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;
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, struct list_head *txlist, int status)
91 while (!list_empty(txlist)) {
92 tx = list_entry(txlist->next, kib_tx_t, tx_list);
94 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;
106 struct list_head *node;
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_nfrags == 0);
130 kiblnd_drop_rx(kib_rx_t *rx)
132 kib_conn_t *conn = rx->rx_conn;
133 struct kib_sched_info *sched = conn->ibc_sched;
136 spin_lock_irqsave(&sched->ibs_lock, flags);
137 LASSERT(conn->ibc_nrx > 0);
139 spin_unlock_irqrestore(&sched->ibs_lock, flags);
141 kiblnd_conn_decref(conn);
145 kiblnd_post_rx(kib_rx_t *rx, int credit)
147 kib_conn_t *conn = rx->rx_conn;
148 kib_net_t *net = conn->ibc_peer->ibp_ni->ni_data;
149 struct ib_recv_wr *bad_wrq = NULL;
153 LASSERT(net != NULL);
154 LASSERT(!in_interrupt());
155 LASSERT(credit == IBLND_POSTRX_NO_CREDIT ||
156 credit == IBLND_POSTRX_PEER_CREDIT ||
157 credit == IBLND_POSTRX_RSRVD_CREDIT);
159 mr = kiblnd_find_dma_mr(conn->ibc_hdev, rx->rx_msgaddr, IBLND_MSG_SIZE);
162 rx->rx_sge.lkey = mr->lkey;
163 rx->rx_sge.addr = rx->rx_msgaddr;
164 rx->rx_sge.length = IBLND_MSG_SIZE;
166 rx->rx_wrq.next = NULL;
167 rx->rx_wrq.sg_list = &rx->rx_sge;
168 rx->rx_wrq.num_sge = 1;
169 rx->rx_wrq.wr_id = kiblnd_ptr2wreqid(rx, IBLND_WID_RX);
171 LASSERT(conn->ibc_state >= IBLND_CONN_INIT);
172 LASSERT(rx->rx_nob >= 0); /* not posted */
174 if (conn->ibc_state > IBLND_CONN_ESTABLISHED) {
175 kiblnd_drop_rx(rx); /* No more posts for this rx */
179 rx->rx_nob = -1; /* flag posted */
181 rc = ib_post_recv(conn->ibc_cmid->qp, &rx->rx_wrq, &bad_wrq);
183 CERROR("Can't post rx for %s: %d, bad_wrq: %p\n",
184 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc, bad_wrq);
188 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) /* Initial post */
192 kiblnd_close_conn(conn, rc);
193 kiblnd_drop_rx(rx); /* No more posts for this rx */
197 if (credit == IBLND_POSTRX_NO_CREDIT)
200 spin_lock(&conn->ibc_lock);
201 if (credit == IBLND_POSTRX_PEER_CREDIT)
202 conn->ibc_outstanding_credits++;
204 conn->ibc_reserved_credits++;
205 spin_unlock(&conn->ibc_lock);
207 kiblnd_check_sends(conn);
212 kiblnd_find_waiting_tx_locked(kib_conn_t *conn, int txtype, __u64 cookie)
214 struct list_head *tmp;
216 list_for_each(tmp, &conn->ibc_active_txs) {
217 kib_tx_t *tx = list_entry(tmp, kib_tx_t, tx_list);
219 LASSERT(!tx->tx_queued);
220 LASSERT(tx->tx_sending != 0 || tx->tx_waiting);
222 if (tx->tx_cookie != cookie)
225 if (tx->tx_waiting &&
226 tx->tx_msg->ibm_type == txtype)
229 CWARN("Bad completion: %swaiting, type %x (wanted %x)\n",
230 tx->tx_waiting ? "" : "NOT ",
231 tx->tx_msg->ibm_type, txtype);
237 kiblnd_handle_completion(kib_conn_t *conn, int txtype, int status, __u64 cookie)
240 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
243 spin_lock(&conn->ibc_lock);
245 tx = kiblnd_find_waiting_tx_locked(conn, txtype, cookie);
247 spin_unlock(&conn->ibc_lock);
249 CWARN("Unmatched completion type %x cookie %#llx from %s\n",
250 txtype, cookie, libcfs_nid2str(conn->ibc_peer->ibp_nid));
251 kiblnd_close_conn(conn, -EPROTO);
255 if (tx->tx_status == 0) { /* success so far */
256 if (status < 0) { /* failed? */
257 tx->tx_status = status;
258 } else if (txtype == IBLND_MSG_GET_REQ) {
259 lnet_set_reply_msg_len(ni, tx->tx_lntmsg[1], status);
265 idle = !tx->tx_queued && (tx->tx_sending == 0);
267 list_del(&tx->tx_list);
269 spin_unlock(&conn->ibc_lock);
272 kiblnd_tx_done(ni, tx);
276 kiblnd_send_completion(kib_conn_t *conn, int type, int status, __u64 cookie)
278 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
279 kib_tx_t *tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
282 CERROR("Can't get tx for completion %x for %s\n",
283 type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
287 tx->tx_msg->ibm_u.completion.ibcm_status = status;
288 tx->tx_msg->ibm_u.completion.ibcm_cookie = cookie;
289 kiblnd_init_tx_msg(ni, tx, type, sizeof(kib_completion_msg_t));
291 kiblnd_queue_tx(tx, conn);
295 kiblnd_handle_rx(kib_rx_t *rx)
297 kib_msg_t *msg = rx->rx_msg;
298 kib_conn_t *conn = rx->rx_conn;
299 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
300 int credits = msg->ibm_credits;
306 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
308 CDEBUG(D_NET, "Received %x[%d] from %s\n",
309 msg->ibm_type, credits,
310 libcfs_nid2str(conn->ibc_peer->ibp_nid));
313 /* Have I received credits that will let me send? */
314 spin_lock(&conn->ibc_lock);
316 if (conn->ibc_credits + credits >
317 IBLND_MSG_QUEUE_SIZE(conn->ibc_version)) {
318 rc2 = conn->ibc_credits;
319 spin_unlock(&conn->ibc_lock);
321 CERROR("Bad credits from %s: %d + %d > %d\n",
322 libcfs_nid2str(conn->ibc_peer->ibp_nid),
324 IBLND_MSG_QUEUE_SIZE(conn->ibc_version));
326 kiblnd_close_conn(conn, -EPROTO);
327 kiblnd_post_rx(rx, IBLND_POSTRX_NO_CREDIT);
331 conn->ibc_credits += credits;
333 /* This ensures the credit taken by NOOP can be returned */
334 if (msg->ibm_type == IBLND_MSG_NOOP &&
335 !IBLND_OOB_CAPABLE(conn->ibc_version)) /* v1 only */
336 conn->ibc_outstanding_credits++;
338 spin_unlock(&conn->ibc_lock);
339 kiblnd_check_sends(conn);
342 switch (msg->ibm_type) {
344 CERROR("Bad IBLND message type %x from %s\n",
345 msg->ibm_type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
346 post_credit = IBLND_POSTRX_NO_CREDIT;
351 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
352 post_credit = IBLND_POSTRX_NO_CREDIT;
356 if (credits != 0) /* credit already posted */
357 post_credit = IBLND_POSTRX_NO_CREDIT;
358 else /* a keepalive NOOP */
359 post_credit = IBLND_POSTRX_PEER_CREDIT;
362 case IBLND_MSG_IMMEDIATE:
363 post_credit = IBLND_POSTRX_DONT_POST;
364 rc = lnet_parse(ni, &msg->ibm_u.immediate.ibim_hdr,
365 msg->ibm_srcnid, rx, 0);
366 if (rc < 0) /* repost on error */
367 post_credit = IBLND_POSTRX_PEER_CREDIT;
370 case IBLND_MSG_PUT_REQ:
371 post_credit = IBLND_POSTRX_DONT_POST;
372 rc = lnet_parse(ni, &msg->ibm_u.putreq.ibprm_hdr,
373 msg->ibm_srcnid, rx, 1);
374 if (rc < 0) /* repost on error */
375 post_credit = IBLND_POSTRX_PEER_CREDIT;
378 case IBLND_MSG_PUT_NAK:
379 CWARN("PUT_NACK from %s\n",
380 libcfs_nid2str(conn->ibc_peer->ibp_nid));
381 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
382 kiblnd_handle_completion(conn, IBLND_MSG_PUT_REQ,
383 msg->ibm_u.completion.ibcm_status,
384 msg->ibm_u.completion.ibcm_cookie);
387 case IBLND_MSG_PUT_ACK:
388 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
390 spin_lock(&conn->ibc_lock);
391 tx = kiblnd_find_waiting_tx_locked(conn, IBLND_MSG_PUT_REQ,
392 msg->ibm_u.putack.ibpam_src_cookie);
394 list_del(&tx->tx_list);
395 spin_unlock(&conn->ibc_lock);
398 CERROR("Unmatched PUT_ACK from %s\n",
399 libcfs_nid2str(conn->ibc_peer->ibp_nid));
404 LASSERT(tx->tx_waiting);
405 /* CAVEAT EMPTOR: I could be racing with tx_complete, but...
406 * (a) I can overwrite tx_msg since my peer has received it!
407 * (b) tx_waiting set tells tx_complete() it's not done. */
409 tx->tx_nwrq = 0; /* overwrite PUT_REQ */
411 rc2 = kiblnd_init_rdma(conn, tx, IBLND_MSG_PUT_DONE,
412 kiblnd_rd_size(&msg->ibm_u.putack.ibpam_rd),
413 &msg->ibm_u.putack.ibpam_rd,
414 msg->ibm_u.putack.ibpam_dst_cookie);
416 CERROR("Can't setup rdma for PUT to %s: %d\n",
417 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc2);
419 spin_lock(&conn->ibc_lock);
420 tx->tx_waiting = 0; /* clear waiting and queue atomically */
421 kiblnd_queue_tx_locked(tx, conn);
422 spin_unlock(&conn->ibc_lock);
425 case IBLND_MSG_PUT_DONE:
426 post_credit = IBLND_POSTRX_PEER_CREDIT;
427 kiblnd_handle_completion(conn, IBLND_MSG_PUT_ACK,
428 msg->ibm_u.completion.ibcm_status,
429 msg->ibm_u.completion.ibcm_cookie);
432 case IBLND_MSG_GET_REQ:
433 post_credit = IBLND_POSTRX_DONT_POST;
434 rc = lnet_parse(ni, &msg->ibm_u.get.ibgm_hdr,
435 msg->ibm_srcnid, rx, 1);
436 if (rc < 0) /* repost on error */
437 post_credit = IBLND_POSTRX_PEER_CREDIT;
440 case IBLND_MSG_GET_DONE:
441 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
442 kiblnd_handle_completion(conn, IBLND_MSG_GET_REQ,
443 msg->ibm_u.completion.ibcm_status,
444 msg->ibm_u.completion.ibcm_cookie);
448 if (rc < 0) /* protocol error */
449 kiblnd_close_conn(conn, rc);
451 if (post_credit != IBLND_POSTRX_DONT_POST)
452 kiblnd_post_rx(rx, post_credit);
456 kiblnd_rx_complete(kib_rx_t *rx, int status, int nob)
458 kib_msg_t *msg = rx->rx_msg;
459 kib_conn_t *conn = rx->rx_conn;
460 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
461 kib_net_t *net = ni->ni_data;
465 LASSERT(net != NULL);
466 LASSERT(rx->rx_nob < 0); /* was posted */
467 rx->rx_nob = 0; /* isn't now */
469 if (conn->ibc_state > IBLND_CONN_ESTABLISHED)
472 if (status != IB_WC_SUCCESS) {
473 CNETERR("Rx from %s failed: %d\n",
474 libcfs_nid2str(conn->ibc_peer->ibp_nid), status);
481 rc = kiblnd_unpack_msg(msg, rx->rx_nob);
483 CERROR("Error %d unpacking rx from %s\n",
484 rc, libcfs_nid2str(conn->ibc_peer->ibp_nid));
488 if (msg->ibm_srcnid != conn->ibc_peer->ibp_nid ||
489 msg->ibm_dstnid != ni->ni_nid ||
490 msg->ibm_srcstamp != conn->ibc_incarnation ||
491 msg->ibm_dststamp != net->ibn_incarnation) {
492 CERROR("Stale rx from %s\n",
493 libcfs_nid2str(conn->ibc_peer->ibp_nid));
498 /* set time last known alive */
499 kiblnd_peer_alive(conn->ibc_peer);
501 /* racing with connection establishment/teardown! */
503 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
504 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
507 write_lock_irqsave(g_lock, flags);
508 /* must check holding global lock to eliminate race */
509 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
510 list_add_tail(&rx->rx_list, &conn->ibc_early_rxs);
511 write_unlock_irqrestore(g_lock, flags);
514 write_unlock_irqrestore(g_lock, flags);
516 kiblnd_handle_rx(rx);
520 CDEBUG(D_NET, "rx %p conn %p\n", rx, conn);
521 kiblnd_close_conn(conn, err);
523 kiblnd_drop_rx(rx); /* Don't re-post rx. */
527 kiblnd_kvaddr_to_page(unsigned long vaddr)
531 if (is_vmalloc_addr((void *)vaddr)) {
532 page = vmalloc_to_page((void *)vaddr);
533 LASSERT(page != NULL);
536 #ifdef CONFIG_HIGHMEM
537 if (vaddr >= PKMAP_BASE &&
538 vaddr < (PKMAP_BASE + LAST_PKMAP * PAGE_SIZE)) {
539 /* No highmem pages only used for bulk (kiov) I/O */
540 CERROR("find page for address in highmem\n");
544 page = virt_to_page(vaddr);
545 LASSERT(page != NULL);
550 kiblnd_fmr_map_tx(kib_net_t *net, kib_tx_t *tx, kib_rdma_desc_t *rd, int nob)
553 __u64 *pages = tx->tx_pages;
554 kib_fmr_poolset_t *fps;
561 LASSERT(tx->tx_pool != NULL);
562 LASSERT(tx->tx_pool->tpo_pool.po_owner != NULL);
564 hdev = tx->tx_pool->tpo_hdev;
566 for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
567 for (size = 0; size < rd->rd_frags[i].rf_nob;
568 size += hdev->ibh_page_size) {
569 pages[npages++] = (rd->rd_frags[i].rf_addr &
570 hdev->ibh_page_mask) + size;
574 cpt = tx->tx_pool->tpo_pool.po_owner->ps_cpt;
576 fps = net->ibn_fmr_ps[cpt];
577 rc = kiblnd_fmr_pool_map(fps, pages, npages, 0, &tx->fmr);
579 CERROR("Can't map %d pages: %d\n", npages, rc);
583 /* If rd is not tx_rd, it's going to get sent to a peer, who will need
585 rd->rd_key = (rd != tx->tx_rd) ? tx->fmr.fmr_pfmr->fmr->rkey :
586 tx->fmr.fmr_pfmr->fmr->lkey;
587 rd->rd_frags[0].rf_addr &= ~hdev->ibh_page_mask;
588 rd->rd_frags[0].rf_nob = nob;
595 kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx)
597 kib_net_t *net = ni->ni_data;
599 LASSERT(net != NULL);
601 if (net->ibn_fmr_ps && tx->fmr.fmr_pfmr) {
602 kiblnd_fmr_pool_unmap(&tx->fmr, tx->tx_status);
603 tx->fmr.fmr_pfmr = NULL;
606 if (tx->tx_nfrags != 0) {
607 kiblnd_dma_unmap_sg(tx->tx_pool->tpo_hdev->ibh_ibdev,
608 tx->tx_frags, tx->tx_nfrags, tx->tx_dmadir);
614 kiblnd_map_tx(lnet_ni_t *ni, kib_tx_t *tx,
615 kib_rdma_desc_t *rd, int nfrags)
617 kib_hca_dev_t *hdev = tx->tx_pool->tpo_hdev;
618 kib_net_t *net = ni->ni_data;
619 struct ib_mr *mr = NULL;
623 /* If rd is not tx_rd, it's going to get sent to a peer and I'm the
625 tx->tx_dmadir = (rd != tx->tx_rd) ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
626 tx->tx_nfrags = nfrags;
628 rd->rd_nfrags = kiblnd_dma_map_sg(hdev->ibh_ibdev, tx->tx_frags,
629 tx->tx_nfrags, tx->tx_dmadir);
631 for (i = 0, nob = 0; i < rd->rd_nfrags; i++) {
632 rd->rd_frags[i].rf_nob = kiblnd_sg_dma_len(
633 hdev->ibh_ibdev, &tx->tx_frags[i]);
634 rd->rd_frags[i].rf_addr = kiblnd_sg_dma_address(
635 hdev->ibh_ibdev, &tx->tx_frags[i]);
636 nob += rd->rd_frags[i].rf_nob;
639 /* looking for pre-mapping MR */
640 mr = kiblnd_find_rd_dma_mr(hdev, rd);
642 /* found pre-mapping MR */
643 rd->rd_key = (rd != tx->tx_rd) ? mr->rkey : mr->lkey;
647 if (net->ibn_fmr_ps != NULL)
648 return kiblnd_fmr_map_tx(net, tx, rd, nob);
655 kiblnd_setup_rd_iov(lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd,
656 unsigned int niov, struct kvec *iov, int offset, int nob)
658 kib_net_t *net = ni->ni_data;
660 struct scatterlist *sg;
667 LASSERT(net != NULL);
669 while (offset >= iov->iov_len) {
670 offset -= iov->iov_len;
680 vaddr = ((unsigned long)iov->iov_base) + offset;
681 page_offset = vaddr & (PAGE_SIZE - 1);
682 page = kiblnd_kvaddr_to_page(vaddr);
684 CERROR("Can't find page\n");
688 fragnob = min((int)(iov->iov_len - offset), nob);
689 fragnob = min(fragnob, (int)PAGE_SIZE - page_offset);
691 sg_set_page(sg, page, fragnob, page_offset);
694 if (offset + fragnob < iov->iov_len) {
704 return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags);
708 kiblnd_setup_rd_kiov(lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd,
709 int nkiov, lnet_kiov_t *kiov, int offset, int nob)
711 kib_net_t *net = ni->ni_data;
712 struct scatterlist *sg;
715 CDEBUG(D_NET, "niov %d offset %d nob %d\n", nkiov, offset, nob);
719 LASSERT(net != NULL);
721 while (offset >= kiov->kiov_len) {
722 offset -= kiov->kiov_len;
732 fragnob = min((int)(kiov->kiov_len - offset), nob);
734 sg_set_page(sg, kiov->kiov_page, fragnob,
735 kiov->kiov_offset + offset);
744 return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags);
748 kiblnd_post_tx_locked(kib_conn_t *conn, kib_tx_t *tx, int credit)
749 __releases(conn->ibc_lock)
750 __acquires(conn->ibc_lock)
752 kib_msg_t *msg = tx->tx_msg;
753 kib_peer_t *peer = conn->ibc_peer;
754 int ver = conn->ibc_version;
757 struct ib_send_wr *bad_wrq;
759 LASSERT(tx->tx_queued);
760 /* We rely on this for QP sizing */
761 LASSERT(tx->tx_nwrq > 0);
762 LASSERT(tx->tx_nwrq <= 1 + IBLND_RDMA_FRAGS(ver));
764 LASSERT(credit == 0 || credit == 1);
765 LASSERT(conn->ibc_outstanding_credits >= 0);
766 LASSERT(conn->ibc_outstanding_credits <= IBLND_MSG_QUEUE_SIZE(ver));
767 LASSERT(conn->ibc_credits >= 0);
768 LASSERT(conn->ibc_credits <= IBLND_MSG_QUEUE_SIZE(ver));
770 if (conn->ibc_nsends_posted == IBLND_CONCURRENT_SENDS(ver)) {
771 /* tx completions outstanding... */
772 CDEBUG(D_NET, "%s: posted enough\n",
773 libcfs_nid2str(peer->ibp_nid));
777 if (credit != 0 && conn->ibc_credits == 0) { /* no credits */
778 CDEBUG(D_NET, "%s: no credits\n",
779 libcfs_nid2str(peer->ibp_nid));
783 if (credit != 0 && !IBLND_OOB_CAPABLE(ver) &&
784 conn->ibc_credits == 1 && /* last credit reserved */
785 msg->ibm_type != IBLND_MSG_NOOP) { /* for NOOP */
786 CDEBUG(D_NET, "%s: not using last credit\n",
787 libcfs_nid2str(peer->ibp_nid));
791 /* NB don't drop ibc_lock before bumping tx_sending */
792 list_del(&tx->tx_list);
795 if (msg->ibm_type == IBLND_MSG_NOOP &&
796 (!kiblnd_need_noop(conn) || /* redundant NOOP */
797 (IBLND_OOB_CAPABLE(ver) && /* posted enough NOOP */
798 conn->ibc_noops_posted == IBLND_OOB_MSGS(ver)))) {
799 /* OK to drop when posted enough NOOPs, since
800 * kiblnd_check_sends will queue NOOP again when
801 * posted NOOPs complete */
802 spin_unlock(&conn->ibc_lock);
803 kiblnd_tx_done(peer->ibp_ni, tx);
804 spin_lock(&conn->ibc_lock);
805 CDEBUG(D_NET, "%s(%d): redundant or enough NOOP\n",
806 libcfs_nid2str(peer->ibp_nid),
807 conn->ibc_noops_posted);
811 kiblnd_pack_msg(peer->ibp_ni, msg, ver, conn->ibc_outstanding_credits,
812 peer->ibp_nid, conn->ibc_incarnation);
814 conn->ibc_credits -= credit;
815 conn->ibc_outstanding_credits = 0;
816 conn->ibc_nsends_posted++;
817 if (msg->ibm_type == IBLND_MSG_NOOP)
818 conn->ibc_noops_posted++;
820 /* CAVEAT EMPTOR! This tx could be the PUT_DONE of an RDMA
821 * PUT. If so, it was first queued here as a PUT_REQ, sent and
822 * stashed on ibc_active_txs, matched by an incoming PUT_ACK,
823 * and then re-queued here. It's (just) possible that
824 * tx_sending is non-zero if we've not done the tx_complete()
825 * from the first send; hence the ++ rather than = below. */
827 list_add(&tx->tx_list, &conn->ibc_active_txs);
829 /* I'm still holding ibc_lock! */
830 if (conn->ibc_state != IBLND_CONN_ESTABLISHED) {
832 } else if (tx->tx_pool->tpo_pool.po_failed ||
833 conn->ibc_hdev != tx->tx_pool->tpo_hdev) {
834 /* close_conn will launch failover */
837 rc = ib_post_send(conn->ibc_cmid->qp, tx->tx_wrq, &bad_wrq);
840 conn->ibc_last_send = jiffies;
845 /* NB credits are transferred in the actual
846 * message, which can only be the last work item */
847 conn->ibc_credits += credit;
848 conn->ibc_outstanding_credits += msg->ibm_credits;
849 conn->ibc_nsends_posted--;
850 if (msg->ibm_type == IBLND_MSG_NOOP)
851 conn->ibc_noops_posted--;
857 done = (tx->tx_sending == 0);
859 list_del(&tx->tx_list);
861 spin_unlock(&conn->ibc_lock);
863 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
864 CERROR("Error %d posting transmit to %s\n",
865 rc, libcfs_nid2str(peer->ibp_nid));
867 CDEBUG(D_NET, "Error %d posting transmit to %s\n",
868 rc, libcfs_nid2str(peer->ibp_nid));
870 kiblnd_close_conn(conn, rc);
873 kiblnd_tx_done(peer->ibp_ni, tx);
875 spin_lock(&conn->ibc_lock);
881 kiblnd_check_sends(kib_conn_t *conn)
883 int ver = conn->ibc_version;
884 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
887 /* Don't send anything until after the connection is established */
888 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
889 CDEBUG(D_NET, "%s too soon\n",
890 libcfs_nid2str(conn->ibc_peer->ibp_nid));
894 spin_lock(&conn->ibc_lock);
896 LASSERT(conn->ibc_nsends_posted <= IBLND_CONCURRENT_SENDS(ver));
897 LASSERT(!IBLND_OOB_CAPABLE(ver) ||
898 conn->ibc_noops_posted <= IBLND_OOB_MSGS(ver));
899 LASSERT(conn->ibc_reserved_credits >= 0);
901 while (conn->ibc_reserved_credits > 0 &&
902 !list_empty(&conn->ibc_tx_queue_rsrvd)) {
903 tx = list_entry(conn->ibc_tx_queue_rsrvd.next,
905 list_del(&tx->tx_list);
906 list_add_tail(&tx->tx_list, &conn->ibc_tx_queue);
907 conn->ibc_reserved_credits--;
910 if (kiblnd_need_noop(conn)) {
911 spin_unlock(&conn->ibc_lock);
913 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
915 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_NOOP, 0);
917 spin_lock(&conn->ibc_lock);
919 kiblnd_queue_tx_locked(tx, conn);
922 kiblnd_conn_addref(conn); /* 1 ref for me.... (see b21911) */
927 if (!list_empty(&conn->ibc_tx_queue_nocred)) {
929 tx = list_entry(conn->ibc_tx_queue_nocred.next,
931 } else if (!list_empty(&conn->ibc_tx_noops)) {
932 LASSERT(!IBLND_OOB_CAPABLE(ver));
934 tx = list_entry(conn->ibc_tx_noops.next,
936 } else if (!list_empty(&conn->ibc_tx_queue)) {
938 tx = list_entry(conn->ibc_tx_queue.next,
943 if (kiblnd_post_tx_locked(conn, tx, credit) != 0)
947 spin_unlock(&conn->ibc_lock);
949 kiblnd_conn_decref(conn); /* ...until here */
953 kiblnd_tx_complete(kib_tx_t *tx, int status)
955 int failed = (status != IB_WC_SUCCESS);
956 kib_conn_t *conn = tx->tx_conn;
959 LASSERT(tx->tx_sending > 0);
962 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
963 CNETERR("Tx -> %s cookie %#llx sending %d waiting %d: failed %d\n",
964 libcfs_nid2str(conn->ibc_peer->ibp_nid),
965 tx->tx_cookie, tx->tx_sending, tx->tx_waiting,
968 kiblnd_close_conn(conn, -EIO);
970 kiblnd_peer_alive(conn->ibc_peer);
973 spin_lock(&conn->ibc_lock);
975 /* I could be racing with rdma completion. Whoever makes 'tx' idle
976 * gets to free it, which also drops its ref on 'conn'. */
979 conn->ibc_nsends_posted--;
980 if (tx->tx_msg->ibm_type == IBLND_MSG_NOOP)
981 conn->ibc_noops_posted--;
984 tx->tx_waiting = 0; /* don't wait for peer */
985 tx->tx_status = -EIO;
988 idle = (tx->tx_sending == 0) && /* This is the final callback */
989 !tx->tx_waiting && /* Not waiting for peer */
990 !tx->tx_queued; /* Not re-queued (PUT_DONE) */
992 list_del(&tx->tx_list);
994 kiblnd_conn_addref(conn); /* 1 ref for me.... */
996 spin_unlock(&conn->ibc_lock);
999 kiblnd_tx_done(conn->ibc_peer->ibp_ni, tx);
1001 kiblnd_check_sends(conn);
1003 kiblnd_conn_decref(conn); /* ...until here */
1007 kiblnd_init_tx_msg(lnet_ni_t *ni, kib_tx_t *tx, int type, int body_nob)
1009 kib_hca_dev_t *hdev = tx->tx_pool->tpo_hdev;
1010 struct ib_sge *sge = &tx->tx_sge[tx->tx_nwrq];
1011 struct ib_send_wr *wrq = &tx->tx_wrq[tx->tx_nwrq];
1012 int nob = offsetof(kib_msg_t, ibm_u) + body_nob;
1015 LASSERT(tx->tx_nwrq >= 0);
1016 LASSERT(tx->tx_nwrq < IBLND_MAX_RDMA_FRAGS + 1);
1017 LASSERT(nob <= IBLND_MSG_SIZE);
1019 kiblnd_init_msg(tx->tx_msg, type, body_nob);
1021 mr = kiblnd_find_dma_mr(hdev, tx->tx_msgaddr, nob);
1022 LASSERT(mr != NULL);
1024 sge->lkey = mr->lkey;
1025 sge->addr = tx->tx_msgaddr;
1028 memset(wrq, 0, sizeof(*wrq));
1031 wrq->wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_TX);
1034 wrq->opcode = IB_WR_SEND;
1035 wrq->send_flags = IB_SEND_SIGNALED;
1041 kiblnd_init_rdma(kib_conn_t *conn, kib_tx_t *tx, int type,
1042 int resid, kib_rdma_desc_t *dstrd, __u64 dstcookie)
1044 kib_msg_t *ibmsg = tx->tx_msg;
1045 kib_rdma_desc_t *srcrd = tx->tx_rd;
1046 struct ib_sge *sge = &tx->tx_sge[0];
1047 struct ib_send_wr *wrq = &tx->tx_wrq[0];
1053 LASSERT(!in_interrupt());
1054 LASSERT(tx->tx_nwrq == 0);
1055 LASSERT(type == IBLND_MSG_GET_DONE ||
1056 type == IBLND_MSG_PUT_DONE);
1058 srcidx = dstidx = 0;
1061 if (srcidx >= srcrd->rd_nfrags) {
1062 CERROR("Src buffer exhausted: %d frags\n", srcidx);
1067 if (dstidx == dstrd->rd_nfrags) {
1068 CERROR("Dst buffer exhausted: %d frags\n", dstidx);
1073 if (tx->tx_nwrq == IBLND_RDMA_FRAGS(conn->ibc_version)) {
1074 CERROR("RDMA too fragmented for %s (%d): %d/%d src %d/%d dst frags\n",
1075 libcfs_nid2str(conn->ibc_peer->ibp_nid),
1076 IBLND_RDMA_FRAGS(conn->ibc_version),
1077 srcidx, srcrd->rd_nfrags,
1078 dstidx, dstrd->rd_nfrags);
1083 wrknob = min(min(kiblnd_rd_frag_size(srcrd, srcidx),
1084 kiblnd_rd_frag_size(dstrd, dstidx)),
1087 sge = &tx->tx_sge[tx->tx_nwrq];
1088 sge->addr = kiblnd_rd_frag_addr(srcrd, srcidx);
1089 sge->lkey = kiblnd_rd_frag_key(srcrd, srcidx);
1090 sge->length = wrknob;
1092 wrq = &tx->tx_wrq[tx->tx_nwrq];
1094 wrq->next = wrq + 1;
1095 wrq->wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_RDMA);
1098 wrq->opcode = IB_WR_RDMA_WRITE;
1099 wrq->send_flags = 0;
1101 wrq->wr.rdma.remote_addr = kiblnd_rd_frag_addr(dstrd, dstidx);
1102 wrq->wr.rdma.rkey = kiblnd_rd_frag_key(dstrd, dstidx);
1104 srcidx = kiblnd_rd_consume_frag(srcrd, srcidx, wrknob);
1105 dstidx = kiblnd_rd_consume_frag(dstrd, dstidx, wrknob);
1114 if (rc < 0) /* no RDMA if completing with failure */
1117 ibmsg->ibm_u.completion.ibcm_status = rc;
1118 ibmsg->ibm_u.completion.ibcm_cookie = dstcookie;
1119 kiblnd_init_tx_msg(conn->ibc_peer->ibp_ni, tx,
1120 type, sizeof(kib_completion_msg_t));
1126 kiblnd_queue_tx_locked(kib_tx_t *tx, kib_conn_t *conn)
1128 struct list_head *q;
1130 LASSERT(tx->tx_nwrq > 0); /* work items set up */
1131 LASSERT(!tx->tx_queued); /* not queued for sending already */
1132 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1135 tx->tx_deadline = jiffies + (*kiblnd_tunables.kib_timeout * HZ);
1137 if (tx->tx_conn == NULL) {
1138 kiblnd_conn_addref(conn);
1140 LASSERT(tx->tx_msg->ibm_type != IBLND_MSG_PUT_DONE);
1142 /* PUT_DONE first attached to conn as a PUT_REQ */
1143 LASSERT(tx->tx_conn == conn);
1144 LASSERT(tx->tx_msg->ibm_type == IBLND_MSG_PUT_DONE);
1147 switch (tx->tx_msg->ibm_type) {
1151 case IBLND_MSG_PUT_REQ:
1152 case IBLND_MSG_GET_REQ:
1153 q = &conn->ibc_tx_queue_rsrvd;
1156 case IBLND_MSG_PUT_NAK:
1157 case IBLND_MSG_PUT_ACK:
1158 case IBLND_MSG_PUT_DONE:
1159 case IBLND_MSG_GET_DONE:
1160 q = &conn->ibc_tx_queue_nocred;
1163 case IBLND_MSG_NOOP:
1164 if (IBLND_OOB_CAPABLE(conn->ibc_version))
1165 q = &conn->ibc_tx_queue_nocred;
1167 q = &conn->ibc_tx_noops;
1170 case IBLND_MSG_IMMEDIATE:
1171 q = &conn->ibc_tx_queue;
1175 list_add_tail(&tx->tx_list, q);
1179 kiblnd_queue_tx(kib_tx_t *tx, kib_conn_t *conn)
1181 spin_lock(&conn->ibc_lock);
1182 kiblnd_queue_tx_locked(tx, conn);
1183 spin_unlock(&conn->ibc_lock);
1185 kiblnd_check_sends(conn);
1188 static int kiblnd_resolve_addr(struct rdma_cm_id *cmid,
1189 struct sockaddr_in *srcaddr,
1190 struct sockaddr_in *dstaddr,
1193 unsigned short port;
1196 /* allow the port to be reused */
1197 rc = rdma_set_reuseaddr(cmid, 1);
1199 CERROR("Unable to set reuse on cmid: %d\n", rc);
1203 /* look for a free privileged port */
1204 for (port = PROT_SOCK-1; port > 0; port--) {
1205 srcaddr->sin_port = htons(port);
1206 rc = rdma_resolve_addr(cmid,
1207 (struct sockaddr *)srcaddr,
1208 (struct sockaddr *)dstaddr,
1211 CDEBUG(D_NET, "bound to port %hu\n", port);
1213 } else if (rc == -EADDRINUSE || rc == -EADDRNOTAVAIL) {
1214 CDEBUG(D_NET, "bind to port %hu failed: %d\n",
1221 CERROR("Failed to bind to a free privileged port\n");
1226 kiblnd_connect_peer(kib_peer_t *peer)
1228 struct rdma_cm_id *cmid;
1230 kib_net_t *net = peer->ibp_ni->ni_data;
1231 struct sockaddr_in srcaddr;
1232 struct sockaddr_in dstaddr;
1235 LASSERT(net != NULL);
1236 LASSERT(peer->ibp_connecting > 0);
1238 cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, peer, RDMA_PS_TCP,
1242 CERROR("Can't create CMID for %s: %ld\n",
1243 libcfs_nid2str(peer->ibp_nid), PTR_ERR(cmid));
1249 memset(&srcaddr, 0, sizeof(srcaddr));
1250 srcaddr.sin_family = AF_INET;
1251 srcaddr.sin_addr.s_addr = htonl(dev->ibd_ifip);
1253 memset(&dstaddr, 0, sizeof(dstaddr));
1254 dstaddr.sin_family = AF_INET;
1255 dstaddr.sin_port = htons(*kiblnd_tunables.kib_service);
1256 dstaddr.sin_addr.s_addr = htonl(LNET_NIDADDR(peer->ibp_nid));
1258 kiblnd_peer_addref(peer); /* cmid's ref */
1260 if (*kiblnd_tunables.kib_use_priv_port) {
1261 rc = kiblnd_resolve_addr(cmid, &srcaddr, &dstaddr,
1262 *kiblnd_tunables.kib_timeout * 1000);
1264 rc = rdma_resolve_addr(cmid,
1265 (struct sockaddr *)&srcaddr,
1266 (struct sockaddr *)&dstaddr,
1267 *kiblnd_tunables.kib_timeout * 1000);
1270 /* Can't initiate address resolution: */
1271 CERROR("Can't resolve addr for %s: %d\n",
1272 libcfs_nid2str(peer->ibp_nid), rc);
1276 LASSERT(cmid->device != NULL);
1277 CDEBUG(D_NET, "%s: connection bound to %s:%pI4h:%s\n",
1278 libcfs_nid2str(peer->ibp_nid), dev->ibd_ifname,
1279 &dev->ibd_ifip, cmid->device->name);
1284 kiblnd_peer_decref(peer); /* cmid's ref */
1285 rdma_destroy_id(cmid);
1287 kiblnd_peer_connect_failed(peer, 1, rc);
1291 kiblnd_launch_tx(lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid)
1296 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
1297 unsigned long flags;
1300 /* If I get here, I've committed to send, so I complete the tx with
1301 * failure on any problems */
1303 LASSERT(tx == NULL || tx->tx_conn == NULL); /* only set when assigned a conn */
1304 LASSERT(tx == NULL || tx->tx_nwrq > 0); /* work items have been set up */
1306 /* First time, just use a read lock since I expect to find my peer
1308 read_lock_irqsave(g_lock, flags);
1310 peer = kiblnd_find_peer_locked(nid);
1311 if (peer != NULL && !list_empty(&peer->ibp_conns)) {
1312 /* Found a peer with an established connection */
1313 conn = kiblnd_get_conn_locked(peer);
1314 kiblnd_conn_addref(conn); /* 1 ref for me... */
1316 read_unlock_irqrestore(g_lock, flags);
1319 kiblnd_queue_tx(tx, conn);
1320 kiblnd_conn_decref(conn); /* ...to here */
1324 read_unlock(g_lock);
1325 /* Re-try with a write lock */
1328 peer = kiblnd_find_peer_locked(nid);
1330 if (list_empty(&peer->ibp_conns)) {
1331 /* found a peer, but it's still connecting... */
1332 LASSERT(peer->ibp_connecting != 0 ||
1333 peer->ibp_accepting != 0);
1335 list_add_tail(&tx->tx_list,
1336 &peer->ibp_tx_queue);
1337 write_unlock_irqrestore(g_lock, flags);
1339 conn = kiblnd_get_conn_locked(peer);
1340 kiblnd_conn_addref(conn); /* 1 ref for me... */
1342 write_unlock_irqrestore(g_lock, flags);
1345 kiblnd_queue_tx(tx, conn);
1346 kiblnd_conn_decref(conn); /* ...to here */
1351 write_unlock_irqrestore(g_lock, flags);
1353 /* Allocate a peer ready to add to the peer table and retry */
1354 rc = kiblnd_create_peer(ni, &peer, nid);
1356 CERROR("Can't create peer %s\n", libcfs_nid2str(nid));
1358 tx->tx_status = -EHOSTUNREACH;
1360 kiblnd_tx_done(ni, tx);
1365 write_lock_irqsave(g_lock, flags);
1367 peer2 = kiblnd_find_peer_locked(nid);
1368 if (peer2 != NULL) {
1369 if (list_empty(&peer2->ibp_conns)) {
1370 /* found a peer, but it's still connecting... */
1371 LASSERT(peer2->ibp_connecting != 0 ||
1372 peer2->ibp_accepting != 0);
1374 list_add_tail(&tx->tx_list,
1375 &peer2->ibp_tx_queue);
1376 write_unlock_irqrestore(g_lock, flags);
1378 conn = kiblnd_get_conn_locked(peer2);
1379 kiblnd_conn_addref(conn); /* 1 ref for me... */
1381 write_unlock_irqrestore(g_lock, flags);
1384 kiblnd_queue_tx(tx, conn);
1385 kiblnd_conn_decref(conn); /* ...to here */
1388 kiblnd_peer_decref(peer);
1392 /* Brand new peer */
1393 LASSERT(peer->ibp_connecting == 0);
1394 peer->ibp_connecting = 1;
1396 /* always called with a ref on ni, which prevents ni being shutdown */
1397 LASSERT(((kib_net_t *)ni->ni_data)->ibn_shutdown == 0);
1400 list_add_tail(&tx->tx_list, &peer->ibp_tx_queue);
1402 kiblnd_peer_addref(peer);
1403 list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid));
1405 write_unlock_irqrestore(g_lock, flags);
1407 kiblnd_connect_peer(peer);
1408 kiblnd_peer_decref(peer);
1412 kiblnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg)
1414 lnet_hdr_t *hdr = &lntmsg->msg_hdr;
1415 int type = lntmsg->msg_type;
1416 lnet_process_id_t target = lntmsg->msg_target;
1417 int target_is_router = lntmsg->msg_target_is_router;
1418 int routing = lntmsg->msg_routing;
1419 unsigned int payload_niov = lntmsg->msg_niov;
1420 struct kvec *payload_iov = lntmsg->msg_iov;
1421 lnet_kiov_t *payload_kiov = lntmsg->msg_kiov;
1422 unsigned int payload_offset = lntmsg->msg_offset;
1423 unsigned int payload_nob = lntmsg->msg_len;
1429 /* NB 'private' is different depending on what we're sending.... */
1431 CDEBUG(D_NET, "sending %d bytes in %d frags to %s\n",
1432 payload_nob, payload_niov, libcfs_id2str(target));
1434 LASSERT(payload_nob == 0 || payload_niov > 0);
1435 LASSERT(payload_niov <= LNET_MAX_IOV);
1437 /* Thread context */
1438 LASSERT(!in_interrupt());
1439 /* payload is either all vaddrs or all pages */
1440 LASSERT(!(payload_kiov != NULL && payload_iov != NULL));
1448 LASSERT(payload_nob == 0);
1452 if (routing || target_is_router)
1453 break; /* send IMMEDIATE */
1455 /* is the REPLY message too small for RDMA? */
1456 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[lntmsg->msg_md->md_length]);
1457 if (nob <= IBLND_MSG_SIZE)
1458 break; /* send IMMEDIATE */
1460 tx = kiblnd_get_idle_tx(ni, target.nid);
1462 CERROR("Can't allocate txd for GET to %s\n",
1463 libcfs_nid2str(target.nid));
1469 if ((lntmsg->msg_md->md_options & LNET_MD_KIOV) == 0)
1470 rc = kiblnd_setup_rd_iov(ni, tx,
1471 &ibmsg->ibm_u.get.ibgm_rd,
1472 lntmsg->msg_md->md_niov,
1473 lntmsg->msg_md->md_iov.iov,
1474 0, lntmsg->msg_md->md_length);
1476 rc = kiblnd_setup_rd_kiov(ni, tx,
1477 &ibmsg->ibm_u.get.ibgm_rd,
1478 lntmsg->msg_md->md_niov,
1479 lntmsg->msg_md->md_iov.kiov,
1480 0, lntmsg->msg_md->md_length);
1482 CERROR("Can't setup GET sink for %s: %d\n",
1483 libcfs_nid2str(target.nid), rc);
1484 kiblnd_tx_done(ni, tx);
1488 nob = offsetof(kib_get_msg_t, ibgm_rd.rd_frags[tx->tx_nfrags]);
1489 ibmsg->ibm_u.get.ibgm_cookie = tx->tx_cookie;
1490 ibmsg->ibm_u.get.ibgm_hdr = *hdr;
1492 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_GET_REQ, nob);
1494 tx->tx_lntmsg[1] = lnet_create_reply_msg(ni, lntmsg);
1495 if (tx->tx_lntmsg[1] == NULL) {
1496 CERROR("Can't create reply for GET -> %s\n",
1497 libcfs_nid2str(target.nid));
1498 kiblnd_tx_done(ni, tx);
1502 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg[0,1] on completion */
1503 tx->tx_waiting = 1; /* waiting for GET_DONE */
1504 kiblnd_launch_tx(ni, tx, target.nid);
1507 case LNET_MSG_REPLY:
1509 /* Is the payload small enough not to need RDMA? */
1510 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob]);
1511 if (nob <= IBLND_MSG_SIZE)
1512 break; /* send IMMEDIATE */
1514 tx = kiblnd_get_idle_tx(ni, target.nid);
1516 CERROR("Can't allocate %s txd for %s\n",
1517 type == LNET_MSG_PUT ? "PUT" : "REPLY",
1518 libcfs_nid2str(target.nid));
1522 if (payload_kiov == NULL)
1523 rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1524 payload_niov, payload_iov,
1525 payload_offset, payload_nob);
1527 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1528 payload_niov, payload_kiov,
1529 payload_offset, payload_nob);
1531 CERROR("Can't setup PUT src for %s: %d\n",
1532 libcfs_nid2str(target.nid), rc);
1533 kiblnd_tx_done(ni, tx);
1538 ibmsg->ibm_u.putreq.ibprm_hdr = *hdr;
1539 ibmsg->ibm_u.putreq.ibprm_cookie = tx->tx_cookie;
1540 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_REQ, sizeof(kib_putreq_msg_t));
1542 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1543 tx->tx_waiting = 1; /* waiting for PUT_{ACK,NAK} */
1544 kiblnd_launch_tx(ni, tx, target.nid);
1548 /* send IMMEDIATE */
1550 LASSERT(offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob])
1553 tx = kiblnd_get_idle_tx(ni, target.nid);
1555 CERROR("Can't send %d to %s: tx descs exhausted\n",
1556 type, libcfs_nid2str(target.nid));
1561 ibmsg->ibm_u.immediate.ibim_hdr = *hdr;
1563 if (payload_kiov != NULL)
1564 lnet_copy_kiov2flat(IBLND_MSG_SIZE, ibmsg,
1565 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1566 payload_niov, payload_kiov,
1567 payload_offset, payload_nob);
1569 lnet_copy_iov2flat(IBLND_MSG_SIZE, ibmsg,
1570 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1571 payload_niov, payload_iov,
1572 payload_offset, payload_nob);
1574 nob = offsetof(kib_immediate_msg_t, ibim_payload[payload_nob]);
1575 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_IMMEDIATE, nob);
1577 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1578 kiblnd_launch_tx(ni, tx, target.nid);
1583 kiblnd_reply(lnet_ni_t *ni, kib_rx_t *rx, lnet_msg_t *lntmsg)
1585 lnet_process_id_t target = lntmsg->msg_target;
1586 unsigned int niov = lntmsg->msg_niov;
1587 struct kvec *iov = lntmsg->msg_iov;
1588 lnet_kiov_t *kiov = lntmsg->msg_kiov;
1589 unsigned int offset = lntmsg->msg_offset;
1590 unsigned int nob = lntmsg->msg_len;
1594 tx = kiblnd_get_idle_tx(ni, rx->rx_conn->ibc_peer->ibp_nid);
1596 CERROR("Can't get tx for REPLY to %s\n",
1597 libcfs_nid2str(target.nid));
1603 else if (kiov == NULL)
1604 rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1605 niov, iov, offset, nob);
1607 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1608 niov, kiov, offset, nob);
1611 CERROR("Can't setup GET src for %s: %d\n",
1612 libcfs_nid2str(target.nid), rc);
1616 rc = kiblnd_init_rdma(rx->rx_conn, tx,
1617 IBLND_MSG_GET_DONE, nob,
1618 &rx->rx_msg->ibm_u.get.ibgm_rd,
1619 rx->rx_msg->ibm_u.get.ibgm_cookie);
1621 CERROR("Can't setup rdma for GET from %s: %d\n",
1622 libcfs_nid2str(target.nid), rc);
1627 /* No RDMA: local completion may happen now! */
1628 lnet_finalize(ni, lntmsg, 0);
1630 /* RDMA: lnet_finalize(lntmsg) when it
1632 tx->tx_lntmsg[0] = lntmsg;
1635 kiblnd_queue_tx(tx, rx->rx_conn);
1639 kiblnd_tx_done(ni, tx);
1641 lnet_finalize(ni, lntmsg, -EIO);
1645 kiblnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed,
1646 unsigned int niov, struct kvec *iov, lnet_kiov_t *kiov,
1647 unsigned int offset, unsigned int mlen, unsigned int rlen)
1649 kib_rx_t *rx = private;
1650 kib_msg_t *rxmsg = rx->rx_msg;
1651 kib_conn_t *conn = rx->rx_conn;
1655 int post_credit = IBLND_POSTRX_PEER_CREDIT;
1658 LASSERT(mlen <= rlen);
1659 LASSERT(!in_interrupt());
1660 /* Either all pages or all vaddrs */
1661 LASSERT(!(kiov != NULL && iov != NULL));
1663 switch (rxmsg->ibm_type) {
1667 case IBLND_MSG_IMMEDIATE:
1668 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[rlen]);
1669 if (nob > rx->rx_nob) {
1670 CERROR("Immediate message from %s too big: %d(%d)\n",
1671 libcfs_nid2str(rxmsg->ibm_u.immediate.ibim_hdr.src_nid),
1678 lnet_copy_flat2kiov(niov, kiov, offset,
1679 IBLND_MSG_SIZE, rxmsg,
1680 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1683 lnet_copy_flat2iov(niov, iov, offset,
1684 IBLND_MSG_SIZE, rxmsg,
1685 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1687 lnet_finalize(ni, lntmsg, 0);
1690 case IBLND_MSG_PUT_REQ:
1692 lnet_finalize(ni, lntmsg, 0);
1693 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK, 0,
1694 rxmsg->ibm_u.putreq.ibprm_cookie);
1698 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
1700 CERROR("Can't allocate tx for %s\n",
1701 libcfs_nid2str(conn->ibc_peer->ibp_nid));
1702 /* Not replying will break the connection */
1709 rc = kiblnd_setup_rd_iov(ni, tx,
1710 &txmsg->ibm_u.putack.ibpam_rd,
1711 niov, iov, offset, mlen);
1713 rc = kiblnd_setup_rd_kiov(ni, tx,
1714 &txmsg->ibm_u.putack.ibpam_rd,
1715 niov, kiov, offset, mlen);
1717 CERROR("Can't setup PUT sink for %s: %d\n",
1718 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
1719 kiblnd_tx_done(ni, tx);
1720 /* tell peer it's over */
1721 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK, rc,
1722 rxmsg->ibm_u.putreq.ibprm_cookie);
1726 nob = offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[tx->tx_nfrags]);
1727 txmsg->ibm_u.putack.ibpam_src_cookie = rxmsg->ibm_u.putreq.ibprm_cookie;
1728 txmsg->ibm_u.putack.ibpam_dst_cookie = tx->tx_cookie;
1730 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_ACK, nob);
1732 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1733 tx->tx_waiting = 1; /* waiting for PUT_DONE */
1734 kiblnd_queue_tx(tx, conn);
1736 /* reposted buffer reserved for PUT_DONE */
1737 post_credit = IBLND_POSTRX_NO_CREDIT;
1740 case IBLND_MSG_GET_REQ:
1741 if (lntmsg != NULL) {
1742 /* Optimized GET; RDMA lntmsg's payload */
1743 kiblnd_reply(ni, rx, lntmsg);
1745 /* GET didn't match anything */
1746 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_GET_DONE,
1748 rxmsg->ibm_u.get.ibgm_cookie);
1753 kiblnd_post_rx(rx, post_credit);
1758 kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name)
1760 struct task_struct *task = kthread_run(fn, arg, "%s", name);
1763 return PTR_ERR(task);
1765 atomic_inc(&kiblnd_data.kib_nthreads);
1770 kiblnd_thread_fini(void)
1772 atomic_dec(&kiblnd_data.kib_nthreads);
1776 kiblnd_peer_alive(kib_peer_t *peer)
1778 /* This is racy, but everyone's only writing cfs_time_current() */
1779 peer->ibp_last_alive = cfs_time_current();
1784 kiblnd_peer_notify(kib_peer_t *peer)
1787 unsigned long last_alive = 0;
1788 unsigned long flags;
1790 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1792 if (list_empty(&peer->ibp_conns) &&
1793 peer->ibp_accepting == 0 &&
1794 peer->ibp_connecting == 0 &&
1795 peer->ibp_error != 0) {
1796 error = peer->ibp_error;
1797 peer->ibp_error = 0;
1799 last_alive = peer->ibp_last_alive;
1802 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1805 lnet_notify(peer->ibp_ni,
1806 peer->ibp_nid, 0, last_alive);
1810 kiblnd_close_conn_locked(kib_conn_t *conn, int error)
1812 /* This just does the immediate housekeeping. 'error' is zero for a
1813 * normal shutdown which can happen only after the connection has been
1814 * established. If the connection is established, schedule the
1815 * connection to be finished off by the connd. Otherwise the connd is
1816 * already dealing with it (either to set it up or tear it down).
1817 * Caller holds kib_global_lock exclusively in irq context */
1818 kib_peer_t *peer = conn->ibc_peer;
1820 unsigned long flags;
1822 LASSERT(error != 0 || conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1824 if (error != 0 && conn->ibc_comms_error == 0)
1825 conn->ibc_comms_error = error;
1827 if (conn->ibc_state != IBLND_CONN_ESTABLISHED)
1828 return; /* already being handled */
1831 list_empty(&conn->ibc_tx_noops) &&
1832 list_empty(&conn->ibc_tx_queue) &&
1833 list_empty(&conn->ibc_tx_queue_rsrvd) &&
1834 list_empty(&conn->ibc_tx_queue_nocred) &&
1835 list_empty(&conn->ibc_active_txs)) {
1836 CDEBUG(D_NET, "closing conn to %s\n",
1837 libcfs_nid2str(peer->ibp_nid));
1839 CNETERR("Closing conn to %s: error %d%s%s%s%s%s\n",
1840 libcfs_nid2str(peer->ibp_nid), error,
1841 list_empty(&conn->ibc_tx_queue) ? "" : "(sending)",
1842 list_empty(&conn->ibc_tx_noops) ? "" : "(sending_noops)",
1843 list_empty(&conn->ibc_tx_queue_rsrvd) ? "" : "(sending_rsrvd)",
1844 list_empty(&conn->ibc_tx_queue_nocred) ? "" : "(sending_nocred)",
1845 list_empty(&conn->ibc_active_txs) ? "" : "(waiting)");
1848 dev = ((kib_net_t *)peer->ibp_ni->ni_data)->ibn_dev;
1849 list_del(&conn->ibc_list);
1850 /* connd (see below) takes over ibc_list's ref */
1852 if (list_empty(&peer->ibp_conns) && /* no more conns */
1853 kiblnd_peer_active(peer)) { /* still in peer table */
1854 kiblnd_unlink_peer_locked(peer);
1856 /* set/clear error on last conn */
1857 peer->ibp_error = conn->ibc_comms_error;
1860 kiblnd_set_conn_state(conn, IBLND_CONN_CLOSING);
1863 kiblnd_dev_can_failover(dev)) {
1864 list_add_tail(&dev->ibd_fail_list,
1865 &kiblnd_data.kib_failed_devs);
1866 wake_up(&kiblnd_data.kib_failover_waitq);
1869 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
1871 list_add_tail(&conn->ibc_list, &kiblnd_data.kib_connd_conns);
1872 wake_up(&kiblnd_data.kib_connd_waitq);
1874 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
1878 kiblnd_close_conn(kib_conn_t *conn, int error)
1880 unsigned long flags;
1882 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1884 kiblnd_close_conn_locked(conn, error);
1886 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1890 kiblnd_handle_early_rxs(kib_conn_t *conn)
1892 unsigned long flags;
1896 LASSERT(!in_interrupt());
1897 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1899 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1900 list_for_each_entry_safe(rx, tmp, &conn->ibc_early_rxs, rx_list) {
1901 list_del(&rx->rx_list);
1902 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1904 kiblnd_handle_rx(rx);
1906 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1908 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1912 kiblnd_abort_txs(kib_conn_t *conn, struct list_head *txs)
1915 struct list_head *tmp;
1916 struct list_head *nxt;
1919 spin_lock(&conn->ibc_lock);
1921 list_for_each_safe(tmp, nxt, txs) {
1922 tx = list_entry(tmp, kib_tx_t, tx_list);
1924 if (txs == &conn->ibc_active_txs) {
1925 LASSERT(!tx->tx_queued);
1926 LASSERT(tx->tx_waiting ||
1927 tx->tx_sending != 0);
1929 LASSERT(tx->tx_queued);
1932 tx->tx_status = -ECONNABORTED;
1935 if (tx->tx_sending == 0) {
1937 list_del(&tx->tx_list);
1938 list_add(&tx->tx_list, &zombies);
1942 spin_unlock(&conn->ibc_lock);
1944 kiblnd_txlist_done(conn->ibc_peer->ibp_ni, &zombies, -ECONNABORTED);
1948 kiblnd_finalise_conn(kib_conn_t *conn)
1950 LASSERT(!in_interrupt());
1951 LASSERT(conn->ibc_state > IBLND_CONN_INIT);
1953 kiblnd_set_conn_state(conn, IBLND_CONN_DISCONNECTED);
1955 /* abort_receives moves QP state to IB_QPS_ERR. This is only required
1956 * for connections that didn't get as far as being connected, because
1957 * rdma_disconnect() does this for free. */
1958 kiblnd_abort_receives(conn);
1960 /* Complete all tx descs not waiting for sends to complete.
1961 * NB we should be safe from RDMA now that the QP has changed state */
1963 kiblnd_abort_txs(conn, &conn->ibc_tx_noops);
1964 kiblnd_abort_txs(conn, &conn->ibc_tx_queue);
1965 kiblnd_abort_txs(conn, &conn->ibc_tx_queue_rsrvd);
1966 kiblnd_abort_txs(conn, &conn->ibc_tx_queue_nocred);
1967 kiblnd_abort_txs(conn, &conn->ibc_active_txs);
1969 kiblnd_handle_early_rxs(conn);
1973 kiblnd_peer_connect_failed(kib_peer_t *peer, int active, int error)
1976 unsigned long flags;
1978 LASSERT(error != 0);
1979 LASSERT(!in_interrupt());
1981 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1984 LASSERT(peer->ibp_connecting > 0);
1985 peer->ibp_connecting--;
1987 LASSERT(peer->ibp_accepting > 0);
1988 peer->ibp_accepting--;
1991 if (peer->ibp_connecting != 0 ||
1992 peer->ibp_accepting != 0) {
1993 /* another connection attempt under way... */
1994 write_unlock_irqrestore(&kiblnd_data.kib_global_lock,
1999 if (list_empty(&peer->ibp_conns)) {
2000 /* Take peer's blocked transmits to complete with error */
2001 list_add(&zombies, &peer->ibp_tx_queue);
2002 list_del_init(&peer->ibp_tx_queue);
2004 if (kiblnd_peer_active(peer))
2005 kiblnd_unlink_peer_locked(peer);
2007 peer->ibp_error = error;
2009 /* Can't have blocked transmits if there are connections */
2010 LASSERT(list_empty(&peer->ibp_tx_queue));
2013 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2015 kiblnd_peer_notify(peer);
2017 if (list_empty(&zombies))
2020 CNETERR("Deleting messages for %s: connection failed\n",
2021 libcfs_nid2str(peer->ibp_nid));
2023 kiblnd_txlist_done(peer->ibp_ni, &zombies, -EHOSTUNREACH);
2027 kiblnd_connreq_done(kib_conn_t *conn, int status)
2029 kib_peer_t *peer = conn->ibc_peer;
2032 struct list_head txs;
2033 unsigned long flags;
2036 active = (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2038 CDEBUG(D_NET, "%s: active(%d), version(%x), status(%d)\n",
2039 libcfs_nid2str(peer->ibp_nid), active,
2040 conn->ibc_version, status);
2042 LASSERT(!in_interrupt());
2043 LASSERT((conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT &&
2044 peer->ibp_connecting > 0) ||
2045 (conn->ibc_state == IBLND_CONN_PASSIVE_WAIT &&
2046 peer->ibp_accepting > 0));
2048 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
2049 conn->ibc_connvars = NULL;
2052 /* failed to establish connection */
2053 kiblnd_peer_connect_failed(peer, active, status);
2054 kiblnd_finalise_conn(conn);
2058 /* connection established */
2059 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2061 conn->ibc_last_send = jiffies;
2062 kiblnd_set_conn_state(conn, IBLND_CONN_ESTABLISHED);
2063 kiblnd_peer_alive(peer);
2065 /* Add conn to peer's list and nuke any dangling conns from a different
2066 * peer instance... */
2067 kiblnd_conn_addref(conn); /* +1 ref for ibc_list */
2068 list_add(&conn->ibc_list, &peer->ibp_conns);
2070 peer->ibp_connecting--;
2072 peer->ibp_accepting--;
2074 if (peer->ibp_version == 0) {
2075 peer->ibp_version = conn->ibc_version;
2076 peer->ibp_incarnation = conn->ibc_incarnation;
2079 if (peer->ibp_version != conn->ibc_version ||
2080 peer->ibp_incarnation != conn->ibc_incarnation) {
2081 kiblnd_close_stale_conns_locked(peer, conn->ibc_version,
2082 conn->ibc_incarnation);
2083 peer->ibp_version = conn->ibc_version;
2084 peer->ibp_incarnation = conn->ibc_incarnation;
2087 /* grab pending txs while I have the lock */
2088 list_add(&txs, &peer->ibp_tx_queue);
2089 list_del_init(&peer->ibp_tx_queue);
2091 if (!kiblnd_peer_active(peer) || /* peer has been deleted */
2092 conn->ibc_comms_error != 0) { /* error has happened already */
2093 lnet_ni_t *ni = peer->ibp_ni;
2095 /* start to shut down connection */
2096 kiblnd_close_conn_locked(conn, -ECONNABORTED);
2097 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2099 kiblnd_txlist_done(ni, &txs, -ECONNABORTED);
2104 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2106 /* Schedule blocked txs */
2107 spin_lock(&conn->ibc_lock);
2108 list_for_each_entry_safe(tx, tmp, &txs, tx_list) {
2109 list_del(&tx->tx_list);
2111 kiblnd_queue_tx_locked(tx, conn);
2113 spin_unlock(&conn->ibc_lock);
2115 kiblnd_check_sends(conn);
2117 /* schedule blocked rxs */
2118 kiblnd_handle_early_rxs(conn);
2122 kiblnd_reject(struct rdma_cm_id *cmid, kib_rej_t *rej)
2126 rc = rdma_reject(cmid, rej, sizeof(*rej));
2129 CWARN("Error %d sending reject\n", rc);
2133 kiblnd_passive_connect(struct rdma_cm_id *cmid, void *priv, int priv_nob)
2135 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2136 kib_msg_t *reqmsg = priv;
2142 lnet_ni_t *ni = NULL;
2143 kib_net_t *net = NULL;
2145 struct rdma_conn_param cp;
2147 int version = IBLND_MSG_VERSION;
2148 unsigned long flags;
2150 struct sockaddr_in *peer_addr;
2151 LASSERT(!in_interrupt());
2153 /* cmid inherits 'context' from the corresponding listener id */
2154 ibdev = (kib_dev_t *)cmid->context;
2155 LASSERT(ibdev != NULL);
2157 memset(&rej, 0, sizeof(rej));
2158 rej.ibr_magic = IBLND_MSG_MAGIC;
2159 rej.ibr_why = IBLND_REJECT_FATAL;
2160 rej.ibr_cp.ibcp_max_msg_size = IBLND_MSG_SIZE;
2162 peer_addr = (struct sockaddr_in *)&(cmid->route.addr.dst_addr);
2163 if (*kiblnd_tunables.kib_require_priv_port &&
2164 ntohs(peer_addr->sin_port) >= PROT_SOCK) {
2165 __u32 ip = ntohl(peer_addr->sin_addr.s_addr);
2166 CERROR("Peer's port (%pI4h:%hu) is not privileged\n",
2167 &ip, ntohs(peer_addr->sin_port));
2171 if (priv_nob < offsetof(kib_msg_t, ibm_type)) {
2172 CERROR("Short connection request\n");
2176 /* Future protocol version compatibility support! If the
2177 * o2iblnd-specific protocol changes, or when LNET unifies
2178 * protocols over all LNDs, the initial connection will
2179 * negotiate a protocol version. I trap this here to avoid
2180 * console errors; the reject tells the peer which protocol I
2182 if (reqmsg->ibm_magic == LNET_PROTO_MAGIC ||
2183 reqmsg->ibm_magic == __swab32(LNET_PROTO_MAGIC))
2185 if (reqmsg->ibm_magic == IBLND_MSG_MAGIC &&
2186 reqmsg->ibm_version != IBLND_MSG_VERSION &&
2187 reqmsg->ibm_version != IBLND_MSG_VERSION_1)
2189 if (reqmsg->ibm_magic == __swab32(IBLND_MSG_MAGIC) &&
2190 reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION) &&
2191 reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION_1))
2194 rc = kiblnd_unpack_msg(reqmsg, priv_nob);
2196 CERROR("Can't parse connection request: %d\n", rc);
2200 nid = reqmsg->ibm_srcnid;
2201 ni = lnet_net2ni(LNET_NIDNET(reqmsg->ibm_dstnid));
2204 net = (kib_net_t *)ni->ni_data;
2205 rej.ibr_incarnation = net->ibn_incarnation;
2208 if (ni == NULL || /* no matching net */
2209 ni->ni_nid != reqmsg->ibm_dstnid || /* right NET, wrong NID! */
2210 net->ibn_dev != ibdev) { /* wrong device */
2211 CERROR("Can't accept %s on %s (%s:%d:%pI4h): bad dst nid %s\n",
2212 libcfs_nid2str(nid),
2213 ni == NULL ? "NA" : libcfs_nid2str(ni->ni_nid),
2214 ibdev->ibd_ifname, ibdev->ibd_nnets,
2216 libcfs_nid2str(reqmsg->ibm_dstnid));
2221 /* check time stamp as soon as possible */
2222 if (reqmsg->ibm_dststamp != 0 &&
2223 reqmsg->ibm_dststamp != net->ibn_incarnation) {
2224 CWARN("Stale connection request\n");
2225 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2229 /* I can accept peer's version */
2230 version = reqmsg->ibm_version;
2232 if (reqmsg->ibm_type != IBLND_MSG_CONNREQ) {
2233 CERROR("Unexpected connreq msg type: %x from %s\n",
2234 reqmsg->ibm_type, libcfs_nid2str(nid));
2238 if (reqmsg->ibm_u.connparams.ibcp_queue_depth !=
2239 IBLND_MSG_QUEUE_SIZE(version)) {
2240 CERROR("Can't accept %s: incompatible queue depth %d (%d wanted)\n",
2241 libcfs_nid2str(nid), reqmsg->ibm_u.connparams.ibcp_queue_depth,
2242 IBLND_MSG_QUEUE_SIZE(version));
2244 if (version == IBLND_MSG_VERSION)
2245 rej.ibr_why = IBLND_REJECT_MSG_QUEUE_SIZE;
2250 if (reqmsg->ibm_u.connparams.ibcp_max_frags !=
2251 IBLND_RDMA_FRAGS(version)) {
2252 CERROR("Can't accept %s(version %x): incompatible max_frags %d (%d wanted)\n",
2253 libcfs_nid2str(nid), version,
2254 reqmsg->ibm_u.connparams.ibcp_max_frags,
2255 IBLND_RDMA_FRAGS(version));
2257 if (version == IBLND_MSG_VERSION)
2258 rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2264 if (reqmsg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2265 CERROR("Can't accept %s: message size %d too big (%d max)\n",
2266 libcfs_nid2str(nid),
2267 reqmsg->ibm_u.connparams.ibcp_max_msg_size,
2272 /* assume 'nid' is a new peer; create */
2273 rc = kiblnd_create_peer(ni, &peer, nid);
2275 CERROR("Can't create peer for %s\n", libcfs_nid2str(nid));
2276 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2280 write_lock_irqsave(g_lock, flags);
2282 peer2 = kiblnd_find_peer_locked(nid);
2283 if (peer2 != NULL) {
2284 if (peer2->ibp_version == 0) {
2285 peer2->ibp_version = version;
2286 peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2289 /* not the guy I've talked with */
2290 if (peer2->ibp_incarnation != reqmsg->ibm_srcstamp ||
2291 peer2->ibp_version != version) {
2292 kiblnd_close_peer_conns_locked(peer2, -ESTALE);
2293 write_unlock_irqrestore(g_lock, flags);
2295 CWARN("Conn stale %s [old ver: %x, new ver: %x]\n",
2296 libcfs_nid2str(nid), peer2->ibp_version, version);
2298 kiblnd_peer_decref(peer);
2299 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2303 /* tie-break connection race in favour of the higher NID */
2304 if (peer2->ibp_connecting != 0 &&
2306 write_unlock_irqrestore(g_lock, flags);
2308 CWARN("Conn race %s\n", libcfs_nid2str(peer2->ibp_nid));
2310 kiblnd_peer_decref(peer);
2311 rej.ibr_why = IBLND_REJECT_CONN_RACE;
2315 peer2->ibp_accepting++;
2316 kiblnd_peer_addref(peer2);
2318 write_unlock_irqrestore(g_lock, flags);
2319 kiblnd_peer_decref(peer);
2322 /* Brand new peer */
2323 LASSERT(peer->ibp_accepting == 0);
2324 LASSERT(peer->ibp_version == 0 &&
2325 peer->ibp_incarnation == 0);
2327 peer->ibp_accepting = 1;
2328 peer->ibp_version = version;
2329 peer->ibp_incarnation = reqmsg->ibm_srcstamp;
2331 /* I have a ref on ni that prevents it being shutdown */
2332 LASSERT(net->ibn_shutdown == 0);
2334 kiblnd_peer_addref(peer);
2335 list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid));
2337 write_unlock_irqrestore(g_lock, flags);
2340 conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_PASSIVE_WAIT, version);
2342 kiblnd_peer_connect_failed(peer, 0, -ENOMEM);
2343 kiblnd_peer_decref(peer);
2344 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2348 /* conn now "owns" cmid, so I return success from here on to ensure the
2349 * CM callback doesn't destroy cmid. */
2351 conn->ibc_incarnation = reqmsg->ibm_srcstamp;
2352 conn->ibc_credits = IBLND_MSG_QUEUE_SIZE(version);
2353 conn->ibc_reserved_credits = IBLND_MSG_QUEUE_SIZE(version);
2354 LASSERT(conn->ibc_credits + conn->ibc_reserved_credits + IBLND_OOB_MSGS(version)
2355 <= IBLND_RX_MSGS(version));
2357 ackmsg = &conn->ibc_connvars->cv_msg;
2358 memset(ackmsg, 0, sizeof(*ackmsg));
2360 kiblnd_init_msg(ackmsg, IBLND_MSG_CONNACK,
2361 sizeof(ackmsg->ibm_u.connparams));
2362 ackmsg->ibm_u.connparams.ibcp_queue_depth = IBLND_MSG_QUEUE_SIZE(version);
2363 ackmsg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2364 ackmsg->ibm_u.connparams.ibcp_max_frags = IBLND_RDMA_FRAGS(version);
2366 kiblnd_pack_msg(ni, ackmsg, version, 0, nid, reqmsg->ibm_srcstamp);
2368 memset(&cp, 0, sizeof(cp));
2369 cp.private_data = ackmsg;
2370 cp.private_data_len = ackmsg->ibm_nob;
2371 cp.responder_resources = 0; /* No atomic ops or RDMA reads */
2372 cp.initiator_depth = 0;
2373 cp.flow_control = 1;
2374 cp.retry_count = *kiblnd_tunables.kib_retry_count;
2375 cp.rnr_retry_count = *kiblnd_tunables.kib_rnr_retry_count;
2377 CDEBUG(D_NET, "Accept %s\n", libcfs_nid2str(nid));
2379 rc = rdma_accept(cmid, &cp);
2381 CERROR("Can't accept %s: %d\n", libcfs_nid2str(nid), rc);
2382 rej.ibr_version = version;
2383 rej.ibr_why = IBLND_REJECT_FATAL;
2385 kiblnd_reject(cmid, &rej);
2386 kiblnd_connreq_done(conn, rc);
2387 kiblnd_conn_decref(conn);
2397 rej.ibr_version = version;
2398 rej.ibr_cp.ibcp_queue_depth = IBLND_MSG_QUEUE_SIZE(version);
2399 rej.ibr_cp.ibcp_max_frags = IBLND_RDMA_FRAGS(version);
2400 kiblnd_reject(cmid, &rej);
2402 return -ECONNREFUSED;
2406 kiblnd_reconnect(kib_conn_t *conn, int version,
2407 __u64 incarnation, int why, kib_connparams_t *cp)
2409 kib_peer_t *peer = conn->ibc_peer;
2412 unsigned long flags;
2414 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2415 LASSERT(peer->ibp_connecting > 0); /* 'conn' at least */
2417 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2419 /* retry connection if it's still needed and no other connection
2420 * attempts (active or passive) are in progress
2421 * NB: reconnect is still needed even when ibp_tx_queue is
2422 * empty if ibp_version != version because reconnect may be
2423 * initiated by kiblnd_query() */
2424 if ((!list_empty(&peer->ibp_tx_queue) ||
2425 peer->ibp_version != version) &&
2426 peer->ibp_connecting == 1 &&
2427 peer->ibp_accepting == 0) {
2429 peer->ibp_connecting++;
2431 peer->ibp_version = version;
2432 peer->ibp_incarnation = incarnation;
2435 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2445 case IBLND_REJECT_CONN_STALE:
2449 case IBLND_REJECT_CONN_RACE:
2450 reason = "conn race";
2453 case IBLND_REJECT_CONN_UNCOMPAT:
2454 reason = "version negotiation";
2458 CNETERR("%s: retrying (%s), %x, %x, queue_dep: %d, max_frag: %d, msg_size: %d\n",
2459 libcfs_nid2str(peer->ibp_nid),
2460 reason, IBLND_MSG_VERSION, version,
2461 cp != NULL ? cp->ibcp_queue_depth : IBLND_MSG_QUEUE_SIZE(version),
2462 cp != NULL ? cp->ibcp_max_frags : IBLND_RDMA_FRAGS(version),
2463 cp != NULL ? cp->ibcp_max_msg_size : IBLND_MSG_SIZE);
2465 kiblnd_connect_peer(peer);
2469 kiblnd_rejected(kib_conn_t *conn, int reason, void *priv, int priv_nob)
2471 kib_peer_t *peer = conn->ibc_peer;
2473 LASSERT(!in_interrupt());
2474 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2477 case IB_CM_REJ_STALE_CONN:
2478 kiblnd_reconnect(conn, IBLND_MSG_VERSION, 0,
2479 IBLND_REJECT_CONN_STALE, NULL);
2482 case IB_CM_REJ_INVALID_SERVICE_ID:
2483 CNETERR("%s rejected: no listener at %d\n",
2484 libcfs_nid2str(peer->ibp_nid),
2485 *kiblnd_tunables.kib_service);
2488 case IB_CM_REJ_CONSUMER_DEFINED:
2489 if (priv_nob >= offsetof(kib_rej_t, ibr_padding)) {
2490 kib_rej_t *rej = priv;
2491 kib_connparams_t *cp = NULL;
2493 __u64 incarnation = -1;
2495 /* NB. default incarnation is -1 because:
2496 * a) V1 will ignore dst incarnation in connreq.
2497 * b) V2 will provide incarnation while rejecting me,
2498 * -1 will be overwrote.
2500 * if I try to connect to a V1 peer with V2 protocol,
2501 * it rejected me then upgrade to V2, I have no idea
2502 * about the upgrading and try to reconnect with V1,
2503 * in this case upgraded V2 can find out I'm trying to
2504 * talk to the old guy and reject me(incarnation is -1).
2507 if (rej->ibr_magic == __swab32(IBLND_MSG_MAGIC) ||
2508 rej->ibr_magic == __swab32(LNET_PROTO_MAGIC)) {
2509 __swab32s(&rej->ibr_magic);
2510 __swab16s(&rej->ibr_version);
2514 if (priv_nob >= sizeof(kib_rej_t) &&
2515 rej->ibr_version > IBLND_MSG_VERSION_1) {
2516 /* priv_nob is always 148 in current version
2517 * of OFED, so we still need to check version.
2518 * (define of IB_CM_REJ_PRIVATE_DATA_SIZE) */
2522 __swab64s(&rej->ibr_incarnation);
2523 __swab16s(&cp->ibcp_queue_depth);
2524 __swab16s(&cp->ibcp_max_frags);
2525 __swab32s(&cp->ibcp_max_msg_size);
2528 incarnation = rej->ibr_incarnation;
2531 if (rej->ibr_magic != IBLND_MSG_MAGIC &&
2532 rej->ibr_magic != LNET_PROTO_MAGIC) {
2533 CERROR("%s rejected: consumer defined fatal error\n",
2534 libcfs_nid2str(peer->ibp_nid));
2538 if (rej->ibr_version != IBLND_MSG_VERSION &&
2539 rej->ibr_version != IBLND_MSG_VERSION_1) {
2540 CERROR("%s rejected: o2iblnd version %x error\n",
2541 libcfs_nid2str(peer->ibp_nid),
2546 if (rej->ibr_why == IBLND_REJECT_FATAL &&
2547 rej->ibr_version == IBLND_MSG_VERSION_1) {
2548 CDEBUG(D_NET, "rejected by old version peer %s: %x\n",
2549 libcfs_nid2str(peer->ibp_nid), rej->ibr_version);
2551 if (conn->ibc_version != IBLND_MSG_VERSION_1)
2552 rej->ibr_why = IBLND_REJECT_CONN_UNCOMPAT;
2555 switch (rej->ibr_why) {
2556 case IBLND_REJECT_CONN_RACE:
2557 case IBLND_REJECT_CONN_STALE:
2558 case IBLND_REJECT_CONN_UNCOMPAT:
2559 kiblnd_reconnect(conn, rej->ibr_version,
2560 incarnation, rej->ibr_why, cp);
2563 case IBLND_REJECT_MSG_QUEUE_SIZE:
2564 CERROR("%s rejected: incompatible message queue depth %d, %d\n",
2565 libcfs_nid2str(peer->ibp_nid),
2566 cp != NULL ? cp->ibcp_queue_depth :
2567 IBLND_MSG_QUEUE_SIZE(rej->ibr_version),
2568 IBLND_MSG_QUEUE_SIZE(conn->ibc_version));
2571 case IBLND_REJECT_RDMA_FRAGS:
2572 CERROR("%s rejected: incompatible # of RDMA fragments %d, %d\n",
2573 libcfs_nid2str(peer->ibp_nid),
2574 cp != NULL ? cp->ibcp_max_frags :
2575 IBLND_RDMA_FRAGS(rej->ibr_version),
2576 IBLND_RDMA_FRAGS(conn->ibc_version));
2579 case IBLND_REJECT_NO_RESOURCES:
2580 CERROR("%s rejected: o2iblnd no resources\n",
2581 libcfs_nid2str(peer->ibp_nid));
2584 case IBLND_REJECT_FATAL:
2585 CERROR("%s rejected: o2iblnd fatal error\n",
2586 libcfs_nid2str(peer->ibp_nid));
2590 CERROR("%s rejected: o2iblnd reason %d\n",
2591 libcfs_nid2str(peer->ibp_nid),
2599 CNETERR("%s rejected: reason %d, size %d\n",
2600 libcfs_nid2str(peer->ibp_nid), reason, priv_nob);
2604 kiblnd_connreq_done(conn, -ECONNREFUSED);
2608 kiblnd_check_connreply(kib_conn_t *conn, void *priv, int priv_nob)
2610 kib_peer_t *peer = conn->ibc_peer;
2611 lnet_ni_t *ni = peer->ibp_ni;
2612 kib_net_t *net = ni->ni_data;
2613 kib_msg_t *msg = priv;
2614 int ver = conn->ibc_version;
2615 int rc = kiblnd_unpack_msg(msg, priv_nob);
2616 unsigned long flags;
2618 LASSERT(net != NULL);
2621 CERROR("Can't unpack connack from %s: %d\n",
2622 libcfs_nid2str(peer->ibp_nid), rc);
2626 if (msg->ibm_type != IBLND_MSG_CONNACK) {
2627 CERROR("Unexpected message %d from %s\n",
2628 msg->ibm_type, libcfs_nid2str(peer->ibp_nid));
2633 if (ver != msg->ibm_version) {
2634 CERROR("%s replied version %x is different with requested version %x\n",
2635 libcfs_nid2str(peer->ibp_nid), msg->ibm_version, ver);
2640 if (msg->ibm_u.connparams.ibcp_queue_depth !=
2641 IBLND_MSG_QUEUE_SIZE(ver)) {
2642 CERROR("%s has incompatible queue depth %d(%d wanted)\n",
2643 libcfs_nid2str(peer->ibp_nid),
2644 msg->ibm_u.connparams.ibcp_queue_depth,
2645 IBLND_MSG_QUEUE_SIZE(ver));
2650 if (msg->ibm_u.connparams.ibcp_max_frags !=
2651 IBLND_RDMA_FRAGS(ver)) {
2652 CERROR("%s has incompatible max_frags %d (%d wanted)\n",
2653 libcfs_nid2str(peer->ibp_nid),
2654 msg->ibm_u.connparams.ibcp_max_frags,
2655 IBLND_RDMA_FRAGS(ver));
2660 if (msg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2661 CERROR("%s max message size %d too big (%d max)\n",
2662 libcfs_nid2str(peer->ibp_nid),
2663 msg->ibm_u.connparams.ibcp_max_msg_size,
2669 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2670 if (msg->ibm_dstnid == ni->ni_nid &&
2671 msg->ibm_dststamp == net->ibn_incarnation)
2675 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2678 CERROR("Bad connection reply from %s, rc = %d, version: %x max_frags: %d\n",
2679 libcfs_nid2str(peer->ibp_nid), rc,
2680 msg->ibm_version, msg->ibm_u.connparams.ibcp_max_frags);
2684 conn->ibc_incarnation = msg->ibm_srcstamp;
2686 conn->ibc_reserved_credits = IBLND_MSG_QUEUE_SIZE(ver);
2687 LASSERT(conn->ibc_credits + conn->ibc_reserved_credits + IBLND_OOB_MSGS(ver)
2688 <= IBLND_RX_MSGS(ver));
2690 kiblnd_connreq_done(conn, 0);
2694 /* NB My QP has already established itself, so I handle anything going
2695 * wrong here by setting ibc_comms_error.
2696 * kiblnd_connreq_done(0) moves the conn state to ESTABLISHED, but then
2697 * immediately tears it down. */
2700 conn->ibc_comms_error = rc;
2701 kiblnd_connreq_done(conn, 0);
2705 kiblnd_active_connect(struct rdma_cm_id *cmid)
2707 kib_peer_t *peer = (kib_peer_t *)cmid->context;
2710 struct rdma_conn_param cp;
2713 unsigned long flags;
2716 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2718 incarnation = peer->ibp_incarnation;
2719 version = (peer->ibp_version == 0) ? IBLND_MSG_VERSION :
2722 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2724 conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_ACTIVE_CONNECT, version);
2726 kiblnd_peer_connect_failed(peer, 1, -ENOMEM);
2727 kiblnd_peer_decref(peer); /* lose cmid's ref */
2731 /* conn "owns" cmid now, so I return success from here on to ensure the
2732 * CM callback doesn't destroy cmid. conn also takes over cmid's ref
2735 msg = &conn->ibc_connvars->cv_msg;
2737 memset(msg, 0, sizeof(*msg));
2738 kiblnd_init_msg(msg, IBLND_MSG_CONNREQ, sizeof(msg->ibm_u.connparams));
2739 msg->ibm_u.connparams.ibcp_queue_depth = IBLND_MSG_QUEUE_SIZE(version);
2740 msg->ibm_u.connparams.ibcp_max_frags = IBLND_RDMA_FRAGS(version);
2741 msg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2743 kiblnd_pack_msg(peer->ibp_ni, msg, version,
2744 0, peer->ibp_nid, incarnation);
2746 memset(&cp, 0, sizeof(cp));
2747 cp.private_data = msg;
2748 cp.private_data_len = msg->ibm_nob;
2749 cp.responder_resources = 0; /* No atomic ops or RDMA reads */
2750 cp.initiator_depth = 0;
2751 cp.flow_control = 1;
2752 cp.retry_count = *kiblnd_tunables.kib_retry_count;
2753 cp.rnr_retry_count = *kiblnd_tunables.kib_rnr_retry_count;
2755 LASSERT(cmid->context == (void *)conn);
2756 LASSERT(conn->ibc_cmid == cmid);
2758 rc = rdma_connect(cmid, &cp);
2760 CERROR("Can't connect to %s: %d\n",
2761 libcfs_nid2str(peer->ibp_nid), rc);
2762 kiblnd_connreq_done(conn, rc);
2763 kiblnd_conn_decref(conn);
2770 kiblnd_cm_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2776 switch (event->event) {
2778 CERROR("Unexpected event: %d, status: %d\n",
2779 event->event, event->status);
2782 case RDMA_CM_EVENT_CONNECT_REQUEST:
2783 /* destroy cmid on failure */
2784 rc = kiblnd_passive_connect(cmid,
2785 (void *)KIBLND_CONN_PARAM(event),
2786 KIBLND_CONN_PARAM_LEN(event));
2787 CDEBUG(D_NET, "connreq: %d\n", rc);
2790 case RDMA_CM_EVENT_ADDR_ERROR:
2791 peer = (kib_peer_t *)cmid->context;
2792 CNETERR("%s: ADDR ERROR %d\n",
2793 libcfs_nid2str(peer->ibp_nid), event->status);
2794 kiblnd_peer_connect_failed(peer, 1, -EHOSTUNREACH);
2795 kiblnd_peer_decref(peer);
2796 return -EHOSTUNREACH; /* rc != 0 destroys cmid */
2798 case RDMA_CM_EVENT_ADDR_RESOLVED:
2799 peer = (kib_peer_t *)cmid->context;
2801 CDEBUG(D_NET, "%s Addr resolved: %d\n",
2802 libcfs_nid2str(peer->ibp_nid), event->status);
2804 if (event->status != 0) {
2805 CNETERR("Can't resolve address for %s: %d\n",
2806 libcfs_nid2str(peer->ibp_nid), event->status);
2809 rc = rdma_resolve_route(
2810 cmid, *kiblnd_tunables.kib_timeout * 1000);
2813 /* Can't initiate route resolution */
2814 CERROR("Can't resolve route for %s: %d\n",
2815 libcfs_nid2str(peer->ibp_nid), rc);
2817 kiblnd_peer_connect_failed(peer, 1, rc);
2818 kiblnd_peer_decref(peer);
2819 return rc; /* rc != 0 destroys cmid */
2821 case RDMA_CM_EVENT_ROUTE_ERROR:
2822 peer = (kib_peer_t *)cmid->context;
2823 CNETERR("%s: ROUTE ERROR %d\n",
2824 libcfs_nid2str(peer->ibp_nid), event->status);
2825 kiblnd_peer_connect_failed(peer, 1, -EHOSTUNREACH);
2826 kiblnd_peer_decref(peer);
2827 return -EHOSTUNREACH; /* rc != 0 destroys cmid */
2829 case RDMA_CM_EVENT_ROUTE_RESOLVED:
2830 peer = (kib_peer_t *)cmid->context;
2831 CDEBUG(D_NET, "%s Route resolved: %d\n",
2832 libcfs_nid2str(peer->ibp_nid), event->status);
2834 if (event->status == 0)
2835 return kiblnd_active_connect(cmid);
2837 CNETERR("Can't resolve route for %s: %d\n",
2838 libcfs_nid2str(peer->ibp_nid), event->status);
2839 kiblnd_peer_connect_failed(peer, 1, event->status);
2840 kiblnd_peer_decref(peer);
2841 return event->status; /* rc != 0 destroys cmid */
2843 case RDMA_CM_EVENT_UNREACHABLE:
2844 conn = (kib_conn_t *)cmid->context;
2845 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
2846 conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
2847 CNETERR("%s: UNREACHABLE %d\n",
2848 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
2849 kiblnd_connreq_done(conn, -ENETDOWN);
2850 kiblnd_conn_decref(conn);
2853 case RDMA_CM_EVENT_CONNECT_ERROR:
2854 conn = (kib_conn_t *)cmid->context;
2855 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
2856 conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
2857 CNETERR("%s: CONNECT ERROR %d\n",
2858 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
2859 kiblnd_connreq_done(conn, -ENOTCONN);
2860 kiblnd_conn_decref(conn);
2863 case RDMA_CM_EVENT_REJECTED:
2864 conn = (kib_conn_t *)cmid->context;
2865 switch (conn->ibc_state) {
2869 case IBLND_CONN_PASSIVE_WAIT:
2870 CERROR("%s: REJECTED %d\n",
2871 libcfs_nid2str(conn->ibc_peer->ibp_nid),
2873 kiblnd_connreq_done(conn, -ECONNRESET);
2876 case IBLND_CONN_ACTIVE_CONNECT:
2877 kiblnd_rejected(conn, event->status,
2878 (void *)KIBLND_CONN_PARAM(event),
2879 KIBLND_CONN_PARAM_LEN(event));
2882 kiblnd_conn_decref(conn);
2885 case RDMA_CM_EVENT_ESTABLISHED:
2886 conn = (kib_conn_t *)cmid->context;
2887 switch (conn->ibc_state) {
2891 case IBLND_CONN_PASSIVE_WAIT:
2892 CDEBUG(D_NET, "ESTABLISHED (passive): %s\n",
2893 libcfs_nid2str(conn->ibc_peer->ibp_nid));
2894 kiblnd_connreq_done(conn, 0);
2897 case IBLND_CONN_ACTIVE_CONNECT:
2898 CDEBUG(D_NET, "ESTABLISHED(active): %s\n",
2899 libcfs_nid2str(conn->ibc_peer->ibp_nid));
2900 kiblnd_check_connreply(conn,
2901 (void *)KIBLND_CONN_PARAM(event),
2902 KIBLND_CONN_PARAM_LEN(event));
2905 /* net keeps its ref on conn! */
2908 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
2909 CDEBUG(D_NET, "Ignore TIMEWAIT_EXIT event\n");
2911 case RDMA_CM_EVENT_DISCONNECTED:
2912 conn = (kib_conn_t *)cmid->context;
2913 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
2914 CERROR("%s DISCONNECTED\n",
2915 libcfs_nid2str(conn->ibc_peer->ibp_nid));
2916 kiblnd_connreq_done(conn, -ECONNRESET);
2918 kiblnd_close_conn(conn, 0);
2920 kiblnd_conn_decref(conn);
2921 cmid->context = NULL;
2924 case RDMA_CM_EVENT_DEVICE_REMOVAL:
2925 LCONSOLE_ERROR_MSG(0x131,
2926 "Received notification of device removal\n"
2927 "Please shutdown LNET to allow this to proceed\n");
2928 /* Can't remove network from underneath LNET for now, so I have
2932 case RDMA_CM_EVENT_ADDR_CHANGE:
2933 LCONSOLE_INFO("Physical link changed (eg hca/port)\n");
2939 kiblnd_check_txs_locked(kib_conn_t *conn, struct list_head *txs)
2942 struct list_head *ttmp;
2944 list_for_each(ttmp, txs) {
2945 tx = list_entry(ttmp, kib_tx_t, tx_list);
2947 if (txs != &conn->ibc_active_txs) {
2948 LASSERT(tx->tx_queued);
2950 LASSERT(!tx->tx_queued);
2951 LASSERT(tx->tx_waiting || tx->tx_sending != 0);
2954 if (cfs_time_aftereq(jiffies, tx->tx_deadline)) {
2955 CERROR("Timed out tx: %s, %lu seconds\n",
2956 kiblnd_queue2str(conn, txs),
2957 cfs_duration_sec(jiffies - tx->tx_deadline));
2966 kiblnd_conn_timed_out_locked(kib_conn_t *conn)
2968 return kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue) ||
2969 kiblnd_check_txs_locked(conn, &conn->ibc_tx_noops) ||
2970 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_rsrvd) ||
2971 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_nocred) ||
2972 kiblnd_check_txs_locked(conn, &conn->ibc_active_txs);
2976 kiblnd_check_conns(int idx)
2979 LIST_HEAD(checksends);
2980 struct list_head *peers = &kiblnd_data.kib_peers[idx];
2981 struct list_head *ptmp;
2985 struct list_head *ctmp;
2986 unsigned long flags;
2988 /* NB. We expect to have a look at all the peers and not find any
2989 * RDMAs to time out, so we just use a shared lock while we
2991 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2993 list_for_each(ptmp, peers) {
2994 peer = list_entry(ptmp, kib_peer_t, ibp_list);
2996 list_for_each(ctmp, &peer->ibp_conns) {
3000 conn = list_entry(ctmp, kib_conn_t, ibc_list);
3002 LASSERT(conn->ibc_state == IBLND_CONN_ESTABLISHED);
3004 spin_lock(&conn->ibc_lock);
3006 sendnoop = kiblnd_need_noop(conn);
3007 timedout = kiblnd_conn_timed_out_locked(conn);
3008 if (!sendnoop && !timedout) {
3009 spin_unlock(&conn->ibc_lock);
3014 CERROR("Timed out RDMA with %s (%lu): c: %u, oc: %u, rc: %u\n",
3015 libcfs_nid2str(peer->ibp_nid),
3016 cfs_duration_sec(cfs_time_current() -
3017 peer->ibp_last_alive),
3019 conn->ibc_outstanding_credits,
3020 conn->ibc_reserved_credits);
3021 list_add(&conn->ibc_connd_list, &closes);
3023 list_add(&conn->ibc_connd_list,
3026 /* +ref for 'closes' or 'checksends' */
3027 kiblnd_conn_addref(conn);
3029 spin_unlock(&conn->ibc_lock);
3033 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3035 /* Handle timeout by closing the whole
3036 * connection. We can only be sure RDMA activity
3037 * has ceased once the QP has been modified. */
3038 list_for_each_entry_safe(conn, tmp, &closes, ibc_connd_list) {
3039 list_del(&conn->ibc_connd_list);
3040 kiblnd_close_conn(conn, -ETIMEDOUT);
3041 kiblnd_conn_decref(conn);
3044 /* In case we have enough credits to return via a
3045 * NOOP, but there were no non-blocking tx descs
3046 * free to do it last time... */
3047 while (!list_empty(&checksends)) {
3048 conn = list_entry(checksends.next,
3049 kib_conn_t, ibc_connd_list);
3050 list_del(&conn->ibc_connd_list);
3051 kiblnd_check_sends(conn);
3052 kiblnd_conn_decref(conn);
3057 kiblnd_disconnect_conn(kib_conn_t *conn)
3059 LASSERT(!in_interrupt());
3060 LASSERT(current == kiblnd_data.kib_connd);
3061 LASSERT(conn->ibc_state == IBLND_CONN_CLOSING);
3063 rdma_disconnect(conn->ibc_cmid);
3064 kiblnd_finalise_conn(conn);
3066 kiblnd_peer_notify(conn->ibc_peer);
3070 kiblnd_connd(void *arg)
3073 unsigned long flags;
3079 unsigned long deadline = jiffies;
3081 cfs_block_allsigs();
3083 init_waitqueue_entry(&wait, current);
3084 kiblnd_data.kib_connd = current;
3086 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3088 while (!kiblnd_data.kib_shutdown) {
3092 if (!list_empty(&kiblnd_data.kib_connd_zombies)) {
3093 conn = list_entry(kiblnd_data.kib_connd_zombies.next,
3094 kib_conn_t, ibc_list);
3095 list_del(&conn->ibc_list);
3097 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock,
3101 kiblnd_destroy_conn(conn);
3103 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3106 if (!list_empty(&kiblnd_data.kib_connd_conns)) {
3107 conn = list_entry(kiblnd_data.kib_connd_conns.next,
3108 kib_conn_t, ibc_list);
3109 list_del(&conn->ibc_list);
3111 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock,
3115 kiblnd_disconnect_conn(conn);
3116 kiblnd_conn_decref(conn);
3118 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3121 /* careful with the jiffy wrap... */
3122 timeout = (int)(deadline - jiffies);
3126 int chunk = kiblnd_data.kib_peer_hash_size;
3128 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
3131 /* Time to check for RDMA timeouts on a few more
3132 * peers: I do checks every 'p' seconds on a
3133 * proportion of the peer table and I need to check
3134 * every connection 'n' times within a timeout
3135 * interval, to ensure I detect a timeout on any
3136 * connection within (n+1)/n times the timeout
3139 if (*kiblnd_tunables.kib_timeout > n * p)
3140 chunk = (chunk * n * p) /
3141 *kiblnd_tunables.kib_timeout;
3145 for (i = 0; i < chunk; i++) {
3146 kiblnd_check_conns(peer_index);
3147 peer_index = (peer_index + 1) %
3148 kiblnd_data.kib_peer_hash_size;
3152 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3158 /* Nothing to do for 'timeout' */
3159 set_current_state(TASK_INTERRUPTIBLE);
3160 add_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3161 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
3163 schedule_timeout(timeout);
3165 remove_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3166 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3169 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
3171 kiblnd_thread_fini();
3176 kiblnd_qp_event(struct ib_event *event, void *arg)
3178 kib_conn_t *conn = arg;
3180 switch (event->event) {
3181 case IB_EVENT_COMM_EST:
3182 CDEBUG(D_NET, "%s established\n",
3183 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3187 CERROR("%s: Async QP event type %d\n",
3188 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3194 kiblnd_complete(struct ib_wc *wc)
3196 switch (kiblnd_wreqid2type(wc->wr_id)) {
3200 case IBLND_WID_RDMA:
3201 /* We only get RDMA completion notification if it fails. All
3202 * subsequent work items, including the final SEND will fail
3203 * too. However we can't print out any more info about the
3204 * failing RDMA because 'tx' might be back on the idle list or
3205 * even reused already if we didn't manage to post all our work
3207 CNETERR("RDMA (tx: %p) failed: %d\n",
3208 kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3212 kiblnd_tx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3216 kiblnd_rx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status,
3223 kiblnd_cq_completion(struct ib_cq *cq, void *arg)
3225 /* NB I'm not allowed to schedule this conn once its refcount has
3226 * reached 0. Since fundamentally I'm racing with scheduler threads
3227 * consuming my CQ I could be called after all completions have
3228 * occurred. But in this case, ibc_nrx == 0 && ibc_nsends_posted == 0
3229 * and this CQ is about to be destroyed so I NOOP. */
3230 kib_conn_t *conn = (kib_conn_t *)arg;
3231 struct kib_sched_info *sched = conn->ibc_sched;
3232 unsigned long flags;
3234 LASSERT(cq == conn->ibc_cq);
3236 spin_lock_irqsave(&sched->ibs_lock, flags);
3238 conn->ibc_ready = 1;
3240 if (!conn->ibc_scheduled &&
3241 (conn->ibc_nrx > 0 ||
3242 conn->ibc_nsends_posted > 0)) {
3243 kiblnd_conn_addref(conn); /* +1 ref for sched_conns */
3244 conn->ibc_scheduled = 1;
3245 list_add_tail(&conn->ibc_sched_list, &sched->ibs_conns);
3247 if (waitqueue_active(&sched->ibs_waitq))
3248 wake_up(&sched->ibs_waitq);
3251 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3255 kiblnd_cq_event(struct ib_event *event, void *arg)
3257 kib_conn_t *conn = arg;
3259 CERROR("%s: async CQ event type %d\n",
3260 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3264 kiblnd_scheduler(void *arg)
3266 long id = (long)arg;
3267 struct kib_sched_info *sched;
3270 unsigned long flags;
3276 cfs_block_allsigs();
3278 init_waitqueue_entry(&wait, current);
3280 sched = kiblnd_data.kib_scheds[KIB_THREAD_CPT(id)];
3282 rc = cfs_cpt_bind(lnet_cpt_table(), sched->ibs_cpt);
3284 CWARN("Failed to bind on CPT %d, please verify whether all CPUs are healthy and reload modules if necessary, otherwise your system might under risk of low performance\n",
3288 spin_lock_irqsave(&sched->ibs_lock, flags);
3290 while (!kiblnd_data.kib_shutdown) {
3291 if (busy_loops++ >= IBLND_RESCHED) {
3292 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3297 spin_lock_irqsave(&sched->ibs_lock, flags);
3302 if (!list_empty(&sched->ibs_conns)) {
3303 conn = list_entry(sched->ibs_conns.next,
3304 kib_conn_t, ibc_sched_list);
3305 /* take over kib_sched_conns' ref on conn... */
3306 LASSERT(conn->ibc_scheduled);
3307 list_del(&conn->ibc_sched_list);
3308 conn->ibc_ready = 0;
3310 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3312 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3314 rc = ib_req_notify_cq(conn->ibc_cq,
3317 CWARN("%s: ib_req_notify_cq failed: %d, closing connection\n",
3318 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
3319 kiblnd_close_conn(conn, -EIO);
3320 kiblnd_conn_decref(conn);
3321 spin_lock_irqsave(&sched->ibs_lock,
3326 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3330 CWARN("%s: ib_poll_cq failed: %d, closing connection\n",
3331 libcfs_nid2str(conn->ibc_peer->ibp_nid),
3333 kiblnd_close_conn(conn, -EIO);
3334 kiblnd_conn_decref(conn);
3335 spin_lock_irqsave(&sched->ibs_lock, flags);
3339 spin_lock_irqsave(&sched->ibs_lock, flags);
3341 if (rc != 0 || conn->ibc_ready) {
3342 /* There may be another completion waiting; get
3343 * another scheduler to check while I handle
3345 /* +1 ref for sched_conns */
3346 kiblnd_conn_addref(conn);
3347 list_add_tail(&conn->ibc_sched_list,
3349 if (waitqueue_active(&sched->ibs_waitq))
3350 wake_up(&sched->ibs_waitq);
3352 conn->ibc_scheduled = 0;
3356 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3357 kiblnd_complete(&wc);
3359 spin_lock_irqsave(&sched->ibs_lock, flags);
3362 kiblnd_conn_decref(conn); /* ...drop my ref from above */
3369 set_current_state(TASK_INTERRUPTIBLE);
3370 add_wait_queue_exclusive(&sched->ibs_waitq, &wait);
3371 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3376 remove_wait_queue(&sched->ibs_waitq, &wait);
3377 spin_lock_irqsave(&sched->ibs_lock, flags);
3380 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3382 kiblnd_thread_fini();
3387 kiblnd_failover_thread(void *arg)
3389 rwlock_t *glock = &kiblnd_data.kib_global_lock;
3392 unsigned long flags;
3395 LASSERT(*kiblnd_tunables.kib_dev_failover != 0);
3397 cfs_block_allsigs();
3399 init_waitqueue_entry(&wait, current);
3400 write_lock_irqsave(glock, flags);
3402 while (!kiblnd_data.kib_shutdown) {
3403 int do_failover = 0;
3406 list_for_each_entry(dev, &kiblnd_data.kib_failed_devs,
3408 if (time_before(cfs_time_current(),
3409 dev->ibd_next_failover))
3416 list_del_init(&dev->ibd_fail_list);
3417 dev->ibd_failover = 1;
3418 write_unlock_irqrestore(glock, flags);
3420 rc = kiblnd_dev_failover(dev);
3422 write_lock_irqsave(glock, flags);
3424 LASSERT(dev->ibd_failover);
3425 dev->ibd_failover = 0;
3426 if (rc >= 0) { /* Device is OK or failover succeed */
3427 dev->ibd_next_failover = cfs_time_shift(3);
3431 /* failed to failover, retry later */
3432 dev->ibd_next_failover =
3433 cfs_time_shift(min(dev->ibd_failed_failover, 10));
3434 if (kiblnd_dev_can_failover(dev)) {
3435 list_add_tail(&dev->ibd_fail_list,
3436 &kiblnd_data.kib_failed_devs);
3442 /* long sleep if no more pending failover */
3443 long_sleep = list_empty(&kiblnd_data.kib_failed_devs);
3445 set_current_state(TASK_INTERRUPTIBLE);
3446 add_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3447 write_unlock_irqrestore(glock, flags);
3449 rc = schedule_timeout(long_sleep ? cfs_time_seconds(10) :
3450 cfs_time_seconds(1));
3451 remove_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3452 write_lock_irqsave(glock, flags);
3454 if (!long_sleep || rc != 0)
3457 /* have a long sleep, routine check all active devices,
3458 * we need checking like this because if there is not active
3459 * connection on the dev and no SEND from local, we may listen
3460 * on wrong HCA for ever while there is a bonding failover */
3461 list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3462 if (kiblnd_dev_can_failover(dev)) {
3463 list_add_tail(&dev->ibd_fail_list,
3464 &kiblnd_data.kib_failed_devs);
3469 write_unlock_irqrestore(glock, flags);
3471 kiblnd_thread_fini();