1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
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>
44 kiblnd_tx_done (lnet_ni_t *ni, kib_tx_t *tx)
46 lnet_msg_t *lntmsg[2];
47 kib_net_t *net = ni->ni_data;
51 LASSERT (net != NULL);
52 LASSERT (!cfs_in_interrupt());
53 LASSERT (!tx->tx_queued); /* mustn't be queued for sending */
54 LASSERT (tx->tx_sending == 0); /* mustn't be awaiting sent callback */
55 LASSERT (!tx->tx_waiting); /* mustn't be awaiting peer response */
56 LASSERT (tx->tx_pool != NULL);
58 kiblnd_unmap_tx(ni, tx);
60 /* tx may have up to 2 lnet msgs to finalise */
61 lntmsg[0] = tx->tx_lntmsg[0]; tx->tx_lntmsg[0] = NULL;
62 lntmsg[1] = tx->tx_lntmsg[1]; tx->tx_lntmsg[1] = NULL;
65 if (tx->tx_conn != NULL) {
66 LASSERT (ni == tx->tx_conn->ibc_peer->ibp_ni);
68 kiblnd_conn_decref(tx->tx_conn);
75 kiblnd_pool_free_node(&tx->tx_pool->tpo_pool, &tx->tx_list);
77 /* delay finalize until my descs have been freed */
78 for (i = 0; i < 2; i++) {
79 if (lntmsg[i] == NULL)
82 lnet_finalize(ni, lntmsg[i], rc);
87 kiblnd_txlist_done (lnet_ni_t *ni, cfs_list_t *txlist, int status)
91 while (!cfs_list_empty (txlist)) {
92 tx = cfs_list_entry (txlist->next, kib_tx_t, tx_list);
94 cfs_list_del(&tx->tx_list);
97 tx->tx_status = status;
98 kiblnd_tx_done(ni, tx);
103 kiblnd_get_idle_tx (lnet_ni_t *ni)
105 kib_net_t *net = (kib_net_t *)ni->ni_data;
109 node = kiblnd_pool_alloc_node(&net->ibn_tx_ps.tps_poolset);
112 tx = container_of(node, kib_tx_t, tx_list);
114 LASSERT (tx->tx_nwrq == 0);
115 LASSERT (!tx->tx_queued);
116 LASSERT (tx->tx_sending == 0);
117 LASSERT (!tx->tx_waiting);
118 LASSERT (tx->tx_status == 0);
119 LASSERT (tx->tx_conn == NULL);
120 LASSERT (tx->tx_lntmsg[0] == NULL);
121 LASSERT (tx->tx_lntmsg[1] == NULL);
122 LASSERT (tx->tx_u.pmr == NULL);
123 LASSERT (tx->tx_nfrags == 0);
129 kiblnd_drop_rx (kib_rx_t *rx)
131 kib_conn_t *conn = rx->rx_conn;
134 cfs_spin_lock_irqsave(&kiblnd_data.kib_sched_lock, flags);
135 LASSERT (conn->ibc_nrx > 0);
137 cfs_spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock, flags);
139 kiblnd_conn_decref(conn);
143 kiblnd_post_rx (kib_rx_t *rx, int credit)
145 kib_conn_t *conn = rx->rx_conn;
146 kib_net_t *net = conn->ibc_peer->ibp_ni->ni_data;
147 struct ib_recv_wr *bad_wrq = NULL;
151 LASSERT (net != NULL);
152 LASSERT (!cfs_in_interrupt());
153 LASSERT (credit == IBLND_POSTRX_NO_CREDIT ||
154 credit == IBLND_POSTRX_PEER_CREDIT ||
155 credit == IBLND_POSTRX_RSRVD_CREDIT);
157 mr = kiblnd_find_dma_mr(net, rx->rx_msgaddr, IBLND_MSG_SIZE);
158 LASSERT (mr != NULL);
160 rx->rx_sge.lkey = mr->lkey;
161 rx->rx_sge.addr = rx->rx_msgaddr;
162 rx->rx_sge.length = IBLND_MSG_SIZE;
164 rx->rx_wrq.next = NULL;
165 rx->rx_wrq.sg_list = &rx->rx_sge;
166 rx->rx_wrq.num_sge = 1;
167 rx->rx_wrq.wr_id = kiblnd_ptr2wreqid(rx, IBLND_WID_RX);
169 LASSERT (conn->ibc_state >= IBLND_CONN_INIT);
170 LASSERT (rx->rx_nob >= 0); /* not posted */
172 if (conn->ibc_state > IBLND_CONN_ESTABLISHED) {
173 kiblnd_drop_rx(rx); /* No more posts for this rx */
177 rx->rx_nob = -1; /* flag posted */
179 rc = ib_post_recv(conn->ibc_cmid->qp, &rx->rx_wrq, &bad_wrq);
181 CERROR("Can't post rx for %s: %d, bad_wrq: %p\n",
182 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc, bad_wrq);
186 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) /* Initial post */
190 kiblnd_close_conn(conn, rc);
191 kiblnd_drop_rx(rx); /* No more posts for this rx */
195 if (credit == IBLND_POSTRX_NO_CREDIT)
198 cfs_spin_lock(&conn->ibc_lock);
199 if (credit == IBLND_POSTRX_PEER_CREDIT)
200 conn->ibc_outstanding_credits++;
202 conn->ibc_reserved_credits++;
203 cfs_spin_unlock(&conn->ibc_lock);
205 kiblnd_check_sends(conn);
210 kiblnd_find_waiting_tx_locked(kib_conn_t *conn, int txtype, __u64 cookie)
214 cfs_list_for_each(tmp, &conn->ibc_active_txs) {
215 kib_tx_t *tx = cfs_list_entry(tmp, kib_tx_t, tx_list);
217 LASSERT (!tx->tx_queued);
218 LASSERT (tx->tx_sending != 0 || tx->tx_waiting);
220 if (tx->tx_cookie != cookie)
223 if (tx->tx_waiting &&
224 tx->tx_msg->ibm_type == txtype)
227 CWARN("Bad completion: %swaiting, type %x (wanted %x)\n",
228 tx->tx_waiting ? "" : "NOT ",
229 tx->tx_msg->ibm_type, txtype);
235 kiblnd_handle_completion(kib_conn_t *conn, int txtype, int status, __u64 cookie)
238 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
241 cfs_spin_lock(&conn->ibc_lock);
243 tx = kiblnd_find_waiting_tx_locked(conn, txtype, cookie);
245 cfs_spin_unlock(&conn->ibc_lock);
247 CWARN("Unmatched completion type %x cookie "LPX64" from %s\n",
248 txtype, cookie, libcfs_nid2str(conn->ibc_peer->ibp_nid));
249 kiblnd_close_conn(conn, -EPROTO);
253 if (tx->tx_status == 0) { /* success so far */
254 if (status < 0) { /* failed? */
255 tx->tx_status = status;
256 } else if (txtype == IBLND_MSG_GET_REQ) {
257 lnet_set_reply_msg_len(ni, tx->tx_lntmsg[1], status);
263 idle = !tx->tx_queued && (tx->tx_sending == 0);
265 cfs_list_del(&tx->tx_list);
267 cfs_spin_unlock(&conn->ibc_lock);
270 kiblnd_tx_done(ni, tx);
274 kiblnd_send_completion (kib_conn_t *conn, int type, int status, __u64 cookie)
276 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
277 kib_tx_t *tx = kiblnd_get_idle_tx(ni);
280 CERROR("Can't get tx for completion %x for %s\n",
281 type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
285 tx->tx_msg->ibm_u.completion.ibcm_status = status;
286 tx->tx_msg->ibm_u.completion.ibcm_cookie = cookie;
287 kiblnd_init_tx_msg(ni, tx, type, sizeof(kib_completion_msg_t));
289 kiblnd_queue_tx(tx, conn);
293 kiblnd_handle_rx (kib_rx_t *rx)
295 kib_msg_t *msg = rx->rx_msg;
296 kib_conn_t *conn = rx->rx_conn;
297 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
298 int credits = msg->ibm_credits;
304 LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
306 CDEBUG (D_NET, "Received %x[%d] from %s\n",
307 msg->ibm_type, credits,
308 libcfs_nid2str(conn->ibc_peer->ibp_nid));
311 /* Have I received credits that will let me send? */
312 cfs_spin_lock(&conn->ibc_lock);
314 if (conn->ibc_credits + credits >
315 IBLND_MSG_QUEUE_SIZE(conn->ibc_version)) {
316 rc2 = conn->ibc_credits;
317 cfs_spin_unlock(&conn->ibc_lock);
319 CERROR("Bad credits from %s: %d + %d > %d\n",
320 libcfs_nid2str(conn->ibc_peer->ibp_nid),
322 IBLND_MSG_QUEUE_SIZE(conn->ibc_version));
324 kiblnd_close_conn(conn, -EPROTO);
325 kiblnd_post_rx(rx, IBLND_POSTRX_NO_CREDIT);
329 conn->ibc_credits += credits;
331 cfs_spin_unlock(&conn->ibc_lock);
332 kiblnd_check_sends(conn);
335 switch (msg->ibm_type) {
337 CERROR("Bad IBLND message type %x from %s\n",
338 msg->ibm_type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
339 post_credit = IBLND_POSTRX_NO_CREDIT;
344 if (IBLND_OOB_CAPABLE(conn->ibc_version))
345 post_credit = IBLND_POSTRX_NO_CREDIT;
347 post_credit = IBLND_POSTRX_PEER_CREDIT;
350 case IBLND_MSG_IMMEDIATE:
351 post_credit = IBLND_POSTRX_DONT_POST;
352 rc = lnet_parse(ni, &msg->ibm_u.immediate.ibim_hdr,
353 msg->ibm_srcnid, rx, 0);
354 if (rc < 0) /* repost on error */
355 post_credit = IBLND_POSTRX_PEER_CREDIT;
358 case IBLND_MSG_PUT_REQ:
359 post_credit = IBLND_POSTRX_DONT_POST;
360 rc = lnet_parse(ni, &msg->ibm_u.putreq.ibprm_hdr,
361 msg->ibm_srcnid, rx, 1);
362 if (rc < 0) /* repost on error */
363 post_credit = IBLND_POSTRX_PEER_CREDIT;
366 case IBLND_MSG_PUT_NAK:
367 CWARN ("PUT_NACK from %s\n",
368 libcfs_nid2str(conn->ibc_peer->ibp_nid));
369 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
370 kiblnd_handle_completion(conn, IBLND_MSG_PUT_REQ,
371 msg->ibm_u.completion.ibcm_status,
372 msg->ibm_u.completion.ibcm_cookie);
375 case IBLND_MSG_PUT_ACK:
376 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
378 cfs_spin_lock(&conn->ibc_lock);
379 tx = kiblnd_find_waiting_tx_locked(conn, IBLND_MSG_PUT_REQ,
380 msg->ibm_u.putack.ibpam_src_cookie);
382 cfs_list_del(&tx->tx_list);
383 cfs_spin_unlock(&conn->ibc_lock);
386 CERROR("Unmatched PUT_ACK from %s\n",
387 libcfs_nid2str(conn->ibc_peer->ibp_nid));
392 LASSERT (tx->tx_waiting);
393 /* CAVEAT EMPTOR: I could be racing with tx_complete, but...
394 * (a) I can overwrite tx_msg since my peer has received it!
395 * (b) tx_waiting set tells tx_complete() it's not done. */
397 tx->tx_nwrq = 0; /* overwrite PUT_REQ */
399 rc2 = kiblnd_init_rdma(conn, tx, IBLND_MSG_PUT_DONE,
400 kiblnd_rd_size(&msg->ibm_u.putack.ibpam_rd),
401 &msg->ibm_u.putack.ibpam_rd,
402 msg->ibm_u.putack.ibpam_dst_cookie);
404 CERROR("Can't setup rdma for PUT to %s: %d\n",
405 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc2);
407 cfs_spin_lock(&conn->ibc_lock);
408 tx->tx_waiting = 0; /* clear waiting and queue atomically */
409 kiblnd_queue_tx_locked(tx, conn);
410 cfs_spin_unlock(&conn->ibc_lock);
413 case IBLND_MSG_PUT_DONE:
414 post_credit = IBLND_POSTRX_PEER_CREDIT;
415 kiblnd_handle_completion(conn, IBLND_MSG_PUT_ACK,
416 msg->ibm_u.completion.ibcm_status,
417 msg->ibm_u.completion.ibcm_cookie);
420 case IBLND_MSG_GET_REQ:
421 post_credit = IBLND_POSTRX_DONT_POST;
422 rc = lnet_parse(ni, &msg->ibm_u.get.ibgm_hdr,
423 msg->ibm_srcnid, rx, 1);
424 if (rc < 0) /* repost on error */
425 post_credit = IBLND_POSTRX_PEER_CREDIT;
428 case IBLND_MSG_GET_DONE:
429 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
430 kiblnd_handle_completion(conn, IBLND_MSG_GET_REQ,
431 msg->ibm_u.completion.ibcm_status,
432 msg->ibm_u.completion.ibcm_cookie);
436 if (rc < 0) /* protocol error */
437 kiblnd_close_conn(conn, rc);
439 if (post_credit != IBLND_POSTRX_DONT_POST)
440 kiblnd_post_rx(rx, post_credit);
444 kiblnd_rx_complete (kib_rx_t *rx, int status, int nob)
446 kib_msg_t *msg = rx->rx_msg;
447 kib_conn_t *conn = rx->rx_conn;
448 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
449 kib_net_t *net = ni->ni_data;
453 LASSERT (net != NULL);
454 LASSERT (rx->rx_nob < 0); /* was posted */
455 rx->rx_nob = 0; /* isn't now */
457 if (conn->ibc_state > IBLND_CONN_ESTABLISHED)
460 if (status != IB_WC_SUCCESS) {
461 CDEBUG(D_NETERROR, "Rx from %s failed: %d\n",
462 libcfs_nid2str(conn->ibc_peer->ibp_nid), status);
469 rc = kiblnd_unpack_msg(msg, rx->rx_nob);
471 CERROR ("Error %d unpacking rx from %s\n",
472 rc, libcfs_nid2str(conn->ibc_peer->ibp_nid));
476 if (msg->ibm_srcnid != conn->ibc_peer->ibp_nid ||
477 msg->ibm_dstnid != ni->ni_nid ||
478 msg->ibm_srcstamp != conn->ibc_incarnation ||
479 msg->ibm_dststamp != net->ibn_incarnation) {
480 CERROR ("Stale rx from %s\n",
481 libcfs_nid2str(conn->ibc_peer->ibp_nid));
486 /* set time last known alive */
487 kiblnd_peer_alive(conn->ibc_peer);
489 /* racing with connection establishment/teardown! */
491 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
492 cfs_rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
495 cfs_write_lock_irqsave(g_lock, flags);
496 /* must check holding global lock to eliminate race */
497 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
498 cfs_list_add_tail(&rx->rx_list, &conn->ibc_early_rxs);
499 cfs_write_unlock_irqrestore(g_lock, flags);
502 cfs_write_unlock_irqrestore(g_lock, flags);
504 kiblnd_handle_rx(rx);
508 CDEBUG(D_NET, "rx %p conn %p\n", rx, conn);
509 kiblnd_close_conn(conn, err);
511 kiblnd_drop_rx(rx); /* Don't re-post rx. */
515 kiblnd_kvaddr_to_page (unsigned long vaddr)
519 if (vaddr >= VMALLOC_START &&
520 vaddr < VMALLOC_END) {
521 page = vmalloc_to_page ((void *)vaddr);
522 LASSERT (page != NULL);
525 #ifdef CONFIG_HIGHMEM
526 if (vaddr >= PKMAP_BASE &&
527 vaddr < (PKMAP_BASE + LAST_PKMAP * PAGE_SIZE)) {
528 /* No highmem pages only used for bulk (kiov) I/O */
529 CERROR("find page for address in highmem\n");
533 page = virt_to_page (vaddr);
534 LASSERT (page != NULL);
539 kiblnd_fmr_map_tx(kib_net_t *net, kib_tx_t *tx, kib_rdma_desc_t *rd, int nob)
541 kib_dev_t *ibdev = net->ibn_dev;
542 __u64 *pages = tx->tx_pages;
548 for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
549 for (size = 0; size < rd->rd_frags[i].rf_nob;
550 size += ibdev->ibd_page_size) {
551 pages[npages ++] = (rd->rd_frags[i].rf_addr &
552 ibdev->ibd_page_mask) + size;
556 rc = kiblnd_fmr_pool_map(&net->ibn_fmr_ps, pages, npages, 0, &tx->tx_u.fmr);
558 CERROR ("Can't map %d pages: %d\n", npages, rc);
562 /* If rd is not tx_rd, it's going to get sent to a peer, who will need
564 rd->rd_key = (rd != tx->tx_rd) ? tx->tx_u.fmr.fmr_pfmr->fmr->rkey :
565 tx->tx_u.fmr.fmr_pfmr->fmr->lkey;
566 rd->rd_frags[0].rf_addr &= ~ibdev->ibd_page_mask;
567 rd->rd_frags[0].rf_nob = nob;
574 kiblnd_pmr_map_tx(kib_net_t *net, kib_tx_t *tx, kib_rdma_desc_t *rd, int nob)
579 iova = rd->rd_frags[0].rf_addr & ~net->ibn_dev->ibd_page_mask;
581 rc = kiblnd_pmr_pool_map(&net->ibn_pmr_ps, rd, &iova, &tx->tx_u.pmr);
583 CERROR("Failed to create MR by phybuf: %d\n", rc);
587 /* If rd is not tx_rd, it's going to get sent to a peer, who will need
589 rd->rd_key = (rd != tx->tx_rd) ? tx->tx_u.pmr->pmr_mr->rkey :
590 tx->tx_u.pmr->pmr_mr->lkey;
592 rd->rd_frags[0].rf_addr = iova;
593 rd->rd_frags[0].rf_nob = nob;
599 kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx)
601 kib_net_t *net = ni->ni_data;
603 LASSERT (net != NULL);
605 if (net->ibn_with_fmr && tx->tx_u.fmr.fmr_pfmr != NULL) {
606 kiblnd_fmr_pool_unmap(&tx->tx_u.fmr, tx->tx_status);
607 tx->tx_u.fmr.fmr_pfmr = NULL;
608 } else if (net->ibn_with_pmr && tx->tx_u.pmr != NULL) {
609 kiblnd_pmr_pool_unmap(tx->tx_u.pmr);
613 if (tx->tx_nfrags != 0) {
614 kiblnd_dma_unmap_sg(net->ibn_dev->ibd_cmid->device,
615 tx->tx_frags, tx->tx_nfrags, tx->tx_dmadir);
621 kiblnd_map_tx(lnet_ni_t *ni, kib_tx_t *tx,
622 kib_rdma_desc_t *rd, int nfrags)
624 kib_net_t *net = ni->ni_data;
625 struct ib_mr *mr = NULL;
629 /* If rd is not tx_rd, it's going to get sent to a peer and I'm the
631 tx->tx_dmadir = (rd != tx->tx_rd) ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
632 tx->tx_nfrags = nfrags;
635 kiblnd_dma_map_sg(net->ibn_dev->ibd_cmid->device,
636 tx->tx_frags, tx->tx_nfrags, tx->tx_dmadir);
638 for (i = 0, nob = 0; i < rd->rd_nfrags; i++) {
639 rd->rd_frags[i].rf_nob = kiblnd_sg_dma_len(
640 net->ibn_dev->ibd_cmid->device, &tx->tx_frags[i]);
641 rd->rd_frags[i].rf_addr = kiblnd_sg_dma_address(
642 net->ibn_dev->ibd_cmid->device, &tx->tx_frags[i]);
643 nob += rd->rd_frags[i].rf_nob;
646 /* looking for pre-mapping MR */
647 mr = kiblnd_find_rd_dma_mr(net, rd);
649 /* found pre-mapping MR */
650 rd->rd_key = (rd != tx->tx_rd) ? mr->rkey : mr->lkey;
654 if (net->ibn_with_fmr)
655 return kiblnd_fmr_map_tx(net, tx, rd, nob);
656 else if (net->ibn_with_pmr)
657 return kiblnd_pmr_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 iovec *iov, int offset, int nob)
667 kib_net_t *net = ni->ni_data;
669 struct scatterlist *sg;
676 LASSERT (net != NULL);
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 get_kiov_length (int nkiov, lnet_kiov_t *kiov, int offset, int nob)
725 fragnob = min((int)(kiov->kiov_len - offset), nob);
739 kiblnd_setup_rd_kiov (lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd,
740 int nkiov, lnet_kiov_t *kiov, int offset, int nob)
742 kib_net_t *net = ni->ni_data;
743 struct scatterlist *sg;
746 CDEBUG(D_NET, "niov %d offset %d nob %d\n", nkiov, offset, nob);
750 LASSERT (net != NULL);
752 while (offset >= kiov->kiov_len) {
753 offset -= kiov->kiov_len;
760 tx->tx_nfrags = get_kiov_length(nkiov, kiov, offset, nob);
761 sg_init_table(sg, tx->tx_nfrags);
765 fragnob = min((int)(kiov->kiov_len - offset), nob);
767 sg_set_page(sg, kiov->kiov_page, fragnob,
768 kiov->kiov_offset + offset);
777 return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags);
781 kiblnd_post_tx_locked (kib_conn_t *conn, kib_tx_t *tx, int credit)
783 kib_msg_t *msg = tx->tx_msg;
784 kib_peer_t *peer = conn->ibc_peer;
785 int ver = conn->ibc_version;
788 struct ib_send_wr *bad_wrq;
790 LASSERT (tx->tx_queued);
791 /* We rely on this for QP sizing */
792 LASSERT (tx->tx_nwrq > 0);
793 LASSERT (tx->tx_nwrq <= 1 + IBLND_RDMA_FRAGS(ver));
795 LASSERT (credit == 0 || credit == 1);
796 LASSERT (conn->ibc_outstanding_credits >= 0);
797 LASSERT (conn->ibc_outstanding_credits <= IBLND_MSG_QUEUE_SIZE(ver));
798 LASSERT (conn->ibc_credits >= 0);
799 LASSERT (conn->ibc_credits <= IBLND_MSG_QUEUE_SIZE(ver));
801 if (conn->ibc_nsends_posted == IBLND_CONCURRENT_SENDS(ver)) {
802 /* tx completions outstanding... */
803 CDEBUG(D_NET, "%s: posted enough\n",
804 libcfs_nid2str(peer->ibp_nid));
808 if (credit != 0 && conn->ibc_credits == 0) { /* no credits */
809 CDEBUG(D_NET, "%s: no credits\n",
810 libcfs_nid2str(peer->ibp_nid));
814 if (credit != 0 && !IBLND_OOB_CAPABLE(ver) &&
815 conn->ibc_credits == 1 && /* last credit reserved for */
816 conn->ibc_outstanding_credits == 0) { /* giving back credits */
817 CDEBUG(D_NET, "%s: not using last credit\n",
818 libcfs_nid2str(peer->ibp_nid));
822 /* NB don't drop ibc_lock before bumping tx_sending */
823 cfs_list_del(&tx->tx_list);
826 if (msg->ibm_type == IBLND_MSG_NOOP &&
827 (!kiblnd_send_noop(conn) || /* redundant NOOP */
828 (IBLND_OOB_CAPABLE(ver) && /* posted enough NOOP */
829 conn->ibc_noops_posted == IBLND_OOB_MSGS(ver)))) {
830 /* OK to drop when posted enough NOOPs, since
831 * kiblnd_check_sends will queue NOOP again when
832 * posted NOOPs complete */
833 cfs_spin_unlock(&conn->ibc_lock);
834 kiblnd_tx_done(peer->ibp_ni, tx);
835 cfs_spin_lock(&conn->ibc_lock);
836 CDEBUG(D_NET, "%s(%d): redundant or enough NOOP\n",
837 libcfs_nid2str(peer->ibp_nid),
838 conn->ibc_noops_posted);
842 kiblnd_pack_msg(peer->ibp_ni, msg, ver, conn->ibc_outstanding_credits,
843 peer->ibp_nid, conn->ibc_incarnation);
845 conn->ibc_credits -= credit;
846 conn->ibc_outstanding_credits = 0;
847 conn->ibc_nsends_posted++;
848 if (msg->ibm_type == IBLND_MSG_NOOP)
849 conn->ibc_noops_posted++;
851 /* CAVEAT EMPTOR! This tx could be the PUT_DONE of an RDMA
852 * PUT. If so, it was first queued here as a PUT_REQ, sent and
853 * stashed on ibc_active_txs, matched by an incoming PUT_ACK,
854 * and then re-queued here. It's (just) possible that
855 * tx_sending is non-zero if we've not done the tx_complete()
856 * from the first send; hence the ++ rather than = below. */
858 cfs_list_add(&tx->tx_list, &conn->ibc_active_txs);
860 /* I'm still holding ibc_lock! */
861 if (conn->ibc_state != IBLND_CONN_ESTABLISHED)
864 rc = ib_post_send(conn->ibc_cmid->qp,
865 tx->tx_wrq, &bad_wrq);
866 conn->ibc_last_send = jiffies;
871 /* NB credits are transferred in the actual
872 * message, which can only be the last work item */
873 conn->ibc_credits += credit;
874 conn->ibc_outstanding_credits += msg->ibm_credits;
875 conn->ibc_nsends_posted--;
876 if (msg->ibm_type == IBLND_MSG_NOOP)
877 conn->ibc_noops_posted--;
883 done = (tx->tx_sending == 0);
885 cfs_list_del(&tx->tx_list);
887 cfs_spin_unlock(&conn->ibc_lock);
889 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
890 CERROR("Error %d posting transmit to %s\n",
891 rc, libcfs_nid2str(peer->ibp_nid));
893 CDEBUG(D_NET, "Error %d posting transmit to %s\n",
894 rc, libcfs_nid2str(peer->ibp_nid));
896 kiblnd_close_conn(conn, rc);
899 kiblnd_tx_done(peer->ibp_ni, tx);
901 cfs_spin_lock(&conn->ibc_lock);
907 kiblnd_check_sends (kib_conn_t *conn)
909 int ver = conn->ibc_version;
910 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
913 /* Don't send anything until after the connection is established */
914 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
915 CDEBUG(D_NET, "%s too soon\n",
916 libcfs_nid2str(conn->ibc_peer->ibp_nid));
920 cfs_spin_lock(&conn->ibc_lock);
922 LASSERT (conn->ibc_nsends_posted <= IBLND_CONCURRENT_SENDS(ver));
923 LASSERT (!IBLND_OOB_CAPABLE(ver) ||
924 conn->ibc_noops_posted <= IBLND_OOB_MSGS(ver));
925 LASSERT (conn->ibc_reserved_credits >= 0);
927 while (conn->ibc_reserved_credits > 0 &&
928 !cfs_list_empty(&conn->ibc_tx_queue_rsrvd)) {
929 tx = cfs_list_entry(conn->ibc_tx_queue_rsrvd.next,
931 cfs_list_del(&tx->tx_list);
932 cfs_list_add_tail(&tx->tx_list, &conn->ibc_tx_queue);
933 conn->ibc_reserved_credits--;
936 if (kiblnd_send_noop(conn)) {
937 cfs_spin_unlock(&conn->ibc_lock);
939 tx = kiblnd_get_idle_tx(ni);
941 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_NOOP, 0);
943 cfs_spin_lock(&conn->ibc_lock);
945 kiblnd_queue_tx_locked(tx, conn);
948 kiblnd_conn_addref(conn); /* 1 ref for me.... (see b21911) */
953 if (!cfs_list_empty(&conn->ibc_tx_queue_nocred)) {
955 tx = cfs_list_entry(conn->ibc_tx_queue_nocred.next,
957 } else if (!cfs_list_empty(&conn->ibc_tx_queue)) {
959 tx = cfs_list_entry(conn->ibc_tx_queue.next,
964 if (kiblnd_post_tx_locked(conn, tx, credit) != 0)
968 cfs_spin_unlock(&conn->ibc_lock);
970 kiblnd_conn_decref(conn); /* ...until here */
974 kiblnd_tx_complete (kib_tx_t *tx, int status)
976 int failed = (status != IB_WC_SUCCESS);
977 kib_conn_t *conn = tx->tx_conn;
980 LASSERT (tx->tx_sending > 0);
983 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
984 CDEBUG(D_NETERROR, "Tx -> %s cookie "LPX64
985 " sending %d waiting %d: failed %d\n",
986 libcfs_nid2str(conn->ibc_peer->ibp_nid),
987 tx->tx_cookie, tx->tx_sending, tx->tx_waiting,
990 kiblnd_close_conn(conn, -EIO);
992 kiblnd_peer_alive(conn->ibc_peer);
995 cfs_spin_lock(&conn->ibc_lock);
997 /* I could be racing with rdma completion. Whoever makes 'tx' idle
998 * gets to free it, which also drops its ref on 'conn'. */
1001 conn->ibc_nsends_posted--;
1002 if (tx->tx_msg->ibm_type == IBLND_MSG_NOOP)
1003 conn->ibc_noops_posted--;
1006 tx->tx_waiting = 0; /* don't wait for peer */
1007 tx->tx_status = -EIO;
1010 idle = (tx->tx_sending == 0) && /* This is the final callback */
1011 !tx->tx_waiting && /* Not waiting for peer */
1012 !tx->tx_queued; /* Not re-queued (PUT_DONE) */
1014 cfs_list_del(&tx->tx_list);
1016 kiblnd_conn_addref(conn); /* 1 ref for me.... */
1018 cfs_spin_unlock(&conn->ibc_lock);
1021 kiblnd_tx_done(conn->ibc_peer->ibp_ni, tx);
1023 kiblnd_check_sends(conn);
1025 kiblnd_conn_decref(conn); /* ...until here */
1029 kiblnd_init_tx_msg (lnet_ni_t *ni, kib_tx_t *tx, int type, int body_nob)
1031 kib_net_t *net = ni->ni_data;
1032 struct ib_sge *sge = &tx->tx_sge[tx->tx_nwrq];
1033 struct ib_send_wr *wrq = &tx->tx_wrq[tx->tx_nwrq];
1034 int nob = offsetof (kib_msg_t, ibm_u) + body_nob;
1037 LASSERT (net != NULL);
1038 LASSERT (tx->tx_nwrq >= 0);
1039 LASSERT (tx->tx_nwrq < IBLND_MAX_RDMA_FRAGS + 1);
1040 LASSERT (nob <= IBLND_MSG_SIZE);
1042 kiblnd_init_msg(tx->tx_msg, type, body_nob);
1044 mr = kiblnd_find_dma_mr(net, tx->tx_msgaddr, nob);
1045 LASSERT (mr != NULL);
1047 sge->lkey = mr->lkey;
1048 sge->addr = tx->tx_msgaddr;
1051 memset(wrq, 0, sizeof(*wrq));
1054 wrq->wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_TX);
1057 wrq->opcode = IB_WR_SEND;
1058 wrq->send_flags = IB_SEND_SIGNALED;
1064 kiblnd_init_rdma (kib_conn_t *conn, kib_tx_t *tx, int type,
1065 int resid, kib_rdma_desc_t *dstrd, __u64 dstcookie)
1067 kib_msg_t *ibmsg = tx->tx_msg;
1068 kib_rdma_desc_t *srcrd = tx->tx_rd;
1069 struct ib_sge *sge = &tx->tx_sge[0];
1070 struct ib_send_wr *wrq = &tx->tx_wrq[0];
1076 LASSERT (!cfs_in_interrupt());
1077 LASSERT (tx->tx_nwrq == 0);
1078 LASSERT (type == IBLND_MSG_GET_DONE ||
1079 type == IBLND_MSG_PUT_DONE);
1081 srcidx = dstidx = 0;
1084 if (srcidx >= srcrd->rd_nfrags) {
1085 CERROR("Src buffer exhausted: %d frags\n", srcidx);
1090 if (dstidx == dstrd->rd_nfrags) {
1091 CERROR("Dst buffer exhausted: %d frags\n", dstidx);
1096 if (tx->tx_nwrq == IBLND_RDMA_FRAGS(conn->ibc_version)) {
1097 CERROR("RDMA too fragmented for %s (%d): "
1098 "%d/%d src %d/%d dst frags\n",
1099 libcfs_nid2str(conn->ibc_peer->ibp_nid),
1100 IBLND_RDMA_FRAGS(conn->ibc_version),
1101 srcidx, srcrd->rd_nfrags,
1102 dstidx, dstrd->rd_nfrags);
1107 wrknob = MIN(MIN(kiblnd_rd_frag_size(srcrd, srcidx),
1108 kiblnd_rd_frag_size(dstrd, dstidx)), resid);
1110 sge = &tx->tx_sge[tx->tx_nwrq];
1111 sge->addr = kiblnd_rd_frag_addr(srcrd, srcidx);
1112 sge->lkey = kiblnd_rd_frag_key(srcrd, srcidx);
1113 sge->length = wrknob;
1115 wrq = &tx->tx_wrq[tx->tx_nwrq];
1117 wrq->next = wrq + 1;
1118 wrq->wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_RDMA);
1121 wrq->opcode = IB_WR_RDMA_WRITE;
1122 wrq->send_flags = 0;
1124 wrq->wr.rdma.remote_addr = kiblnd_rd_frag_addr(dstrd, dstidx);
1125 wrq->wr.rdma.rkey = kiblnd_rd_frag_key(dstrd, dstidx);
1127 srcidx = kiblnd_rd_consume_frag(srcrd, srcidx, wrknob);
1128 dstidx = kiblnd_rd_consume_frag(dstrd, dstidx, wrknob);
1137 if (rc < 0) /* no RDMA if completing with failure */
1140 ibmsg->ibm_u.completion.ibcm_status = rc;
1141 ibmsg->ibm_u.completion.ibcm_cookie = dstcookie;
1142 kiblnd_init_tx_msg(conn->ibc_peer->ibp_ni, tx,
1143 type, sizeof (kib_completion_msg_t));
1149 kiblnd_queue_tx_locked (kib_tx_t *tx, kib_conn_t *conn)
1153 LASSERT (tx->tx_nwrq > 0); /* work items set up */
1154 LASSERT (!tx->tx_queued); /* not queued for sending already */
1155 LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1158 tx->tx_deadline = jiffies + (*kiblnd_tunables.kib_timeout * CFS_HZ);
1160 if (tx->tx_conn == NULL) {
1161 kiblnd_conn_addref(conn);
1163 LASSERT (tx->tx_msg->ibm_type != IBLND_MSG_PUT_DONE);
1165 /* PUT_DONE first attached to conn as a PUT_REQ */
1166 LASSERT (tx->tx_conn == conn);
1167 LASSERT (tx->tx_msg->ibm_type == IBLND_MSG_PUT_DONE);
1170 switch (tx->tx_msg->ibm_type) {
1174 case IBLND_MSG_PUT_REQ:
1175 case IBLND_MSG_GET_REQ:
1176 q = &conn->ibc_tx_queue_rsrvd;
1179 case IBLND_MSG_PUT_NAK:
1180 case IBLND_MSG_PUT_ACK:
1181 case IBLND_MSG_PUT_DONE:
1182 case IBLND_MSG_GET_DONE:
1183 q = &conn->ibc_tx_queue_nocred;
1186 case IBLND_MSG_NOOP:
1187 if (IBLND_OOB_CAPABLE(conn->ibc_version))
1188 q = &conn->ibc_tx_queue_nocred;
1190 q = &conn->ibc_tx_queue;
1193 case IBLND_MSG_IMMEDIATE:
1194 q = &conn->ibc_tx_queue;
1198 cfs_list_add_tail(&tx->tx_list, q);
1202 kiblnd_queue_tx (kib_tx_t *tx, kib_conn_t *conn)
1204 cfs_spin_lock(&conn->ibc_lock);
1205 kiblnd_queue_tx_locked(tx, conn);
1206 cfs_spin_unlock(&conn->ibc_lock);
1208 kiblnd_check_sends(conn);
1212 kiblnd_connect_peer (kib_peer_t *peer)
1214 struct rdma_cm_id *cmid;
1216 kib_net_t *net = peer->ibp_ni->ni_data;
1217 struct sockaddr_in srcaddr;
1218 struct sockaddr_in dstaddr;
1221 LASSERT (net != NULL);
1222 LASSERT (peer->ibp_connecting > 0);
1224 cmid = rdma_create_id(kiblnd_cm_callback, peer, RDMA_PS_TCP);
1226 CERROR("Can't create CMID for %s: %ld\n",
1227 libcfs_nid2str(peer->ibp_nid), PTR_ERR(cmid));
1233 memset(&srcaddr, 0, sizeof(srcaddr));
1234 srcaddr.sin_family = AF_INET;
1235 srcaddr.sin_addr.s_addr = htonl(dev->ibd_ifip);
1237 memset(&dstaddr, 0, sizeof(dstaddr));
1238 dstaddr.sin_family = AF_INET;
1239 dstaddr.sin_port = htons(*kiblnd_tunables.kib_service);
1240 dstaddr.sin_addr.s_addr = htonl(LNET_NIDADDR(peer->ibp_nid));
1242 kiblnd_peer_addref(peer); /* cmid's ref */
1244 rc = rdma_resolve_addr(cmid,
1245 (struct sockaddr *)&srcaddr,
1246 (struct sockaddr *)&dstaddr,
1247 *kiblnd_tunables.kib_timeout * 1000);
1249 LASSERT (cmid->device != NULL);
1250 CDEBUG(D_NET, "%s: connection bound to %s:%u.%u.%u.%u:%s\n",
1251 libcfs_nid2str(peer->ibp_nid), dev->ibd_ifname,
1252 HIPQUAD(dev->ibd_ifip), cmid->device->name);
1256 /* Can't initiate address resolution: */
1257 CERROR("Can't resolve addr for %s: %d\n",
1258 libcfs_nid2str(peer->ibp_nid), rc);
1260 kiblnd_peer_decref(peer); /* cmid's ref */
1261 rdma_destroy_id(cmid);
1263 kiblnd_peer_connect_failed(peer, 1, rc);
1267 kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid)
1272 cfs_rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
1273 unsigned long flags;
1276 /* If I get here, I've committed to send, so I complete the tx with
1277 * failure on any problems */
1279 LASSERT (tx == NULL || tx->tx_conn == NULL); /* only set when assigned a conn */
1280 LASSERT (tx == NULL || tx->tx_nwrq > 0); /* work items have been set up */
1282 /* First time, just use a read lock since I expect to find my peer
1284 cfs_read_lock_irqsave(g_lock, flags);
1286 peer = kiblnd_find_peer_locked(nid);
1287 if (peer != NULL && !cfs_list_empty(&peer->ibp_conns)) {
1288 /* Found a peer with an established connection */
1289 conn = kiblnd_get_conn_locked(peer);
1290 kiblnd_conn_addref(conn); /* 1 ref for me... */
1292 cfs_read_unlock_irqrestore(g_lock, flags);
1295 kiblnd_queue_tx(tx, conn);
1296 kiblnd_conn_decref(conn); /* ...to here */
1300 cfs_read_unlock(g_lock);
1301 /* Re-try with a write lock */
1302 cfs_write_lock(g_lock);
1304 peer = kiblnd_find_peer_locked(nid);
1306 if (cfs_list_empty(&peer->ibp_conns)) {
1307 /* found a peer, but it's still connecting... */
1308 LASSERT (peer->ibp_connecting != 0 ||
1309 peer->ibp_accepting != 0);
1311 cfs_list_add_tail(&tx->tx_list,
1312 &peer->ibp_tx_queue);
1313 cfs_write_unlock_irqrestore(g_lock, flags);
1315 conn = kiblnd_get_conn_locked(peer);
1316 kiblnd_conn_addref(conn); /* 1 ref for me... */
1318 cfs_write_unlock_irqrestore(g_lock, flags);
1321 kiblnd_queue_tx(tx, conn);
1322 kiblnd_conn_decref(conn); /* ...to here */
1327 cfs_write_unlock_irqrestore(g_lock, flags);
1329 /* Allocate a peer ready to add to the peer table and retry */
1330 rc = kiblnd_create_peer(ni, &peer, nid);
1332 CERROR("Can't create peer %s\n", libcfs_nid2str(nid));
1334 tx->tx_status = -EHOSTUNREACH;
1336 kiblnd_tx_done(ni, tx);
1341 cfs_write_lock_irqsave(g_lock, flags);
1343 peer2 = kiblnd_find_peer_locked(nid);
1344 if (peer2 != NULL) {
1345 if (cfs_list_empty(&peer2->ibp_conns)) {
1346 /* found a peer, but it's still connecting... */
1347 LASSERT (peer2->ibp_connecting != 0 ||
1348 peer2->ibp_accepting != 0);
1350 cfs_list_add_tail(&tx->tx_list,
1351 &peer2->ibp_tx_queue);
1352 cfs_write_unlock_irqrestore(g_lock, flags);
1354 conn = kiblnd_get_conn_locked(peer2);
1355 kiblnd_conn_addref(conn); /* 1 ref for me... */
1357 cfs_write_unlock_irqrestore(g_lock, flags);
1360 kiblnd_queue_tx(tx, conn);
1361 kiblnd_conn_decref(conn); /* ...to here */
1364 kiblnd_peer_decref(peer);
1368 /* Brand new peer */
1369 LASSERT (peer->ibp_connecting == 0);
1370 peer->ibp_connecting = 1;
1372 /* always called with a ref on ni, which prevents ni being shutdown */
1373 LASSERT (((kib_net_t *)ni->ni_data)->ibn_shutdown == 0);
1376 cfs_list_add_tail(&tx->tx_list, &peer->ibp_tx_queue);
1378 kiblnd_peer_addref(peer);
1379 cfs_list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid));
1381 cfs_write_unlock_irqrestore(g_lock, flags);
1383 kiblnd_connect_peer(peer);
1384 kiblnd_peer_decref(peer);
1388 kiblnd_send (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg)
1390 lnet_hdr_t *hdr = &lntmsg->msg_hdr;
1391 int type = lntmsg->msg_type;
1392 lnet_process_id_t target = lntmsg->msg_target;
1393 int target_is_router = lntmsg->msg_target_is_router;
1394 int routing = lntmsg->msg_routing;
1395 unsigned int payload_niov = lntmsg->msg_niov;
1396 struct iovec *payload_iov = lntmsg->msg_iov;
1397 lnet_kiov_t *payload_kiov = lntmsg->msg_kiov;
1398 unsigned int payload_offset = lntmsg->msg_offset;
1399 unsigned int payload_nob = lntmsg->msg_len;
1405 /* NB 'private' is different depending on what we're sending.... */
1407 CDEBUG(D_NET, "sending %d bytes in %d frags to %s\n",
1408 payload_nob, payload_niov, libcfs_id2str(target));
1410 LASSERT (payload_nob == 0 || payload_niov > 0);
1411 LASSERT (payload_niov <= LNET_MAX_IOV);
1413 /* Thread context */
1414 LASSERT (!cfs_in_interrupt());
1415 /* payload is either all vaddrs or all pages */
1416 LASSERT (!(payload_kiov != NULL && payload_iov != NULL));
1424 LASSERT (payload_nob == 0);
1428 if (routing || target_is_router)
1429 break; /* send IMMEDIATE */
1431 /* is the REPLY message too small for RDMA? */
1432 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[lntmsg->msg_md->md_length]);
1433 if (nob <= IBLND_MSG_SIZE)
1434 break; /* send IMMEDIATE */
1436 tx = kiblnd_get_idle_tx(ni);
1438 CERROR("Can't allocate txd for GET to %s: \n",
1439 libcfs_nid2str(target.nid));
1445 if ((lntmsg->msg_md->md_options & LNET_MD_KIOV) == 0)
1446 rc = kiblnd_setup_rd_iov(ni, tx,
1447 &ibmsg->ibm_u.get.ibgm_rd,
1448 lntmsg->msg_md->md_niov,
1449 lntmsg->msg_md->md_iov.iov,
1450 0, lntmsg->msg_md->md_length);
1452 rc = kiblnd_setup_rd_kiov(ni, tx,
1453 &ibmsg->ibm_u.get.ibgm_rd,
1454 lntmsg->msg_md->md_niov,
1455 lntmsg->msg_md->md_iov.kiov,
1456 0, lntmsg->msg_md->md_length);
1458 CERROR("Can't setup GET sink for %s: %d\n",
1459 libcfs_nid2str(target.nid), rc);
1460 kiblnd_tx_done(ni, tx);
1464 nob = offsetof(kib_get_msg_t, ibgm_rd.rd_frags[tx->tx_nfrags]);
1465 ibmsg->ibm_u.get.ibgm_cookie = tx->tx_cookie;
1466 ibmsg->ibm_u.get.ibgm_hdr = *hdr;
1468 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_GET_REQ, nob);
1470 tx->tx_lntmsg[1] = lnet_create_reply_msg(ni, lntmsg);
1471 if (tx->tx_lntmsg[1] == NULL) {
1472 CERROR("Can't create reply for GET -> %s\n",
1473 libcfs_nid2str(target.nid));
1474 kiblnd_tx_done(ni, tx);
1478 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg[0,1] on completion */
1479 tx->tx_waiting = 1; /* waiting for GET_DONE */
1480 kiblnd_launch_tx(ni, tx, target.nid);
1483 case LNET_MSG_REPLY:
1485 /* Is the payload small enough not to need RDMA? */
1486 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob]);
1487 if (nob <= IBLND_MSG_SIZE)
1488 break; /* send IMMEDIATE */
1490 tx = kiblnd_get_idle_tx(ni);
1492 CERROR("Can't allocate %s txd for %s\n",
1493 type == LNET_MSG_PUT ? "PUT" : "REPLY",
1494 libcfs_nid2str(target.nid));
1498 if (payload_kiov == NULL)
1499 rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1500 payload_niov, payload_iov,
1501 payload_offset, payload_nob);
1503 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1504 payload_niov, payload_kiov,
1505 payload_offset, payload_nob);
1507 CERROR("Can't setup PUT src for %s: %d\n",
1508 libcfs_nid2str(target.nid), rc);
1509 kiblnd_tx_done(ni, tx);
1514 ibmsg->ibm_u.putreq.ibprm_hdr = *hdr;
1515 ibmsg->ibm_u.putreq.ibprm_cookie = tx->tx_cookie;
1516 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_REQ, sizeof(kib_putreq_msg_t));
1518 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1519 tx->tx_waiting = 1; /* waiting for PUT_{ACK,NAK} */
1520 kiblnd_launch_tx(ni, tx, target.nid);
1524 /* send IMMEDIATE */
1526 LASSERT (offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob])
1529 tx = kiblnd_get_idle_tx(ni);
1531 CERROR ("Can't send %d to %s: tx descs exhausted\n",
1532 type, libcfs_nid2str(target.nid));
1537 ibmsg->ibm_u.immediate.ibim_hdr = *hdr;
1539 if (payload_kiov != NULL)
1540 lnet_copy_kiov2flat(IBLND_MSG_SIZE, ibmsg,
1541 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1542 payload_niov, payload_kiov,
1543 payload_offset, payload_nob);
1545 lnet_copy_iov2flat(IBLND_MSG_SIZE, ibmsg,
1546 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1547 payload_niov, payload_iov,
1548 payload_offset, payload_nob);
1550 nob = offsetof(kib_immediate_msg_t, ibim_payload[payload_nob]);
1551 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_IMMEDIATE, nob);
1553 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1554 kiblnd_launch_tx(ni, tx, target.nid);
1559 kiblnd_reply (lnet_ni_t *ni, kib_rx_t *rx, lnet_msg_t *lntmsg)
1561 lnet_process_id_t target = lntmsg->msg_target;
1562 unsigned int niov = lntmsg->msg_niov;
1563 struct iovec *iov = lntmsg->msg_iov;
1564 lnet_kiov_t *kiov = lntmsg->msg_kiov;
1565 unsigned int offset = lntmsg->msg_offset;
1566 unsigned int nob = lntmsg->msg_len;
1570 tx = kiblnd_get_idle_tx(ni);
1572 CERROR("Can't get tx for REPLY to %s\n",
1573 libcfs_nid2str(target.nid));
1579 else if (kiov == NULL)
1580 rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1581 niov, iov, offset, nob);
1583 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1584 niov, kiov, offset, nob);
1587 CERROR("Can't setup GET src for %s: %d\n",
1588 libcfs_nid2str(target.nid), rc);
1592 rc = kiblnd_init_rdma(rx->rx_conn, tx,
1593 IBLND_MSG_GET_DONE, nob,
1594 &rx->rx_msg->ibm_u.get.ibgm_rd,
1595 rx->rx_msg->ibm_u.get.ibgm_cookie);
1597 CERROR("Can't setup rdma for GET from %s: %d\n",
1598 libcfs_nid2str(target.nid), rc);
1603 /* No RDMA: local completion may happen now! */
1604 lnet_finalize(ni, lntmsg, 0);
1606 /* RDMA: lnet_finalize(lntmsg) when it
1608 tx->tx_lntmsg[0] = lntmsg;
1611 kiblnd_queue_tx(tx, rx->rx_conn);
1615 kiblnd_tx_done(ni, tx);
1617 lnet_finalize(ni, lntmsg, -EIO);
1621 kiblnd_recv (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed,
1622 unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov,
1623 unsigned int offset, unsigned int mlen, unsigned int rlen)
1625 kib_rx_t *rx = private;
1626 kib_msg_t *rxmsg = rx->rx_msg;
1627 kib_conn_t *conn = rx->rx_conn;
1631 int post_credit = IBLND_POSTRX_PEER_CREDIT;
1634 LASSERT (mlen <= rlen);
1635 LASSERT (!cfs_in_interrupt());
1636 /* Either all pages or all vaddrs */
1637 LASSERT (!(kiov != NULL && iov != NULL));
1639 switch (rxmsg->ibm_type) {
1643 case IBLND_MSG_IMMEDIATE:
1644 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[rlen]);
1645 if (nob > rx->rx_nob) {
1646 CERROR ("Immediate message from %s too big: %d(%d)\n",
1647 libcfs_nid2str(rxmsg->ibm_u.immediate.ibim_hdr.src_nid),
1654 lnet_copy_flat2kiov(niov, kiov, offset,
1655 IBLND_MSG_SIZE, rxmsg,
1656 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1659 lnet_copy_flat2iov(niov, iov, offset,
1660 IBLND_MSG_SIZE, rxmsg,
1661 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1663 lnet_finalize (ni, lntmsg, 0);
1666 case IBLND_MSG_PUT_REQ:
1668 lnet_finalize(ni, lntmsg, 0);
1669 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK, 0,
1670 rxmsg->ibm_u.putreq.ibprm_cookie);
1674 tx = kiblnd_get_idle_tx(ni);
1676 CERROR("Can't allocate tx for %s\n",
1677 libcfs_nid2str(conn->ibc_peer->ibp_nid));
1678 /* Not replying will break the connection */
1685 rc = kiblnd_setup_rd_iov(ni, tx,
1686 &txmsg->ibm_u.putack.ibpam_rd,
1687 niov, iov, offset, mlen);
1689 rc = kiblnd_setup_rd_kiov(ni, tx,
1690 &txmsg->ibm_u.putack.ibpam_rd,
1691 niov, kiov, offset, mlen);
1693 CERROR("Can't setup PUT sink for %s: %d\n",
1694 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
1695 kiblnd_tx_done(ni, tx);
1696 /* tell peer it's over */
1697 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK, rc,
1698 rxmsg->ibm_u.putreq.ibprm_cookie);
1702 nob = offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[tx->tx_nfrags]);
1703 txmsg->ibm_u.putack.ibpam_src_cookie = rxmsg->ibm_u.putreq.ibprm_cookie;
1704 txmsg->ibm_u.putack.ibpam_dst_cookie = tx->tx_cookie;
1706 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_ACK, nob);
1708 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1709 tx->tx_waiting = 1; /* waiting for PUT_DONE */
1710 kiblnd_queue_tx(tx, conn);
1712 /* reposted buffer reserved for PUT_DONE */
1713 post_credit = IBLND_POSTRX_NO_CREDIT;
1716 case IBLND_MSG_GET_REQ:
1717 if (lntmsg != NULL) {
1718 /* Optimized GET; RDMA lntmsg's payload */
1719 kiblnd_reply(ni, rx, lntmsg);
1721 /* GET didn't match anything */
1722 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_GET_DONE,
1724 rxmsg->ibm_u.get.ibgm_cookie);
1729 kiblnd_post_rx(rx, post_credit);
1734 kiblnd_thread_start (int (*fn)(void *arg), void *arg)
1736 long pid = cfs_kernel_thread (fn, arg, 0);
1741 cfs_atomic_inc (&kiblnd_data.kib_nthreads);
1746 kiblnd_thread_fini (void)
1748 cfs_atomic_dec (&kiblnd_data.kib_nthreads);
1752 kiblnd_peer_alive (kib_peer_t *peer)
1754 /* This is racy, but everyone's only writing cfs_time_current() */
1755 peer->ibp_last_alive = cfs_time_current();
1760 kiblnd_peer_notify (kib_peer_t *peer)
1763 cfs_time_t last_alive = 0;
1764 unsigned long flags;
1766 cfs_read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1768 if (cfs_list_empty(&peer->ibp_conns) &&
1769 peer->ibp_accepting == 0 &&
1770 peer->ibp_connecting == 0 &&
1771 peer->ibp_error != 0) {
1772 error = peer->ibp_error;
1773 peer->ibp_error = 0;
1775 last_alive = peer->ibp_last_alive;
1778 cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1781 lnet_notify(peer->ibp_ni,
1782 peer->ibp_nid, 0, last_alive);
1786 kiblnd_close_conn_locked (kib_conn_t *conn, int error)
1788 /* This just does the immediate housekeeping. 'error' is zero for a
1789 * normal shutdown which can happen only after the connection has been
1790 * established. If the connection is established, schedule the
1791 * connection to be finished off by the connd. Otherwise the connd is
1792 * already dealing with it (either to set it up or tear it down).
1793 * Caller holds kib_global_lock exclusively in irq context */
1794 unsigned long flags;
1795 kib_peer_t *peer = conn->ibc_peer;
1797 LASSERT (error != 0 || conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1799 if (error != 0 && conn->ibc_comms_error == 0)
1800 conn->ibc_comms_error = error;
1802 if (conn->ibc_state != IBLND_CONN_ESTABLISHED)
1803 return; /* already being handled */
1806 cfs_list_empty(&conn->ibc_tx_queue) &&
1807 cfs_list_empty(&conn->ibc_tx_queue_rsrvd) &&
1808 cfs_list_empty(&conn->ibc_tx_queue_nocred) &&
1809 cfs_list_empty(&conn->ibc_active_txs)) {
1810 CDEBUG(D_NET, "closing conn to %s\n",
1811 libcfs_nid2str(peer->ibp_nid));
1813 CDEBUG(D_NETERROR, "Closing conn to %s: error %d%s%s%s%s\n",
1814 libcfs_nid2str(peer->ibp_nid), error,
1815 cfs_list_empty(&conn->ibc_tx_queue) ? "" : "(sending)",
1816 cfs_list_empty(&conn->ibc_tx_queue_rsrvd) ? "" : "(sending_rsrvd)",
1817 cfs_list_empty(&conn->ibc_tx_queue_nocred) ? "" : "(sending_nocred)",
1818 cfs_list_empty(&conn->ibc_active_txs) ? "" : "(waiting)");
1821 cfs_list_del(&conn->ibc_list);
1822 /* connd (see below) takes over ibc_list's ref */
1824 if (cfs_list_empty (&peer->ibp_conns) && /* no more conns */
1825 kiblnd_peer_active(peer)) { /* still in peer table */
1826 kiblnd_unlink_peer_locked(peer);
1828 /* set/clear error on last conn */
1829 peer->ibp_error = conn->ibc_comms_error;
1832 kiblnd_set_conn_state(conn, IBLND_CONN_CLOSING);
1834 cfs_spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
1836 cfs_list_add_tail (&conn->ibc_list, &kiblnd_data.kib_connd_conns);
1837 cfs_waitq_signal (&kiblnd_data.kib_connd_waitq);
1839 cfs_spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
1843 kiblnd_close_conn (kib_conn_t *conn, int error)
1845 unsigned long flags;
1847 cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1849 kiblnd_close_conn_locked(conn, error);
1851 cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1855 kiblnd_handle_early_rxs(kib_conn_t *conn)
1857 unsigned long flags;
1860 LASSERT (!cfs_in_interrupt());
1861 LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1863 cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1864 while (!cfs_list_empty(&conn->ibc_early_rxs)) {
1865 rx = cfs_list_entry(conn->ibc_early_rxs.next,
1867 cfs_list_del(&rx->rx_list);
1868 cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock,
1871 kiblnd_handle_rx(rx);
1873 cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1875 cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1879 kiblnd_abort_txs(kib_conn_t *conn, cfs_list_t *txs)
1881 CFS_LIST_HEAD (zombies);
1886 cfs_spin_lock(&conn->ibc_lock);
1888 cfs_list_for_each_safe (tmp, nxt, txs) {
1889 tx = cfs_list_entry (tmp, kib_tx_t, tx_list);
1891 if (txs == &conn->ibc_active_txs) {
1892 LASSERT (!tx->tx_queued);
1893 LASSERT (tx->tx_waiting ||
1894 tx->tx_sending != 0);
1896 LASSERT (tx->tx_queued);
1899 tx->tx_status = -ECONNABORTED;
1902 if (tx->tx_sending == 0) {
1904 cfs_list_del (&tx->tx_list);
1905 cfs_list_add (&tx->tx_list, &zombies);
1909 cfs_spin_unlock(&conn->ibc_lock);
1911 kiblnd_txlist_done(conn->ibc_peer->ibp_ni,
1912 &zombies, -ECONNABORTED);
1916 kiblnd_finalise_conn (kib_conn_t *conn)
1918 LASSERT (!cfs_in_interrupt());
1919 LASSERT (conn->ibc_state > IBLND_CONN_INIT);
1921 kiblnd_set_conn_state(conn, IBLND_CONN_DISCONNECTED);
1923 /* abort_receives moves QP state to IB_QPS_ERR. This is only required
1924 * for connections that didn't get as far as being connected, because
1925 * rdma_disconnect() does this for free. */
1926 kiblnd_abort_receives(conn);
1928 /* Complete all tx descs not waiting for sends to complete.
1929 * NB we should be safe from RDMA now that the QP has changed state */
1931 kiblnd_abort_txs(conn, &conn->ibc_tx_queue);
1932 kiblnd_abort_txs(conn, &conn->ibc_tx_queue_rsrvd);
1933 kiblnd_abort_txs(conn, &conn->ibc_tx_queue_nocred);
1934 kiblnd_abort_txs(conn, &conn->ibc_active_txs);
1936 kiblnd_handle_early_rxs(conn);
1940 kiblnd_peer_connect_failed (kib_peer_t *peer, int active, int error)
1942 CFS_LIST_HEAD (zombies);
1943 unsigned long flags;
1945 LASSERT (error != 0);
1946 LASSERT (!cfs_in_interrupt());
1948 cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1951 LASSERT (peer->ibp_connecting > 0);
1952 peer->ibp_connecting--;
1954 LASSERT (peer->ibp_accepting > 0);
1955 peer->ibp_accepting--;
1958 if (peer->ibp_connecting != 0 ||
1959 peer->ibp_accepting != 0) {
1960 /* another connection attempt under way... */
1961 cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock,
1966 if (cfs_list_empty(&peer->ibp_conns)) {
1967 /* Take peer's blocked transmits to complete with error */
1968 cfs_list_add(&zombies, &peer->ibp_tx_queue);
1969 cfs_list_del_init(&peer->ibp_tx_queue);
1971 if (kiblnd_peer_active(peer))
1972 kiblnd_unlink_peer_locked(peer);
1974 peer->ibp_error = error;
1976 /* Can't have blocked transmits if there are connections */
1977 LASSERT (cfs_list_empty(&peer->ibp_tx_queue));
1980 cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1982 kiblnd_peer_notify(peer);
1984 if (cfs_list_empty (&zombies))
1987 CDEBUG (D_NETERROR, "Deleting messages for %s: connection failed\n",
1988 libcfs_nid2str(peer->ibp_nid));
1990 kiblnd_txlist_done(peer->ibp_ni, &zombies, -EHOSTUNREACH);
1994 kiblnd_connreq_done(kib_conn_t *conn, int status)
1996 kib_peer_t *peer = conn->ibc_peer;
1999 unsigned long flags;
2002 active = (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2004 CDEBUG(D_NET,"%s: active(%d), version(%x), status(%d)\n",
2005 libcfs_nid2str(peer->ibp_nid), active,
2006 conn->ibc_version, status);
2008 LASSERT (!cfs_in_interrupt());
2009 LASSERT ((conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT &&
2010 peer->ibp_connecting > 0) ||
2011 (conn->ibc_state == IBLND_CONN_PASSIVE_WAIT &&
2012 peer->ibp_accepting > 0));
2014 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
2015 conn->ibc_connvars = NULL;
2018 /* failed to establish connection */
2019 kiblnd_peer_connect_failed(peer, active, status);
2020 kiblnd_finalise_conn(conn);
2024 /* connection established */
2025 cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2027 conn->ibc_last_send = jiffies;
2028 kiblnd_set_conn_state(conn, IBLND_CONN_ESTABLISHED);
2029 kiblnd_peer_alive(peer);
2031 /* Add conn to peer's list and nuke any dangling conns from a different
2032 * peer instance... */
2033 kiblnd_conn_addref(conn); /* +1 ref for ibc_list */
2034 cfs_list_add(&conn->ibc_list, &peer->ibp_conns);
2036 peer->ibp_connecting--;
2038 peer->ibp_accepting--;
2040 if (peer->ibp_version == 0) {
2041 peer->ibp_version = conn->ibc_version;
2042 peer->ibp_incarnation = conn->ibc_incarnation;
2045 if (peer->ibp_version != conn->ibc_version ||
2046 peer->ibp_incarnation != conn->ibc_incarnation) {
2047 kiblnd_close_stale_conns_locked(peer, conn->ibc_version,
2048 conn->ibc_incarnation);
2049 peer->ibp_version = conn->ibc_version;
2050 peer->ibp_incarnation = conn->ibc_incarnation;
2053 /* grab pending txs while I have the lock */
2054 cfs_list_add(&txs, &peer->ibp_tx_queue);
2055 cfs_list_del_init(&peer->ibp_tx_queue);
2057 if (!kiblnd_peer_active(peer) || /* peer has been deleted */
2058 conn->ibc_comms_error != 0) { /* error has happened already */
2059 lnet_ni_t *ni = peer->ibp_ni;
2061 /* start to shut down connection */
2062 kiblnd_close_conn_locked(conn, -ECONNABORTED);
2063 cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2066 kiblnd_txlist_done(ni, &txs, -ECONNABORTED);
2071 cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2073 /* Schedule blocked txs */
2074 cfs_spin_lock (&conn->ibc_lock);
2075 while (!cfs_list_empty (&txs)) {
2076 tx = cfs_list_entry (txs.next, kib_tx_t, tx_list);
2077 cfs_list_del(&tx->tx_list);
2079 kiblnd_queue_tx_locked(tx, conn);
2081 cfs_spin_unlock (&conn->ibc_lock);
2083 kiblnd_check_sends(conn);
2085 /* schedule blocked rxs */
2086 kiblnd_handle_early_rxs(conn);
2090 kiblnd_reject(struct rdma_cm_id *cmid, kib_rej_t *rej)
2094 rc = rdma_reject(cmid, rej, sizeof(*rej));
2097 CWARN("Error %d sending reject\n", rc);
2101 kiblnd_passive_connect (struct rdma_cm_id *cmid, void *priv, int priv_nob)
2103 cfs_rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2104 kib_msg_t *reqmsg = priv;
2110 lnet_ni_t *ni = NULL;
2111 kib_net_t *net = NULL;
2113 struct rdma_conn_param cp;
2115 int version = IBLND_MSG_VERSION;
2116 unsigned long flags;
2119 LASSERT (!cfs_in_interrupt());
2121 /* cmid inherits 'context' from the corresponding listener id */
2122 ibdev = (kib_dev_t *)cmid->context;
2123 LASSERT (ibdev != NULL);
2125 memset(&rej, 0, sizeof(rej));
2126 rej.ibr_magic = IBLND_MSG_MAGIC;
2127 rej.ibr_why = IBLND_REJECT_FATAL;
2128 rej.ibr_cp.ibcp_max_msg_size = IBLND_MSG_SIZE;
2130 if (priv_nob < offsetof(kib_msg_t, ibm_type)) {
2131 CERROR("Short connection request\n");
2135 /* Future protocol version compatibility support! If the
2136 * o2iblnd-specific protocol changes, or when LNET unifies
2137 * protocols over all LNDs, the initial connection will
2138 * negotiate a protocol version. I trap this here to avoid
2139 * console errors; the reject tells the peer which protocol I
2141 if (reqmsg->ibm_magic == LNET_PROTO_MAGIC ||
2142 reqmsg->ibm_magic == __swab32(LNET_PROTO_MAGIC))
2144 if (reqmsg->ibm_magic == IBLND_MSG_MAGIC &&
2145 reqmsg->ibm_version != IBLND_MSG_VERSION &&
2146 reqmsg->ibm_version != IBLND_MSG_VERSION_1)
2148 if (reqmsg->ibm_magic == __swab32(IBLND_MSG_MAGIC) &&
2149 reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION) &&
2150 reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION_1))
2153 rc = kiblnd_unpack_msg(reqmsg, priv_nob);
2155 CERROR("Can't parse connection request: %d\n", rc);
2159 nid = reqmsg->ibm_srcnid;
2160 ni = lnet_net2ni(LNET_NIDNET(reqmsg->ibm_dstnid));
2163 net = (kib_net_t *)ni->ni_data;
2164 rej.ibr_incarnation = net->ibn_incarnation;
2167 if (ni == NULL || /* no matching net */
2168 ni->ni_nid != reqmsg->ibm_dstnid || /* right NET, wrong NID! */
2169 net->ibn_dev != ibdev) { /* wrong device */
2170 CERROR("Can't accept %s on %s (%s:%d:%u.%u.%u.%u): "
2171 "bad dst nid %s\n", libcfs_nid2str(nid),
2172 ni == NULL ? "NA" : libcfs_nid2str(ni->ni_nid),
2173 ibdev->ibd_ifname, ibdev->ibd_nnets,
2174 HIPQUAD(ibdev->ibd_ifip),
2175 libcfs_nid2str(reqmsg->ibm_dstnid));
2180 /* check time stamp as soon as possible */
2181 if (reqmsg->ibm_dststamp != 0 &&
2182 reqmsg->ibm_dststamp != net->ibn_incarnation) {
2183 CWARN("Stale connection request\n");
2184 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2188 /* I can accept peer's version */
2189 version = reqmsg->ibm_version;
2191 if (reqmsg->ibm_type != IBLND_MSG_CONNREQ) {
2192 CERROR("Unexpected connreq msg type: %x from %s\n",
2193 reqmsg->ibm_type, libcfs_nid2str(nid));
2197 if (reqmsg->ibm_u.connparams.ibcp_queue_depth !=
2198 IBLND_MSG_QUEUE_SIZE(version)) {
2199 CERROR("Can't accept %s: incompatible queue depth %d (%d wanted)\n",
2200 libcfs_nid2str(nid), reqmsg->ibm_u.connparams.ibcp_queue_depth,
2201 IBLND_MSG_QUEUE_SIZE(version));
2203 if (version == IBLND_MSG_VERSION)
2204 rej.ibr_why = IBLND_REJECT_MSG_QUEUE_SIZE;
2209 if (reqmsg->ibm_u.connparams.ibcp_max_frags !=
2210 IBLND_RDMA_FRAGS(version)) {
2211 CERROR("Can't accept %s(version %x): "
2212 "incompatible max_frags %d (%d wanted)\n",
2213 libcfs_nid2str(nid), version,
2214 reqmsg->ibm_u.connparams.ibcp_max_frags,
2215 IBLND_RDMA_FRAGS(version));
2217 if (version == IBLND_MSG_VERSION)
2218 rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2224 if (reqmsg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2225 CERROR("Can't accept %s: message size %d too big (%d max)\n",
2226 libcfs_nid2str(nid),
2227 reqmsg->ibm_u.connparams.ibcp_max_msg_size,
2232 /* assume 'nid' is a new peer; create */
2233 rc = kiblnd_create_peer(ni, &peer, nid);
2235 CERROR("Can't create peer for %s\n", libcfs_nid2str(nid));
2236 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2240 cfs_write_lock_irqsave(g_lock, flags);
2242 peer2 = kiblnd_find_peer_locked(nid);
2243 if (peer2 != NULL) {
2244 if (peer2->ibp_version == 0) {
2245 peer2->ibp_version = version;
2246 peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2249 /* not the guy I've talked with */
2250 if (peer2->ibp_incarnation != reqmsg->ibm_srcstamp ||
2251 peer2->ibp_version != version) {
2252 kiblnd_close_peer_conns_locked(peer2, -ESTALE);
2253 cfs_write_unlock_irqrestore(g_lock, flags);
2255 CWARN("Conn stale %s [old ver: %x, new ver: %x]\n",
2256 libcfs_nid2str(nid), peer2->ibp_version, version);
2258 kiblnd_peer_decref(peer);
2259 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2263 /* tie-break connection race in favour of the higher NID */
2264 if (peer2->ibp_connecting != 0 &&
2266 cfs_write_unlock_irqrestore(g_lock, flags);
2268 CWARN("Conn race %s\n", libcfs_nid2str(peer2->ibp_nid));
2270 kiblnd_peer_decref(peer);
2271 rej.ibr_why = IBLND_REJECT_CONN_RACE;
2275 peer2->ibp_accepting++;
2276 kiblnd_peer_addref(peer2);
2278 cfs_write_unlock_irqrestore(g_lock, flags);
2279 kiblnd_peer_decref(peer);
2282 /* Brand new peer */
2283 LASSERT (peer->ibp_accepting == 0);
2284 LASSERT (peer->ibp_version == 0 &&
2285 peer->ibp_incarnation == 0);
2287 peer->ibp_accepting = 1;
2288 peer->ibp_version = version;
2289 peer->ibp_incarnation = reqmsg->ibm_srcstamp;
2291 /* I have a ref on ni that prevents it being shutdown */
2292 LASSERT (net->ibn_shutdown == 0);
2294 kiblnd_peer_addref(peer);
2295 cfs_list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid));
2297 cfs_write_unlock_irqrestore(g_lock, flags);
2300 conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_PASSIVE_WAIT, version);
2302 kiblnd_peer_connect_failed(peer, 0, -ENOMEM);
2303 kiblnd_peer_decref(peer);
2304 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2308 /* conn now "owns" cmid, so I return success from here on to ensure the
2309 * CM callback doesn't destroy cmid. */
2311 conn->ibc_incarnation = reqmsg->ibm_srcstamp;
2312 conn->ibc_credits = IBLND_MSG_QUEUE_SIZE(version);
2313 conn->ibc_reserved_credits = IBLND_MSG_QUEUE_SIZE(version);
2314 LASSERT (conn->ibc_credits + conn->ibc_reserved_credits + IBLND_OOB_MSGS(version)
2315 <= IBLND_RX_MSGS(version));
2317 ackmsg = &conn->ibc_connvars->cv_msg;
2318 memset(ackmsg, 0, sizeof(*ackmsg));
2320 kiblnd_init_msg(ackmsg, IBLND_MSG_CONNACK,
2321 sizeof(ackmsg->ibm_u.connparams));
2322 ackmsg->ibm_u.connparams.ibcp_queue_depth = IBLND_MSG_QUEUE_SIZE(version);
2323 ackmsg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2324 ackmsg->ibm_u.connparams.ibcp_max_frags = IBLND_RDMA_FRAGS(version);
2326 kiblnd_pack_msg(ni, ackmsg, version, 0, nid, reqmsg->ibm_srcstamp);
2328 memset(&cp, 0, sizeof(cp));
2329 cp.private_data = ackmsg;
2330 cp.private_data_len = ackmsg->ibm_nob;
2331 cp.responder_resources = 0; /* No atomic ops or RDMA reads */
2332 cp.initiator_depth = 0;
2333 cp.flow_control = 1;
2334 cp.retry_count = *kiblnd_tunables.kib_retry_count;
2335 cp.rnr_retry_count = *kiblnd_tunables.kib_rnr_retry_count;
2337 CDEBUG(D_NET, "Accept %s\n", libcfs_nid2str(nid));
2339 rc = rdma_accept(cmid, &cp);
2341 CERROR("Can't accept %s: %d\n", libcfs_nid2str(nid), rc);
2342 rej.ibr_version = version;
2343 rej.ibr_why = IBLND_REJECT_FATAL;
2345 kiblnd_reject(cmid, &rej);
2346 kiblnd_connreq_done(conn, rc);
2347 kiblnd_conn_decref(conn);
2357 rej.ibr_version = version;
2358 rej.ibr_cp.ibcp_queue_depth = IBLND_MSG_QUEUE_SIZE(version);
2359 rej.ibr_cp.ibcp_max_frags = IBLND_RDMA_FRAGS(version);
2360 kiblnd_reject(cmid, &rej);
2362 return -ECONNREFUSED;
2366 kiblnd_reconnect (kib_conn_t *conn, int version,
2367 __u64 incarnation, int why, kib_connparams_t *cp)
2369 kib_peer_t *peer = conn->ibc_peer;
2372 unsigned long flags;
2374 LASSERT (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2375 LASSERT (peer->ibp_connecting > 0); /* 'conn' at least */
2377 cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2379 /* retry connection if it's still needed and no other connection
2380 * attempts (active or passive) are in progress
2381 * NB: reconnect is still needed even when ibp_tx_queue is
2382 * empty if ibp_version != version because reconnect may be
2383 * initiated by kiblnd_query() */
2384 if ((!cfs_list_empty(&peer->ibp_tx_queue) ||
2385 peer->ibp_version != version) &&
2386 peer->ibp_connecting == 1 &&
2387 peer->ibp_accepting == 0) {
2389 peer->ibp_connecting++;
2391 peer->ibp_version = version;
2392 peer->ibp_incarnation = incarnation;
2395 cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2405 case IBLND_REJECT_CONN_STALE:
2409 case IBLND_REJECT_CONN_RACE:
2410 reason = "conn race";
2413 case IBLND_REJECT_CONN_UNCOMPAT:
2414 reason = "version negotiation";
2418 CDEBUG(D_NETERROR, "%s: retrying (%s), %x, %x, "
2419 "queue_dep: %d, max_frag: %d, msg_size: %d\n",
2420 libcfs_nid2str(peer->ibp_nid),
2421 reason, IBLND_MSG_VERSION, version,
2422 cp != NULL ? cp->ibcp_queue_depth : IBLND_MSG_QUEUE_SIZE(version),
2423 cp != NULL ? cp->ibcp_max_frags : IBLND_RDMA_FRAGS(version),
2424 cp != NULL ? cp->ibcp_max_msg_size: IBLND_MSG_SIZE);
2426 kiblnd_connect_peer(peer);
2430 kiblnd_rejected (kib_conn_t *conn, int reason, void *priv, int priv_nob)
2432 kib_peer_t *peer = conn->ibc_peer;
2434 LASSERT (!cfs_in_interrupt());
2435 LASSERT (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2438 case IB_CM_REJ_STALE_CONN:
2439 kiblnd_reconnect(conn, IBLND_MSG_VERSION, 0,
2440 IBLND_REJECT_CONN_STALE, NULL);
2443 case IB_CM_REJ_INVALID_SERVICE_ID:
2444 CDEBUG(D_NETERROR, "%s rejected: no listener at %d\n",
2445 libcfs_nid2str(peer->ibp_nid),
2446 *kiblnd_tunables.kib_service);
2449 case IB_CM_REJ_CONSUMER_DEFINED:
2450 if (priv_nob >= offsetof(kib_rej_t, ibr_padding)) {
2451 kib_rej_t *rej = priv;
2452 kib_connparams_t *cp = NULL;
2454 __u64 incarnation = -1;
2456 /* NB. default incarnation is -1 because:
2457 * a) V1 will ignore dst incarnation in connreq.
2458 * b) V2 will provide incarnation while rejecting me,
2459 * -1 will be overwrote.
2461 * if I try to connect to a V1 peer with V2 protocol,
2462 * it rejected me then upgrade to V2, I have no idea
2463 * about the upgrading and try to reconnect with V1,
2464 * in this case upgraded V2 can find out I'm trying to
2465 * talk to the old guy and reject me(incarnation is -1).
2468 if (rej->ibr_magic == __swab32(IBLND_MSG_MAGIC) ||
2469 rej->ibr_magic == __swab32(LNET_PROTO_MAGIC)) {
2470 __swab32s(&rej->ibr_magic);
2471 __swab16s(&rej->ibr_version);
2475 if (priv_nob >= sizeof(kib_rej_t) &&
2476 rej->ibr_version > IBLND_MSG_VERSION_1) {
2477 /* priv_nob is always 148 in current version
2478 * of OFED, so we still need to check version.
2479 * (define of IB_CM_REJ_PRIVATE_DATA_SIZE) */
2483 __swab64s(&rej->ibr_incarnation);
2484 __swab16s(&cp->ibcp_queue_depth);
2485 __swab16s(&cp->ibcp_max_frags);
2486 __swab32s(&cp->ibcp_max_msg_size);
2489 incarnation = rej->ibr_incarnation;
2492 if (rej->ibr_magic != IBLND_MSG_MAGIC &&
2493 rej->ibr_magic != LNET_PROTO_MAGIC) {
2494 CERROR("%s rejected: consumer defined fatal error\n",
2495 libcfs_nid2str(peer->ibp_nid));
2499 if (rej->ibr_version != IBLND_MSG_VERSION &&
2500 rej->ibr_version != IBLND_MSG_VERSION_1) {
2501 CERROR("%s rejected: o2iblnd version %x error\n",
2502 libcfs_nid2str(peer->ibp_nid),
2507 if (rej->ibr_why == IBLND_REJECT_FATAL &&
2508 rej->ibr_version == IBLND_MSG_VERSION_1) {
2509 CDEBUG(D_NET, "rejected by old version peer %s: %x\n",
2510 libcfs_nid2str(peer->ibp_nid), rej->ibr_version);
2512 if (conn->ibc_version != IBLND_MSG_VERSION_1)
2513 rej->ibr_why = IBLND_REJECT_CONN_UNCOMPAT;
2516 switch (rej->ibr_why) {
2517 case IBLND_REJECT_CONN_RACE:
2518 case IBLND_REJECT_CONN_STALE:
2519 case IBLND_REJECT_CONN_UNCOMPAT:
2520 kiblnd_reconnect(conn, rej->ibr_version,
2521 incarnation, rej->ibr_why, cp);
2524 case IBLND_REJECT_MSG_QUEUE_SIZE:
2525 CERROR("%s rejected: incompatible message queue depth %d, %d\n",
2526 libcfs_nid2str(peer->ibp_nid), cp->ibcp_queue_depth,
2527 IBLND_MSG_QUEUE_SIZE(conn->ibc_version));
2530 case IBLND_REJECT_RDMA_FRAGS:
2531 CERROR("%s rejected: incompatible # of RDMA fragments %d, %d\n",
2532 libcfs_nid2str(peer->ibp_nid), cp->ibcp_max_frags,
2533 IBLND_RDMA_FRAGS(conn->ibc_version));
2536 case IBLND_REJECT_NO_RESOURCES:
2537 CERROR("%s rejected: o2iblnd no resources\n",
2538 libcfs_nid2str(peer->ibp_nid));
2541 case IBLND_REJECT_FATAL:
2542 CERROR("%s rejected: o2iblnd fatal error\n",
2543 libcfs_nid2str(peer->ibp_nid));
2547 CERROR("%s rejected: o2iblnd reason %d\n",
2548 libcfs_nid2str(peer->ibp_nid),
2556 CDEBUG(D_NETERROR, "%s rejected: reason %d, size %d\n",
2557 libcfs_nid2str(peer->ibp_nid), reason, priv_nob);
2561 kiblnd_connreq_done(conn, -ECONNREFUSED);
2565 kiblnd_check_connreply (kib_conn_t *conn, void *priv, int priv_nob)
2567 kib_peer_t *peer = conn->ibc_peer;
2568 lnet_ni_t *ni = peer->ibp_ni;
2569 kib_net_t *net = ni->ni_data;
2570 kib_msg_t *msg = priv;
2571 int ver = conn->ibc_version;
2572 int rc = kiblnd_unpack_msg(msg, priv_nob);
2573 unsigned long flags;
2575 LASSERT (net != NULL);
2578 CERROR("Can't unpack connack from %s: %d\n",
2579 libcfs_nid2str(peer->ibp_nid), rc);
2583 if (msg->ibm_type != IBLND_MSG_CONNACK) {
2584 CERROR("Unexpected message %d from %s\n",
2585 msg->ibm_type, libcfs_nid2str(peer->ibp_nid));
2590 if (ver != msg->ibm_version) {
2591 CERROR("%s replied version %x is different with "
2592 "requested version %x\n",
2593 libcfs_nid2str(peer->ibp_nid), msg->ibm_version, ver);
2598 if (msg->ibm_u.connparams.ibcp_queue_depth !=
2599 IBLND_MSG_QUEUE_SIZE(ver)) {
2600 CERROR("%s has incompatible queue depth %d(%d wanted)\n",
2601 libcfs_nid2str(peer->ibp_nid),
2602 msg->ibm_u.connparams.ibcp_queue_depth,
2603 IBLND_MSG_QUEUE_SIZE(ver));
2608 if (msg->ibm_u.connparams.ibcp_max_frags !=
2609 IBLND_RDMA_FRAGS(ver)) {
2610 CERROR("%s has incompatible max_frags %d (%d wanted)\n",
2611 libcfs_nid2str(peer->ibp_nid),
2612 msg->ibm_u.connparams.ibcp_max_frags,
2613 IBLND_RDMA_FRAGS(ver));
2618 if (msg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2619 CERROR("%s max message size %d too big (%d max)\n",
2620 libcfs_nid2str(peer->ibp_nid),
2621 msg->ibm_u.connparams.ibcp_max_msg_size,
2627 cfs_read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2628 if (msg->ibm_dstnid == ni->ni_nid &&
2629 msg->ibm_dststamp == net->ibn_incarnation)
2633 cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2636 CERROR("Bad connection reply from %s, rc = %d, "
2637 "version: %x max_frags: %d\n",
2638 libcfs_nid2str(peer->ibp_nid), rc,
2639 msg->ibm_version, msg->ibm_u.connparams.ibcp_max_frags);
2643 conn->ibc_incarnation = msg->ibm_srcstamp;
2645 conn->ibc_reserved_credits = IBLND_MSG_QUEUE_SIZE(ver);
2646 LASSERT (conn->ibc_credits + conn->ibc_reserved_credits + IBLND_OOB_MSGS(ver)
2647 <= IBLND_RX_MSGS(ver));
2649 kiblnd_connreq_done(conn, 0);
2653 /* NB My QP has already established itself, so I handle anything going
2654 * wrong here by setting ibc_comms_error.
2655 * kiblnd_connreq_done(0) moves the conn state to ESTABLISHED, but then
2656 * immediately tears it down. */
2659 conn->ibc_comms_error = rc;
2660 kiblnd_connreq_done(conn, 0);
2664 kiblnd_active_connect (struct rdma_cm_id *cmid)
2666 kib_peer_t *peer = (kib_peer_t *)cmid->context;
2669 struct rdma_conn_param cp;
2672 unsigned long flags;
2675 cfs_read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2677 incarnation = peer->ibp_incarnation;
2678 version = (peer->ibp_version == 0) ? IBLND_MSG_VERSION : peer->ibp_version;
2680 cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2682 conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_ACTIVE_CONNECT, version);
2684 kiblnd_peer_connect_failed(peer, 1, -ENOMEM);
2685 kiblnd_peer_decref(peer); /* lose cmid's ref */
2689 /* conn "owns" cmid now, so I return success from here on to ensure the
2690 * CM callback doesn't destroy cmid. conn also takes over cmid's ref
2693 msg = &conn->ibc_connvars->cv_msg;
2695 memset(msg, 0, sizeof(*msg));
2696 kiblnd_init_msg(msg, IBLND_MSG_CONNREQ, sizeof(msg->ibm_u.connparams));
2697 msg->ibm_u.connparams.ibcp_queue_depth = IBLND_MSG_QUEUE_SIZE(version);
2698 msg->ibm_u.connparams.ibcp_max_frags = IBLND_RDMA_FRAGS(version);
2699 msg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2701 kiblnd_pack_msg(peer->ibp_ni, msg, version,
2702 0, peer->ibp_nid, incarnation);
2704 memset(&cp, 0, sizeof(cp));
2705 cp.private_data = msg;
2706 cp.private_data_len = msg->ibm_nob;
2707 cp.responder_resources = 0; /* No atomic ops or RDMA reads */
2708 cp.initiator_depth = 0;
2709 cp.flow_control = 1;
2710 cp.retry_count = *kiblnd_tunables.kib_retry_count;
2711 cp.rnr_retry_count = *kiblnd_tunables.kib_rnr_retry_count;
2713 LASSERT(cmid->context == (void *)conn);
2714 LASSERT(conn->ibc_cmid == cmid);
2716 rc = rdma_connect(cmid, &cp);
2718 CERROR("Can't connect to %s: %d\n",
2719 libcfs_nid2str(peer->ibp_nid), rc);
2720 kiblnd_connreq_done(conn, rc);
2721 kiblnd_conn_decref(conn);
2728 kiblnd_cm_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2734 switch (event->event) {
2736 CERROR("Unexpected event: %d, status: %d\n",
2737 event->event, event->status);
2740 case RDMA_CM_EVENT_CONNECT_REQUEST:
2741 /* destroy cmid on failure */
2742 rc = kiblnd_passive_connect(cmid,
2743 (void *)KIBLND_CONN_PARAM(event),
2744 KIBLND_CONN_PARAM_LEN(event));
2745 CDEBUG(D_NET, "connreq: %d\n", rc);
2748 case RDMA_CM_EVENT_ADDR_ERROR:
2749 peer = (kib_peer_t *)cmid->context;
2750 CDEBUG(D_NETERROR, "%s: ADDR ERROR %d\n",
2751 libcfs_nid2str(peer->ibp_nid), event->status);
2752 kiblnd_peer_connect_failed(peer, 1, -EHOSTUNREACH);
2753 kiblnd_peer_decref(peer);
2754 return -EHOSTUNREACH; /* rc != 0 destroys cmid */
2756 case RDMA_CM_EVENT_ADDR_RESOLVED:
2757 peer = (kib_peer_t *)cmid->context;
2759 CDEBUG(D_NET,"%s Addr resolved: %d\n",
2760 libcfs_nid2str(peer->ibp_nid), event->status);
2762 if (event->status != 0) {
2763 CDEBUG(D_NETERROR, "Can't resolve address for %s: %d\n",
2764 libcfs_nid2str(peer->ibp_nid), event->status);
2767 rc = rdma_resolve_route(
2768 cmid, *kiblnd_tunables.kib_timeout * 1000);
2771 /* Can't initiate route resolution */
2772 CERROR("Can't resolve route for %s: %d\n",
2773 libcfs_nid2str(peer->ibp_nid), rc);
2775 kiblnd_peer_connect_failed(peer, 1, rc);
2776 kiblnd_peer_decref(peer);
2777 return rc; /* rc != 0 destroys cmid */
2779 case RDMA_CM_EVENT_ROUTE_ERROR:
2780 peer = (kib_peer_t *)cmid->context;
2781 CDEBUG(D_NETERROR, "%s: ROUTE ERROR %d\n",
2782 libcfs_nid2str(peer->ibp_nid), event->status);
2783 kiblnd_peer_connect_failed(peer, 1, -EHOSTUNREACH);
2784 kiblnd_peer_decref(peer);
2785 return -EHOSTUNREACH; /* rc != 0 destroys cmid */
2787 case RDMA_CM_EVENT_ROUTE_RESOLVED:
2788 peer = (kib_peer_t *)cmid->context;
2789 CDEBUG(D_NET,"%s Route resolved: %d\n",
2790 libcfs_nid2str(peer->ibp_nid), event->status);
2792 if (event->status == 0)
2793 return kiblnd_active_connect(cmid);
2795 CDEBUG(D_NETERROR, "Can't resolve route for %s: %d\n",
2796 libcfs_nid2str(peer->ibp_nid), event->status);
2797 kiblnd_peer_connect_failed(peer, 1, event->status);
2798 kiblnd_peer_decref(peer);
2799 return event->status; /* rc != 0 destroys cmid */
2801 case RDMA_CM_EVENT_UNREACHABLE:
2802 conn = (kib_conn_t *)cmid->context;
2803 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
2804 conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
2805 CDEBUG(D_NETERROR, "%s: UNREACHABLE %d\n",
2806 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
2807 kiblnd_connreq_done(conn, -ENETDOWN);
2808 kiblnd_conn_decref(conn);
2811 case RDMA_CM_EVENT_CONNECT_ERROR:
2812 conn = (kib_conn_t *)cmid->context;
2813 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
2814 conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
2815 CDEBUG(D_NETERROR, "%s: CONNECT ERROR %d\n",
2816 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
2817 kiblnd_connreq_done(conn, -ENOTCONN);
2818 kiblnd_conn_decref(conn);
2821 case RDMA_CM_EVENT_REJECTED:
2822 conn = (kib_conn_t *)cmid->context;
2823 switch (conn->ibc_state) {
2827 case IBLND_CONN_PASSIVE_WAIT:
2828 CERROR ("%s: REJECTED %d\n",
2829 libcfs_nid2str(conn->ibc_peer->ibp_nid),
2831 kiblnd_connreq_done(conn, -ECONNRESET);
2834 case IBLND_CONN_ACTIVE_CONNECT:
2835 kiblnd_rejected(conn, event->status,
2836 (void *)KIBLND_CONN_PARAM(event),
2837 KIBLND_CONN_PARAM_LEN(event));
2840 kiblnd_conn_decref(conn);
2843 case RDMA_CM_EVENT_ESTABLISHED:
2844 conn = (kib_conn_t *)cmid->context;
2845 switch (conn->ibc_state) {
2849 case IBLND_CONN_PASSIVE_WAIT:
2850 CDEBUG(D_NET, "ESTABLISHED (passive): %s\n",
2851 libcfs_nid2str(conn->ibc_peer->ibp_nid));
2852 kiblnd_connreq_done(conn, 0);
2855 case IBLND_CONN_ACTIVE_CONNECT:
2856 CDEBUG(D_NET, "ESTABLISHED(active): %s\n",
2857 libcfs_nid2str(conn->ibc_peer->ibp_nid));
2858 kiblnd_check_connreply(conn,
2859 (void *)KIBLND_CONN_PARAM(event),
2860 KIBLND_CONN_PARAM_LEN(event));
2863 /* net keeps its ref on conn! */
2866 #ifdef HAVE_OFED_RDMA_CMEV_TIMEWAIT_EXIT
2867 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
2868 CDEBUG(D_NET, "Ignore TIMEWAIT_EXIT event\n");
2871 case RDMA_CM_EVENT_DISCONNECTED:
2872 conn = (kib_conn_t *)cmid->context;
2873 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
2874 CERROR("%s DISCONNECTED\n",
2875 libcfs_nid2str(conn->ibc_peer->ibp_nid));
2876 kiblnd_connreq_done(conn, -ECONNRESET);
2878 kiblnd_close_conn(conn, 0);
2880 kiblnd_conn_decref(conn);
2881 cmid->context = NULL;
2884 case RDMA_CM_EVENT_DEVICE_REMOVAL:
2885 LCONSOLE_ERROR_MSG(0x131,
2886 "Received notification of device removal\n"
2887 "Please shutdown LNET to allow this to proceed\n");
2888 /* Can't remove network from underneath LNET for now, so I have
2892 #ifdef HAVE_OFED_RDMA_CMEV_ADDRCHANGE
2893 case RDMA_CM_EVENT_ADDR_CHANGE:
2894 LCONSOLE_INFO("Physical link changed (eg hca/port)\n");
2901 kiblnd_check_txs (kib_conn_t *conn, cfs_list_t *txs)
2907 cfs_spin_lock(&conn->ibc_lock);
2909 cfs_list_for_each (ttmp, txs) {
2910 tx = cfs_list_entry (ttmp, kib_tx_t, tx_list);
2912 if (txs != &conn->ibc_active_txs) {
2913 LASSERT (tx->tx_queued);
2915 LASSERT (!tx->tx_queued);
2916 LASSERT (tx->tx_waiting || tx->tx_sending != 0);
2919 if (cfs_time_aftereq (jiffies, tx->tx_deadline)) {
2921 CERROR("Timed out tx: %s, %lu seconds\n",
2922 kiblnd_queue2str(conn, txs),
2923 cfs_duration_sec(jiffies - tx->tx_deadline));
2928 cfs_spin_unlock(&conn->ibc_lock);
2933 kiblnd_conn_timed_out (kib_conn_t *conn)
2935 return kiblnd_check_txs(conn, &conn->ibc_tx_queue) ||
2936 kiblnd_check_txs(conn, &conn->ibc_tx_queue_rsrvd) ||
2937 kiblnd_check_txs(conn, &conn->ibc_tx_queue_nocred) ||
2938 kiblnd_check_txs(conn, &conn->ibc_active_txs);
2942 kiblnd_check_conns (int idx)
2944 cfs_list_t *peers = &kiblnd_data.kib_peers[idx];
2949 unsigned long flags;
2952 /* NB. We expect to have a look at all the peers and not find any
2953 * rdmas to time out, so we just use a shared lock while we
2955 cfs_read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2957 cfs_list_for_each (ptmp, peers) {
2958 peer = cfs_list_entry (ptmp, kib_peer_t, ibp_list);
2960 cfs_list_for_each (ctmp, &peer->ibp_conns) {
2961 conn = cfs_list_entry (ctmp, kib_conn_t, ibc_list);
2963 LASSERT (conn->ibc_state == IBLND_CONN_ESTABLISHED);
2965 /* In case we have enough credits to return via a
2966 * NOOP, but there were no non-blocking tx descs
2967 * free to do it last time... */
2968 kiblnd_check_sends(conn);
2970 if (!kiblnd_conn_timed_out(conn))
2973 /* Handle timeout by closing the whole connection. We
2974 * can only be sure RDMA activity has ceased once the
2975 * QP has been modified. */
2977 kiblnd_conn_addref(conn); /* 1 ref for me... */
2979 cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2982 CERROR("Timed out RDMA with %s (%lu)\n",
2983 libcfs_nid2str(peer->ibp_nid),
2984 cfs_duration_sec(cfs_time_current() -
2985 peer->ibp_last_alive));
2987 kiblnd_close_conn(conn, -ETIMEDOUT);
2988 kiblnd_conn_decref(conn); /* ...until here */
2990 /* start again now I've dropped the lock */
2995 cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2999 kiblnd_disconnect_conn (kib_conn_t *conn)
3001 LASSERT (!cfs_in_interrupt());
3002 LASSERT (current == kiblnd_data.kib_connd);
3003 LASSERT (conn->ibc_state == IBLND_CONN_CLOSING);
3005 rdma_disconnect(conn->ibc_cmid);
3006 kiblnd_finalise_conn(conn);
3008 kiblnd_peer_notify(conn->ibc_peer);
3012 kiblnd_connd (void *arg)
3014 cfs_waitlink_t wait;
3015 unsigned long flags;
3021 unsigned long deadline = jiffies;
3023 cfs_daemonize ("kiblnd_connd");
3024 cfs_block_allsigs ();
3026 cfs_waitlink_init (&wait);
3027 kiblnd_data.kib_connd = current;
3029 cfs_spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3031 while (!kiblnd_data.kib_shutdown) {
3035 if (!cfs_list_empty (&kiblnd_data.kib_connd_zombies)) {
3036 conn = cfs_list_entry(kiblnd_data. \
3037 kib_connd_zombies.next,
3038 kib_conn_t, ibc_list);
3039 cfs_list_del(&conn->ibc_list);
3041 cfs_spin_unlock_irqrestore (&kiblnd_data.kib_connd_lock,
3045 kiblnd_destroy_conn(conn);
3047 cfs_spin_lock_irqsave (&kiblnd_data.kib_connd_lock,
3051 if (!cfs_list_empty (&kiblnd_data.kib_connd_conns)) {
3052 conn = cfs_list_entry (kiblnd_data.kib_connd_conns.next,
3053 kib_conn_t, ibc_list);
3054 cfs_list_del(&conn->ibc_list);
3056 cfs_spin_unlock_irqrestore (&kiblnd_data.kib_connd_lock,
3060 kiblnd_disconnect_conn(conn);
3061 kiblnd_conn_decref(conn);
3063 cfs_spin_lock_irqsave (&kiblnd_data.kib_connd_lock,
3067 /* careful with the jiffy wrap... */
3068 timeout = (int)(deadline - jiffies);
3072 int chunk = kiblnd_data.kib_peer_hash_size;
3074 cfs_spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
3077 /* Time to check for RDMA timeouts on a few more
3078 * peers: I do checks every 'p' seconds on a
3079 * proportion of the peer table and I need to check
3080 * every connection 'n' times within a timeout
3081 * interval, to ensure I detect a timeout on any
3082 * connection within (n+1)/n times the timeout
3085 if (*kiblnd_tunables.kib_timeout > n * p)
3086 chunk = (chunk * n * p) /
3087 *kiblnd_tunables.kib_timeout;
3091 for (i = 0; i < chunk; i++) {
3092 kiblnd_check_conns(peer_index);
3093 peer_index = (peer_index + 1) %
3094 kiblnd_data.kib_peer_hash_size;
3097 deadline += p * CFS_HZ;
3098 cfs_spin_lock_irqsave(&kiblnd_data.kib_connd_lock,
3105 /* Nothing to do for 'timeout' */
3106 cfs_set_current_state (CFS_TASK_INTERRUPTIBLE);
3107 cfs_waitq_add (&kiblnd_data.kib_connd_waitq, &wait);
3108 cfs_spin_unlock_irqrestore (&kiblnd_data.kib_connd_lock, flags);
3110 cfs_waitq_timedwait(&wait, CFS_TASK_INTERRUPTIBLE, timeout);
3112 cfs_set_current_state (CFS_TASK_RUNNING);
3113 cfs_waitq_del (&kiblnd_data.kib_connd_waitq, &wait);
3114 cfs_spin_lock_irqsave (&kiblnd_data.kib_connd_lock, flags);
3117 cfs_spin_unlock_irqrestore (&kiblnd_data.kib_connd_lock, flags);
3119 kiblnd_thread_fini();
3124 kiblnd_qp_event(struct ib_event *event, void *arg)
3126 kib_conn_t *conn = arg;
3128 switch (event->event) {
3129 case IB_EVENT_COMM_EST:
3130 CDEBUG(D_NET, "%s established\n",
3131 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3135 CERROR("%s: Async QP event type %d\n",
3136 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3142 kiblnd_complete (struct ib_wc *wc)
3144 switch (kiblnd_wreqid2type(wc->wr_id)) {
3148 case IBLND_WID_RDMA:
3149 /* We only get RDMA completion notification if it fails. All
3150 * subsequent work items, including the final SEND will fail
3151 * too. However we can't print out any more info about the
3152 * failing RDMA because 'tx' might be back on the idle list or
3153 * even reused already if we didn't manage to post all our work
3155 CDEBUG(D_NETERROR, "RDMA (tx: %p) failed: %d\n",
3156 kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3160 kiblnd_tx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3164 kiblnd_rx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status,
3171 kiblnd_cq_completion (struct ib_cq *cq, void *arg)
3173 /* NB I'm not allowed to schedule this conn once its refcount has
3174 * reached 0. Since fundamentally I'm racing with scheduler threads
3175 * consuming my CQ I could be called after all completions have
3176 * occurred. But in this case, ibc_nrx == 0 && ibc_nsends_posted == 0
3177 * and this CQ is about to be destroyed so I NOOP. */
3178 kib_conn_t *conn = (kib_conn_t *)arg;
3179 unsigned long flags;
3181 LASSERT (cq == conn->ibc_cq);
3183 cfs_spin_lock_irqsave(&kiblnd_data.kib_sched_lock, flags);
3185 conn->ibc_ready = 1;
3187 if (!conn->ibc_scheduled &&
3188 (conn->ibc_nrx > 0 ||
3189 conn->ibc_nsends_posted > 0)) {
3190 kiblnd_conn_addref(conn); /* +1 ref for sched_conns */
3191 conn->ibc_scheduled = 1;
3192 cfs_list_add_tail(&conn->ibc_sched_list,
3193 &kiblnd_data.kib_sched_conns);
3194 cfs_waitq_signal(&kiblnd_data.kib_sched_waitq);
3197 cfs_spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock, flags);
3201 kiblnd_cq_event(struct ib_event *event, void *arg)
3203 kib_conn_t *conn = arg;
3205 CERROR("%s: async CQ event type %d\n",
3206 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3210 kiblnd_scheduler(void *arg)
3212 long id = (long)arg;
3213 cfs_waitlink_t wait;
3215 unsigned long flags;
3222 snprintf(name, sizeof(name), "kiblnd_sd_%02ld", id);
3223 cfs_daemonize(name);
3224 cfs_block_allsigs();
3226 cfs_waitlink_init(&wait);
3228 cfs_spin_lock_irqsave(&kiblnd_data.kib_sched_lock, flags);
3230 while (!kiblnd_data.kib_shutdown) {
3231 if (busy_loops++ >= IBLND_RESCHED) {
3232 cfs_spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock,
3238 cfs_spin_lock_irqsave(&kiblnd_data.kib_sched_lock,
3244 if (!cfs_list_empty(&kiblnd_data.kib_sched_conns)) {
3245 conn = cfs_list_entry(kiblnd_data.kib_sched_conns.next,
3246 kib_conn_t, ibc_sched_list);
3247 /* take over kib_sched_conns' ref on conn... */
3248 LASSERT(conn->ibc_scheduled);
3249 cfs_list_del(&conn->ibc_sched_list);
3250 conn->ibc_ready = 0;
3252 cfs_spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock,
3255 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3257 rc = ib_req_notify_cq(conn->ibc_cq,
3260 CWARN("%s: ib_req_notify_cq failed: %d, "
3261 "closing connection\n",
3262 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
3263 kiblnd_close_conn(conn, -EIO);
3264 kiblnd_conn_decref(conn);
3265 cfs_spin_lock_irqsave(&kiblnd_data. \
3271 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3275 CWARN("%s: ib_poll_cq failed: %d, "
3276 "closing connection\n",
3277 libcfs_nid2str(conn->ibc_peer->ibp_nid),
3279 kiblnd_close_conn(conn, -EIO);
3280 kiblnd_conn_decref(conn);
3281 cfs_spin_lock_irqsave(&kiblnd_data. \
3282 kib_sched_lock, flags);
3286 cfs_spin_lock_irqsave(&kiblnd_data.kib_sched_lock,
3289 if (rc != 0 || conn->ibc_ready) {
3290 /* There may be another completion waiting; get
3291 * another scheduler to check while I handle
3293 kiblnd_conn_addref(conn); /* +1 ref for sched_conns */
3294 cfs_list_add_tail(&conn->ibc_sched_list,
3295 &kiblnd_data.kib_sched_conns);
3296 cfs_waitq_signal(&kiblnd_data.kib_sched_waitq);
3298 conn->ibc_scheduled = 0;
3302 cfs_spin_unlock_irqrestore(&kiblnd_data. \
3306 kiblnd_complete(&wc);
3308 cfs_spin_lock_irqsave(&kiblnd_data. \
3313 kiblnd_conn_decref(conn); /* ...drop my ref from above */
3320 cfs_set_current_state(CFS_TASK_INTERRUPTIBLE);
3321 cfs_waitq_add_exclusive(&kiblnd_data.kib_sched_waitq, &wait);
3322 cfs_spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock, flags);
3324 cfs_waitq_wait(&wait, CFS_TASK_INTERRUPTIBLE);
3327 cfs_waitq_del(&kiblnd_data.kib_sched_waitq, &wait);
3328 cfs_set_current_state(CFS_TASK_RUNNING);
3329 cfs_spin_lock_irqsave(&kiblnd_data.kib_sched_lock, flags);
3332 cfs_spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock, flags);
3334 kiblnd_thread_fini();