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, 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_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)
215 struct list_head *tmp;
217 list_for_each(tmp, &conn->ibc_active_txs) {
218 kib_tx_t *tx = 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 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 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 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 (is_vmalloc_addr((void *)vaddr)) {
533 page = vmalloc_to_page ((void *)vaddr);
534 LASSERT (page != NULL);
537 #ifdef CONFIG_HIGHMEM
538 if (vaddr >= PKMAP_BASE &&
539 vaddr < (PKMAP_BASE + LAST_PKMAP * PAGE_SIZE)) {
540 /* No highmem pages only used for bulk (kiov) I/O */
541 CERROR("find page for address in highmem\n");
545 page = virt_to_page (vaddr);
546 LASSERT (page != NULL);
551 kiblnd_fmr_map_tx(kib_net_t *net, kib_tx_t *tx, kib_rdma_desc_t *rd, int nob)
554 __u64 *pages = tx->tx_pages;
555 kib_fmr_poolset_t *fps;
562 LASSERT(tx->tx_pool != NULL);
563 LASSERT(tx->tx_pool->tpo_pool.po_owner != NULL);
565 hdev = tx->tx_pool->tpo_hdev;
567 for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
568 for (size = 0; size < rd->rd_frags[i].rf_nob;
569 size += hdev->ibh_page_size) {
570 pages[npages ++] = (rd->rd_frags[i].rf_addr &
571 hdev->ibh_page_mask) + size;
575 cpt = tx->tx_pool->tpo_pool.po_owner->ps_cpt;
577 fps = net->ibn_fmr_ps[cpt];
578 rc = kiblnd_fmr_pool_map(fps, pages, npages, 0, &tx->tx_u.fmr);
580 CERROR ("Can't map %d pages: %d\n", npages, rc);
584 /* If rd is not tx_rd, it's going to get sent to a peer, who will need
586 rd->rd_key = (rd != tx->tx_rd) ? tx->tx_u.fmr.fmr_pfmr->fmr->rkey :
587 tx->tx_u.fmr.fmr_pfmr->fmr->lkey;
588 rd->rd_frags[0].rf_addr &= ~hdev->ibh_page_mask;
589 rd->rd_frags[0].rf_nob = nob;
596 kiblnd_pmr_map_tx(kib_net_t *net, kib_tx_t *tx, kib_rdma_desc_t *rd, int nob)
599 kib_pmr_poolset_t *pps;
604 LASSERT(tx->tx_pool != NULL);
605 LASSERT(tx->tx_pool->tpo_pool.po_owner != NULL);
607 hdev = tx->tx_pool->tpo_hdev;
609 iova = rd->rd_frags[0].rf_addr & ~hdev->ibh_page_mask;
611 cpt = tx->tx_pool->tpo_pool.po_owner->ps_cpt;
613 pps = net->ibn_pmr_ps[cpt];
614 rc = kiblnd_pmr_pool_map(pps, hdev, rd, &iova, &tx->tx_u.pmr);
616 CERROR("Failed to create MR by phybuf: %d\n", rc);
620 /* If rd is not tx_rd, it's going to get sent to a peer, who will need
622 rd->rd_key = (rd != tx->tx_rd) ? tx->tx_u.pmr->pmr_mr->rkey :
623 tx->tx_u.pmr->pmr_mr->lkey;
625 rd->rd_frags[0].rf_addr = iova;
626 rd->rd_frags[0].rf_nob = nob;
632 kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx)
634 kib_net_t *net = ni->ni_data;
636 LASSERT(net != NULL);
638 if (net->ibn_fmr_ps != NULL && tx->tx_u.fmr.fmr_pfmr != NULL) {
639 kiblnd_fmr_pool_unmap(&tx->tx_u.fmr, tx->tx_status);
640 tx->tx_u.fmr.fmr_pfmr = NULL;
642 } else if (net->ibn_pmr_ps != NULL && tx->tx_u.pmr != NULL) {
643 kiblnd_pmr_pool_unmap(tx->tx_u.pmr);
647 if (tx->tx_nfrags != 0) {
648 kiblnd_dma_unmap_sg(tx->tx_pool->tpo_hdev->ibh_ibdev,
649 tx->tx_frags, tx->tx_nfrags, tx->tx_dmadir);
655 kiblnd_map_tx(lnet_ni_t *ni, kib_tx_t *tx,
656 kib_rdma_desc_t *rd, int nfrags)
658 kib_hca_dev_t *hdev = tx->tx_pool->tpo_hdev;
659 kib_net_t *net = ni->ni_data;
660 struct ib_mr *mr = NULL;
664 /* If rd is not tx_rd, it's going to get sent to a peer and I'm the
666 tx->tx_dmadir = (rd != tx->tx_rd) ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
667 tx->tx_nfrags = nfrags;
670 kiblnd_dma_map_sg(hdev->ibh_ibdev,
671 tx->tx_frags, tx->tx_nfrags, tx->tx_dmadir);
673 for (i = 0, nob = 0; i < rd->rd_nfrags; i++) {
674 rd->rd_frags[i].rf_nob = kiblnd_sg_dma_len(
675 hdev->ibh_ibdev, &tx->tx_frags[i]);
676 rd->rd_frags[i].rf_addr = kiblnd_sg_dma_address(
677 hdev->ibh_ibdev, &tx->tx_frags[i]);
678 nob += rd->rd_frags[i].rf_nob;
681 /* looking for pre-mapping MR */
682 mr = kiblnd_find_rd_dma_mr(hdev, rd);
684 /* found pre-mapping MR */
685 rd->rd_key = (rd != tx->tx_rd) ? mr->rkey : mr->lkey;
689 if (net->ibn_fmr_ps != NULL)
690 return kiblnd_fmr_map_tx(net, tx, rd, nob);
691 else if (net->ibn_pmr_ps != NULL)
692 return kiblnd_pmr_map_tx(net, tx, rd, nob);
699 kiblnd_setup_rd_iov(lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd,
700 unsigned int niov, struct iovec *iov, int offset, int nob)
702 kib_net_t *net = ni->ni_data;
704 struct scatterlist *sg;
711 LASSERT (net != NULL);
713 while (offset >= iov->iov_len) {
714 offset -= iov->iov_len;
724 vaddr = ((unsigned long)iov->iov_base) + offset;
725 page_offset = vaddr & (PAGE_SIZE - 1);
726 page = kiblnd_kvaddr_to_page(vaddr);
728 CERROR ("Can't find page\n");
732 fragnob = min((int)(iov->iov_len - offset), nob);
733 fragnob = min(fragnob, (int)PAGE_SIZE - page_offset);
735 sg_set_page(sg, page, fragnob, page_offset);
738 if (offset + fragnob < iov->iov_len) {
748 return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags);
752 kiblnd_setup_rd_kiov (lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd,
753 int nkiov, lnet_kiov_t *kiov, int offset, int nob)
755 kib_net_t *net = ni->ni_data;
756 struct scatterlist *sg;
759 CDEBUG(D_NET, "niov %d offset %d nob %d\n", nkiov, offset, nob);
763 LASSERT (net != NULL);
765 while (offset >= kiov->kiov_len) {
766 offset -= kiov->kiov_len;
776 fragnob = min((int)(kiov->kiov_len - offset), nob);
778 sg_set_page(sg, kiov->kiov_page, fragnob,
779 kiov->kiov_offset + offset);
788 return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags);
792 kiblnd_post_tx_locked (kib_conn_t *conn, kib_tx_t *tx, int credit)
794 kib_msg_t *msg = tx->tx_msg;
795 kib_peer_t *peer = conn->ibc_peer;
796 int ver = conn->ibc_version;
799 struct ib_send_wr *bad_wrq;
801 LASSERT (tx->tx_queued);
802 /* We rely on this for QP sizing */
803 LASSERT (tx->tx_nwrq > 0);
804 LASSERT (tx->tx_nwrq <= 1 + IBLND_RDMA_FRAGS(ver));
806 LASSERT (credit == 0 || credit == 1);
807 LASSERT (conn->ibc_outstanding_credits >= 0);
808 LASSERT (conn->ibc_outstanding_credits <= IBLND_MSG_QUEUE_SIZE(ver));
809 LASSERT (conn->ibc_credits >= 0);
810 LASSERT (conn->ibc_credits <= IBLND_MSG_QUEUE_SIZE(ver));
812 if (conn->ibc_nsends_posted == IBLND_CONCURRENT_SENDS(ver)) {
813 /* tx completions outstanding... */
814 CDEBUG(D_NET, "%s: posted enough\n",
815 libcfs_nid2str(peer->ibp_nid));
819 if (credit != 0 && conn->ibc_credits == 0) { /* no credits */
820 CDEBUG(D_NET, "%s: no credits\n",
821 libcfs_nid2str(peer->ibp_nid));
825 if (credit != 0 && !IBLND_OOB_CAPABLE(ver) &&
826 conn->ibc_credits == 1 && /* last credit reserved */
827 msg->ibm_type != IBLND_MSG_NOOP) { /* for NOOP */
828 CDEBUG(D_NET, "%s: not using last credit\n",
829 libcfs_nid2str(peer->ibp_nid));
833 /* NB don't drop ibc_lock before bumping tx_sending */
834 list_del(&tx->tx_list);
837 if (msg->ibm_type == IBLND_MSG_NOOP &&
838 (!kiblnd_need_noop(conn) || /* redundant NOOP */
839 (IBLND_OOB_CAPABLE(ver) && /* posted enough NOOP */
840 conn->ibc_noops_posted == IBLND_OOB_MSGS(ver)))) {
841 /* OK to drop when posted enough NOOPs, since
842 * kiblnd_check_sends will queue NOOP again when
843 * posted NOOPs complete */
844 spin_unlock(&conn->ibc_lock);
845 kiblnd_tx_done(peer->ibp_ni, tx);
846 spin_lock(&conn->ibc_lock);
847 CDEBUG(D_NET, "%s(%d): redundant or enough NOOP\n",
848 libcfs_nid2str(peer->ibp_nid),
849 conn->ibc_noops_posted);
853 kiblnd_pack_msg(peer->ibp_ni, msg, ver, conn->ibc_outstanding_credits,
854 peer->ibp_nid, conn->ibc_incarnation);
856 conn->ibc_credits -= credit;
857 conn->ibc_outstanding_credits = 0;
858 conn->ibc_nsends_posted++;
859 if (msg->ibm_type == IBLND_MSG_NOOP)
860 conn->ibc_noops_posted++;
862 /* CAVEAT EMPTOR! This tx could be the PUT_DONE of an RDMA
863 * PUT. If so, it was first queued here as a PUT_REQ, sent and
864 * stashed on ibc_active_txs, matched by an incoming PUT_ACK,
865 * and then re-queued here. It's (just) possible that
866 * tx_sending is non-zero if we've not done the tx_complete()
867 * from the first send; hence the ++ rather than = below. */
869 list_add(&tx->tx_list, &conn->ibc_active_txs);
871 /* I'm still holding ibc_lock! */
872 if (conn->ibc_state != IBLND_CONN_ESTABLISHED) {
874 } else if (tx->tx_pool->tpo_pool.po_failed ||
875 conn->ibc_hdev != tx->tx_pool->tpo_hdev) {
876 /* close_conn will launch failover */
879 rc = ib_post_send(conn->ibc_cmid->qp,
880 tx->tx_wrq, &bad_wrq);
883 conn->ibc_last_send = jiffies;
888 /* NB credits are transferred in the actual
889 * message, which can only be the last work item */
890 conn->ibc_credits += credit;
891 conn->ibc_outstanding_credits += msg->ibm_credits;
892 conn->ibc_nsends_posted--;
893 if (msg->ibm_type == IBLND_MSG_NOOP)
894 conn->ibc_noops_posted--;
900 done = (tx->tx_sending == 0);
902 list_del(&tx->tx_list);
904 spin_unlock(&conn->ibc_lock);
906 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
907 CERROR("Error %d posting transmit to %s\n",
908 rc, libcfs_nid2str(peer->ibp_nid));
910 CDEBUG(D_NET, "Error %d posting transmit to %s\n",
911 rc, libcfs_nid2str(peer->ibp_nid));
913 kiblnd_close_conn(conn, rc);
916 kiblnd_tx_done(peer->ibp_ni, tx);
918 spin_lock(&conn->ibc_lock);
924 kiblnd_check_sends (kib_conn_t *conn)
926 int ver = conn->ibc_version;
927 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
930 /* Don't send anything until after the connection is established */
931 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
932 CDEBUG(D_NET, "%s too soon\n",
933 libcfs_nid2str(conn->ibc_peer->ibp_nid));
937 spin_lock(&conn->ibc_lock);
939 LASSERT (conn->ibc_nsends_posted <= IBLND_CONCURRENT_SENDS(ver));
940 LASSERT (!IBLND_OOB_CAPABLE(ver) ||
941 conn->ibc_noops_posted <= IBLND_OOB_MSGS(ver));
942 LASSERT (conn->ibc_reserved_credits >= 0);
944 while (conn->ibc_reserved_credits > 0 &&
945 !list_empty(&conn->ibc_tx_queue_rsrvd)) {
946 tx = list_entry(conn->ibc_tx_queue_rsrvd.next,
948 list_del(&tx->tx_list);
949 list_add_tail(&tx->tx_list, &conn->ibc_tx_queue);
950 conn->ibc_reserved_credits--;
953 if (kiblnd_need_noop(conn)) {
954 spin_unlock(&conn->ibc_lock);
956 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
958 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_NOOP, 0);
960 spin_lock(&conn->ibc_lock);
962 kiblnd_queue_tx_locked(tx, conn);
965 kiblnd_conn_addref(conn); /* 1 ref for me.... (see b21911) */
970 if (!list_empty(&conn->ibc_tx_queue_nocred)) {
972 tx = list_entry(conn->ibc_tx_queue_nocred.next,
974 } else if (!list_empty(&conn->ibc_tx_noops)) {
975 LASSERT (!IBLND_OOB_CAPABLE(ver));
977 tx = list_entry(conn->ibc_tx_noops.next,
979 } else if (!list_empty(&conn->ibc_tx_queue)) {
981 tx = list_entry(conn->ibc_tx_queue.next,
986 if (kiblnd_post_tx_locked(conn, tx, credit) != 0)
990 spin_unlock(&conn->ibc_lock);
992 kiblnd_conn_decref(conn); /* ...until here */
996 kiblnd_tx_complete (kib_tx_t *tx, int status)
998 int failed = (status != IB_WC_SUCCESS);
999 kib_conn_t *conn = tx->tx_conn;
1002 LASSERT (tx->tx_sending > 0);
1005 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
1006 CNETERR("Tx -> %s cookie "LPX64
1007 " sending %d waiting %d: failed %d\n",
1008 libcfs_nid2str(conn->ibc_peer->ibp_nid),
1009 tx->tx_cookie, tx->tx_sending, tx->tx_waiting,
1012 kiblnd_close_conn(conn, -EIO);
1014 kiblnd_peer_alive(conn->ibc_peer);
1017 spin_lock(&conn->ibc_lock);
1019 /* I could be racing with rdma completion. Whoever makes 'tx' idle
1020 * gets to free it, which also drops its ref on 'conn'. */
1023 conn->ibc_nsends_posted--;
1024 if (tx->tx_msg->ibm_type == IBLND_MSG_NOOP)
1025 conn->ibc_noops_posted--;
1028 tx->tx_waiting = 0; /* don't wait for peer */
1029 tx->tx_status = -EIO;
1032 idle = (tx->tx_sending == 0) && /* This is the final callback */
1033 !tx->tx_waiting && /* Not waiting for peer */
1034 !tx->tx_queued; /* Not re-queued (PUT_DONE) */
1036 list_del(&tx->tx_list);
1038 kiblnd_conn_addref(conn); /* 1 ref for me.... */
1040 spin_unlock(&conn->ibc_lock);
1043 kiblnd_tx_done(conn->ibc_peer->ibp_ni, tx);
1045 kiblnd_check_sends(conn);
1047 kiblnd_conn_decref(conn); /* ...until here */
1051 kiblnd_init_tx_msg (lnet_ni_t *ni, kib_tx_t *tx, int type, int body_nob)
1053 kib_hca_dev_t *hdev = tx->tx_pool->tpo_hdev;
1054 struct ib_sge *sge = &tx->tx_sge[tx->tx_nwrq];
1055 struct ib_send_wr *wrq = &tx->tx_wrq[tx->tx_nwrq];
1056 int nob = offsetof (kib_msg_t, ibm_u) + body_nob;
1059 LASSERT (tx->tx_nwrq >= 0);
1060 LASSERT (tx->tx_nwrq < IBLND_MAX_RDMA_FRAGS + 1);
1061 LASSERT (nob <= IBLND_MSG_SIZE);
1063 kiblnd_init_msg(tx->tx_msg, type, body_nob);
1065 mr = kiblnd_find_dma_mr(hdev, tx->tx_msgaddr, nob);
1066 LASSERT (mr != NULL);
1068 sge->lkey = mr->lkey;
1069 sge->addr = tx->tx_msgaddr;
1072 memset(wrq, 0, sizeof(*wrq));
1075 wrq->wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_TX);
1078 wrq->opcode = IB_WR_SEND;
1079 wrq->send_flags = IB_SEND_SIGNALED;
1085 kiblnd_init_rdma (kib_conn_t *conn, kib_tx_t *tx, int type,
1086 int resid, kib_rdma_desc_t *dstrd, __u64 dstcookie)
1088 kib_msg_t *ibmsg = tx->tx_msg;
1089 kib_rdma_desc_t *srcrd = tx->tx_rd;
1090 struct ib_sge *sge = &tx->tx_sge[0];
1091 struct ib_send_wr *wrq = &tx->tx_wrq[0];
1097 LASSERT (!in_interrupt());
1098 LASSERT (tx->tx_nwrq == 0);
1099 LASSERT (type == IBLND_MSG_GET_DONE ||
1100 type == IBLND_MSG_PUT_DONE);
1102 srcidx = dstidx = 0;
1105 if (srcidx >= srcrd->rd_nfrags) {
1106 CERROR("Src buffer exhausted: %d frags\n", srcidx);
1111 if (dstidx == dstrd->rd_nfrags) {
1112 CERROR("Dst buffer exhausted: %d frags\n", dstidx);
1117 if (tx->tx_nwrq == IBLND_RDMA_FRAGS(conn->ibc_version)) {
1118 CERROR("RDMA too fragmented for %s (%d): "
1119 "%d/%d src %d/%d dst frags\n",
1120 libcfs_nid2str(conn->ibc_peer->ibp_nid),
1121 IBLND_RDMA_FRAGS(conn->ibc_version),
1122 srcidx, srcrd->rd_nfrags,
1123 dstidx, dstrd->rd_nfrags);
1128 wrknob = MIN(MIN(kiblnd_rd_frag_size(srcrd, srcidx),
1129 kiblnd_rd_frag_size(dstrd, dstidx)), resid);
1131 sge = &tx->tx_sge[tx->tx_nwrq];
1132 sge->addr = kiblnd_rd_frag_addr(srcrd, srcidx);
1133 sge->lkey = kiblnd_rd_frag_key(srcrd, srcidx);
1134 sge->length = wrknob;
1136 wrq = &tx->tx_wrq[tx->tx_nwrq];
1138 wrq->next = wrq + 1;
1139 wrq->wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_RDMA);
1142 wrq->opcode = IB_WR_RDMA_WRITE;
1143 wrq->send_flags = 0;
1145 wrq->wr.rdma.remote_addr = kiblnd_rd_frag_addr(dstrd, dstidx);
1146 wrq->wr.rdma.rkey = kiblnd_rd_frag_key(dstrd, dstidx);
1148 srcidx = kiblnd_rd_consume_frag(srcrd, srcidx, wrknob);
1149 dstidx = kiblnd_rd_consume_frag(dstrd, dstidx, wrknob);
1158 if (rc < 0) /* no RDMA if completing with failure */
1161 ibmsg->ibm_u.completion.ibcm_status = rc;
1162 ibmsg->ibm_u.completion.ibcm_cookie = dstcookie;
1163 kiblnd_init_tx_msg(conn->ibc_peer->ibp_ni, tx,
1164 type, sizeof (kib_completion_msg_t));
1170 kiblnd_queue_tx_locked (kib_tx_t *tx, kib_conn_t *conn)
1172 struct list_head *q;
1174 LASSERT(tx->tx_nwrq > 0); /* work items set up */
1175 LASSERT(!tx->tx_queued); /* not queued for sending already */
1176 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1179 tx->tx_deadline = jiffies + (*kiblnd_tunables.kib_timeout * HZ);
1181 if (tx->tx_conn == NULL) {
1182 kiblnd_conn_addref(conn);
1184 LASSERT (tx->tx_msg->ibm_type != IBLND_MSG_PUT_DONE);
1186 /* PUT_DONE first attached to conn as a PUT_REQ */
1187 LASSERT (tx->tx_conn == conn);
1188 LASSERT (tx->tx_msg->ibm_type == IBLND_MSG_PUT_DONE);
1191 switch (tx->tx_msg->ibm_type) {
1195 case IBLND_MSG_PUT_REQ:
1196 case IBLND_MSG_GET_REQ:
1197 q = &conn->ibc_tx_queue_rsrvd;
1200 case IBLND_MSG_PUT_NAK:
1201 case IBLND_MSG_PUT_ACK:
1202 case IBLND_MSG_PUT_DONE:
1203 case IBLND_MSG_GET_DONE:
1204 q = &conn->ibc_tx_queue_nocred;
1207 case IBLND_MSG_NOOP:
1208 if (IBLND_OOB_CAPABLE(conn->ibc_version))
1209 q = &conn->ibc_tx_queue_nocred;
1211 q = &conn->ibc_tx_noops;
1214 case IBLND_MSG_IMMEDIATE:
1215 q = &conn->ibc_tx_queue;
1219 list_add_tail(&tx->tx_list, q);
1223 kiblnd_queue_tx (kib_tx_t *tx, kib_conn_t *conn)
1225 spin_lock(&conn->ibc_lock);
1226 kiblnd_queue_tx_locked(tx, conn);
1227 spin_unlock(&conn->ibc_lock);
1229 kiblnd_check_sends(conn);
1232 static int kiblnd_resolve_addr(struct rdma_cm_id *cmid,
1233 struct sockaddr_in *srcaddr,
1234 struct sockaddr_in *dstaddr,
1237 unsigned short port;
1240 #ifdef HAVE_OFED_RDMA_SET_REUSEADDR
1241 /* allow the port to be reused */
1242 rc = rdma_set_reuseaddr(cmid, 1);
1244 CERROR("Unable to set reuse on cmid: %d\n", rc);
1249 /* look for a free privileged port */
1250 for (port = PROT_SOCK-1; port > 0; port--) {
1251 srcaddr->sin_port = htons(port);
1252 rc = rdma_resolve_addr(cmid,
1253 (struct sockaddr *)srcaddr,
1254 (struct sockaddr *)dstaddr,
1257 CDEBUG(D_NET, "bound to port %hu\n", port);
1259 } else if (rc == -EADDRINUSE || rc == -EADDRNOTAVAIL) {
1260 CDEBUG(D_NET, "bind to port %hu failed: %d\n",
1267 CERROR("Failed to bind to a free privileged port\n");
1268 #ifndef HAVE_OFED_RDMA_SET_REUSEADDR
1269 CERROR("You may need IB verbs that supports rdma_set_reuseaddr()\n");
1275 kiblnd_connect_peer (kib_peer_t *peer)
1277 struct rdma_cm_id *cmid;
1279 kib_net_t *net = peer->ibp_ni->ni_data;
1280 struct sockaddr_in srcaddr;
1281 struct sockaddr_in dstaddr;
1284 LASSERT (net != NULL);
1285 LASSERT (peer->ibp_connecting > 0);
1287 cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, peer, RDMA_PS_TCP,
1291 CERROR("Can't create CMID for %s: %ld\n",
1292 libcfs_nid2str(peer->ibp_nid), PTR_ERR(cmid));
1298 memset(&srcaddr, 0, sizeof(srcaddr));
1299 srcaddr.sin_family = AF_INET;
1300 srcaddr.sin_addr.s_addr = htonl(dev->ibd_ifip);
1302 memset(&dstaddr, 0, sizeof(dstaddr));
1303 dstaddr.sin_family = AF_INET;
1304 dstaddr.sin_port = htons(*kiblnd_tunables.kib_service);
1305 dstaddr.sin_addr.s_addr = htonl(LNET_NIDADDR(peer->ibp_nid));
1307 kiblnd_peer_addref(peer); /* cmid's ref */
1309 if (*kiblnd_tunables.kib_use_priv_port) {
1310 rc = kiblnd_resolve_addr(cmid, &srcaddr, &dstaddr,
1311 *kiblnd_tunables.kib_timeout * 1000);
1313 rc = rdma_resolve_addr(cmid,
1314 (struct sockaddr *)&srcaddr,
1315 (struct sockaddr *)&dstaddr,
1316 *kiblnd_tunables.kib_timeout * 1000);
1319 /* Can't initiate address resolution: */
1320 CERROR("Can't resolve addr for %s: %d\n",
1321 libcfs_nid2str(peer->ibp_nid), rc);
1325 LASSERT (cmid->device != NULL);
1326 CDEBUG(D_NET, "%s: connection bound to %s:%u.%u.%u.%u:%s\n",
1327 libcfs_nid2str(peer->ibp_nid), dev->ibd_ifname,
1328 HIPQUAD(dev->ibd_ifip), cmid->device->name);
1333 kiblnd_peer_decref(peer); /* cmid's ref */
1334 rdma_destroy_id(cmid);
1336 kiblnd_peer_connect_failed(peer, 1, rc);
1340 kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid)
1345 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
1346 unsigned long flags;
1349 /* If I get here, I've committed to send, so I complete the tx with
1350 * failure on any problems */
1352 LASSERT (tx == NULL || tx->tx_conn == NULL); /* only set when assigned a conn */
1353 LASSERT (tx == NULL || tx->tx_nwrq > 0); /* work items have been set up */
1355 /* First time, just use a read lock since I expect to find my peer
1357 read_lock_irqsave(g_lock, flags);
1359 peer = kiblnd_find_peer_locked(nid);
1360 if (peer != NULL && !list_empty(&peer->ibp_conns)) {
1361 /* Found a peer with an established connection */
1362 conn = kiblnd_get_conn_locked(peer);
1363 kiblnd_conn_addref(conn); /* 1 ref for me... */
1365 read_unlock_irqrestore(g_lock, flags);
1368 kiblnd_queue_tx(tx, conn);
1369 kiblnd_conn_decref(conn); /* ...to here */
1373 read_unlock(g_lock);
1374 /* Re-try with a write lock */
1377 peer = kiblnd_find_peer_locked(nid);
1379 if (list_empty(&peer->ibp_conns)) {
1380 /* found a peer, but it's still connecting... */
1381 LASSERT (peer->ibp_connecting != 0 ||
1382 peer->ibp_accepting != 0);
1384 list_add_tail(&tx->tx_list,
1385 &peer->ibp_tx_queue);
1386 write_unlock_irqrestore(g_lock, flags);
1388 conn = kiblnd_get_conn_locked(peer);
1389 kiblnd_conn_addref(conn); /* 1 ref for me... */
1391 write_unlock_irqrestore(g_lock, flags);
1394 kiblnd_queue_tx(tx, conn);
1395 kiblnd_conn_decref(conn); /* ...to here */
1400 write_unlock_irqrestore(g_lock, flags);
1402 /* Allocate a peer ready to add to the peer table and retry */
1403 rc = kiblnd_create_peer(ni, &peer, nid);
1405 CERROR("Can't create peer %s\n", libcfs_nid2str(nid));
1407 tx->tx_status = -EHOSTUNREACH;
1409 kiblnd_tx_done(ni, tx);
1414 write_lock_irqsave(g_lock, flags);
1416 peer2 = kiblnd_find_peer_locked(nid);
1417 if (peer2 != NULL) {
1418 if (list_empty(&peer2->ibp_conns)) {
1419 /* found a peer, but it's still connecting... */
1420 LASSERT (peer2->ibp_connecting != 0 ||
1421 peer2->ibp_accepting != 0);
1423 list_add_tail(&tx->tx_list,
1424 &peer2->ibp_tx_queue);
1425 write_unlock_irqrestore(g_lock, flags);
1427 conn = kiblnd_get_conn_locked(peer2);
1428 kiblnd_conn_addref(conn); /* 1 ref for me... */
1430 write_unlock_irqrestore(g_lock, flags);
1433 kiblnd_queue_tx(tx, conn);
1434 kiblnd_conn_decref(conn); /* ...to here */
1437 kiblnd_peer_decref(peer);
1441 /* Brand new peer */
1442 LASSERT (peer->ibp_connecting == 0);
1443 peer->ibp_connecting = 1;
1445 /* always called with a ref on ni, which prevents ni being shutdown */
1446 LASSERT (((kib_net_t *)ni->ni_data)->ibn_shutdown == 0);
1449 list_add_tail(&tx->tx_list, &peer->ibp_tx_queue);
1451 kiblnd_peer_addref(peer);
1452 list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid));
1454 write_unlock_irqrestore(g_lock, flags);
1456 kiblnd_connect_peer(peer);
1457 kiblnd_peer_decref(peer);
1461 kiblnd_send (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg)
1463 lnet_hdr_t *hdr = &lntmsg->msg_hdr;
1464 int type = lntmsg->msg_type;
1465 lnet_process_id_t target = lntmsg->msg_target;
1466 int target_is_router = lntmsg->msg_target_is_router;
1467 int routing = lntmsg->msg_routing;
1468 unsigned int payload_niov = lntmsg->msg_niov;
1469 struct iovec *payload_iov = lntmsg->msg_iov;
1470 lnet_kiov_t *payload_kiov = lntmsg->msg_kiov;
1471 unsigned int payload_offset = lntmsg->msg_offset;
1472 unsigned int payload_nob = lntmsg->msg_len;
1478 /* NB 'private' is different depending on what we're sending.... */
1480 CDEBUG(D_NET, "sending %d bytes in %d frags to %s\n",
1481 payload_nob, payload_niov, libcfs_id2str(target));
1483 LASSERT (payload_nob == 0 || payload_niov > 0);
1484 LASSERT (payload_niov <= LNET_MAX_IOV);
1486 /* Thread context */
1487 LASSERT (!in_interrupt());
1488 /* payload is either all vaddrs or all pages */
1489 LASSERT (!(payload_kiov != NULL && payload_iov != NULL));
1497 LASSERT (payload_nob == 0);
1501 if (routing || target_is_router)
1502 break; /* send IMMEDIATE */
1504 /* is the REPLY message too small for RDMA? */
1505 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[lntmsg->msg_md->md_length]);
1506 if (nob <= IBLND_MSG_SIZE)
1507 break; /* send IMMEDIATE */
1509 tx = kiblnd_get_idle_tx(ni, target.nid);
1511 CERROR("Can't allocate txd for GET to %s\n",
1512 libcfs_nid2str(target.nid));
1518 if ((lntmsg->msg_md->md_options & LNET_MD_KIOV) == 0)
1519 rc = kiblnd_setup_rd_iov(ni, tx,
1520 &ibmsg->ibm_u.get.ibgm_rd,
1521 lntmsg->msg_md->md_niov,
1522 lntmsg->msg_md->md_iov.iov,
1523 0, lntmsg->msg_md->md_length);
1525 rc = kiblnd_setup_rd_kiov(ni, tx,
1526 &ibmsg->ibm_u.get.ibgm_rd,
1527 lntmsg->msg_md->md_niov,
1528 lntmsg->msg_md->md_iov.kiov,
1529 0, lntmsg->msg_md->md_length);
1531 CERROR("Can't setup GET sink for %s: %d\n",
1532 libcfs_nid2str(target.nid), rc);
1533 kiblnd_tx_done(ni, tx);
1537 nob = offsetof(kib_get_msg_t, ibgm_rd.rd_frags[tx->tx_nfrags]);
1538 ibmsg->ibm_u.get.ibgm_cookie = tx->tx_cookie;
1539 ibmsg->ibm_u.get.ibgm_hdr = *hdr;
1541 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_GET_REQ, nob);
1543 tx->tx_lntmsg[1] = lnet_create_reply_msg(ni, lntmsg);
1544 if (tx->tx_lntmsg[1] == NULL) {
1545 CERROR("Can't create reply for GET -> %s\n",
1546 libcfs_nid2str(target.nid));
1547 kiblnd_tx_done(ni, tx);
1551 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg[0,1] on completion */
1552 tx->tx_waiting = 1; /* waiting for GET_DONE */
1553 kiblnd_launch_tx(ni, tx, target.nid);
1556 case LNET_MSG_REPLY:
1558 /* Is the payload small enough not to need RDMA? */
1559 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob]);
1560 if (nob <= IBLND_MSG_SIZE)
1561 break; /* send IMMEDIATE */
1563 tx = kiblnd_get_idle_tx(ni, target.nid);
1565 CERROR("Can't allocate %s txd for %s\n",
1566 type == LNET_MSG_PUT ? "PUT" : "REPLY",
1567 libcfs_nid2str(target.nid));
1571 if (payload_kiov == NULL)
1572 rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1573 payload_niov, payload_iov,
1574 payload_offset, payload_nob);
1576 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1577 payload_niov, payload_kiov,
1578 payload_offset, payload_nob);
1580 CERROR("Can't setup PUT src for %s: %d\n",
1581 libcfs_nid2str(target.nid), rc);
1582 kiblnd_tx_done(ni, tx);
1587 ibmsg->ibm_u.putreq.ibprm_hdr = *hdr;
1588 ibmsg->ibm_u.putreq.ibprm_cookie = tx->tx_cookie;
1589 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_REQ, sizeof(kib_putreq_msg_t));
1591 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1592 tx->tx_waiting = 1; /* waiting for PUT_{ACK,NAK} */
1593 kiblnd_launch_tx(ni, tx, target.nid);
1597 /* send IMMEDIATE */
1599 LASSERT (offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob])
1602 tx = kiblnd_get_idle_tx(ni, target.nid);
1604 CERROR ("Can't send %d to %s: tx descs exhausted\n",
1605 type, libcfs_nid2str(target.nid));
1610 ibmsg->ibm_u.immediate.ibim_hdr = *hdr;
1612 if (payload_kiov != NULL)
1613 lnet_copy_kiov2flat(IBLND_MSG_SIZE, ibmsg,
1614 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1615 payload_niov, payload_kiov,
1616 payload_offset, payload_nob);
1618 lnet_copy_iov2flat(IBLND_MSG_SIZE, ibmsg,
1619 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1620 payload_niov, payload_iov,
1621 payload_offset, payload_nob);
1623 nob = offsetof(kib_immediate_msg_t, ibim_payload[payload_nob]);
1624 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_IMMEDIATE, nob);
1626 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1627 kiblnd_launch_tx(ni, tx, target.nid);
1632 kiblnd_reply (lnet_ni_t *ni, kib_rx_t *rx, lnet_msg_t *lntmsg)
1634 lnet_process_id_t target = lntmsg->msg_target;
1635 unsigned int niov = lntmsg->msg_niov;
1636 struct iovec *iov = lntmsg->msg_iov;
1637 lnet_kiov_t *kiov = lntmsg->msg_kiov;
1638 unsigned int offset = lntmsg->msg_offset;
1639 unsigned int nob = lntmsg->msg_len;
1643 tx = kiblnd_get_idle_tx(ni, rx->rx_conn->ibc_peer->ibp_nid);
1645 CERROR("Can't get tx for REPLY to %s\n",
1646 libcfs_nid2str(target.nid));
1652 else if (kiov == NULL)
1653 rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1654 niov, iov, offset, nob);
1656 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1657 niov, kiov, offset, nob);
1660 CERROR("Can't setup GET src for %s: %d\n",
1661 libcfs_nid2str(target.nid), rc);
1665 rc = kiblnd_init_rdma(rx->rx_conn, tx,
1666 IBLND_MSG_GET_DONE, nob,
1667 &rx->rx_msg->ibm_u.get.ibgm_rd,
1668 rx->rx_msg->ibm_u.get.ibgm_cookie);
1670 CERROR("Can't setup rdma for GET from %s: %d\n",
1671 libcfs_nid2str(target.nid), rc);
1676 /* No RDMA: local completion may happen now! */
1677 lnet_finalize(ni, lntmsg, 0);
1679 /* RDMA: lnet_finalize(lntmsg) when it
1681 tx->tx_lntmsg[0] = lntmsg;
1684 kiblnd_queue_tx(tx, rx->rx_conn);
1688 kiblnd_tx_done(ni, tx);
1690 lnet_finalize(ni, lntmsg, -EIO);
1694 kiblnd_recv (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed,
1695 unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov,
1696 unsigned int offset, unsigned int mlen, unsigned int rlen)
1698 kib_rx_t *rx = private;
1699 kib_msg_t *rxmsg = rx->rx_msg;
1700 kib_conn_t *conn = rx->rx_conn;
1704 int post_credit = IBLND_POSTRX_PEER_CREDIT;
1707 LASSERT (mlen <= rlen);
1708 LASSERT (!in_interrupt());
1709 /* Either all pages or all vaddrs */
1710 LASSERT (!(kiov != NULL && iov != NULL));
1712 switch (rxmsg->ibm_type) {
1716 case IBLND_MSG_IMMEDIATE:
1717 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[rlen]);
1718 if (nob > rx->rx_nob) {
1719 CERROR ("Immediate message from %s too big: %d(%d)\n",
1720 libcfs_nid2str(rxmsg->ibm_u.immediate.ibim_hdr.src_nid),
1727 lnet_copy_flat2kiov(niov, kiov, offset,
1728 IBLND_MSG_SIZE, rxmsg,
1729 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1732 lnet_copy_flat2iov(niov, iov, offset,
1733 IBLND_MSG_SIZE, rxmsg,
1734 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1736 lnet_finalize (ni, lntmsg, 0);
1739 case IBLND_MSG_PUT_REQ:
1741 lnet_finalize(ni, lntmsg, 0);
1742 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK, 0,
1743 rxmsg->ibm_u.putreq.ibprm_cookie);
1747 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
1749 CERROR("Can't allocate tx for %s\n",
1750 libcfs_nid2str(conn->ibc_peer->ibp_nid));
1751 /* Not replying will break the connection */
1758 rc = kiblnd_setup_rd_iov(ni, tx,
1759 &txmsg->ibm_u.putack.ibpam_rd,
1760 niov, iov, offset, mlen);
1762 rc = kiblnd_setup_rd_kiov(ni, tx,
1763 &txmsg->ibm_u.putack.ibpam_rd,
1764 niov, kiov, offset, mlen);
1766 CERROR("Can't setup PUT sink for %s: %d\n",
1767 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
1768 kiblnd_tx_done(ni, tx);
1769 /* tell peer it's over */
1770 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK, rc,
1771 rxmsg->ibm_u.putreq.ibprm_cookie);
1775 nob = offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[tx->tx_nfrags]);
1776 txmsg->ibm_u.putack.ibpam_src_cookie = rxmsg->ibm_u.putreq.ibprm_cookie;
1777 txmsg->ibm_u.putack.ibpam_dst_cookie = tx->tx_cookie;
1779 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_ACK, nob);
1781 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1782 tx->tx_waiting = 1; /* waiting for PUT_DONE */
1783 kiblnd_queue_tx(tx, conn);
1785 /* reposted buffer reserved for PUT_DONE */
1786 post_credit = IBLND_POSTRX_NO_CREDIT;
1789 case IBLND_MSG_GET_REQ:
1790 if (lntmsg != NULL) {
1791 /* Optimized GET; RDMA lntmsg's payload */
1792 kiblnd_reply(ni, rx, lntmsg);
1794 /* GET didn't match anything */
1795 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_GET_DONE,
1797 rxmsg->ibm_u.get.ibgm_cookie);
1802 kiblnd_post_rx(rx, post_credit);
1807 kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name)
1809 struct task_struct *task = kthread_run(fn, arg, name);
1812 return PTR_ERR(task);
1814 atomic_inc(&kiblnd_data.kib_nthreads);
1819 kiblnd_thread_fini (void)
1821 atomic_dec (&kiblnd_data.kib_nthreads);
1825 kiblnd_peer_alive (kib_peer_t *peer)
1827 /* This is racy, but everyone's only writing cfs_time_current() */
1828 peer->ibp_last_alive = cfs_time_current();
1833 kiblnd_peer_notify (kib_peer_t *peer)
1836 cfs_time_t last_alive = 0;
1837 unsigned long flags;
1839 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1841 if (list_empty(&peer->ibp_conns) &&
1842 peer->ibp_accepting == 0 &&
1843 peer->ibp_connecting == 0 &&
1844 peer->ibp_error != 0) {
1845 error = peer->ibp_error;
1846 peer->ibp_error = 0;
1848 last_alive = peer->ibp_last_alive;
1851 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1854 lnet_notify(peer->ibp_ni,
1855 peer->ibp_nid, 0, last_alive);
1859 kiblnd_close_conn_locked (kib_conn_t *conn, int error)
1861 /* This just does the immediate housekeeping. 'error' is zero for a
1862 * normal shutdown which can happen only after the connection has been
1863 * established. If the connection is established, schedule the
1864 * connection to be finished off by the connd. Otherwise the connd is
1865 * already dealing with it (either to set it up or tear it down).
1866 * Caller holds kib_global_lock exclusively in irq context */
1867 kib_peer_t *peer = conn->ibc_peer;
1869 unsigned long flags;
1871 LASSERT (error != 0 || conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1873 if (error != 0 && conn->ibc_comms_error == 0)
1874 conn->ibc_comms_error = error;
1876 if (conn->ibc_state != IBLND_CONN_ESTABLISHED)
1877 return; /* already being handled */
1880 list_empty(&conn->ibc_tx_noops) &&
1881 list_empty(&conn->ibc_tx_queue) &&
1882 list_empty(&conn->ibc_tx_queue_rsrvd) &&
1883 list_empty(&conn->ibc_tx_queue_nocred) &&
1884 list_empty(&conn->ibc_active_txs)) {
1885 CDEBUG(D_NET, "closing conn to %s\n",
1886 libcfs_nid2str(peer->ibp_nid));
1888 CNETERR("Closing conn to %s: error %d%s%s%s%s%s\n",
1889 libcfs_nid2str(peer->ibp_nid), error,
1890 list_empty(&conn->ibc_tx_queue) ? "" : "(sending)",
1891 list_empty(&conn->ibc_tx_noops) ? "" : "(sending_noops)",
1892 list_empty(&conn->ibc_tx_queue_rsrvd) ?
1893 "" : "(sending_rsrvd)",
1894 list_empty(&conn->ibc_tx_queue_nocred) ?
1895 "" : "(sending_nocred)",
1896 list_empty(&conn->ibc_active_txs) ? "" : "(waiting)");
1899 dev = ((kib_net_t *)peer->ibp_ni->ni_data)->ibn_dev;
1900 list_del(&conn->ibc_list);
1901 /* connd (see below) takes over ibc_list's ref */
1903 if (list_empty(&peer->ibp_conns) && /* no more conns */
1904 kiblnd_peer_active(peer)) { /* still in peer table */
1905 kiblnd_unlink_peer_locked(peer);
1907 /* set/clear error on last conn */
1908 peer->ibp_error = conn->ibc_comms_error;
1911 kiblnd_set_conn_state(conn, IBLND_CONN_CLOSING);
1914 kiblnd_dev_can_failover(dev)) {
1915 list_add_tail(&dev->ibd_fail_list,
1916 &kiblnd_data.kib_failed_devs);
1917 wake_up(&kiblnd_data.kib_failover_waitq);
1920 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
1922 list_add_tail(&conn->ibc_list, &kiblnd_data.kib_connd_conns);
1923 wake_up(&kiblnd_data.kib_connd_waitq);
1925 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
1929 kiblnd_close_conn(kib_conn_t *conn, int error)
1931 unsigned long flags;
1933 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1935 kiblnd_close_conn_locked(conn, error);
1937 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1941 kiblnd_handle_early_rxs(kib_conn_t *conn)
1943 unsigned long flags;
1946 LASSERT(!in_interrupt());
1947 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1949 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1950 while (!list_empty(&conn->ibc_early_rxs)) {
1951 rx = list_entry(conn->ibc_early_rxs.next,
1953 list_del(&rx->rx_list);
1954 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1956 kiblnd_handle_rx(rx);
1958 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1960 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1964 kiblnd_abort_txs(kib_conn_t *conn, struct list_head *txs)
1966 struct list_head zombies = LIST_HEAD_INIT(zombies);
1967 struct list_head *tmp;
1968 struct list_head *nxt;
1971 spin_lock(&conn->ibc_lock);
1973 list_for_each_safe(tmp, nxt, txs) {
1974 tx = list_entry(tmp, kib_tx_t, tx_list);
1976 if (txs == &conn->ibc_active_txs) {
1977 LASSERT(!tx->tx_queued);
1978 LASSERT(tx->tx_waiting ||
1979 tx->tx_sending != 0);
1981 LASSERT(tx->tx_queued);
1984 tx->tx_status = -ECONNABORTED;
1987 if (tx->tx_sending == 0) {
1989 list_del(&tx->tx_list);
1990 list_add(&tx->tx_list, &zombies);
1994 spin_unlock(&conn->ibc_lock);
1996 kiblnd_txlist_done(conn->ibc_peer->ibp_ni, &zombies, -ECONNABORTED);
2000 kiblnd_finalise_conn (kib_conn_t *conn)
2002 LASSERT (!in_interrupt());
2003 LASSERT (conn->ibc_state > IBLND_CONN_INIT);
2005 kiblnd_set_conn_state(conn, IBLND_CONN_DISCONNECTED);
2007 /* abort_receives moves QP state to IB_QPS_ERR. This is only required
2008 * for connections that didn't get as far as being connected, because
2009 * rdma_disconnect() does this for free. */
2010 kiblnd_abort_receives(conn);
2012 /* Complete all tx descs not waiting for sends to complete.
2013 * NB we should be safe from RDMA now that the QP has changed state */
2015 kiblnd_abort_txs(conn, &conn->ibc_tx_noops);
2016 kiblnd_abort_txs(conn, &conn->ibc_tx_queue);
2017 kiblnd_abort_txs(conn, &conn->ibc_tx_queue_rsrvd);
2018 kiblnd_abort_txs(conn, &conn->ibc_tx_queue_nocred);
2019 kiblnd_abort_txs(conn, &conn->ibc_active_txs);
2021 kiblnd_handle_early_rxs(conn);
2025 kiblnd_peer_connect_failed (kib_peer_t *peer, int active, int error)
2027 struct list_head zombies = LIST_HEAD_INIT(zombies);
2028 unsigned long flags;
2030 LASSERT (error != 0);
2031 LASSERT (!in_interrupt());
2033 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2036 LASSERT (peer->ibp_connecting > 0);
2037 peer->ibp_connecting--;
2039 LASSERT (peer->ibp_accepting > 0);
2040 peer->ibp_accepting--;
2043 if (peer->ibp_connecting != 0 ||
2044 peer->ibp_accepting != 0) {
2045 /* another connection attempt under way... */
2046 write_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2051 if (list_empty(&peer->ibp_conns)) {
2052 /* Take peer's blocked transmits to complete with error */
2053 list_add(&zombies, &peer->ibp_tx_queue);
2054 list_del_init(&peer->ibp_tx_queue);
2056 if (kiblnd_peer_active(peer))
2057 kiblnd_unlink_peer_locked(peer);
2059 peer->ibp_error = error;
2061 /* Can't have blocked transmits if there are connections */
2062 LASSERT(list_empty(&peer->ibp_tx_queue));
2065 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2067 kiblnd_peer_notify(peer);
2069 if (list_empty(&zombies))
2072 CNETERR("Deleting messages for %s: connection failed\n",
2073 libcfs_nid2str(peer->ibp_nid));
2075 kiblnd_txlist_done(peer->ibp_ni, &zombies, -EHOSTUNREACH);
2079 kiblnd_connreq_done(kib_conn_t *conn, int status)
2081 kib_peer_t *peer = conn->ibc_peer;
2083 struct list_head txs;
2084 unsigned long flags;
2087 active = (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2089 CDEBUG(D_NET,"%s: active(%d), version(%x), status(%d)\n",
2090 libcfs_nid2str(peer->ibp_nid), active,
2091 conn->ibc_version, status);
2093 LASSERT (!in_interrupt());
2094 LASSERT ((conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT &&
2095 peer->ibp_connecting > 0) ||
2096 (conn->ibc_state == IBLND_CONN_PASSIVE_WAIT &&
2097 peer->ibp_accepting > 0));
2099 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
2100 conn->ibc_connvars = NULL;
2103 /* failed to establish connection */
2104 kiblnd_peer_connect_failed(peer, active, status);
2105 kiblnd_finalise_conn(conn);
2109 /* connection established */
2110 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2112 conn->ibc_last_send = jiffies;
2113 kiblnd_set_conn_state(conn, IBLND_CONN_ESTABLISHED);
2114 kiblnd_peer_alive(peer);
2116 /* Add conn to peer's list and nuke any dangling conns from a different
2117 * peer instance... */
2118 kiblnd_conn_addref(conn); /* +1 ref for ibc_list */
2119 list_add(&conn->ibc_list, &peer->ibp_conns);
2121 peer->ibp_connecting--;
2123 peer->ibp_accepting--;
2125 if (peer->ibp_version == 0) {
2126 peer->ibp_version = conn->ibc_version;
2127 peer->ibp_incarnation = conn->ibc_incarnation;
2130 if (peer->ibp_version != conn->ibc_version ||
2131 peer->ibp_incarnation != conn->ibc_incarnation) {
2132 kiblnd_close_stale_conns_locked(peer, conn->ibc_version,
2133 conn->ibc_incarnation);
2134 peer->ibp_version = conn->ibc_version;
2135 peer->ibp_incarnation = conn->ibc_incarnation;
2138 /* grab pending txs while I have the lock */
2139 list_add(&txs, &peer->ibp_tx_queue);
2140 list_del_init(&peer->ibp_tx_queue);
2142 if (!kiblnd_peer_active(peer) || /* peer has been deleted */
2143 conn->ibc_comms_error != 0) { /* error has happened already */
2144 lnet_ni_t *ni = peer->ibp_ni;
2146 /* start to shut down connection */
2147 kiblnd_close_conn_locked(conn, -ECONNABORTED);
2148 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2150 kiblnd_txlist_done(ni, &txs, -ECONNABORTED);
2155 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2157 /* Schedule blocked txs */
2158 spin_lock(&conn->ibc_lock);
2159 while (!list_empty(&txs)) {
2160 tx = list_entry(txs.next, kib_tx_t, tx_list);
2161 list_del(&tx->tx_list);
2163 kiblnd_queue_tx_locked(tx, conn);
2165 spin_unlock(&conn->ibc_lock);
2167 kiblnd_check_sends(conn);
2169 /* schedule blocked rxs */
2170 kiblnd_handle_early_rxs(conn);
2174 kiblnd_reject(struct rdma_cm_id *cmid, kib_rej_t *rej)
2178 rc = rdma_reject(cmid, rej, sizeof(*rej));
2181 CWARN("Error %d sending reject\n", rc);
2185 kiblnd_passive_connect (struct rdma_cm_id *cmid, void *priv, int priv_nob)
2187 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2188 kib_msg_t *reqmsg = priv;
2194 lnet_ni_t *ni = NULL;
2195 kib_net_t *net = NULL;
2197 struct rdma_conn_param cp;
2199 int version = IBLND_MSG_VERSION;
2200 unsigned long flags;
2202 struct sockaddr_in *peer_addr;
2203 LASSERT (!in_interrupt());
2205 /* cmid inherits 'context' from the corresponding listener id */
2206 ibdev = (kib_dev_t *)cmid->context;
2207 LASSERT (ibdev != NULL);
2209 memset(&rej, 0, sizeof(rej));
2210 rej.ibr_magic = IBLND_MSG_MAGIC;
2211 rej.ibr_why = IBLND_REJECT_FATAL;
2212 rej.ibr_cp.ibcp_max_msg_size = IBLND_MSG_SIZE;
2214 peer_addr = (struct sockaddr_in *)&(cmid->route.addr.dst_addr);
2215 if (*kiblnd_tunables.kib_require_priv_port &&
2216 ntohs(peer_addr->sin_port) >= PROT_SOCK) {
2217 __u32 ip = ntohl(peer_addr->sin_addr.s_addr);
2218 CERROR("Peer's port (%u.%u.%u.%u:%hu) is not privileged\n",
2219 HIPQUAD(ip), ntohs(peer_addr->sin_port));
2223 if (priv_nob < offsetof(kib_msg_t, ibm_type)) {
2224 CERROR("Short connection request\n");
2228 /* Future protocol version compatibility support! If the
2229 * o2iblnd-specific protocol changes, or when LNET unifies
2230 * protocols over all LNDs, the initial connection will
2231 * negotiate a protocol version. I trap this here to avoid
2232 * console errors; the reject tells the peer which protocol I
2234 if (reqmsg->ibm_magic == LNET_PROTO_MAGIC ||
2235 reqmsg->ibm_magic == __swab32(LNET_PROTO_MAGIC))
2237 if (reqmsg->ibm_magic == IBLND_MSG_MAGIC &&
2238 reqmsg->ibm_version != IBLND_MSG_VERSION &&
2239 reqmsg->ibm_version != IBLND_MSG_VERSION_1)
2241 if (reqmsg->ibm_magic == __swab32(IBLND_MSG_MAGIC) &&
2242 reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION) &&
2243 reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION_1))
2246 rc = kiblnd_unpack_msg(reqmsg, priv_nob);
2248 CERROR("Can't parse connection request: %d\n", rc);
2252 nid = reqmsg->ibm_srcnid;
2253 ni = lnet_net2ni(LNET_NIDNET(reqmsg->ibm_dstnid));
2256 net = (kib_net_t *)ni->ni_data;
2257 rej.ibr_incarnation = net->ibn_incarnation;
2260 if (ni == NULL || /* no matching net */
2261 ni->ni_nid != reqmsg->ibm_dstnid || /* right NET, wrong NID! */
2262 net->ibn_dev != ibdev) { /* wrong device */
2263 CERROR("Can't accept %s on %s (%s:%d:%u.%u.%u.%u): "
2264 "bad dst nid %s\n", libcfs_nid2str(nid),
2265 ni == NULL ? "NA" : libcfs_nid2str(ni->ni_nid),
2266 ibdev->ibd_ifname, ibdev->ibd_nnets,
2267 HIPQUAD(ibdev->ibd_ifip),
2268 libcfs_nid2str(reqmsg->ibm_dstnid));
2273 /* check time stamp as soon as possible */
2274 if (reqmsg->ibm_dststamp != 0 &&
2275 reqmsg->ibm_dststamp != net->ibn_incarnation) {
2276 CWARN("Stale connection request\n");
2277 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2281 /* I can accept peer's version */
2282 version = reqmsg->ibm_version;
2284 if (reqmsg->ibm_type != IBLND_MSG_CONNREQ) {
2285 CERROR("Unexpected connreq msg type: %x from %s\n",
2286 reqmsg->ibm_type, libcfs_nid2str(nid));
2290 if (reqmsg->ibm_u.connparams.ibcp_queue_depth !=
2291 IBLND_MSG_QUEUE_SIZE(version)) {
2292 CERROR("Can't accept %s: incompatible queue depth %d (%d wanted)\n",
2293 libcfs_nid2str(nid), reqmsg->ibm_u.connparams.ibcp_queue_depth,
2294 IBLND_MSG_QUEUE_SIZE(version));
2296 if (version == IBLND_MSG_VERSION)
2297 rej.ibr_why = IBLND_REJECT_MSG_QUEUE_SIZE;
2302 if (reqmsg->ibm_u.connparams.ibcp_max_frags !=
2303 IBLND_RDMA_FRAGS(version)) {
2304 CERROR("Can't accept %s(version %x): "
2305 "incompatible max_frags %d (%d wanted)\n",
2306 libcfs_nid2str(nid), version,
2307 reqmsg->ibm_u.connparams.ibcp_max_frags,
2308 IBLND_RDMA_FRAGS(version));
2310 if (version == IBLND_MSG_VERSION)
2311 rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2317 if (reqmsg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2318 CERROR("Can't accept %s: message size %d too big (%d max)\n",
2319 libcfs_nid2str(nid),
2320 reqmsg->ibm_u.connparams.ibcp_max_msg_size,
2325 /* assume 'nid' is a new peer; create */
2326 rc = kiblnd_create_peer(ni, &peer, nid);
2328 CERROR("Can't create peer for %s\n", libcfs_nid2str(nid));
2329 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2333 write_lock_irqsave(g_lock, flags);
2335 peer2 = kiblnd_find_peer_locked(nid);
2336 if (peer2 != NULL) {
2337 if (peer2->ibp_version == 0) {
2338 peer2->ibp_version = version;
2339 peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2342 /* not the guy I've talked with */
2343 if (peer2->ibp_incarnation != reqmsg->ibm_srcstamp ||
2344 peer2->ibp_version != version) {
2345 kiblnd_close_peer_conns_locked(peer2, -ESTALE);
2346 write_unlock_irqrestore(g_lock, flags);
2348 CWARN("Conn stale %s [old ver: %x, new ver: %x]\n",
2349 libcfs_nid2str(nid), peer2->ibp_version, version);
2351 kiblnd_peer_decref(peer);
2352 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2356 /* tie-break connection race in favour of the higher NID */
2357 if (peer2->ibp_connecting != 0 &&
2359 write_unlock_irqrestore(g_lock, flags);
2361 CWARN("Conn race %s\n", libcfs_nid2str(peer2->ibp_nid));
2363 kiblnd_peer_decref(peer);
2364 rej.ibr_why = IBLND_REJECT_CONN_RACE;
2368 peer2->ibp_accepting++;
2369 kiblnd_peer_addref(peer2);
2371 write_unlock_irqrestore(g_lock, flags);
2372 kiblnd_peer_decref(peer);
2375 /* Brand new peer */
2376 LASSERT (peer->ibp_accepting == 0);
2377 LASSERT (peer->ibp_version == 0 &&
2378 peer->ibp_incarnation == 0);
2380 peer->ibp_accepting = 1;
2381 peer->ibp_version = version;
2382 peer->ibp_incarnation = reqmsg->ibm_srcstamp;
2384 /* I have a ref on ni that prevents it being shutdown */
2385 LASSERT (net->ibn_shutdown == 0);
2387 kiblnd_peer_addref(peer);
2388 list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid));
2390 write_unlock_irqrestore(g_lock, flags);
2393 conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_PASSIVE_WAIT, version);
2395 kiblnd_peer_connect_failed(peer, 0, -ENOMEM);
2396 kiblnd_peer_decref(peer);
2397 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2401 /* conn now "owns" cmid, so I return success from here on to ensure the
2402 * CM callback doesn't destroy cmid. */
2404 conn->ibc_incarnation = reqmsg->ibm_srcstamp;
2405 conn->ibc_credits = IBLND_MSG_QUEUE_SIZE(version);
2406 conn->ibc_reserved_credits = IBLND_MSG_QUEUE_SIZE(version);
2407 LASSERT (conn->ibc_credits + conn->ibc_reserved_credits + IBLND_OOB_MSGS(version)
2408 <= IBLND_RX_MSGS(version));
2410 ackmsg = &conn->ibc_connvars->cv_msg;
2411 memset(ackmsg, 0, sizeof(*ackmsg));
2413 kiblnd_init_msg(ackmsg, IBLND_MSG_CONNACK,
2414 sizeof(ackmsg->ibm_u.connparams));
2415 ackmsg->ibm_u.connparams.ibcp_queue_depth = IBLND_MSG_QUEUE_SIZE(version);
2416 ackmsg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2417 ackmsg->ibm_u.connparams.ibcp_max_frags = IBLND_RDMA_FRAGS(version);
2419 kiblnd_pack_msg(ni, ackmsg, version, 0, nid, reqmsg->ibm_srcstamp);
2421 memset(&cp, 0, sizeof(cp));
2422 cp.private_data = ackmsg;
2423 cp.private_data_len = ackmsg->ibm_nob;
2424 cp.responder_resources = 0; /* No atomic ops or RDMA reads */
2425 cp.initiator_depth = 0;
2426 cp.flow_control = 1;
2427 cp.retry_count = *kiblnd_tunables.kib_retry_count;
2428 cp.rnr_retry_count = *kiblnd_tunables.kib_rnr_retry_count;
2430 CDEBUG(D_NET, "Accept %s\n", libcfs_nid2str(nid));
2432 rc = rdma_accept(cmid, &cp);
2434 CERROR("Can't accept %s: %d\n", libcfs_nid2str(nid), rc);
2435 rej.ibr_version = version;
2436 rej.ibr_why = IBLND_REJECT_FATAL;
2438 kiblnd_reject(cmid, &rej);
2439 kiblnd_connreq_done(conn, rc);
2440 kiblnd_conn_decref(conn);
2450 rej.ibr_version = version;
2451 rej.ibr_cp.ibcp_queue_depth = IBLND_MSG_QUEUE_SIZE(version);
2452 rej.ibr_cp.ibcp_max_frags = IBLND_RDMA_FRAGS(version);
2453 kiblnd_reject(cmid, &rej);
2455 return -ECONNREFUSED;
2459 kiblnd_reconnect (kib_conn_t *conn, int version,
2460 __u64 incarnation, int why, kib_connparams_t *cp)
2462 kib_peer_t *peer = conn->ibc_peer;
2465 unsigned long flags;
2467 LASSERT (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2468 LASSERT (peer->ibp_connecting > 0); /* 'conn' at least */
2470 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2472 /* retry connection if it's still needed and no other connection
2473 * attempts (active or passive) are in progress
2474 * NB: reconnect is still needed even when ibp_tx_queue is
2475 * empty if ibp_version != version because reconnect may be
2476 * initiated by kiblnd_query() */
2477 if ((!list_empty(&peer->ibp_tx_queue) ||
2478 peer->ibp_version != version) &&
2479 peer->ibp_connecting == 1 &&
2480 peer->ibp_accepting == 0) {
2482 peer->ibp_connecting++;
2484 peer->ibp_version = version;
2485 peer->ibp_incarnation = incarnation;
2488 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2498 case IBLND_REJECT_CONN_STALE:
2502 case IBLND_REJECT_CONN_RACE:
2503 reason = "conn race";
2506 case IBLND_REJECT_CONN_UNCOMPAT:
2507 reason = "version negotiation";
2511 CNETERR("%s: retrying (%s), %x, %x, "
2512 "queue_dep: %d, max_frag: %d, msg_size: %d\n",
2513 libcfs_nid2str(peer->ibp_nid),
2514 reason, IBLND_MSG_VERSION, version,
2515 cp != NULL? cp->ibcp_queue_depth :IBLND_MSG_QUEUE_SIZE(version),
2516 cp != NULL? cp->ibcp_max_frags : IBLND_RDMA_FRAGS(version),
2517 cp != NULL? cp->ibcp_max_msg_size: IBLND_MSG_SIZE);
2519 kiblnd_connect_peer(peer);
2523 kiblnd_rejected (kib_conn_t *conn, int reason, void *priv, int priv_nob)
2525 kib_peer_t *peer = conn->ibc_peer;
2527 LASSERT (!in_interrupt());
2528 LASSERT (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2531 case IB_CM_REJ_STALE_CONN:
2532 kiblnd_reconnect(conn, IBLND_MSG_VERSION, 0,
2533 IBLND_REJECT_CONN_STALE, NULL);
2536 case IB_CM_REJ_INVALID_SERVICE_ID:
2537 CNETERR("%s rejected: no listener at %d\n",
2538 libcfs_nid2str(peer->ibp_nid),
2539 *kiblnd_tunables.kib_service);
2542 case IB_CM_REJ_CONSUMER_DEFINED:
2543 if (priv_nob >= offsetof(kib_rej_t, ibr_padding)) {
2544 kib_rej_t *rej = priv;
2545 kib_connparams_t *cp = NULL;
2547 __u64 incarnation = -1;
2549 /* NB. default incarnation is -1 because:
2550 * a) V1 will ignore dst incarnation in connreq.
2551 * b) V2 will provide incarnation while rejecting me,
2552 * -1 will be overwrote.
2554 * if I try to connect to a V1 peer with V2 protocol,
2555 * it rejected me then upgrade to V2, I have no idea
2556 * about the upgrading and try to reconnect with V1,
2557 * in this case upgraded V2 can find out I'm trying to
2558 * talk to the old guy and reject me(incarnation is -1).
2561 if (rej->ibr_magic == __swab32(IBLND_MSG_MAGIC) ||
2562 rej->ibr_magic == __swab32(LNET_PROTO_MAGIC)) {
2563 __swab32s(&rej->ibr_magic);
2564 __swab16s(&rej->ibr_version);
2568 if (priv_nob >= sizeof(kib_rej_t) &&
2569 rej->ibr_version > IBLND_MSG_VERSION_1) {
2570 /* priv_nob is always 148 in current version
2571 * of OFED, so we still need to check version.
2572 * (define of IB_CM_REJ_PRIVATE_DATA_SIZE) */
2576 __swab64s(&rej->ibr_incarnation);
2577 __swab16s(&cp->ibcp_queue_depth);
2578 __swab16s(&cp->ibcp_max_frags);
2579 __swab32s(&cp->ibcp_max_msg_size);
2582 incarnation = rej->ibr_incarnation;
2585 if (rej->ibr_magic != IBLND_MSG_MAGIC &&
2586 rej->ibr_magic != LNET_PROTO_MAGIC) {
2587 CERROR("%s rejected: consumer defined fatal error\n",
2588 libcfs_nid2str(peer->ibp_nid));
2592 if (rej->ibr_version != IBLND_MSG_VERSION &&
2593 rej->ibr_version != IBLND_MSG_VERSION_1) {
2594 CERROR("%s rejected: o2iblnd version %x error\n",
2595 libcfs_nid2str(peer->ibp_nid),
2600 if (rej->ibr_why == IBLND_REJECT_FATAL &&
2601 rej->ibr_version == IBLND_MSG_VERSION_1) {
2602 CDEBUG(D_NET, "rejected by old version peer %s: %x\n",
2603 libcfs_nid2str(peer->ibp_nid), rej->ibr_version);
2605 if (conn->ibc_version != IBLND_MSG_VERSION_1)
2606 rej->ibr_why = IBLND_REJECT_CONN_UNCOMPAT;
2609 switch (rej->ibr_why) {
2610 case IBLND_REJECT_CONN_RACE:
2611 case IBLND_REJECT_CONN_STALE:
2612 case IBLND_REJECT_CONN_UNCOMPAT:
2613 kiblnd_reconnect(conn, rej->ibr_version,
2614 incarnation, rej->ibr_why, cp);
2617 case IBLND_REJECT_MSG_QUEUE_SIZE:
2618 CERROR("%s rejected: incompatible message queue depth %d, %d\n",
2619 libcfs_nid2str(peer->ibp_nid),
2620 cp != NULL ? cp->ibcp_queue_depth :
2621 IBLND_MSG_QUEUE_SIZE(rej->ibr_version),
2622 IBLND_MSG_QUEUE_SIZE(conn->ibc_version));
2625 case IBLND_REJECT_RDMA_FRAGS:
2626 CERROR("%s rejected: incompatible # of RDMA fragments %d, %d\n",
2627 libcfs_nid2str(peer->ibp_nid),
2628 cp != NULL ? cp->ibcp_max_frags :
2629 IBLND_RDMA_FRAGS(rej->ibr_version),
2630 IBLND_RDMA_FRAGS(conn->ibc_version));
2633 case IBLND_REJECT_NO_RESOURCES:
2634 CERROR("%s rejected: o2iblnd no resources\n",
2635 libcfs_nid2str(peer->ibp_nid));
2638 case IBLND_REJECT_FATAL:
2639 CERROR("%s rejected: o2iblnd fatal error\n",
2640 libcfs_nid2str(peer->ibp_nid));
2644 CERROR("%s rejected: o2iblnd reason %d\n",
2645 libcfs_nid2str(peer->ibp_nid),
2653 CNETERR("%s rejected: reason %d, size %d\n",
2654 libcfs_nid2str(peer->ibp_nid), reason, priv_nob);
2658 kiblnd_connreq_done(conn, -ECONNREFUSED);
2662 kiblnd_check_connreply (kib_conn_t *conn, void *priv, int priv_nob)
2664 kib_peer_t *peer = conn->ibc_peer;
2665 lnet_ni_t *ni = peer->ibp_ni;
2666 kib_net_t *net = ni->ni_data;
2667 kib_msg_t *msg = priv;
2668 int ver = conn->ibc_version;
2669 int rc = kiblnd_unpack_msg(msg, priv_nob);
2670 unsigned long flags;
2672 LASSERT (net != NULL);
2675 CERROR("Can't unpack connack from %s: %d\n",
2676 libcfs_nid2str(peer->ibp_nid), rc);
2680 if (msg->ibm_type != IBLND_MSG_CONNACK) {
2681 CERROR("Unexpected message %d from %s\n",
2682 msg->ibm_type, libcfs_nid2str(peer->ibp_nid));
2687 if (ver != msg->ibm_version) {
2688 CERROR("%s replied version %x is different with "
2689 "requested version %x\n",
2690 libcfs_nid2str(peer->ibp_nid), msg->ibm_version, ver);
2695 if (msg->ibm_u.connparams.ibcp_queue_depth !=
2696 IBLND_MSG_QUEUE_SIZE(ver)) {
2697 CERROR("%s has incompatible queue depth %d(%d wanted)\n",
2698 libcfs_nid2str(peer->ibp_nid),
2699 msg->ibm_u.connparams.ibcp_queue_depth,
2700 IBLND_MSG_QUEUE_SIZE(ver));
2705 if (msg->ibm_u.connparams.ibcp_max_frags !=
2706 IBLND_RDMA_FRAGS(ver)) {
2707 CERROR("%s has incompatible max_frags %d (%d wanted)\n",
2708 libcfs_nid2str(peer->ibp_nid),
2709 msg->ibm_u.connparams.ibcp_max_frags,
2710 IBLND_RDMA_FRAGS(ver));
2715 if (msg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2716 CERROR("%s max message size %d too big (%d max)\n",
2717 libcfs_nid2str(peer->ibp_nid),
2718 msg->ibm_u.connparams.ibcp_max_msg_size,
2724 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2725 if (msg->ibm_dstnid == ni->ni_nid &&
2726 msg->ibm_dststamp == net->ibn_incarnation)
2730 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2733 CERROR("Bad connection reply from %s, rc = %d, "
2734 "version: %x max_frags: %d\n",
2735 libcfs_nid2str(peer->ibp_nid), rc,
2736 msg->ibm_version, msg->ibm_u.connparams.ibcp_max_frags);
2740 conn->ibc_incarnation = msg->ibm_srcstamp;
2742 conn->ibc_reserved_credits = IBLND_MSG_QUEUE_SIZE(ver);
2743 LASSERT (conn->ibc_credits + conn->ibc_reserved_credits + IBLND_OOB_MSGS(ver)
2744 <= IBLND_RX_MSGS(ver));
2746 kiblnd_connreq_done(conn, 0);
2750 /* NB My QP has already established itself, so I handle anything going
2751 * wrong here by setting ibc_comms_error.
2752 * kiblnd_connreq_done(0) moves the conn state to ESTABLISHED, but then
2753 * immediately tears it down. */
2756 conn->ibc_comms_error = rc;
2757 kiblnd_connreq_done(conn, 0);
2761 kiblnd_active_connect (struct rdma_cm_id *cmid)
2763 kib_peer_t *peer = (kib_peer_t *)cmid->context;
2766 struct rdma_conn_param cp;
2769 unsigned long flags;
2772 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2774 incarnation = peer->ibp_incarnation;
2775 version = (peer->ibp_version == 0) ? IBLND_MSG_VERSION :
2778 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2780 conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_ACTIVE_CONNECT, version);
2782 kiblnd_peer_connect_failed(peer, 1, -ENOMEM);
2783 kiblnd_peer_decref(peer); /* lose cmid's ref */
2787 /* conn "owns" cmid now, so I return success from here on to ensure the
2788 * CM callback doesn't destroy cmid. conn also takes over cmid's ref
2791 msg = &conn->ibc_connvars->cv_msg;
2793 memset(msg, 0, sizeof(*msg));
2794 kiblnd_init_msg(msg, IBLND_MSG_CONNREQ, sizeof(msg->ibm_u.connparams));
2795 msg->ibm_u.connparams.ibcp_queue_depth = IBLND_MSG_QUEUE_SIZE(version);
2796 msg->ibm_u.connparams.ibcp_max_frags = IBLND_RDMA_FRAGS(version);
2797 msg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2799 kiblnd_pack_msg(peer->ibp_ni, msg, version,
2800 0, peer->ibp_nid, incarnation);
2802 memset(&cp, 0, sizeof(cp));
2803 cp.private_data = msg;
2804 cp.private_data_len = msg->ibm_nob;
2805 cp.responder_resources = 0; /* No atomic ops or RDMA reads */
2806 cp.initiator_depth = 0;
2807 cp.flow_control = 1;
2808 cp.retry_count = *kiblnd_tunables.kib_retry_count;
2809 cp.rnr_retry_count = *kiblnd_tunables.kib_rnr_retry_count;
2811 LASSERT(cmid->context == (void *)conn);
2812 LASSERT(conn->ibc_cmid == cmid);
2814 rc = rdma_connect(cmid, &cp);
2816 CERROR("Can't connect to %s: %d\n",
2817 libcfs_nid2str(peer->ibp_nid), rc);
2818 kiblnd_connreq_done(conn, rc);
2819 kiblnd_conn_decref(conn);
2826 kiblnd_cm_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2832 switch (event->event) {
2834 CERROR("Unexpected event: %d, status: %d\n",
2835 event->event, event->status);
2838 case RDMA_CM_EVENT_CONNECT_REQUEST:
2839 /* destroy cmid on failure */
2840 rc = kiblnd_passive_connect(cmid,
2841 (void *)KIBLND_CONN_PARAM(event),
2842 KIBLND_CONN_PARAM_LEN(event));
2843 CDEBUG(D_NET, "connreq: %d\n", rc);
2846 case RDMA_CM_EVENT_ADDR_ERROR:
2847 peer = (kib_peer_t *)cmid->context;
2848 CNETERR("%s: ADDR ERROR %d\n",
2849 libcfs_nid2str(peer->ibp_nid), event->status);
2850 kiblnd_peer_connect_failed(peer, 1, -EHOSTUNREACH);
2851 kiblnd_peer_decref(peer);
2852 return -EHOSTUNREACH; /* rc != 0 destroys cmid */
2854 case RDMA_CM_EVENT_ADDR_RESOLVED:
2855 peer = (kib_peer_t *)cmid->context;
2857 CDEBUG(D_NET,"%s Addr resolved: %d\n",
2858 libcfs_nid2str(peer->ibp_nid), event->status);
2860 if (event->status != 0) {
2861 CNETERR("Can't resolve address for %s: %d\n",
2862 libcfs_nid2str(peer->ibp_nid), event->status);
2865 rc = rdma_resolve_route(
2866 cmid, *kiblnd_tunables.kib_timeout * 1000);
2869 /* Can't initiate route resolution */
2870 CERROR("Can't resolve route for %s: %d\n",
2871 libcfs_nid2str(peer->ibp_nid), rc);
2873 kiblnd_peer_connect_failed(peer, 1, rc);
2874 kiblnd_peer_decref(peer);
2875 return rc; /* rc != 0 destroys cmid */
2877 case RDMA_CM_EVENT_ROUTE_ERROR:
2878 peer = (kib_peer_t *)cmid->context;
2879 CNETERR("%s: ROUTE ERROR %d\n",
2880 libcfs_nid2str(peer->ibp_nid), event->status);
2881 kiblnd_peer_connect_failed(peer, 1, -EHOSTUNREACH);
2882 kiblnd_peer_decref(peer);
2883 return -EHOSTUNREACH; /* rc != 0 destroys cmid */
2885 case RDMA_CM_EVENT_ROUTE_RESOLVED:
2886 peer = (kib_peer_t *)cmid->context;
2887 CDEBUG(D_NET,"%s Route resolved: %d\n",
2888 libcfs_nid2str(peer->ibp_nid), event->status);
2890 if (event->status == 0)
2891 return kiblnd_active_connect(cmid);
2893 CNETERR("Can't resolve route for %s: %d\n",
2894 libcfs_nid2str(peer->ibp_nid), event->status);
2895 kiblnd_peer_connect_failed(peer, 1, event->status);
2896 kiblnd_peer_decref(peer);
2897 return event->status; /* rc != 0 destroys cmid */
2899 case RDMA_CM_EVENT_UNREACHABLE:
2900 conn = (kib_conn_t *)cmid->context;
2901 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
2902 conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
2903 CNETERR("%s: UNREACHABLE %d\n",
2904 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
2905 kiblnd_connreq_done(conn, -ENETDOWN);
2906 kiblnd_conn_decref(conn);
2909 case RDMA_CM_EVENT_CONNECT_ERROR:
2910 conn = (kib_conn_t *)cmid->context;
2911 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
2912 conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
2913 CNETERR("%s: CONNECT ERROR %d\n",
2914 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
2915 kiblnd_connreq_done(conn, -ENOTCONN);
2916 kiblnd_conn_decref(conn);
2919 case RDMA_CM_EVENT_REJECTED:
2920 conn = (kib_conn_t *)cmid->context;
2921 switch (conn->ibc_state) {
2925 case IBLND_CONN_PASSIVE_WAIT:
2926 CERROR ("%s: REJECTED %d\n",
2927 libcfs_nid2str(conn->ibc_peer->ibp_nid),
2929 kiblnd_connreq_done(conn, -ECONNRESET);
2932 case IBLND_CONN_ACTIVE_CONNECT:
2933 kiblnd_rejected(conn, event->status,
2934 (void *)KIBLND_CONN_PARAM(event),
2935 KIBLND_CONN_PARAM_LEN(event));
2938 kiblnd_conn_decref(conn);
2941 case RDMA_CM_EVENT_ESTABLISHED:
2942 conn = (kib_conn_t *)cmid->context;
2943 switch (conn->ibc_state) {
2947 case IBLND_CONN_PASSIVE_WAIT:
2948 CDEBUG(D_NET, "ESTABLISHED (passive): %s\n",
2949 libcfs_nid2str(conn->ibc_peer->ibp_nid));
2950 kiblnd_connreq_done(conn, 0);
2953 case IBLND_CONN_ACTIVE_CONNECT:
2954 CDEBUG(D_NET, "ESTABLISHED(active): %s\n",
2955 libcfs_nid2str(conn->ibc_peer->ibp_nid));
2956 kiblnd_check_connreply(conn,
2957 (void *)KIBLND_CONN_PARAM(event),
2958 KIBLND_CONN_PARAM_LEN(event));
2961 /* net keeps its ref on conn! */
2964 #ifdef HAVE_OFED_RDMA_CMEV_TIMEWAIT_EXIT
2965 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
2966 CDEBUG(D_NET, "Ignore TIMEWAIT_EXIT event\n");
2969 case RDMA_CM_EVENT_DISCONNECTED:
2970 conn = (kib_conn_t *)cmid->context;
2971 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
2972 CERROR("%s DISCONNECTED\n",
2973 libcfs_nid2str(conn->ibc_peer->ibp_nid));
2974 kiblnd_connreq_done(conn, -ECONNRESET);
2976 kiblnd_close_conn(conn, 0);
2978 kiblnd_conn_decref(conn);
2979 cmid->context = NULL;
2982 case RDMA_CM_EVENT_DEVICE_REMOVAL:
2983 LCONSOLE_ERROR_MSG(0x131,
2984 "Received notification of device removal\n"
2985 "Please shutdown LNET to allow this to proceed\n");
2986 /* Can't remove network from underneath LNET for now, so I have
2990 #ifdef HAVE_OFED_RDMA_CMEV_ADDRCHANGE
2991 case RDMA_CM_EVENT_ADDR_CHANGE:
2992 LCONSOLE_INFO("Physical link changed (eg hca/port)\n");
2999 kiblnd_check_txs_locked(kib_conn_t *conn, struct list_head *txs)
3002 struct list_head *ttmp;
3004 list_for_each(ttmp, txs) {
3005 tx = list_entry(ttmp, kib_tx_t, tx_list);
3007 if (txs != &conn->ibc_active_txs) {
3008 LASSERT(tx->tx_queued);
3010 LASSERT(!tx->tx_queued);
3011 LASSERT(tx->tx_waiting || tx->tx_sending != 0);
3014 if (cfs_time_aftereq(jiffies, tx->tx_deadline)) {
3015 CERROR("Timed out tx: %s, %lu seconds\n",
3016 kiblnd_queue2str(conn, txs),
3017 cfs_duration_sec(jiffies - tx->tx_deadline));
3026 kiblnd_conn_timed_out_locked(kib_conn_t *conn)
3028 return kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue) ||
3029 kiblnd_check_txs_locked(conn, &conn->ibc_tx_noops) ||
3030 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_rsrvd) ||
3031 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_nocred) ||
3032 kiblnd_check_txs_locked(conn, &conn->ibc_active_txs);
3036 kiblnd_check_conns (int idx)
3038 struct list_head closes = LIST_HEAD_INIT(closes);
3039 struct list_head checksends = LIST_HEAD_INIT(checksends);
3040 struct list_head *peers = &kiblnd_data.kib_peers[idx];
3041 struct list_head *ptmp;
3044 struct list_head *ctmp;
3045 unsigned long flags;
3047 /* NB. We expect to have a look at all the peers and not find any
3048 * RDMAs to time out, so we just use a shared lock while we
3050 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3052 list_for_each(ptmp, peers) {
3053 peer = list_entry(ptmp, kib_peer_t, ibp_list);
3055 list_for_each(ctmp, &peer->ibp_conns) {
3059 conn = list_entry(ctmp, kib_conn_t, ibc_list);
3061 LASSERT(conn->ibc_state == IBLND_CONN_ESTABLISHED);
3063 spin_lock(&conn->ibc_lock);
3065 sendnoop = kiblnd_need_noop(conn);
3066 timedout = kiblnd_conn_timed_out_locked(conn);
3067 if (!sendnoop && !timedout) {
3068 spin_unlock(&conn->ibc_lock);
3073 CERROR("Timed out RDMA with %s (%lu): "
3074 "c: %u, oc: %u, rc: %u\n",
3075 libcfs_nid2str(peer->ibp_nid),
3076 cfs_duration_sec(cfs_time_current() -
3077 peer->ibp_last_alive),
3079 conn->ibc_outstanding_credits,
3080 conn->ibc_reserved_credits);
3081 list_add(&conn->ibc_connd_list, &closes);
3083 list_add(&conn->ibc_connd_list, &checksends);
3085 /* +ref for 'closes' or 'checksends' */
3086 kiblnd_conn_addref(conn);
3088 spin_unlock(&conn->ibc_lock);
3092 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3094 /* Handle timeout by closing the whole
3095 * connection. We can only be sure RDMA activity
3096 * has ceased once the QP has been modified. */
3097 while (!list_empty(&closes)) {
3098 conn = list_entry(closes.next,
3099 kib_conn_t, ibc_connd_list);
3100 list_del(&conn->ibc_connd_list);
3101 kiblnd_close_conn(conn, -ETIMEDOUT);
3102 kiblnd_conn_decref(conn);
3105 /* In case we have enough credits to return via a
3106 * NOOP, but there were no non-blocking tx descs
3107 * free to do it last time... */
3108 while (!list_empty(&checksends)) {
3109 conn = list_entry(checksends.next,
3110 kib_conn_t, ibc_connd_list);
3111 list_del(&conn->ibc_connd_list);
3112 kiblnd_check_sends(conn);
3113 kiblnd_conn_decref(conn);
3118 kiblnd_disconnect_conn (kib_conn_t *conn)
3120 LASSERT (!in_interrupt());
3121 LASSERT (current == kiblnd_data.kib_connd);
3122 LASSERT (conn->ibc_state == IBLND_CONN_CLOSING);
3124 rdma_disconnect(conn->ibc_cmid);
3125 kiblnd_finalise_conn(conn);
3127 kiblnd_peer_notify(conn->ibc_peer);
3131 kiblnd_connd (void *arg)
3134 unsigned long flags;
3140 unsigned long deadline = jiffies;
3142 cfs_block_allsigs ();
3144 init_waitqueue_entry_current (&wait);
3145 kiblnd_data.kib_connd = current;
3147 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3149 while (!kiblnd_data.kib_shutdown) {
3153 if (!list_empty(&kiblnd_data.kib_connd_zombies)) {
3154 conn = list_entry(kiblnd_data. \
3155 kib_connd_zombies.next,
3156 kib_conn_t, ibc_list);
3157 list_del(&conn->ibc_list);
3159 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock,
3163 kiblnd_destroy_conn(conn);
3165 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3168 if (!list_empty(&kiblnd_data.kib_connd_conns)) {
3169 conn = list_entry(kiblnd_data.kib_connd_conns.next,
3170 kib_conn_t, ibc_list);
3171 list_del(&conn->ibc_list);
3173 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock,
3177 kiblnd_disconnect_conn(conn);
3178 kiblnd_conn_decref(conn);
3180 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3183 /* careful with the jiffy wrap... */
3184 timeout = (int)(deadline - jiffies);
3188 int chunk = kiblnd_data.kib_peer_hash_size;
3190 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
3193 /* Time to check for RDMA timeouts on a few more
3194 * peers: I do checks every 'p' seconds on a
3195 * proportion of the peer table and I need to check
3196 * every connection 'n' times within a timeout
3197 * interval, to ensure I detect a timeout on any
3198 * connection within (n+1)/n times the timeout
3201 if (*kiblnd_tunables.kib_timeout > n * p)
3202 chunk = (chunk * n * p) /
3203 *kiblnd_tunables.kib_timeout;
3207 for (i = 0; i < chunk; i++) {
3208 kiblnd_check_conns(peer_index);
3209 peer_index = (peer_index + 1) %
3210 kiblnd_data.kib_peer_hash_size;
3214 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3220 /* Nothing to do for 'timeout' */
3221 set_current_state(TASK_INTERRUPTIBLE);
3222 add_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3223 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
3225 waitq_timedwait(&wait, TASK_INTERRUPTIBLE, timeout);
3227 set_current_state(TASK_RUNNING);
3228 remove_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3229 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3232 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
3234 kiblnd_thread_fini();
3239 kiblnd_qp_event(struct ib_event *event, void *arg)
3241 kib_conn_t *conn = arg;
3243 switch (event->event) {
3244 case IB_EVENT_COMM_EST:
3245 CDEBUG(D_NET, "%s established\n",
3246 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3250 CERROR("%s: Async QP event type %d\n",
3251 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3257 kiblnd_complete (struct ib_wc *wc)
3259 switch (kiblnd_wreqid2type(wc->wr_id)) {
3263 case IBLND_WID_RDMA:
3264 /* We only get RDMA completion notification if it fails. All
3265 * subsequent work items, including the final SEND will fail
3266 * too. However we can't print out any more info about the
3267 * failing RDMA because 'tx' might be back on the idle list or
3268 * even reused already if we didn't manage to post all our work
3270 CNETERR("RDMA (tx: %p) failed: %d\n",
3271 kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3275 kiblnd_tx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3279 kiblnd_rx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status,
3286 kiblnd_cq_completion(struct ib_cq *cq, void *arg)
3288 /* NB I'm not allowed to schedule this conn once its refcount has
3289 * reached 0. Since fundamentally I'm racing with scheduler threads
3290 * consuming my CQ I could be called after all completions have
3291 * occurred. But in this case, ibc_nrx == 0 && ibc_nsends_posted == 0
3292 * and this CQ is about to be destroyed so I NOOP. */
3293 kib_conn_t *conn = (kib_conn_t *)arg;
3294 struct kib_sched_info *sched = conn->ibc_sched;
3295 unsigned long flags;
3297 LASSERT(cq == conn->ibc_cq);
3299 spin_lock_irqsave(&sched->ibs_lock, flags);
3301 conn->ibc_ready = 1;
3303 if (!conn->ibc_scheduled &&
3304 (conn->ibc_nrx > 0 ||
3305 conn->ibc_nsends_posted > 0)) {
3306 kiblnd_conn_addref(conn); /* +1 ref for sched_conns */
3307 conn->ibc_scheduled = 1;
3308 list_add_tail(&conn->ibc_sched_list, &sched->ibs_conns);
3310 if (waitqueue_active(&sched->ibs_waitq))
3311 wake_up(&sched->ibs_waitq);
3314 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3318 kiblnd_cq_event(struct ib_event *event, void *arg)
3320 kib_conn_t *conn = arg;
3322 CERROR("%s: async CQ event type %d\n",
3323 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3327 kiblnd_scheduler(void *arg)
3329 long id = (long)arg;
3330 struct kib_sched_info *sched;
3333 unsigned long flags;
3339 cfs_block_allsigs();
3341 init_waitqueue_entry_current(&wait);
3343 sched = kiblnd_data.kib_scheds[KIB_THREAD_CPT(id)];
3345 rc = cfs_cpt_bind(lnet_cpt_table(), sched->ibs_cpt);
3347 CWARN("Failed to bind on CPT %d, please verify whether "
3348 "all CPUs are healthy and reload modules if necessary, "
3349 "otherwise your system might under risk of low "
3350 "performance\n", sched->ibs_cpt);
3353 spin_lock_irqsave(&sched->ibs_lock, flags);
3355 while (!kiblnd_data.kib_shutdown) {
3356 if (busy_loops++ >= IBLND_RESCHED) {
3357 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3362 spin_lock_irqsave(&sched->ibs_lock, flags);
3367 if (!list_empty(&sched->ibs_conns)) {
3368 conn = list_entry(sched->ibs_conns.next,
3369 kib_conn_t, ibc_sched_list);
3370 /* take over kib_sched_conns' ref on conn... */
3371 LASSERT(conn->ibc_scheduled);
3372 list_del(&conn->ibc_sched_list);
3373 conn->ibc_ready = 0;
3375 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3377 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3379 rc = ib_req_notify_cq(conn->ibc_cq,
3382 CWARN("%s: ib_req_notify_cq failed: %d, "
3383 "closing connection\n",
3384 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
3385 kiblnd_close_conn(conn, -EIO);
3386 kiblnd_conn_decref(conn);
3387 spin_lock_irqsave(&sched->ibs_lock,
3392 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3396 CWARN("%s: ib_poll_cq failed: %d, "
3397 "closing connection\n",
3398 libcfs_nid2str(conn->ibc_peer->ibp_nid),
3400 kiblnd_close_conn(conn, -EIO);
3401 kiblnd_conn_decref(conn);
3402 spin_lock_irqsave(&sched->ibs_lock, flags);
3406 spin_lock_irqsave(&sched->ibs_lock, flags);
3408 if (rc != 0 || conn->ibc_ready) {
3409 /* There may be another completion waiting; get
3410 * another scheduler to check while I handle
3412 /* +1 ref for sched_conns */
3413 kiblnd_conn_addref(conn);
3414 list_add_tail(&conn->ibc_sched_list,
3416 if (waitqueue_active(&sched->ibs_waitq))
3417 wake_up(&sched->ibs_waitq);
3419 conn->ibc_scheduled = 0;
3423 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3424 kiblnd_complete(&wc);
3426 spin_lock_irqsave(&sched->ibs_lock, flags);
3429 kiblnd_conn_decref(conn); /* ...drop my ref from above */
3436 set_current_state(TASK_INTERRUPTIBLE);
3437 add_wait_queue_exclusive(&sched->ibs_waitq, &wait);
3438 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3440 waitq_wait(&wait, TASK_INTERRUPTIBLE);
3443 remove_wait_queue(&sched->ibs_waitq, &wait);
3444 set_current_state(TASK_RUNNING);
3445 spin_lock_irqsave(&sched->ibs_lock, flags);
3448 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3450 kiblnd_thread_fini();
3455 kiblnd_failover_thread(void *arg)
3457 rwlock_t *glock = &kiblnd_data.kib_global_lock;
3460 unsigned long flags;
3463 LASSERT (*kiblnd_tunables.kib_dev_failover != 0);
3465 cfs_block_allsigs ();
3467 init_waitqueue_entry_current(&wait);
3468 write_lock_irqsave(glock, flags);
3470 while (!kiblnd_data.kib_shutdown) {
3471 int do_failover = 0;
3474 list_for_each_entry(dev, &kiblnd_data.kib_failed_devs,
3476 if (cfs_time_before(cfs_time_current(),
3477 dev->ibd_next_failover))
3484 list_del_init(&dev->ibd_fail_list);
3485 dev->ibd_failover = 1;
3486 write_unlock_irqrestore(glock, flags);
3488 rc = kiblnd_dev_failover(dev);
3490 write_lock_irqsave(glock, flags);
3492 LASSERT (dev->ibd_failover);
3493 dev->ibd_failover = 0;
3494 if (rc >= 0) { /* Device is OK or failover succeed */
3495 dev->ibd_next_failover = cfs_time_shift(3);
3499 /* failed to failover, retry later */
3500 dev->ibd_next_failover =
3501 cfs_time_shift(min(dev->ibd_failed_failover, 10));
3502 if (kiblnd_dev_can_failover(dev)) {
3503 list_add_tail(&dev->ibd_fail_list,
3504 &kiblnd_data.kib_failed_devs);
3510 /* long sleep if no more pending failover */
3511 long_sleep = list_empty(&kiblnd_data.kib_failed_devs);
3513 set_current_state(TASK_INTERRUPTIBLE);
3514 add_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3515 write_unlock_irqrestore(glock, flags);
3517 rc = schedule_timeout(long_sleep ? cfs_time_seconds(10) :
3518 cfs_time_seconds(1));
3519 set_current_state(TASK_RUNNING);
3520 remove_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3521 write_lock_irqsave(glock, flags);
3523 if (!long_sleep || rc != 0)
3526 /* have a long sleep, routine check all active devices,
3527 * we need checking like this because if there is not active
3528 * connection on the dev and no SEND from local, we may listen
3529 * on wrong HCA for ever while there is a bonding failover */
3530 list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3531 if (kiblnd_dev_can_failover(dev)) {
3532 list_add_tail(&dev->ibd_fail_list,
3533 &kiblnd_data.kib_failed_devs);
3538 write_unlock_irqrestore(glock, flags);
3540 kiblnd_thread_fini();