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.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2012, 2015, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lnet/klnds/o2iblnd/o2iblnd_cb.c
38 * Author: Eric Barton <eric@bartonsoftware.com>
43 static void kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx);
46 kiblnd_tx_done(lnet_ni_t *ni, kib_tx_t *tx)
48 lnet_msg_t *lntmsg[2];
49 kib_net_t *net = ni->ni_data;
54 LASSERT(!in_interrupt());
55 LASSERT(!tx->tx_queued); /* mustn't be queued for sending */
56 LASSERT(!tx->tx_sending); /* mustn't be awaiting sent callback */
57 LASSERT(!tx->tx_waiting); /* mustn't be awaiting peer response */
60 kiblnd_unmap_tx(ni, tx);
62 /* tx may have up to 2 lnet msgs to finalise */
63 lntmsg[0] = tx->tx_lntmsg[0]; tx->tx_lntmsg[0] = NULL;
64 lntmsg[1] = tx->tx_lntmsg[1]; tx->tx_lntmsg[1] = NULL;
68 LASSERT(ni == tx->tx_conn->ibc_peer->ibp_ni);
70 kiblnd_conn_decref(tx->tx_conn);
77 kiblnd_pool_free_node(&tx->tx_pool->tpo_pool, &tx->tx_list);
79 /* delay finalize until my descs have been freed */
80 for (i = 0; i < 2; i++) {
84 lnet_finalize(ni, lntmsg[i], rc);
89 kiblnd_txlist_done(lnet_ni_t *ni, struct list_head *txlist, int status)
93 while (!list_empty(txlist)) {
94 tx = list_entry(txlist->next, kib_tx_t, tx_list);
96 list_del(&tx->tx_list);
99 tx->tx_status = status;
100 kiblnd_tx_done(ni, tx);
105 kiblnd_get_idle_tx(lnet_ni_t *ni, lnet_nid_t target)
107 kib_net_t *net = (kib_net_t *)ni->ni_data;
108 struct list_head *node;
110 kib_tx_poolset_t *tps;
112 tps = net->ibn_tx_ps[lnet_cpt_of_nid(target)];
113 node = kiblnd_pool_alloc_node(&tps->tps_poolset);
116 tx = list_entry(node, kib_tx_t, tx_list);
118 LASSERT(!tx->tx_nwrq);
119 LASSERT(!tx->tx_queued);
120 LASSERT(!tx->tx_sending);
121 LASSERT(!tx->tx_waiting);
122 LASSERT(!tx->tx_status);
123 LASSERT(!tx->tx_conn);
124 LASSERT(!tx->tx_lntmsg[0]);
125 LASSERT(!tx->tx_lntmsg[1]);
126 LASSERT(!tx->tx_nfrags);
132 kiblnd_drop_rx(kib_rx_t *rx)
134 kib_conn_t *conn = rx->rx_conn;
135 struct kib_sched_info *sched = conn->ibc_sched;
138 spin_lock_irqsave(&sched->ibs_lock, flags);
139 LASSERT(conn->ibc_nrx > 0);
141 spin_unlock_irqrestore(&sched->ibs_lock, flags);
143 kiblnd_conn_decref(conn);
147 kiblnd_post_rx(kib_rx_t *rx, int credit)
149 kib_conn_t *conn = rx->rx_conn;
150 kib_net_t *net = conn->ibc_peer->ibp_ni->ni_data;
151 struct ib_recv_wr *bad_wrq = NULL;
156 LASSERT(!in_interrupt());
157 LASSERT(credit == IBLND_POSTRX_NO_CREDIT ||
158 credit == IBLND_POSTRX_PEER_CREDIT ||
159 credit == IBLND_POSTRX_RSRVD_CREDIT);
161 mr = kiblnd_find_dma_mr(conn->ibc_hdev, rx->rx_msgaddr, IBLND_MSG_SIZE);
164 rx->rx_sge.lkey = mr->lkey;
165 rx->rx_sge.addr = rx->rx_msgaddr;
166 rx->rx_sge.length = IBLND_MSG_SIZE;
168 rx->rx_wrq.next = NULL;
169 rx->rx_wrq.sg_list = &rx->rx_sge;
170 rx->rx_wrq.num_sge = 1;
171 rx->rx_wrq.wr_id = kiblnd_ptr2wreqid(rx, IBLND_WID_RX);
173 LASSERT(conn->ibc_state >= IBLND_CONN_INIT);
174 LASSERT(rx->rx_nob >= 0); /* not posted */
176 if (conn->ibc_state > IBLND_CONN_ESTABLISHED) {
177 kiblnd_drop_rx(rx); /* No more posts for this rx */
181 rx->rx_nob = -1; /* flag posted */
183 /* NB: need an extra reference after ib_post_recv because we don't
184 * own this rx (and rx::rx_conn) anymore, LU-5678.
186 kiblnd_conn_addref(conn);
187 rc = ib_post_recv(conn->ibc_cmid->qp, &rx->rx_wrq, &bad_wrq);
189 CERROR("Can't post rx for %s: %d, bad_wrq: %p\n",
190 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc, bad_wrq);
194 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) /* Initial post */
198 kiblnd_close_conn(conn, rc);
199 kiblnd_drop_rx(rx); /* No more posts for this rx */
203 if (credit == IBLND_POSTRX_NO_CREDIT)
206 spin_lock(&conn->ibc_lock);
207 if (credit == IBLND_POSTRX_PEER_CREDIT)
208 conn->ibc_outstanding_credits++;
210 conn->ibc_reserved_credits++;
211 spin_unlock(&conn->ibc_lock);
213 kiblnd_check_sends(conn);
215 kiblnd_conn_decref(conn);
220 kiblnd_find_waiting_tx_locked(kib_conn_t *conn, int txtype, __u64 cookie)
222 struct list_head *tmp;
224 list_for_each(tmp, &conn->ibc_active_txs) {
225 kib_tx_t *tx = list_entry(tmp, kib_tx_t, tx_list);
227 LASSERT(!tx->tx_queued);
228 LASSERT(tx->tx_sending || tx->tx_waiting);
230 if (tx->tx_cookie != cookie)
233 if (tx->tx_waiting &&
234 tx->tx_msg->ibm_type == txtype)
237 CWARN("Bad completion: %swaiting, type %x (wanted %x)\n",
238 tx->tx_waiting ? "" : "NOT ",
239 tx->tx_msg->ibm_type, txtype);
245 kiblnd_handle_completion(kib_conn_t *conn, int txtype, int status, __u64 cookie)
248 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
251 spin_lock(&conn->ibc_lock);
253 tx = kiblnd_find_waiting_tx_locked(conn, txtype, cookie);
255 spin_unlock(&conn->ibc_lock);
257 CWARN("Unmatched completion type %x cookie %#llx from %s\n",
258 txtype, cookie, libcfs_nid2str(conn->ibc_peer->ibp_nid));
259 kiblnd_close_conn(conn, -EPROTO);
263 if (!tx->tx_status) { /* success so far */
264 if (status < 0) /* failed? */
265 tx->tx_status = status;
266 else if (txtype == IBLND_MSG_GET_REQ)
267 lnet_set_reply_msg_len(ni, tx->tx_lntmsg[1], status);
272 idle = !tx->tx_queued && !tx->tx_sending;
274 list_del(&tx->tx_list);
276 spin_unlock(&conn->ibc_lock);
279 kiblnd_tx_done(ni, tx);
283 kiblnd_send_completion(kib_conn_t *conn, int type, int status, __u64 cookie)
285 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
286 kib_tx_t *tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
289 CERROR("Can't get tx for completion %x for %s\n",
290 type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
294 tx->tx_msg->ibm_u.completion.ibcm_status = status;
295 tx->tx_msg->ibm_u.completion.ibcm_cookie = cookie;
296 kiblnd_init_tx_msg(ni, tx, type, sizeof(kib_completion_msg_t));
298 kiblnd_queue_tx(tx, conn);
302 kiblnd_handle_rx(kib_rx_t *rx)
304 kib_msg_t *msg = rx->rx_msg;
305 kib_conn_t *conn = rx->rx_conn;
306 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
307 int credits = msg->ibm_credits;
313 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
315 CDEBUG(D_NET, "Received %x[%d] from %s\n",
316 msg->ibm_type, credits,
317 libcfs_nid2str(conn->ibc_peer->ibp_nid));
320 /* Have I received credits that will let me send? */
321 spin_lock(&conn->ibc_lock);
323 if (conn->ibc_credits + credits >
324 IBLND_MSG_QUEUE_SIZE(conn->ibc_version)) {
325 rc2 = conn->ibc_credits;
326 spin_unlock(&conn->ibc_lock);
328 CERROR("Bad credits from %s: %d + %d > %d\n",
329 libcfs_nid2str(conn->ibc_peer->ibp_nid),
331 IBLND_MSG_QUEUE_SIZE(conn->ibc_version));
333 kiblnd_close_conn(conn, -EPROTO);
334 kiblnd_post_rx(rx, IBLND_POSTRX_NO_CREDIT);
338 conn->ibc_credits += credits;
340 /* This ensures the credit taken by NOOP can be returned */
341 if (msg->ibm_type == IBLND_MSG_NOOP &&
342 !IBLND_OOB_CAPABLE(conn->ibc_version)) /* v1 only */
343 conn->ibc_outstanding_credits++;
345 spin_unlock(&conn->ibc_lock);
346 kiblnd_check_sends(conn);
349 switch (msg->ibm_type) {
351 CERROR("Bad IBLND message type %x from %s\n",
352 msg->ibm_type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
353 post_credit = IBLND_POSTRX_NO_CREDIT;
358 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
359 post_credit = IBLND_POSTRX_NO_CREDIT;
363 if (credits) /* credit already posted */
364 post_credit = IBLND_POSTRX_NO_CREDIT;
365 else /* a keepalive NOOP */
366 post_credit = IBLND_POSTRX_PEER_CREDIT;
369 case IBLND_MSG_IMMEDIATE:
370 post_credit = IBLND_POSTRX_DONT_POST;
371 rc = lnet_parse(ni, &msg->ibm_u.immediate.ibim_hdr,
372 msg->ibm_srcnid, rx, 0);
373 if (rc < 0) /* repost on error */
374 post_credit = IBLND_POSTRX_PEER_CREDIT;
377 case IBLND_MSG_PUT_REQ:
378 post_credit = IBLND_POSTRX_DONT_POST;
379 rc = lnet_parse(ni, &msg->ibm_u.putreq.ibprm_hdr,
380 msg->ibm_srcnid, rx, 1);
381 if (rc < 0) /* repost on error */
382 post_credit = IBLND_POSTRX_PEER_CREDIT;
385 case IBLND_MSG_PUT_NAK:
386 CWARN("PUT_NACK from %s\n",
387 libcfs_nid2str(conn->ibc_peer->ibp_nid));
388 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
389 kiblnd_handle_completion(conn, IBLND_MSG_PUT_REQ,
390 msg->ibm_u.completion.ibcm_status,
391 msg->ibm_u.completion.ibcm_cookie);
394 case IBLND_MSG_PUT_ACK:
395 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
397 spin_lock(&conn->ibc_lock);
398 tx = kiblnd_find_waiting_tx_locked(conn, IBLND_MSG_PUT_REQ,
399 msg->ibm_u.putack.ibpam_src_cookie);
401 list_del(&tx->tx_list);
402 spin_unlock(&conn->ibc_lock);
405 CERROR("Unmatched PUT_ACK from %s\n",
406 libcfs_nid2str(conn->ibc_peer->ibp_nid));
411 LASSERT(tx->tx_waiting);
413 * CAVEAT EMPTOR: I could be racing with tx_complete, but...
414 * (a) I can overwrite tx_msg since my peer has received it!
415 * (b) tx_waiting set tells tx_complete() it's not done.
417 tx->tx_nwrq = 0; /* overwrite PUT_REQ */
419 rc2 = kiblnd_init_rdma(conn, tx, IBLND_MSG_PUT_DONE,
420 kiblnd_rd_size(&msg->ibm_u.putack.ibpam_rd),
421 &msg->ibm_u.putack.ibpam_rd,
422 msg->ibm_u.putack.ibpam_dst_cookie);
424 CERROR("Can't setup rdma for PUT to %s: %d\n",
425 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc2);
427 spin_lock(&conn->ibc_lock);
428 tx->tx_waiting = 0; /* clear waiting and queue atomically */
429 kiblnd_queue_tx_locked(tx, conn);
430 spin_unlock(&conn->ibc_lock);
433 case IBLND_MSG_PUT_DONE:
434 post_credit = IBLND_POSTRX_PEER_CREDIT;
435 kiblnd_handle_completion(conn, IBLND_MSG_PUT_ACK,
436 msg->ibm_u.completion.ibcm_status,
437 msg->ibm_u.completion.ibcm_cookie);
440 case IBLND_MSG_GET_REQ:
441 post_credit = IBLND_POSTRX_DONT_POST;
442 rc = lnet_parse(ni, &msg->ibm_u.get.ibgm_hdr,
443 msg->ibm_srcnid, rx, 1);
444 if (rc < 0) /* repost on error */
445 post_credit = IBLND_POSTRX_PEER_CREDIT;
448 case IBLND_MSG_GET_DONE:
449 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
450 kiblnd_handle_completion(conn, IBLND_MSG_GET_REQ,
451 msg->ibm_u.completion.ibcm_status,
452 msg->ibm_u.completion.ibcm_cookie);
456 if (rc < 0) /* protocol error */
457 kiblnd_close_conn(conn, rc);
459 if (post_credit != IBLND_POSTRX_DONT_POST)
460 kiblnd_post_rx(rx, post_credit);
464 kiblnd_rx_complete(kib_rx_t *rx, int status, int nob)
466 kib_msg_t *msg = rx->rx_msg;
467 kib_conn_t *conn = rx->rx_conn;
468 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
469 kib_net_t *net = ni->ni_data;
474 LASSERT(rx->rx_nob < 0); /* was posted */
475 rx->rx_nob = 0; /* isn't now */
477 if (conn->ibc_state > IBLND_CONN_ESTABLISHED)
480 if (status != IB_WC_SUCCESS) {
481 CNETERR("Rx from %s failed: %d\n",
482 libcfs_nid2str(conn->ibc_peer->ibp_nid), status);
489 rc = kiblnd_unpack_msg(msg, rx->rx_nob);
491 CERROR("Error %d unpacking rx from %s\n",
492 rc, libcfs_nid2str(conn->ibc_peer->ibp_nid));
496 if (msg->ibm_srcnid != conn->ibc_peer->ibp_nid ||
497 msg->ibm_dstnid != ni->ni_nid ||
498 msg->ibm_srcstamp != conn->ibc_incarnation ||
499 msg->ibm_dststamp != net->ibn_incarnation) {
500 CERROR("Stale rx from %s\n",
501 libcfs_nid2str(conn->ibc_peer->ibp_nid));
506 /* set time last known alive */
507 kiblnd_peer_alive(conn->ibc_peer);
509 /* racing with connection establishment/teardown! */
511 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
512 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
515 write_lock_irqsave(g_lock, flags);
516 /* must check holding global lock to eliminate race */
517 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
518 list_add_tail(&rx->rx_list, &conn->ibc_early_rxs);
519 write_unlock_irqrestore(g_lock, flags);
522 write_unlock_irqrestore(g_lock, flags);
524 kiblnd_handle_rx(rx);
528 CDEBUG(D_NET, "rx %p conn %p\n", rx, conn);
529 kiblnd_close_conn(conn, err);
531 kiblnd_drop_rx(rx); /* Don't re-post rx. */
535 kiblnd_kvaddr_to_page(unsigned long vaddr)
539 if (is_vmalloc_addr((void *)vaddr)) {
540 page = vmalloc_to_page((void *)vaddr);
544 #ifdef CONFIG_HIGHMEM
545 if (vaddr >= PKMAP_BASE &&
546 vaddr < (PKMAP_BASE + LAST_PKMAP * PAGE_SIZE)) {
547 /* No highmem pages only used for bulk (kiov) I/O */
548 CERROR("find page for address in highmem\n");
552 page = virt_to_page(vaddr);
558 kiblnd_fmr_map_tx(kib_net_t *net, kib_tx_t *tx, kib_rdma_desc_t *rd, int nob)
561 __u64 *pages = tx->tx_pages;
562 kib_fmr_poolset_t *fps;
569 LASSERT(tx->tx_pool);
570 LASSERT(tx->tx_pool->tpo_pool.po_owner);
572 hdev = tx->tx_pool->tpo_hdev;
574 for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
575 for (size = 0; size < rd->rd_frags[i].rf_nob;
576 size += hdev->ibh_page_size) {
577 pages[npages++] = (rd->rd_frags[i].rf_addr &
578 hdev->ibh_page_mask) + size;
582 cpt = tx->tx_pool->tpo_pool.po_owner->ps_cpt;
584 fps = net->ibn_fmr_ps[cpt];
585 rc = kiblnd_fmr_pool_map(fps, pages, npages, 0, &tx->fmr);
587 CERROR("Can't map %d pages: %d\n", npages, rc);
592 * If rd is not tx_rd, it's going to get sent to a peer, who will need
595 rd->rd_key = (rd != tx->tx_rd) ? tx->fmr.fmr_pfmr->fmr->rkey :
596 tx->fmr.fmr_pfmr->fmr->lkey;
597 rd->rd_frags[0].rf_addr &= ~hdev->ibh_page_mask;
598 rd->rd_frags[0].rf_nob = nob;
604 static void kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx)
606 kib_net_t *net = ni->ni_data;
610 if (net->ibn_fmr_ps && tx->fmr.fmr_pfmr) {
611 kiblnd_fmr_pool_unmap(&tx->fmr, tx->tx_status);
612 tx->fmr.fmr_pfmr = NULL;
616 kiblnd_dma_unmap_sg(tx->tx_pool->tpo_hdev->ibh_ibdev,
617 tx->tx_frags, tx->tx_nfrags, tx->tx_dmadir);
622 static int kiblnd_map_tx(lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd,
625 kib_hca_dev_t *hdev = tx->tx_pool->tpo_hdev;
626 kib_net_t *net = ni->ni_data;
627 struct ib_mr *mr = NULL;
632 * If rd is not tx_rd, it's going to get sent to a peer and I'm the
635 tx->tx_dmadir = (rd != tx->tx_rd) ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
636 tx->tx_nfrags = nfrags;
638 rd->rd_nfrags = kiblnd_dma_map_sg(hdev->ibh_ibdev, tx->tx_frags,
639 tx->tx_nfrags, tx->tx_dmadir);
641 for (i = 0, nob = 0; i < rd->rd_nfrags; i++) {
642 rd->rd_frags[i].rf_nob = kiblnd_sg_dma_len(
643 hdev->ibh_ibdev, &tx->tx_frags[i]);
644 rd->rd_frags[i].rf_addr = kiblnd_sg_dma_address(
645 hdev->ibh_ibdev, &tx->tx_frags[i]);
646 nob += rd->rd_frags[i].rf_nob;
649 /* looking for pre-mapping MR */
650 mr = kiblnd_find_rd_dma_mr(hdev, rd);
652 /* found pre-mapping MR */
653 rd->rd_key = (rd != tx->tx_rd) ? mr->rkey : mr->lkey;
658 return kiblnd_fmr_map_tx(net, tx, rd, nob);
664 kiblnd_setup_rd_iov(lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd,
665 unsigned int niov, struct kvec *iov, int offset, int nob)
667 kib_net_t *net = ni->ni_data;
669 struct scatterlist *sg;
678 while (offset >= iov->iov_len) {
679 offset -= iov->iov_len;
689 vaddr = ((unsigned long)iov->iov_base) + offset;
690 page_offset = vaddr & (PAGE_SIZE - 1);
691 page = kiblnd_kvaddr_to_page(vaddr);
693 CERROR("Can't find page\n");
697 fragnob = min((int)(iov->iov_len - offset), nob);
698 fragnob = min(fragnob, (int)PAGE_SIZE - page_offset);
700 sg_set_page(sg, page, fragnob, page_offset);
703 if (offset + fragnob < iov->iov_len) {
713 return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags);
717 kiblnd_setup_rd_kiov(lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd,
718 int nkiov, lnet_kiov_t *kiov, int offset, int nob)
720 kib_net_t *net = ni->ni_data;
721 struct scatterlist *sg;
724 CDEBUG(D_NET, "niov %d offset %d nob %d\n", nkiov, offset, nob);
730 while (offset >= kiov->kiov_len) {
731 offset -= kiov->kiov_len;
741 fragnob = min((int)(kiov->kiov_len - offset), nob);
743 sg_set_page(sg, kiov->kiov_page, fragnob,
744 kiov->kiov_offset + offset);
753 return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags);
757 kiblnd_post_tx_locked(kib_conn_t *conn, kib_tx_t *tx, int credit)
758 __must_hold(&conn->ibc_lock)
760 kib_msg_t *msg = tx->tx_msg;
761 kib_peer_t *peer = conn->ibc_peer;
762 int ver = conn->ibc_version;
765 struct ib_send_wr *bad_wrq;
767 LASSERT(tx->tx_queued);
768 /* We rely on this for QP sizing */
769 LASSERT(tx->tx_nwrq > 0);
770 LASSERT(tx->tx_nwrq <= 1 + IBLND_RDMA_FRAGS(ver));
772 LASSERT(!credit || credit == 1);
773 LASSERT(conn->ibc_outstanding_credits >= 0);
774 LASSERT(conn->ibc_outstanding_credits <= IBLND_MSG_QUEUE_SIZE(ver));
775 LASSERT(conn->ibc_credits >= 0);
776 LASSERT(conn->ibc_credits <= IBLND_MSG_QUEUE_SIZE(ver));
778 if (conn->ibc_nsends_posted == IBLND_CONCURRENT_SENDS(ver)) {
779 /* tx completions outstanding... */
780 CDEBUG(D_NET, "%s: posted enough\n",
781 libcfs_nid2str(peer->ibp_nid));
785 if (credit && !conn->ibc_credits) { /* no credits */
786 CDEBUG(D_NET, "%s: no credits\n",
787 libcfs_nid2str(peer->ibp_nid));
791 if (credit && !IBLND_OOB_CAPABLE(ver) &&
792 conn->ibc_credits == 1 && /* last credit reserved */
793 msg->ibm_type != IBLND_MSG_NOOP) { /* for NOOP */
794 CDEBUG(D_NET, "%s: not using last credit\n",
795 libcfs_nid2str(peer->ibp_nid));
799 /* NB don't drop ibc_lock before bumping tx_sending */
800 list_del(&tx->tx_list);
803 if (msg->ibm_type == IBLND_MSG_NOOP &&
804 (!kiblnd_need_noop(conn) || /* redundant NOOP */
805 (IBLND_OOB_CAPABLE(ver) && /* posted enough NOOP */
806 conn->ibc_noops_posted == IBLND_OOB_MSGS(ver)))) {
808 * OK to drop when posted enough NOOPs, since
809 * kiblnd_check_sends will queue NOOP again when
810 * posted NOOPs complete
812 spin_unlock(&conn->ibc_lock);
813 kiblnd_tx_done(peer->ibp_ni, tx);
814 spin_lock(&conn->ibc_lock);
815 CDEBUG(D_NET, "%s(%d): redundant or enough NOOP\n",
816 libcfs_nid2str(peer->ibp_nid),
817 conn->ibc_noops_posted);
821 kiblnd_pack_msg(peer->ibp_ni, msg, ver, conn->ibc_outstanding_credits,
822 peer->ibp_nid, conn->ibc_incarnation);
824 conn->ibc_credits -= credit;
825 conn->ibc_outstanding_credits = 0;
826 conn->ibc_nsends_posted++;
827 if (msg->ibm_type == IBLND_MSG_NOOP)
828 conn->ibc_noops_posted++;
831 * CAVEAT EMPTOR! This tx could be the PUT_DONE of an RDMA
832 * PUT. If so, it was first queued here as a PUT_REQ, sent and
833 * stashed on ibc_active_txs, matched by an incoming PUT_ACK,
834 * and then re-queued here. It's (just) possible that
835 * tx_sending is non-zero if we've not done the tx_complete()
836 * from the first send; hence the ++ rather than = below.
839 list_add(&tx->tx_list, &conn->ibc_active_txs);
841 /* I'm still holding ibc_lock! */
842 if (conn->ibc_state != IBLND_CONN_ESTABLISHED) {
844 } else if (tx->tx_pool->tpo_pool.po_failed ||
845 conn->ibc_hdev != tx->tx_pool->tpo_hdev) {
846 /* close_conn will launch failover */
849 rc = ib_post_send(conn->ibc_cmid->qp, &tx->tx_wrq->wr, &bad_wrq);
852 conn->ibc_last_send = jiffies;
858 * NB credits are transferred in the actual
859 * message, which can only be the last work item
861 conn->ibc_credits += credit;
862 conn->ibc_outstanding_credits += msg->ibm_credits;
863 conn->ibc_nsends_posted--;
864 if (msg->ibm_type == IBLND_MSG_NOOP)
865 conn->ibc_noops_posted--;
871 done = !tx->tx_sending;
873 list_del(&tx->tx_list);
875 spin_unlock(&conn->ibc_lock);
877 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
878 CERROR("Error %d posting transmit to %s\n",
879 rc, libcfs_nid2str(peer->ibp_nid));
881 CDEBUG(D_NET, "Error %d posting transmit to %s\n",
882 rc, libcfs_nid2str(peer->ibp_nid));
884 kiblnd_close_conn(conn, rc);
887 kiblnd_tx_done(peer->ibp_ni, tx);
889 spin_lock(&conn->ibc_lock);
895 kiblnd_check_sends(kib_conn_t *conn)
897 int ver = conn->ibc_version;
898 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
901 /* Don't send anything until after the connection is established */
902 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
903 CDEBUG(D_NET, "%s too soon\n",
904 libcfs_nid2str(conn->ibc_peer->ibp_nid));
908 spin_lock(&conn->ibc_lock);
910 LASSERT(conn->ibc_nsends_posted <= IBLND_CONCURRENT_SENDS(ver));
911 LASSERT(!IBLND_OOB_CAPABLE(ver) ||
912 conn->ibc_noops_posted <= IBLND_OOB_MSGS(ver));
913 LASSERT(conn->ibc_reserved_credits >= 0);
915 while (conn->ibc_reserved_credits > 0 &&
916 !list_empty(&conn->ibc_tx_queue_rsrvd)) {
917 tx = list_entry(conn->ibc_tx_queue_rsrvd.next,
919 list_del(&tx->tx_list);
920 list_add_tail(&tx->tx_list, &conn->ibc_tx_queue);
921 conn->ibc_reserved_credits--;
924 if (kiblnd_need_noop(conn)) {
925 spin_unlock(&conn->ibc_lock);
927 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
929 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_NOOP, 0);
931 spin_lock(&conn->ibc_lock);
933 kiblnd_queue_tx_locked(tx, conn);
936 kiblnd_conn_addref(conn); /* 1 ref for me.... (see b21911) */
941 if (!list_empty(&conn->ibc_tx_queue_nocred)) {
943 tx = list_entry(conn->ibc_tx_queue_nocred.next,
945 } else if (!list_empty(&conn->ibc_tx_noops)) {
946 LASSERT(!IBLND_OOB_CAPABLE(ver));
948 tx = list_entry(conn->ibc_tx_noops.next,
950 } else if (!list_empty(&conn->ibc_tx_queue)) {
952 tx = list_entry(conn->ibc_tx_queue.next,
958 if (kiblnd_post_tx_locked(conn, tx, credit))
962 spin_unlock(&conn->ibc_lock);
964 kiblnd_conn_decref(conn); /* ...until here */
968 kiblnd_tx_complete(kib_tx_t *tx, int status)
970 int failed = (status != IB_WC_SUCCESS);
971 kib_conn_t *conn = tx->tx_conn;
974 LASSERT(tx->tx_sending > 0);
977 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
978 CNETERR("Tx -> %s cookie %#llx sending %d waiting %d: failed %d\n",
979 libcfs_nid2str(conn->ibc_peer->ibp_nid),
980 tx->tx_cookie, tx->tx_sending, tx->tx_waiting,
983 kiblnd_close_conn(conn, -EIO);
985 kiblnd_peer_alive(conn->ibc_peer);
988 spin_lock(&conn->ibc_lock);
991 * I could be racing with rdma completion. Whoever makes 'tx' idle
992 * gets to free it, which also drops its ref on 'conn'.
995 conn->ibc_nsends_posted--;
996 if (tx->tx_msg->ibm_type == IBLND_MSG_NOOP)
997 conn->ibc_noops_posted--;
1000 tx->tx_waiting = 0; /* don't wait for peer */
1001 tx->tx_status = -EIO;
1004 idle = !tx->tx_sending && /* This is the final callback */
1005 !tx->tx_waiting && /* Not waiting for peer */
1006 !tx->tx_queued; /* Not re-queued (PUT_DONE) */
1008 list_del(&tx->tx_list);
1010 kiblnd_conn_addref(conn); /* 1 ref for me.... */
1012 spin_unlock(&conn->ibc_lock);
1015 kiblnd_tx_done(conn->ibc_peer->ibp_ni, tx);
1017 kiblnd_check_sends(conn);
1019 kiblnd_conn_decref(conn); /* ...until here */
1023 kiblnd_init_tx_msg(lnet_ni_t *ni, kib_tx_t *tx, int type, int body_nob)
1025 kib_hca_dev_t *hdev = tx->tx_pool->tpo_hdev;
1026 struct ib_sge *sge = &tx->tx_sge[tx->tx_nwrq];
1027 struct ib_rdma_wr *wrq = &tx->tx_wrq[tx->tx_nwrq];
1028 int nob = offsetof(kib_msg_t, ibm_u) + body_nob;
1031 LASSERT(tx->tx_nwrq >= 0);
1032 LASSERT(tx->tx_nwrq < IBLND_MAX_RDMA_FRAGS + 1);
1033 LASSERT(nob <= IBLND_MSG_SIZE);
1035 kiblnd_init_msg(tx->tx_msg, type, body_nob);
1037 mr = kiblnd_find_dma_mr(hdev, tx->tx_msgaddr, nob);
1040 sge->lkey = mr->lkey;
1041 sge->addr = tx->tx_msgaddr;
1044 memset(wrq, 0, sizeof(*wrq));
1046 wrq->wr.next = NULL;
1047 wrq->wr.wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_TX);
1048 wrq->wr.sg_list = sge;
1049 wrq->wr.num_sge = 1;
1050 wrq->wr.opcode = IB_WR_SEND;
1051 wrq->wr.send_flags = IB_SEND_SIGNALED;
1057 kiblnd_init_rdma(kib_conn_t *conn, kib_tx_t *tx, int type,
1058 int resid, kib_rdma_desc_t *dstrd, __u64 dstcookie)
1060 kib_msg_t *ibmsg = tx->tx_msg;
1061 kib_rdma_desc_t *srcrd = tx->tx_rd;
1062 struct ib_sge *sge = &tx->tx_sge[0];
1063 struct ib_rdma_wr *wrq = &tx->tx_wrq[0], *next;
1069 LASSERT(!in_interrupt());
1070 LASSERT(!tx->tx_nwrq);
1071 LASSERT(type == IBLND_MSG_GET_DONE ||
1072 type == IBLND_MSG_PUT_DONE);
1075 if (srcidx >= srcrd->rd_nfrags) {
1076 CERROR("Src buffer exhausted: %d frags\n", srcidx);
1081 if (dstidx == dstrd->rd_nfrags) {
1082 CERROR("Dst buffer exhausted: %d frags\n", dstidx);
1087 if (tx->tx_nwrq == IBLND_RDMA_FRAGS(conn->ibc_version)) {
1088 CERROR("RDMA too fragmented for %s (%d): %d/%d src %d/%d dst frags\n",
1089 libcfs_nid2str(conn->ibc_peer->ibp_nid),
1090 IBLND_RDMA_FRAGS(conn->ibc_version),
1091 srcidx, srcrd->rd_nfrags,
1092 dstidx, dstrd->rd_nfrags);
1097 wrknob = min(min(kiblnd_rd_frag_size(srcrd, srcidx),
1098 kiblnd_rd_frag_size(dstrd, dstidx)),
1101 sge = &tx->tx_sge[tx->tx_nwrq];
1102 sge->addr = kiblnd_rd_frag_addr(srcrd, srcidx);
1103 sge->lkey = kiblnd_rd_frag_key(srcrd, srcidx);
1104 sge->length = wrknob;
1106 wrq = &tx->tx_wrq[tx->tx_nwrq];
1109 wrq->wr.next = &next->wr;
1110 wrq->wr.wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_RDMA);
1111 wrq->wr.sg_list = sge;
1112 wrq->wr.num_sge = 1;
1113 wrq->wr.opcode = IB_WR_RDMA_WRITE;
1114 wrq->wr.send_flags = 0;
1116 wrq->remote_addr = kiblnd_rd_frag_addr(dstrd, dstidx);
1117 wrq->rkey = kiblnd_rd_frag_key(dstrd, dstidx);
1119 srcidx = kiblnd_rd_consume_frag(srcrd, srcidx, wrknob);
1120 dstidx = kiblnd_rd_consume_frag(dstrd, dstidx, wrknob);
1129 if (rc < 0) /* no RDMA if completing with failure */
1132 ibmsg->ibm_u.completion.ibcm_status = rc;
1133 ibmsg->ibm_u.completion.ibcm_cookie = dstcookie;
1134 kiblnd_init_tx_msg(conn->ibc_peer->ibp_ni, tx,
1135 type, sizeof(kib_completion_msg_t));
1141 kiblnd_queue_tx_locked(kib_tx_t *tx, kib_conn_t *conn)
1143 struct list_head *q;
1145 LASSERT(tx->tx_nwrq > 0); /* work items set up */
1146 LASSERT(!tx->tx_queued); /* not queued for sending already */
1147 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1150 tx->tx_deadline = jiffies +
1151 msecs_to_jiffies(*kiblnd_tunables.kib_timeout *
1155 kiblnd_conn_addref(conn);
1157 LASSERT(tx->tx_msg->ibm_type != IBLND_MSG_PUT_DONE);
1159 /* PUT_DONE first attached to conn as a PUT_REQ */
1160 LASSERT(tx->tx_conn == conn);
1161 LASSERT(tx->tx_msg->ibm_type == IBLND_MSG_PUT_DONE);
1164 switch (tx->tx_msg->ibm_type) {
1168 case IBLND_MSG_PUT_REQ:
1169 case IBLND_MSG_GET_REQ:
1170 q = &conn->ibc_tx_queue_rsrvd;
1173 case IBLND_MSG_PUT_NAK:
1174 case IBLND_MSG_PUT_ACK:
1175 case IBLND_MSG_PUT_DONE:
1176 case IBLND_MSG_GET_DONE:
1177 q = &conn->ibc_tx_queue_nocred;
1180 case IBLND_MSG_NOOP:
1181 if (IBLND_OOB_CAPABLE(conn->ibc_version))
1182 q = &conn->ibc_tx_queue_nocred;
1184 q = &conn->ibc_tx_noops;
1187 case IBLND_MSG_IMMEDIATE:
1188 q = &conn->ibc_tx_queue;
1192 list_add_tail(&tx->tx_list, q);
1196 kiblnd_queue_tx(kib_tx_t *tx, kib_conn_t *conn)
1198 spin_lock(&conn->ibc_lock);
1199 kiblnd_queue_tx_locked(tx, conn);
1200 spin_unlock(&conn->ibc_lock);
1202 kiblnd_check_sends(conn);
1205 static int kiblnd_resolve_addr(struct rdma_cm_id *cmid,
1206 struct sockaddr_in *srcaddr,
1207 struct sockaddr_in *dstaddr,
1210 unsigned short port;
1213 /* allow the port to be reused */
1214 rc = rdma_set_reuseaddr(cmid, 1);
1216 CERROR("Unable to set reuse on cmid: %d\n", rc);
1220 /* look for a free privileged port */
1221 for (port = PROT_SOCK - 1; port > 0; port--) {
1222 srcaddr->sin_port = htons(port);
1223 rc = rdma_resolve_addr(cmid,
1224 (struct sockaddr *)srcaddr,
1225 (struct sockaddr *)dstaddr,
1228 CDEBUG(D_NET, "bound to port %hu\n", port);
1230 } else if (rc == -EADDRINUSE || rc == -EADDRNOTAVAIL) {
1231 CDEBUG(D_NET, "bind to port %hu failed: %d\n",
1238 CERROR("Failed to bind to a free privileged port\n");
1243 kiblnd_connect_peer(kib_peer_t *peer)
1245 struct rdma_cm_id *cmid;
1247 kib_net_t *net = peer->ibp_ni->ni_data;
1248 struct sockaddr_in srcaddr;
1249 struct sockaddr_in dstaddr;
1253 LASSERT(peer->ibp_connecting > 0);
1255 cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, peer, RDMA_PS_TCP,
1259 CERROR("Can't create CMID for %s: %ld\n",
1260 libcfs_nid2str(peer->ibp_nid), PTR_ERR(cmid));
1266 memset(&srcaddr, 0, sizeof(srcaddr));
1267 srcaddr.sin_family = AF_INET;
1268 srcaddr.sin_addr.s_addr = htonl(dev->ibd_ifip);
1270 memset(&dstaddr, 0, sizeof(dstaddr));
1271 dstaddr.sin_family = AF_INET;
1272 dstaddr.sin_port = htons(*kiblnd_tunables.kib_service);
1273 dstaddr.sin_addr.s_addr = htonl(LNET_NIDADDR(peer->ibp_nid));
1275 kiblnd_peer_addref(peer); /* cmid's ref */
1277 if (*kiblnd_tunables.kib_use_priv_port) {
1278 rc = kiblnd_resolve_addr(cmid, &srcaddr, &dstaddr,
1279 *kiblnd_tunables.kib_timeout * 1000);
1281 rc = rdma_resolve_addr(cmid,
1282 (struct sockaddr *)&srcaddr,
1283 (struct sockaddr *)&dstaddr,
1284 *kiblnd_tunables.kib_timeout * 1000);
1287 /* Can't initiate address resolution: */
1288 CERROR("Can't resolve addr for %s: %d\n",
1289 libcfs_nid2str(peer->ibp_nid), rc);
1293 LASSERT(cmid->device);
1294 CDEBUG(D_NET, "%s: connection bound to %s:%pI4h:%s\n",
1295 libcfs_nid2str(peer->ibp_nid), dev->ibd_ifname,
1296 &dev->ibd_ifip, cmid->device->name);
1301 kiblnd_peer_decref(peer); /* cmid's ref */
1302 rdma_destroy_id(cmid);
1304 kiblnd_peer_connect_failed(peer, 1, rc);
1308 kiblnd_launch_tx(lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid)
1313 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
1314 unsigned long flags;
1318 * If I get here, I've committed to send, so I complete the tx with
1319 * failure on any problems
1321 LASSERT(!tx || !tx->tx_conn); /* only set when assigned a conn */
1322 LASSERT(!tx || tx->tx_nwrq > 0); /* work items have been set up */
1325 * First time, just use a read lock since I expect to find my peer
1328 read_lock_irqsave(g_lock, flags);
1330 peer = kiblnd_find_peer_locked(nid);
1331 if (peer && !list_empty(&peer->ibp_conns)) {
1332 /* Found a peer with an established connection */
1333 conn = kiblnd_get_conn_locked(peer);
1334 kiblnd_conn_addref(conn); /* 1 ref for me... */
1336 read_unlock_irqrestore(g_lock, flags);
1339 kiblnd_queue_tx(tx, conn);
1340 kiblnd_conn_decref(conn); /* ...to here */
1344 read_unlock(g_lock);
1345 /* Re-try with a write lock */
1348 peer = kiblnd_find_peer_locked(nid);
1350 if (list_empty(&peer->ibp_conns)) {
1351 /* found a peer, but it's still connecting... */
1352 LASSERT(peer->ibp_connecting ||
1353 peer->ibp_accepting);
1355 list_add_tail(&tx->tx_list,
1356 &peer->ibp_tx_queue);
1357 write_unlock_irqrestore(g_lock, flags);
1359 conn = kiblnd_get_conn_locked(peer);
1360 kiblnd_conn_addref(conn); /* 1 ref for me... */
1362 write_unlock_irqrestore(g_lock, flags);
1365 kiblnd_queue_tx(tx, conn);
1366 kiblnd_conn_decref(conn); /* ...to here */
1371 write_unlock_irqrestore(g_lock, flags);
1373 /* Allocate a peer ready to add to the peer table and retry */
1374 rc = kiblnd_create_peer(ni, &peer, nid);
1376 CERROR("Can't create peer %s\n", libcfs_nid2str(nid));
1378 tx->tx_status = -EHOSTUNREACH;
1380 kiblnd_tx_done(ni, tx);
1385 write_lock_irqsave(g_lock, flags);
1387 peer2 = kiblnd_find_peer_locked(nid);
1389 if (list_empty(&peer2->ibp_conns)) {
1390 /* found a peer, but it's still connecting... */
1391 LASSERT(peer2->ibp_connecting ||
1392 peer2->ibp_accepting);
1394 list_add_tail(&tx->tx_list,
1395 &peer2->ibp_tx_queue);
1396 write_unlock_irqrestore(g_lock, flags);
1398 conn = kiblnd_get_conn_locked(peer2);
1399 kiblnd_conn_addref(conn); /* 1 ref for me... */
1401 write_unlock_irqrestore(g_lock, flags);
1404 kiblnd_queue_tx(tx, conn);
1405 kiblnd_conn_decref(conn); /* ...to here */
1408 kiblnd_peer_decref(peer);
1412 /* Brand new peer */
1413 LASSERT(!peer->ibp_connecting);
1414 peer->ibp_connecting = 1;
1416 /* always called with a ref on ni, which prevents ni being shutdown */
1417 LASSERT(!((kib_net_t *)ni->ni_data)->ibn_shutdown);
1420 list_add_tail(&tx->tx_list, &peer->ibp_tx_queue);
1422 kiblnd_peer_addref(peer);
1423 list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid));
1425 write_unlock_irqrestore(g_lock, flags);
1427 kiblnd_connect_peer(peer);
1428 kiblnd_peer_decref(peer);
1432 kiblnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg)
1434 lnet_hdr_t *hdr = &lntmsg->msg_hdr;
1435 int type = lntmsg->msg_type;
1436 lnet_process_id_t target = lntmsg->msg_target;
1437 int target_is_router = lntmsg->msg_target_is_router;
1438 int routing = lntmsg->msg_routing;
1439 unsigned int payload_niov = lntmsg->msg_niov;
1440 struct kvec *payload_iov = lntmsg->msg_iov;
1441 lnet_kiov_t *payload_kiov = lntmsg->msg_kiov;
1442 unsigned int payload_offset = lntmsg->msg_offset;
1443 unsigned int payload_nob = lntmsg->msg_len;
1445 kib_rdma_desc_t *rd;
1450 /* NB 'private' is different depending on what we're sending.... */
1452 CDEBUG(D_NET, "sending %d bytes in %d frags to %s\n",
1453 payload_nob, payload_niov, libcfs_id2str(target));
1455 LASSERT(!payload_nob || payload_niov > 0);
1456 LASSERT(payload_niov <= LNET_MAX_IOV);
1458 /* Thread context */
1459 LASSERT(!in_interrupt());
1460 /* payload is either all vaddrs or all pages */
1461 LASSERT(!(payload_kiov && payload_iov));
1469 LASSERT(!payload_nob);
1473 if (routing || target_is_router)
1474 break; /* send IMMEDIATE */
1476 /* is the REPLY message too small for RDMA? */
1477 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[lntmsg->msg_md->md_length]);
1478 if (nob <= IBLND_MSG_SIZE)
1479 break; /* send IMMEDIATE */
1481 tx = kiblnd_get_idle_tx(ni, target.nid);
1483 CERROR("Can't allocate txd for GET to %s\n",
1484 libcfs_nid2str(target.nid));
1489 rd = &ibmsg->ibm_u.get.ibgm_rd;
1490 if (!(lntmsg->msg_md->md_options & LNET_MD_KIOV))
1491 rc = kiblnd_setup_rd_iov(ni, tx, rd,
1492 lntmsg->msg_md->md_niov,
1493 lntmsg->msg_md->md_iov.iov,
1494 0, lntmsg->msg_md->md_length);
1496 rc = kiblnd_setup_rd_kiov(ni, tx, rd,
1497 lntmsg->msg_md->md_niov,
1498 lntmsg->msg_md->md_iov.kiov,
1499 0, lntmsg->msg_md->md_length);
1501 CERROR("Can't setup GET sink for %s: %d\n",
1502 libcfs_nid2str(target.nid), rc);
1503 kiblnd_tx_done(ni, tx);
1507 nob = offsetof(kib_get_msg_t, ibgm_rd.rd_frags[rd->rd_nfrags]);
1508 ibmsg->ibm_u.get.ibgm_cookie = tx->tx_cookie;
1509 ibmsg->ibm_u.get.ibgm_hdr = *hdr;
1511 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_GET_REQ, nob);
1513 tx->tx_lntmsg[1] = lnet_create_reply_msg(ni, lntmsg);
1514 if (!tx->tx_lntmsg[1]) {
1515 CERROR("Can't create reply for GET -> %s\n",
1516 libcfs_nid2str(target.nid));
1517 kiblnd_tx_done(ni, tx);
1521 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg[0,1] on completion */
1522 tx->tx_waiting = 1; /* waiting for GET_DONE */
1523 kiblnd_launch_tx(ni, tx, target.nid);
1526 case LNET_MSG_REPLY:
1528 /* Is the payload small enough not to need RDMA? */
1529 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob]);
1530 if (nob <= IBLND_MSG_SIZE)
1531 break; /* send IMMEDIATE */
1533 tx = kiblnd_get_idle_tx(ni, target.nid);
1535 CERROR("Can't allocate %s txd for %s\n",
1536 type == LNET_MSG_PUT ? "PUT" : "REPLY",
1537 libcfs_nid2str(target.nid));
1542 rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1543 payload_niov, payload_iov,
1544 payload_offset, payload_nob);
1546 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1547 payload_niov, payload_kiov,
1548 payload_offset, payload_nob);
1550 CERROR("Can't setup PUT src for %s: %d\n",
1551 libcfs_nid2str(target.nid), rc);
1552 kiblnd_tx_done(ni, tx);
1557 ibmsg->ibm_u.putreq.ibprm_hdr = *hdr;
1558 ibmsg->ibm_u.putreq.ibprm_cookie = tx->tx_cookie;
1559 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_REQ, sizeof(kib_putreq_msg_t));
1561 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1562 tx->tx_waiting = 1; /* waiting for PUT_{ACK,NAK} */
1563 kiblnd_launch_tx(ni, tx, target.nid);
1567 /* send IMMEDIATE */
1569 LASSERT(offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob])
1572 tx = kiblnd_get_idle_tx(ni, target.nid);
1574 CERROR("Can't send %d to %s: tx descs exhausted\n",
1575 type, libcfs_nid2str(target.nid));
1580 ibmsg->ibm_u.immediate.ibim_hdr = *hdr;
1583 lnet_copy_kiov2flat(IBLND_MSG_SIZE, ibmsg,
1584 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1585 payload_niov, payload_kiov,
1586 payload_offset, payload_nob);
1588 lnet_copy_iov2flat(IBLND_MSG_SIZE, ibmsg,
1589 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1590 payload_niov, payload_iov,
1591 payload_offset, payload_nob);
1593 nob = offsetof(kib_immediate_msg_t, ibim_payload[payload_nob]);
1594 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_IMMEDIATE, nob);
1596 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1597 kiblnd_launch_tx(ni, tx, target.nid);
1602 kiblnd_reply(lnet_ni_t *ni, kib_rx_t *rx, lnet_msg_t *lntmsg)
1604 lnet_process_id_t target = lntmsg->msg_target;
1605 unsigned int niov = lntmsg->msg_niov;
1606 struct kvec *iov = lntmsg->msg_iov;
1607 lnet_kiov_t *kiov = lntmsg->msg_kiov;
1608 unsigned int offset = lntmsg->msg_offset;
1609 unsigned int nob = lntmsg->msg_len;
1613 tx = kiblnd_get_idle_tx(ni, rx->rx_conn->ibc_peer->ibp_nid);
1615 CERROR("Can't get tx for REPLY to %s\n",
1616 libcfs_nid2str(target.nid));
1623 rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1624 niov, iov, offset, nob);
1626 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1627 niov, kiov, offset, nob);
1630 CERROR("Can't setup GET src for %s: %d\n",
1631 libcfs_nid2str(target.nid), rc);
1635 rc = kiblnd_init_rdma(rx->rx_conn, tx,
1636 IBLND_MSG_GET_DONE, nob,
1637 &rx->rx_msg->ibm_u.get.ibgm_rd,
1638 rx->rx_msg->ibm_u.get.ibgm_cookie);
1640 CERROR("Can't setup rdma for GET from %s: %d\n",
1641 libcfs_nid2str(target.nid), rc);
1646 /* No RDMA: local completion may happen now! */
1647 lnet_finalize(ni, lntmsg, 0);
1649 /* RDMA: lnet_finalize(lntmsg) when it completes */
1650 tx->tx_lntmsg[0] = lntmsg;
1653 kiblnd_queue_tx(tx, rx->rx_conn);
1657 kiblnd_tx_done(ni, tx);
1659 lnet_finalize(ni, lntmsg, -EIO);
1663 kiblnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed,
1664 unsigned int niov, struct kvec *iov, lnet_kiov_t *kiov,
1665 unsigned int offset, unsigned int mlen, unsigned int rlen)
1667 kib_rx_t *rx = private;
1668 kib_msg_t *rxmsg = rx->rx_msg;
1669 kib_conn_t *conn = rx->rx_conn;
1672 int post_credit = IBLND_POSTRX_PEER_CREDIT;
1675 LASSERT(mlen <= rlen);
1676 LASSERT(!in_interrupt());
1677 /* Either all pages or all vaddrs */
1678 LASSERT(!(kiov && iov));
1680 switch (rxmsg->ibm_type) {
1684 case IBLND_MSG_IMMEDIATE:
1685 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[rlen]);
1686 if (nob > rx->rx_nob) {
1687 CERROR("Immediate message from %s too big: %d(%d)\n",
1688 libcfs_nid2str(rxmsg->ibm_u.immediate.ibim_hdr.src_nid),
1695 lnet_copy_flat2kiov(niov, kiov, offset,
1696 IBLND_MSG_SIZE, rxmsg,
1697 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1700 lnet_copy_flat2iov(niov, iov, offset,
1701 IBLND_MSG_SIZE, rxmsg,
1702 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1704 lnet_finalize(ni, lntmsg, 0);
1707 case IBLND_MSG_PUT_REQ: {
1709 kib_rdma_desc_t *rd;
1712 lnet_finalize(ni, lntmsg, 0);
1713 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK, 0,
1714 rxmsg->ibm_u.putreq.ibprm_cookie);
1718 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
1720 CERROR("Can't allocate tx for %s\n",
1721 libcfs_nid2str(conn->ibc_peer->ibp_nid));
1722 /* Not replying will break the connection */
1728 rd = &txmsg->ibm_u.putack.ibpam_rd;
1730 rc = kiblnd_setup_rd_iov(ni, tx, rd,
1731 niov, iov, offset, mlen);
1733 rc = kiblnd_setup_rd_kiov(ni, tx, rd,
1734 niov, kiov, offset, mlen);
1736 CERROR("Can't setup PUT sink for %s: %d\n",
1737 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
1738 kiblnd_tx_done(ni, tx);
1739 /* tell peer it's over */
1740 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK, rc,
1741 rxmsg->ibm_u.putreq.ibprm_cookie);
1745 nob = offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[rd->rd_nfrags]);
1746 txmsg->ibm_u.putack.ibpam_src_cookie = rxmsg->ibm_u.putreq.ibprm_cookie;
1747 txmsg->ibm_u.putack.ibpam_dst_cookie = tx->tx_cookie;
1749 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_ACK, nob);
1751 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1752 tx->tx_waiting = 1; /* waiting for PUT_DONE */
1753 kiblnd_queue_tx(tx, conn);
1755 /* reposted buffer reserved for PUT_DONE */
1756 post_credit = IBLND_POSTRX_NO_CREDIT;
1760 case IBLND_MSG_GET_REQ:
1762 /* Optimized GET; RDMA lntmsg's payload */
1763 kiblnd_reply(ni, rx, lntmsg);
1765 /* GET didn't match anything */
1766 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_GET_DONE,
1768 rxmsg->ibm_u.get.ibgm_cookie);
1773 kiblnd_post_rx(rx, post_credit);
1778 kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name)
1780 struct task_struct *task = kthread_run(fn, arg, "%s", name);
1783 return PTR_ERR(task);
1785 atomic_inc(&kiblnd_data.kib_nthreads);
1790 kiblnd_thread_fini(void)
1792 atomic_dec(&kiblnd_data.kib_nthreads);
1796 kiblnd_peer_alive(kib_peer_t *peer)
1798 /* This is racy, but everyone's only writing cfs_time_current() */
1799 peer->ibp_last_alive = cfs_time_current();
1804 kiblnd_peer_notify(kib_peer_t *peer)
1807 unsigned long last_alive = 0;
1808 unsigned long flags;
1810 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1812 if (list_empty(&peer->ibp_conns) &&
1813 !peer->ibp_accepting &&
1814 !peer->ibp_connecting &&
1816 error = peer->ibp_error;
1817 peer->ibp_error = 0;
1819 last_alive = peer->ibp_last_alive;
1822 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1825 lnet_notify(peer->ibp_ni,
1826 peer->ibp_nid, 0, last_alive);
1830 kiblnd_close_conn_locked(kib_conn_t *conn, int error)
1833 * This just does the immediate housekeeping. 'error' is zero for a
1834 * normal shutdown which can happen only after the connection has been
1835 * established. If the connection is established, schedule the
1836 * connection to be finished off by the connd. Otherwise the connd is
1837 * already dealing with it (either to set it up or tear it down).
1838 * Caller holds kib_global_lock exclusively in irq context
1840 kib_peer_t *peer = conn->ibc_peer;
1842 unsigned long flags;
1844 LASSERT(error || conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1846 if (error && !conn->ibc_comms_error)
1847 conn->ibc_comms_error = error;
1849 if (conn->ibc_state != IBLND_CONN_ESTABLISHED)
1850 return; /* already being handled */
1853 list_empty(&conn->ibc_tx_noops) &&
1854 list_empty(&conn->ibc_tx_queue) &&
1855 list_empty(&conn->ibc_tx_queue_rsrvd) &&
1856 list_empty(&conn->ibc_tx_queue_nocred) &&
1857 list_empty(&conn->ibc_active_txs)) {
1858 CDEBUG(D_NET, "closing conn to %s\n",
1859 libcfs_nid2str(peer->ibp_nid));
1861 CNETERR("Closing conn to %s: error %d%s%s%s%s%s\n",
1862 libcfs_nid2str(peer->ibp_nid), error,
1863 list_empty(&conn->ibc_tx_queue) ? "" : "(sending)",
1864 list_empty(&conn->ibc_tx_noops) ? "" : "(sending_noops)",
1865 list_empty(&conn->ibc_tx_queue_rsrvd) ? "" : "(sending_rsrvd)",
1866 list_empty(&conn->ibc_tx_queue_nocred) ? "" : "(sending_nocred)",
1867 list_empty(&conn->ibc_active_txs) ? "" : "(waiting)");
1870 dev = ((kib_net_t *)peer->ibp_ni->ni_data)->ibn_dev;
1871 list_del(&conn->ibc_list);
1872 /* connd (see below) takes over ibc_list's ref */
1874 if (list_empty(&peer->ibp_conns) && /* no more conns */
1875 kiblnd_peer_active(peer)) { /* still in peer table */
1876 kiblnd_unlink_peer_locked(peer);
1878 /* set/clear error on last conn */
1879 peer->ibp_error = conn->ibc_comms_error;
1882 kiblnd_set_conn_state(conn, IBLND_CONN_CLOSING);
1885 kiblnd_dev_can_failover(dev)) {
1886 list_add_tail(&dev->ibd_fail_list,
1887 &kiblnd_data.kib_failed_devs);
1888 wake_up(&kiblnd_data.kib_failover_waitq);
1891 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
1893 list_add_tail(&conn->ibc_list, &kiblnd_data.kib_connd_conns);
1894 wake_up(&kiblnd_data.kib_connd_waitq);
1896 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
1900 kiblnd_close_conn(kib_conn_t *conn, int error)
1902 unsigned long flags;
1904 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1906 kiblnd_close_conn_locked(conn, error);
1908 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1912 kiblnd_handle_early_rxs(kib_conn_t *conn)
1914 unsigned long flags;
1918 LASSERT(!in_interrupt());
1919 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1921 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1922 list_for_each_entry_safe(rx, tmp, &conn->ibc_early_rxs, rx_list) {
1923 list_del(&rx->rx_list);
1924 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1926 kiblnd_handle_rx(rx);
1928 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1930 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1934 kiblnd_abort_txs(kib_conn_t *conn, struct list_head *txs)
1937 struct list_head *tmp;
1938 struct list_head *nxt;
1941 spin_lock(&conn->ibc_lock);
1943 list_for_each_safe(tmp, nxt, txs) {
1944 tx = list_entry(tmp, kib_tx_t, tx_list);
1946 if (txs == &conn->ibc_active_txs) {
1947 LASSERT(!tx->tx_queued);
1948 LASSERT(tx->tx_waiting || tx->tx_sending);
1950 LASSERT(tx->tx_queued);
1953 tx->tx_status = -ECONNABORTED;
1956 if (!tx->tx_sending) {
1958 list_del(&tx->tx_list);
1959 list_add(&tx->tx_list, &zombies);
1963 spin_unlock(&conn->ibc_lock);
1965 kiblnd_txlist_done(conn->ibc_peer->ibp_ni, &zombies, -ECONNABORTED);
1969 kiblnd_finalise_conn(kib_conn_t *conn)
1971 LASSERT(!in_interrupt());
1972 LASSERT(conn->ibc_state > IBLND_CONN_INIT);
1974 kiblnd_set_conn_state(conn, IBLND_CONN_DISCONNECTED);
1977 * abort_receives moves QP state to IB_QPS_ERR. This is only required
1978 * for connections that didn't get as far as being connected, because
1979 * rdma_disconnect() does this for free.
1981 kiblnd_abort_receives(conn);
1984 * Complete all tx descs not waiting for sends to complete.
1985 * NB we should be safe from RDMA now that the QP has changed state
1987 kiblnd_abort_txs(conn, &conn->ibc_tx_noops);
1988 kiblnd_abort_txs(conn, &conn->ibc_tx_queue);
1989 kiblnd_abort_txs(conn, &conn->ibc_tx_queue_rsrvd);
1990 kiblnd_abort_txs(conn, &conn->ibc_tx_queue_nocred);
1991 kiblnd_abort_txs(conn, &conn->ibc_active_txs);
1993 kiblnd_handle_early_rxs(conn);
1997 kiblnd_peer_connect_failed(kib_peer_t *peer, int active, int error)
2000 unsigned long flags;
2003 LASSERT(!in_interrupt());
2005 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2008 LASSERT(peer->ibp_connecting > 0);
2009 peer->ibp_connecting--;
2011 LASSERT(peer->ibp_accepting > 0);
2012 peer->ibp_accepting--;
2015 if (peer->ibp_connecting ||
2016 peer->ibp_accepting) {
2017 /* another connection attempt under way... */
2018 write_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2023 if (list_empty(&peer->ibp_conns)) {
2024 /* Take peer's blocked transmits to complete with error */
2025 list_add(&zombies, &peer->ibp_tx_queue);
2026 list_del_init(&peer->ibp_tx_queue);
2028 if (kiblnd_peer_active(peer))
2029 kiblnd_unlink_peer_locked(peer);
2031 peer->ibp_error = error;
2033 /* Can't have blocked transmits if there are connections */
2034 LASSERT(list_empty(&peer->ibp_tx_queue));
2037 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2039 kiblnd_peer_notify(peer);
2041 if (list_empty(&zombies))
2044 CNETERR("Deleting messages for %s: connection failed\n",
2045 libcfs_nid2str(peer->ibp_nid));
2047 kiblnd_txlist_done(peer->ibp_ni, &zombies, -EHOSTUNREACH);
2051 kiblnd_connreq_done(kib_conn_t *conn, int status)
2053 kib_peer_t *peer = conn->ibc_peer;
2056 struct list_head txs;
2057 unsigned long flags;
2060 active = (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2062 CDEBUG(D_NET, "%s: active(%d), version(%x), status(%d)\n",
2063 libcfs_nid2str(peer->ibp_nid), active,
2064 conn->ibc_version, status);
2066 LASSERT(!in_interrupt());
2067 LASSERT((conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT &&
2068 peer->ibp_connecting > 0) ||
2069 (conn->ibc_state == IBLND_CONN_PASSIVE_WAIT &&
2070 peer->ibp_accepting > 0));
2072 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
2073 conn->ibc_connvars = NULL;
2076 /* failed to establish connection */
2077 kiblnd_peer_connect_failed(peer, active, status);
2078 kiblnd_finalise_conn(conn);
2082 /* connection established */
2083 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2085 conn->ibc_last_send = jiffies;
2086 kiblnd_set_conn_state(conn, IBLND_CONN_ESTABLISHED);
2087 kiblnd_peer_alive(peer);
2090 * Add conn to peer's list and nuke any dangling conns from a different
2093 kiblnd_conn_addref(conn); /* +1 ref for ibc_list */
2094 list_add(&conn->ibc_list, &peer->ibp_conns);
2096 peer->ibp_connecting--;
2098 peer->ibp_accepting--;
2100 if (!peer->ibp_version) {
2101 peer->ibp_version = conn->ibc_version;
2102 peer->ibp_incarnation = conn->ibc_incarnation;
2105 if (peer->ibp_version != conn->ibc_version ||
2106 peer->ibp_incarnation != conn->ibc_incarnation) {
2107 kiblnd_close_stale_conns_locked(peer, conn->ibc_version,
2108 conn->ibc_incarnation);
2109 peer->ibp_version = conn->ibc_version;
2110 peer->ibp_incarnation = conn->ibc_incarnation;
2113 /* grab pending txs while I have the lock */
2114 list_add(&txs, &peer->ibp_tx_queue);
2115 list_del_init(&peer->ibp_tx_queue);
2117 if (!kiblnd_peer_active(peer) || /* peer has been deleted */
2118 conn->ibc_comms_error) { /* error has happened already */
2119 lnet_ni_t *ni = peer->ibp_ni;
2121 /* start to shut down connection */
2122 kiblnd_close_conn_locked(conn, -ECONNABORTED);
2123 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2125 kiblnd_txlist_done(ni, &txs, -ECONNABORTED);
2130 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2132 /* Schedule blocked txs */
2133 spin_lock(&conn->ibc_lock);
2134 list_for_each_entry_safe(tx, tmp, &txs, tx_list) {
2135 list_del(&tx->tx_list);
2137 kiblnd_queue_tx_locked(tx, conn);
2139 spin_unlock(&conn->ibc_lock);
2141 kiblnd_check_sends(conn);
2143 /* schedule blocked rxs */
2144 kiblnd_handle_early_rxs(conn);
2148 kiblnd_reject(struct rdma_cm_id *cmid, kib_rej_t *rej)
2152 rc = rdma_reject(cmid, rej, sizeof(*rej));
2155 CWARN("Error %d sending reject\n", rc);
2159 kiblnd_passive_connect(struct rdma_cm_id *cmid, void *priv, int priv_nob)
2161 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2162 kib_msg_t *reqmsg = priv;
2168 lnet_ni_t *ni = NULL;
2169 kib_net_t *net = NULL;
2171 struct rdma_conn_param cp;
2173 int version = IBLND_MSG_VERSION;
2174 unsigned long flags;
2176 struct sockaddr_in *peer_addr;
2178 LASSERT(!in_interrupt());
2180 /* cmid inherits 'context' from the corresponding listener id */
2181 ibdev = (kib_dev_t *)cmid->context;
2184 memset(&rej, 0, sizeof(rej));
2185 rej.ibr_magic = IBLND_MSG_MAGIC;
2186 rej.ibr_why = IBLND_REJECT_FATAL;
2187 rej.ibr_cp.ibcp_max_msg_size = IBLND_MSG_SIZE;
2189 peer_addr = (struct sockaddr_in *)&cmid->route.addr.dst_addr;
2190 if (*kiblnd_tunables.kib_require_priv_port &&
2191 ntohs(peer_addr->sin_port) >= PROT_SOCK) {
2192 __u32 ip = ntohl(peer_addr->sin_addr.s_addr);
2194 CERROR("Peer's port (%pI4h:%hu) is not privileged\n",
2195 &ip, ntohs(peer_addr->sin_port));
2199 if (priv_nob < offsetof(kib_msg_t, ibm_type)) {
2200 CERROR("Short connection request\n");
2205 * Future protocol version compatibility support! If the
2206 * o2iblnd-specific protocol changes, or when LNET unifies
2207 * protocols over all LNDs, the initial connection will
2208 * negotiate a protocol version. I trap this here to avoid
2209 * console errors; the reject tells the peer which protocol I
2212 if (reqmsg->ibm_magic == LNET_PROTO_MAGIC ||
2213 reqmsg->ibm_magic == __swab32(LNET_PROTO_MAGIC))
2215 if (reqmsg->ibm_magic == IBLND_MSG_MAGIC &&
2216 reqmsg->ibm_version != IBLND_MSG_VERSION &&
2217 reqmsg->ibm_version != IBLND_MSG_VERSION_1)
2219 if (reqmsg->ibm_magic == __swab32(IBLND_MSG_MAGIC) &&
2220 reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION) &&
2221 reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION_1))
2224 rc = kiblnd_unpack_msg(reqmsg, priv_nob);
2226 CERROR("Can't parse connection request: %d\n", rc);
2230 nid = reqmsg->ibm_srcnid;
2231 ni = lnet_net2ni(LNET_NIDNET(reqmsg->ibm_dstnid));
2234 net = (kib_net_t *)ni->ni_data;
2235 rej.ibr_incarnation = net->ibn_incarnation;
2238 if (!ni || /* no matching net */
2239 ni->ni_nid != reqmsg->ibm_dstnid || /* right NET, wrong NID! */
2240 net->ibn_dev != ibdev) { /* wrong device */
2241 CERROR("Can't accept %s on %s (%s:%d:%pI4h): bad dst nid %s\n",
2242 libcfs_nid2str(nid),
2243 !ni ? "NA" : libcfs_nid2str(ni->ni_nid),
2244 ibdev->ibd_ifname, ibdev->ibd_nnets,
2246 libcfs_nid2str(reqmsg->ibm_dstnid));
2251 /* check time stamp as soon as possible */
2252 if (reqmsg->ibm_dststamp &&
2253 reqmsg->ibm_dststamp != net->ibn_incarnation) {
2254 CWARN("Stale connection request\n");
2255 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2259 /* I can accept peer's version */
2260 version = reqmsg->ibm_version;
2262 if (reqmsg->ibm_type != IBLND_MSG_CONNREQ) {
2263 CERROR("Unexpected connreq msg type: %x from %s\n",
2264 reqmsg->ibm_type, libcfs_nid2str(nid));
2268 if (reqmsg->ibm_u.connparams.ibcp_queue_depth !=
2269 IBLND_MSG_QUEUE_SIZE(version)) {
2270 CERROR("Can't accept %s: incompatible queue depth %d (%d wanted)\n",
2271 libcfs_nid2str(nid), reqmsg->ibm_u.connparams.ibcp_queue_depth,
2272 IBLND_MSG_QUEUE_SIZE(version));
2274 if (version == IBLND_MSG_VERSION)
2275 rej.ibr_why = IBLND_REJECT_MSG_QUEUE_SIZE;
2280 if (reqmsg->ibm_u.connparams.ibcp_max_frags !=
2281 IBLND_RDMA_FRAGS(version)) {
2282 CERROR("Can't accept %s(version %x): incompatible max_frags %d (%d wanted)\n",
2283 libcfs_nid2str(nid), version,
2284 reqmsg->ibm_u.connparams.ibcp_max_frags,
2285 IBLND_RDMA_FRAGS(version));
2287 if (version == IBLND_MSG_VERSION)
2288 rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2293 if (reqmsg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2294 CERROR("Can't accept %s: message size %d too big (%d max)\n",
2295 libcfs_nid2str(nid),
2296 reqmsg->ibm_u.connparams.ibcp_max_msg_size,
2301 /* assume 'nid' is a new peer; create */
2302 rc = kiblnd_create_peer(ni, &peer, nid);
2304 CERROR("Can't create peer for %s\n", libcfs_nid2str(nid));
2305 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2309 write_lock_irqsave(g_lock, flags);
2311 peer2 = kiblnd_find_peer_locked(nid);
2313 if (!peer2->ibp_version) {
2314 peer2->ibp_version = version;
2315 peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2318 /* not the guy I've talked with */
2319 if (peer2->ibp_incarnation != reqmsg->ibm_srcstamp ||
2320 peer2->ibp_version != version) {
2321 kiblnd_close_peer_conns_locked(peer2, -ESTALE);
2322 write_unlock_irqrestore(g_lock, flags);
2324 CWARN("Conn stale %s [old ver: %x, new ver: %x]\n",
2325 libcfs_nid2str(nid), peer2->ibp_version, version);
2327 kiblnd_peer_decref(peer);
2328 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2332 /* tie-break connection race in favour of the higher NID */
2333 if (peer2->ibp_connecting &&
2335 write_unlock_irqrestore(g_lock, flags);
2337 CWARN("Conn race %s\n", libcfs_nid2str(peer2->ibp_nid));
2339 kiblnd_peer_decref(peer);
2340 rej.ibr_why = IBLND_REJECT_CONN_RACE;
2344 peer2->ibp_accepting++;
2345 kiblnd_peer_addref(peer2);
2347 write_unlock_irqrestore(g_lock, flags);
2348 kiblnd_peer_decref(peer);
2351 /* Brand new peer */
2352 LASSERT(!peer->ibp_accepting);
2353 LASSERT(!peer->ibp_version &&
2354 !peer->ibp_incarnation);
2356 peer->ibp_accepting = 1;
2357 peer->ibp_version = version;
2358 peer->ibp_incarnation = reqmsg->ibm_srcstamp;
2360 /* I have a ref on ni that prevents it being shutdown */
2361 LASSERT(!net->ibn_shutdown);
2363 kiblnd_peer_addref(peer);
2364 list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid));
2366 write_unlock_irqrestore(g_lock, flags);
2369 conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_PASSIVE_WAIT, version);
2371 kiblnd_peer_connect_failed(peer, 0, -ENOMEM);
2372 kiblnd_peer_decref(peer);
2373 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2378 * conn now "owns" cmid, so I return success from here on to ensure the
2379 * CM callback doesn't destroy cmid.
2381 conn->ibc_incarnation = reqmsg->ibm_srcstamp;
2382 conn->ibc_credits = IBLND_MSG_QUEUE_SIZE(version);
2383 conn->ibc_reserved_credits = IBLND_MSG_QUEUE_SIZE(version);
2384 LASSERT(conn->ibc_credits + conn->ibc_reserved_credits + IBLND_OOB_MSGS(version)
2385 <= IBLND_RX_MSGS(version));
2387 ackmsg = &conn->ibc_connvars->cv_msg;
2388 memset(ackmsg, 0, sizeof(*ackmsg));
2390 kiblnd_init_msg(ackmsg, IBLND_MSG_CONNACK,
2391 sizeof(ackmsg->ibm_u.connparams));
2392 ackmsg->ibm_u.connparams.ibcp_queue_depth = IBLND_MSG_QUEUE_SIZE(version);
2393 ackmsg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2394 ackmsg->ibm_u.connparams.ibcp_max_frags = IBLND_RDMA_FRAGS(version);
2396 kiblnd_pack_msg(ni, ackmsg, version, 0, nid, reqmsg->ibm_srcstamp);
2398 memset(&cp, 0, sizeof(cp));
2399 cp.private_data = ackmsg;
2400 cp.private_data_len = ackmsg->ibm_nob;
2401 cp.responder_resources = 0; /* No atomic ops or RDMA reads */
2402 cp.initiator_depth = 0;
2403 cp.flow_control = 1;
2404 cp.retry_count = *kiblnd_tunables.kib_retry_count;
2405 cp.rnr_retry_count = *kiblnd_tunables.kib_rnr_retry_count;
2407 CDEBUG(D_NET, "Accept %s\n", libcfs_nid2str(nid));
2409 rc = rdma_accept(cmid, &cp);
2411 CERROR("Can't accept %s: %d\n", libcfs_nid2str(nid), rc);
2412 rej.ibr_version = version;
2413 rej.ibr_why = IBLND_REJECT_FATAL;
2415 kiblnd_reject(cmid, &rej);
2416 kiblnd_connreq_done(conn, rc);
2417 kiblnd_conn_decref(conn);
2427 rej.ibr_version = version;
2428 rej.ibr_cp.ibcp_queue_depth = IBLND_MSG_QUEUE_SIZE(version);
2429 rej.ibr_cp.ibcp_max_frags = IBLND_RDMA_FRAGS(version);
2430 kiblnd_reject(cmid, &rej);
2432 return -ECONNREFUSED;
2436 kiblnd_reconnect(kib_conn_t *conn, int version,
2437 __u64 incarnation, int why, kib_connparams_t *cp)
2439 kib_peer_t *peer = conn->ibc_peer;
2442 unsigned long flags;
2444 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2445 LASSERT(peer->ibp_connecting > 0); /* 'conn' at least */
2447 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2450 * retry connection if it's still needed and no other connection
2451 * attempts (active or passive) are in progress
2452 * NB: reconnect is still needed even when ibp_tx_queue is
2453 * empty if ibp_version != version because reconnect may be
2454 * initiated by kiblnd_query()
2456 if ((!list_empty(&peer->ibp_tx_queue) ||
2457 peer->ibp_version != version) &&
2458 peer->ibp_connecting == 1 &&
2459 !peer->ibp_accepting) {
2461 peer->ibp_connecting++;
2463 peer->ibp_version = version;
2464 peer->ibp_incarnation = incarnation;
2467 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2477 case IBLND_REJECT_CONN_STALE:
2481 case IBLND_REJECT_CONN_RACE:
2482 reason = "conn race";
2485 case IBLND_REJECT_CONN_UNCOMPAT:
2486 reason = "version negotiation";
2490 CNETERR("%s: retrying (%s), %x, %x, queue_dep: %d, max_frag: %d, msg_size: %d\n",
2491 libcfs_nid2str(peer->ibp_nid),
2492 reason, IBLND_MSG_VERSION, version,
2493 cp ? cp->ibcp_queue_depth : IBLND_MSG_QUEUE_SIZE(version),
2494 cp ? cp->ibcp_max_frags : IBLND_RDMA_FRAGS(version),
2495 cp ? cp->ibcp_max_msg_size : IBLND_MSG_SIZE);
2497 kiblnd_connect_peer(peer);
2501 kiblnd_rejected(kib_conn_t *conn, int reason, void *priv, int priv_nob)
2503 kib_peer_t *peer = conn->ibc_peer;
2505 LASSERT(!in_interrupt());
2506 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2509 case IB_CM_REJ_STALE_CONN:
2510 kiblnd_reconnect(conn, IBLND_MSG_VERSION, 0,
2511 IBLND_REJECT_CONN_STALE, NULL);
2514 case IB_CM_REJ_INVALID_SERVICE_ID:
2515 CNETERR("%s rejected: no listener at %d\n",
2516 libcfs_nid2str(peer->ibp_nid),
2517 *kiblnd_tunables.kib_service);
2520 case IB_CM_REJ_CONSUMER_DEFINED:
2521 if (priv_nob >= offsetof(kib_rej_t, ibr_padding)) {
2522 kib_rej_t *rej = priv;
2523 kib_connparams_t *cp = NULL;
2525 __u64 incarnation = -1;
2527 /* NB. default incarnation is -1 because:
2528 * a) V1 will ignore dst incarnation in connreq.
2529 * b) V2 will provide incarnation while rejecting me,
2530 * -1 will be overwrote.
2532 * if I try to connect to a V1 peer with V2 protocol,
2533 * it rejected me then upgrade to V2, I have no idea
2534 * about the upgrading and try to reconnect with V1,
2535 * in this case upgraded V2 can find out I'm trying to
2536 * talk to the old guy and reject me(incarnation is -1).
2539 if (rej->ibr_magic == __swab32(IBLND_MSG_MAGIC) ||
2540 rej->ibr_magic == __swab32(LNET_PROTO_MAGIC)) {
2541 __swab32s(&rej->ibr_magic);
2542 __swab16s(&rej->ibr_version);
2546 if (priv_nob >= sizeof(kib_rej_t) &&
2547 rej->ibr_version > IBLND_MSG_VERSION_1) {
2549 * priv_nob is always 148 in current version
2550 * of OFED, so we still need to check version.
2551 * (define of IB_CM_REJ_PRIVATE_DATA_SIZE)
2556 __swab64s(&rej->ibr_incarnation);
2557 __swab16s(&cp->ibcp_queue_depth);
2558 __swab16s(&cp->ibcp_max_frags);
2559 __swab32s(&cp->ibcp_max_msg_size);
2562 incarnation = rej->ibr_incarnation;
2565 if (rej->ibr_magic != IBLND_MSG_MAGIC &&
2566 rej->ibr_magic != LNET_PROTO_MAGIC) {
2567 CERROR("%s rejected: consumer defined fatal error\n",
2568 libcfs_nid2str(peer->ibp_nid));
2572 if (rej->ibr_version != IBLND_MSG_VERSION &&
2573 rej->ibr_version != IBLND_MSG_VERSION_1) {
2574 CERROR("%s rejected: o2iblnd version %x error\n",
2575 libcfs_nid2str(peer->ibp_nid),
2580 if (rej->ibr_why == IBLND_REJECT_FATAL &&
2581 rej->ibr_version == IBLND_MSG_VERSION_1) {
2582 CDEBUG(D_NET, "rejected by old version peer %s: %x\n",
2583 libcfs_nid2str(peer->ibp_nid), rej->ibr_version);
2585 if (conn->ibc_version != IBLND_MSG_VERSION_1)
2586 rej->ibr_why = IBLND_REJECT_CONN_UNCOMPAT;
2589 switch (rej->ibr_why) {
2590 case IBLND_REJECT_CONN_RACE:
2591 case IBLND_REJECT_CONN_STALE:
2592 case IBLND_REJECT_CONN_UNCOMPAT:
2593 kiblnd_reconnect(conn, rej->ibr_version,
2594 incarnation, rej->ibr_why, cp);
2597 case IBLND_REJECT_MSG_QUEUE_SIZE:
2598 CERROR("%s rejected: incompatible message queue depth %d, %d\n",
2599 libcfs_nid2str(peer->ibp_nid),
2600 cp ? cp->ibcp_queue_depth :
2601 IBLND_MSG_QUEUE_SIZE(rej->ibr_version),
2602 IBLND_MSG_QUEUE_SIZE(conn->ibc_version));
2605 case IBLND_REJECT_RDMA_FRAGS:
2606 CERROR("%s rejected: incompatible # of RDMA fragments %d, %d\n",
2607 libcfs_nid2str(peer->ibp_nid),
2608 cp ? cp->ibcp_max_frags :
2609 IBLND_RDMA_FRAGS(rej->ibr_version),
2610 IBLND_RDMA_FRAGS(conn->ibc_version));
2613 case IBLND_REJECT_NO_RESOURCES:
2614 CERROR("%s rejected: o2iblnd no resources\n",
2615 libcfs_nid2str(peer->ibp_nid));
2618 case IBLND_REJECT_FATAL:
2619 CERROR("%s rejected: o2iblnd fatal error\n",
2620 libcfs_nid2str(peer->ibp_nid));
2624 CERROR("%s rejected: o2iblnd reason %d\n",
2625 libcfs_nid2str(peer->ibp_nid),
2633 CNETERR("%s rejected: reason %d, size %d\n",
2634 libcfs_nid2str(peer->ibp_nid), reason, priv_nob);
2638 kiblnd_connreq_done(conn, -ECONNREFUSED);
2642 kiblnd_check_connreply(kib_conn_t *conn, void *priv, int priv_nob)
2644 kib_peer_t *peer = conn->ibc_peer;
2645 lnet_ni_t *ni = peer->ibp_ni;
2646 kib_net_t *net = ni->ni_data;
2647 kib_msg_t *msg = priv;
2648 int ver = conn->ibc_version;
2649 int rc = kiblnd_unpack_msg(msg, priv_nob);
2650 unsigned long flags;
2655 CERROR("Can't unpack connack from %s: %d\n",
2656 libcfs_nid2str(peer->ibp_nid), rc);
2660 if (msg->ibm_type != IBLND_MSG_CONNACK) {
2661 CERROR("Unexpected message %d from %s\n",
2662 msg->ibm_type, libcfs_nid2str(peer->ibp_nid));
2667 if (ver != msg->ibm_version) {
2668 CERROR("%s replied version %x is different with requested version %x\n",
2669 libcfs_nid2str(peer->ibp_nid), msg->ibm_version, ver);
2674 if (msg->ibm_u.connparams.ibcp_queue_depth !=
2675 IBLND_MSG_QUEUE_SIZE(ver)) {
2676 CERROR("%s has incompatible queue depth %d(%d wanted)\n",
2677 libcfs_nid2str(peer->ibp_nid),
2678 msg->ibm_u.connparams.ibcp_queue_depth,
2679 IBLND_MSG_QUEUE_SIZE(ver));
2684 if (msg->ibm_u.connparams.ibcp_max_frags !=
2685 IBLND_RDMA_FRAGS(ver)) {
2686 CERROR("%s has incompatible max_frags %d (%d wanted)\n",
2687 libcfs_nid2str(peer->ibp_nid),
2688 msg->ibm_u.connparams.ibcp_max_frags,
2689 IBLND_RDMA_FRAGS(ver));
2694 if (msg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2695 CERROR("%s max message size %d too big (%d max)\n",
2696 libcfs_nid2str(peer->ibp_nid),
2697 msg->ibm_u.connparams.ibcp_max_msg_size,
2703 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2704 if (msg->ibm_dstnid == ni->ni_nid &&
2705 msg->ibm_dststamp == net->ibn_incarnation)
2709 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2712 CERROR("Bad connection reply from %s, rc = %d, version: %x max_frags: %d\n",
2713 libcfs_nid2str(peer->ibp_nid), rc,
2714 msg->ibm_version, msg->ibm_u.connparams.ibcp_max_frags);
2718 conn->ibc_incarnation = msg->ibm_srcstamp;
2720 conn->ibc_reserved_credits = IBLND_MSG_QUEUE_SIZE(ver);
2721 LASSERT(conn->ibc_credits + conn->ibc_reserved_credits + IBLND_OOB_MSGS(ver)
2722 <= IBLND_RX_MSGS(ver));
2724 kiblnd_connreq_done(conn, 0);
2729 * NB My QP has already established itself, so I handle anything going
2730 * wrong here by setting ibc_comms_error.
2731 * kiblnd_connreq_done(0) moves the conn state to ESTABLISHED, but then
2732 * immediately tears it down.
2735 conn->ibc_comms_error = rc;
2736 kiblnd_connreq_done(conn, 0);
2740 kiblnd_active_connect(struct rdma_cm_id *cmid)
2742 kib_peer_t *peer = (kib_peer_t *)cmid->context;
2745 struct rdma_conn_param cp;
2748 unsigned long flags;
2751 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2753 incarnation = peer->ibp_incarnation;
2754 version = !peer->ibp_version ? IBLND_MSG_VERSION :
2757 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2759 conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_ACTIVE_CONNECT, version);
2761 kiblnd_peer_connect_failed(peer, 1, -ENOMEM);
2762 kiblnd_peer_decref(peer); /* lose cmid's ref */
2767 * conn "owns" cmid now, so I return success from here on to ensure the
2768 * CM callback doesn't destroy cmid. conn also takes over cmid's ref
2771 msg = &conn->ibc_connvars->cv_msg;
2773 memset(msg, 0, sizeof(*msg));
2774 kiblnd_init_msg(msg, IBLND_MSG_CONNREQ, sizeof(msg->ibm_u.connparams));
2775 msg->ibm_u.connparams.ibcp_queue_depth = IBLND_MSG_QUEUE_SIZE(version);
2776 msg->ibm_u.connparams.ibcp_max_frags = IBLND_RDMA_FRAGS(version);
2777 msg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2779 kiblnd_pack_msg(peer->ibp_ni, msg, version,
2780 0, peer->ibp_nid, incarnation);
2782 memset(&cp, 0, sizeof(cp));
2783 cp.private_data = msg;
2784 cp.private_data_len = msg->ibm_nob;
2785 cp.responder_resources = 0; /* No atomic ops or RDMA reads */
2786 cp.initiator_depth = 0;
2787 cp.flow_control = 1;
2788 cp.retry_count = *kiblnd_tunables.kib_retry_count;
2789 cp.rnr_retry_count = *kiblnd_tunables.kib_rnr_retry_count;
2791 LASSERT(cmid->context == (void *)conn);
2792 LASSERT(conn->ibc_cmid == cmid);
2794 rc = rdma_connect(cmid, &cp);
2796 CERROR("Can't connect to %s: %d\n",
2797 libcfs_nid2str(peer->ibp_nid), rc);
2798 kiblnd_connreq_done(conn, rc);
2799 kiblnd_conn_decref(conn);
2806 kiblnd_cm_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2812 switch (event->event) {
2814 CERROR("Unexpected event: %d, status: %d\n",
2815 event->event, event->status);
2818 case RDMA_CM_EVENT_CONNECT_REQUEST:
2819 /* destroy cmid on failure */
2820 rc = kiblnd_passive_connect(cmid,
2821 (void *)KIBLND_CONN_PARAM(event),
2822 KIBLND_CONN_PARAM_LEN(event));
2823 CDEBUG(D_NET, "connreq: %d\n", rc);
2826 case RDMA_CM_EVENT_ADDR_ERROR:
2827 peer = (kib_peer_t *)cmid->context;
2828 CNETERR("%s: ADDR ERROR %d\n",
2829 libcfs_nid2str(peer->ibp_nid), event->status);
2830 kiblnd_peer_connect_failed(peer, 1, -EHOSTUNREACH);
2831 kiblnd_peer_decref(peer);
2832 return -EHOSTUNREACH; /* rc destroys cmid */
2834 case RDMA_CM_EVENT_ADDR_RESOLVED:
2835 peer = (kib_peer_t *)cmid->context;
2837 CDEBUG(D_NET, "%s Addr resolved: %d\n",
2838 libcfs_nid2str(peer->ibp_nid), event->status);
2840 if (event->status) {
2841 CNETERR("Can't resolve address for %s: %d\n",
2842 libcfs_nid2str(peer->ibp_nid), event->status);
2845 rc = rdma_resolve_route(
2846 cmid, *kiblnd_tunables.kib_timeout * 1000);
2849 /* Can't initiate route resolution */
2850 CERROR("Can't resolve route for %s: %d\n",
2851 libcfs_nid2str(peer->ibp_nid), rc);
2853 kiblnd_peer_connect_failed(peer, 1, rc);
2854 kiblnd_peer_decref(peer);
2855 return rc; /* rc destroys cmid */
2857 case RDMA_CM_EVENT_ROUTE_ERROR:
2858 peer = (kib_peer_t *)cmid->context;
2859 CNETERR("%s: ROUTE ERROR %d\n",
2860 libcfs_nid2str(peer->ibp_nid), event->status);
2861 kiblnd_peer_connect_failed(peer, 1, -EHOSTUNREACH);
2862 kiblnd_peer_decref(peer);
2863 return -EHOSTUNREACH; /* rc destroys cmid */
2865 case RDMA_CM_EVENT_ROUTE_RESOLVED:
2866 peer = (kib_peer_t *)cmid->context;
2867 CDEBUG(D_NET, "%s Route resolved: %d\n",
2868 libcfs_nid2str(peer->ibp_nid), event->status);
2871 return kiblnd_active_connect(cmid);
2873 CNETERR("Can't resolve route for %s: %d\n",
2874 libcfs_nid2str(peer->ibp_nid), event->status);
2875 kiblnd_peer_connect_failed(peer, 1, event->status);
2876 kiblnd_peer_decref(peer);
2877 return event->status; /* rc destroys cmid */
2879 case RDMA_CM_EVENT_UNREACHABLE:
2880 conn = (kib_conn_t *)cmid->context;
2881 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
2882 conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
2883 CNETERR("%s: UNREACHABLE %d\n",
2884 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
2885 kiblnd_connreq_done(conn, -ENETDOWN);
2886 kiblnd_conn_decref(conn);
2889 case RDMA_CM_EVENT_CONNECT_ERROR:
2890 conn = (kib_conn_t *)cmid->context;
2891 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
2892 conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
2893 CNETERR("%s: CONNECT ERROR %d\n",
2894 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
2895 kiblnd_connreq_done(conn, -ENOTCONN);
2896 kiblnd_conn_decref(conn);
2899 case RDMA_CM_EVENT_REJECTED:
2900 conn = (kib_conn_t *)cmid->context;
2901 switch (conn->ibc_state) {
2905 case IBLND_CONN_PASSIVE_WAIT:
2906 CERROR("%s: REJECTED %d\n",
2907 libcfs_nid2str(conn->ibc_peer->ibp_nid),
2909 kiblnd_connreq_done(conn, -ECONNRESET);
2912 case IBLND_CONN_ACTIVE_CONNECT:
2913 kiblnd_rejected(conn, event->status,
2914 (void *)KIBLND_CONN_PARAM(event),
2915 KIBLND_CONN_PARAM_LEN(event));
2918 kiblnd_conn_decref(conn);
2921 case RDMA_CM_EVENT_ESTABLISHED:
2922 conn = (kib_conn_t *)cmid->context;
2923 switch (conn->ibc_state) {
2927 case IBLND_CONN_PASSIVE_WAIT:
2928 CDEBUG(D_NET, "ESTABLISHED (passive): %s\n",
2929 libcfs_nid2str(conn->ibc_peer->ibp_nid));
2930 kiblnd_connreq_done(conn, 0);
2933 case IBLND_CONN_ACTIVE_CONNECT:
2934 CDEBUG(D_NET, "ESTABLISHED(active): %s\n",
2935 libcfs_nid2str(conn->ibc_peer->ibp_nid));
2936 kiblnd_check_connreply(conn,
2937 (void *)KIBLND_CONN_PARAM(event),
2938 KIBLND_CONN_PARAM_LEN(event));
2941 /* net keeps its ref on conn! */
2944 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
2945 CDEBUG(D_NET, "Ignore TIMEWAIT_EXIT event\n");
2947 case RDMA_CM_EVENT_DISCONNECTED:
2948 conn = (kib_conn_t *)cmid->context;
2949 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
2950 CERROR("%s DISCONNECTED\n",
2951 libcfs_nid2str(conn->ibc_peer->ibp_nid));
2952 kiblnd_connreq_done(conn, -ECONNRESET);
2954 kiblnd_close_conn(conn, 0);
2956 kiblnd_conn_decref(conn);
2957 cmid->context = NULL;
2960 case RDMA_CM_EVENT_DEVICE_REMOVAL:
2961 LCONSOLE_ERROR_MSG(0x131,
2962 "Received notification of device removal\n"
2963 "Please shutdown LNET to allow this to proceed\n");
2965 * Can't remove network from underneath LNET for now, so I have
2970 case RDMA_CM_EVENT_ADDR_CHANGE:
2971 LCONSOLE_INFO("Physical link changed (eg hca/port)\n");
2977 kiblnd_check_txs_locked(kib_conn_t *conn, struct list_head *txs)
2980 struct list_head *ttmp;
2982 list_for_each(ttmp, txs) {
2983 tx = list_entry(ttmp, kib_tx_t, tx_list);
2985 if (txs != &conn->ibc_active_txs) {
2986 LASSERT(tx->tx_queued);
2988 LASSERT(!tx->tx_queued);
2989 LASSERT(tx->tx_waiting || tx->tx_sending);
2992 if (cfs_time_aftereq(jiffies, tx->tx_deadline)) {
2993 CERROR("Timed out tx: %s, %lu seconds\n",
2994 kiblnd_queue2str(conn, txs),
2995 cfs_duration_sec(jiffies - tx->tx_deadline));
3004 kiblnd_conn_timed_out_locked(kib_conn_t *conn)
3006 return kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue) ||
3007 kiblnd_check_txs_locked(conn, &conn->ibc_tx_noops) ||
3008 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_rsrvd) ||
3009 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_nocred) ||
3010 kiblnd_check_txs_locked(conn, &conn->ibc_active_txs);
3014 kiblnd_check_conns(int idx)
3017 LIST_HEAD(checksends);
3018 struct list_head *peers = &kiblnd_data.kib_peers[idx];
3019 struct list_head *ptmp;
3023 struct list_head *ctmp;
3024 unsigned long flags;
3027 * NB. We expect to have a look at all the peers and not find any
3028 * RDMAs to time out, so we just use a shared lock while we
3031 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3033 list_for_each(ptmp, peers) {
3034 peer = list_entry(ptmp, kib_peer_t, ibp_list);
3036 list_for_each(ctmp, &peer->ibp_conns) {
3040 conn = list_entry(ctmp, kib_conn_t, ibc_list);
3042 LASSERT(conn->ibc_state == IBLND_CONN_ESTABLISHED);
3044 spin_lock(&conn->ibc_lock);
3046 sendnoop = kiblnd_need_noop(conn);
3047 timedout = kiblnd_conn_timed_out_locked(conn);
3048 if (!sendnoop && !timedout) {
3049 spin_unlock(&conn->ibc_lock);
3054 CERROR("Timed out RDMA with %s (%lu): c: %u, oc: %u, rc: %u\n",
3055 libcfs_nid2str(peer->ibp_nid),
3056 cfs_duration_sec(cfs_time_current() -
3057 peer->ibp_last_alive),
3059 conn->ibc_outstanding_credits,
3060 conn->ibc_reserved_credits);
3061 list_add(&conn->ibc_connd_list, &closes);
3063 list_add(&conn->ibc_connd_list, &checksends);
3065 /* +ref for 'closes' or 'checksends' */
3066 kiblnd_conn_addref(conn);
3068 spin_unlock(&conn->ibc_lock);
3072 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3075 * Handle timeout by closing the whole
3076 * connection. We can only be sure RDMA activity
3077 * has ceased once the QP has been modified.
3079 list_for_each_entry_safe(conn, tmp, &closes, ibc_connd_list) {
3080 list_del(&conn->ibc_connd_list);
3081 kiblnd_close_conn(conn, -ETIMEDOUT);
3082 kiblnd_conn_decref(conn);
3086 * In case we have enough credits to return via a
3087 * NOOP, but there were no non-blocking tx descs
3088 * free to do it last time...
3090 while (!list_empty(&checksends)) {
3091 conn = list_entry(checksends.next, kib_conn_t, ibc_connd_list);
3092 list_del(&conn->ibc_connd_list);
3093 kiblnd_check_sends(conn);
3094 kiblnd_conn_decref(conn);
3099 kiblnd_disconnect_conn(kib_conn_t *conn)
3101 LASSERT(!in_interrupt());
3102 LASSERT(current == kiblnd_data.kib_connd);
3103 LASSERT(conn->ibc_state == IBLND_CONN_CLOSING);
3105 rdma_disconnect(conn->ibc_cmid);
3106 kiblnd_finalise_conn(conn);
3108 kiblnd_peer_notify(conn->ibc_peer);
3112 kiblnd_connd(void *arg)
3115 unsigned long flags;
3121 unsigned long deadline = jiffies;
3123 cfs_block_allsigs();
3125 init_waitqueue_entry(&wait, current);
3126 kiblnd_data.kib_connd = current;
3128 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3130 while (!kiblnd_data.kib_shutdown) {
3133 if (!list_empty(&kiblnd_data.kib_connd_zombies)) {
3134 conn = list_entry(kiblnd_data.kib_connd_zombies.next,
3135 kib_conn_t, ibc_list);
3136 list_del(&conn->ibc_list);
3138 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock,
3142 kiblnd_destroy_conn(conn);
3144 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3147 if (!list_empty(&kiblnd_data.kib_connd_conns)) {
3148 conn = list_entry(kiblnd_data.kib_connd_conns.next,
3149 kib_conn_t, ibc_list);
3150 list_del(&conn->ibc_list);
3152 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock,
3156 kiblnd_disconnect_conn(conn);
3157 kiblnd_conn_decref(conn);
3159 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3162 /* careful with the jiffy wrap... */
3163 timeout = (int)(deadline - jiffies);
3167 int chunk = kiblnd_data.kib_peer_hash_size;
3169 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
3173 * Time to check for RDMA timeouts on a few more
3174 * peers: I do checks every 'p' seconds on a
3175 * proportion of the peer table and I need to check
3176 * every connection 'n' times within a timeout
3177 * interval, to ensure I detect a timeout on any
3178 * connection within (n+1)/n times the timeout
3181 if (*kiblnd_tunables.kib_timeout > n * p)
3182 chunk = (chunk * n * p) /
3183 *kiblnd_tunables.kib_timeout;
3187 for (i = 0; i < chunk; i++) {
3188 kiblnd_check_conns(peer_index);
3189 peer_index = (peer_index + 1) %
3190 kiblnd_data.kib_peer_hash_size;
3193 deadline += msecs_to_jiffies(p * MSEC_PER_SEC);
3194 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3200 /* Nothing to do for 'timeout' */
3201 set_current_state(TASK_INTERRUPTIBLE);
3202 add_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3203 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
3205 schedule_timeout(timeout);
3207 remove_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3208 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3211 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
3213 kiblnd_thread_fini();
3218 kiblnd_qp_event(struct ib_event *event, void *arg)
3220 kib_conn_t *conn = arg;
3222 switch (event->event) {
3223 case IB_EVENT_COMM_EST:
3224 CDEBUG(D_NET, "%s established\n",
3225 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3229 CERROR("%s: Async QP event type %d\n",
3230 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3236 kiblnd_complete(struct ib_wc *wc)
3238 switch (kiblnd_wreqid2type(wc->wr_id)) {
3242 case IBLND_WID_RDMA:
3244 * We only get RDMA completion notification if it fails. All
3245 * subsequent work items, including the final SEND will fail
3246 * too. However we can't print out any more info about the
3247 * failing RDMA because 'tx' might be back on the idle list or
3248 * even reused already if we didn't manage to post all our work
3251 CNETERR("RDMA (tx: %p) failed: %d\n",
3252 kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3256 kiblnd_tx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3260 kiblnd_rx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status,
3267 kiblnd_cq_completion(struct ib_cq *cq, void *arg)
3270 * NB I'm not allowed to schedule this conn once its refcount has
3271 * reached 0. Since fundamentally I'm racing with scheduler threads
3272 * consuming my CQ I could be called after all completions have
3273 * occurred. But in this case, !ibc_nrx && !ibc_nsends_posted
3274 * and this CQ is about to be destroyed so I NOOP.
3276 kib_conn_t *conn = arg;
3277 struct kib_sched_info *sched = conn->ibc_sched;
3278 unsigned long flags;
3280 LASSERT(cq == conn->ibc_cq);
3282 spin_lock_irqsave(&sched->ibs_lock, flags);
3284 conn->ibc_ready = 1;
3286 if (!conn->ibc_scheduled &&
3287 (conn->ibc_nrx > 0 ||
3288 conn->ibc_nsends_posted > 0)) {
3289 kiblnd_conn_addref(conn); /* +1 ref for sched_conns */
3290 conn->ibc_scheduled = 1;
3291 list_add_tail(&conn->ibc_sched_list, &sched->ibs_conns);
3293 if (waitqueue_active(&sched->ibs_waitq))
3294 wake_up(&sched->ibs_waitq);
3297 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3301 kiblnd_cq_event(struct ib_event *event, void *arg)
3303 kib_conn_t *conn = arg;
3305 CERROR("%s: async CQ event type %d\n",
3306 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3310 kiblnd_scheduler(void *arg)
3312 long id = (long)arg;
3313 struct kib_sched_info *sched;
3316 unsigned long flags;
3322 cfs_block_allsigs();
3324 init_waitqueue_entry(&wait, current);
3326 sched = kiblnd_data.kib_scheds[KIB_THREAD_CPT(id)];
3328 rc = cfs_cpt_bind(lnet_cpt_table(), sched->ibs_cpt);
3330 CWARN("Failed to bind on CPT %d, please verify whether all CPUs are healthy and reload modules if necessary, otherwise your system might under risk of low performance\n",
3334 spin_lock_irqsave(&sched->ibs_lock, flags);
3336 while (!kiblnd_data.kib_shutdown) {
3337 if (busy_loops++ >= IBLND_RESCHED) {
3338 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3343 spin_lock_irqsave(&sched->ibs_lock, flags);
3348 if (!list_empty(&sched->ibs_conns)) {
3349 conn = list_entry(sched->ibs_conns.next, kib_conn_t,
3351 /* take over kib_sched_conns' ref on conn... */
3352 LASSERT(conn->ibc_scheduled);
3353 list_del(&conn->ibc_sched_list);
3354 conn->ibc_ready = 0;
3356 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3358 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3360 rc = ib_req_notify_cq(conn->ibc_cq,
3363 CWARN("%s: ib_req_notify_cq failed: %d, closing connection\n",
3364 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
3365 kiblnd_close_conn(conn, -EIO);
3366 kiblnd_conn_decref(conn);
3367 spin_lock_irqsave(&sched->ibs_lock,
3372 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3376 CWARN("%s: ib_poll_cq failed: %d, closing connection\n",
3377 libcfs_nid2str(conn->ibc_peer->ibp_nid),
3379 kiblnd_close_conn(conn, -EIO);
3380 kiblnd_conn_decref(conn);
3381 spin_lock_irqsave(&sched->ibs_lock, flags);
3385 spin_lock_irqsave(&sched->ibs_lock, flags);
3387 if (rc || conn->ibc_ready) {
3389 * There may be another completion waiting; get
3390 * another scheduler to check while I handle
3393 /* +1 ref for sched_conns */
3394 kiblnd_conn_addref(conn);
3395 list_add_tail(&conn->ibc_sched_list,
3397 if (waitqueue_active(&sched->ibs_waitq))
3398 wake_up(&sched->ibs_waitq);
3400 conn->ibc_scheduled = 0;
3404 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3405 kiblnd_complete(&wc);
3407 spin_lock_irqsave(&sched->ibs_lock, flags);
3410 kiblnd_conn_decref(conn); /* ...drop my ref from above */
3417 set_current_state(TASK_INTERRUPTIBLE);
3418 add_wait_queue_exclusive(&sched->ibs_waitq, &wait);
3419 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3424 remove_wait_queue(&sched->ibs_waitq, &wait);
3425 spin_lock_irqsave(&sched->ibs_lock, flags);
3428 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3430 kiblnd_thread_fini();
3435 kiblnd_failover_thread(void *arg)
3437 rwlock_t *glock = &kiblnd_data.kib_global_lock;
3440 unsigned long flags;
3443 LASSERT(*kiblnd_tunables.kib_dev_failover);
3445 cfs_block_allsigs();
3447 init_waitqueue_entry(&wait, current);
3448 write_lock_irqsave(glock, flags);
3450 while (!kiblnd_data.kib_shutdown) {
3451 int do_failover = 0;
3454 list_for_each_entry(dev, &kiblnd_data.kib_failed_devs,
3456 if (time_before(cfs_time_current(),
3457 dev->ibd_next_failover))
3464 list_del_init(&dev->ibd_fail_list);
3465 dev->ibd_failover = 1;
3466 write_unlock_irqrestore(glock, flags);
3468 rc = kiblnd_dev_failover(dev);
3470 write_lock_irqsave(glock, flags);
3472 LASSERT(dev->ibd_failover);
3473 dev->ibd_failover = 0;
3474 if (rc >= 0) { /* Device is OK or failover succeed */
3475 dev->ibd_next_failover = cfs_time_shift(3);
3479 /* failed to failover, retry later */
3480 dev->ibd_next_failover =
3481 cfs_time_shift(min(dev->ibd_failed_failover, 10));
3482 if (kiblnd_dev_can_failover(dev)) {
3483 list_add_tail(&dev->ibd_fail_list,
3484 &kiblnd_data.kib_failed_devs);
3490 /* long sleep if no more pending failover */
3491 long_sleep = list_empty(&kiblnd_data.kib_failed_devs);
3493 set_current_state(TASK_INTERRUPTIBLE);
3494 add_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3495 write_unlock_irqrestore(glock, flags);
3497 rc = schedule_timeout(long_sleep ? cfs_time_seconds(10) :
3498 cfs_time_seconds(1));
3499 remove_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3500 write_lock_irqsave(glock, flags);
3502 if (!long_sleep || rc)
3506 * have a long sleep, routine check all active devices,
3507 * we need checking like this because if there is not active
3508 * connection on the dev and no SEND from local, we may listen
3509 * on wrong HCA for ever while there is a bonding failover
3511 list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3512 if (kiblnd_dev_can_failover(dev)) {
3513 list_add_tail(&dev->ibd_fail_list,
3514 &kiblnd_data.kib_failed_devs);
3519 write_unlock_irqrestore(glock, flags);
3521 kiblnd_thread_fini();