3511 peer_index = (peer_index + 1) %
3512 kiblnd_data.kib_peer_hash_size;
3515 deadline += msecs_to_jiffies(p * MSEC_PER_SEC);
3516 spin_lock_irqsave(lock, flags);
3522 /* Nothing to do for 'timeout' */
3523 set_current_state(TASK_INTERRUPTIBLE);
3524 add_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3525 spin_unlock_irqrestore(lock, flags);
3527 schedule_timeout(timeout);
3529 set_current_state(TASK_RUNNING);
3530 remove_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3531 spin_lock_irqsave(lock, flags);
3534 spin_unlock_irqrestore(lock, flags);
3536 kiblnd_thread_fini();
3541 kiblnd_qp_event(struct ib_event *event, void *arg)
3543 kib_conn_t *conn = arg;
3545 switch (event->event) {
3546 case IB_EVENT_COMM_EST:
3547 CDEBUG(D_NET, "%s established\n",
3548 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3549 /* We received a packet but connection isn't established
3550 * probably handshake packet was lost, so free to
3551 * force make connection established */
3552 rdma_notify(conn->ibc_cmid, IB_EVENT_COMM_EST);
3556 CERROR("%s: Async QP event type %d\n",
3557 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3563 kiblnd_complete (struct ib_wc *wc)
3565 switch (kiblnd_wreqid2type(wc->wr_id)) {
3570 if (wc->status != IB_WC_SUCCESS &&
3571 wc->status != IB_WC_WR_FLUSH_ERR)
3572 CNETERR("FastReg failed: %d\n", wc->status);
3575 case IBLND_WID_RDMA:
3576 /* We only get RDMA completion notification if it fails. All
3577 * subsequent work items, including the final SEND will fail
3578 * too. However we can't print out any more info about the
3579 * failing RDMA because 'tx' might be back on the idle list or
3580 * even reused already if we didn't manage to post all our work
3582 CNETERR("RDMA (tx: %p) failed: %d\n",
3583 kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3587 kiblnd_tx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3591 kiblnd_rx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status,
3598 kiblnd_cq_completion(struct ib_cq *cq, void *arg)
3600 /* NB I'm not allowed to schedule this conn once its refcount has
3601 * reached 0. Since fundamentally I'm racing with scheduler threads
3602 * consuming my CQ I could be called after all completions have
3603 * occurred. But in this case, ibc_nrx == 0 && ibc_nsends_posted == 0
3604 * and this CQ is about to be destroyed so I NOOP. */
3605 kib_conn_t *conn = (kib_conn_t *)arg;
3606 struct kib_sched_info *sched = conn->ibc_sched;
3607 unsigned long flags;
3609 LASSERT(cq == conn->ibc_cq);
3611 spin_lock_irqsave(&sched->ibs_lock, flags);
3613 conn->ibc_ready = 1;
3615 if (!conn->ibc_scheduled &&
3616 (conn->ibc_nrx > 0 ||
3617 conn->ibc_nsends_posted > 0)) {
3618 kiblnd_conn_addref(conn); /* +1 ref for sched_conns */
3619 conn->ibc_scheduled = 1;
3620 list_add_tail(&conn->ibc_sched_list, &sched->ibs_conns);
3622 if (waitqueue_active(&sched->ibs_waitq))
3623 wake_up(&sched->ibs_waitq);
3626 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3630 kiblnd_cq_event(struct ib_event *event, void *arg)
3632 kib_conn_t *conn = arg;
3634 CERROR("%s: async CQ event type %d\n",
3635 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3639 kiblnd_scheduler(void *arg)
3641 long id = (long)arg;
3642 struct kib_sched_info *sched;
3645 unsigned long flags;
3651 cfs_block_allsigs();
3653 init_waitqueue_entry(&wait, current);
3655 sched = kiblnd_data.kib_scheds[KIB_THREAD_CPT(id)];
3657 rc = cfs_cpt_bind(lnet_cpt_table(), sched->ibs_cpt);
3659 CWARN("Unable to bind on CPU partition %d, please verify "
3660 "whether all CPUs are healthy and reload modules if "
3661 "necessary, otherwise your system might under risk of "
3662 "low performance\n", sched->ibs_cpt);
3665 spin_lock_irqsave(&sched->ibs_lock, flags);
3667 while (!kiblnd_data.kib_shutdown) {
3668 if (busy_loops++ >= IBLND_RESCHED) {
3669 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3674 spin_lock_irqsave(&sched->ibs_lock, flags);
3679 if (!list_empty(&sched->ibs_conns)) {
3680 conn = list_entry(sched->ibs_conns.next,
3681 kib_conn_t, ibc_sched_list);
3682 /* take over kib_sched_conns' ref on conn... */
3683 LASSERT(conn->ibc_scheduled);
3684 list_del(&conn->ibc_sched_list);
3685 conn->ibc_ready = 0;
3687 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3689 wc.wr_id = IBLND_WID_INVAL;
3691 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3693 rc = ib_req_notify_cq(conn->ibc_cq,
3696 CWARN("%s: ib_req_notify_cq failed: %d, "
3697 "closing connection\n",
3698 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
3699 kiblnd_close_conn(conn, -EIO);
3700 kiblnd_conn_decref(conn);
3701 spin_lock_irqsave(&sched->ibs_lock,
3706 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3709 if (unlikely(rc > 0 && wc.wr_id == IBLND_WID_INVAL)) {
3711 "ib_poll_cq (rc: %d) returned invalid "
3712 "wr_id, opcode %d, status: %d, "
3713 "vendor_err: %d, conn: %s status: %d\n"
3714 "please upgrade firmware and OFED or "
3715 "contact vendor.\n", rc,
3716 wc.opcode, wc.status, wc.vendor_err,
3717 libcfs_nid2str(conn->ibc_peer->ibp_nid),
3723 CWARN("%s: ib_poll_cq failed: %d, "
3724 "closing connection\n",
3725 libcfs_nid2str(conn->ibc_peer->ibp_nid),
3727 kiblnd_close_conn(conn, -EIO);
3728 kiblnd_conn_decref(conn);
3729 spin_lock_irqsave(&sched->ibs_lock, flags);
3733 spin_lock_irqsave(&sched->ibs_lock, flags);
3735 if (rc != 0 || conn->ibc_ready) {
3736 /* There may be another completion waiting; get
3737 * another scheduler to check while I handle
3739 /* +1 ref for sched_conns */
3740 kiblnd_conn_addref(conn);
3741 list_add_tail(&conn->ibc_sched_list,
3743 if (waitqueue_active(&sched->ibs_waitq))
3744 wake_up(&sched->ibs_waitq);
3746 conn->ibc_scheduled = 0;
3750 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3751 kiblnd_complete(&wc);
3753 spin_lock_irqsave(&sched->ibs_lock, flags);
3756 kiblnd_conn_decref(conn); /* ...drop my ref from above */
3763 set_current_state(TASK_INTERRUPTIBLE);
3764 add_wait_queue_exclusive(&sched->ibs_waitq, &wait);
3765 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3770 remove_wait_queue(&sched->ibs_waitq, &wait);
3771 set_current_state(TASK_RUNNING);
3772 spin_lock_irqsave(&sched->ibs_lock, flags);
3775 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3777 kiblnd_thread_fini();
3782 kiblnd_failover_thread(void *arg)
3784 rwlock_t *glock = &kiblnd_data.kib_global_lock;
3787 unsigned long flags;
3790 LASSERT(*kiblnd_tunables.kib_dev_failover != 0);
3792 cfs_block_allsigs();
3794 init_waitqueue_entry(&wait, current);
3795 write_lock_irqsave(glock, flags);
3797 while (!kiblnd_data.kib_shutdown) {
3798 int do_failover = 0;
3801 list_for_each_entry(dev, &kiblnd_data.kib_failed_devs,
3803 if (cfs_time_before(cfs_time_current(),
3804 dev->ibd_next_failover))
3811 list_del_init(&dev->ibd_fail_list);
3812 dev->ibd_failover = 1;
3813 write_unlock_irqrestore(glock, flags);
3815 rc = kiblnd_dev_failover(dev);
3817 write_lock_irqsave(glock, flags);
3819 LASSERT (dev->ibd_failover);
3820 dev->ibd_failover = 0;
3821 if (rc >= 0) { /* Device is OK or failover succeed */
3822 dev->ibd_next_failover = cfs_time_shift(3);
3826 /* failed to failover, retry later */
3827 dev->ibd_next_failover =
3828 cfs_time_shift(min(dev->ibd_failed_failover, 10));
3829 if (kiblnd_dev_can_failover(dev)) {
3830 list_add_tail(&dev->ibd_fail_list,
3831 &kiblnd_data.kib_failed_devs);
3837 /* long sleep if no more pending failover */
3838 long_sleep = list_empty(&kiblnd_data.kib_failed_devs);
3840 set_current_state(TASK_INTERRUPTIBLE);
3841 add_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3842 write_unlock_irqrestore(glock, flags);
3844 rc = schedule_timeout(long_sleep ? cfs_time_seconds(10) :
3845 cfs_time_seconds(1));
3846 set_current_state(TASK_RUNNING);
3847 remove_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3848 write_lock_irqsave(glock, flags);
3850 if (!long_sleep || rc != 0)
3853 /* have a long sleep, routine check all active devices,
3854 * we need checking like this because if there is not active
3855 * connection on the dev and no SEND from local, we may listen
3856 * on wrong HCA for ever while there is a bonding failover */
3857 list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3858 if (kiblnd_dev_can_failover(dev)) {
3859 list_add_tail(&dev->ibd_fail_list,
3860 &kiblnd_data.kib_failed_devs);
3865 write_unlock_irqrestore(glock, flags);
3867 kiblnd_thread_fini();