4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2012, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lnet/klnds/o2iblnd/o2iblnd_cb.c
38 * Author: Eric Barton <eric@bartonsoftware.com>
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 (!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, lnet_nid_t target)
105 kib_net_t *net = (kib_net_t *)ni->ni_data;
108 kib_tx_poolset_t *tps;
110 tps = net->ibn_tx_ps[lnet_cpt_of_nid(target)];
111 node = kiblnd_pool_alloc_node(&tps->tps_poolset);
114 tx = container_of(node, kib_tx_t, tx_list);
116 LASSERT (tx->tx_nwrq == 0);
117 LASSERT (!tx->tx_queued);
118 LASSERT (tx->tx_sending == 0);
119 LASSERT (!tx->tx_waiting);
120 LASSERT (tx->tx_status == 0);
121 LASSERT (tx->tx_conn == NULL);
122 LASSERT (tx->tx_lntmsg[0] == NULL);
123 LASSERT (tx->tx_lntmsg[1] == NULL);
124 LASSERT (tx->tx_u.pmr == NULL);
125 LASSERT (tx->tx_nfrags == 0);
131 kiblnd_drop_rx(kib_rx_t *rx)
133 kib_conn_t *conn = rx->rx_conn;
134 struct kib_sched_info *sched = conn->ibc_sched;
137 spin_lock_irqsave(&sched->ibs_lock, flags);
138 LASSERT(conn->ibc_nrx > 0);
140 spin_unlock_irqrestore(&sched->ibs_lock, flags);
142 kiblnd_conn_decref(conn);
146 kiblnd_post_rx (kib_rx_t *rx, int credit)
148 kib_conn_t *conn = rx->rx_conn;
149 kib_net_t *net = conn->ibc_peer->ibp_ni->ni_data;
150 struct ib_recv_wr *bad_wrq = NULL;
154 LASSERT (net != NULL);
155 LASSERT (!in_interrupt());
156 LASSERT (credit == IBLND_POSTRX_NO_CREDIT ||
157 credit == IBLND_POSTRX_PEER_CREDIT ||
158 credit == IBLND_POSTRX_RSRVD_CREDIT);
160 mr = kiblnd_find_dma_mr(conn->ibc_hdev, rx->rx_msgaddr, IBLND_MSG_SIZE);
161 LASSERT (mr != NULL);
163 rx->rx_sge.lkey = mr->lkey;
164 rx->rx_sge.addr = rx->rx_msgaddr;
165 rx->rx_sge.length = IBLND_MSG_SIZE;
167 rx->rx_wrq.next = NULL;
168 rx->rx_wrq.sg_list = &rx->rx_sge;
169 rx->rx_wrq.num_sge = 1;
170 rx->rx_wrq.wr_id = kiblnd_ptr2wreqid(rx, IBLND_WID_RX);
172 LASSERT (conn->ibc_state >= IBLND_CONN_INIT);
173 LASSERT (rx->rx_nob >= 0); /* not posted */
175 if (conn->ibc_state > IBLND_CONN_ESTABLISHED) {
176 kiblnd_drop_rx(rx); /* No more posts for this rx */
180 rx->rx_nob = -1; /* flag posted */
182 rc = ib_post_recv(conn->ibc_cmid->qp, &rx->rx_wrq, &bad_wrq);
184 CERROR("Can't post rx for %s: %d, bad_wrq: %p\n",
185 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc, bad_wrq);
189 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) /* Initial post */
193 kiblnd_close_conn(conn, rc);
194 kiblnd_drop_rx(rx); /* No more posts for this rx */
198 if (credit == IBLND_POSTRX_NO_CREDIT)
201 spin_lock(&conn->ibc_lock);
202 if (credit == IBLND_POSTRX_PEER_CREDIT)
203 conn->ibc_outstanding_credits++;
205 conn->ibc_reserved_credits++;
206 spin_unlock(&conn->ibc_lock);
208 kiblnd_check_sends(conn);
213 kiblnd_find_waiting_tx_locked(kib_conn_t *conn, int txtype, __u64 cookie)
217 cfs_list_for_each(tmp, &conn->ibc_active_txs) {
218 kib_tx_t *tx = cfs_list_entry(tmp, kib_tx_t, tx_list);
220 LASSERT (!tx->tx_queued);
221 LASSERT (tx->tx_sending != 0 || tx->tx_waiting);
223 if (tx->tx_cookie != cookie)
226 if (tx->tx_waiting &&
227 tx->tx_msg->ibm_type == txtype)
230 CWARN("Bad completion: %swaiting, type %x (wanted %x)\n",
231 tx->tx_waiting ? "" : "NOT ",
232 tx->tx_msg->ibm_type, txtype);
238 kiblnd_handle_completion(kib_conn_t *conn, int txtype, int status, __u64 cookie)
241 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
244 spin_lock(&conn->ibc_lock);
246 tx = kiblnd_find_waiting_tx_locked(conn, txtype, cookie);
248 spin_unlock(&conn->ibc_lock);
250 CWARN("Unmatched completion type %x cookie "LPX64" from %s\n",
251 txtype, cookie, libcfs_nid2str(conn->ibc_peer->ibp_nid));
252 kiblnd_close_conn(conn, -EPROTO);
256 if (tx->tx_status == 0) { /* success so far */
257 if (status < 0) { /* failed? */
258 tx->tx_status = status;
259 } else if (txtype == IBLND_MSG_GET_REQ) {
260 lnet_set_reply_msg_len(ni, tx->tx_lntmsg[1], status);
266 idle = !tx->tx_queued && (tx->tx_sending == 0);
268 cfs_list_del(&tx->tx_list);
270 spin_unlock(&conn->ibc_lock);
273 kiblnd_tx_done(ni, tx);
277 kiblnd_send_completion(kib_conn_t *conn, int type, int status, __u64 cookie)
279 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
280 kib_tx_t *tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
283 CERROR("Can't get tx for completion %x for %s\n",
284 type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
288 tx->tx_msg->ibm_u.completion.ibcm_status = status;
289 tx->tx_msg->ibm_u.completion.ibcm_cookie = cookie;
290 kiblnd_init_tx_msg(ni, tx, type, sizeof(kib_completion_msg_t));
292 kiblnd_queue_tx(tx, conn);
296 kiblnd_handle_rx (kib_rx_t *rx)
298 kib_msg_t *msg = rx->rx_msg;
299 kib_conn_t *conn = rx->rx_conn;
300 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
301 int credits = msg->ibm_credits;
307 LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
309 CDEBUG (D_NET, "Received %x[%d] from %s\n",
310 msg->ibm_type, credits,
311 libcfs_nid2str(conn->ibc_peer->ibp_nid));
314 /* Have I received credits that will let me send? */
315 spin_lock(&conn->ibc_lock);
317 if (conn->ibc_credits + credits >
318 IBLND_MSG_QUEUE_SIZE(conn->ibc_version)) {
319 rc2 = conn->ibc_credits;
320 spin_unlock(&conn->ibc_lock);
322 CERROR("Bad credits from %s: %d + %d > %d\n",
323 libcfs_nid2str(conn->ibc_peer->ibp_nid),
325 IBLND_MSG_QUEUE_SIZE(conn->ibc_version));
327 kiblnd_close_conn(conn, -EPROTO);
328 kiblnd_post_rx(rx, IBLND_POSTRX_NO_CREDIT);
332 conn->ibc_credits += credits;
334 /* This ensures the credit taken by NOOP can be returned */
335 if (msg->ibm_type == IBLND_MSG_NOOP &&
336 !IBLND_OOB_CAPABLE(conn->ibc_version)) /* v1 only */
337 conn->ibc_outstanding_credits++;
339 spin_unlock(&conn->ibc_lock);
340 kiblnd_check_sends(conn);
343 switch (msg->ibm_type) {
345 CERROR("Bad IBLND message type %x from %s\n",
346 msg->ibm_type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
347 post_credit = IBLND_POSTRX_NO_CREDIT;
352 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
353 post_credit = IBLND_POSTRX_NO_CREDIT;
357 if (credits != 0) /* credit already posted */
358 post_credit = IBLND_POSTRX_NO_CREDIT;
359 else /* a keepalive NOOP */
360 post_credit = IBLND_POSTRX_PEER_CREDIT;
363 case IBLND_MSG_IMMEDIATE:
364 post_credit = IBLND_POSTRX_DONT_POST;
365 rc = lnet_parse(ni, &msg->ibm_u.immediate.ibim_hdr,
366 msg->ibm_srcnid, rx, 0);
367 if (rc < 0) /* repost on error */
368 post_credit = IBLND_POSTRX_PEER_CREDIT;
371 case IBLND_MSG_PUT_REQ:
372 post_credit = IBLND_POSTRX_DONT_POST;
373 rc = lnet_parse(ni, &msg->ibm_u.putreq.ibprm_hdr,
374 msg->ibm_srcnid, rx, 1);
375 if (rc < 0) /* repost on error */
376 post_credit = IBLND_POSTRX_PEER_CREDIT;
379 case IBLND_MSG_PUT_NAK:
380 CWARN ("PUT_NACK from %s\n",
381 libcfs_nid2str(conn->ibc_peer->ibp_nid));
382 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
383 kiblnd_handle_completion(conn, IBLND_MSG_PUT_REQ,
384 msg->ibm_u.completion.ibcm_status,
385 msg->ibm_u.completion.ibcm_cookie);
388 case IBLND_MSG_PUT_ACK:
389 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
391 spin_lock(&conn->ibc_lock);
392 tx = kiblnd_find_waiting_tx_locked(conn, IBLND_MSG_PUT_REQ,
393 msg->ibm_u.putack.ibpam_src_cookie);
395 cfs_list_del(&tx->tx_list);
396 spin_unlock(&conn->ibc_lock);
399 CERROR("Unmatched PUT_ACK from %s\n",
400 libcfs_nid2str(conn->ibc_peer->ibp_nid));
405 LASSERT (tx->tx_waiting);
406 /* CAVEAT EMPTOR: I could be racing with tx_complete, but...
407 * (a) I can overwrite tx_msg since my peer has received it!
408 * (b) tx_waiting set tells tx_complete() it's not done. */
410 tx->tx_nwrq = 0; /* overwrite PUT_REQ */
412 rc2 = kiblnd_init_rdma(conn, tx, IBLND_MSG_PUT_DONE,
413 kiblnd_rd_size(&msg->ibm_u.putack.ibpam_rd),
414 &msg->ibm_u.putack.ibpam_rd,
415 msg->ibm_u.putack.ibpam_dst_cookie);
417 CERROR("Can't setup rdma for PUT to %s: %d\n",
418 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc2);
420 spin_lock(&conn->ibc_lock);
421 tx->tx_waiting = 0; /* clear waiting and queue atomically */
422 kiblnd_queue_tx_locked(tx, conn);
423 spin_unlock(&conn->ibc_lock);
426 case IBLND_MSG_PUT_DONE:
427 post_credit = IBLND_POSTRX_PEER_CREDIT;
428 kiblnd_handle_completion(conn, IBLND_MSG_PUT_ACK,
429 msg->ibm_u.completion.ibcm_status,
430 msg->ibm_u.completion.ibcm_cookie);
433 case IBLND_MSG_GET_REQ:
434 post_credit = IBLND_POSTRX_DONT_POST;
435 rc = lnet_parse(ni, &msg->ibm_u.get.ibgm_hdr,
436 msg->ibm_srcnid, rx, 1);
437 if (rc < 0) /* repost on error */
438 post_credit = IBLND_POSTRX_PEER_CREDIT;
441 case IBLND_MSG_GET_DONE:
442 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
443 kiblnd_handle_completion(conn, IBLND_MSG_GET_REQ,
444 msg->ibm_u.completion.ibcm_status,
445 msg->ibm_u.completion.ibcm_cookie);
449 if (rc < 0) /* protocol error */
450 kiblnd_close_conn(conn, rc);
452 if (post_credit != IBLND_POSTRX_DONT_POST)
453 kiblnd_post_rx(rx, post_credit);
457 kiblnd_rx_complete (kib_rx_t *rx, int status, int nob)
459 kib_msg_t *msg = rx->rx_msg;
460 kib_conn_t *conn = rx->rx_conn;
461 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
462 kib_net_t *net = ni->ni_data;
466 LASSERT (net != NULL);
467 LASSERT (rx->rx_nob < 0); /* was posted */
468 rx->rx_nob = 0; /* isn't now */
470 if (conn->ibc_state > IBLND_CONN_ESTABLISHED)
473 if (status != IB_WC_SUCCESS) {
474 CNETERR("Rx from %s failed: %d\n",
475 libcfs_nid2str(conn->ibc_peer->ibp_nid), status);
482 rc = kiblnd_unpack_msg(msg, rx->rx_nob);
484 CERROR ("Error %d unpacking rx from %s\n",
485 rc, libcfs_nid2str(conn->ibc_peer->ibp_nid));
489 if (msg->ibm_srcnid != conn->ibc_peer->ibp_nid ||
490 msg->ibm_dstnid != ni->ni_nid ||
491 msg->ibm_srcstamp != conn->ibc_incarnation ||
492 msg->ibm_dststamp != net->ibn_incarnation) {
493 CERROR ("Stale rx from %s\n",
494 libcfs_nid2str(conn->ibc_peer->ibp_nid));
499 /* set time last known alive */
500 kiblnd_peer_alive(conn->ibc_peer);
502 /* racing with connection establishment/teardown! */
504 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
505 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
508 write_lock_irqsave(g_lock, flags);
509 /* must check holding global lock to eliminate race */
510 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
511 cfs_list_add_tail(&rx->rx_list, &conn->ibc_early_rxs);
512 write_unlock_irqrestore(g_lock, flags);
515 write_unlock_irqrestore(g_lock, flags);
517 kiblnd_handle_rx(rx);
521 CDEBUG(D_NET, "rx %p conn %p\n", rx, conn);
522 kiblnd_close_conn(conn, err);
524 kiblnd_drop_rx(rx); /* Don't re-post rx. */
528 kiblnd_kvaddr_to_page (unsigned long vaddr)
532 if (vaddr >= VMALLOC_START &&
533 vaddr < VMALLOC_END) {
534 page = vmalloc_to_page ((void *)vaddr);
535 LASSERT (page != NULL);
538 #ifdef CONFIG_HIGHMEM
539 if (vaddr >= PKMAP_BASE &&
540 vaddr < (PKMAP_BASE + LAST_PKMAP * PAGE_SIZE)) {
541 /* No highmem pages only used for bulk (kiov) I/O */
542 CERROR("find page for address in highmem\n");
546 page = virt_to_page (vaddr);
547 LASSERT (page != NULL);
552 kiblnd_fmr_map_tx(kib_net_t *net, kib_tx_t *tx, kib_rdma_desc_t *rd, int nob)
555 __u64 *pages = tx->tx_pages;
556 kib_fmr_poolset_t *fps;
563 LASSERT(tx->tx_pool != NULL);
564 LASSERT(tx->tx_pool->tpo_pool.po_owner != NULL);
566 hdev = tx->tx_pool->tpo_hdev;
568 for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
569 for (size = 0; size < rd->rd_frags[i].rf_nob;
570 size += hdev->ibh_page_size) {
571 pages[npages ++] = (rd->rd_frags[i].rf_addr &
572 hdev->ibh_page_mask) + size;
576 cpt = tx->tx_pool->tpo_pool.po_owner->ps_cpt;
578 fps = net->ibn_fmr_ps[cpt];
579 rc = kiblnd_fmr_pool_map(fps, pages, npages, 0, &tx->tx_u.fmr);
581 CERROR ("Can't map %d pages: %d\n", npages, rc);
585 /* If rd is not tx_rd, it's going to get sent to a peer, who will need
587 rd->rd_key = (rd != tx->tx_rd) ? tx->tx_u.fmr.fmr_pfmr->fmr->rkey :
588 tx->tx_u.fmr.fmr_pfmr->fmr->lkey;
589 rd->rd_frags[0].rf_addr &= ~hdev->ibh_page_mask;
590 rd->rd_frags[0].rf_nob = nob;
597 kiblnd_pmr_map_tx(kib_net_t *net, kib_tx_t *tx, kib_rdma_desc_t *rd, int nob)
600 kib_pmr_poolset_t *pps;
605 LASSERT(tx->tx_pool != NULL);
606 LASSERT(tx->tx_pool->tpo_pool.po_owner != NULL);
608 hdev = tx->tx_pool->tpo_hdev;
610 iova = rd->rd_frags[0].rf_addr & ~hdev->ibh_page_mask;
612 cpt = tx->tx_pool->tpo_pool.po_owner->ps_cpt;
614 pps = net->ibn_pmr_ps[cpt];
615 rc = kiblnd_pmr_pool_map(pps, hdev, rd, &iova, &tx->tx_u.pmr);
617 CERROR("Failed to create MR by phybuf: %d\n", rc);
621 /* If rd is not tx_rd, it's going to get sent to a peer, who will need
623 rd->rd_key = (rd != tx->tx_rd) ? tx->tx_u.pmr->pmr_mr->rkey :
624 tx->tx_u.pmr->pmr_mr->lkey;
626 rd->rd_frags[0].rf_addr = iova;
627 rd->rd_frags[0].rf_nob = nob;
633 kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx)
635 kib_net_t *net = ni->ni_data;
637 LASSERT(net != NULL);
639 if (net->ibn_fmr_ps != NULL && tx->tx_u.fmr.fmr_pfmr != NULL) {
640 kiblnd_fmr_pool_unmap(&tx->tx_u.fmr, tx->tx_status);
641 tx->tx_u.fmr.fmr_pfmr = NULL;
643 } else if (net->ibn_pmr_ps != NULL && tx->tx_u.pmr != NULL) {
644 kiblnd_pmr_pool_unmap(tx->tx_u.pmr);
648 if (tx->tx_nfrags != 0) {
649 kiblnd_dma_unmap_sg(tx->tx_pool->tpo_hdev->ibh_ibdev,
650 tx->tx_frags, tx->tx_nfrags, tx->tx_dmadir);
656 kiblnd_map_tx(lnet_ni_t *ni, kib_tx_t *tx,
657 kib_rdma_desc_t *rd, int nfrags)
659 kib_hca_dev_t *hdev = tx->tx_pool->tpo_hdev;
660 kib_net_t *net = ni->ni_data;
661 struct ib_mr *mr = NULL;
665 /* If rd is not tx_rd, it's going to get sent to a peer and I'm the
667 tx->tx_dmadir = (rd != tx->tx_rd) ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
668 tx->tx_nfrags = nfrags;
671 kiblnd_dma_map_sg(hdev->ibh_ibdev,
672 tx->tx_frags, tx->tx_nfrags, tx->tx_dmadir);
674 for (i = 0, nob = 0; i < rd->rd_nfrags; i++) {
675 rd->rd_frags[i].rf_nob = kiblnd_sg_dma_len(
676 hdev->ibh_ibdev, &tx->tx_frags[i]);
677 rd->rd_frags[i].rf_addr = kiblnd_sg_dma_address(
678 hdev->ibh_ibdev, &tx->tx_frags[i]);
679 nob += rd->rd_frags[i].rf_nob;
682 /* looking for pre-mapping MR */
683 mr = kiblnd_find_rd_dma_mr(hdev, rd);
685 /* found pre-mapping MR */
686 rd->rd_key = (rd != tx->tx_rd) ? mr->rkey : mr->lkey;
690 if (net->ibn_fmr_ps != NULL)
691 return kiblnd_fmr_map_tx(net, tx, rd, nob);
692 else if (net->ibn_pmr_ps != NULL)
693 return kiblnd_pmr_map_tx(net, tx, rd, nob);
700 kiblnd_setup_rd_iov(lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd,
701 unsigned int niov, struct iovec *iov, int offset, int nob)
703 kib_net_t *net = ni->ni_data;
705 struct scatterlist *sg;
712 LASSERT (net != NULL);
714 while (offset >= iov->iov_len) {
715 offset -= iov->iov_len;
725 vaddr = ((unsigned long)iov->iov_base) + offset;
726 page_offset = vaddr & (PAGE_SIZE - 1);
727 page = kiblnd_kvaddr_to_page(vaddr);
729 CERROR ("Can't find page\n");
733 fragnob = min((int)(iov->iov_len - offset), nob);
734 fragnob = min(fragnob, (int)PAGE_SIZE - page_offset);
736 sg_set_page(sg, page, fragnob, page_offset);
739 if (offset + fragnob < iov->iov_len) {
749 return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags);
753 kiblnd_setup_rd_kiov (lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd,
754 int nkiov, lnet_kiov_t *kiov, int offset, int nob)
756 kib_net_t *net = ni->ni_data;
757 struct scatterlist *sg;
760 CDEBUG(D_NET, "niov %d offset %d nob %d\n", nkiov, offset, nob);
764 LASSERT (net != NULL);
766 while (offset >= kiov->kiov_len) {
767 offset -= kiov->kiov_len;
777 fragnob = min((int)(kiov->kiov_len - offset), nob);
779 sg_set_page(sg, kiov->kiov_page, fragnob,
780 kiov->kiov_offset + offset);
789 return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags);
793 kiblnd_post_tx_locked (kib_conn_t *conn, kib_tx_t *tx, int credit)
795 kib_msg_t *msg = tx->tx_msg;
796 kib_peer_t *peer = conn->ibc_peer;
797 int ver = conn->ibc_version;
800 struct ib_send_wr *bad_wrq;
802 LASSERT (tx->tx_queued);
803 /* We rely on this for QP sizing */
804 LASSERT (tx->tx_nwrq > 0);
805 LASSERT (tx->tx_nwrq <= 1 + IBLND_RDMA_FRAGS(ver));
807 LASSERT (credit == 0 || credit == 1);
808 LASSERT (conn->ibc_outstanding_credits >= 0);
809 LASSERT (conn->ibc_outstanding_credits <= IBLND_MSG_QUEUE_SIZE(ver));
810 LASSERT (conn->ibc_credits >= 0);
811 LASSERT (conn->ibc_credits <= IBLND_MSG_QUEUE_SIZE(ver));
813 if (conn->ibc_nsends_posted == IBLND_CONCURRENT_SENDS(ver)) {
814 /* tx completions outstanding... */
815 CDEBUG(D_NET, "%s: posted enough\n",
816 libcfs_nid2str(peer->ibp_nid));
820 if (credit != 0 && conn->ibc_credits == 0) { /* no credits */
821 CDEBUG(D_NET, "%s: no credits\n",
822 libcfs_nid2str(peer->ibp_nid));
826 if (credit != 0 && !IBLND_OOB_CAPABLE(ver) &&
827 conn->ibc_credits == 1 && /* last credit reserved */
828 msg->ibm_type != IBLND_MSG_NOOP) { /* for NOOP */
829 CDEBUG(D_NET, "%s: not using last credit\n",
830 libcfs_nid2str(peer->ibp_nid));
834 /* NB don't drop ibc_lock before bumping tx_sending */
835 cfs_list_del(&tx->tx_list);
838 if (msg->ibm_type == IBLND_MSG_NOOP &&
839 (!kiblnd_need_noop(conn) || /* redundant NOOP */
840 (IBLND_OOB_CAPABLE(ver) && /* posted enough NOOP */
841 conn->ibc_noops_posted == IBLND_OOB_MSGS(ver)))) {
842 /* OK to drop when posted enough NOOPs, since
843 * kiblnd_check_sends will queue NOOP again when
844 * posted NOOPs complete */
845 spin_unlock(&conn->ibc_lock);
846 kiblnd_tx_done(peer->ibp_ni, tx);
847 spin_lock(&conn->ibc_lock);
848 CDEBUG(D_NET, "%s(%d): redundant or enough NOOP\n",
849 libcfs_nid2str(peer->ibp_nid),
850 conn->ibc_noops_posted);
854 kiblnd_pack_msg(peer->ibp_ni, msg, ver, conn->ibc_outstanding_credits,
855 peer->ibp_nid, conn->ibc_incarnation);
857 conn->ibc_credits -= credit;
858 conn->ibc_outstanding_credits = 0;
859 conn->ibc_nsends_posted++;
860 if (msg->ibm_type == IBLND_MSG_NOOP)
861 conn->ibc_noops_posted++;
863 /* CAVEAT EMPTOR! This tx could be the PUT_DONE of an RDMA
864 * PUT. If so, it was first queued here as a PUT_REQ, sent and
865 * stashed on ibc_active_txs, matched by an incoming PUT_ACK,
866 * and then re-queued here. It's (just) possible that
867 * tx_sending is non-zero if we've not done the tx_complete()
868 * from the first send; hence the ++ rather than = below. */
870 cfs_list_add(&tx->tx_list, &conn->ibc_active_txs);
872 /* I'm still holding ibc_lock! */
873 if (conn->ibc_state != IBLND_CONN_ESTABLISHED) {
875 } else if (tx->tx_pool->tpo_pool.po_failed ||
876 conn->ibc_hdev != tx->tx_pool->tpo_hdev) {
877 /* close_conn will launch failover */
880 rc = ib_post_send(conn->ibc_cmid->qp,
881 tx->tx_wrq, &bad_wrq);
884 conn->ibc_last_send = jiffies;
889 /* NB credits are transferred in the actual
890 * message, which can only be the last work item */
891 conn->ibc_credits += credit;
892 conn->ibc_outstanding_credits += msg->ibm_credits;
893 conn->ibc_nsends_posted--;
894 if (msg->ibm_type == IBLND_MSG_NOOP)
895 conn->ibc_noops_posted--;
901 done = (tx->tx_sending == 0);
903 cfs_list_del(&tx->tx_list);
905 spin_unlock(&conn->ibc_lock);
907 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
908 CERROR("Error %d posting transmit to %s\n",
909 rc, libcfs_nid2str(peer->ibp_nid));
911 CDEBUG(D_NET, "Error %d posting transmit to %s\n",
912 rc, libcfs_nid2str(peer->ibp_nid));
914 kiblnd_close_conn(conn, rc);
917 kiblnd_tx_done(peer->ibp_ni, tx);
919 spin_lock(&conn->ibc_lock);
925 kiblnd_check_sends (kib_conn_t *conn)
927 int ver = conn->ibc_version;
928 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
931 /* Don't send anything until after the connection is established */
932 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
933 CDEBUG(D_NET, "%s too soon\n",
934 libcfs_nid2str(conn->ibc_peer->ibp_nid));
938 spin_lock(&conn->ibc_lock);
940 LASSERT (conn->ibc_nsends_posted <= IBLND_CONCURRENT_SENDS(ver));
941 LASSERT (!IBLND_OOB_CAPABLE(ver) ||
942 conn->ibc_noops_posted <= IBLND_OOB_MSGS(ver));
943 LASSERT (conn->ibc_reserved_credits >= 0);
945 while (conn->ibc_reserved_credits > 0 &&
946 !cfs_list_empty(&conn->ibc_tx_queue_rsrvd)) {
947 tx = cfs_list_entry(conn->ibc_tx_queue_rsrvd.next,
949 cfs_list_del(&tx->tx_list);
950 cfs_list_add_tail(&tx->tx_list, &conn->ibc_tx_queue);
951 conn->ibc_reserved_credits--;
954 if (kiblnd_need_noop(conn)) {
955 spin_unlock(&conn->ibc_lock);
957 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
959 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_NOOP, 0);
961 spin_lock(&conn->ibc_lock);
963 kiblnd_queue_tx_locked(tx, conn);
966 kiblnd_conn_addref(conn); /* 1 ref for me.... (see b21911) */
971 if (!cfs_list_empty(&conn->ibc_tx_queue_nocred)) {
973 tx = cfs_list_entry(conn->ibc_tx_queue_nocred.next,
975 } else if (!cfs_list_empty(&conn->ibc_tx_noops)) {
976 LASSERT (!IBLND_OOB_CAPABLE(ver));
978 tx = cfs_list_entry(conn->ibc_tx_noops.next,
980 } else if (!cfs_list_empty(&conn->ibc_tx_queue)) {
982 tx = cfs_list_entry(conn->ibc_tx_queue.next,
987 if (kiblnd_post_tx_locked(conn, tx, credit) != 0)
991 spin_unlock(&conn->ibc_lock);
993 kiblnd_conn_decref(conn); /* ...until here */
997 kiblnd_tx_complete (kib_tx_t *tx, int status)
999 int failed = (status != IB_WC_SUCCESS);
1000 kib_conn_t *conn = tx->tx_conn;
1003 LASSERT (tx->tx_sending > 0);
1006 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
1007 CNETERR("Tx -> %s cookie "LPX64
1008 " sending %d waiting %d: failed %d\n",
1009 libcfs_nid2str(conn->ibc_peer->ibp_nid),
1010 tx->tx_cookie, tx->tx_sending, tx->tx_waiting,
1013 kiblnd_close_conn(conn, -EIO);
1015 kiblnd_peer_alive(conn->ibc_peer);
1018 spin_lock(&conn->ibc_lock);
1020 /* I could be racing with rdma completion. Whoever makes 'tx' idle
1021 * gets to free it, which also drops its ref on 'conn'. */
1024 conn->ibc_nsends_posted--;
1025 if (tx->tx_msg->ibm_type == IBLND_MSG_NOOP)
1026 conn->ibc_noops_posted--;
1029 tx->tx_waiting = 0; /* don't wait for peer */
1030 tx->tx_status = -EIO;
1033 idle = (tx->tx_sending == 0) && /* This is the final callback */
1034 !tx->tx_waiting && /* Not waiting for peer */
1035 !tx->tx_queued; /* Not re-queued (PUT_DONE) */
1037 cfs_list_del(&tx->tx_list);
1039 kiblnd_conn_addref(conn); /* 1 ref for me.... */
1041 spin_unlock(&conn->ibc_lock);
1044 kiblnd_tx_done(conn->ibc_peer->ibp_ni, tx);
1046 kiblnd_check_sends(conn);
1048 kiblnd_conn_decref(conn); /* ...until here */
1052 kiblnd_init_tx_msg (lnet_ni_t *ni, kib_tx_t *tx, int type, int body_nob)
1054 kib_hca_dev_t *hdev = tx->tx_pool->tpo_hdev;
1055 struct ib_sge *sge = &tx->tx_sge[tx->tx_nwrq];
1056 struct ib_send_wr *wrq = &tx->tx_wrq[tx->tx_nwrq];
1057 int nob = offsetof (kib_msg_t, ibm_u) + body_nob;
1060 LASSERT (tx->tx_nwrq >= 0);
1061 LASSERT (tx->tx_nwrq < IBLND_MAX_RDMA_FRAGS + 1);
1062 LASSERT (nob <= IBLND_MSG_SIZE);
1064 kiblnd_init_msg(tx->tx_msg, type, body_nob);
1066 mr = kiblnd_find_dma_mr(hdev, tx->tx_msgaddr, nob);
1067 LASSERT (mr != NULL);
1069 sge->lkey = mr->lkey;
1070 sge->addr = tx->tx_msgaddr;
1073 memset(wrq, 0, sizeof(*wrq));
1076 wrq->wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_TX);
1079 wrq->opcode = IB_WR_SEND;
1080 wrq->send_flags = IB_SEND_SIGNALED;
1086 kiblnd_init_rdma (kib_conn_t *conn, kib_tx_t *tx, int type,
1087 int resid, kib_rdma_desc_t *dstrd, __u64 dstcookie)
1089 kib_msg_t *ibmsg = tx->tx_msg;
1090 kib_rdma_desc_t *srcrd = tx->tx_rd;
1091 struct ib_sge *sge = &tx->tx_sge[0];
1092 struct ib_send_wr *wrq = &tx->tx_wrq[0];
1098 LASSERT (!in_interrupt());
1099 LASSERT (tx->tx_nwrq == 0);
1100 LASSERT (type == IBLND_MSG_GET_DONE ||
1101 type == IBLND_MSG_PUT_DONE);
1103 srcidx = dstidx = 0;
1106 if (srcidx >= srcrd->rd_nfrags) {
1107 CERROR("Src buffer exhausted: %d frags\n", srcidx);
1112 if (dstidx == dstrd->rd_nfrags) {
1113 CERROR("Dst buffer exhausted: %d frags\n", dstidx);
1118 if (tx->tx_nwrq == IBLND_RDMA_FRAGS(conn->ibc_version)) {
1119 CERROR("RDMA too fragmented for %s (%d): "
1120 "%d/%d src %d/%d dst frags\n",
1121 libcfs_nid2str(conn->ibc_peer->ibp_nid),
1122 IBLND_RDMA_FRAGS(conn->ibc_version),
1123 srcidx, srcrd->rd_nfrags,
1124 dstidx, dstrd->rd_nfrags);
1129 wrknob = MIN(MIN(kiblnd_rd_frag_size(srcrd, srcidx),
1130 kiblnd_rd_frag_size(dstrd, dstidx)), resid);
1132 sge = &tx->tx_sge[tx->tx_nwrq];
1133 sge->addr = kiblnd_rd_frag_addr(srcrd, srcidx);
1134 sge->lkey = kiblnd_rd_frag_key(srcrd, srcidx);
1135 sge->length = wrknob;
1137 wrq = &tx->tx_wrq[tx->tx_nwrq];
1139 wrq->next = wrq + 1;
1140 wrq->wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_RDMA);
1143 wrq->opcode = IB_WR_RDMA_WRITE;
1144 wrq->send_flags = 0;
1146 wrq->wr.rdma.remote_addr = kiblnd_rd_frag_addr(dstrd, dstidx);
1147 wrq->wr.rdma.rkey = kiblnd_rd_frag_key(dstrd, dstidx);
1149 srcidx = kiblnd_rd_consume_frag(srcrd, srcidx, wrknob);
1150 dstidx = kiblnd_rd_consume_frag(dstrd, dstidx, wrknob);
1159 if (rc < 0) /* no RDMA if completing with failure */
1162 ibmsg->ibm_u.completion.ibcm_status = rc;
1163 ibmsg->ibm_u.completion.ibcm_cookie = dstcookie;
1164 kiblnd_init_tx_msg(conn->ibc_peer->ibp_ni, tx,
1165 type, sizeof (kib_completion_msg_t));
1171 kiblnd_queue_tx_locked (kib_tx_t *tx, kib_conn_t *conn)
1175 LASSERT (tx->tx_nwrq > 0); /* work items set up */
1176 LASSERT (!tx->tx_queued); /* not queued for sending already */
1177 LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1180 tx->tx_deadline = jiffies + (*kiblnd_tunables.kib_timeout * HZ);
1182 if (tx->tx_conn == NULL) {
1183 kiblnd_conn_addref(conn);
1185 LASSERT (tx->tx_msg->ibm_type != IBLND_MSG_PUT_DONE);
1187 /* PUT_DONE first attached to conn as a PUT_REQ */
1188 LASSERT (tx->tx_conn == conn);
1189 LASSERT (tx->tx_msg->ibm_type == IBLND_MSG_PUT_DONE);
1192 switch (tx->tx_msg->ibm_type) {
1196 case IBLND_MSG_PUT_REQ:
1197 case IBLND_MSG_GET_REQ:
1198 q = &conn->ibc_tx_queue_rsrvd;
1201 case IBLND_MSG_PUT_NAK:
1202 case IBLND_MSG_PUT_ACK:
1203 case IBLND_MSG_PUT_DONE:
1204 case IBLND_MSG_GET_DONE:
1205 q = &conn->ibc_tx_queue_nocred;
1208 case IBLND_MSG_NOOP:
1209 if (IBLND_OOB_CAPABLE(conn->ibc_version))
1210 q = &conn->ibc_tx_queue_nocred;
1212 q = &conn->ibc_tx_noops;
1215 case IBLND_MSG_IMMEDIATE:
1216 q = &conn->ibc_tx_queue;
1220 cfs_list_add_tail(&tx->tx_list, q);
1224 kiblnd_queue_tx (kib_tx_t *tx, kib_conn_t *conn)
1226 spin_lock(&conn->ibc_lock);
1227 kiblnd_queue_tx_locked(tx, conn);
1228 spin_unlock(&conn->ibc_lock);
1230 kiblnd_check_sends(conn);
1233 static int kiblnd_resolve_addr(struct rdma_cm_id *cmid,
1234 struct sockaddr_in *srcaddr,
1235 struct sockaddr_in *dstaddr,
1238 unsigned short port;
1241 #ifdef HAVE_OFED_RDMA_SET_REUSEADDR
1242 /* allow the port to be reused */
1243 rc = rdma_set_reuseaddr(cmid, 1);
1245 CERROR("Unable to set reuse on cmid: %d\n", rc);
1250 /* look for a free privileged port */
1251 for (port = PROT_SOCK-1; port > 0; port--) {
1252 srcaddr->sin_port = htons(port);
1253 rc = rdma_resolve_addr(cmid,
1254 (struct sockaddr *)srcaddr,
1255 (struct sockaddr *)dstaddr,
1258 CDEBUG(D_NET, "bound to port %hu\n", port);
1260 } else if (rc == -EADDRINUSE || rc == -EADDRNOTAVAIL) {
1261 CDEBUG(D_NET, "bind to port %hu failed: %d\n",
1268 CERROR("Failed to bind to a free privileged port\n");
1269 #ifndef HAVE_OFED_RDMA_SET_REUSEADDR
1270 CERROR("You may need IB verbs that supports rdma_set_reuseaddr()\n");
1276 kiblnd_connect_peer (kib_peer_t *peer)
1278 struct rdma_cm_id *cmid;
1280 kib_net_t *net = peer->ibp_ni->ni_data;
1281 struct sockaddr_in srcaddr;
1282 struct sockaddr_in dstaddr;
1285 LASSERT (net != NULL);
1286 LASSERT (peer->ibp_connecting > 0);
1288 cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, peer, RDMA_PS_TCP,
1292 CERROR("Can't create CMID for %s: %ld\n",
1293 libcfs_nid2str(peer->ibp_nid), PTR_ERR(cmid));
1299 memset(&srcaddr, 0, sizeof(srcaddr));
1300 srcaddr.sin_family = AF_INET;
1301 srcaddr.sin_addr.s_addr = htonl(dev->ibd_ifip);
1303 memset(&dstaddr, 0, sizeof(dstaddr));
1304 dstaddr.sin_family = AF_INET;
1305 dstaddr.sin_port = htons(*kiblnd_tunables.kib_service);
1306 dstaddr.sin_addr.s_addr = htonl(LNET_NIDADDR(peer->ibp_nid));
1308 kiblnd_peer_addref(peer); /* cmid's ref */
1310 if (*kiblnd_tunables.kib_use_priv_port) {
1311 rc = kiblnd_resolve_addr(cmid, &srcaddr, &dstaddr,
1312 *kiblnd_tunables.kib_timeout * 1000);
1314 rc = rdma_resolve_addr(cmid,
1315 (struct sockaddr *)&srcaddr,
1316 (struct sockaddr *)&dstaddr,
1317 *kiblnd_tunables.kib_timeout * 1000);
1320 /* Can't initiate address resolution: */
1321 CERROR("Can't resolve addr for %s: %d\n",
1322 libcfs_nid2str(peer->ibp_nid), rc);
1326 LASSERT (cmid->device != NULL);
1327 CDEBUG(D_NET, "%s: connection bound to %s:%u.%u.%u.%u:%s\n",
1328 libcfs_nid2str(peer->ibp_nid), dev->ibd_ifname,
1329 HIPQUAD(dev->ibd_ifip), cmid->device->name);
1334 kiblnd_peer_decref(peer); /* cmid's ref */
1335 rdma_destroy_id(cmid);
1337 kiblnd_peer_connect_failed(peer, 1, rc);
1341 kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid)
1346 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
1347 unsigned long flags;
1350 /* If I get here, I've committed to send, so I complete the tx with
1351 * failure on any problems */
1353 LASSERT (tx == NULL || tx->tx_conn == NULL); /* only set when assigned a conn */
1354 LASSERT (tx == NULL || tx->tx_nwrq > 0); /* work items have been set up */
1356 /* First time, just use a read lock since I expect to find my peer
1358 read_lock_irqsave(g_lock, flags);
1360 peer = kiblnd_find_peer_locked(nid);
1361 if (peer != NULL && !cfs_list_empty(&peer->ibp_conns)) {
1362 /* Found a peer with an established connection */
1363 conn = kiblnd_get_conn_locked(peer);
1364 kiblnd_conn_addref(conn); /* 1 ref for me... */
1366 read_unlock_irqrestore(g_lock, flags);
1369 kiblnd_queue_tx(tx, conn);
1370 kiblnd_conn_decref(conn); /* ...to here */
1374 read_unlock(g_lock);
1375 /* Re-try with a write lock */
1378 peer = kiblnd_find_peer_locked(nid);
1380 if (cfs_list_empty(&peer->ibp_conns)) {
1381 /* found a peer, but it's still connecting... */
1382 LASSERT (peer->ibp_connecting != 0 ||
1383 peer->ibp_accepting != 0);
1385 cfs_list_add_tail(&tx->tx_list,
1386 &peer->ibp_tx_queue);
1387 write_unlock_irqrestore(g_lock, flags);
1389 conn = kiblnd_get_conn_locked(peer);
1390 kiblnd_conn_addref(conn); /* 1 ref for me... */
1392 write_unlock_irqrestore(g_lock, flags);
1395 kiblnd_queue_tx(tx, conn);
1396 kiblnd_conn_decref(conn); /* ...to here */
1401 write_unlock_irqrestore(g_lock, flags);
1403 /* Allocate a peer ready to add to the peer table and retry */
1404 rc = kiblnd_create_peer(ni, &peer, nid);
1406 CERROR("Can't create peer %s\n", libcfs_nid2str(nid));
1408 tx->tx_status = -EHOSTUNREACH;
1410 kiblnd_tx_done(ni, tx);
1415 write_lock_irqsave(g_lock, flags);
1417 peer2 = kiblnd_find_peer_locked(nid);
1418 if (peer2 != NULL) {
1419 if (cfs_list_empty(&peer2->ibp_conns)) {
1420 /* found a peer, but it's still connecting... */
1421 LASSERT (peer2->ibp_connecting != 0 ||
1422 peer2->ibp_accepting != 0);
1424 cfs_list_add_tail(&tx->tx_list,
1425 &peer2->ibp_tx_queue);
1426 write_unlock_irqrestore(g_lock, flags);
1428 conn = kiblnd_get_conn_locked(peer2);
1429 kiblnd_conn_addref(conn); /* 1 ref for me... */
1431 write_unlock_irqrestore(g_lock, flags);
1434 kiblnd_queue_tx(tx, conn);
1435 kiblnd_conn_decref(conn); /* ...to here */
1438 kiblnd_peer_decref(peer);
1442 /* Brand new peer */
1443 LASSERT (peer->ibp_connecting == 0);
1444 peer->ibp_connecting = 1;
1446 /* always called with a ref on ni, which prevents ni being shutdown */
1447 LASSERT (((kib_net_t *)ni->ni_data)->ibn_shutdown == 0);
1450 cfs_list_add_tail(&tx->tx_list, &peer->ibp_tx_queue);
1452 kiblnd_peer_addref(peer);
1453 cfs_list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid));
1455 write_unlock_irqrestore(g_lock, flags);
1457 kiblnd_connect_peer(peer);
1458 kiblnd_peer_decref(peer);
1462 kiblnd_send (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg)
1464 lnet_hdr_t *hdr = &lntmsg->msg_hdr;
1465 int type = lntmsg->msg_type;
1466 lnet_process_id_t target = lntmsg->msg_target;
1467 int target_is_router = lntmsg->msg_target_is_router;
1468 int routing = lntmsg->msg_routing;
1469 unsigned int payload_niov = lntmsg->msg_niov;
1470 struct iovec *payload_iov = lntmsg->msg_iov;
1471 lnet_kiov_t *payload_kiov = lntmsg->msg_kiov;
1472 unsigned int payload_offset = lntmsg->msg_offset;
1473 unsigned int payload_nob = lntmsg->msg_len;
1479 /* NB 'private' is different depending on what we're sending.... */
1481 CDEBUG(D_NET, "sending %d bytes in %d frags to %s\n",
1482 payload_nob, payload_niov, libcfs_id2str(target));
1484 LASSERT (payload_nob == 0 || payload_niov > 0);
1485 LASSERT (payload_niov <= LNET_MAX_IOV);
1487 /* Thread context */
1488 LASSERT (!in_interrupt());
1489 /* payload is either all vaddrs or all pages */
1490 LASSERT (!(payload_kiov != NULL && payload_iov != NULL));
1498 LASSERT (payload_nob == 0);
1502 if (routing || target_is_router)
1503 break; /* send IMMEDIATE */
1505 /* is the REPLY message too small for RDMA? */
1506 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[lntmsg->msg_md->md_length]);
1507 if (nob <= IBLND_MSG_SIZE)
1508 break; /* send IMMEDIATE */
1510 tx = kiblnd_get_idle_tx(ni, target.nid);
1512 CERROR("Can't allocate txd for GET to %s\n",
1513 libcfs_nid2str(target.nid));
1519 if ((lntmsg->msg_md->md_options & LNET_MD_KIOV) == 0)
1520 rc = kiblnd_setup_rd_iov(ni, tx,
1521 &ibmsg->ibm_u.get.ibgm_rd,
1522 lntmsg->msg_md->md_niov,
1523 lntmsg->msg_md->md_iov.iov,
1524 0, lntmsg->msg_md->md_length);
1526 rc = kiblnd_setup_rd_kiov(ni, tx,
1527 &ibmsg->ibm_u.get.ibgm_rd,
1528 lntmsg->msg_md->md_niov,
1529 lntmsg->msg_md->md_iov.kiov,
1530 0, lntmsg->msg_md->md_length);
1532 CERROR("Can't setup GET sink for %s: %d\n",
1533 libcfs_nid2str(target.nid), rc);
1534 kiblnd_tx_done(ni, tx);
1538 nob = offsetof(kib_get_msg_t, ibgm_rd.rd_frags[tx->tx_nfrags]);
1539 ibmsg->ibm_u.get.ibgm_cookie = tx->tx_cookie;
1540 ibmsg->ibm_u.get.ibgm_hdr = *hdr;
1542 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_GET_REQ, nob);
1544 tx->tx_lntmsg[1] = lnet_create_reply_msg(ni, lntmsg);
1545 if (tx->tx_lntmsg[1] == NULL) {
1546 CERROR("Can't create reply for GET -> %s\n",
1547 libcfs_nid2str(target.nid));
1548 kiblnd_tx_done(ni, tx);
1552 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg[0,1] on completion */
1553 tx->tx_waiting = 1; /* waiting for GET_DONE */
1554 kiblnd_launch_tx(ni, tx, target.nid);
1557 case LNET_MSG_REPLY:
1559 /* Is the payload small enough not to need RDMA? */
1560 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob]);
1561 if (nob <= IBLND_MSG_SIZE)
1562 break; /* send IMMEDIATE */
1564 tx = kiblnd_get_idle_tx(ni, target.nid);
1566 CERROR("Can't allocate %s txd for %s\n",
1567 type == LNET_MSG_PUT ? "PUT" : "REPLY",
1568 libcfs_nid2str(target.nid));
1572 if (payload_kiov == NULL)
1573 rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1574 payload_niov, payload_iov,
1575 payload_offset, payload_nob);
1577 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1578 payload_niov, payload_kiov,
1579 payload_offset, payload_nob);
1581 CERROR("Can't setup PUT src for %s: %d\n",
1582 libcfs_nid2str(target.nid), rc);
1583 kiblnd_tx_done(ni, tx);
1588 ibmsg->ibm_u.putreq.ibprm_hdr = *hdr;
1589 ibmsg->ibm_u.putreq.ibprm_cookie = tx->tx_cookie;
1590 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_REQ, sizeof(kib_putreq_msg_t));
1592 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1593 tx->tx_waiting = 1; /* waiting for PUT_{ACK,NAK} */
1594 kiblnd_launch_tx(ni, tx, target.nid);
1598 /* send IMMEDIATE */
1600 LASSERT (offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob])
1603 tx = kiblnd_get_idle_tx(ni, target.nid);
1605 CERROR ("Can't send %d to %s: tx descs exhausted\n",
1606 type, libcfs_nid2str(target.nid));
1611 ibmsg->ibm_u.immediate.ibim_hdr = *hdr;
1613 if (payload_kiov != NULL)
1614 lnet_copy_kiov2flat(IBLND_MSG_SIZE, ibmsg,
1615 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1616 payload_niov, payload_kiov,
1617 payload_offset, payload_nob);
1619 lnet_copy_iov2flat(IBLND_MSG_SIZE, ibmsg,
1620 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1621 payload_niov, payload_iov,
1622 payload_offset, payload_nob);
1624 nob = offsetof(kib_immediate_msg_t, ibim_payload[payload_nob]);
1625 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_IMMEDIATE, nob);
1627 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1628 kiblnd_launch_tx(ni, tx, target.nid);
1633 kiblnd_reply (lnet_ni_t *ni, kib_rx_t *rx, lnet_msg_t *lntmsg)
1635 lnet_process_id_t target = lntmsg->msg_target;
1636 unsigned int niov = lntmsg->msg_niov;
1637 struct iovec *iov = lntmsg->msg_iov;
1638 lnet_kiov_t *kiov = lntmsg->msg_kiov;
1639 unsigned int offset = lntmsg->msg_offset;
1640 unsigned int nob = lntmsg->msg_len;
1644 tx = kiblnd_get_idle_tx(ni, rx->rx_conn->ibc_peer->ibp_nid);
1646 CERROR("Can't get tx for REPLY to %s\n",
1647 libcfs_nid2str(target.nid));
1653 else if (kiov == NULL)
1654 rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1655 niov, iov, offset, nob);
1657 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1658 niov, kiov, offset, nob);
1661 CERROR("Can't setup GET src for %s: %d\n",
1662 libcfs_nid2str(target.nid), rc);
1666 rc = kiblnd_init_rdma(rx->rx_conn, tx,
1667 IBLND_MSG_GET_DONE, nob,
1668 &rx->rx_msg->ibm_u.get.ibgm_rd,
1669 rx->rx_msg->ibm_u.get.ibgm_cookie);
1671 CERROR("Can't setup rdma for GET from %s: %d\n",
1672 libcfs_nid2str(target.nid), rc);
1677 /* No RDMA: local completion may happen now! */
1678 lnet_finalize(ni, lntmsg, 0);
1680 /* RDMA: lnet_finalize(lntmsg) when it
1682 tx->tx_lntmsg[0] = lntmsg;
1685 kiblnd_queue_tx(tx, rx->rx_conn);
1689 kiblnd_tx_done(ni, tx);
1691 lnet_finalize(ni, lntmsg, -EIO);
1695 kiblnd_recv (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed,
1696 unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov,
1697 unsigned int offset, unsigned int mlen, unsigned int rlen)
1699 kib_rx_t *rx = private;
1700 kib_msg_t *rxmsg = rx->rx_msg;
1701 kib_conn_t *conn = rx->rx_conn;
1705 int post_credit = IBLND_POSTRX_PEER_CREDIT;
1708 LASSERT (mlen <= rlen);
1709 LASSERT (!in_interrupt());
1710 /* Either all pages or all vaddrs */
1711 LASSERT (!(kiov != NULL && iov != NULL));
1713 switch (rxmsg->ibm_type) {
1717 case IBLND_MSG_IMMEDIATE:
1718 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[rlen]);
1719 if (nob > rx->rx_nob) {
1720 CERROR ("Immediate message from %s too big: %d(%d)\n",
1721 libcfs_nid2str(rxmsg->ibm_u.immediate.ibim_hdr.src_nid),
1728 lnet_copy_flat2kiov(niov, kiov, offset,
1729 IBLND_MSG_SIZE, rxmsg,
1730 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1733 lnet_copy_flat2iov(niov, iov, offset,
1734 IBLND_MSG_SIZE, rxmsg,
1735 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1737 lnet_finalize (ni, lntmsg, 0);
1740 case IBLND_MSG_PUT_REQ:
1742 lnet_finalize(ni, lntmsg, 0);
1743 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK, 0,
1744 rxmsg->ibm_u.putreq.ibprm_cookie);
1748 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
1750 CERROR("Can't allocate tx for %s\n",
1751 libcfs_nid2str(conn->ibc_peer->ibp_nid));
1752 /* Not replying will break the connection */
1759 rc = kiblnd_setup_rd_iov(ni, tx,
1760 &txmsg->ibm_u.putack.ibpam_rd,
1761 niov, iov, offset, mlen);
1763 rc = kiblnd_setup_rd_kiov(ni, tx,
1764 &txmsg->ibm_u.putack.ibpam_rd,
1765 niov, kiov, offset, mlen);
1767 CERROR("Can't setup PUT sink for %s: %d\n",
1768 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
1769 kiblnd_tx_done(ni, tx);
1770 /* tell peer it's over */
1771 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK, rc,
1772 rxmsg->ibm_u.putreq.ibprm_cookie);
1776 nob = offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[tx->tx_nfrags]);
1777 txmsg->ibm_u.putack.ibpam_src_cookie = rxmsg->ibm_u.putreq.ibprm_cookie;
1778 txmsg->ibm_u.putack.ibpam_dst_cookie = tx->tx_cookie;
1780 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_ACK, nob);
1782 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1783 tx->tx_waiting = 1; /* waiting for PUT_DONE */
1784 kiblnd_queue_tx(tx, conn);
1786 /* reposted buffer reserved for PUT_DONE */
1787 post_credit = IBLND_POSTRX_NO_CREDIT;
1790 case IBLND_MSG_GET_REQ:
1791 if (lntmsg != NULL) {
1792 /* Optimized GET; RDMA lntmsg's payload */
1793 kiblnd_reply(ni, rx, lntmsg);
1795 /* GET didn't match anything */
1796 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_GET_DONE,
1798 rxmsg->ibm_u.get.ibgm_cookie);
1803 kiblnd_post_rx(rx, post_credit);
1808 kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name)
1810 struct task_struct *task = kthread_run(fn, arg, name);
1813 return PTR_ERR(task);
1815 atomic_inc(&kiblnd_data.kib_nthreads);
1820 kiblnd_thread_fini (void)
1822 atomic_dec (&kiblnd_data.kib_nthreads);
1826 kiblnd_peer_alive (kib_peer_t *peer)
1828 /* This is racy, but everyone's only writing cfs_time_current() */
1829 peer->ibp_last_alive = cfs_time_current();
1834 kiblnd_peer_notify (kib_peer_t *peer)
1837 cfs_time_t last_alive = 0;
1838 unsigned long flags;
1840 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1842 if (cfs_list_empty(&peer->ibp_conns) &&
1843 peer->ibp_accepting == 0 &&
1844 peer->ibp_connecting == 0 &&
1845 peer->ibp_error != 0) {
1846 error = peer->ibp_error;
1847 peer->ibp_error = 0;
1849 last_alive = peer->ibp_last_alive;
1852 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1855 lnet_notify(peer->ibp_ni,
1856 peer->ibp_nid, 0, last_alive);
1860 kiblnd_close_conn_locked (kib_conn_t *conn, int error)
1862 /* This just does the immediate housekeeping. 'error' is zero for a
1863 * normal shutdown which can happen only after the connection has been
1864 * established. If the connection is established, schedule the
1865 * connection to be finished off by the connd. Otherwise the connd is
1866 * already dealing with it (either to set it up or tear it down).
1867 * Caller holds kib_global_lock exclusively in irq context */
1868 kib_peer_t *peer = conn->ibc_peer;
1870 unsigned long flags;
1872 LASSERT (error != 0 || conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1874 if (error != 0 && conn->ibc_comms_error == 0)
1875 conn->ibc_comms_error = error;
1877 if (conn->ibc_state != IBLND_CONN_ESTABLISHED)
1878 return; /* already being handled */
1881 cfs_list_empty(&conn->ibc_tx_noops) &&
1882 cfs_list_empty(&conn->ibc_tx_queue) &&
1883 cfs_list_empty(&conn->ibc_tx_queue_rsrvd) &&
1884 cfs_list_empty(&conn->ibc_tx_queue_nocred) &&
1885 cfs_list_empty(&conn->ibc_active_txs)) {
1886 CDEBUG(D_NET, "closing conn to %s\n",
1887 libcfs_nid2str(peer->ibp_nid));
1889 CNETERR("Closing conn to %s: error %d%s%s%s%s%s\n",
1890 libcfs_nid2str(peer->ibp_nid), error,
1891 cfs_list_empty(&conn->ibc_tx_queue) ? "" : "(sending)",
1892 cfs_list_empty(&conn->ibc_tx_noops) ? "" : "(sending_noops)",
1893 cfs_list_empty(&conn->ibc_tx_queue_rsrvd) ? "" : "(sending_rsrvd)",
1894 cfs_list_empty(&conn->ibc_tx_queue_nocred) ? "" : "(sending_nocred)",
1895 cfs_list_empty(&conn->ibc_active_txs) ? "" : "(waiting)");
1898 dev = ((kib_net_t *)peer->ibp_ni->ni_data)->ibn_dev;
1899 cfs_list_del(&conn->ibc_list);
1900 /* connd (see below) takes over ibc_list's ref */
1902 if (cfs_list_empty (&peer->ibp_conns) && /* no more conns */
1903 kiblnd_peer_active(peer)) { /* still in peer table */
1904 kiblnd_unlink_peer_locked(peer);
1906 /* set/clear error on last conn */
1907 peer->ibp_error = conn->ibc_comms_error;
1910 kiblnd_set_conn_state(conn, IBLND_CONN_CLOSING);
1913 kiblnd_dev_can_failover(dev)) {
1914 cfs_list_add_tail(&dev->ibd_fail_list,
1915 &kiblnd_data.kib_failed_devs);
1916 wake_up(&kiblnd_data.kib_failover_waitq);
1919 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
1921 cfs_list_add_tail(&conn->ibc_list, &kiblnd_data.kib_connd_conns);
1922 wake_up(&kiblnd_data.kib_connd_waitq);
1924 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
1928 kiblnd_close_conn(kib_conn_t *conn, int error)
1930 unsigned long flags;
1932 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1934 kiblnd_close_conn_locked(conn, error);
1936 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1940 kiblnd_handle_early_rxs(kib_conn_t *conn)
1942 unsigned long flags;
1945 LASSERT(!in_interrupt());
1946 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1948 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1949 while (!cfs_list_empty(&conn->ibc_early_rxs)) {
1950 rx = cfs_list_entry(conn->ibc_early_rxs.next,
1952 cfs_list_del(&rx->rx_list);
1953 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1955 kiblnd_handle_rx(rx);
1957 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1959 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1963 kiblnd_abort_txs(kib_conn_t *conn, cfs_list_t *txs)
1965 CFS_LIST_HEAD (zombies);
1970 spin_lock(&conn->ibc_lock);
1972 cfs_list_for_each_safe (tmp, nxt, txs) {
1973 tx = cfs_list_entry (tmp, kib_tx_t, tx_list);
1975 if (txs == &conn->ibc_active_txs) {
1976 LASSERT (!tx->tx_queued);
1977 LASSERT (tx->tx_waiting ||
1978 tx->tx_sending != 0);
1980 LASSERT (tx->tx_queued);
1983 tx->tx_status = -ECONNABORTED;
1986 if (tx->tx_sending == 0) {
1988 cfs_list_del (&tx->tx_list);
1989 cfs_list_add (&tx->tx_list, &zombies);
1993 spin_unlock(&conn->ibc_lock);
1995 kiblnd_txlist_done(conn->ibc_peer->ibp_ni, &zombies, -ECONNABORTED);
1999 kiblnd_finalise_conn (kib_conn_t *conn)
2001 LASSERT (!in_interrupt());
2002 LASSERT (conn->ibc_state > IBLND_CONN_INIT);
2004 kiblnd_set_conn_state(conn, IBLND_CONN_DISCONNECTED);
2006 /* abort_receives moves QP state to IB_QPS_ERR. This is only required
2007 * for connections that didn't get as far as being connected, because
2008 * rdma_disconnect() does this for free. */
2009 kiblnd_abort_receives(conn);
2011 /* Complete all tx descs not waiting for sends to complete.
2012 * NB we should be safe from RDMA now that the QP has changed state */
2014 kiblnd_abort_txs(conn, &conn->ibc_tx_noops);
2015 kiblnd_abort_txs(conn, &conn->ibc_tx_queue);
2016 kiblnd_abort_txs(conn, &conn->ibc_tx_queue_rsrvd);
2017 kiblnd_abort_txs(conn, &conn->ibc_tx_queue_nocred);
2018 kiblnd_abort_txs(conn, &conn->ibc_active_txs);
2020 kiblnd_handle_early_rxs(conn);
2024 kiblnd_peer_connect_failed (kib_peer_t *peer, int active, int error)
2026 CFS_LIST_HEAD (zombies);
2027 unsigned long flags;
2029 LASSERT (error != 0);
2030 LASSERT (!in_interrupt());
2032 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2035 LASSERT (peer->ibp_connecting > 0);
2036 peer->ibp_connecting--;
2038 LASSERT (peer->ibp_accepting > 0);
2039 peer->ibp_accepting--;
2042 if (peer->ibp_connecting != 0 ||
2043 peer->ibp_accepting != 0) {
2044 /* another connection attempt under way... */
2045 write_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2050 if (cfs_list_empty(&peer->ibp_conns)) {
2051 /* Take peer's blocked transmits to complete with error */
2052 cfs_list_add(&zombies, &peer->ibp_tx_queue);
2053 cfs_list_del_init(&peer->ibp_tx_queue);
2055 if (kiblnd_peer_active(peer))
2056 kiblnd_unlink_peer_locked(peer);
2058 peer->ibp_error = error;
2060 /* Can't have blocked transmits if there are connections */
2061 LASSERT (cfs_list_empty(&peer->ibp_tx_queue));
2064 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2066 kiblnd_peer_notify(peer);
2068 if (cfs_list_empty (&zombies))
2071 CNETERR("Deleting messages for %s: connection failed\n",
2072 libcfs_nid2str(peer->ibp_nid));
2074 kiblnd_txlist_done(peer->ibp_ni, &zombies, -EHOSTUNREACH);
2078 kiblnd_connreq_done(kib_conn_t *conn, int status)
2080 kib_peer_t *peer = conn->ibc_peer;
2083 unsigned long flags;
2086 active = (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2088 CDEBUG(D_NET,"%s: active(%d), version(%x), status(%d)\n",
2089 libcfs_nid2str(peer->ibp_nid), active,
2090 conn->ibc_version, status);
2092 LASSERT (!in_interrupt());
2093 LASSERT ((conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT &&
2094 peer->ibp_connecting > 0) ||
2095 (conn->ibc_state == IBLND_CONN_PASSIVE_WAIT &&
2096 peer->ibp_accepting > 0));
2098 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
2099 conn->ibc_connvars = NULL;
2102 /* failed to establish connection */
2103 kiblnd_peer_connect_failed(peer, active, status);
2104 kiblnd_finalise_conn(conn);
2108 /* connection established */
2109 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2111 conn->ibc_last_send = jiffies;
2112 kiblnd_set_conn_state(conn, IBLND_CONN_ESTABLISHED);
2113 kiblnd_peer_alive(peer);
2115 /* Add conn to peer's list and nuke any dangling conns from a different
2116 * peer instance... */
2117 kiblnd_conn_addref(conn); /* +1 ref for ibc_list */
2118 cfs_list_add(&conn->ibc_list, &peer->ibp_conns);
2120 peer->ibp_connecting--;
2122 peer->ibp_accepting--;
2124 if (peer->ibp_version == 0) {
2125 peer->ibp_version = conn->ibc_version;
2126 peer->ibp_incarnation = conn->ibc_incarnation;
2129 if (peer->ibp_version != conn->ibc_version ||
2130 peer->ibp_incarnation != conn->ibc_incarnation) {
2131 kiblnd_close_stale_conns_locked(peer, conn->ibc_version,
2132 conn->ibc_incarnation);
2133 peer->ibp_version = conn->ibc_version;
2134 peer->ibp_incarnation = conn->ibc_incarnation;
2137 /* grab pending txs while I have the lock */
2138 cfs_list_add(&txs, &peer->ibp_tx_queue);
2139 cfs_list_del_init(&peer->ibp_tx_queue);
2141 if (!kiblnd_peer_active(peer) || /* peer has been deleted */
2142 conn->ibc_comms_error != 0) { /* error has happened already */
2143 lnet_ni_t *ni = peer->ibp_ni;
2145 /* start to shut down connection */
2146 kiblnd_close_conn_locked(conn, -ECONNABORTED);
2147 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2149 kiblnd_txlist_done(ni, &txs, -ECONNABORTED);
2154 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2156 /* Schedule blocked txs */
2157 spin_lock(&conn->ibc_lock);
2158 while (!cfs_list_empty(&txs)) {
2159 tx = cfs_list_entry(txs.next, kib_tx_t, tx_list);
2160 cfs_list_del(&tx->tx_list);
2162 kiblnd_queue_tx_locked(tx, conn);
2164 spin_unlock(&conn->ibc_lock);
2166 kiblnd_check_sends(conn);
2168 /* schedule blocked rxs */
2169 kiblnd_handle_early_rxs(conn);
2173 kiblnd_reject(struct rdma_cm_id *cmid, kib_rej_t *rej)
2177 rc = rdma_reject(cmid, rej, sizeof(*rej));
2180 CWARN("Error %d sending reject\n", rc);
2184 kiblnd_passive_connect (struct rdma_cm_id *cmid, void *priv, int priv_nob)
2186 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2187 kib_msg_t *reqmsg = priv;
2193 lnet_ni_t *ni = NULL;
2194 kib_net_t *net = NULL;
2196 struct rdma_conn_param cp;
2198 int version = IBLND_MSG_VERSION;
2199 unsigned long flags;
2201 struct sockaddr_in *peer_addr;
2202 LASSERT (!in_interrupt());
2204 /* cmid inherits 'context' from the corresponding listener id */
2205 ibdev = (kib_dev_t *)cmid->context;
2206 LASSERT (ibdev != NULL);
2208 memset(&rej, 0, sizeof(rej));
2209 rej.ibr_magic = IBLND_MSG_MAGIC;
2210 rej.ibr_why = IBLND_REJECT_FATAL;
2211 rej.ibr_cp.ibcp_max_msg_size = IBLND_MSG_SIZE;
2213 peer_addr = (struct sockaddr_in *)&(cmid->route.addr.dst_addr);
2214 if (*kiblnd_tunables.kib_require_priv_port &&
2215 ntohs(peer_addr->sin_port) >= PROT_SOCK) {
2216 __u32 ip = ntohl(peer_addr->sin_addr.s_addr);
2217 CERROR("Peer's port (%u.%u.%u.%u:%hu) is not privileged\n",
2218 HIPQUAD(ip), ntohs(peer_addr->sin_port));
2222 if (priv_nob < offsetof(kib_msg_t, ibm_type)) {
2223 CERROR("Short connection request\n");
2227 /* Future protocol version compatibility support! If the
2228 * o2iblnd-specific protocol changes, or when LNET unifies
2229 * protocols over all LNDs, the initial connection will
2230 * negotiate a protocol version. I trap this here to avoid
2231 * console errors; the reject tells the peer which protocol I
2233 if (reqmsg->ibm_magic == LNET_PROTO_MAGIC ||
2234 reqmsg->ibm_magic == __swab32(LNET_PROTO_MAGIC))
2236 if (reqmsg->ibm_magic == IBLND_MSG_MAGIC &&
2237 reqmsg->ibm_version != IBLND_MSG_VERSION &&
2238 reqmsg->ibm_version != IBLND_MSG_VERSION_1)
2240 if (reqmsg->ibm_magic == __swab32(IBLND_MSG_MAGIC) &&
2241 reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION) &&
2242 reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION_1))
2245 rc = kiblnd_unpack_msg(reqmsg, priv_nob);
2247 CERROR("Can't parse connection request: %d\n", rc);
2251 nid = reqmsg->ibm_srcnid;
2252 ni = lnet_net2ni(LNET_NIDNET(reqmsg->ibm_dstnid));
2255 net = (kib_net_t *)ni->ni_data;
2256 rej.ibr_incarnation = net->ibn_incarnation;
2259 if (ni == NULL || /* no matching net */
2260 ni->ni_nid != reqmsg->ibm_dstnid || /* right NET, wrong NID! */
2261 net->ibn_dev != ibdev) { /* wrong device */
2262 CERROR("Can't accept %s on %s (%s:%d:%u.%u.%u.%u): "
2263 "bad dst nid %s\n", libcfs_nid2str(nid),
2264 ni == NULL ? "NA" : libcfs_nid2str(ni->ni_nid),
2265 ibdev->ibd_ifname, ibdev->ibd_nnets,
2266 HIPQUAD(ibdev->ibd_ifip),
2267 libcfs_nid2str(reqmsg->ibm_dstnid));
2272 /* check time stamp as soon as possible */
2273 if (reqmsg->ibm_dststamp != 0 &&
2274 reqmsg->ibm_dststamp != net->ibn_incarnation) {
2275 CWARN("Stale connection request\n");
2276 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2280 /* I can accept peer's version */
2281 version = reqmsg->ibm_version;
2283 if (reqmsg->ibm_type != IBLND_MSG_CONNREQ) {
2284 CERROR("Unexpected connreq msg type: %x from %s\n",
2285 reqmsg->ibm_type, libcfs_nid2str(nid));
2289 if (reqmsg->ibm_u.connparams.ibcp_queue_depth !=
2290 IBLND_MSG_QUEUE_SIZE(version)) {
2291 CERROR("Can't accept %s: incompatible queue depth %d (%d wanted)\n",
2292 libcfs_nid2str(nid), reqmsg->ibm_u.connparams.ibcp_queue_depth,
2293 IBLND_MSG_QUEUE_SIZE(version));
2295 if (version == IBLND_MSG_VERSION)
2296 rej.ibr_why = IBLND_REJECT_MSG_QUEUE_SIZE;
2301 if (reqmsg->ibm_u.connparams.ibcp_max_frags !=
2302 IBLND_RDMA_FRAGS(version)) {
2303 CERROR("Can't accept %s(version %x): "
2304 "incompatible max_frags %d (%d wanted)\n",
2305 libcfs_nid2str(nid), version,
2306 reqmsg->ibm_u.connparams.ibcp_max_frags,
2307 IBLND_RDMA_FRAGS(version));
2309 if (version == IBLND_MSG_VERSION)
2310 rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2316 if (reqmsg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2317 CERROR("Can't accept %s: message size %d too big (%d max)\n",
2318 libcfs_nid2str(nid),
2319 reqmsg->ibm_u.connparams.ibcp_max_msg_size,
2324 /* assume 'nid' is a new peer; create */
2325 rc = kiblnd_create_peer(ni, &peer, nid);
2327 CERROR("Can't create peer for %s\n", libcfs_nid2str(nid));
2328 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2332 write_lock_irqsave(g_lock, flags);
2334 peer2 = kiblnd_find_peer_locked(nid);
2335 if (peer2 != NULL) {
2336 if (peer2->ibp_version == 0) {
2337 peer2->ibp_version = version;
2338 peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2341 /* not the guy I've talked with */
2342 if (peer2->ibp_incarnation != reqmsg->ibm_srcstamp ||
2343 peer2->ibp_version != version) {
2344 kiblnd_close_peer_conns_locked(peer2, -ESTALE);
2345 write_unlock_irqrestore(g_lock, flags);
2347 CWARN("Conn stale %s [old ver: %x, new ver: %x]\n",
2348 libcfs_nid2str(nid), peer2->ibp_version, version);
2350 kiblnd_peer_decref(peer);
2351 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2355 /* tie-break connection race in favour of the higher NID */
2356 if (peer2->ibp_connecting != 0 &&
2358 write_unlock_irqrestore(g_lock, flags);
2360 CWARN("Conn race %s\n", libcfs_nid2str(peer2->ibp_nid));
2362 kiblnd_peer_decref(peer);
2363 rej.ibr_why = IBLND_REJECT_CONN_RACE;
2367 peer2->ibp_accepting++;
2368 kiblnd_peer_addref(peer2);
2370 write_unlock_irqrestore(g_lock, flags);
2371 kiblnd_peer_decref(peer);
2374 /* Brand new peer */
2375 LASSERT (peer->ibp_accepting == 0);
2376 LASSERT (peer->ibp_version == 0 &&
2377 peer->ibp_incarnation == 0);
2379 peer->ibp_accepting = 1;
2380 peer->ibp_version = version;
2381 peer->ibp_incarnation = reqmsg->ibm_srcstamp;
2383 /* I have a ref on ni that prevents it being shutdown */
2384 LASSERT (net->ibn_shutdown == 0);
2386 kiblnd_peer_addref(peer);
2387 cfs_list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid));
2389 write_unlock_irqrestore(g_lock, flags);
2392 conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_PASSIVE_WAIT, version);
2394 kiblnd_peer_connect_failed(peer, 0, -ENOMEM);
2395 kiblnd_peer_decref(peer);
2396 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2400 /* conn now "owns" cmid, so I return success from here on to ensure the
2401 * CM callback doesn't destroy cmid. */
2403 conn->ibc_incarnation = reqmsg->ibm_srcstamp;
2404 conn->ibc_credits = IBLND_MSG_QUEUE_SIZE(version);
2405 conn->ibc_reserved_credits = IBLND_MSG_QUEUE_SIZE(version);
2406 LASSERT (conn->ibc_credits + conn->ibc_reserved_credits + IBLND_OOB_MSGS(version)
2407 <= IBLND_RX_MSGS(version));
2409 ackmsg = &conn->ibc_connvars->cv_msg;
2410 memset(ackmsg, 0, sizeof(*ackmsg));
2412 kiblnd_init_msg(ackmsg, IBLND_MSG_CONNACK,
2413 sizeof(ackmsg->ibm_u.connparams));
2414 ackmsg->ibm_u.connparams.ibcp_queue_depth = IBLND_MSG_QUEUE_SIZE(version);
2415 ackmsg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2416 ackmsg->ibm_u.connparams.ibcp_max_frags = IBLND_RDMA_FRAGS(version);
2418 kiblnd_pack_msg(ni, ackmsg, version, 0, nid, reqmsg->ibm_srcstamp);
2420 memset(&cp, 0, sizeof(cp));
2421 cp.private_data = ackmsg;
2422 cp.private_data_len = ackmsg->ibm_nob;
2423 cp.responder_resources = 0; /* No atomic ops or RDMA reads */
2424 cp.initiator_depth = 0;
2425 cp.flow_control = 1;
2426 cp.retry_count = *kiblnd_tunables.kib_retry_count;
2427 cp.rnr_retry_count = *kiblnd_tunables.kib_rnr_retry_count;
2429 CDEBUG(D_NET, "Accept %s\n", libcfs_nid2str(nid));
2431 rc = rdma_accept(cmid, &cp);
2433 CERROR("Can't accept %s: %d\n", libcfs_nid2str(nid), rc);
2434 rej.ibr_version = version;
2435 rej.ibr_why = IBLND_REJECT_FATAL;
2437 kiblnd_reject(cmid, &rej);
2438 kiblnd_connreq_done(conn, rc);
2439 kiblnd_conn_decref(conn);
2449 rej.ibr_version = version;
2450 rej.ibr_cp.ibcp_queue_depth = IBLND_MSG_QUEUE_SIZE(version);
2451 rej.ibr_cp.ibcp_max_frags = IBLND_RDMA_FRAGS(version);
2452 kiblnd_reject(cmid, &rej);
2454 return -ECONNREFUSED;
2458 kiblnd_reconnect (kib_conn_t *conn, int version,
2459 __u64 incarnation, int why, kib_connparams_t *cp)
2461 kib_peer_t *peer = conn->ibc_peer;
2464 unsigned long flags;
2466 LASSERT (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2467 LASSERT (peer->ibp_connecting > 0); /* 'conn' at least */
2469 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2471 /* retry connection if it's still needed and no other connection
2472 * attempts (active or passive) are in progress
2473 * NB: reconnect is still needed even when ibp_tx_queue is
2474 * empty if ibp_version != version because reconnect may be
2475 * initiated by kiblnd_query() */
2476 if ((!cfs_list_empty(&peer->ibp_tx_queue) ||
2477 peer->ibp_version != version) &&
2478 peer->ibp_connecting == 1 &&
2479 peer->ibp_accepting == 0) {
2481 peer->ibp_connecting++;
2483 peer->ibp_version = version;
2484 peer->ibp_incarnation = incarnation;
2487 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2497 case IBLND_REJECT_CONN_STALE:
2501 case IBLND_REJECT_CONN_RACE:
2502 reason = "conn race";
2505 case IBLND_REJECT_CONN_UNCOMPAT:
2506 reason = "version negotiation";
2510 CNETERR("%s: retrying (%s), %x, %x, "
2511 "queue_dep: %d, max_frag: %d, msg_size: %d\n",
2512 libcfs_nid2str(peer->ibp_nid),
2513 reason, IBLND_MSG_VERSION, version,
2514 cp != NULL? cp->ibcp_queue_depth :IBLND_MSG_QUEUE_SIZE(version),
2515 cp != NULL? cp->ibcp_max_frags : IBLND_RDMA_FRAGS(version),
2516 cp != NULL? cp->ibcp_max_msg_size: IBLND_MSG_SIZE);
2518 kiblnd_connect_peer(peer);
2522 kiblnd_rejected (kib_conn_t *conn, int reason, void *priv, int priv_nob)
2524 kib_peer_t *peer = conn->ibc_peer;
2526 LASSERT (!in_interrupt());
2527 LASSERT (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2530 case IB_CM_REJ_STALE_CONN:
2531 kiblnd_reconnect(conn, IBLND_MSG_VERSION, 0,
2532 IBLND_REJECT_CONN_STALE, NULL);
2535 case IB_CM_REJ_INVALID_SERVICE_ID:
2536 CNETERR("%s rejected: no listener at %d\n",
2537 libcfs_nid2str(peer->ibp_nid),
2538 *kiblnd_tunables.kib_service);
2541 case IB_CM_REJ_CONSUMER_DEFINED:
2542 if (priv_nob >= offsetof(kib_rej_t, ibr_padding)) {
2543 kib_rej_t *rej = priv;
2544 kib_connparams_t *cp = NULL;
2546 __u64 incarnation = -1;
2548 /* NB. default incarnation is -1 because:
2549 * a) V1 will ignore dst incarnation in connreq.
2550 * b) V2 will provide incarnation while rejecting me,
2551 * -1 will be overwrote.
2553 * if I try to connect to a V1 peer with V2 protocol,
2554 * it rejected me then upgrade to V2, I have no idea
2555 * about the upgrading and try to reconnect with V1,
2556 * in this case upgraded V2 can find out I'm trying to
2557 * talk to the old guy and reject me(incarnation is -1).
2560 if (rej->ibr_magic == __swab32(IBLND_MSG_MAGIC) ||
2561 rej->ibr_magic == __swab32(LNET_PROTO_MAGIC)) {
2562 __swab32s(&rej->ibr_magic);
2563 __swab16s(&rej->ibr_version);
2567 if (priv_nob >= sizeof(kib_rej_t) &&
2568 rej->ibr_version > IBLND_MSG_VERSION_1) {
2569 /* priv_nob is always 148 in current version
2570 * of OFED, so we still need to check version.
2571 * (define of IB_CM_REJ_PRIVATE_DATA_SIZE) */
2575 __swab64s(&rej->ibr_incarnation);
2576 __swab16s(&cp->ibcp_queue_depth);
2577 __swab16s(&cp->ibcp_max_frags);
2578 __swab32s(&cp->ibcp_max_msg_size);
2581 incarnation = rej->ibr_incarnation;
2584 if (rej->ibr_magic != IBLND_MSG_MAGIC &&
2585 rej->ibr_magic != LNET_PROTO_MAGIC) {
2586 CERROR("%s rejected: consumer defined fatal error\n",
2587 libcfs_nid2str(peer->ibp_nid));
2591 if (rej->ibr_version != IBLND_MSG_VERSION &&
2592 rej->ibr_version != IBLND_MSG_VERSION_1) {
2593 CERROR("%s rejected: o2iblnd version %x error\n",
2594 libcfs_nid2str(peer->ibp_nid),
2599 if (rej->ibr_why == IBLND_REJECT_FATAL &&
2600 rej->ibr_version == IBLND_MSG_VERSION_1) {
2601 CDEBUG(D_NET, "rejected by old version peer %s: %x\n",
2602 libcfs_nid2str(peer->ibp_nid), rej->ibr_version);
2604 if (conn->ibc_version != IBLND_MSG_VERSION_1)
2605 rej->ibr_why = IBLND_REJECT_CONN_UNCOMPAT;
2608 switch (rej->ibr_why) {
2609 case IBLND_REJECT_CONN_RACE:
2610 case IBLND_REJECT_CONN_STALE:
2611 case IBLND_REJECT_CONN_UNCOMPAT:
2612 kiblnd_reconnect(conn, rej->ibr_version,
2613 incarnation, rej->ibr_why, cp);
2616 case IBLND_REJECT_MSG_QUEUE_SIZE:
2617 CERROR("%s rejected: incompatible message queue depth %d, %d\n",
2618 libcfs_nid2str(peer->ibp_nid), cp->ibcp_queue_depth,
2619 IBLND_MSG_QUEUE_SIZE(conn->ibc_version));
2622 case IBLND_REJECT_RDMA_FRAGS:
2623 CERROR("%s rejected: incompatible # of RDMA fragments %d, %d\n",
2624 libcfs_nid2str(peer->ibp_nid), cp->ibcp_max_frags,
2625 IBLND_RDMA_FRAGS(conn->ibc_version));
2628 case IBLND_REJECT_NO_RESOURCES:
2629 CERROR("%s rejected: o2iblnd no resources\n",
2630 libcfs_nid2str(peer->ibp_nid));
2633 case IBLND_REJECT_FATAL:
2634 CERROR("%s rejected: o2iblnd fatal error\n",
2635 libcfs_nid2str(peer->ibp_nid));
2639 CERROR("%s rejected: o2iblnd reason %d\n",
2640 libcfs_nid2str(peer->ibp_nid),
2648 CNETERR("%s rejected: reason %d, size %d\n",
2649 libcfs_nid2str(peer->ibp_nid), reason, priv_nob);
2653 kiblnd_connreq_done(conn, -ECONNREFUSED);
2657 kiblnd_check_connreply (kib_conn_t *conn, void *priv, int priv_nob)
2659 kib_peer_t *peer = conn->ibc_peer;
2660 lnet_ni_t *ni = peer->ibp_ni;
2661 kib_net_t *net = ni->ni_data;
2662 kib_msg_t *msg = priv;
2663 int ver = conn->ibc_version;
2664 int rc = kiblnd_unpack_msg(msg, priv_nob);
2665 unsigned long flags;
2667 LASSERT (net != NULL);
2670 CERROR("Can't unpack connack from %s: %d\n",
2671 libcfs_nid2str(peer->ibp_nid), rc);
2675 if (msg->ibm_type != IBLND_MSG_CONNACK) {
2676 CERROR("Unexpected message %d from %s\n",
2677 msg->ibm_type, libcfs_nid2str(peer->ibp_nid));
2682 if (ver != msg->ibm_version) {
2683 CERROR("%s replied version %x is different with "
2684 "requested version %x\n",
2685 libcfs_nid2str(peer->ibp_nid), msg->ibm_version, ver);
2690 if (msg->ibm_u.connparams.ibcp_queue_depth !=
2691 IBLND_MSG_QUEUE_SIZE(ver)) {
2692 CERROR("%s has incompatible queue depth %d(%d wanted)\n",
2693 libcfs_nid2str(peer->ibp_nid),
2694 msg->ibm_u.connparams.ibcp_queue_depth,
2695 IBLND_MSG_QUEUE_SIZE(ver));
2700 if (msg->ibm_u.connparams.ibcp_max_frags !=
2701 IBLND_RDMA_FRAGS(ver)) {
2702 CERROR("%s has incompatible max_frags %d (%d wanted)\n",
2703 libcfs_nid2str(peer->ibp_nid),
2704 msg->ibm_u.connparams.ibcp_max_frags,
2705 IBLND_RDMA_FRAGS(ver));
2710 if (msg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2711 CERROR("%s max message size %d too big (%d max)\n",
2712 libcfs_nid2str(peer->ibp_nid),
2713 msg->ibm_u.connparams.ibcp_max_msg_size,
2719 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2720 if (msg->ibm_dstnid == ni->ni_nid &&
2721 msg->ibm_dststamp == net->ibn_incarnation)
2725 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2728 CERROR("Bad connection reply from %s, rc = %d, "
2729 "version: %x max_frags: %d\n",
2730 libcfs_nid2str(peer->ibp_nid), rc,
2731 msg->ibm_version, msg->ibm_u.connparams.ibcp_max_frags);
2735 conn->ibc_incarnation = msg->ibm_srcstamp;
2737 conn->ibc_reserved_credits = IBLND_MSG_QUEUE_SIZE(ver);
2738 LASSERT (conn->ibc_credits + conn->ibc_reserved_credits + IBLND_OOB_MSGS(ver)
2739 <= IBLND_RX_MSGS(ver));
2741 kiblnd_connreq_done(conn, 0);
2745 /* NB My QP has already established itself, so I handle anything going
2746 * wrong here by setting ibc_comms_error.
2747 * kiblnd_connreq_done(0) moves the conn state to ESTABLISHED, but then
2748 * immediately tears it down. */
2751 conn->ibc_comms_error = rc;
2752 kiblnd_connreq_done(conn, 0);
2756 kiblnd_active_connect (struct rdma_cm_id *cmid)
2758 kib_peer_t *peer = (kib_peer_t *)cmid->context;
2761 struct rdma_conn_param cp;
2764 unsigned long flags;
2767 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2769 incarnation = peer->ibp_incarnation;
2770 version = (peer->ibp_version == 0) ? IBLND_MSG_VERSION :
2773 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2775 conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_ACTIVE_CONNECT, version);
2777 kiblnd_peer_connect_failed(peer, 1, -ENOMEM);
2778 kiblnd_peer_decref(peer); /* lose cmid's ref */
2782 /* conn "owns" cmid now, so I return success from here on to ensure the
2783 * CM callback doesn't destroy cmid. conn also takes over cmid's ref
2786 msg = &conn->ibc_connvars->cv_msg;
2788 memset(msg, 0, sizeof(*msg));
2789 kiblnd_init_msg(msg, IBLND_MSG_CONNREQ, sizeof(msg->ibm_u.connparams));
2790 msg->ibm_u.connparams.ibcp_queue_depth = IBLND_MSG_QUEUE_SIZE(version);
2791 msg->ibm_u.connparams.ibcp_max_frags = IBLND_RDMA_FRAGS(version);
2792 msg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2794 kiblnd_pack_msg(peer->ibp_ni, msg, version,
2795 0, peer->ibp_nid, incarnation);
2797 memset(&cp, 0, sizeof(cp));
2798 cp.private_data = msg;
2799 cp.private_data_len = msg->ibm_nob;
2800 cp.responder_resources = 0; /* No atomic ops or RDMA reads */
2801 cp.initiator_depth = 0;
2802 cp.flow_control = 1;
2803 cp.retry_count = *kiblnd_tunables.kib_retry_count;
2804 cp.rnr_retry_count = *kiblnd_tunables.kib_rnr_retry_count;
2806 LASSERT(cmid->context == (void *)conn);
2807 LASSERT(conn->ibc_cmid == cmid);
2809 rc = rdma_connect(cmid, &cp);
2811 CERROR("Can't connect to %s: %d\n",
2812 libcfs_nid2str(peer->ibp_nid), rc);
2813 kiblnd_connreq_done(conn, rc);
2814 kiblnd_conn_decref(conn);
2821 kiblnd_cm_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2827 switch (event->event) {
2829 CERROR("Unexpected event: %d, status: %d\n",
2830 event->event, event->status);
2833 case RDMA_CM_EVENT_CONNECT_REQUEST:
2834 /* destroy cmid on failure */
2835 rc = kiblnd_passive_connect(cmid,
2836 (void *)KIBLND_CONN_PARAM(event),
2837 KIBLND_CONN_PARAM_LEN(event));
2838 CDEBUG(D_NET, "connreq: %d\n", rc);
2841 case RDMA_CM_EVENT_ADDR_ERROR:
2842 peer = (kib_peer_t *)cmid->context;
2843 CNETERR("%s: ADDR ERROR %d\n",
2844 libcfs_nid2str(peer->ibp_nid), event->status);
2845 kiblnd_peer_connect_failed(peer, 1, -EHOSTUNREACH);
2846 kiblnd_peer_decref(peer);
2847 return -EHOSTUNREACH; /* rc != 0 destroys cmid */
2849 case RDMA_CM_EVENT_ADDR_RESOLVED:
2850 peer = (kib_peer_t *)cmid->context;
2852 CDEBUG(D_NET,"%s Addr resolved: %d\n",
2853 libcfs_nid2str(peer->ibp_nid), event->status);
2855 if (event->status != 0) {
2856 CNETERR("Can't resolve address for %s: %d\n",
2857 libcfs_nid2str(peer->ibp_nid), event->status);
2860 rc = rdma_resolve_route(
2861 cmid, *kiblnd_tunables.kib_timeout * 1000);
2864 /* Can't initiate route resolution */
2865 CERROR("Can't resolve route for %s: %d\n",
2866 libcfs_nid2str(peer->ibp_nid), rc);
2868 kiblnd_peer_connect_failed(peer, 1, rc);
2869 kiblnd_peer_decref(peer);
2870 return rc; /* rc != 0 destroys cmid */
2872 case RDMA_CM_EVENT_ROUTE_ERROR:
2873 peer = (kib_peer_t *)cmid->context;
2874 CNETERR("%s: ROUTE ERROR %d\n",
2875 libcfs_nid2str(peer->ibp_nid), event->status);
2876 kiblnd_peer_connect_failed(peer, 1, -EHOSTUNREACH);
2877 kiblnd_peer_decref(peer);
2878 return -EHOSTUNREACH; /* rc != 0 destroys cmid */
2880 case RDMA_CM_EVENT_ROUTE_RESOLVED:
2881 peer = (kib_peer_t *)cmid->context;
2882 CDEBUG(D_NET,"%s Route resolved: %d\n",
2883 libcfs_nid2str(peer->ibp_nid), event->status);
2885 if (event->status == 0)
2886 return kiblnd_active_connect(cmid);
2888 CNETERR("Can't resolve route for %s: %d\n",
2889 libcfs_nid2str(peer->ibp_nid), event->status);
2890 kiblnd_peer_connect_failed(peer, 1, event->status);
2891 kiblnd_peer_decref(peer);
2892 return event->status; /* rc != 0 destroys cmid */
2894 case RDMA_CM_EVENT_UNREACHABLE:
2895 conn = (kib_conn_t *)cmid->context;
2896 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
2897 conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
2898 CNETERR("%s: UNREACHABLE %d\n",
2899 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
2900 kiblnd_connreq_done(conn, -ENETDOWN);
2901 kiblnd_conn_decref(conn);
2904 case RDMA_CM_EVENT_CONNECT_ERROR:
2905 conn = (kib_conn_t *)cmid->context;
2906 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
2907 conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
2908 CNETERR("%s: CONNECT ERROR %d\n",
2909 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
2910 kiblnd_connreq_done(conn, -ENOTCONN);
2911 kiblnd_conn_decref(conn);
2914 case RDMA_CM_EVENT_REJECTED:
2915 conn = (kib_conn_t *)cmid->context;
2916 switch (conn->ibc_state) {
2920 case IBLND_CONN_PASSIVE_WAIT:
2921 CERROR ("%s: REJECTED %d\n",
2922 libcfs_nid2str(conn->ibc_peer->ibp_nid),
2924 kiblnd_connreq_done(conn, -ECONNRESET);
2927 case IBLND_CONN_ACTIVE_CONNECT:
2928 kiblnd_rejected(conn, event->status,
2929 (void *)KIBLND_CONN_PARAM(event),
2930 KIBLND_CONN_PARAM_LEN(event));
2933 kiblnd_conn_decref(conn);
2936 case RDMA_CM_EVENT_ESTABLISHED:
2937 conn = (kib_conn_t *)cmid->context;
2938 switch (conn->ibc_state) {
2942 case IBLND_CONN_PASSIVE_WAIT:
2943 CDEBUG(D_NET, "ESTABLISHED (passive): %s\n",
2944 libcfs_nid2str(conn->ibc_peer->ibp_nid));
2945 kiblnd_connreq_done(conn, 0);
2948 case IBLND_CONN_ACTIVE_CONNECT:
2949 CDEBUG(D_NET, "ESTABLISHED(active): %s\n",
2950 libcfs_nid2str(conn->ibc_peer->ibp_nid));
2951 kiblnd_check_connreply(conn,
2952 (void *)KIBLND_CONN_PARAM(event),
2953 KIBLND_CONN_PARAM_LEN(event));
2956 /* net keeps its ref on conn! */
2959 #ifdef HAVE_OFED_RDMA_CMEV_TIMEWAIT_EXIT
2960 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
2961 CDEBUG(D_NET, "Ignore TIMEWAIT_EXIT event\n");
2964 case RDMA_CM_EVENT_DISCONNECTED:
2965 conn = (kib_conn_t *)cmid->context;
2966 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
2967 CERROR("%s DISCONNECTED\n",
2968 libcfs_nid2str(conn->ibc_peer->ibp_nid));
2969 kiblnd_connreq_done(conn, -ECONNRESET);
2971 kiblnd_close_conn(conn, 0);
2973 kiblnd_conn_decref(conn);
2974 cmid->context = NULL;
2977 case RDMA_CM_EVENT_DEVICE_REMOVAL:
2978 LCONSOLE_ERROR_MSG(0x131,
2979 "Received notification of device removal\n"
2980 "Please shutdown LNET to allow this to proceed\n");
2981 /* Can't remove network from underneath LNET for now, so I have
2985 #ifdef HAVE_OFED_RDMA_CMEV_ADDRCHANGE
2986 case RDMA_CM_EVENT_ADDR_CHANGE:
2987 LCONSOLE_INFO("Physical link changed (eg hca/port)\n");
2994 kiblnd_check_txs_locked(kib_conn_t *conn, cfs_list_t *txs)
2999 cfs_list_for_each (ttmp, txs) {
3000 tx = cfs_list_entry (ttmp, kib_tx_t, tx_list);
3002 if (txs != &conn->ibc_active_txs) {
3003 LASSERT (tx->tx_queued);
3005 LASSERT (!tx->tx_queued);
3006 LASSERT (tx->tx_waiting || tx->tx_sending != 0);
3009 if (cfs_time_aftereq (jiffies, tx->tx_deadline)) {
3010 CERROR("Timed out tx: %s, %lu seconds\n",
3011 kiblnd_queue2str(conn, txs),
3012 cfs_duration_sec(jiffies - tx->tx_deadline));
3021 kiblnd_conn_timed_out_locked(kib_conn_t *conn)
3023 return kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue) ||
3024 kiblnd_check_txs_locked(conn, &conn->ibc_tx_noops) ||
3025 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_rsrvd) ||
3026 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_nocred) ||
3027 kiblnd_check_txs_locked(conn, &conn->ibc_active_txs);
3031 kiblnd_check_conns (int idx)
3033 CFS_LIST_HEAD (closes);
3034 CFS_LIST_HEAD (checksends);
3035 cfs_list_t *peers = &kiblnd_data.kib_peers[idx];
3040 unsigned long flags;
3042 /* NB. We expect to have a look at all the peers and not find any
3043 * RDMAs to time out, so we just use a shared lock while we
3045 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3047 cfs_list_for_each (ptmp, peers) {
3048 peer = cfs_list_entry (ptmp, kib_peer_t, ibp_list);
3050 cfs_list_for_each (ctmp, &peer->ibp_conns) {
3054 conn = cfs_list_entry(ctmp, kib_conn_t, ibc_list);
3056 LASSERT (conn->ibc_state == IBLND_CONN_ESTABLISHED);
3058 spin_lock(&conn->ibc_lock);
3060 sendnoop = kiblnd_need_noop(conn);
3061 timedout = kiblnd_conn_timed_out_locked(conn);
3062 if (!sendnoop && !timedout) {
3063 spin_unlock(&conn->ibc_lock);
3068 CERROR("Timed out RDMA with %s (%lu): "
3069 "c: %u, oc: %u, rc: %u\n",
3070 libcfs_nid2str(peer->ibp_nid),
3071 cfs_duration_sec(cfs_time_current() -
3072 peer->ibp_last_alive),
3074 conn->ibc_outstanding_credits,
3075 conn->ibc_reserved_credits);
3076 cfs_list_add(&conn->ibc_connd_list, &closes);
3078 cfs_list_add(&conn->ibc_connd_list,
3081 /* +ref for 'closes' or 'checksends' */
3082 kiblnd_conn_addref(conn);
3084 spin_unlock(&conn->ibc_lock);
3088 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3090 /* Handle timeout by closing the whole
3091 * connection. We can only be sure RDMA activity
3092 * has ceased once the QP has been modified. */
3093 while (!cfs_list_empty(&closes)) {
3094 conn = cfs_list_entry(closes.next,
3095 kib_conn_t, ibc_connd_list);
3096 cfs_list_del(&conn->ibc_connd_list);
3097 kiblnd_close_conn(conn, -ETIMEDOUT);
3098 kiblnd_conn_decref(conn);
3101 /* In case we have enough credits to return via a
3102 * NOOP, but there were no non-blocking tx descs
3103 * free to do it last time... */
3104 while (!cfs_list_empty(&checksends)) {
3105 conn = cfs_list_entry(checksends.next,
3106 kib_conn_t, ibc_connd_list);
3107 cfs_list_del(&conn->ibc_connd_list);
3108 kiblnd_check_sends(conn);
3109 kiblnd_conn_decref(conn);
3114 kiblnd_disconnect_conn (kib_conn_t *conn)
3116 LASSERT (!in_interrupt());
3117 LASSERT (current == kiblnd_data.kib_connd);
3118 LASSERT (conn->ibc_state == IBLND_CONN_CLOSING);
3120 rdma_disconnect(conn->ibc_cmid);
3121 kiblnd_finalise_conn(conn);
3123 kiblnd_peer_notify(conn->ibc_peer);
3127 kiblnd_connd (void *arg)
3130 unsigned long flags;
3136 unsigned long deadline = jiffies;
3138 cfs_block_allsigs ();
3140 init_waitqueue_entry_current (&wait);
3141 kiblnd_data.kib_connd = current;
3143 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3145 while (!kiblnd_data.kib_shutdown) {
3149 if (!cfs_list_empty (&kiblnd_data.kib_connd_zombies)) {
3150 conn = cfs_list_entry(kiblnd_data. \
3151 kib_connd_zombies.next,
3152 kib_conn_t, ibc_list);
3153 cfs_list_del(&conn->ibc_list);
3155 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock,
3159 kiblnd_destroy_conn(conn);
3161 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3164 if (!cfs_list_empty(&kiblnd_data.kib_connd_conns)) {
3165 conn = cfs_list_entry(kiblnd_data.kib_connd_conns.next,
3166 kib_conn_t, ibc_list);
3167 cfs_list_del(&conn->ibc_list);
3169 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock,
3173 kiblnd_disconnect_conn(conn);
3174 kiblnd_conn_decref(conn);
3176 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3179 /* careful with the jiffy wrap... */
3180 timeout = (int)(deadline - jiffies);
3184 int chunk = kiblnd_data.kib_peer_hash_size;
3186 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
3189 /* Time to check for RDMA timeouts on a few more
3190 * peers: I do checks every 'p' seconds on a
3191 * proportion of the peer table and I need to check
3192 * every connection 'n' times within a timeout
3193 * interval, to ensure I detect a timeout on any
3194 * connection within (n+1)/n times the timeout
3197 if (*kiblnd_tunables.kib_timeout > n * p)
3198 chunk = (chunk * n * p) /
3199 *kiblnd_tunables.kib_timeout;
3203 for (i = 0; i < chunk; i++) {
3204 kiblnd_check_conns(peer_index);
3205 peer_index = (peer_index + 1) %
3206 kiblnd_data.kib_peer_hash_size;
3210 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3216 /* Nothing to do for 'timeout' */
3217 set_current_state(TASK_INTERRUPTIBLE);
3218 add_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3219 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
3221 waitq_timedwait(&wait, TASK_INTERRUPTIBLE, timeout);
3223 set_current_state(TASK_RUNNING);
3224 remove_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3225 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3228 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
3230 kiblnd_thread_fini();
3235 kiblnd_qp_event(struct ib_event *event, void *arg)
3237 kib_conn_t *conn = arg;
3239 switch (event->event) {
3240 case IB_EVENT_COMM_EST:
3241 CDEBUG(D_NET, "%s established\n",
3242 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3246 CERROR("%s: Async QP event type %d\n",
3247 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3253 kiblnd_complete (struct ib_wc *wc)
3255 switch (kiblnd_wreqid2type(wc->wr_id)) {
3259 case IBLND_WID_RDMA:
3260 /* We only get RDMA completion notification if it fails. All
3261 * subsequent work items, including the final SEND will fail
3262 * too. However we can't print out any more info about the
3263 * failing RDMA because 'tx' might be back on the idle list or
3264 * even reused already if we didn't manage to post all our work
3266 CNETERR("RDMA (tx: %p) failed: %d\n",
3267 kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3271 kiblnd_tx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3275 kiblnd_rx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status,
3282 kiblnd_cq_completion(struct ib_cq *cq, void *arg)
3284 /* NB I'm not allowed to schedule this conn once its refcount has
3285 * reached 0. Since fundamentally I'm racing with scheduler threads
3286 * consuming my CQ I could be called after all completions have
3287 * occurred. But in this case, ibc_nrx == 0 && ibc_nsends_posted == 0
3288 * and this CQ is about to be destroyed so I NOOP. */
3289 kib_conn_t *conn = (kib_conn_t *)arg;
3290 struct kib_sched_info *sched = conn->ibc_sched;
3291 unsigned long flags;
3293 LASSERT(cq == conn->ibc_cq);
3295 spin_lock_irqsave(&sched->ibs_lock, flags);
3297 conn->ibc_ready = 1;
3299 if (!conn->ibc_scheduled &&
3300 (conn->ibc_nrx > 0 ||
3301 conn->ibc_nsends_posted > 0)) {
3302 kiblnd_conn_addref(conn); /* +1 ref for sched_conns */
3303 conn->ibc_scheduled = 1;
3304 cfs_list_add_tail(&conn->ibc_sched_list, &sched->ibs_conns);
3306 if (waitqueue_active(&sched->ibs_waitq))
3307 wake_up(&sched->ibs_waitq);
3310 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3314 kiblnd_cq_event(struct ib_event *event, void *arg)
3316 kib_conn_t *conn = arg;
3318 CERROR("%s: async CQ event type %d\n",
3319 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3323 kiblnd_scheduler(void *arg)
3325 long id = (long)arg;
3326 struct kib_sched_info *sched;
3329 unsigned long flags;
3335 cfs_block_allsigs();
3337 init_waitqueue_entry_current(&wait);
3339 sched = kiblnd_data.kib_scheds[KIB_THREAD_CPT(id)];
3341 rc = cfs_cpt_bind(lnet_cpt_table(), sched->ibs_cpt);
3343 CWARN("Failed to bind on CPT %d, please verify whether "
3344 "all CPUs are healthy and reload modules if necessary, "
3345 "otherwise your system might under risk of low "
3346 "performance\n", sched->ibs_cpt);
3349 spin_lock_irqsave(&sched->ibs_lock, flags);
3351 while (!kiblnd_data.kib_shutdown) {
3352 if (busy_loops++ >= IBLND_RESCHED) {
3353 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3358 spin_lock_irqsave(&sched->ibs_lock, flags);
3363 if (!cfs_list_empty(&sched->ibs_conns)) {
3364 conn = cfs_list_entry(sched->ibs_conns.next,
3365 kib_conn_t, ibc_sched_list);
3366 /* take over kib_sched_conns' ref on conn... */
3367 LASSERT(conn->ibc_scheduled);
3368 cfs_list_del(&conn->ibc_sched_list);
3369 conn->ibc_ready = 0;
3371 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3373 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3375 rc = ib_req_notify_cq(conn->ibc_cq,
3378 CWARN("%s: ib_req_notify_cq failed: %d, "
3379 "closing connection\n",
3380 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
3381 kiblnd_close_conn(conn, -EIO);
3382 kiblnd_conn_decref(conn);
3383 spin_lock_irqsave(&sched->ibs_lock,
3388 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3392 CWARN("%s: ib_poll_cq failed: %d, "
3393 "closing connection\n",
3394 libcfs_nid2str(conn->ibc_peer->ibp_nid),
3396 kiblnd_close_conn(conn, -EIO);
3397 kiblnd_conn_decref(conn);
3398 spin_lock_irqsave(&sched->ibs_lock, flags);
3402 spin_lock_irqsave(&sched->ibs_lock, flags);
3404 if (rc != 0 || conn->ibc_ready) {
3405 /* There may be another completion waiting; get
3406 * another scheduler to check while I handle
3408 /* +1 ref for sched_conns */
3409 kiblnd_conn_addref(conn);
3410 cfs_list_add_tail(&conn->ibc_sched_list,
3412 if (waitqueue_active(&sched->ibs_waitq))
3413 wake_up(&sched->ibs_waitq);
3415 conn->ibc_scheduled = 0;
3419 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3420 kiblnd_complete(&wc);
3422 spin_lock_irqsave(&sched->ibs_lock, flags);
3425 kiblnd_conn_decref(conn); /* ...drop my ref from above */
3432 set_current_state(TASK_INTERRUPTIBLE);
3433 add_wait_queue_exclusive(&sched->ibs_waitq, &wait);
3434 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3436 waitq_wait(&wait, TASK_INTERRUPTIBLE);
3439 remove_wait_queue(&sched->ibs_waitq, &wait);
3440 set_current_state(TASK_RUNNING);
3441 spin_lock_irqsave(&sched->ibs_lock, flags);
3444 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3446 kiblnd_thread_fini();
3451 kiblnd_failover_thread(void *arg)
3453 rwlock_t *glock = &kiblnd_data.kib_global_lock;
3456 unsigned long flags;
3459 LASSERT (*kiblnd_tunables.kib_dev_failover != 0);
3461 cfs_block_allsigs ();
3463 init_waitqueue_entry_current(&wait);
3464 write_lock_irqsave(glock, flags);
3466 while (!kiblnd_data.kib_shutdown) {
3467 int do_failover = 0;
3470 cfs_list_for_each_entry(dev, &kiblnd_data.kib_failed_devs,
3472 if (cfs_time_before(cfs_time_current(),
3473 dev->ibd_next_failover))
3480 cfs_list_del_init(&dev->ibd_fail_list);
3481 dev->ibd_failover = 1;
3482 write_unlock_irqrestore(glock, flags);
3484 rc = kiblnd_dev_failover(dev);
3486 write_lock_irqsave(glock, flags);
3488 LASSERT (dev->ibd_failover);
3489 dev->ibd_failover = 0;
3490 if (rc >= 0) { /* Device is OK or failover succeed */
3491 dev->ibd_next_failover = cfs_time_shift(3);
3495 /* failed to failover, retry later */
3496 dev->ibd_next_failover =
3497 cfs_time_shift(min(dev->ibd_failed_failover, 10));
3498 if (kiblnd_dev_can_failover(dev)) {
3499 cfs_list_add_tail(&dev->ibd_fail_list,
3500 &kiblnd_data.kib_failed_devs);
3506 /* long sleep if no more pending failover */
3507 long_sleep = cfs_list_empty(&kiblnd_data.kib_failed_devs);
3509 set_current_state(TASK_INTERRUPTIBLE);
3510 add_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3511 write_unlock_irqrestore(glock, flags);
3513 rc = schedule_timeout(long_sleep ? cfs_time_seconds(10) :
3514 cfs_time_seconds(1));
3515 set_current_state(TASK_RUNNING);
3516 remove_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3517 write_lock_irqsave(glock, flags);
3519 if (!long_sleep || rc != 0)
3522 /* have a long sleep, routine check all active devices,
3523 * we need checking like this because if there is not active
3524 * connection on the dev and no SEND from local, we may listen
3525 * on wrong HCA for ever while there is a bonding failover */
3526 cfs_list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3527 if (kiblnd_dev_can_failover(dev)) {
3528 cfs_list_add_tail(&dev->ibd_fail_list,
3529 &kiblnd_data.kib_failed_devs);
3534 write_unlock_irqrestore(glock, flags);
3536 kiblnd_thread_fini();