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(conn->ibc_hdev, 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 /* This ensures the credit taken by NOOP can be returned */
332 if (msg->ibm_type == IBLND_MSG_NOOP &&
333 !IBLND_OOB_CAPABLE(conn->ibc_version)) /* v1 only */
334 conn->ibc_outstanding_credits++;
336 cfs_spin_unlock(&conn->ibc_lock);
337 kiblnd_check_sends(conn);
340 switch (msg->ibm_type) {
342 CERROR("Bad IBLND message type %x from %s\n",
343 msg->ibm_type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
344 post_credit = IBLND_POSTRX_NO_CREDIT;
349 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
350 post_credit = IBLND_POSTRX_NO_CREDIT;
354 if (credits != 0) /* credit already posted */
355 post_credit = IBLND_POSTRX_NO_CREDIT;
356 else /* a keepalive NOOP */
357 post_credit = IBLND_POSTRX_PEER_CREDIT;
360 case IBLND_MSG_IMMEDIATE:
361 post_credit = IBLND_POSTRX_DONT_POST;
362 rc = lnet_parse(ni, &msg->ibm_u.immediate.ibim_hdr,
363 msg->ibm_srcnid, rx, 0);
364 if (rc < 0) /* repost on error */
365 post_credit = IBLND_POSTRX_PEER_CREDIT;
368 case IBLND_MSG_PUT_REQ:
369 post_credit = IBLND_POSTRX_DONT_POST;
370 rc = lnet_parse(ni, &msg->ibm_u.putreq.ibprm_hdr,
371 msg->ibm_srcnid, rx, 1);
372 if (rc < 0) /* repost on error */
373 post_credit = IBLND_POSTRX_PEER_CREDIT;
376 case IBLND_MSG_PUT_NAK:
377 CWARN ("PUT_NACK from %s\n",
378 libcfs_nid2str(conn->ibc_peer->ibp_nid));
379 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
380 kiblnd_handle_completion(conn, IBLND_MSG_PUT_REQ,
381 msg->ibm_u.completion.ibcm_status,
382 msg->ibm_u.completion.ibcm_cookie);
385 case IBLND_MSG_PUT_ACK:
386 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
388 cfs_spin_lock(&conn->ibc_lock);
389 tx = kiblnd_find_waiting_tx_locked(conn, IBLND_MSG_PUT_REQ,
390 msg->ibm_u.putack.ibpam_src_cookie);
392 cfs_list_del(&tx->tx_list);
393 cfs_spin_unlock(&conn->ibc_lock);
396 CERROR("Unmatched PUT_ACK from %s\n",
397 libcfs_nid2str(conn->ibc_peer->ibp_nid));
402 LASSERT (tx->tx_waiting);
403 /* CAVEAT EMPTOR: I could be racing with tx_complete, but...
404 * (a) I can overwrite tx_msg since my peer has received it!
405 * (b) tx_waiting set tells tx_complete() it's not done. */
407 tx->tx_nwrq = 0; /* overwrite PUT_REQ */
409 rc2 = kiblnd_init_rdma(conn, tx, IBLND_MSG_PUT_DONE,
410 kiblnd_rd_size(&msg->ibm_u.putack.ibpam_rd),
411 &msg->ibm_u.putack.ibpam_rd,
412 msg->ibm_u.putack.ibpam_dst_cookie);
414 CERROR("Can't setup rdma for PUT to %s: %d\n",
415 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc2);
417 cfs_spin_lock(&conn->ibc_lock);
418 tx->tx_waiting = 0; /* clear waiting and queue atomically */
419 kiblnd_queue_tx_locked(tx, conn);
420 cfs_spin_unlock(&conn->ibc_lock);
423 case IBLND_MSG_PUT_DONE:
424 post_credit = IBLND_POSTRX_PEER_CREDIT;
425 kiblnd_handle_completion(conn, IBLND_MSG_PUT_ACK,
426 msg->ibm_u.completion.ibcm_status,
427 msg->ibm_u.completion.ibcm_cookie);
430 case IBLND_MSG_GET_REQ:
431 post_credit = IBLND_POSTRX_DONT_POST;
432 rc = lnet_parse(ni, &msg->ibm_u.get.ibgm_hdr,
433 msg->ibm_srcnid, rx, 1);
434 if (rc < 0) /* repost on error */
435 post_credit = IBLND_POSTRX_PEER_CREDIT;
438 case IBLND_MSG_GET_DONE:
439 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
440 kiblnd_handle_completion(conn, IBLND_MSG_GET_REQ,
441 msg->ibm_u.completion.ibcm_status,
442 msg->ibm_u.completion.ibcm_cookie);
446 if (rc < 0) /* protocol error */
447 kiblnd_close_conn(conn, rc);
449 if (post_credit != IBLND_POSTRX_DONT_POST)
450 kiblnd_post_rx(rx, post_credit);
454 kiblnd_rx_complete (kib_rx_t *rx, int status, int nob)
456 kib_msg_t *msg = rx->rx_msg;
457 kib_conn_t *conn = rx->rx_conn;
458 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
459 kib_net_t *net = ni->ni_data;
463 LASSERT (net != NULL);
464 LASSERT (rx->rx_nob < 0); /* was posted */
465 rx->rx_nob = 0; /* isn't now */
467 if (conn->ibc_state > IBLND_CONN_ESTABLISHED)
470 if (status != IB_WC_SUCCESS) {
471 CNETERR("Rx from %s failed: %d\n",
472 libcfs_nid2str(conn->ibc_peer->ibp_nid), status);
479 rc = kiblnd_unpack_msg(msg, rx->rx_nob);
481 CERROR ("Error %d unpacking rx from %s\n",
482 rc, libcfs_nid2str(conn->ibc_peer->ibp_nid));
486 if (msg->ibm_srcnid != conn->ibc_peer->ibp_nid ||
487 msg->ibm_dstnid != ni->ni_nid ||
488 msg->ibm_srcstamp != conn->ibc_incarnation ||
489 msg->ibm_dststamp != net->ibn_incarnation) {
490 CERROR ("Stale rx from %s\n",
491 libcfs_nid2str(conn->ibc_peer->ibp_nid));
496 /* set time last known alive */
497 kiblnd_peer_alive(conn->ibc_peer);
499 /* racing with connection establishment/teardown! */
501 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
502 cfs_rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
505 cfs_write_lock_irqsave(g_lock, flags);
506 /* must check holding global lock to eliminate race */
507 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
508 cfs_list_add_tail(&rx->rx_list, &conn->ibc_early_rxs);
509 cfs_write_unlock_irqrestore(g_lock, flags);
512 cfs_write_unlock_irqrestore(g_lock, flags);
514 kiblnd_handle_rx(rx);
518 CDEBUG(D_NET, "rx %p conn %p\n", rx, conn);
519 kiblnd_close_conn(conn, err);
521 kiblnd_drop_rx(rx); /* Don't re-post rx. */
525 kiblnd_kvaddr_to_page (unsigned long vaddr)
529 if (vaddr >= VMALLOC_START &&
530 vaddr < VMALLOC_END) {
531 page = vmalloc_to_page ((void *)vaddr);
532 LASSERT (page != NULL);
535 #ifdef CONFIG_HIGHMEM
536 if (vaddr >= PKMAP_BASE &&
537 vaddr < (PKMAP_BASE + LAST_PKMAP * PAGE_SIZE)) {
538 /* No highmem pages only used for bulk (kiov) I/O */
539 CERROR("find page for address in highmem\n");
543 page = virt_to_page (vaddr);
544 LASSERT (page != NULL);
549 kiblnd_fmr_map_tx(kib_net_t *net, kib_tx_t *tx, kib_rdma_desc_t *rd, int nob)
551 kib_hca_dev_t *hdev = tx->tx_pool->tpo_hdev;
552 __u64 *pages = tx->tx_pages;
558 for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
559 for (size = 0; size < rd->rd_frags[i].rf_nob;
560 size += hdev->ibh_page_size) {
561 pages[npages ++] = (rd->rd_frags[i].rf_addr &
562 hdev->ibh_page_mask) + size;
566 rc = kiblnd_fmr_pool_map(&net->ibn_fmr_ps, pages, npages, 0, &tx->tx_u.fmr);
568 CERROR ("Can't map %d pages: %d\n", npages, rc);
572 /* If rd is not tx_rd, it's going to get sent to a peer, who will need
574 rd->rd_key = (rd != tx->tx_rd) ? tx->tx_u.fmr.fmr_pfmr->fmr->rkey :
575 tx->tx_u.fmr.fmr_pfmr->fmr->lkey;
576 rd->rd_frags[0].rf_addr &= ~hdev->ibh_page_mask;
577 rd->rd_frags[0].rf_nob = nob;
584 kiblnd_pmr_map_tx(kib_net_t *net, kib_tx_t *tx, kib_rdma_desc_t *rd, int nob)
586 kib_hca_dev_t *hdev = tx->tx_pool->tpo_hdev;
590 iova = rd->rd_frags[0].rf_addr & ~hdev->ibh_page_mask;
592 rc = kiblnd_pmr_pool_map(&net->ibn_pmr_ps, hdev, rd, &iova, &tx->tx_u.pmr);
594 CERROR("Failed to create MR by phybuf: %d\n", rc);
598 /* If rd is not tx_rd, it's going to get sent to a peer, who will need
600 rd->rd_key = (rd != tx->tx_rd) ? tx->tx_u.pmr->pmr_mr->rkey :
601 tx->tx_u.pmr->pmr_mr->lkey;
603 rd->rd_frags[0].rf_addr = iova;
604 rd->rd_frags[0].rf_nob = nob;
610 kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx)
612 kib_net_t *net = ni->ni_data;
614 LASSERT (net != NULL);
616 if (net->ibn_with_fmr && tx->tx_u.fmr.fmr_pfmr != NULL) {
617 kiblnd_fmr_pool_unmap(&tx->tx_u.fmr, tx->tx_status);
618 tx->tx_u.fmr.fmr_pfmr = NULL;
619 } else if (net->ibn_with_pmr && tx->tx_u.pmr != NULL) {
620 kiblnd_pmr_pool_unmap(tx->tx_u.pmr);
624 if (tx->tx_nfrags != 0) {
625 kiblnd_dma_unmap_sg(tx->tx_pool->tpo_hdev->ibh_ibdev,
626 tx->tx_frags, tx->tx_nfrags, tx->tx_dmadir);
632 kiblnd_map_tx(lnet_ni_t *ni, kib_tx_t *tx,
633 kib_rdma_desc_t *rd, int nfrags)
635 kib_hca_dev_t *hdev = tx->tx_pool->tpo_hdev;
636 kib_net_t *net = ni->ni_data;
637 struct ib_mr *mr = NULL;
641 /* If rd is not tx_rd, it's going to get sent to a peer and I'm the
643 tx->tx_dmadir = (rd != tx->tx_rd) ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
644 tx->tx_nfrags = nfrags;
647 kiblnd_dma_map_sg(hdev->ibh_ibdev,
648 tx->tx_frags, tx->tx_nfrags, tx->tx_dmadir);
650 for (i = 0, nob = 0; i < rd->rd_nfrags; i++) {
651 rd->rd_frags[i].rf_nob = kiblnd_sg_dma_len(
652 hdev->ibh_ibdev, &tx->tx_frags[i]);
653 rd->rd_frags[i].rf_addr = kiblnd_sg_dma_address(
654 hdev->ibh_ibdev, &tx->tx_frags[i]);
655 nob += rd->rd_frags[i].rf_nob;
658 /* looking for pre-mapping MR */
659 mr = kiblnd_find_rd_dma_mr(hdev, rd);
661 /* found pre-mapping MR */
662 rd->rd_key = (rd != tx->tx_rd) ? mr->rkey : mr->lkey;
666 if (net->ibn_with_fmr)
667 return kiblnd_fmr_map_tx(net, tx, rd, nob);
668 else if (net->ibn_with_pmr)
669 return kiblnd_pmr_map_tx(net, tx, rd, nob);
676 kiblnd_setup_rd_iov(lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd,
677 unsigned int niov, struct iovec *iov, int offset, int nob)
679 kib_net_t *net = ni->ni_data;
681 struct scatterlist *sg;
688 LASSERT (net != NULL);
690 while (offset >= iov->iov_len) {
691 offset -= iov->iov_len;
701 vaddr = ((unsigned long)iov->iov_base) + offset;
702 page_offset = vaddr & (PAGE_SIZE - 1);
703 page = kiblnd_kvaddr_to_page(vaddr);
705 CERROR ("Can't find page\n");
709 fragnob = min((int)(iov->iov_len - offset), nob);
710 fragnob = min(fragnob, (int)PAGE_SIZE - page_offset);
712 sg_set_page(sg, page, fragnob, page_offset);
715 if (offset + fragnob < iov->iov_len) {
725 return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags);
729 kiblnd_setup_rd_kiov (lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd,
730 int nkiov, lnet_kiov_t *kiov, int offset, int nob)
732 kib_net_t *net = ni->ni_data;
733 struct scatterlist *sg;
736 CDEBUG(D_NET, "niov %d offset %d nob %d\n", nkiov, offset, nob);
740 LASSERT (net != NULL);
742 while (offset >= kiov->kiov_len) {
743 offset -= kiov->kiov_len;
753 fragnob = min((int)(kiov->kiov_len - offset), nob);
755 memset(sg, 0, sizeof(*sg));
756 sg_set_page(sg, kiov->kiov_page, fragnob,
757 kiov->kiov_offset + offset);
766 return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags);
770 kiblnd_post_tx_locked (kib_conn_t *conn, kib_tx_t *tx, int credit)
772 kib_msg_t *msg = tx->tx_msg;
773 kib_peer_t *peer = conn->ibc_peer;
774 int ver = conn->ibc_version;
777 struct ib_send_wr *bad_wrq;
779 LASSERT (tx->tx_queued);
780 /* We rely on this for QP sizing */
781 LASSERT (tx->tx_nwrq > 0);
782 LASSERT (tx->tx_nwrq <= 1 + IBLND_RDMA_FRAGS(ver));
784 LASSERT (credit == 0 || credit == 1);
785 LASSERT (conn->ibc_outstanding_credits >= 0);
786 LASSERT (conn->ibc_outstanding_credits <= IBLND_MSG_QUEUE_SIZE(ver));
787 LASSERT (conn->ibc_credits >= 0);
788 LASSERT (conn->ibc_credits <= IBLND_MSG_QUEUE_SIZE(ver));
790 if (conn->ibc_nsends_posted == IBLND_CONCURRENT_SENDS(ver)) {
791 /* tx completions outstanding... */
792 CDEBUG(D_NET, "%s: posted enough\n",
793 libcfs_nid2str(peer->ibp_nid));
797 if (credit != 0 && conn->ibc_credits == 0) { /* no credits */
798 CDEBUG(D_NET, "%s: no credits\n",
799 libcfs_nid2str(peer->ibp_nid));
803 if (credit != 0 && !IBLND_OOB_CAPABLE(ver) &&
804 conn->ibc_credits == 1 && /* last credit reserved */
805 msg->ibm_type != IBLND_MSG_NOOP) { /* for NOOP */
806 CDEBUG(D_NET, "%s: not using last credit\n",
807 libcfs_nid2str(peer->ibp_nid));
811 /* NB don't drop ibc_lock before bumping tx_sending */
812 cfs_list_del(&tx->tx_list);
815 if (msg->ibm_type == IBLND_MSG_NOOP &&
816 (!kiblnd_send_noop(conn) || /* redundant NOOP */
817 (IBLND_OOB_CAPABLE(ver) && /* posted enough NOOP */
818 conn->ibc_noops_posted == IBLND_OOB_MSGS(ver)))) {
819 /* OK to drop when posted enough NOOPs, since
820 * kiblnd_check_sends will queue NOOP again when
821 * posted NOOPs complete */
822 cfs_spin_unlock(&conn->ibc_lock);
823 kiblnd_tx_done(peer->ibp_ni, tx);
824 cfs_spin_lock(&conn->ibc_lock);
825 CDEBUG(D_NET, "%s(%d): redundant or enough NOOP\n",
826 libcfs_nid2str(peer->ibp_nid),
827 conn->ibc_noops_posted);
831 kiblnd_pack_msg(peer->ibp_ni, msg, ver, conn->ibc_outstanding_credits,
832 peer->ibp_nid, conn->ibc_incarnation);
834 conn->ibc_credits -= credit;
835 conn->ibc_outstanding_credits = 0;
836 conn->ibc_nsends_posted++;
837 if (msg->ibm_type == IBLND_MSG_NOOP)
838 conn->ibc_noops_posted++;
840 /* CAVEAT EMPTOR! This tx could be the PUT_DONE of an RDMA
841 * PUT. If so, it was first queued here as a PUT_REQ, sent and
842 * stashed on ibc_active_txs, matched by an incoming PUT_ACK,
843 * and then re-queued here. It's (just) possible that
844 * tx_sending is non-zero if we've not done the tx_complete()
845 * from the first send; hence the ++ rather than = below. */
847 cfs_list_add(&tx->tx_list, &conn->ibc_active_txs);
849 /* I'm still holding ibc_lock! */
850 if (conn->ibc_state != IBLND_CONN_ESTABLISHED) {
852 } else if (tx->tx_pool->tpo_pool.po_failed ||
853 conn->ibc_hdev != tx->tx_pool->tpo_hdev) {
854 /* close_conn will launch failover */
857 rc = ib_post_send(conn->ibc_cmid->qp,
858 tx->tx_wrq, &bad_wrq);
861 conn->ibc_last_send = jiffies;
866 /* NB credits are transferred in the actual
867 * message, which can only be the last work item */
868 conn->ibc_credits += credit;
869 conn->ibc_outstanding_credits += msg->ibm_credits;
870 conn->ibc_nsends_posted--;
871 if (msg->ibm_type == IBLND_MSG_NOOP)
872 conn->ibc_noops_posted--;
878 done = (tx->tx_sending == 0);
880 cfs_list_del(&tx->tx_list);
882 cfs_spin_unlock(&conn->ibc_lock);
884 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
885 CERROR("Error %d posting transmit to %s\n",
886 rc, libcfs_nid2str(peer->ibp_nid));
888 CDEBUG(D_NET, "Error %d posting transmit to %s\n",
889 rc, libcfs_nid2str(peer->ibp_nid));
891 kiblnd_close_conn(conn, rc);
894 kiblnd_tx_done(peer->ibp_ni, tx);
896 cfs_spin_lock(&conn->ibc_lock);
902 kiblnd_check_sends (kib_conn_t *conn)
904 int ver = conn->ibc_version;
905 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
908 /* Don't send anything until after the connection is established */
909 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
910 CDEBUG(D_NET, "%s too soon\n",
911 libcfs_nid2str(conn->ibc_peer->ibp_nid));
915 cfs_spin_lock(&conn->ibc_lock);
917 LASSERT (conn->ibc_nsends_posted <= IBLND_CONCURRENT_SENDS(ver));
918 LASSERT (!IBLND_OOB_CAPABLE(ver) ||
919 conn->ibc_noops_posted <= IBLND_OOB_MSGS(ver));
920 LASSERT (conn->ibc_reserved_credits >= 0);
922 while (conn->ibc_reserved_credits > 0 &&
923 !cfs_list_empty(&conn->ibc_tx_queue_rsrvd)) {
924 tx = cfs_list_entry(conn->ibc_tx_queue_rsrvd.next,
926 cfs_list_del(&tx->tx_list);
927 cfs_list_add_tail(&tx->tx_list, &conn->ibc_tx_queue);
928 conn->ibc_reserved_credits--;
931 conn->ibc_retry_noop = 0;
932 if (kiblnd_send_noop(conn)) {
933 cfs_spin_unlock(&conn->ibc_lock);
935 tx = kiblnd_get_idle_tx(ni);
937 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_NOOP, 0);
939 cfs_spin_lock(&conn->ibc_lock);
941 kiblnd_queue_tx_locked(tx, conn);
942 else if (kiblnd_send_noop(conn)) /* still... */
943 conn->ibc_retry_noop = 1;
946 kiblnd_conn_addref(conn); /* 1 ref for me.... (see b21911) */
951 if (!cfs_list_empty(&conn->ibc_tx_queue_nocred)) {
953 tx = cfs_list_entry(conn->ibc_tx_queue_nocred.next,
955 } else if (!cfs_list_empty(&conn->ibc_tx_noops)) {
956 LASSERT (!IBLND_OOB_CAPABLE(ver));
958 tx = cfs_list_entry(conn->ibc_tx_noops.next,
960 } else if (!cfs_list_empty(&conn->ibc_tx_queue)) {
962 tx = cfs_list_entry(conn->ibc_tx_queue.next,
967 if (kiblnd_post_tx_locked(conn, tx, credit) != 0)
971 cfs_spin_unlock(&conn->ibc_lock);
973 kiblnd_conn_decref(conn); /* ...until here */
977 kiblnd_tx_complete (kib_tx_t *tx, int status)
979 int failed = (status != IB_WC_SUCCESS);
980 kib_conn_t *conn = tx->tx_conn;
983 LASSERT (tx->tx_sending > 0);
986 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
987 CNETERR("Tx -> %s cookie "LPX64
988 " sending %d waiting %d: failed %d\n",
989 libcfs_nid2str(conn->ibc_peer->ibp_nid),
990 tx->tx_cookie, tx->tx_sending, tx->tx_waiting,
993 kiblnd_close_conn(conn, -EIO);
995 kiblnd_peer_alive(conn->ibc_peer);
998 cfs_spin_lock(&conn->ibc_lock);
1000 /* I could be racing with rdma completion. Whoever makes 'tx' idle
1001 * gets to free it, which also drops its ref on 'conn'. */
1004 conn->ibc_nsends_posted--;
1005 if (tx->tx_msg->ibm_type == IBLND_MSG_NOOP)
1006 conn->ibc_noops_posted--;
1009 tx->tx_waiting = 0; /* don't wait for peer */
1010 tx->tx_status = -EIO;
1013 idle = (tx->tx_sending == 0) && /* This is the final callback */
1014 !tx->tx_waiting && /* Not waiting for peer */
1015 !tx->tx_queued; /* Not re-queued (PUT_DONE) */
1017 cfs_list_del(&tx->tx_list);
1019 kiblnd_conn_addref(conn); /* 1 ref for me.... */
1021 cfs_spin_unlock(&conn->ibc_lock);
1024 kiblnd_tx_done(conn->ibc_peer->ibp_ni, tx);
1026 kiblnd_check_sends(conn);
1028 kiblnd_conn_decref(conn); /* ...until here */
1032 kiblnd_init_tx_msg (lnet_ni_t *ni, kib_tx_t *tx, int type, int body_nob)
1034 kib_hca_dev_t *hdev = tx->tx_pool->tpo_hdev;
1035 struct ib_sge *sge = &tx->tx_sge[tx->tx_nwrq];
1036 struct ib_send_wr *wrq = &tx->tx_wrq[tx->tx_nwrq];
1037 int nob = offsetof (kib_msg_t, ibm_u) + body_nob;
1040 LASSERT (tx->tx_nwrq >= 0);
1041 LASSERT (tx->tx_nwrq < IBLND_MAX_RDMA_FRAGS + 1);
1042 LASSERT (nob <= IBLND_MSG_SIZE);
1044 kiblnd_init_msg(tx->tx_msg, type, body_nob);
1046 mr = kiblnd_find_dma_mr(hdev, tx->tx_msgaddr, nob);
1047 LASSERT (mr != NULL);
1049 sge->lkey = mr->lkey;
1050 sge->addr = tx->tx_msgaddr;
1053 memset(wrq, 0, sizeof(*wrq));
1056 wrq->wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_TX);
1059 wrq->opcode = IB_WR_SEND;
1060 wrq->send_flags = IB_SEND_SIGNALED;
1066 kiblnd_init_rdma (kib_conn_t *conn, kib_tx_t *tx, int type,
1067 int resid, kib_rdma_desc_t *dstrd, __u64 dstcookie)
1069 kib_msg_t *ibmsg = tx->tx_msg;
1070 kib_rdma_desc_t *srcrd = tx->tx_rd;
1071 struct ib_sge *sge = &tx->tx_sge[0];
1072 struct ib_send_wr *wrq = &tx->tx_wrq[0];
1078 LASSERT (!cfs_in_interrupt());
1079 LASSERT (tx->tx_nwrq == 0);
1080 LASSERT (type == IBLND_MSG_GET_DONE ||
1081 type == IBLND_MSG_PUT_DONE);
1083 srcidx = dstidx = 0;
1086 if (srcidx >= srcrd->rd_nfrags) {
1087 CERROR("Src buffer exhausted: %d frags\n", srcidx);
1092 if (dstidx == dstrd->rd_nfrags) {
1093 CERROR("Dst buffer exhausted: %d frags\n", dstidx);
1098 if (tx->tx_nwrq == IBLND_RDMA_FRAGS(conn->ibc_version)) {
1099 CERROR("RDMA too fragmented for %s (%d): "
1100 "%d/%d src %d/%d dst frags\n",
1101 libcfs_nid2str(conn->ibc_peer->ibp_nid),
1102 IBLND_RDMA_FRAGS(conn->ibc_version),
1103 srcidx, srcrd->rd_nfrags,
1104 dstidx, dstrd->rd_nfrags);
1109 wrknob = MIN(MIN(kiblnd_rd_frag_size(srcrd, srcidx),
1110 kiblnd_rd_frag_size(dstrd, dstidx)), resid);
1112 sge = &tx->tx_sge[tx->tx_nwrq];
1113 sge->addr = kiblnd_rd_frag_addr(srcrd, srcidx);
1114 sge->lkey = kiblnd_rd_frag_key(srcrd, srcidx);
1115 sge->length = wrknob;
1117 wrq = &tx->tx_wrq[tx->tx_nwrq];
1119 wrq->next = wrq + 1;
1120 wrq->wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_RDMA);
1123 wrq->opcode = IB_WR_RDMA_WRITE;
1124 wrq->send_flags = 0;
1126 wrq->wr.rdma.remote_addr = kiblnd_rd_frag_addr(dstrd, dstidx);
1127 wrq->wr.rdma.rkey = kiblnd_rd_frag_key(dstrd, dstidx);
1129 srcidx = kiblnd_rd_consume_frag(srcrd, srcidx, wrknob);
1130 dstidx = kiblnd_rd_consume_frag(dstrd, dstidx, wrknob);
1139 if (rc < 0) /* no RDMA if completing with failure */
1142 ibmsg->ibm_u.completion.ibcm_status = rc;
1143 ibmsg->ibm_u.completion.ibcm_cookie = dstcookie;
1144 kiblnd_init_tx_msg(conn->ibc_peer->ibp_ni, tx,
1145 type, sizeof (kib_completion_msg_t));
1151 kiblnd_queue_tx_locked (kib_tx_t *tx, kib_conn_t *conn)
1155 LASSERT (tx->tx_nwrq > 0); /* work items set up */
1156 LASSERT (!tx->tx_queued); /* not queued for sending already */
1157 LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1160 tx->tx_deadline = jiffies + (*kiblnd_tunables.kib_timeout * CFS_HZ);
1162 if (tx->tx_conn == NULL) {
1163 kiblnd_conn_addref(conn);
1165 LASSERT (tx->tx_msg->ibm_type != IBLND_MSG_PUT_DONE);
1167 /* PUT_DONE first attached to conn as a PUT_REQ */
1168 LASSERT (tx->tx_conn == conn);
1169 LASSERT (tx->tx_msg->ibm_type == IBLND_MSG_PUT_DONE);
1172 switch (tx->tx_msg->ibm_type) {
1176 case IBLND_MSG_PUT_REQ:
1177 case IBLND_MSG_GET_REQ:
1178 q = &conn->ibc_tx_queue_rsrvd;
1181 case IBLND_MSG_PUT_NAK:
1182 case IBLND_MSG_PUT_ACK:
1183 case IBLND_MSG_PUT_DONE:
1184 case IBLND_MSG_GET_DONE:
1185 q = &conn->ibc_tx_queue_nocred;
1188 case IBLND_MSG_NOOP:
1189 if (IBLND_OOB_CAPABLE(conn->ibc_version))
1190 q = &conn->ibc_tx_queue_nocred;
1192 q = &conn->ibc_tx_noops;
1195 case IBLND_MSG_IMMEDIATE:
1196 q = &conn->ibc_tx_queue;
1200 cfs_list_add_tail(&tx->tx_list, q);
1204 kiblnd_queue_tx (kib_tx_t *tx, kib_conn_t *conn)
1206 cfs_spin_lock(&conn->ibc_lock);
1207 kiblnd_queue_tx_locked(tx, conn);
1208 cfs_spin_unlock(&conn->ibc_lock);
1210 kiblnd_check_sends(conn);
1214 kiblnd_connect_peer (kib_peer_t *peer)
1216 struct rdma_cm_id *cmid;
1218 kib_net_t *net = peer->ibp_ni->ni_data;
1219 struct sockaddr_in srcaddr;
1220 struct sockaddr_in dstaddr;
1223 LASSERT (net != NULL);
1224 LASSERT (peer->ibp_connecting > 0);
1226 cmid = rdma_create_id(kiblnd_cm_callback, peer, RDMA_PS_TCP);
1228 CERROR("Can't create CMID for %s: %ld\n",
1229 libcfs_nid2str(peer->ibp_nid), PTR_ERR(cmid));
1235 memset(&srcaddr, 0, sizeof(srcaddr));
1236 srcaddr.sin_family = AF_INET;
1237 srcaddr.sin_addr.s_addr = htonl(dev->ibd_ifip);
1239 memset(&dstaddr, 0, sizeof(dstaddr));
1240 dstaddr.sin_family = AF_INET;
1241 dstaddr.sin_port = htons(*kiblnd_tunables.kib_service);
1242 dstaddr.sin_addr.s_addr = htonl(LNET_NIDADDR(peer->ibp_nid));
1244 kiblnd_peer_addref(peer); /* cmid's ref */
1246 rc = rdma_resolve_addr(cmid,
1247 (struct sockaddr *)&srcaddr,
1248 (struct sockaddr *)&dstaddr,
1249 *kiblnd_tunables.kib_timeout * 1000);
1251 LASSERT (cmid->device != NULL);
1252 CDEBUG(D_NET, "%s: connection bound to %s:%u.%u.%u.%u:%s\n",
1253 libcfs_nid2str(peer->ibp_nid), dev->ibd_ifname,
1254 HIPQUAD(dev->ibd_ifip), cmid->device->name);
1258 /* Can't initiate address resolution: */
1259 CERROR("Can't resolve addr for %s: %d\n",
1260 libcfs_nid2str(peer->ibp_nid), rc);
1262 kiblnd_peer_decref(peer); /* cmid's ref */
1263 rdma_destroy_id(cmid);
1265 kiblnd_peer_connect_failed(peer, 1, rc);
1269 kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid)
1274 cfs_rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
1275 unsigned long flags;
1278 /* If I get here, I've committed to send, so I complete the tx with
1279 * failure on any problems */
1281 LASSERT (tx == NULL || tx->tx_conn == NULL); /* only set when assigned a conn */
1282 LASSERT (tx == NULL || tx->tx_nwrq > 0); /* work items have been set up */
1284 /* First time, just use a read lock since I expect to find my peer
1286 cfs_read_lock_irqsave(g_lock, flags);
1288 peer = kiblnd_find_peer_locked(nid);
1289 if (peer != NULL && !cfs_list_empty(&peer->ibp_conns)) {
1290 /* Found a peer with an established connection */
1291 conn = kiblnd_get_conn_locked(peer);
1292 kiblnd_conn_addref(conn); /* 1 ref for me... */
1294 cfs_read_unlock_irqrestore(g_lock, flags);
1297 kiblnd_queue_tx(tx, conn);
1298 kiblnd_conn_decref(conn); /* ...to here */
1302 cfs_read_unlock(g_lock);
1303 /* Re-try with a write lock */
1304 cfs_write_lock(g_lock);
1306 peer = kiblnd_find_peer_locked(nid);
1308 if (cfs_list_empty(&peer->ibp_conns)) {
1309 /* found a peer, but it's still connecting... */
1310 LASSERT (peer->ibp_connecting != 0 ||
1311 peer->ibp_accepting != 0);
1313 cfs_list_add_tail(&tx->tx_list,
1314 &peer->ibp_tx_queue);
1315 cfs_write_unlock_irqrestore(g_lock, flags);
1317 conn = kiblnd_get_conn_locked(peer);
1318 kiblnd_conn_addref(conn); /* 1 ref for me... */
1320 cfs_write_unlock_irqrestore(g_lock, flags);
1323 kiblnd_queue_tx(tx, conn);
1324 kiblnd_conn_decref(conn); /* ...to here */
1329 cfs_write_unlock_irqrestore(g_lock, flags);
1331 /* Allocate a peer ready to add to the peer table and retry */
1332 rc = kiblnd_create_peer(ni, &peer, nid);
1334 CERROR("Can't create peer %s\n", libcfs_nid2str(nid));
1336 tx->tx_status = -EHOSTUNREACH;
1338 kiblnd_tx_done(ni, tx);
1343 cfs_write_lock_irqsave(g_lock, flags);
1345 peer2 = kiblnd_find_peer_locked(nid);
1346 if (peer2 != NULL) {
1347 if (cfs_list_empty(&peer2->ibp_conns)) {
1348 /* found a peer, but it's still connecting... */
1349 LASSERT (peer2->ibp_connecting != 0 ||
1350 peer2->ibp_accepting != 0);
1352 cfs_list_add_tail(&tx->tx_list,
1353 &peer2->ibp_tx_queue);
1354 cfs_write_unlock_irqrestore(g_lock, flags);
1356 conn = kiblnd_get_conn_locked(peer2);
1357 kiblnd_conn_addref(conn); /* 1 ref for me... */
1359 cfs_write_unlock_irqrestore(g_lock, flags);
1362 kiblnd_queue_tx(tx, conn);
1363 kiblnd_conn_decref(conn); /* ...to here */
1366 kiblnd_peer_decref(peer);
1370 /* Brand new peer */
1371 LASSERT (peer->ibp_connecting == 0);
1372 peer->ibp_connecting = 1;
1374 /* always called with a ref on ni, which prevents ni being shutdown */
1375 LASSERT (((kib_net_t *)ni->ni_data)->ibn_shutdown == 0);
1378 cfs_list_add_tail(&tx->tx_list, &peer->ibp_tx_queue);
1380 kiblnd_peer_addref(peer);
1381 cfs_list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid));
1383 cfs_write_unlock_irqrestore(g_lock, flags);
1385 kiblnd_connect_peer(peer);
1386 kiblnd_peer_decref(peer);
1390 kiblnd_send (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg)
1392 lnet_hdr_t *hdr = &lntmsg->msg_hdr;
1393 int type = lntmsg->msg_type;
1394 lnet_process_id_t target = lntmsg->msg_target;
1395 int target_is_router = lntmsg->msg_target_is_router;
1396 int routing = lntmsg->msg_routing;
1397 unsigned int payload_niov = lntmsg->msg_niov;
1398 struct iovec *payload_iov = lntmsg->msg_iov;
1399 lnet_kiov_t *payload_kiov = lntmsg->msg_kiov;
1400 unsigned int payload_offset = lntmsg->msg_offset;
1401 unsigned int payload_nob = lntmsg->msg_len;
1407 /* NB 'private' is different depending on what we're sending.... */
1409 CDEBUG(D_NET, "sending %d bytes in %d frags to %s\n",
1410 payload_nob, payload_niov, libcfs_id2str(target));
1412 LASSERT (payload_nob == 0 || payload_niov > 0);
1413 LASSERT (payload_niov <= LNET_MAX_IOV);
1415 /* Thread context */
1416 LASSERT (!cfs_in_interrupt());
1417 /* payload is either all vaddrs or all pages */
1418 LASSERT (!(payload_kiov != NULL && payload_iov != NULL));
1426 LASSERT (payload_nob == 0);
1430 if (routing || target_is_router)
1431 break; /* send IMMEDIATE */
1433 /* is the REPLY message too small for RDMA? */
1434 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[lntmsg->msg_md->md_length]);
1435 if (nob <= IBLND_MSG_SIZE)
1436 break; /* send IMMEDIATE */
1438 tx = kiblnd_get_idle_tx(ni);
1440 CERROR("Can't allocate txd for GET to %s: \n",
1441 libcfs_nid2str(target.nid));
1447 if ((lntmsg->msg_md->md_options & LNET_MD_KIOV) == 0)
1448 rc = kiblnd_setup_rd_iov(ni, tx,
1449 &ibmsg->ibm_u.get.ibgm_rd,
1450 lntmsg->msg_md->md_niov,
1451 lntmsg->msg_md->md_iov.iov,
1452 0, lntmsg->msg_md->md_length);
1454 rc = kiblnd_setup_rd_kiov(ni, tx,
1455 &ibmsg->ibm_u.get.ibgm_rd,
1456 lntmsg->msg_md->md_niov,
1457 lntmsg->msg_md->md_iov.kiov,
1458 0, lntmsg->msg_md->md_length);
1460 CERROR("Can't setup GET sink for %s: %d\n",
1461 libcfs_nid2str(target.nid), rc);
1462 kiblnd_tx_done(ni, tx);
1466 nob = offsetof(kib_get_msg_t, ibgm_rd.rd_frags[tx->tx_nfrags]);
1467 ibmsg->ibm_u.get.ibgm_cookie = tx->tx_cookie;
1468 ibmsg->ibm_u.get.ibgm_hdr = *hdr;
1470 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_GET_REQ, nob);
1472 tx->tx_lntmsg[1] = lnet_create_reply_msg(ni, lntmsg);
1473 if (tx->tx_lntmsg[1] == NULL) {
1474 CERROR("Can't create reply for GET -> %s\n",
1475 libcfs_nid2str(target.nid));
1476 kiblnd_tx_done(ni, tx);
1480 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg[0,1] on completion */
1481 tx->tx_waiting = 1; /* waiting for GET_DONE */
1482 kiblnd_launch_tx(ni, tx, target.nid);
1485 case LNET_MSG_REPLY:
1487 /* Is the payload small enough not to need RDMA? */
1488 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob]);
1489 if (nob <= IBLND_MSG_SIZE)
1490 break; /* send IMMEDIATE */
1492 tx = kiblnd_get_idle_tx(ni);
1494 CERROR("Can't allocate %s txd for %s\n",
1495 type == LNET_MSG_PUT ? "PUT" : "REPLY",
1496 libcfs_nid2str(target.nid));
1500 if (payload_kiov == NULL)
1501 rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1502 payload_niov, payload_iov,
1503 payload_offset, payload_nob);
1505 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1506 payload_niov, payload_kiov,
1507 payload_offset, payload_nob);
1509 CERROR("Can't setup PUT src for %s: %d\n",
1510 libcfs_nid2str(target.nid), rc);
1511 kiblnd_tx_done(ni, tx);
1516 ibmsg->ibm_u.putreq.ibprm_hdr = *hdr;
1517 ibmsg->ibm_u.putreq.ibprm_cookie = tx->tx_cookie;
1518 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_REQ, sizeof(kib_putreq_msg_t));
1520 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1521 tx->tx_waiting = 1; /* waiting for PUT_{ACK,NAK} */
1522 kiblnd_launch_tx(ni, tx, target.nid);
1526 /* send IMMEDIATE */
1528 LASSERT (offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob])
1531 tx = kiblnd_get_idle_tx(ni);
1533 CERROR ("Can't send %d to %s: tx descs exhausted\n",
1534 type, libcfs_nid2str(target.nid));
1539 ibmsg->ibm_u.immediate.ibim_hdr = *hdr;
1541 if (payload_kiov != NULL)
1542 lnet_copy_kiov2flat(IBLND_MSG_SIZE, ibmsg,
1543 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1544 payload_niov, payload_kiov,
1545 payload_offset, payload_nob);
1547 lnet_copy_iov2flat(IBLND_MSG_SIZE, ibmsg,
1548 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1549 payload_niov, payload_iov,
1550 payload_offset, payload_nob);
1552 nob = offsetof(kib_immediate_msg_t, ibim_payload[payload_nob]);
1553 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_IMMEDIATE, nob);
1555 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1556 kiblnd_launch_tx(ni, tx, target.nid);
1561 kiblnd_reply (lnet_ni_t *ni, kib_rx_t *rx, lnet_msg_t *lntmsg)
1563 lnet_process_id_t target = lntmsg->msg_target;
1564 unsigned int niov = lntmsg->msg_niov;
1565 struct iovec *iov = lntmsg->msg_iov;
1566 lnet_kiov_t *kiov = lntmsg->msg_kiov;
1567 unsigned int offset = lntmsg->msg_offset;
1568 unsigned int nob = lntmsg->msg_len;
1572 tx = kiblnd_get_idle_tx(ni);
1574 CERROR("Can't get tx for REPLY to %s\n",
1575 libcfs_nid2str(target.nid));
1581 else if (kiov == NULL)
1582 rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1583 niov, iov, offset, nob);
1585 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1586 niov, kiov, offset, nob);
1589 CERROR("Can't setup GET src for %s: %d\n",
1590 libcfs_nid2str(target.nid), rc);
1594 rc = kiblnd_init_rdma(rx->rx_conn, tx,
1595 IBLND_MSG_GET_DONE, nob,
1596 &rx->rx_msg->ibm_u.get.ibgm_rd,
1597 rx->rx_msg->ibm_u.get.ibgm_cookie);
1599 CERROR("Can't setup rdma for GET from %s: %d\n",
1600 libcfs_nid2str(target.nid), rc);
1605 /* No RDMA: local completion may happen now! */
1606 lnet_finalize(ni, lntmsg, 0);
1608 /* RDMA: lnet_finalize(lntmsg) when it
1610 tx->tx_lntmsg[0] = lntmsg;
1613 kiblnd_queue_tx(tx, rx->rx_conn);
1617 kiblnd_tx_done(ni, tx);
1619 lnet_finalize(ni, lntmsg, -EIO);
1623 kiblnd_recv (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed,
1624 unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov,
1625 unsigned int offset, unsigned int mlen, unsigned int rlen)
1627 kib_rx_t *rx = private;
1628 kib_msg_t *rxmsg = rx->rx_msg;
1629 kib_conn_t *conn = rx->rx_conn;
1633 int post_credit = IBLND_POSTRX_PEER_CREDIT;
1636 LASSERT (mlen <= rlen);
1637 LASSERT (!cfs_in_interrupt());
1638 /* Either all pages or all vaddrs */
1639 LASSERT (!(kiov != NULL && iov != NULL));
1641 switch (rxmsg->ibm_type) {
1645 case IBLND_MSG_IMMEDIATE:
1646 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[rlen]);
1647 if (nob > rx->rx_nob) {
1648 CERROR ("Immediate message from %s too big: %d(%d)\n",
1649 libcfs_nid2str(rxmsg->ibm_u.immediate.ibim_hdr.src_nid),
1656 lnet_copy_flat2kiov(niov, kiov, offset,
1657 IBLND_MSG_SIZE, rxmsg,
1658 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1661 lnet_copy_flat2iov(niov, iov, offset,
1662 IBLND_MSG_SIZE, rxmsg,
1663 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1665 lnet_finalize (ni, lntmsg, 0);
1668 case IBLND_MSG_PUT_REQ:
1670 lnet_finalize(ni, lntmsg, 0);
1671 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK, 0,
1672 rxmsg->ibm_u.putreq.ibprm_cookie);
1676 tx = kiblnd_get_idle_tx(ni);
1678 CERROR("Can't allocate tx for %s\n",
1679 libcfs_nid2str(conn->ibc_peer->ibp_nid));
1680 /* Not replying will break the connection */
1687 rc = kiblnd_setup_rd_iov(ni, tx,
1688 &txmsg->ibm_u.putack.ibpam_rd,
1689 niov, iov, offset, mlen);
1691 rc = kiblnd_setup_rd_kiov(ni, tx,
1692 &txmsg->ibm_u.putack.ibpam_rd,
1693 niov, kiov, offset, mlen);
1695 CERROR("Can't setup PUT sink for %s: %d\n",
1696 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
1697 kiblnd_tx_done(ni, tx);
1698 /* tell peer it's over */
1699 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK, rc,
1700 rxmsg->ibm_u.putreq.ibprm_cookie);
1704 nob = offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[tx->tx_nfrags]);
1705 txmsg->ibm_u.putack.ibpam_src_cookie = rxmsg->ibm_u.putreq.ibprm_cookie;
1706 txmsg->ibm_u.putack.ibpam_dst_cookie = tx->tx_cookie;
1708 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_ACK, nob);
1710 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1711 tx->tx_waiting = 1; /* waiting for PUT_DONE */
1712 kiblnd_queue_tx(tx, conn);
1714 /* reposted buffer reserved for PUT_DONE */
1715 post_credit = IBLND_POSTRX_NO_CREDIT;
1718 case IBLND_MSG_GET_REQ:
1719 if (lntmsg != NULL) {
1720 /* Optimized GET; RDMA lntmsg's payload */
1721 kiblnd_reply(ni, rx, lntmsg);
1723 /* GET didn't match anything */
1724 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_GET_DONE,
1726 rxmsg->ibm_u.get.ibgm_cookie);
1731 kiblnd_post_rx(rx, post_credit);
1736 kiblnd_thread_start (int (*fn)(void *arg), void *arg)
1738 long pid = cfs_kernel_thread (fn, arg, 0);
1743 cfs_atomic_inc (&kiblnd_data.kib_nthreads);
1748 kiblnd_thread_fini (void)
1750 cfs_atomic_dec (&kiblnd_data.kib_nthreads);
1754 kiblnd_peer_alive (kib_peer_t *peer)
1756 /* This is racy, but everyone's only writing cfs_time_current() */
1757 peer->ibp_last_alive = cfs_time_current();
1762 kiblnd_peer_notify (kib_peer_t *peer)
1765 cfs_time_t last_alive = 0;
1766 unsigned long flags;
1768 cfs_read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1770 if (cfs_list_empty(&peer->ibp_conns) &&
1771 peer->ibp_accepting == 0 &&
1772 peer->ibp_connecting == 0 &&
1773 peer->ibp_error != 0) {
1774 error = peer->ibp_error;
1775 peer->ibp_error = 0;
1777 last_alive = peer->ibp_last_alive;
1780 cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1783 lnet_notify(peer->ibp_ni,
1784 peer->ibp_nid, 0, last_alive);
1788 kiblnd_close_conn_locked (kib_conn_t *conn, int error)
1790 /* This just does the immediate housekeeping. 'error' is zero for a
1791 * normal shutdown which can happen only after the connection has been
1792 * established. If the connection is established, schedule the
1793 * connection to be finished off by the connd. Otherwise the connd is
1794 * already dealing with it (either to set it up or tear it down).
1795 * Caller holds kib_global_lock exclusively in irq context */
1796 kib_peer_t *peer = conn->ibc_peer;
1798 unsigned long flags;
1800 LASSERT (error != 0 || conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1802 if (error != 0 && conn->ibc_comms_error == 0)
1803 conn->ibc_comms_error = error;
1805 if (conn->ibc_state != IBLND_CONN_ESTABLISHED)
1806 return; /* already being handled */
1809 cfs_list_empty(&conn->ibc_tx_noops) &&
1810 cfs_list_empty(&conn->ibc_tx_queue) &&
1811 cfs_list_empty(&conn->ibc_tx_queue_rsrvd) &&
1812 cfs_list_empty(&conn->ibc_tx_queue_nocred) &&
1813 cfs_list_empty(&conn->ibc_active_txs)) {
1814 CDEBUG(D_NET, "closing conn to %s\n",
1815 libcfs_nid2str(peer->ibp_nid));
1817 CNETERR("Closing conn to %s: error %d%s%s%s%s%s\n",
1818 libcfs_nid2str(peer->ibp_nid), error,
1819 cfs_list_empty(&conn->ibc_tx_queue) ? "" : "(sending)",
1820 cfs_list_empty(&conn->ibc_tx_noops) ? "" : "(sending_noops)",
1821 cfs_list_empty(&conn->ibc_tx_queue_rsrvd) ? "" : "(sending_rsrvd)",
1822 cfs_list_empty(&conn->ibc_tx_queue_nocred) ? "" : "(sending_nocred)",
1823 cfs_list_empty(&conn->ibc_active_txs) ? "" : "(waiting)");
1826 dev = ((kib_net_t *)peer->ibp_ni->ni_data)->ibn_dev;
1827 cfs_list_del(&conn->ibc_list);
1828 /* connd (see below) takes over ibc_list's ref */
1830 if (cfs_list_empty (&peer->ibp_conns) && /* no more conns */
1831 kiblnd_peer_active(peer)) { /* still in peer table */
1832 kiblnd_unlink_peer_locked(peer);
1834 /* set/clear error on last conn */
1835 peer->ibp_error = conn->ibc_comms_error;
1838 kiblnd_set_conn_state(conn, IBLND_CONN_CLOSING);
1841 kiblnd_dev_can_failover(dev)) {
1842 cfs_list_add_tail(&dev->ibd_fail_list,
1843 &kiblnd_data.kib_failed_devs);
1844 cfs_waitq_signal(&kiblnd_data.kib_failover_waitq);
1847 cfs_spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
1849 cfs_list_add_tail (&conn->ibc_list, &kiblnd_data.kib_connd_conns);
1850 cfs_waitq_signal (&kiblnd_data.kib_connd_waitq);
1852 cfs_spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
1856 kiblnd_close_conn (kib_conn_t *conn, int error)
1858 unsigned long flags;
1860 cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1862 kiblnd_close_conn_locked(conn, error);
1864 cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1868 kiblnd_handle_early_rxs(kib_conn_t *conn)
1870 unsigned long flags;
1873 LASSERT (!cfs_in_interrupt());
1874 LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1876 cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1877 while (!cfs_list_empty(&conn->ibc_early_rxs)) {
1878 rx = cfs_list_entry(conn->ibc_early_rxs.next,
1880 cfs_list_del(&rx->rx_list);
1881 cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock,
1884 kiblnd_handle_rx(rx);
1886 cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1888 cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1892 kiblnd_abort_txs(kib_conn_t *conn, cfs_list_t *txs)
1894 CFS_LIST_HEAD (zombies);
1899 cfs_spin_lock(&conn->ibc_lock);
1901 cfs_list_for_each_safe (tmp, nxt, txs) {
1902 tx = cfs_list_entry (tmp, kib_tx_t, tx_list);
1904 if (txs == &conn->ibc_active_txs) {
1905 LASSERT (!tx->tx_queued);
1906 LASSERT (tx->tx_waiting ||
1907 tx->tx_sending != 0);
1909 LASSERT (tx->tx_queued);
1912 tx->tx_status = -ECONNABORTED;
1915 if (tx->tx_sending == 0) {
1917 cfs_list_del (&tx->tx_list);
1918 cfs_list_add (&tx->tx_list, &zombies);
1922 cfs_spin_unlock(&conn->ibc_lock);
1924 kiblnd_txlist_done(conn->ibc_peer->ibp_ni,
1925 &zombies, -ECONNABORTED);
1929 kiblnd_finalise_conn (kib_conn_t *conn)
1931 LASSERT (!cfs_in_interrupt());
1932 LASSERT (conn->ibc_state > IBLND_CONN_INIT);
1934 kiblnd_set_conn_state(conn, IBLND_CONN_DISCONNECTED);
1936 /* abort_receives moves QP state to IB_QPS_ERR. This is only required
1937 * for connections that didn't get as far as being connected, because
1938 * rdma_disconnect() does this for free. */
1939 kiblnd_abort_receives(conn);
1941 /* Complete all tx descs not waiting for sends to complete.
1942 * NB we should be safe from RDMA now that the QP has changed state */
1944 kiblnd_abort_txs(conn, &conn->ibc_tx_noops);
1945 kiblnd_abort_txs(conn, &conn->ibc_tx_queue);
1946 kiblnd_abort_txs(conn, &conn->ibc_tx_queue_rsrvd);
1947 kiblnd_abort_txs(conn, &conn->ibc_tx_queue_nocred);
1948 kiblnd_abort_txs(conn, &conn->ibc_active_txs);
1950 kiblnd_handle_early_rxs(conn);
1954 kiblnd_peer_connect_failed (kib_peer_t *peer, int active, int error)
1956 CFS_LIST_HEAD (zombies);
1957 unsigned long flags;
1959 LASSERT (error != 0);
1960 LASSERT (!cfs_in_interrupt());
1962 cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1965 LASSERT (peer->ibp_connecting > 0);
1966 peer->ibp_connecting--;
1968 LASSERT (peer->ibp_accepting > 0);
1969 peer->ibp_accepting--;
1972 if (peer->ibp_connecting != 0 ||
1973 peer->ibp_accepting != 0) {
1974 /* another connection attempt under way... */
1975 cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock,
1980 if (cfs_list_empty(&peer->ibp_conns)) {
1981 /* Take peer's blocked transmits to complete with error */
1982 cfs_list_add(&zombies, &peer->ibp_tx_queue);
1983 cfs_list_del_init(&peer->ibp_tx_queue);
1985 if (kiblnd_peer_active(peer))
1986 kiblnd_unlink_peer_locked(peer);
1988 peer->ibp_error = error;
1990 /* Can't have blocked transmits if there are connections */
1991 LASSERT (cfs_list_empty(&peer->ibp_tx_queue));
1994 cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1996 kiblnd_peer_notify(peer);
1998 if (cfs_list_empty (&zombies))
2001 CNETERR("Deleting messages for %s: connection failed\n",
2002 libcfs_nid2str(peer->ibp_nid));
2004 kiblnd_txlist_done(peer->ibp_ni, &zombies, -EHOSTUNREACH);
2008 kiblnd_connreq_done(kib_conn_t *conn, int status)
2010 kib_peer_t *peer = conn->ibc_peer;
2013 unsigned long flags;
2016 active = (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2018 CDEBUG(D_NET,"%s: active(%d), version(%x), status(%d)\n",
2019 libcfs_nid2str(peer->ibp_nid), active,
2020 conn->ibc_version, status);
2022 LASSERT (!cfs_in_interrupt());
2023 LASSERT ((conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT &&
2024 peer->ibp_connecting > 0) ||
2025 (conn->ibc_state == IBLND_CONN_PASSIVE_WAIT &&
2026 peer->ibp_accepting > 0));
2028 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
2029 conn->ibc_connvars = NULL;
2032 /* failed to establish connection */
2033 kiblnd_peer_connect_failed(peer, active, status);
2034 kiblnd_finalise_conn(conn);
2038 /* connection established */
2039 cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2041 conn->ibc_last_send = jiffies;
2042 kiblnd_set_conn_state(conn, IBLND_CONN_ESTABLISHED);
2043 kiblnd_peer_alive(peer);
2045 /* Add conn to peer's list and nuke any dangling conns from a different
2046 * peer instance... */
2047 kiblnd_conn_addref(conn); /* +1 ref for ibc_list */
2048 cfs_list_add(&conn->ibc_list, &peer->ibp_conns);
2050 peer->ibp_connecting--;
2052 peer->ibp_accepting--;
2054 if (peer->ibp_version == 0) {
2055 peer->ibp_version = conn->ibc_version;
2056 peer->ibp_incarnation = conn->ibc_incarnation;
2059 if (peer->ibp_version != conn->ibc_version ||
2060 peer->ibp_incarnation != conn->ibc_incarnation) {
2061 kiblnd_close_stale_conns_locked(peer, conn->ibc_version,
2062 conn->ibc_incarnation);
2063 peer->ibp_version = conn->ibc_version;
2064 peer->ibp_incarnation = conn->ibc_incarnation;
2067 /* grab pending txs while I have the lock */
2068 cfs_list_add(&txs, &peer->ibp_tx_queue);
2069 cfs_list_del_init(&peer->ibp_tx_queue);
2071 if (!kiblnd_peer_active(peer) || /* peer has been deleted */
2072 conn->ibc_comms_error != 0) { /* error has happened already */
2073 lnet_ni_t *ni = peer->ibp_ni;
2075 /* start to shut down connection */
2076 kiblnd_close_conn_locked(conn, -ECONNABORTED);
2077 cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2080 kiblnd_txlist_done(ni, &txs, -ECONNABORTED);
2085 cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2087 /* Schedule blocked txs */
2088 cfs_spin_lock (&conn->ibc_lock);
2089 while (!cfs_list_empty (&txs)) {
2090 tx = cfs_list_entry (txs.next, kib_tx_t, tx_list);
2091 cfs_list_del(&tx->tx_list);
2093 kiblnd_queue_tx_locked(tx, conn);
2095 cfs_spin_unlock (&conn->ibc_lock);
2097 kiblnd_check_sends(conn);
2099 /* schedule blocked rxs */
2100 kiblnd_handle_early_rxs(conn);
2104 kiblnd_reject(struct rdma_cm_id *cmid, kib_rej_t *rej)
2108 rc = rdma_reject(cmid, rej, sizeof(*rej));
2111 CWARN("Error %d sending reject\n", rc);
2115 kiblnd_passive_connect (struct rdma_cm_id *cmid, void *priv, int priv_nob)
2117 cfs_rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2118 kib_msg_t *reqmsg = priv;
2124 lnet_ni_t *ni = NULL;
2125 kib_net_t *net = NULL;
2127 struct rdma_conn_param cp;
2129 int version = IBLND_MSG_VERSION;
2130 unsigned long flags;
2133 LASSERT (!cfs_in_interrupt());
2135 /* cmid inherits 'context' from the corresponding listener id */
2136 ibdev = (kib_dev_t *)cmid->context;
2137 LASSERT (ibdev != NULL);
2139 memset(&rej, 0, sizeof(rej));
2140 rej.ibr_magic = IBLND_MSG_MAGIC;
2141 rej.ibr_why = IBLND_REJECT_FATAL;
2142 rej.ibr_cp.ibcp_max_msg_size = IBLND_MSG_SIZE;
2144 if (priv_nob < offsetof(kib_msg_t, ibm_type)) {
2145 CERROR("Short connection request\n");
2149 /* Future protocol version compatibility support! If the
2150 * o2iblnd-specific protocol changes, or when LNET unifies
2151 * protocols over all LNDs, the initial connection will
2152 * negotiate a protocol version. I trap this here to avoid
2153 * console errors; the reject tells the peer which protocol I
2155 if (reqmsg->ibm_magic == LNET_PROTO_MAGIC ||
2156 reqmsg->ibm_magic == __swab32(LNET_PROTO_MAGIC))
2158 if (reqmsg->ibm_magic == IBLND_MSG_MAGIC &&
2159 reqmsg->ibm_version != IBLND_MSG_VERSION &&
2160 reqmsg->ibm_version != IBLND_MSG_VERSION_1)
2162 if (reqmsg->ibm_magic == __swab32(IBLND_MSG_MAGIC) &&
2163 reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION) &&
2164 reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION_1))
2167 rc = kiblnd_unpack_msg(reqmsg, priv_nob);
2169 CERROR("Can't parse connection request: %d\n", rc);
2173 nid = reqmsg->ibm_srcnid;
2174 ni = lnet_net2ni(LNET_NIDNET(reqmsg->ibm_dstnid));
2177 net = (kib_net_t *)ni->ni_data;
2178 rej.ibr_incarnation = net->ibn_incarnation;
2181 if (ni == NULL || /* no matching net */
2182 ni->ni_nid != reqmsg->ibm_dstnid || /* right NET, wrong NID! */
2183 net->ibn_dev != ibdev) { /* wrong device */
2184 CERROR("Can't accept %s on %s (%s:%d:%u.%u.%u.%u): "
2185 "bad dst nid %s\n", libcfs_nid2str(nid),
2186 ni == NULL ? "NA" : libcfs_nid2str(ni->ni_nid),
2187 ibdev->ibd_ifname, ibdev->ibd_nnets,
2188 HIPQUAD(ibdev->ibd_ifip),
2189 libcfs_nid2str(reqmsg->ibm_dstnid));
2194 /* check time stamp as soon as possible */
2195 if (reqmsg->ibm_dststamp != 0 &&
2196 reqmsg->ibm_dststamp != net->ibn_incarnation) {
2197 CWARN("Stale connection request\n");
2198 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2202 /* I can accept peer's version */
2203 version = reqmsg->ibm_version;
2205 if (reqmsg->ibm_type != IBLND_MSG_CONNREQ) {
2206 CERROR("Unexpected connreq msg type: %x from %s\n",
2207 reqmsg->ibm_type, libcfs_nid2str(nid));
2211 if (reqmsg->ibm_u.connparams.ibcp_queue_depth !=
2212 IBLND_MSG_QUEUE_SIZE(version)) {
2213 CERROR("Can't accept %s: incompatible queue depth %d (%d wanted)\n",
2214 libcfs_nid2str(nid), reqmsg->ibm_u.connparams.ibcp_queue_depth,
2215 IBLND_MSG_QUEUE_SIZE(version));
2217 if (version == IBLND_MSG_VERSION)
2218 rej.ibr_why = IBLND_REJECT_MSG_QUEUE_SIZE;
2223 if (reqmsg->ibm_u.connparams.ibcp_max_frags !=
2224 IBLND_RDMA_FRAGS(version)) {
2225 CERROR("Can't accept %s(version %x): "
2226 "incompatible max_frags %d (%d wanted)\n",
2227 libcfs_nid2str(nid), version,
2228 reqmsg->ibm_u.connparams.ibcp_max_frags,
2229 IBLND_RDMA_FRAGS(version));
2231 if (version == IBLND_MSG_VERSION)
2232 rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2238 if (reqmsg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2239 CERROR("Can't accept %s: message size %d too big (%d max)\n",
2240 libcfs_nid2str(nid),
2241 reqmsg->ibm_u.connparams.ibcp_max_msg_size,
2246 /* assume 'nid' is a new peer; create */
2247 rc = kiblnd_create_peer(ni, &peer, nid);
2249 CERROR("Can't create peer for %s\n", libcfs_nid2str(nid));
2250 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2254 cfs_write_lock_irqsave(g_lock, flags);
2256 peer2 = kiblnd_find_peer_locked(nid);
2257 if (peer2 != NULL) {
2258 if (peer2->ibp_version == 0) {
2259 peer2->ibp_version = version;
2260 peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2263 /* not the guy I've talked with */
2264 if (peer2->ibp_incarnation != reqmsg->ibm_srcstamp ||
2265 peer2->ibp_version != version) {
2266 kiblnd_close_peer_conns_locked(peer2, -ESTALE);
2267 cfs_write_unlock_irqrestore(g_lock, flags);
2269 CWARN("Conn stale %s [old ver: %x, new ver: %x]\n",
2270 libcfs_nid2str(nid), peer2->ibp_version, version);
2272 kiblnd_peer_decref(peer);
2273 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2277 /* tie-break connection race in favour of the higher NID */
2278 if (peer2->ibp_connecting != 0 &&
2280 cfs_write_unlock_irqrestore(g_lock, flags);
2282 CWARN("Conn race %s\n", libcfs_nid2str(peer2->ibp_nid));
2284 kiblnd_peer_decref(peer);
2285 rej.ibr_why = IBLND_REJECT_CONN_RACE;
2289 peer2->ibp_accepting++;
2290 kiblnd_peer_addref(peer2);
2292 cfs_write_unlock_irqrestore(g_lock, flags);
2293 kiblnd_peer_decref(peer);
2296 /* Brand new peer */
2297 LASSERT (peer->ibp_accepting == 0);
2298 LASSERT (peer->ibp_version == 0 &&
2299 peer->ibp_incarnation == 0);
2301 peer->ibp_accepting = 1;
2302 peer->ibp_version = version;
2303 peer->ibp_incarnation = reqmsg->ibm_srcstamp;
2305 /* I have a ref on ni that prevents it being shutdown */
2306 LASSERT (net->ibn_shutdown == 0);
2308 kiblnd_peer_addref(peer);
2309 cfs_list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid));
2311 cfs_write_unlock_irqrestore(g_lock, flags);
2314 conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_PASSIVE_WAIT, version);
2316 kiblnd_peer_connect_failed(peer, 0, -ENOMEM);
2317 kiblnd_peer_decref(peer);
2318 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2322 /* conn now "owns" cmid, so I return success from here on to ensure the
2323 * CM callback doesn't destroy cmid. */
2325 conn->ibc_incarnation = reqmsg->ibm_srcstamp;
2326 conn->ibc_credits = IBLND_MSG_QUEUE_SIZE(version);
2327 conn->ibc_reserved_credits = IBLND_MSG_QUEUE_SIZE(version);
2328 LASSERT (conn->ibc_credits + conn->ibc_reserved_credits + IBLND_OOB_MSGS(version)
2329 <= IBLND_RX_MSGS(version));
2331 ackmsg = &conn->ibc_connvars->cv_msg;
2332 memset(ackmsg, 0, sizeof(*ackmsg));
2334 kiblnd_init_msg(ackmsg, IBLND_MSG_CONNACK,
2335 sizeof(ackmsg->ibm_u.connparams));
2336 ackmsg->ibm_u.connparams.ibcp_queue_depth = IBLND_MSG_QUEUE_SIZE(version);
2337 ackmsg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2338 ackmsg->ibm_u.connparams.ibcp_max_frags = IBLND_RDMA_FRAGS(version);
2340 kiblnd_pack_msg(ni, ackmsg, version, 0, nid, reqmsg->ibm_srcstamp);
2342 memset(&cp, 0, sizeof(cp));
2343 cp.private_data = ackmsg;
2344 cp.private_data_len = ackmsg->ibm_nob;
2345 cp.responder_resources = 0; /* No atomic ops or RDMA reads */
2346 cp.initiator_depth = 0;
2347 cp.flow_control = 1;
2348 cp.retry_count = *kiblnd_tunables.kib_retry_count;
2349 cp.rnr_retry_count = *kiblnd_tunables.kib_rnr_retry_count;
2351 CDEBUG(D_NET, "Accept %s\n", libcfs_nid2str(nid));
2353 rc = rdma_accept(cmid, &cp);
2355 CERROR("Can't accept %s: %d\n", libcfs_nid2str(nid), rc);
2356 rej.ibr_version = version;
2357 rej.ibr_why = IBLND_REJECT_FATAL;
2359 kiblnd_reject(cmid, &rej);
2360 kiblnd_connreq_done(conn, rc);
2361 kiblnd_conn_decref(conn);
2371 rej.ibr_version = version;
2372 rej.ibr_cp.ibcp_queue_depth = IBLND_MSG_QUEUE_SIZE(version);
2373 rej.ibr_cp.ibcp_max_frags = IBLND_RDMA_FRAGS(version);
2374 kiblnd_reject(cmid, &rej);
2376 return -ECONNREFUSED;
2380 kiblnd_reconnect (kib_conn_t *conn, int version,
2381 __u64 incarnation, int why, kib_connparams_t *cp)
2383 kib_peer_t *peer = conn->ibc_peer;
2386 unsigned long flags;
2388 LASSERT (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2389 LASSERT (peer->ibp_connecting > 0); /* 'conn' at least */
2391 cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2393 /* retry connection if it's still needed and no other connection
2394 * attempts (active or passive) are in progress
2395 * NB: reconnect is still needed even when ibp_tx_queue is
2396 * empty if ibp_version != version because reconnect may be
2397 * initiated by kiblnd_query() */
2398 if ((!cfs_list_empty(&peer->ibp_tx_queue) ||
2399 peer->ibp_version != version) &&
2400 peer->ibp_connecting == 1 &&
2401 peer->ibp_accepting == 0) {
2403 peer->ibp_connecting++;
2405 peer->ibp_version = version;
2406 peer->ibp_incarnation = incarnation;
2409 cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2419 case IBLND_REJECT_CONN_STALE:
2423 case IBLND_REJECT_CONN_RACE:
2424 reason = "conn race";
2427 case IBLND_REJECT_CONN_UNCOMPAT:
2428 reason = "version negotiation";
2432 CNETERR("%s: retrying (%s), %x, %x, "
2433 "queue_dep: %d, max_frag: %d, msg_size: %d\n",
2434 libcfs_nid2str(peer->ibp_nid),
2435 reason, IBLND_MSG_VERSION, version,
2436 cp != NULL? cp->ibcp_queue_depth :IBLND_MSG_QUEUE_SIZE(version),
2437 cp != NULL? cp->ibcp_max_frags : IBLND_RDMA_FRAGS(version),
2438 cp != NULL? cp->ibcp_max_msg_size: IBLND_MSG_SIZE);
2440 kiblnd_connect_peer(peer);
2444 kiblnd_rejected (kib_conn_t *conn, int reason, void *priv, int priv_nob)
2446 kib_peer_t *peer = conn->ibc_peer;
2448 LASSERT (!cfs_in_interrupt());
2449 LASSERT (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2452 case IB_CM_REJ_STALE_CONN:
2453 kiblnd_reconnect(conn, IBLND_MSG_VERSION, 0,
2454 IBLND_REJECT_CONN_STALE, NULL);
2457 case IB_CM_REJ_INVALID_SERVICE_ID:
2458 CNETERR("%s rejected: no listener at %d\n",
2459 libcfs_nid2str(peer->ibp_nid),
2460 *kiblnd_tunables.kib_service);
2463 case IB_CM_REJ_CONSUMER_DEFINED:
2464 if (priv_nob >= offsetof(kib_rej_t, ibr_padding)) {
2465 kib_rej_t *rej = priv;
2466 kib_connparams_t *cp = NULL;
2468 __u64 incarnation = -1;
2470 /* NB. default incarnation is -1 because:
2471 * a) V1 will ignore dst incarnation in connreq.
2472 * b) V2 will provide incarnation while rejecting me,
2473 * -1 will be overwrote.
2475 * if I try to connect to a V1 peer with V2 protocol,
2476 * it rejected me then upgrade to V2, I have no idea
2477 * about the upgrading and try to reconnect with V1,
2478 * in this case upgraded V2 can find out I'm trying to
2479 * talk to the old guy and reject me(incarnation is -1).
2482 if (rej->ibr_magic == __swab32(IBLND_MSG_MAGIC) ||
2483 rej->ibr_magic == __swab32(LNET_PROTO_MAGIC)) {
2484 __swab32s(&rej->ibr_magic);
2485 __swab16s(&rej->ibr_version);
2489 if (priv_nob >= sizeof(kib_rej_t) &&
2490 rej->ibr_version > IBLND_MSG_VERSION_1) {
2491 /* priv_nob is always 148 in current version
2492 * of OFED, so we still need to check version.
2493 * (define of IB_CM_REJ_PRIVATE_DATA_SIZE) */
2497 __swab64s(&rej->ibr_incarnation);
2498 __swab16s(&cp->ibcp_queue_depth);
2499 __swab16s(&cp->ibcp_max_frags);
2500 __swab32s(&cp->ibcp_max_msg_size);
2503 incarnation = rej->ibr_incarnation;
2506 if (rej->ibr_magic != IBLND_MSG_MAGIC &&
2507 rej->ibr_magic != LNET_PROTO_MAGIC) {
2508 CERROR("%s rejected: consumer defined fatal error\n",
2509 libcfs_nid2str(peer->ibp_nid));
2513 if (rej->ibr_version != IBLND_MSG_VERSION &&
2514 rej->ibr_version != IBLND_MSG_VERSION_1) {
2515 CERROR("%s rejected: o2iblnd version %x error\n",
2516 libcfs_nid2str(peer->ibp_nid),
2521 if (rej->ibr_why == IBLND_REJECT_FATAL &&
2522 rej->ibr_version == IBLND_MSG_VERSION_1) {
2523 CDEBUG(D_NET, "rejected by old version peer %s: %x\n",
2524 libcfs_nid2str(peer->ibp_nid), rej->ibr_version);
2526 if (conn->ibc_version != IBLND_MSG_VERSION_1)
2527 rej->ibr_why = IBLND_REJECT_CONN_UNCOMPAT;
2530 switch (rej->ibr_why) {
2531 case IBLND_REJECT_CONN_RACE:
2532 case IBLND_REJECT_CONN_STALE:
2533 case IBLND_REJECT_CONN_UNCOMPAT:
2534 kiblnd_reconnect(conn, rej->ibr_version,
2535 incarnation, rej->ibr_why, cp);
2538 case IBLND_REJECT_MSG_QUEUE_SIZE:
2539 CERROR("%s rejected: incompatible message queue depth %d, %d\n",
2540 libcfs_nid2str(peer->ibp_nid), cp->ibcp_queue_depth,
2541 IBLND_MSG_QUEUE_SIZE(conn->ibc_version));
2544 case IBLND_REJECT_RDMA_FRAGS:
2545 CERROR("%s rejected: incompatible # of RDMA fragments %d, %d\n",
2546 libcfs_nid2str(peer->ibp_nid), cp->ibcp_max_frags,
2547 IBLND_RDMA_FRAGS(conn->ibc_version));
2550 case IBLND_REJECT_NO_RESOURCES:
2551 CERROR("%s rejected: o2iblnd no resources\n",
2552 libcfs_nid2str(peer->ibp_nid));
2555 case IBLND_REJECT_FATAL:
2556 CERROR("%s rejected: o2iblnd fatal error\n",
2557 libcfs_nid2str(peer->ibp_nid));
2561 CERROR("%s rejected: o2iblnd reason %d\n",
2562 libcfs_nid2str(peer->ibp_nid),
2570 CNETERR("%s rejected: reason %d, size %d\n",
2571 libcfs_nid2str(peer->ibp_nid), reason, priv_nob);
2575 kiblnd_connreq_done(conn, -ECONNREFUSED);
2579 kiblnd_check_connreply (kib_conn_t *conn, void *priv, int priv_nob)
2581 kib_peer_t *peer = conn->ibc_peer;
2582 lnet_ni_t *ni = peer->ibp_ni;
2583 kib_net_t *net = ni->ni_data;
2584 kib_msg_t *msg = priv;
2585 int ver = conn->ibc_version;
2586 int rc = kiblnd_unpack_msg(msg, priv_nob);
2587 unsigned long flags;
2589 LASSERT (net != NULL);
2592 CERROR("Can't unpack connack from %s: %d\n",
2593 libcfs_nid2str(peer->ibp_nid), rc);
2597 if (msg->ibm_type != IBLND_MSG_CONNACK) {
2598 CERROR("Unexpected message %d from %s\n",
2599 msg->ibm_type, libcfs_nid2str(peer->ibp_nid));
2604 if (ver != msg->ibm_version) {
2605 CERROR("%s replied version %x is different with "
2606 "requested version %x\n",
2607 libcfs_nid2str(peer->ibp_nid), msg->ibm_version, ver);
2612 if (msg->ibm_u.connparams.ibcp_queue_depth !=
2613 IBLND_MSG_QUEUE_SIZE(ver)) {
2614 CERROR("%s has incompatible queue depth %d(%d wanted)\n",
2615 libcfs_nid2str(peer->ibp_nid),
2616 msg->ibm_u.connparams.ibcp_queue_depth,
2617 IBLND_MSG_QUEUE_SIZE(ver));
2622 if (msg->ibm_u.connparams.ibcp_max_frags !=
2623 IBLND_RDMA_FRAGS(ver)) {
2624 CERROR("%s has incompatible max_frags %d (%d wanted)\n",
2625 libcfs_nid2str(peer->ibp_nid),
2626 msg->ibm_u.connparams.ibcp_max_frags,
2627 IBLND_RDMA_FRAGS(ver));
2632 if (msg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2633 CERROR("%s max message size %d too big (%d max)\n",
2634 libcfs_nid2str(peer->ibp_nid),
2635 msg->ibm_u.connparams.ibcp_max_msg_size,
2641 cfs_read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2642 if (msg->ibm_dstnid == ni->ni_nid &&
2643 msg->ibm_dststamp == net->ibn_incarnation)
2647 cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2650 CERROR("Bad connection reply from %s, rc = %d, "
2651 "version: %x max_frags: %d\n",
2652 libcfs_nid2str(peer->ibp_nid), rc,
2653 msg->ibm_version, msg->ibm_u.connparams.ibcp_max_frags);
2657 conn->ibc_incarnation = msg->ibm_srcstamp;
2659 conn->ibc_reserved_credits = IBLND_MSG_QUEUE_SIZE(ver);
2660 LASSERT (conn->ibc_credits + conn->ibc_reserved_credits + IBLND_OOB_MSGS(ver)
2661 <= IBLND_RX_MSGS(ver));
2663 kiblnd_connreq_done(conn, 0);
2667 /* NB My QP has already established itself, so I handle anything going
2668 * wrong here by setting ibc_comms_error.
2669 * kiblnd_connreq_done(0) moves the conn state to ESTABLISHED, but then
2670 * immediately tears it down. */
2673 conn->ibc_comms_error = rc;
2674 kiblnd_connreq_done(conn, 0);
2678 kiblnd_active_connect (struct rdma_cm_id *cmid)
2680 kib_peer_t *peer = (kib_peer_t *)cmid->context;
2683 struct rdma_conn_param cp;
2686 unsigned long flags;
2689 cfs_read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2691 incarnation = peer->ibp_incarnation;
2692 version = (peer->ibp_version == 0) ? IBLND_MSG_VERSION : peer->ibp_version;
2694 cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2696 conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_ACTIVE_CONNECT, version);
2698 kiblnd_peer_connect_failed(peer, 1, -ENOMEM);
2699 kiblnd_peer_decref(peer); /* lose cmid's ref */
2703 /* conn "owns" cmid now, so I return success from here on to ensure the
2704 * CM callback doesn't destroy cmid. conn also takes over cmid's ref
2707 msg = &conn->ibc_connvars->cv_msg;
2709 memset(msg, 0, sizeof(*msg));
2710 kiblnd_init_msg(msg, IBLND_MSG_CONNREQ, sizeof(msg->ibm_u.connparams));
2711 msg->ibm_u.connparams.ibcp_queue_depth = IBLND_MSG_QUEUE_SIZE(version);
2712 msg->ibm_u.connparams.ibcp_max_frags = IBLND_RDMA_FRAGS(version);
2713 msg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2715 kiblnd_pack_msg(peer->ibp_ni, msg, version,
2716 0, peer->ibp_nid, incarnation);
2718 memset(&cp, 0, sizeof(cp));
2719 cp.private_data = msg;
2720 cp.private_data_len = msg->ibm_nob;
2721 cp.responder_resources = 0; /* No atomic ops or RDMA reads */
2722 cp.initiator_depth = 0;
2723 cp.flow_control = 1;
2724 cp.retry_count = *kiblnd_tunables.kib_retry_count;
2725 cp.rnr_retry_count = *kiblnd_tunables.kib_rnr_retry_count;
2727 LASSERT(cmid->context == (void *)conn);
2728 LASSERT(conn->ibc_cmid == cmid);
2730 rc = rdma_connect(cmid, &cp);
2732 CERROR("Can't connect to %s: %d\n",
2733 libcfs_nid2str(peer->ibp_nid), rc);
2734 kiblnd_connreq_done(conn, rc);
2735 kiblnd_conn_decref(conn);
2742 kiblnd_cm_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2748 switch (event->event) {
2750 CERROR("Unexpected event: %d, status: %d\n",
2751 event->event, event->status);
2754 case RDMA_CM_EVENT_CONNECT_REQUEST:
2755 /* destroy cmid on failure */
2756 rc = kiblnd_passive_connect(cmid,
2757 (void *)KIBLND_CONN_PARAM(event),
2758 KIBLND_CONN_PARAM_LEN(event));
2759 CDEBUG(D_NET, "connreq: %d\n", rc);
2762 case RDMA_CM_EVENT_ADDR_ERROR:
2763 peer = (kib_peer_t *)cmid->context;
2764 CNETERR("%s: ADDR ERROR %d\n",
2765 libcfs_nid2str(peer->ibp_nid), event->status);
2766 kiblnd_peer_connect_failed(peer, 1, -EHOSTUNREACH);
2767 kiblnd_peer_decref(peer);
2768 return -EHOSTUNREACH; /* rc != 0 destroys cmid */
2770 case RDMA_CM_EVENT_ADDR_RESOLVED:
2771 peer = (kib_peer_t *)cmid->context;
2773 CDEBUG(D_NET,"%s Addr resolved: %d\n",
2774 libcfs_nid2str(peer->ibp_nid), event->status);
2776 if (event->status != 0) {
2777 CNETERR("Can't resolve address for %s: %d\n",
2778 libcfs_nid2str(peer->ibp_nid), event->status);
2781 rc = rdma_resolve_route(
2782 cmid, *kiblnd_tunables.kib_timeout * 1000);
2785 /* Can't initiate route resolution */
2786 CERROR("Can't resolve route for %s: %d\n",
2787 libcfs_nid2str(peer->ibp_nid), rc);
2789 kiblnd_peer_connect_failed(peer, 1, rc);
2790 kiblnd_peer_decref(peer);
2791 return rc; /* rc != 0 destroys cmid */
2793 case RDMA_CM_EVENT_ROUTE_ERROR:
2794 peer = (kib_peer_t *)cmid->context;
2795 CNETERR("%s: ROUTE ERROR %d\n",
2796 libcfs_nid2str(peer->ibp_nid), event->status);
2797 kiblnd_peer_connect_failed(peer, 1, -EHOSTUNREACH);
2798 kiblnd_peer_decref(peer);
2799 return -EHOSTUNREACH; /* rc != 0 destroys cmid */
2801 case RDMA_CM_EVENT_ROUTE_RESOLVED:
2802 peer = (kib_peer_t *)cmid->context;
2803 CDEBUG(D_NET,"%s Route resolved: %d\n",
2804 libcfs_nid2str(peer->ibp_nid), event->status);
2806 if (event->status == 0)
2807 return kiblnd_active_connect(cmid);
2809 CNETERR("Can't resolve route for %s: %d\n",
2810 libcfs_nid2str(peer->ibp_nid), event->status);
2811 kiblnd_peer_connect_failed(peer, 1, event->status);
2812 kiblnd_peer_decref(peer);
2813 return event->status; /* rc != 0 destroys cmid */
2815 case RDMA_CM_EVENT_UNREACHABLE:
2816 conn = (kib_conn_t *)cmid->context;
2817 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
2818 conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
2819 CNETERR("%s: UNREACHABLE %d\n",
2820 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
2821 kiblnd_connreq_done(conn, -ENETDOWN);
2822 kiblnd_conn_decref(conn);
2825 case RDMA_CM_EVENT_CONNECT_ERROR:
2826 conn = (kib_conn_t *)cmid->context;
2827 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
2828 conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
2829 CNETERR("%s: CONNECT ERROR %d\n",
2830 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
2831 kiblnd_connreq_done(conn, -ENOTCONN);
2832 kiblnd_conn_decref(conn);
2835 case RDMA_CM_EVENT_REJECTED:
2836 conn = (kib_conn_t *)cmid->context;
2837 switch (conn->ibc_state) {
2841 case IBLND_CONN_PASSIVE_WAIT:
2842 CERROR ("%s: REJECTED %d\n",
2843 libcfs_nid2str(conn->ibc_peer->ibp_nid),
2845 kiblnd_connreq_done(conn, -ECONNRESET);
2848 case IBLND_CONN_ACTIVE_CONNECT:
2849 kiblnd_rejected(conn, event->status,
2850 (void *)KIBLND_CONN_PARAM(event),
2851 KIBLND_CONN_PARAM_LEN(event));
2854 kiblnd_conn_decref(conn);
2857 case RDMA_CM_EVENT_ESTABLISHED:
2858 conn = (kib_conn_t *)cmid->context;
2859 switch (conn->ibc_state) {
2863 case IBLND_CONN_PASSIVE_WAIT:
2864 CDEBUG(D_NET, "ESTABLISHED (passive): %s\n",
2865 libcfs_nid2str(conn->ibc_peer->ibp_nid));
2866 kiblnd_connreq_done(conn, 0);
2869 case IBLND_CONN_ACTIVE_CONNECT:
2870 CDEBUG(D_NET, "ESTABLISHED(active): %s\n",
2871 libcfs_nid2str(conn->ibc_peer->ibp_nid));
2872 kiblnd_check_connreply(conn,
2873 (void *)KIBLND_CONN_PARAM(event),
2874 KIBLND_CONN_PARAM_LEN(event));
2877 /* net keeps its ref on conn! */
2880 #ifdef HAVE_OFED_RDMA_CMEV_TIMEWAIT_EXIT
2881 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
2882 CDEBUG(D_NET, "Ignore TIMEWAIT_EXIT event\n");
2885 case RDMA_CM_EVENT_DISCONNECTED:
2886 conn = (kib_conn_t *)cmid->context;
2887 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
2888 CERROR("%s DISCONNECTED\n",
2889 libcfs_nid2str(conn->ibc_peer->ibp_nid));
2890 kiblnd_connreq_done(conn, -ECONNRESET);
2892 kiblnd_close_conn(conn, 0);
2894 kiblnd_conn_decref(conn);
2895 cmid->context = NULL;
2898 case RDMA_CM_EVENT_DEVICE_REMOVAL:
2899 LCONSOLE_ERROR_MSG(0x131,
2900 "Received notification of device removal\n"
2901 "Please shutdown LNET to allow this to proceed\n");
2902 /* Can't remove network from underneath LNET for now, so I have
2906 #ifdef HAVE_OFED_RDMA_CMEV_ADDRCHANGE
2907 case RDMA_CM_EVENT_ADDR_CHANGE:
2908 LCONSOLE_INFO("Physical link changed (eg hca/port)\n");
2915 kiblnd_check_txs_locked (kib_conn_t *conn, cfs_list_t *txs)
2920 cfs_list_for_each (ttmp, txs) {
2921 tx = cfs_list_entry (ttmp, kib_tx_t, tx_list);
2923 if (txs != &conn->ibc_active_txs) {
2924 LASSERT (tx->tx_queued);
2926 LASSERT (!tx->tx_queued);
2927 LASSERT (tx->tx_waiting || tx->tx_sending != 0);
2930 if (cfs_time_aftereq (jiffies, tx->tx_deadline)) {
2931 CERROR("Timed out tx: %s, %lu seconds\n",
2932 kiblnd_queue2str(conn, txs),
2933 cfs_duration_sec(jiffies - tx->tx_deadline));
2942 kiblnd_conn_timed_out_locked (kib_conn_t *conn)
2944 return kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue) ||
2945 kiblnd_check_txs_locked(conn, &conn->ibc_tx_noops) ||
2946 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_rsrvd) ||
2947 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_nocred) ||
2948 kiblnd_check_txs_locked(conn, &conn->ibc_active_txs);
2952 kiblnd_check_conns (int idx)
2954 CFS_LIST_HEAD (closes);
2955 CFS_LIST_HEAD (checksends);
2956 cfs_list_t *peers = &kiblnd_data.kib_peers[idx];
2961 unsigned long flags;
2963 /* NB. We expect to have a look at all the peers and not find any
2964 * RDMAs to time out, so we just use a shared lock while we
2966 cfs_read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2968 cfs_list_for_each (ptmp, peers) {
2969 peer = cfs_list_entry (ptmp, kib_peer_t, ibp_list);
2971 cfs_list_for_each (ctmp, &peer->ibp_conns) {
2975 conn = cfs_list_entry(ctmp, kib_conn_t, ibc_list);
2977 LASSERT (conn->ibc_state == IBLND_CONN_ESTABLISHED);
2979 cfs_spin_lock(&conn->ibc_lock);
2981 retry_noop = conn->ibc_retry_noop;
2982 timedout = kiblnd_conn_timed_out_locked(conn);
2983 if (!retry_noop && !timedout) {
2984 cfs_spin_unlock(&conn->ibc_lock);
2989 CERROR("Timed out RDMA with %s (%lu): "
2990 "c: %u, oc: %u, rc: %u\n",
2991 libcfs_nid2str(peer->ibp_nid),
2992 cfs_duration_sec(cfs_time_current() -
2993 peer->ibp_last_alive),
2995 conn->ibc_outstanding_credits,
2996 conn->ibc_reserved_credits);
2997 cfs_list_add(&conn->ibc_connd_list, &closes);
2999 cfs_list_add(&conn->ibc_connd_list, &checksends);
3001 /* +ref for 'closes' or 'checksends' */
3002 kiblnd_conn_addref(conn);
3004 cfs_spin_unlock(&conn->ibc_lock);
3008 cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3010 /* Handle timeout by closing the whole
3011 * connection. We can only be sure RDMA activity
3012 * has ceased once the QP has been modified. */
3013 while (!cfs_list_empty(&closes)) {
3014 conn = cfs_list_entry(closes.next, kib_conn_t, ibc_connd_list);
3015 cfs_list_del(&conn->ibc_connd_list);
3016 kiblnd_close_conn(conn, -ETIMEDOUT);
3017 kiblnd_conn_decref(conn);
3020 /* In case we have enough credits to return via a
3021 * NOOP, but there were no non-blocking tx descs
3022 * free to do it last time... */
3023 while (!cfs_list_empty(&checksends)) {
3024 conn = cfs_list_entry(checksends.next, kib_conn_t, ibc_connd_list);
3025 cfs_list_del(&conn->ibc_connd_list);
3026 kiblnd_check_sends(conn);
3027 kiblnd_conn_decref(conn);
3032 kiblnd_disconnect_conn (kib_conn_t *conn)
3034 LASSERT (!cfs_in_interrupt());
3035 LASSERT (current == kiblnd_data.kib_connd);
3036 LASSERT (conn->ibc_state == IBLND_CONN_CLOSING);
3038 rdma_disconnect(conn->ibc_cmid);
3039 kiblnd_finalise_conn(conn);
3041 kiblnd_peer_notify(conn->ibc_peer);
3045 kiblnd_connd (void *arg)
3047 cfs_waitlink_t wait;
3048 unsigned long flags;
3054 unsigned long deadline = jiffies;
3056 cfs_daemonize ("kiblnd_connd");
3057 cfs_block_allsigs ();
3059 cfs_waitlink_init (&wait);
3060 kiblnd_data.kib_connd = current;
3062 cfs_spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3064 while (!kiblnd_data.kib_shutdown) {
3068 if (!cfs_list_empty (&kiblnd_data.kib_connd_zombies)) {
3069 conn = cfs_list_entry(kiblnd_data. \
3070 kib_connd_zombies.next,
3071 kib_conn_t, ibc_list);
3072 cfs_list_del(&conn->ibc_list);
3074 cfs_spin_unlock_irqrestore (&kiblnd_data.kib_connd_lock,
3078 kiblnd_destroy_conn(conn);
3080 cfs_spin_lock_irqsave (&kiblnd_data.kib_connd_lock,
3084 if (!cfs_list_empty (&kiblnd_data.kib_connd_conns)) {
3085 conn = cfs_list_entry (kiblnd_data.kib_connd_conns.next,
3086 kib_conn_t, ibc_list);
3087 cfs_list_del(&conn->ibc_list);
3089 cfs_spin_unlock_irqrestore (&kiblnd_data.kib_connd_lock,
3093 kiblnd_disconnect_conn(conn);
3094 kiblnd_conn_decref(conn);
3096 cfs_spin_lock_irqsave (&kiblnd_data.kib_connd_lock,
3100 /* careful with the jiffy wrap... */
3101 timeout = (int)(deadline - jiffies);
3105 int chunk = kiblnd_data.kib_peer_hash_size;
3107 cfs_spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
3110 /* Time to check for RDMA timeouts on a few more
3111 * peers: I do checks every 'p' seconds on a
3112 * proportion of the peer table and I need to check
3113 * every connection 'n' times within a timeout
3114 * interval, to ensure I detect a timeout on any
3115 * connection within (n+1)/n times the timeout
3118 if (*kiblnd_tunables.kib_timeout > n * p)
3119 chunk = (chunk * n * p) /
3120 *kiblnd_tunables.kib_timeout;
3124 for (i = 0; i < chunk; i++) {
3125 kiblnd_check_conns(peer_index);
3126 peer_index = (peer_index + 1) %
3127 kiblnd_data.kib_peer_hash_size;
3130 deadline += p * CFS_HZ;
3131 cfs_spin_lock_irqsave(&kiblnd_data.kib_connd_lock,
3138 /* Nothing to do for 'timeout' */
3139 cfs_set_current_state (CFS_TASK_INTERRUPTIBLE);
3140 cfs_waitq_add (&kiblnd_data.kib_connd_waitq, &wait);
3141 cfs_spin_unlock_irqrestore (&kiblnd_data.kib_connd_lock, flags);
3143 cfs_waitq_timedwait(&wait, CFS_TASK_INTERRUPTIBLE, timeout);
3145 cfs_set_current_state (CFS_TASK_RUNNING);
3146 cfs_waitq_del (&kiblnd_data.kib_connd_waitq, &wait);
3147 cfs_spin_lock_irqsave (&kiblnd_data.kib_connd_lock, flags);
3150 cfs_spin_unlock_irqrestore (&kiblnd_data.kib_connd_lock, flags);
3152 kiblnd_thread_fini();
3157 kiblnd_qp_event(struct ib_event *event, void *arg)
3159 kib_conn_t *conn = arg;
3161 switch (event->event) {
3162 case IB_EVENT_COMM_EST:
3163 CDEBUG(D_NET, "%s established\n",
3164 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3168 CERROR("%s: Async QP event type %d\n",
3169 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3175 kiblnd_complete (struct ib_wc *wc)
3177 switch (kiblnd_wreqid2type(wc->wr_id)) {
3181 case IBLND_WID_RDMA:
3182 /* We only get RDMA completion notification if it fails. All
3183 * subsequent work items, including the final SEND will fail
3184 * too. However we can't print out any more info about the
3185 * failing RDMA because 'tx' might be back on the idle list or
3186 * even reused already if we didn't manage to post all our work
3188 CNETERR("RDMA (tx: %p) failed: %d\n",
3189 kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3193 kiblnd_tx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3197 kiblnd_rx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status,
3204 kiblnd_cq_completion (struct ib_cq *cq, void *arg)
3206 /* NB I'm not allowed to schedule this conn once its refcount has
3207 * reached 0. Since fundamentally I'm racing with scheduler threads
3208 * consuming my CQ I could be called after all completions have
3209 * occurred. But in this case, ibc_nrx == 0 && ibc_nsends_posted == 0
3210 * and this CQ is about to be destroyed so I NOOP. */
3211 kib_conn_t *conn = (kib_conn_t *)arg;
3212 unsigned long flags;
3214 LASSERT (cq == conn->ibc_cq);
3216 cfs_spin_lock_irqsave(&kiblnd_data.kib_sched_lock, flags);
3218 conn->ibc_ready = 1;
3220 if (!conn->ibc_scheduled &&
3221 (conn->ibc_nrx > 0 ||
3222 conn->ibc_nsends_posted > 0)) {
3223 kiblnd_conn_addref(conn); /* +1 ref for sched_conns */
3224 conn->ibc_scheduled = 1;
3225 cfs_list_add_tail(&conn->ibc_sched_list,
3226 &kiblnd_data.kib_sched_conns);
3227 cfs_waitq_signal(&kiblnd_data.kib_sched_waitq);
3230 cfs_spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock, flags);
3234 kiblnd_cq_event(struct ib_event *event, void *arg)
3236 kib_conn_t *conn = arg;
3238 CERROR("%s: async CQ event type %d\n",
3239 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3243 kiblnd_scheduler(void *arg)
3245 long id = (long)arg;
3246 cfs_waitlink_t wait;
3248 unsigned long flags;
3255 snprintf(name, sizeof(name), "kiblnd_sd_%02ld", id);
3256 cfs_daemonize(name);
3257 cfs_block_allsigs();
3259 cfs_waitlink_init(&wait);
3261 cfs_spin_lock_irqsave(&kiblnd_data.kib_sched_lock, flags);
3263 while (!kiblnd_data.kib_shutdown) {
3264 if (busy_loops++ >= IBLND_RESCHED) {
3265 cfs_spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock,
3271 cfs_spin_lock_irqsave(&kiblnd_data.kib_sched_lock,
3277 if (!cfs_list_empty(&kiblnd_data.kib_sched_conns)) {
3278 conn = cfs_list_entry(kiblnd_data.kib_sched_conns.next,
3279 kib_conn_t, ibc_sched_list);
3280 /* take over kib_sched_conns' ref on conn... */
3281 LASSERT(conn->ibc_scheduled);
3282 cfs_list_del(&conn->ibc_sched_list);
3283 conn->ibc_ready = 0;
3285 cfs_spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock,
3288 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3290 rc = ib_req_notify_cq(conn->ibc_cq,
3293 CWARN("%s: ib_req_notify_cq failed: %d, "
3294 "closing connection\n",
3295 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
3296 kiblnd_close_conn(conn, -EIO);
3297 kiblnd_conn_decref(conn);
3298 cfs_spin_lock_irqsave(&kiblnd_data. \
3304 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3308 CWARN("%s: ib_poll_cq failed: %d, "
3309 "closing connection\n",
3310 libcfs_nid2str(conn->ibc_peer->ibp_nid),
3312 kiblnd_close_conn(conn, -EIO);
3313 kiblnd_conn_decref(conn);
3314 cfs_spin_lock_irqsave(&kiblnd_data. \
3315 kib_sched_lock, flags);
3319 cfs_spin_lock_irqsave(&kiblnd_data.kib_sched_lock,
3322 if (rc != 0 || conn->ibc_ready) {
3323 /* There may be another completion waiting; get
3324 * another scheduler to check while I handle
3326 kiblnd_conn_addref(conn); /* +1 ref for sched_conns */
3327 cfs_list_add_tail(&conn->ibc_sched_list,
3328 &kiblnd_data.kib_sched_conns);
3329 cfs_waitq_signal(&kiblnd_data.kib_sched_waitq);
3331 conn->ibc_scheduled = 0;
3335 cfs_spin_unlock_irqrestore(&kiblnd_data. \
3339 kiblnd_complete(&wc);
3341 cfs_spin_lock_irqsave(&kiblnd_data. \
3346 kiblnd_conn_decref(conn); /* ...drop my ref from above */
3353 cfs_set_current_state(CFS_TASK_INTERRUPTIBLE);
3354 cfs_waitq_add_exclusive(&kiblnd_data.kib_sched_waitq, &wait);
3355 cfs_spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock, flags);
3357 cfs_waitq_wait(&wait, CFS_TASK_INTERRUPTIBLE);
3360 cfs_waitq_del(&kiblnd_data.kib_sched_waitq, &wait);
3361 cfs_set_current_state(CFS_TASK_RUNNING);
3362 cfs_spin_lock_irqsave(&kiblnd_data.kib_sched_lock, flags);
3365 cfs_spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock, flags);
3367 kiblnd_thread_fini();
3372 kiblnd_failover_thread(void *arg)
3374 cfs_rwlock_t *glock = &kiblnd_data.kib_global_lock;
3376 cfs_waitlink_t wait;
3377 unsigned long flags;
3380 LASSERT (*kiblnd_tunables.kib_dev_failover != 0);
3382 cfs_daemonize ("kiblnd_failover");
3383 cfs_block_allsigs ();
3385 cfs_waitlink_init(&wait);
3386 cfs_write_lock_irqsave(glock, flags);
3388 while (!kiblnd_data.kib_shutdown) {
3389 int do_failover = 0;
3392 cfs_list_for_each_entry(dev, &kiblnd_data.kib_failed_devs,
3394 if (cfs_time_before(cfs_time_current(),
3395 dev->ibd_next_failover))
3402 cfs_list_del_init(&dev->ibd_fail_list);
3403 dev->ibd_failover = 1;
3404 cfs_write_unlock_irqrestore(glock, flags);
3406 rc = kiblnd_dev_failover(dev);
3408 cfs_write_lock_irqsave(glock, flags);
3410 LASSERT (dev->ibd_failover);
3411 dev->ibd_failover = 0;
3412 if (rc >= 0) { /* Device is OK or failover succeed */
3413 dev->ibd_next_failover = cfs_time_shift(3);
3417 /* failed to failover, retry later */
3418 dev->ibd_next_failover =
3419 cfs_time_shift(min(dev->ibd_failed_failover, 10));
3420 if (kiblnd_dev_can_failover(dev)) {
3421 cfs_list_add_tail(&dev->ibd_fail_list,
3422 &kiblnd_data.kib_failed_devs);
3428 /* long sleep if no more pending failover */
3429 long_sleep = cfs_list_empty(&kiblnd_data.kib_failed_devs);
3431 cfs_set_current_state(CFS_TASK_INTERRUPTIBLE);
3432 cfs_waitq_add(&kiblnd_data.kib_failover_waitq, &wait);
3433 cfs_write_unlock_irqrestore(glock, flags);
3435 rc = schedule_timeout(long_sleep ? cfs_time_seconds(10) :
3436 cfs_time_seconds(1));
3437 cfs_set_current_state(CFS_TASK_RUNNING);
3438 cfs_waitq_del(&kiblnd_data.kib_failover_waitq, &wait);
3439 cfs_write_lock_irqsave(glock, flags);
3441 if (!long_sleep || rc != 0)
3444 /* have a long sleep, routine check all active devices,
3445 * we need checking like this because if there is not active
3446 * connection on the dev and no SEND from local, we may listen
3447 * on wrong HCA for ever while there is a bonding failover */
3448 cfs_list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3449 if (kiblnd_dev_can_failover(dev)) {
3450 cfs_list_add_tail(&dev->ibd_fail_list,
3451 &kiblnd_data.kib_failed_devs);
3456 cfs_write_unlock_irqrestore(glock, flags);
3458 kiblnd_thread_fini();