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 (!cfs_in_interrupt());
53 LASSERT (!tx->tx_queued); /* mustn't be queued for sending */
54 LASSERT (tx->tx_sending == 0); /* mustn't be awaiting sent callback */
55 LASSERT (!tx->tx_waiting); /* mustn't be awaiting peer response */
56 LASSERT (tx->tx_pool != NULL);
58 kiblnd_unmap_tx(ni, tx);
60 /* tx may have up to 2 lnet msgs to finalise */
61 lntmsg[0] = tx->tx_lntmsg[0]; tx->tx_lntmsg[0] = NULL;
62 lntmsg[1] = tx->tx_lntmsg[1]; tx->tx_lntmsg[1] = NULL;
65 if (tx->tx_conn != NULL) {
66 LASSERT (ni == tx->tx_conn->ibc_peer->ibp_ni);
68 kiblnd_conn_decref(tx->tx_conn);
75 kiblnd_pool_free_node(&tx->tx_pool->tpo_pool, &tx->tx_list);
77 /* delay finalize until my descs have been freed */
78 for (i = 0; i < 2; i++) {
79 if (lntmsg[i] == NULL)
82 lnet_finalize(ni, lntmsg[i], rc);
87 kiblnd_txlist_done (lnet_ni_t *ni, cfs_list_t *txlist, int status)
91 while (!cfs_list_empty (txlist)) {
92 tx = cfs_list_entry (txlist->next, kib_tx_t, tx_list);
94 cfs_list_del(&tx->tx_list);
97 tx->tx_status = status;
98 kiblnd_tx_done(ni, tx);
103 kiblnd_get_idle_tx(lnet_ni_t *ni, lnet_nid_t target)
105 kib_net_t *net = (kib_net_t *)ni->ni_data;
108 kib_tx_poolset_t *tps;
110 tps = net->ibn_tx_ps[lnet_cpt_of_nid(target)];
111 node = kiblnd_pool_alloc_node(&tps->tps_poolset);
114 tx = container_of(node, kib_tx_t, tx_list);
116 LASSERT (tx->tx_nwrq == 0);
117 LASSERT (!tx->tx_queued);
118 LASSERT (tx->tx_sending == 0);
119 LASSERT (!tx->tx_waiting);
120 LASSERT (tx->tx_status == 0);
121 LASSERT (tx->tx_conn == NULL);
122 LASSERT (tx->tx_lntmsg[0] == NULL);
123 LASSERT (tx->tx_lntmsg[1] == NULL);
124 LASSERT (tx->tx_u.pmr == NULL);
125 LASSERT (tx->tx_nfrags == 0);
131 kiblnd_drop_rx(kib_rx_t *rx)
133 kib_conn_t *conn = rx->rx_conn;
134 struct kib_sched_info *sched = conn->ibc_sched;
137 spin_lock_irqsave(&sched->ibs_lock, flags);
138 LASSERT(conn->ibc_nrx > 0);
140 spin_unlock_irqrestore(&sched->ibs_lock, flags);
142 kiblnd_conn_decref(conn);
146 kiblnd_post_rx (kib_rx_t *rx, int credit)
148 kib_conn_t *conn = rx->rx_conn;
149 kib_net_t *net = conn->ibc_peer->ibp_ni->ni_data;
150 struct ib_recv_wr *bad_wrq = NULL;
154 LASSERT (net != NULL);
155 LASSERT (!cfs_in_interrupt());
156 LASSERT (credit == IBLND_POSTRX_NO_CREDIT ||
157 credit == IBLND_POSTRX_PEER_CREDIT ||
158 credit == IBLND_POSTRX_RSRVD_CREDIT);
160 mr = kiblnd_find_dma_mr(conn->ibc_hdev, rx->rx_msgaddr, IBLND_MSG_SIZE);
161 LASSERT (mr != NULL);
163 rx->rx_sge.lkey = mr->lkey;
164 rx->rx_sge.addr = rx->rx_msgaddr;
165 rx->rx_sge.length = IBLND_MSG_SIZE;
167 rx->rx_wrq.next = NULL;
168 rx->rx_wrq.sg_list = &rx->rx_sge;
169 rx->rx_wrq.num_sge = 1;
170 rx->rx_wrq.wr_id = kiblnd_ptr2wreqid(rx, IBLND_WID_RX);
172 LASSERT (conn->ibc_state >= IBLND_CONN_INIT);
173 LASSERT (rx->rx_nob >= 0); /* not posted */
175 if (conn->ibc_state > IBLND_CONN_ESTABLISHED) {
176 kiblnd_drop_rx(rx); /* No more posts for this rx */
180 rx->rx_nob = -1; /* flag posted */
182 rc = ib_post_recv(conn->ibc_cmid->qp, &rx->rx_wrq, &bad_wrq);
184 CERROR("Can't post rx for %s: %d, bad_wrq: %p\n",
185 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc, bad_wrq);
189 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) /* Initial post */
193 kiblnd_close_conn(conn, rc);
194 kiblnd_drop_rx(rx); /* No more posts for this rx */
198 if (credit == IBLND_POSTRX_NO_CREDIT)
201 spin_lock(&conn->ibc_lock);
202 if (credit == IBLND_POSTRX_PEER_CREDIT)
203 conn->ibc_outstanding_credits++;
205 conn->ibc_reserved_credits++;
206 spin_unlock(&conn->ibc_lock);
208 kiblnd_check_sends(conn);
213 kiblnd_find_waiting_tx_locked(kib_conn_t *conn, int txtype, __u64 cookie)
217 cfs_list_for_each(tmp, &conn->ibc_active_txs) {
218 kib_tx_t *tx = cfs_list_entry(tmp, kib_tx_t, tx_list);
220 LASSERT (!tx->tx_queued);
221 LASSERT (tx->tx_sending != 0 || tx->tx_waiting);
223 if (tx->tx_cookie != cookie)
226 if (tx->tx_waiting &&
227 tx->tx_msg->ibm_type == txtype)
230 CWARN("Bad completion: %swaiting, type %x (wanted %x)\n",
231 tx->tx_waiting ? "" : "NOT ",
232 tx->tx_msg->ibm_type, txtype);
238 kiblnd_handle_completion(kib_conn_t *conn, int txtype, int status, __u64 cookie)
241 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
244 spin_lock(&conn->ibc_lock);
246 tx = kiblnd_find_waiting_tx_locked(conn, txtype, cookie);
248 spin_unlock(&conn->ibc_lock);
250 CWARN("Unmatched completion type %x cookie "LPX64" from %s\n",
251 txtype, cookie, libcfs_nid2str(conn->ibc_peer->ibp_nid));
252 kiblnd_close_conn(conn, -EPROTO);
256 if (tx->tx_status == 0) { /* success so far */
257 if (status < 0) { /* failed? */
258 tx->tx_status = status;
259 } else if (txtype == IBLND_MSG_GET_REQ) {
260 lnet_set_reply_msg_len(ni, tx->tx_lntmsg[1], status);
266 idle = !tx->tx_queued && (tx->tx_sending == 0);
268 cfs_list_del(&tx->tx_list);
270 spin_unlock(&conn->ibc_lock);
273 kiblnd_tx_done(ni, tx);
277 kiblnd_send_completion(kib_conn_t *conn, int type, int status, __u64 cookie)
279 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
280 kib_tx_t *tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
283 CERROR("Can't get tx for completion %x for %s\n",
284 type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
288 tx->tx_msg->ibm_u.completion.ibcm_status = status;
289 tx->tx_msg->ibm_u.completion.ibcm_cookie = cookie;
290 kiblnd_init_tx_msg(ni, tx, type, sizeof(kib_completion_msg_t));
292 kiblnd_queue_tx(tx, conn);
296 kiblnd_handle_rx (kib_rx_t *rx)
298 kib_msg_t *msg = rx->rx_msg;
299 kib_conn_t *conn = rx->rx_conn;
300 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
301 int credits = msg->ibm_credits;
307 LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
309 CDEBUG (D_NET, "Received %x[%d] from %s\n",
310 msg->ibm_type, credits,
311 libcfs_nid2str(conn->ibc_peer->ibp_nid));
314 /* Have I received credits that will let me send? */
315 spin_lock(&conn->ibc_lock);
317 if (conn->ibc_credits + credits >
318 IBLND_MSG_QUEUE_SIZE(conn->ibc_version)) {
319 rc2 = conn->ibc_credits;
320 spin_unlock(&conn->ibc_lock);
322 CERROR("Bad credits from %s: %d + %d > %d\n",
323 libcfs_nid2str(conn->ibc_peer->ibp_nid),
325 IBLND_MSG_QUEUE_SIZE(conn->ibc_version));
327 kiblnd_close_conn(conn, -EPROTO);
328 kiblnd_post_rx(rx, IBLND_POSTRX_NO_CREDIT);
332 conn->ibc_credits += credits;
334 /* This ensures the credit taken by NOOP can be returned */
335 if (msg->ibm_type == IBLND_MSG_NOOP &&
336 !IBLND_OOB_CAPABLE(conn->ibc_version)) /* v1 only */
337 conn->ibc_outstanding_credits++;
339 spin_unlock(&conn->ibc_lock);
340 kiblnd_check_sends(conn);
343 switch (msg->ibm_type) {
345 CERROR("Bad IBLND message type %x from %s\n",
346 msg->ibm_type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
347 post_credit = IBLND_POSTRX_NO_CREDIT;
352 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
353 post_credit = IBLND_POSTRX_NO_CREDIT;
357 if (credits != 0) /* credit already posted */
358 post_credit = IBLND_POSTRX_NO_CREDIT;
359 else /* a keepalive NOOP */
360 post_credit = IBLND_POSTRX_PEER_CREDIT;
363 case IBLND_MSG_IMMEDIATE:
364 post_credit = IBLND_POSTRX_DONT_POST;
365 rc = lnet_parse(ni, &msg->ibm_u.immediate.ibim_hdr,
366 msg->ibm_srcnid, rx, 0);
367 if (rc < 0) /* repost on error */
368 post_credit = IBLND_POSTRX_PEER_CREDIT;
371 case IBLND_MSG_PUT_REQ:
372 post_credit = IBLND_POSTRX_DONT_POST;
373 rc = lnet_parse(ni, &msg->ibm_u.putreq.ibprm_hdr,
374 msg->ibm_srcnid, rx, 1);
375 if (rc < 0) /* repost on error */
376 post_credit = IBLND_POSTRX_PEER_CREDIT;
379 case IBLND_MSG_PUT_NAK:
380 CWARN ("PUT_NACK from %s\n",
381 libcfs_nid2str(conn->ibc_peer->ibp_nid));
382 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
383 kiblnd_handle_completion(conn, IBLND_MSG_PUT_REQ,
384 msg->ibm_u.completion.ibcm_status,
385 msg->ibm_u.completion.ibcm_cookie);
388 case IBLND_MSG_PUT_ACK:
389 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
391 spin_lock(&conn->ibc_lock);
392 tx = kiblnd_find_waiting_tx_locked(conn, IBLND_MSG_PUT_REQ,
393 msg->ibm_u.putack.ibpam_src_cookie);
395 cfs_list_del(&tx->tx_list);
396 spin_unlock(&conn->ibc_lock);
399 CERROR("Unmatched PUT_ACK from %s\n",
400 libcfs_nid2str(conn->ibc_peer->ibp_nid));
405 LASSERT (tx->tx_waiting);
406 /* CAVEAT EMPTOR: I could be racing with tx_complete, but...
407 * (a) I can overwrite tx_msg since my peer has received it!
408 * (b) tx_waiting set tells tx_complete() it's not done. */
410 tx->tx_nwrq = 0; /* overwrite PUT_REQ */
412 rc2 = kiblnd_init_rdma(conn, tx, IBLND_MSG_PUT_DONE,
413 kiblnd_rd_size(&msg->ibm_u.putack.ibpam_rd),
414 &msg->ibm_u.putack.ibpam_rd,
415 msg->ibm_u.putack.ibpam_dst_cookie);
417 CERROR("Can't setup rdma for PUT to %s: %d\n",
418 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc2);
420 spin_lock(&conn->ibc_lock);
421 tx->tx_waiting = 0; /* clear waiting and queue atomically */
422 kiblnd_queue_tx_locked(tx, conn);
423 spin_unlock(&conn->ibc_lock);
426 case IBLND_MSG_PUT_DONE:
427 post_credit = IBLND_POSTRX_PEER_CREDIT;
428 kiblnd_handle_completion(conn, IBLND_MSG_PUT_ACK,
429 msg->ibm_u.completion.ibcm_status,
430 msg->ibm_u.completion.ibcm_cookie);
433 case IBLND_MSG_GET_REQ:
434 post_credit = IBLND_POSTRX_DONT_POST;
435 rc = lnet_parse(ni, &msg->ibm_u.get.ibgm_hdr,
436 msg->ibm_srcnid, rx, 1);
437 if (rc < 0) /* repost on error */
438 post_credit = IBLND_POSTRX_PEER_CREDIT;
441 case IBLND_MSG_GET_DONE:
442 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
443 kiblnd_handle_completion(conn, IBLND_MSG_GET_REQ,
444 msg->ibm_u.completion.ibcm_status,
445 msg->ibm_u.completion.ibcm_cookie);
449 if (rc < 0) /* protocol error */
450 kiblnd_close_conn(conn, rc);
452 if (post_credit != IBLND_POSTRX_DONT_POST)
453 kiblnd_post_rx(rx, post_credit);
457 kiblnd_rx_complete (kib_rx_t *rx, int status, int nob)
459 kib_msg_t *msg = rx->rx_msg;
460 kib_conn_t *conn = rx->rx_conn;
461 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
462 kib_net_t *net = ni->ni_data;
466 LASSERT (net != NULL);
467 LASSERT (rx->rx_nob < 0); /* was posted */
468 rx->rx_nob = 0; /* isn't now */
470 if (conn->ibc_state > IBLND_CONN_ESTABLISHED)
473 if (status != IB_WC_SUCCESS) {
474 CNETERR("Rx from %s failed: %d\n",
475 libcfs_nid2str(conn->ibc_peer->ibp_nid), status);
482 rc = kiblnd_unpack_msg(msg, rx->rx_nob);
484 CERROR ("Error %d unpacking rx from %s\n",
485 rc, libcfs_nid2str(conn->ibc_peer->ibp_nid));
489 if (msg->ibm_srcnid != conn->ibc_peer->ibp_nid ||
490 msg->ibm_dstnid != ni->ni_nid ||
491 msg->ibm_srcstamp != conn->ibc_incarnation ||
492 msg->ibm_dststamp != net->ibn_incarnation) {
493 CERROR ("Stale rx from %s\n",
494 libcfs_nid2str(conn->ibc_peer->ibp_nid));
499 /* set time last known alive */
500 kiblnd_peer_alive(conn->ibc_peer);
502 /* racing with connection establishment/teardown! */
504 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
505 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
508 write_lock_irqsave(g_lock, flags);
509 /* must check holding global lock to eliminate race */
510 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
511 cfs_list_add_tail(&rx->rx_list, &conn->ibc_early_rxs);
512 write_unlock_irqrestore(g_lock, flags);
515 write_unlock_irqrestore(g_lock, flags);
517 kiblnd_handle_rx(rx);
521 CDEBUG(D_NET, "rx %p conn %p\n", rx, conn);
522 kiblnd_close_conn(conn, err);
524 kiblnd_drop_rx(rx); /* Don't re-post rx. */
528 kiblnd_kvaddr_to_page (unsigned long vaddr)
532 if (vaddr >= VMALLOC_START &&
533 vaddr < VMALLOC_END) {
534 page = vmalloc_to_page ((void *)vaddr);
535 LASSERT (page != NULL);
538 #ifdef CONFIG_HIGHMEM
539 if (vaddr >= PKMAP_BASE &&
540 vaddr < (PKMAP_BASE + LAST_PKMAP * PAGE_SIZE)) {
541 /* No highmem pages only used for bulk (kiov) I/O */
542 CERROR("find page for address in highmem\n");
546 page = virt_to_page (vaddr);
547 LASSERT (page != NULL);
552 kiblnd_fmr_map_tx(kib_net_t *net, kib_tx_t *tx, kib_rdma_desc_t *rd, int nob)
554 kib_hca_dev_t *hdev = tx->tx_pool->tpo_hdev;
555 __u64 *pages = tx->tx_pages;
556 kib_fmr_poolset_t *fps;
563 for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
564 for (size = 0; size < rd->rd_frags[i].rf_nob;
565 size += hdev->ibh_page_size) {
566 pages[npages ++] = (rd->rd_frags[i].rf_addr &
567 hdev->ibh_page_mask) + size;
571 LASSERT(tx->tx_pool != NULL);
572 LASSERT(tx->tx_pool->tpo_pool.po_owner != NULL);
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->tx_u.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->tx_u.fmr.fmr_pfmr->fmr->rkey :
586 tx->tx_u.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_pmr_map_tx(kib_net_t *net, kib_tx_t *tx, kib_rdma_desc_t *rd, int nob)
597 kib_hca_dev_t *hdev = tx->tx_pool->tpo_hdev;
598 kib_pmr_poolset_t *pps;
603 iova = rd->rd_frags[0].rf_addr & ~hdev->ibh_page_mask;
605 LASSERT(tx->tx_pool != NULL);
606 LASSERT(tx->tx_pool->tpo_pool.po_owner != NULL);
608 cpt = tx->tx_pool->tpo_pool.po_owner->ps_cpt;
610 pps = net->ibn_pmr_ps[cpt];
611 rc = kiblnd_pmr_pool_map(pps, hdev, rd, &iova, &tx->tx_u.pmr);
613 CERROR("Failed to create MR by phybuf: %d\n", rc);
617 /* If rd is not tx_rd, it's going to get sent to a peer, who will need
619 rd->rd_key = (rd != tx->tx_rd) ? tx->tx_u.pmr->pmr_mr->rkey :
620 tx->tx_u.pmr->pmr_mr->lkey;
622 rd->rd_frags[0].rf_addr = iova;
623 rd->rd_frags[0].rf_nob = nob;
629 kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx)
631 kib_net_t *net = ni->ni_data;
633 LASSERT(net != NULL);
635 if (net->ibn_fmr_ps != NULL && tx->tx_u.fmr.fmr_pfmr != NULL) {
636 kiblnd_fmr_pool_unmap(&tx->tx_u.fmr, tx->tx_status);
637 tx->tx_u.fmr.fmr_pfmr = NULL;
639 } else if (net->ibn_pmr_ps != NULL && tx->tx_u.pmr != NULL) {
640 kiblnd_pmr_pool_unmap(tx->tx_u.pmr);
644 if (tx->tx_nfrags != 0) {
645 kiblnd_dma_unmap_sg(tx->tx_pool->tpo_hdev->ibh_ibdev,
646 tx->tx_frags, tx->tx_nfrags, tx->tx_dmadir);
652 kiblnd_map_tx(lnet_ni_t *ni, kib_tx_t *tx,
653 kib_rdma_desc_t *rd, int nfrags)
655 kib_hca_dev_t *hdev = tx->tx_pool->tpo_hdev;
656 kib_net_t *net = ni->ni_data;
657 struct ib_mr *mr = NULL;
661 /* If rd is not tx_rd, it's going to get sent to a peer and I'm the
663 tx->tx_dmadir = (rd != tx->tx_rd) ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
664 tx->tx_nfrags = nfrags;
667 kiblnd_dma_map_sg(hdev->ibh_ibdev,
668 tx->tx_frags, tx->tx_nfrags, tx->tx_dmadir);
670 for (i = 0, nob = 0; i < rd->rd_nfrags; i++) {
671 rd->rd_frags[i].rf_nob = kiblnd_sg_dma_len(
672 hdev->ibh_ibdev, &tx->tx_frags[i]);
673 rd->rd_frags[i].rf_addr = kiblnd_sg_dma_address(
674 hdev->ibh_ibdev, &tx->tx_frags[i]);
675 nob += rd->rd_frags[i].rf_nob;
678 /* looking for pre-mapping MR */
679 mr = kiblnd_find_rd_dma_mr(hdev, rd);
681 /* found pre-mapping MR */
682 rd->rd_key = (rd != tx->tx_rd) ? mr->rkey : mr->lkey;
686 if (net->ibn_fmr_ps != NULL)
687 return kiblnd_fmr_map_tx(net, tx, rd, nob);
688 else if (net->ibn_pmr_ps != NULL)
689 return kiblnd_pmr_map_tx(net, tx, rd, nob);
696 kiblnd_setup_rd_iov(lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd,
697 unsigned int niov, struct iovec *iov, int offset, int nob)
699 kib_net_t *net = ni->ni_data;
701 struct scatterlist *sg;
708 LASSERT (net != NULL);
710 while (offset >= iov->iov_len) {
711 offset -= iov->iov_len;
721 vaddr = ((unsigned long)iov->iov_base) + offset;
722 page_offset = vaddr & (PAGE_SIZE - 1);
723 page = kiblnd_kvaddr_to_page(vaddr);
725 CERROR ("Can't find page\n");
729 fragnob = min((int)(iov->iov_len - offset), nob);
730 fragnob = min(fragnob, (int)PAGE_SIZE - page_offset);
732 sg_set_page(sg, page, fragnob, page_offset);
735 if (offset + fragnob < iov->iov_len) {
745 return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags);
749 kiblnd_setup_rd_kiov (lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd,
750 int nkiov, lnet_kiov_t *kiov, int offset, int nob)
752 kib_net_t *net = ni->ni_data;
753 struct scatterlist *sg;
756 CDEBUG(D_NET, "niov %d offset %d nob %d\n", nkiov, offset, nob);
760 LASSERT (net != NULL);
762 while (offset >= kiov->kiov_len) {
763 offset -= kiov->kiov_len;
773 fragnob = min((int)(kiov->kiov_len - offset), nob);
775 sg_set_page(sg, kiov->kiov_page, fragnob,
776 kiov->kiov_offset + offset);
785 return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags);
789 kiblnd_post_tx_locked (kib_conn_t *conn, kib_tx_t *tx, int credit)
791 kib_msg_t *msg = tx->tx_msg;
792 kib_peer_t *peer = conn->ibc_peer;
793 int ver = conn->ibc_version;
796 struct ib_send_wr *bad_wrq;
798 LASSERT (tx->tx_queued);
799 /* We rely on this for QP sizing */
800 LASSERT (tx->tx_nwrq > 0);
801 LASSERT (tx->tx_nwrq <= 1 + IBLND_RDMA_FRAGS(ver));
803 LASSERT (credit == 0 || credit == 1);
804 LASSERT (conn->ibc_outstanding_credits >= 0);
805 LASSERT (conn->ibc_outstanding_credits <= IBLND_MSG_QUEUE_SIZE(ver));
806 LASSERT (conn->ibc_credits >= 0);
807 LASSERT (conn->ibc_credits <= IBLND_MSG_QUEUE_SIZE(ver));
809 if (conn->ibc_nsends_posted == IBLND_CONCURRENT_SENDS(ver)) {
810 /* tx completions outstanding... */
811 CDEBUG(D_NET, "%s: posted enough\n",
812 libcfs_nid2str(peer->ibp_nid));
816 if (credit != 0 && conn->ibc_credits == 0) { /* no credits */
817 CDEBUG(D_NET, "%s: no credits\n",
818 libcfs_nid2str(peer->ibp_nid));
822 if (credit != 0 && !IBLND_OOB_CAPABLE(ver) &&
823 conn->ibc_credits == 1 && /* last credit reserved */
824 msg->ibm_type != IBLND_MSG_NOOP) { /* for NOOP */
825 CDEBUG(D_NET, "%s: not using last credit\n",
826 libcfs_nid2str(peer->ibp_nid));
830 /* NB don't drop ibc_lock before bumping tx_sending */
831 cfs_list_del(&tx->tx_list);
834 if (msg->ibm_type == IBLND_MSG_NOOP &&
835 (!kiblnd_need_noop(conn) || /* redundant NOOP */
836 (IBLND_OOB_CAPABLE(ver) && /* posted enough NOOP */
837 conn->ibc_noops_posted == IBLND_OOB_MSGS(ver)))) {
838 /* OK to drop when posted enough NOOPs, since
839 * kiblnd_check_sends will queue NOOP again when
840 * posted NOOPs complete */
841 spin_unlock(&conn->ibc_lock);
842 kiblnd_tx_done(peer->ibp_ni, tx);
843 spin_lock(&conn->ibc_lock);
844 CDEBUG(D_NET, "%s(%d): redundant or enough NOOP\n",
845 libcfs_nid2str(peer->ibp_nid),
846 conn->ibc_noops_posted);
850 kiblnd_pack_msg(peer->ibp_ni, msg, ver, conn->ibc_outstanding_credits,
851 peer->ibp_nid, conn->ibc_incarnation);
853 conn->ibc_credits -= credit;
854 conn->ibc_outstanding_credits = 0;
855 conn->ibc_nsends_posted++;
856 if (msg->ibm_type == IBLND_MSG_NOOP)
857 conn->ibc_noops_posted++;
859 /* CAVEAT EMPTOR! This tx could be the PUT_DONE of an RDMA
860 * PUT. If so, it was first queued here as a PUT_REQ, sent and
861 * stashed on ibc_active_txs, matched by an incoming PUT_ACK,
862 * and then re-queued here. It's (just) possible that
863 * tx_sending is non-zero if we've not done the tx_complete()
864 * from the first send; hence the ++ rather than = below. */
866 cfs_list_add(&tx->tx_list, &conn->ibc_active_txs);
868 /* I'm still holding ibc_lock! */
869 if (conn->ibc_state != IBLND_CONN_ESTABLISHED) {
871 } else if (tx->tx_pool->tpo_pool.po_failed ||
872 conn->ibc_hdev != tx->tx_pool->tpo_hdev) {
873 /* close_conn will launch failover */
876 rc = ib_post_send(conn->ibc_cmid->qp,
877 tx->tx_wrq, &bad_wrq);
880 conn->ibc_last_send = jiffies;
885 /* NB credits are transferred in the actual
886 * message, which can only be the last work item */
887 conn->ibc_credits += credit;
888 conn->ibc_outstanding_credits += msg->ibm_credits;
889 conn->ibc_nsends_posted--;
890 if (msg->ibm_type == IBLND_MSG_NOOP)
891 conn->ibc_noops_posted--;
897 done = (tx->tx_sending == 0);
899 cfs_list_del(&tx->tx_list);
901 spin_unlock(&conn->ibc_lock);
903 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
904 CERROR("Error %d posting transmit to %s\n",
905 rc, libcfs_nid2str(peer->ibp_nid));
907 CDEBUG(D_NET, "Error %d posting transmit to %s\n",
908 rc, libcfs_nid2str(peer->ibp_nid));
910 kiblnd_close_conn(conn, rc);
913 kiblnd_tx_done(peer->ibp_ni, tx);
915 spin_lock(&conn->ibc_lock);
921 kiblnd_check_sends (kib_conn_t *conn)
923 int ver = conn->ibc_version;
924 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
927 /* Don't send anything until after the connection is established */
928 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
929 CDEBUG(D_NET, "%s too soon\n",
930 libcfs_nid2str(conn->ibc_peer->ibp_nid));
934 spin_lock(&conn->ibc_lock);
936 LASSERT (conn->ibc_nsends_posted <= IBLND_CONCURRENT_SENDS(ver));
937 LASSERT (!IBLND_OOB_CAPABLE(ver) ||
938 conn->ibc_noops_posted <= IBLND_OOB_MSGS(ver));
939 LASSERT (conn->ibc_reserved_credits >= 0);
941 while (conn->ibc_reserved_credits > 0 &&
942 !cfs_list_empty(&conn->ibc_tx_queue_rsrvd)) {
943 tx = cfs_list_entry(conn->ibc_tx_queue_rsrvd.next,
945 cfs_list_del(&tx->tx_list);
946 cfs_list_add_tail(&tx->tx_list, &conn->ibc_tx_queue);
947 conn->ibc_reserved_credits--;
950 if (kiblnd_need_noop(conn)) {
951 spin_unlock(&conn->ibc_lock);
953 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
955 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_NOOP, 0);
957 spin_lock(&conn->ibc_lock);
959 kiblnd_queue_tx_locked(tx, conn);
962 kiblnd_conn_addref(conn); /* 1 ref for me.... (see b21911) */
967 if (!cfs_list_empty(&conn->ibc_tx_queue_nocred)) {
969 tx = cfs_list_entry(conn->ibc_tx_queue_nocred.next,
971 } else if (!cfs_list_empty(&conn->ibc_tx_noops)) {
972 LASSERT (!IBLND_OOB_CAPABLE(ver));
974 tx = cfs_list_entry(conn->ibc_tx_noops.next,
976 } else if (!cfs_list_empty(&conn->ibc_tx_queue)) {
978 tx = cfs_list_entry(conn->ibc_tx_queue.next,
983 if (kiblnd_post_tx_locked(conn, tx, credit) != 0)
987 spin_unlock(&conn->ibc_lock);
989 kiblnd_conn_decref(conn); /* ...until here */
993 kiblnd_tx_complete (kib_tx_t *tx, int status)
995 int failed = (status != IB_WC_SUCCESS);
996 kib_conn_t *conn = tx->tx_conn;
999 LASSERT (tx->tx_sending > 0);
1002 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
1003 CNETERR("Tx -> %s cookie "LPX64
1004 " sending %d waiting %d: failed %d\n",
1005 libcfs_nid2str(conn->ibc_peer->ibp_nid),
1006 tx->tx_cookie, tx->tx_sending, tx->tx_waiting,
1009 kiblnd_close_conn(conn, -EIO);
1011 kiblnd_peer_alive(conn->ibc_peer);
1014 spin_lock(&conn->ibc_lock);
1016 /* I could be racing with rdma completion. Whoever makes 'tx' idle
1017 * gets to free it, which also drops its ref on 'conn'. */
1020 conn->ibc_nsends_posted--;
1021 if (tx->tx_msg->ibm_type == IBLND_MSG_NOOP)
1022 conn->ibc_noops_posted--;
1025 tx->tx_waiting = 0; /* don't wait for peer */
1026 tx->tx_status = -EIO;
1029 idle = (tx->tx_sending == 0) && /* This is the final callback */
1030 !tx->tx_waiting && /* Not waiting for peer */
1031 !tx->tx_queued; /* Not re-queued (PUT_DONE) */
1033 cfs_list_del(&tx->tx_list);
1035 kiblnd_conn_addref(conn); /* 1 ref for me.... */
1037 spin_unlock(&conn->ibc_lock);
1040 kiblnd_tx_done(conn->ibc_peer->ibp_ni, tx);
1042 kiblnd_check_sends(conn);
1044 kiblnd_conn_decref(conn); /* ...until here */
1048 kiblnd_init_tx_msg (lnet_ni_t *ni, kib_tx_t *tx, int type, int body_nob)
1050 kib_hca_dev_t *hdev = tx->tx_pool->tpo_hdev;
1051 struct ib_sge *sge = &tx->tx_sge[tx->tx_nwrq];
1052 struct ib_send_wr *wrq = &tx->tx_wrq[tx->tx_nwrq];
1053 int nob = offsetof (kib_msg_t, ibm_u) + body_nob;
1056 LASSERT (tx->tx_nwrq >= 0);
1057 LASSERT (tx->tx_nwrq < IBLND_MAX_RDMA_FRAGS + 1);
1058 LASSERT (nob <= IBLND_MSG_SIZE);
1060 kiblnd_init_msg(tx->tx_msg, type, body_nob);
1062 mr = kiblnd_find_dma_mr(hdev, tx->tx_msgaddr, nob);
1063 LASSERT (mr != NULL);
1065 sge->lkey = mr->lkey;
1066 sge->addr = tx->tx_msgaddr;
1069 memset(wrq, 0, sizeof(*wrq));
1072 wrq->wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_TX);
1075 wrq->opcode = IB_WR_SEND;
1076 wrq->send_flags = IB_SEND_SIGNALED;
1082 kiblnd_init_rdma (kib_conn_t *conn, kib_tx_t *tx, int type,
1083 int resid, kib_rdma_desc_t *dstrd, __u64 dstcookie)
1085 kib_msg_t *ibmsg = tx->tx_msg;
1086 kib_rdma_desc_t *srcrd = tx->tx_rd;
1087 struct ib_sge *sge = &tx->tx_sge[0];
1088 struct ib_send_wr *wrq = &tx->tx_wrq[0];
1094 LASSERT (!cfs_in_interrupt());
1095 LASSERT (tx->tx_nwrq == 0);
1096 LASSERT (type == IBLND_MSG_GET_DONE ||
1097 type == IBLND_MSG_PUT_DONE);
1099 srcidx = dstidx = 0;
1102 if (srcidx >= srcrd->rd_nfrags) {
1103 CERROR("Src buffer exhausted: %d frags\n", srcidx);
1108 if (dstidx == dstrd->rd_nfrags) {
1109 CERROR("Dst buffer exhausted: %d frags\n", dstidx);
1114 if (tx->tx_nwrq == IBLND_RDMA_FRAGS(conn->ibc_version)) {
1115 CERROR("RDMA too fragmented for %s (%d): "
1116 "%d/%d src %d/%d dst frags\n",
1117 libcfs_nid2str(conn->ibc_peer->ibp_nid),
1118 IBLND_RDMA_FRAGS(conn->ibc_version),
1119 srcidx, srcrd->rd_nfrags,
1120 dstidx, dstrd->rd_nfrags);
1125 wrknob = MIN(MIN(kiblnd_rd_frag_size(srcrd, srcidx),
1126 kiblnd_rd_frag_size(dstrd, dstidx)), resid);
1128 sge = &tx->tx_sge[tx->tx_nwrq];
1129 sge->addr = kiblnd_rd_frag_addr(srcrd, srcidx);
1130 sge->lkey = kiblnd_rd_frag_key(srcrd, srcidx);
1131 sge->length = wrknob;
1133 wrq = &tx->tx_wrq[tx->tx_nwrq];
1135 wrq->next = wrq + 1;
1136 wrq->wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_RDMA);
1139 wrq->opcode = IB_WR_RDMA_WRITE;
1140 wrq->send_flags = 0;
1142 wrq->wr.rdma.remote_addr = kiblnd_rd_frag_addr(dstrd, dstidx);
1143 wrq->wr.rdma.rkey = kiblnd_rd_frag_key(dstrd, dstidx);
1145 srcidx = kiblnd_rd_consume_frag(srcrd, srcidx, wrknob);
1146 dstidx = kiblnd_rd_consume_frag(dstrd, dstidx, wrknob);
1155 if (rc < 0) /* no RDMA if completing with failure */
1158 ibmsg->ibm_u.completion.ibcm_status = rc;
1159 ibmsg->ibm_u.completion.ibcm_cookie = dstcookie;
1160 kiblnd_init_tx_msg(conn->ibc_peer->ibp_ni, tx,
1161 type, sizeof (kib_completion_msg_t));
1167 kiblnd_queue_tx_locked (kib_tx_t *tx, kib_conn_t *conn)
1171 LASSERT (tx->tx_nwrq > 0); /* work items set up */
1172 LASSERT (!tx->tx_queued); /* not queued for sending already */
1173 LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1176 tx->tx_deadline = jiffies + (*kiblnd_tunables.kib_timeout * CFS_HZ);
1178 if (tx->tx_conn == NULL) {
1179 kiblnd_conn_addref(conn);
1181 LASSERT (tx->tx_msg->ibm_type != IBLND_MSG_PUT_DONE);
1183 /* PUT_DONE first attached to conn as a PUT_REQ */
1184 LASSERT (tx->tx_conn == conn);
1185 LASSERT (tx->tx_msg->ibm_type == IBLND_MSG_PUT_DONE);
1188 switch (tx->tx_msg->ibm_type) {
1192 case IBLND_MSG_PUT_REQ:
1193 case IBLND_MSG_GET_REQ:
1194 q = &conn->ibc_tx_queue_rsrvd;
1197 case IBLND_MSG_PUT_NAK:
1198 case IBLND_MSG_PUT_ACK:
1199 case IBLND_MSG_PUT_DONE:
1200 case IBLND_MSG_GET_DONE:
1201 q = &conn->ibc_tx_queue_nocred;
1204 case IBLND_MSG_NOOP:
1205 if (IBLND_OOB_CAPABLE(conn->ibc_version))
1206 q = &conn->ibc_tx_queue_nocred;
1208 q = &conn->ibc_tx_noops;
1211 case IBLND_MSG_IMMEDIATE:
1212 q = &conn->ibc_tx_queue;
1216 cfs_list_add_tail(&tx->tx_list, q);
1220 kiblnd_queue_tx (kib_tx_t *tx, kib_conn_t *conn)
1222 spin_lock(&conn->ibc_lock);
1223 kiblnd_queue_tx_locked(tx, conn);
1224 spin_unlock(&conn->ibc_lock);
1226 kiblnd_check_sends(conn);
1229 static int kiblnd_resolve_addr(struct rdma_cm_id *cmid,
1230 struct sockaddr_in *srcaddr,
1231 struct sockaddr_in *dstaddr,
1234 unsigned short port;
1237 #ifdef HAVE_OFED_RDMA_SET_REUSEADDR
1238 /* allow the port to be reused */
1239 rc = rdma_set_reuseaddr(cmid, 1);
1241 CERROR("Unable to set reuse on cmid: %d\n", rc);
1246 /* look for a free privileged port */
1247 for (port = PROT_SOCK-1; port > 0; port--) {
1248 srcaddr->sin_port = htons(port);
1249 rc = rdma_resolve_addr(cmid,
1250 (struct sockaddr *)srcaddr,
1251 (struct sockaddr *)dstaddr,
1254 CDEBUG(D_NET, "bound to port %hu\n", port);
1256 } else if (rc == -EADDRINUSE || rc == -EADDRNOTAVAIL) {
1257 CDEBUG(D_NET, "bind to port %hu failed: %d\n",
1264 CERROR("Failed to bind to a free privileged port\n");
1265 #ifndef HAVE_OFED_RDMA_SET_REUSEADDR
1266 CERROR("You may need IB verbs that supports rdma_set_reuseaddr()\n");
1272 kiblnd_connect_peer (kib_peer_t *peer)
1274 struct rdma_cm_id *cmid;
1276 kib_net_t *net = peer->ibp_ni->ni_data;
1277 struct sockaddr_in srcaddr;
1278 struct sockaddr_in dstaddr;
1281 LASSERT (net != NULL);
1282 LASSERT (peer->ibp_connecting > 0);
1284 cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, peer, RDMA_PS_TCP,
1288 CERROR("Can't create CMID for %s: %ld\n",
1289 libcfs_nid2str(peer->ibp_nid), PTR_ERR(cmid));
1295 memset(&srcaddr, 0, sizeof(srcaddr));
1296 srcaddr.sin_family = AF_INET;
1297 srcaddr.sin_addr.s_addr = htonl(dev->ibd_ifip);
1299 memset(&dstaddr, 0, sizeof(dstaddr));
1300 dstaddr.sin_family = AF_INET;
1301 dstaddr.sin_port = htons(*kiblnd_tunables.kib_service);
1302 dstaddr.sin_addr.s_addr = htonl(LNET_NIDADDR(peer->ibp_nid));
1304 kiblnd_peer_addref(peer); /* cmid's ref */
1306 if (*kiblnd_tunables.kib_use_priv_port) {
1307 rc = kiblnd_resolve_addr(cmid, &srcaddr, &dstaddr,
1308 *kiblnd_tunables.kib_timeout * 1000);
1310 rc = rdma_resolve_addr(cmid,
1311 (struct sockaddr *)&srcaddr,
1312 (struct sockaddr *)&dstaddr,
1313 *kiblnd_tunables.kib_timeout * 1000);
1316 /* Can't initiate address resolution: */
1317 CERROR("Can't resolve addr for %s: %d\n",
1318 libcfs_nid2str(peer->ibp_nid), rc);
1322 LASSERT (cmid->device != NULL);
1323 CDEBUG(D_NET, "%s: connection bound to %s:%u.%u.%u.%u:%s\n",
1324 libcfs_nid2str(peer->ibp_nid), dev->ibd_ifname,
1325 HIPQUAD(dev->ibd_ifip), cmid->device->name);
1330 kiblnd_peer_decref(peer); /* cmid's ref */
1331 rdma_destroy_id(cmid);
1333 kiblnd_peer_connect_failed(peer, 1, rc);
1337 kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid)
1342 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
1343 unsigned long flags;
1346 /* If I get here, I've committed to send, so I complete the tx with
1347 * failure on any problems */
1349 LASSERT (tx == NULL || tx->tx_conn == NULL); /* only set when assigned a conn */
1350 LASSERT (tx == NULL || tx->tx_nwrq > 0); /* work items have been set up */
1352 /* First time, just use a read lock since I expect to find my peer
1354 read_lock_irqsave(g_lock, flags);
1356 peer = kiblnd_find_peer_locked(nid);
1357 if (peer != NULL && !cfs_list_empty(&peer->ibp_conns)) {
1358 /* Found a peer with an established connection */
1359 conn = kiblnd_get_conn_locked(peer);
1360 kiblnd_conn_addref(conn); /* 1 ref for me... */
1362 read_unlock_irqrestore(g_lock, flags);
1365 kiblnd_queue_tx(tx, conn);
1366 kiblnd_conn_decref(conn); /* ...to here */
1370 read_unlock(g_lock);
1371 /* Re-try with a write lock */
1374 peer = kiblnd_find_peer_locked(nid);
1376 if (cfs_list_empty(&peer->ibp_conns)) {
1377 /* found a peer, but it's still connecting... */
1378 LASSERT (peer->ibp_connecting != 0 ||
1379 peer->ibp_accepting != 0);
1381 cfs_list_add_tail(&tx->tx_list,
1382 &peer->ibp_tx_queue);
1383 write_unlock_irqrestore(g_lock, flags);
1385 conn = kiblnd_get_conn_locked(peer);
1386 kiblnd_conn_addref(conn); /* 1 ref for me... */
1388 write_unlock_irqrestore(g_lock, flags);
1391 kiblnd_queue_tx(tx, conn);
1392 kiblnd_conn_decref(conn); /* ...to here */
1397 write_unlock_irqrestore(g_lock, flags);
1399 /* Allocate a peer ready to add to the peer table and retry */
1400 rc = kiblnd_create_peer(ni, &peer, nid);
1402 CERROR("Can't create peer %s\n", libcfs_nid2str(nid));
1404 tx->tx_status = -EHOSTUNREACH;
1406 kiblnd_tx_done(ni, tx);
1411 write_lock_irqsave(g_lock, flags);
1413 peer2 = kiblnd_find_peer_locked(nid);
1414 if (peer2 != NULL) {
1415 if (cfs_list_empty(&peer2->ibp_conns)) {
1416 /* found a peer, but it's still connecting... */
1417 LASSERT (peer2->ibp_connecting != 0 ||
1418 peer2->ibp_accepting != 0);
1420 cfs_list_add_tail(&tx->tx_list,
1421 &peer2->ibp_tx_queue);
1422 write_unlock_irqrestore(g_lock, flags);
1424 conn = kiblnd_get_conn_locked(peer2);
1425 kiblnd_conn_addref(conn); /* 1 ref for me... */
1427 write_unlock_irqrestore(g_lock, flags);
1430 kiblnd_queue_tx(tx, conn);
1431 kiblnd_conn_decref(conn); /* ...to here */
1434 kiblnd_peer_decref(peer);
1438 /* Brand new peer */
1439 LASSERT (peer->ibp_connecting == 0);
1440 peer->ibp_connecting = 1;
1442 /* always called with a ref on ni, which prevents ni being shutdown */
1443 LASSERT (((kib_net_t *)ni->ni_data)->ibn_shutdown == 0);
1446 cfs_list_add_tail(&tx->tx_list, &peer->ibp_tx_queue);
1448 kiblnd_peer_addref(peer);
1449 cfs_list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid));
1451 write_unlock_irqrestore(g_lock, flags);
1453 kiblnd_connect_peer(peer);
1454 kiblnd_peer_decref(peer);
1458 kiblnd_send (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg)
1460 lnet_hdr_t *hdr = &lntmsg->msg_hdr;
1461 int type = lntmsg->msg_type;
1462 lnet_process_id_t target = lntmsg->msg_target;
1463 int target_is_router = lntmsg->msg_target_is_router;
1464 int routing = lntmsg->msg_routing;
1465 unsigned int payload_niov = lntmsg->msg_niov;
1466 struct iovec *payload_iov = lntmsg->msg_iov;
1467 lnet_kiov_t *payload_kiov = lntmsg->msg_kiov;
1468 unsigned int payload_offset = lntmsg->msg_offset;
1469 unsigned int payload_nob = lntmsg->msg_len;
1475 /* NB 'private' is different depending on what we're sending.... */
1477 CDEBUG(D_NET, "sending %d bytes in %d frags to %s\n",
1478 payload_nob, payload_niov, libcfs_id2str(target));
1480 LASSERT (payload_nob == 0 || payload_niov > 0);
1481 LASSERT (payload_niov <= LNET_MAX_IOV);
1483 /* Thread context */
1484 LASSERT (!cfs_in_interrupt());
1485 /* payload is either all vaddrs or all pages */
1486 LASSERT (!(payload_kiov != NULL && payload_iov != NULL));
1494 LASSERT (payload_nob == 0);
1498 if (routing || target_is_router)
1499 break; /* send IMMEDIATE */
1501 /* is the REPLY message too small for RDMA? */
1502 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[lntmsg->msg_md->md_length]);
1503 if (nob <= IBLND_MSG_SIZE)
1504 break; /* send IMMEDIATE */
1506 tx = kiblnd_get_idle_tx(ni, target.nid);
1508 CERROR("Can't allocate txd for GET to %s\n",
1509 libcfs_nid2str(target.nid));
1515 if ((lntmsg->msg_md->md_options & LNET_MD_KIOV) == 0)
1516 rc = kiblnd_setup_rd_iov(ni, tx,
1517 &ibmsg->ibm_u.get.ibgm_rd,
1518 lntmsg->msg_md->md_niov,
1519 lntmsg->msg_md->md_iov.iov,
1520 0, lntmsg->msg_md->md_length);
1522 rc = kiblnd_setup_rd_kiov(ni, tx,
1523 &ibmsg->ibm_u.get.ibgm_rd,
1524 lntmsg->msg_md->md_niov,
1525 lntmsg->msg_md->md_iov.kiov,
1526 0, lntmsg->msg_md->md_length);
1528 CERROR("Can't setup GET sink for %s: %d\n",
1529 libcfs_nid2str(target.nid), rc);
1530 kiblnd_tx_done(ni, tx);
1534 nob = offsetof(kib_get_msg_t, ibgm_rd.rd_frags[tx->tx_nfrags]);
1535 ibmsg->ibm_u.get.ibgm_cookie = tx->tx_cookie;
1536 ibmsg->ibm_u.get.ibgm_hdr = *hdr;
1538 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_GET_REQ, nob);
1540 tx->tx_lntmsg[1] = lnet_create_reply_msg(ni, lntmsg);
1541 if (tx->tx_lntmsg[1] == NULL) {
1542 CERROR("Can't create reply for GET -> %s\n",
1543 libcfs_nid2str(target.nid));
1544 kiblnd_tx_done(ni, tx);
1548 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg[0,1] on completion */
1549 tx->tx_waiting = 1; /* waiting for GET_DONE */
1550 kiblnd_launch_tx(ni, tx, target.nid);
1553 case LNET_MSG_REPLY:
1555 /* Is the payload small enough not to need RDMA? */
1556 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob]);
1557 if (nob <= IBLND_MSG_SIZE)
1558 break; /* send IMMEDIATE */
1560 tx = kiblnd_get_idle_tx(ni, target.nid);
1562 CERROR("Can't allocate %s txd for %s\n",
1563 type == LNET_MSG_PUT ? "PUT" : "REPLY",
1564 libcfs_nid2str(target.nid));
1568 if (payload_kiov == NULL)
1569 rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1570 payload_niov, payload_iov,
1571 payload_offset, payload_nob);
1573 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1574 payload_niov, payload_kiov,
1575 payload_offset, payload_nob);
1577 CERROR("Can't setup PUT src for %s: %d\n",
1578 libcfs_nid2str(target.nid), rc);
1579 kiblnd_tx_done(ni, tx);
1584 ibmsg->ibm_u.putreq.ibprm_hdr = *hdr;
1585 ibmsg->ibm_u.putreq.ibprm_cookie = tx->tx_cookie;
1586 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_REQ, sizeof(kib_putreq_msg_t));
1588 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1589 tx->tx_waiting = 1; /* waiting for PUT_{ACK,NAK} */
1590 kiblnd_launch_tx(ni, tx, target.nid);
1594 /* send IMMEDIATE */
1596 LASSERT (offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob])
1599 tx = kiblnd_get_idle_tx(ni, target.nid);
1601 CERROR ("Can't send %d to %s: tx descs exhausted\n",
1602 type, libcfs_nid2str(target.nid));
1607 ibmsg->ibm_u.immediate.ibim_hdr = *hdr;
1609 if (payload_kiov != NULL)
1610 lnet_copy_kiov2flat(IBLND_MSG_SIZE, ibmsg,
1611 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1612 payload_niov, payload_kiov,
1613 payload_offset, payload_nob);
1615 lnet_copy_iov2flat(IBLND_MSG_SIZE, ibmsg,
1616 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1617 payload_niov, payload_iov,
1618 payload_offset, payload_nob);
1620 nob = offsetof(kib_immediate_msg_t, ibim_payload[payload_nob]);
1621 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_IMMEDIATE, nob);
1623 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1624 kiblnd_launch_tx(ni, tx, target.nid);
1629 kiblnd_reply (lnet_ni_t *ni, kib_rx_t *rx, lnet_msg_t *lntmsg)
1631 lnet_process_id_t target = lntmsg->msg_target;
1632 unsigned int niov = lntmsg->msg_niov;
1633 struct iovec *iov = lntmsg->msg_iov;
1634 lnet_kiov_t *kiov = lntmsg->msg_kiov;
1635 unsigned int offset = lntmsg->msg_offset;
1636 unsigned int nob = lntmsg->msg_len;
1640 tx = kiblnd_get_idle_tx(ni, rx->rx_conn->ibc_peer->ibp_nid);
1642 CERROR("Can't get tx for REPLY to %s\n",
1643 libcfs_nid2str(target.nid));
1649 else if (kiov == NULL)
1650 rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1651 niov, iov, offset, nob);
1653 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1654 niov, kiov, offset, nob);
1657 CERROR("Can't setup GET src for %s: %d\n",
1658 libcfs_nid2str(target.nid), rc);
1662 rc = kiblnd_init_rdma(rx->rx_conn, tx,
1663 IBLND_MSG_GET_DONE, nob,
1664 &rx->rx_msg->ibm_u.get.ibgm_rd,
1665 rx->rx_msg->ibm_u.get.ibgm_cookie);
1667 CERROR("Can't setup rdma for GET from %s: %d\n",
1668 libcfs_nid2str(target.nid), rc);
1673 /* No RDMA: local completion may happen now! */
1674 lnet_finalize(ni, lntmsg, 0);
1676 /* RDMA: lnet_finalize(lntmsg) when it
1678 tx->tx_lntmsg[0] = lntmsg;
1681 kiblnd_queue_tx(tx, rx->rx_conn);
1685 kiblnd_tx_done(ni, tx);
1687 lnet_finalize(ni, lntmsg, -EIO);
1691 kiblnd_recv (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed,
1692 unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov,
1693 unsigned int offset, unsigned int mlen, unsigned int rlen)
1695 kib_rx_t *rx = private;
1696 kib_msg_t *rxmsg = rx->rx_msg;
1697 kib_conn_t *conn = rx->rx_conn;
1701 int post_credit = IBLND_POSTRX_PEER_CREDIT;
1704 LASSERT (mlen <= rlen);
1705 LASSERT (!cfs_in_interrupt());
1706 /* Either all pages or all vaddrs */
1707 LASSERT (!(kiov != NULL && iov != NULL));
1709 switch (rxmsg->ibm_type) {
1713 case IBLND_MSG_IMMEDIATE:
1714 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[rlen]);
1715 if (nob > rx->rx_nob) {
1716 CERROR ("Immediate message from %s too big: %d(%d)\n",
1717 libcfs_nid2str(rxmsg->ibm_u.immediate.ibim_hdr.src_nid),
1724 lnet_copy_flat2kiov(niov, kiov, offset,
1725 IBLND_MSG_SIZE, rxmsg,
1726 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1729 lnet_copy_flat2iov(niov, iov, offset,
1730 IBLND_MSG_SIZE, rxmsg,
1731 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1733 lnet_finalize (ni, lntmsg, 0);
1736 case IBLND_MSG_PUT_REQ:
1738 lnet_finalize(ni, lntmsg, 0);
1739 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK, 0,
1740 rxmsg->ibm_u.putreq.ibprm_cookie);
1744 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
1746 CERROR("Can't allocate tx for %s\n",
1747 libcfs_nid2str(conn->ibc_peer->ibp_nid));
1748 /* Not replying will break the connection */
1755 rc = kiblnd_setup_rd_iov(ni, tx,
1756 &txmsg->ibm_u.putack.ibpam_rd,
1757 niov, iov, offset, mlen);
1759 rc = kiblnd_setup_rd_kiov(ni, tx,
1760 &txmsg->ibm_u.putack.ibpam_rd,
1761 niov, kiov, offset, mlen);
1763 CERROR("Can't setup PUT sink for %s: %d\n",
1764 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
1765 kiblnd_tx_done(ni, tx);
1766 /* tell peer it's over */
1767 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK, rc,
1768 rxmsg->ibm_u.putreq.ibprm_cookie);
1772 nob = offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[tx->tx_nfrags]);
1773 txmsg->ibm_u.putack.ibpam_src_cookie = rxmsg->ibm_u.putreq.ibprm_cookie;
1774 txmsg->ibm_u.putack.ibpam_dst_cookie = tx->tx_cookie;
1776 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_ACK, nob);
1778 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1779 tx->tx_waiting = 1; /* waiting for PUT_DONE */
1780 kiblnd_queue_tx(tx, conn);
1782 /* reposted buffer reserved for PUT_DONE */
1783 post_credit = IBLND_POSTRX_NO_CREDIT;
1786 case IBLND_MSG_GET_REQ:
1787 if (lntmsg != NULL) {
1788 /* Optimized GET; RDMA lntmsg's payload */
1789 kiblnd_reply(ni, rx, lntmsg);
1791 /* GET didn't match anything */
1792 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_GET_DONE,
1794 rxmsg->ibm_u.get.ibgm_cookie);
1799 kiblnd_post_rx(rx, post_credit);
1804 kiblnd_thread_start (int (*fn)(void *arg), void *arg)
1806 long pid = cfs_create_thread (fn, arg, 0);
1811 cfs_atomic_inc (&kiblnd_data.kib_nthreads);
1816 kiblnd_thread_fini (void)
1818 cfs_atomic_dec (&kiblnd_data.kib_nthreads);
1822 kiblnd_peer_alive (kib_peer_t *peer)
1824 /* This is racy, but everyone's only writing cfs_time_current() */
1825 peer->ibp_last_alive = cfs_time_current();
1830 kiblnd_peer_notify (kib_peer_t *peer)
1833 cfs_time_t last_alive = 0;
1834 unsigned long flags;
1836 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1838 if (cfs_list_empty(&peer->ibp_conns) &&
1839 peer->ibp_accepting == 0 &&
1840 peer->ibp_connecting == 0 &&
1841 peer->ibp_error != 0) {
1842 error = peer->ibp_error;
1843 peer->ibp_error = 0;
1845 last_alive = peer->ibp_last_alive;
1848 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1851 lnet_notify(peer->ibp_ni,
1852 peer->ibp_nid, 0, last_alive);
1856 kiblnd_close_conn_locked (kib_conn_t *conn, int error)
1858 /* This just does the immediate housekeeping. 'error' is zero for a
1859 * normal shutdown which can happen only after the connection has been
1860 * established. If the connection is established, schedule the
1861 * connection to be finished off by the connd. Otherwise the connd is
1862 * already dealing with it (either to set it up or tear it down).
1863 * Caller holds kib_global_lock exclusively in irq context */
1864 kib_peer_t *peer = conn->ibc_peer;
1866 unsigned long flags;
1868 LASSERT (error != 0 || conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1870 if (error != 0 && conn->ibc_comms_error == 0)
1871 conn->ibc_comms_error = error;
1873 if (conn->ibc_state != IBLND_CONN_ESTABLISHED)
1874 return; /* already being handled */
1877 cfs_list_empty(&conn->ibc_tx_noops) &&
1878 cfs_list_empty(&conn->ibc_tx_queue) &&
1879 cfs_list_empty(&conn->ibc_tx_queue_rsrvd) &&
1880 cfs_list_empty(&conn->ibc_tx_queue_nocred) &&
1881 cfs_list_empty(&conn->ibc_active_txs)) {
1882 CDEBUG(D_NET, "closing conn to %s\n",
1883 libcfs_nid2str(peer->ibp_nid));
1885 CNETERR("Closing conn to %s: error %d%s%s%s%s%s\n",
1886 libcfs_nid2str(peer->ibp_nid), error,
1887 cfs_list_empty(&conn->ibc_tx_queue) ? "" : "(sending)",
1888 cfs_list_empty(&conn->ibc_tx_noops) ? "" : "(sending_noops)",
1889 cfs_list_empty(&conn->ibc_tx_queue_rsrvd) ? "" : "(sending_rsrvd)",
1890 cfs_list_empty(&conn->ibc_tx_queue_nocred) ? "" : "(sending_nocred)",
1891 cfs_list_empty(&conn->ibc_active_txs) ? "" : "(waiting)");
1894 dev = ((kib_net_t *)peer->ibp_ni->ni_data)->ibn_dev;
1895 cfs_list_del(&conn->ibc_list);
1896 /* connd (see below) takes over ibc_list's ref */
1898 if (cfs_list_empty (&peer->ibp_conns) && /* no more conns */
1899 kiblnd_peer_active(peer)) { /* still in peer table */
1900 kiblnd_unlink_peer_locked(peer);
1902 /* set/clear error on last conn */
1903 peer->ibp_error = conn->ibc_comms_error;
1906 kiblnd_set_conn_state(conn, IBLND_CONN_CLOSING);
1909 kiblnd_dev_can_failover(dev)) {
1910 cfs_list_add_tail(&dev->ibd_fail_list,
1911 &kiblnd_data.kib_failed_devs);
1912 cfs_waitq_signal(&kiblnd_data.kib_failover_waitq);
1915 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
1917 cfs_list_add_tail(&conn->ibc_list, &kiblnd_data.kib_connd_conns);
1918 cfs_waitq_signal(&kiblnd_data.kib_connd_waitq);
1920 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
1924 kiblnd_close_conn(kib_conn_t *conn, int error)
1926 unsigned long flags;
1928 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1930 kiblnd_close_conn_locked(conn, error);
1932 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1936 kiblnd_handle_early_rxs(kib_conn_t *conn)
1938 unsigned long flags;
1941 LASSERT(!cfs_in_interrupt());
1942 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1944 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1945 while (!cfs_list_empty(&conn->ibc_early_rxs)) {
1946 rx = cfs_list_entry(conn->ibc_early_rxs.next,
1948 cfs_list_del(&rx->rx_list);
1949 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1951 kiblnd_handle_rx(rx);
1953 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1955 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1959 kiblnd_abort_txs(kib_conn_t *conn, cfs_list_t *txs)
1961 CFS_LIST_HEAD (zombies);
1966 spin_lock(&conn->ibc_lock);
1968 cfs_list_for_each_safe (tmp, nxt, txs) {
1969 tx = cfs_list_entry (tmp, kib_tx_t, tx_list);
1971 if (txs == &conn->ibc_active_txs) {
1972 LASSERT (!tx->tx_queued);
1973 LASSERT (tx->tx_waiting ||
1974 tx->tx_sending != 0);
1976 LASSERT (tx->tx_queued);
1979 tx->tx_status = -ECONNABORTED;
1982 if (tx->tx_sending == 0) {
1984 cfs_list_del (&tx->tx_list);
1985 cfs_list_add (&tx->tx_list, &zombies);
1989 spin_unlock(&conn->ibc_lock);
1991 kiblnd_txlist_done(conn->ibc_peer->ibp_ni, &zombies, -ECONNABORTED);
1995 kiblnd_finalise_conn (kib_conn_t *conn)
1997 LASSERT (!cfs_in_interrupt());
1998 LASSERT (conn->ibc_state > IBLND_CONN_INIT);
2000 kiblnd_set_conn_state(conn, IBLND_CONN_DISCONNECTED);
2002 /* abort_receives moves QP state to IB_QPS_ERR. This is only required
2003 * for connections that didn't get as far as being connected, because
2004 * rdma_disconnect() does this for free. */
2005 kiblnd_abort_receives(conn);
2007 /* Complete all tx descs not waiting for sends to complete.
2008 * NB we should be safe from RDMA now that the QP has changed state */
2010 kiblnd_abort_txs(conn, &conn->ibc_tx_noops);
2011 kiblnd_abort_txs(conn, &conn->ibc_tx_queue);
2012 kiblnd_abort_txs(conn, &conn->ibc_tx_queue_rsrvd);
2013 kiblnd_abort_txs(conn, &conn->ibc_tx_queue_nocred);
2014 kiblnd_abort_txs(conn, &conn->ibc_active_txs);
2016 kiblnd_handle_early_rxs(conn);
2020 kiblnd_peer_connect_failed (kib_peer_t *peer, int active, int error)
2022 CFS_LIST_HEAD (zombies);
2023 unsigned long flags;
2025 LASSERT (error != 0);
2026 LASSERT (!cfs_in_interrupt());
2028 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2031 LASSERT (peer->ibp_connecting > 0);
2032 peer->ibp_connecting--;
2034 LASSERT (peer->ibp_accepting > 0);
2035 peer->ibp_accepting--;
2038 if (peer->ibp_connecting != 0 ||
2039 peer->ibp_accepting != 0) {
2040 /* another connection attempt under way... */
2041 write_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2046 if (cfs_list_empty(&peer->ibp_conns)) {
2047 /* Take peer's blocked transmits to complete with error */
2048 cfs_list_add(&zombies, &peer->ibp_tx_queue);
2049 cfs_list_del_init(&peer->ibp_tx_queue);
2051 if (kiblnd_peer_active(peer))
2052 kiblnd_unlink_peer_locked(peer);
2054 peer->ibp_error = error;
2056 /* Can't have blocked transmits if there are connections */
2057 LASSERT (cfs_list_empty(&peer->ibp_tx_queue));
2060 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2062 kiblnd_peer_notify(peer);
2064 if (cfs_list_empty (&zombies))
2067 CNETERR("Deleting messages for %s: connection failed\n",
2068 libcfs_nid2str(peer->ibp_nid));
2070 kiblnd_txlist_done(peer->ibp_ni, &zombies, -EHOSTUNREACH);
2074 kiblnd_connreq_done(kib_conn_t *conn, int status)
2076 kib_peer_t *peer = conn->ibc_peer;
2079 unsigned long flags;
2082 active = (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2084 CDEBUG(D_NET,"%s: active(%d), version(%x), status(%d)\n",
2085 libcfs_nid2str(peer->ibp_nid), active,
2086 conn->ibc_version, status);
2088 LASSERT (!cfs_in_interrupt());
2089 LASSERT ((conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT &&
2090 peer->ibp_connecting > 0) ||
2091 (conn->ibc_state == IBLND_CONN_PASSIVE_WAIT &&
2092 peer->ibp_accepting > 0));
2094 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
2095 conn->ibc_connvars = NULL;
2098 /* failed to establish connection */
2099 kiblnd_peer_connect_failed(peer, active, status);
2100 kiblnd_finalise_conn(conn);
2104 /* connection established */
2105 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2107 conn->ibc_last_send = jiffies;
2108 kiblnd_set_conn_state(conn, IBLND_CONN_ESTABLISHED);
2109 kiblnd_peer_alive(peer);
2111 /* Add conn to peer's list and nuke any dangling conns from a different
2112 * peer instance... */
2113 kiblnd_conn_addref(conn); /* +1 ref for ibc_list */
2114 cfs_list_add(&conn->ibc_list, &peer->ibp_conns);
2116 peer->ibp_connecting--;
2118 peer->ibp_accepting--;
2120 if (peer->ibp_version == 0) {
2121 peer->ibp_version = conn->ibc_version;
2122 peer->ibp_incarnation = conn->ibc_incarnation;
2125 if (peer->ibp_version != conn->ibc_version ||
2126 peer->ibp_incarnation != conn->ibc_incarnation) {
2127 kiblnd_close_stale_conns_locked(peer, conn->ibc_version,
2128 conn->ibc_incarnation);
2129 peer->ibp_version = conn->ibc_version;
2130 peer->ibp_incarnation = conn->ibc_incarnation;
2133 /* grab pending txs while I have the lock */
2134 cfs_list_add(&txs, &peer->ibp_tx_queue);
2135 cfs_list_del_init(&peer->ibp_tx_queue);
2137 if (!kiblnd_peer_active(peer) || /* peer has been deleted */
2138 conn->ibc_comms_error != 0) { /* error has happened already */
2139 lnet_ni_t *ni = peer->ibp_ni;
2141 /* start to shut down connection */
2142 kiblnd_close_conn_locked(conn, -ECONNABORTED);
2143 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2145 kiblnd_txlist_done(ni, &txs, -ECONNABORTED);
2150 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2152 /* Schedule blocked txs */
2153 spin_lock(&conn->ibc_lock);
2154 while (!cfs_list_empty(&txs)) {
2155 tx = cfs_list_entry(txs.next, kib_tx_t, tx_list);
2156 cfs_list_del(&tx->tx_list);
2158 kiblnd_queue_tx_locked(tx, conn);
2160 spin_unlock(&conn->ibc_lock);
2162 kiblnd_check_sends(conn);
2164 /* schedule blocked rxs */
2165 kiblnd_handle_early_rxs(conn);
2169 kiblnd_reject(struct rdma_cm_id *cmid, kib_rej_t *rej)
2173 rc = rdma_reject(cmid, rej, sizeof(*rej));
2176 CWARN("Error %d sending reject\n", rc);
2180 kiblnd_passive_connect (struct rdma_cm_id *cmid, void *priv, int priv_nob)
2182 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2183 kib_msg_t *reqmsg = priv;
2189 lnet_ni_t *ni = NULL;
2190 kib_net_t *net = NULL;
2192 struct rdma_conn_param cp;
2194 int version = IBLND_MSG_VERSION;
2195 unsigned long flags;
2197 struct sockaddr_in *peer_addr;
2198 LASSERT (!cfs_in_interrupt());
2200 /* cmid inherits 'context' from the corresponding listener id */
2201 ibdev = (kib_dev_t *)cmid->context;
2202 LASSERT (ibdev != NULL);
2204 memset(&rej, 0, sizeof(rej));
2205 rej.ibr_magic = IBLND_MSG_MAGIC;
2206 rej.ibr_why = IBLND_REJECT_FATAL;
2207 rej.ibr_cp.ibcp_max_msg_size = IBLND_MSG_SIZE;
2209 peer_addr = (struct sockaddr_in *)&(cmid->route.addr.dst_addr);
2210 if (*kiblnd_tunables.kib_require_priv_port &&
2211 ntohs(peer_addr->sin_port) >= PROT_SOCK) {
2212 __u32 ip = ntohl(peer_addr->sin_addr.s_addr);
2213 CERROR("Peer's port (%u.%u.%u.%u:%hu) is not privileged\n",
2214 HIPQUAD(ip), ntohs(peer_addr->sin_port));
2218 if (priv_nob < offsetof(kib_msg_t, ibm_type)) {
2219 CERROR("Short connection request\n");
2223 /* Future protocol version compatibility support! If the
2224 * o2iblnd-specific protocol changes, or when LNET unifies
2225 * protocols over all LNDs, the initial connection will
2226 * negotiate a protocol version. I trap this here to avoid
2227 * console errors; the reject tells the peer which protocol I
2229 if (reqmsg->ibm_magic == LNET_PROTO_MAGIC ||
2230 reqmsg->ibm_magic == __swab32(LNET_PROTO_MAGIC))
2232 if (reqmsg->ibm_magic == IBLND_MSG_MAGIC &&
2233 reqmsg->ibm_version != IBLND_MSG_VERSION &&
2234 reqmsg->ibm_version != IBLND_MSG_VERSION_1)
2236 if (reqmsg->ibm_magic == __swab32(IBLND_MSG_MAGIC) &&
2237 reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION) &&
2238 reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION_1))
2241 rc = kiblnd_unpack_msg(reqmsg, priv_nob);
2243 CERROR("Can't parse connection request: %d\n", rc);
2247 nid = reqmsg->ibm_srcnid;
2248 ni = lnet_net2ni(LNET_NIDNET(reqmsg->ibm_dstnid));
2251 net = (kib_net_t *)ni->ni_data;
2252 rej.ibr_incarnation = net->ibn_incarnation;
2255 if (ni == NULL || /* no matching net */
2256 ni->ni_nid != reqmsg->ibm_dstnid || /* right NET, wrong NID! */
2257 net->ibn_dev != ibdev) { /* wrong device */
2258 CERROR("Can't accept %s on %s (%s:%d:%u.%u.%u.%u): "
2259 "bad dst nid %s\n", libcfs_nid2str(nid),
2260 ni == NULL ? "NA" : libcfs_nid2str(ni->ni_nid),
2261 ibdev->ibd_ifname, ibdev->ibd_nnets,
2262 HIPQUAD(ibdev->ibd_ifip),
2263 libcfs_nid2str(reqmsg->ibm_dstnid));
2268 /* check time stamp as soon as possible */
2269 if (reqmsg->ibm_dststamp != 0 &&
2270 reqmsg->ibm_dststamp != net->ibn_incarnation) {
2271 CWARN("Stale connection request\n");
2272 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2276 /* I can accept peer's version */
2277 version = reqmsg->ibm_version;
2279 if (reqmsg->ibm_type != IBLND_MSG_CONNREQ) {
2280 CERROR("Unexpected connreq msg type: %x from %s\n",
2281 reqmsg->ibm_type, libcfs_nid2str(nid));
2285 if (reqmsg->ibm_u.connparams.ibcp_queue_depth !=
2286 IBLND_MSG_QUEUE_SIZE(version)) {
2287 CERROR("Can't accept %s: incompatible queue depth %d (%d wanted)\n",
2288 libcfs_nid2str(nid), reqmsg->ibm_u.connparams.ibcp_queue_depth,
2289 IBLND_MSG_QUEUE_SIZE(version));
2291 if (version == IBLND_MSG_VERSION)
2292 rej.ibr_why = IBLND_REJECT_MSG_QUEUE_SIZE;
2297 if (reqmsg->ibm_u.connparams.ibcp_max_frags !=
2298 IBLND_RDMA_FRAGS(version)) {
2299 CERROR("Can't accept %s(version %x): "
2300 "incompatible max_frags %d (%d wanted)\n",
2301 libcfs_nid2str(nid), version,
2302 reqmsg->ibm_u.connparams.ibcp_max_frags,
2303 IBLND_RDMA_FRAGS(version));
2305 if (version == IBLND_MSG_VERSION)
2306 rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2312 if (reqmsg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2313 CERROR("Can't accept %s: message size %d too big (%d max)\n",
2314 libcfs_nid2str(nid),
2315 reqmsg->ibm_u.connparams.ibcp_max_msg_size,
2320 /* assume 'nid' is a new peer; create */
2321 rc = kiblnd_create_peer(ni, &peer, nid);
2323 CERROR("Can't create peer for %s\n", libcfs_nid2str(nid));
2324 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2328 write_lock_irqsave(g_lock, flags);
2330 peer2 = kiblnd_find_peer_locked(nid);
2331 if (peer2 != NULL) {
2332 if (peer2->ibp_version == 0) {
2333 peer2->ibp_version = version;
2334 peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2337 /* not the guy I've talked with */
2338 if (peer2->ibp_incarnation != reqmsg->ibm_srcstamp ||
2339 peer2->ibp_version != version) {
2340 kiblnd_close_peer_conns_locked(peer2, -ESTALE);
2341 write_unlock_irqrestore(g_lock, flags);
2343 CWARN("Conn stale %s [old ver: %x, new ver: %x]\n",
2344 libcfs_nid2str(nid), peer2->ibp_version, version);
2346 kiblnd_peer_decref(peer);
2347 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2351 /* tie-break connection race in favour of the higher NID */
2352 if (peer2->ibp_connecting != 0 &&
2354 write_unlock_irqrestore(g_lock, flags);
2356 CWARN("Conn race %s\n", libcfs_nid2str(peer2->ibp_nid));
2358 kiblnd_peer_decref(peer);
2359 rej.ibr_why = IBLND_REJECT_CONN_RACE;
2363 peer2->ibp_accepting++;
2364 kiblnd_peer_addref(peer2);
2366 write_unlock_irqrestore(g_lock, flags);
2367 kiblnd_peer_decref(peer);
2370 /* Brand new peer */
2371 LASSERT (peer->ibp_accepting == 0);
2372 LASSERT (peer->ibp_version == 0 &&
2373 peer->ibp_incarnation == 0);
2375 peer->ibp_accepting = 1;
2376 peer->ibp_version = version;
2377 peer->ibp_incarnation = reqmsg->ibm_srcstamp;
2379 /* I have a ref on ni that prevents it being shutdown */
2380 LASSERT (net->ibn_shutdown == 0);
2382 kiblnd_peer_addref(peer);
2383 cfs_list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid));
2385 write_unlock_irqrestore(g_lock, flags);
2388 conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_PASSIVE_WAIT, version);
2390 kiblnd_peer_connect_failed(peer, 0, -ENOMEM);
2391 kiblnd_peer_decref(peer);
2392 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2396 /* conn now "owns" cmid, so I return success from here on to ensure the
2397 * CM callback doesn't destroy cmid. */
2399 conn->ibc_incarnation = reqmsg->ibm_srcstamp;
2400 conn->ibc_credits = IBLND_MSG_QUEUE_SIZE(version);
2401 conn->ibc_reserved_credits = IBLND_MSG_QUEUE_SIZE(version);
2402 LASSERT (conn->ibc_credits + conn->ibc_reserved_credits + IBLND_OOB_MSGS(version)
2403 <= IBLND_RX_MSGS(version));
2405 ackmsg = &conn->ibc_connvars->cv_msg;
2406 memset(ackmsg, 0, sizeof(*ackmsg));
2408 kiblnd_init_msg(ackmsg, IBLND_MSG_CONNACK,
2409 sizeof(ackmsg->ibm_u.connparams));
2410 ackmsg->ibm_u.connparams.ibcp_queue_depth = IBLND_MSG_QUEUE_SIZE(version);
2411 ackmsg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2412 ackmsg->ibm_u.connparams.ibcp_max_frags = IBLND_RDMA_FRAGS(version);
2414 kiblnd_pack_msg(ni, ackmsg, version, 0, nid, reqmsg->ibm_srcstamp);
2416 memset(&cp, 0, sizeof(cp));
2417 cp.private_data = ackmsg;
2418 cp.private_data_len = ackmsg->ibm_nob;
2419 cp.responder_resources = 0; /* No atomic ops or RDMA reads */
2420 cp.initiator_depth = 0;
2421 cp.flow_control = 1;
2422 cp.retry_count = *kiblnd_tunables.kib_retry_count;
2423 cp.rnr_retry_count = *kiblnd_tunables.kib_rnr_retry_count;
2425 CDEBUG(D_NET, "Accept %s\n", libcfs_nid2str(nid));
2427 rc = rdma_accept(cmid, &cp);
2429 CERROR("Can't accept %s: %d\n", libcfs_nid2str(nid), rc);
2430 rej.ibr_version = version;
2431 rej.ibr_why = IBLND_REJECT_FATAL;
2433 kiblnd_reject(cmid, &rej);
2434 kiblnd_connreq_done(conn, rc);
2435 kiblnd_conn_decref(conn);
2445 rej.ibr_version = version;
2446 rej.ibr_cp.ibcp_queue_depth = IBLND_MSG_QUEUE_SIZE(version);
2447 rej.ibr_cp.ibcp_max_frags = IBLND_RDMA_FRAGS(version);
2448 kiblnd_reject(cmid, &rej);
2450 return -ECONNREFUSED;
2454 kiblnd_reconnect (kib_conn_t *conn, int version,
2455 __u64 incarnation, int why, kib_connparams_t *cp)
2457 kib_peer_t *peer = conn->ibc_peer;
2460 unsigned long flags;
2462 LASSERT (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2463 LASSERT (peer->ibp_connecting > 0); /* 'conn' at least */
2465 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2467 /* retry connection if it's still needed and no other connection
2468 * attempts (active or passive) are in progress
2469 * NB: reconnect is still needed even when ibp_tx_queue is
2470 * empty if ibp_version != version because reconnect may be
2471 * initiated by kiblnd_query() */
2472 if ((!cfs_list_empty(&peer->ibp_tx_queue) ||
2473 peer->ibp_version != version) &&
2474 peer->ibp_connecting == 1 &&
2475 peer->ibp_accepting == 0) {
2477 peer->ibp_connecting++;
2479 peer->ibp_version = version;
2480 peer->ibp_incarnation = incarnation;
2483 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2493 case IBLND_REJECT_CONN_STALE:
2497 case IBLND_REJECT_CONN_RACE:
2498 reason = "conn race";
2501 case IBLND_REJECT_CONN_UNCOMPAT:
2502 reason = "version negotiation";
2506 CNETERR("%s: retrying (%s), %x, %x, "
2507 "queue_dep: %d, max_frag: %d, msg_size: %d\n",
2508 libcfs_nid2str(peer->ibp_nid),
2509 reason, IBLND_MSG_VERSION, version,
2510 cp != NULL? cp->ibcp_queue_depth :IBLND_MSG_QUEUE_SIZE(version),
2511 cp != NULL? cp->ibcp_max_frags : IBLND_RDMA_FRAGS(version),
2512 cp != NULL? cp->ibcp_max_msg_size: IBLND_MSG_SIZE);
2514 kiblnd_connect_peer(peer);
2518 kiblnd_rejected (kib_conn_t *conn, int reason, void *priv, int priv_nob)
2520 kib_peer_t *peer = conn->ibc_peer;
2522 LASSERT (!cfs_in_interrupt());
2523 LASSERT (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2526 case IB_CM_REJ_STALE_CONN:
2527 kiblnd_reconnect(conn, IBLND_MSG_VERSION, 0,
2528 IBLND_REJECT_CONN_STALE, NULL);
2531 case IB_CM_REJ_INVALID_SERVICE_ID:
2532 CNETERR("%s rejected: no listener at %d\n",
2533 libcfs_nid2str(peer->ibp_nid),
2534 *kiblnd_tunables.kib_service);
2537 case IB_CM_REJ_CONSUMER_DEFINED:
2538 if (priv_nob >= offsetof(kib_rej_t, ibr_padding)) {
2539 kib_rej_t *rej = priv;
2540 kib_connparams_t *cp = NULL;
2542 __u64 incarnation = -1;
2544 /* NB. default incarnation is -1 because:
2545 * a) V1 will ignore dst incarnation in connreq.
2546 * b) V2 will provide incarnation while rejecting me,
2547 * -1 will be overwrote.
2549 * if I try to connect to a V1 peer with V2 protocol,
2550 * it rejected me then upgrade to V2, I have no idea
2551 * about the upgrading and try to reconnect with V1,
2552 * in this case upgraded V2 can find out I'm trying to
2553 * talk to the old guy and reject me(incarnation is -1).
2556 if (rej->ibr_magic == __swab32(IBLND_MSG_MAGIC) ||
2557 rej->ibr_magic == __swab32(LNET_PROTO_MAGIC)) {
2558 __swab32s(&rej->ibr_magic);
2559 __swab16s(&rej->ibr_version);
2563 if (priv_nob >= sizeof(kib_rej_t) &&
2564 rej->ibr_version > IBLND_MSG_VERSION_1) {
2565 /* priv_nob is always 148 in current version
2566 * of OFED, so we still need to check version.
2567 * (define of IB_CM_REJ_PRIVATE_DATA_SIZE) */
2571 __swab64s(&rej->ibr_incarnation);
2572 __swab16s(&cp->ibcp_queue_depth);
2573 __swab16s(&cp->ibcp_max_frags);
2574 __swab32s(&cp->ibcp_max_msg_size);
2577 incarnation = rej->ibr_incarnation;
2580 if (rej->ibr_magic != IBLND_MSG_MAGIC &&
2581 rej->ibr_magic != LNET_PROTO_MAGIC) {
2582 CERROR("%s rejected: consumer defined fatal error\n",
2583 libcfs_nid2str(peer->ibp_nid));
2587 if (rej->ibr_version != IBLND_MSG_VERSION &&
2588 rej->ibr_version != IBLND_MSG_VERSION_1) {
2589 CERROR("%s rejected: o2iblnd version %x error\n",
2590 libcfs_nid2str(peer->ibp_nid),
2595 if (rej->ibr_why == IBLND_REJECT_FATAL &&
2596 rej->ibr_version == IBLND_MSG_VERSION_1) {
2597 CDEBUG(D_NET, "rejected by old version peer %s: %x\n",
2598 libcfs_nid2str(peer->ibp_nid), rej->ibr_version);
2600 if (conn->ibc_version != IBLND_MSG_VERSION_1)
2601 rej->ibr_why = IBLND_REJECT_CONN_UNCOMPAT;
2604 switch (rej->ibr_why) {
2605 case IBLND_REJECT_CONN_RACE:
2606 case IBLND_REJECT_CONN_STALE:
2607 case IBLND_REJECT_CONN_UNCOMPAT:
2608 kiblnd_reconnect(conn, rej->ibr_version,
2609 incarnation, rej->ibr_why, cp);
2612 case IBLND_REJECT_MSG_QUEUE_SIZE:
2613 CERROR("%s rejected: incompatible message queue depth %d, %d\n",
2614 libcfs_nid2str(peer->ibp_nid), cp->ibcp_queue_depth,
2615 IBLND_MSG_QUEUE_SIZE(conn->ibc_version));
2618 case IBLND_REJECT_RDMA_FRAGS:
2619 CERROR("%s rejected: incompatible # of RDMA fragments %d, %d\n",
2620 libcfs_nid2str(peer->ibp_nid), cp->ibcp_max_frags,
2621 IBLND_RDMA_FRAGS(conn->ibc_version));
2624 case IBLND_REJECT_NO_RESOURCES:
2625 CERROR("%s rejected: o2iblnd no resources\n",
2626 libcfs_nid2str(peer->ibp_nid));
2629 case IBLND_REJECT_FATAL:
2630 CERROR("%s rejected: o2iblnd fatal error\n",
2631 libcfs_nid2str(peer->ibp_nid));
2635 CERROR("%s rejected: o2iblnd reason %d\n",
2636 libcfs_nid2str(peer->ibp_nid),
2644 CNETERR("%s rejected: reason %d, size %d\n",
2645 libcfs_nid2str(peer->ibp_nid), reason, priv_nob);
2649 kiblnd_connreq_done(conn, -ECONNREFUSED);
2653 kiblnd_check_connreply (kib_conn_t *conn, void *priv, int priv_nob)
2655 kib_peer_t *peer = conn->ibc_peer;
2656 lnet_ni_t *ni = peer->ibp_ni;
2657 kib_net_t *net = ni->ni_data;
2658 kib_msg_t *msg = priv;
2659 int ver = conn->ibc_version;
2660 int rc = kiblnd_unpack_msg(msg, priv_nob);
2661 unsigned long flags;
2663 LASSERT (net != NULL);
2666 CERROR("Can't unpack connack from %s: %d\n",
2667 libcfs_nid2str(peer->ibp_nid), rc);
2671 if (msg->ibm_type != IBLND_MSG_CONNACK) {
2672 CERROR("Unexpected message %d from %s\n",
2673 msg->ibm_type, libcfs_nid2str(peer->ibp_nid));
2678 if (ver != msg->ibm_version) {
2679 CERROR("%s replied version %x is different with "
2680 "requested version %x\n",
2681 libcfs_nid2str(peer->ibp_nid), msg->ibm_version, ver);
2686 if (msg->ibm_u.connparams.ibcp_queue_depth !=
2687 IBLND_MSG_QUEUE_SIZE(ver)) {
2688 CERROR("%s has incompatible queue depth %d(%d wanted)\n",
2689 libcfs_nid2str(peer->ibp_nid),
2690 msg->ibm_u.connparams.ibcp_queue_depth,
2691 IBLND_MSG_QUEUE_SIZE(ver));
2696 if (msg->ibm_u.connparams.ibcp_max_frags !=
2697 IBLND_RDMA_FRAGS(ver)) {
2698 CERROR("%s has incompatible max_frags %d (%d wanted)\n",
2699 libcfs_nid2str(peer->ibp_nid),
2700 msg->ibm_u.connparams.ibcp_max_frags,
2701 IBLND_RDMA_FRAGS(ver));
2706 if (msg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2707 CERROR("%s max message size %d too big (%d max)\n",
2708 libcfs_nid2str(peer->ibp_nid),
2709 msg->ibm_u.connparams.ibcp_max_msg_size,
2715 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2716 if (msg->ibm_dstnid == ni->ni_nid &&
2717 msg->ibm_dststamp == net->ibn_incarnation)
2721 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2724 CERROR("Bad connection reply from %s, rc = %d, "
2725 "version: %x max_frags: %d\n",
2726 libcfs_nid2str(peer->ibp_nid), rc,
2727 msg->ibm_version, msg->ibm_u.connparams.ibcp_max_frags);
2731 conn->ibc_incarnation = msg->ibm_srcstamp;
2733 conn->ibc_reserved_credits = IBLND_MSG_QUEUE_SIZE(ver);
2734 LASSERT (conn->ibc_credits + conn->ibc_reserved_credits + IBLND_OOB_MSGS(ver)
2735 <= IBLND_RX_MSGS(ver));
2737 kiblnd_connreq_done(conn, 0);
2741 /* NB My QP has already established itself, so I handle anything going
2742 * wrong here by setting ibc_comms_error.
2743 * kiblnd_connreq_done(0) moves the conn state to ESTABLISHED, but then
2744 * immediately tears it down. */
2747 conn->ibc_comms_error = rc;
2748 kiblnd_connreq_done(conn, 0);
2752 kiblnd_active_connect (struct rdma_cm_id *cmid)
2754 kib_peer_t *peer = (kib_peer_t *)cmid->context;
2757 struct rdma_conn_param cp;
2760 unsigned long flags;
2763 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2765 incarnation = peer->ibp_incarnation;
2766 version = (peer->ibp_version == 0) ? IBLND_MSG_VERSION :
2769 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2771 conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_ACTIVE_CONNECT, version);
2773 kiblnd_peer_connect_failed(peer, 1, -ENOMEM);
2774 kiblnd_peer_decref(peer); /* lose cmid's ref */
2778 /* conn "owns" cmid now, so I return success from here on to ensure the
2779 * CM callback doesn't destroy cmid. conn also takes over cmid's ref
2782 msg = &conn->ibc_connvars->cv_msg;
2784 memset(msg, 0, sizeof(*msg));
2785 kiblnd_init_msg(msg, IBLND_MSG_CONNREQ, sizeof(msg->ibm_u.connparams));
2786 msg->ibm_u.connparams.ibcp_queue_depth = IBLND_MSG_QUEUE_SIZE(version);
2787 msg->ibm_u.connparams.ibcp_max_frags = IBLND_RDMA_FRAGS(version);
2788 msg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2790 kiblnd_pack_msg(peer->ibp_ni, msg, version,
2791 0, peer->ibp_nid, incarnation);
2793 memset(&cp, 0, sizeof(cp));
2794 cp.private_data = msg;
2795 cp.private_data_len = msg->ibm_nob;
2796 cp.responder_resources = 0; /* No atomic ops or RDMA reads */
2797 cp.initiator_depth = 0;
2798 cp.flow_control = 1;
2799 cp.retry_count = *kiblnd_tunables.kib_retry_count;
2800 cp.rnr_retry_count = *kiblnd_tunables.kib_rnr_retry_count;
2802 LASSERT(cmid->context == (void *)conn);
2803 LASSERT(conn->ibc_cmid == cmid);
2805 rc = rdma_connect(cmid, &cp);
2807 CERROR("Can't connect to %s: %d\n",
2808 libcfs_nid2str(peer->ibp_nid), rc);
2809 kiblnd_connreq_done(conn, rc);
2810 kiblnd_conn_decref(conn);
2817 kiblnd_cm_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2823 switch (event->event) {
2825 CERROR("Unexpected event: %d, status: %d\n",
2826 event->event, event->status);
2829 case RDMA_CM_EVENT_CONNECT_REQUEST:
2830 /* destroy cmid on failure */
2831 rc = kiblnd_passive_connect(cmid,
2832 (void *)KIBLND_CONN_PARAM(event),
2833 KIBLND_CONN_PARAM_LEN(event));
2834 CDEBUG(D_NET, "connreq: %d\n", rc);
2837 case RDMA_CM_EVENT_ADDR_ERROR:
2838 peer = (kib_peer_t *)cmid->context;
2839 CNETERR("%s: ADDR ERROR %d\n",
2840 libcfs_nid2str(peer->ibp_nid), event->status);
2841 kiblnd_peer_connect_failed(peer, 1, -EHOSTUNREACH);
2842 kiblnd_peer_decref(peer);
2843 return -EHOSTUNREACH; /* rc != 0 destroys cmid */
2845 case RDMA_CM_EVENT_ADDR_RESOLVED:
2846 peer = (kib_peer_t *)cmid->context;
2848 CDEBUG(D_NET,"%s Addr resolved: %d\n",
2849 libcfs_nid2str(peer->ibp_nid), event->status);
2851 if (event->status != 0) {
2852 CNETERR("Can't resolve address for %s: %d\n",
2853 libcfs_nid2str(peer->ibp_nid), event->status);
2856 rc = rdma_resolve_route(
2857 cmid, *kiblnd_tunables.kib_timeout * 1000);
2860 /* Can't initiate route resolution */
2861 CERROR("Can't resolve route for %s: %d\n",
2862 libcfs_nid2str(peer->ibp_nid), rc);
2864 kiblnd_peer_connect_failed(peer, 1, rc);
2865 kiblnd_peer_decref(peer);
2866 return rc; /* rc != 0 destroys cmid */
2868 case RDMA_CM_EVENT_ROUTE_ERROR:
2869 peer = (kib_peer_t *)cmid->context;
2870 CNETERR("%s: ROUTE ERROR %d\n",
2871 libcfs_nid2str(peer->ibp_nid), event->status);
2872 kiblnd_peer_connect_failed(peer, 1, -EHOSTUNREACH);
2873 kiblnd_peer_decref(peer);
2874 return -EHOSTUNREACH; /* rc != 0 destroys cmid */
2876 case RDMA_CM_EVENT_ROUTE_RESOLVED:
2877 peer = (kib_peer_t *)cmid->context;
2878 CDEBUG(D_NET,"%s Route resolved: %d\n",
2879 libcfs_nid2str(peer->ibp_nid), event->status);
2881 if (event->status == 0)
2882 return kiblnd_active_connect(cmid);
2884 CNETERR("Can't resolve route for %s: %d\n",
2885 libcfs_nid2str(peer->ibp_nid), event->status);
2886 kiblnd_peer_connect_failed(peer, 1, event->status);
2887 kiblnd_peer_decref(peer);
2888 return event->status; /* rc != 0 destroys cmid */
2890 case RDMA_CM_EVENT_UNREACHABLE:
2891 conn = (kib_conn_t *)cmid->context;
2892 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
2893 conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
2894 CNETERR("%s: UNREACHABLE %d\n",
2895 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
2896 kiblnd_connreq_done(conn, -ENETDOWN);
2897 kiblnd_conn_decref(conn);
2900 case RDMA_CM_EVENT_CONNECT_ERROR:
2901 conn = (kib_conn_t *)cmid->context;
2902 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
2903 conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
2904 CNETERR("%s: CONNECT ERROR %d\n",
2905 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
2906 kiblnd_connreq_done(conn, -ENOTCONN);
2907 kiblnd_conn_decref(conn);
2910 case RDMA_CM_EVENT_REJECTED:
2911 conn = (kib_conn_t *)cmid->context;
2912 switch (conn->ibc_state) {
2916 case IBLND_CONN_PASSIVE_WAIT:
2917 CERROR ("%s: REJECTED %d\n",
2918 libcfs_nid2str(conn->ibc_peer->ibp_nid),
2920 kiblnd_connreq_done(conn, -ECONNRESET);
2923 case IBLND_CONN_ACTIVE_CONNECT:
2924 kiblnd_rejected(conn, event->status,
2925 (void *)KIBLND_CONN_PARAM(event),
2926 KIBLND_CONN_PARAM_LEN(event));
2929 kiblnd_conn_decref(conn);
2932 case RDMA_CM_EVENT_ESTABLISHED:
2933 conn = (kib_conn_t *)cmid->context;
2934 switch (conn->ibc_state) {
2938 case IBLND_CONN_PASSIVE_WAIT:
2939 CDEBUG(D_NET, "ESTABLISHED (passive): %s\n",
2940 libcfs_nid2str(conn->ibc_peer->ibp_nid));
2941 kiblnd_connreq_done(conn, 0);
2944 case IBLND_CONN_ACTIVE_CONNECT:
2945 CDEBUG(D_NET, "ESTABLISHED(active): %s\n",
2946 libcfs_nid2str(conn->ibc_peer->ibp_nid));
2947 kiblnd_check_connreply(conn,
2948 (void *)KIBLND_CONN_PARAM(event),
2949 KIBLND_CONN_PARAM_LEN(event));
2952 /* net keeps its ref on conn! */
2955 #ifdef HAVE_OFED_RDMA_CMEV_TIMEWAIT_EXIT
2956 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
2957 CDEBUG(D_NET, "Ignore TIMEWAIT_EXIT event\n");
2960 case RDMA_CM_EVENT_DISCONNECTED:
2961 conn = (kib_conn_t *)cmid->context;
2962 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
2963 CERROR("%s DISCONNECTED\n",
2964 libcfs_nid2str(conn->ibc_peer->ibp_nid));
2965 kiblnd_connreq_done(conn, -ECONNRESET);
2967 kiblnd_close_conn(conn, 0);
2969 kiblnd_conn_decref(conn);
2970 cmid->context = NULL;
2973 case RDMA_CM_EVENT_DEVICE_REMOVAL:
2974 LCONSOLE_ERROR_MSG(0x131,
2975 "Received notification of device removal\n"
2976 "Please shutdown LNET to allow this to proceed\n");
2977 /* Can't remove network from underneath LNET for now, so I have
2981 #ifdef HAVE_OFED_RDMA_CMEV_ADDRCHANGE
2982 case RDMA_CM_EVENT_ADDR_CHANGE:
2983 LCONSOLE_INFO("Physical link changed (eg hca/port)\n");
2990 kiblnd_check_txs_locked(kib_conn_t *conn, cfs_list_t *txs)
2995 cfs_list_for_each (ttmp, txs) {
2996 tx = cfs_list_entry (ttmp, kib_tx_t, tx_list);
2998 if (txs != &conn->ibc_active_txs) {
2999 LASSERT (tx->tx_queued);
3001 LASSERT (!tx->tx_queued);
3002 LASSERT (tx->tx_waiting || tx->tx_sending != 0);
3005 if (cfs_time_aftereq (jiffies, tx->tx_deadline)) {
3006 CERROR("Timed out tx: %s, %lu seconds\n",
3007 kiblnd_queue2str(conn, txs),
3008 cfs_duration_sec(jiffies - tx->tx_deadline));
3017 kiblnd_conn_timed_out_locked(kib_conn_t *conn)
3019 return kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue) ||
3020 kiblnd_check_txs_locked(conn, &conn->ibc_tx_noops) ||
3021 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_rsrvd) ||
3022 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_nocred) ||
3023 kiblnd_check_txs_locked(conn, &conn->ibc_active_txs);
3027 kiblnd_check_conns (int idx)
3029 CFS_LIST_HEAD (closes);
3030 CFS_LIST_HEAD (checksends);
3031 cfs_list_t *peers = &kiblnd_data.kib_peers[idx];
3036 unsigned long flags;
3038 /* NB. We expect to have a look at all the peers and not find any
3039 * RDMAs to time out, so we just use a shared lock while we
3041 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3043 cfs_list_for_each (ptmp, peers) {
3044 peer = cfs_list_entry (ptmp, kib_peer_t, ibp_list);
3046 cfs_list_for_each (ctmp, &peer->ibp_conns) {
3050 conn = cfs_list_entry(ctmp, kib_conn_t, ibc_list);
3052 LASSERT (conn->ibc_state == IBLND_CONN_ESTABLISHED);
3054 spin_lock(&conn->ibc_lock);
3056 sendnoop = kiblnd_need_noop(conn);
3057 timedout = kiblnd_conn_timed_out_locked(conn);
3058 if (!sendnoop && !timedout) {
3059 spin_unlock(&conn->ibc_lock);
3064 CERROR("Timed out RDMA with %s (%lu): "
3065 "c: %u, oc: %u, rc: %u\n",
3066 libcfs_nid2str(peer->ibp_nid),
3067 cfs_duration_sec(cfs_time_current() -
3068 peer->ibp_last_alive),
3070 conn->ibc_outstanding_credits,
3071 conn->ibc_reserved_credits);
3072 cfs_list_add(&conn->ibc_connd_list, &closes);
3074 cfs_list_add(&conn->ibc_connd_list,
3077 /* +ref for 'closes' or 'checksends' */
3078 kiblnd_conn_addref(conn);
3080 spin_unlock(&conn->ibc_lock);
3084 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3086 /* Handle timeout by closing the whole
3087 * connection. We can only be sure RDMA activity
3088 * has ceased once the QP has been modified. */
3089 while (!cfs_list_empty(&closes)) {
3090 conn = cfs_list_entry(closes.next,
3091 kib_conn_t, ibc_connd_list);
3092 cfs_list_del(&conn->ibc_connd_list);
3093 kiblnd_close_conn(conn, -ETIMEDOUT);
3094 kiblnd_conn_decref(conn);
3097 /* In case we have enough credits to return via a
3098 * NOOP, but there were no non-blocking tx descs
3099 * free to do it last time... */
3100 while (!cfs_list_empty(&checksends)) {
3101 conn = cfs_list_entry(checksends.next,
3102 kib_conn_t, ibc_connd_list);
3103 cfs_list_del(&conn->ibc_connd_list);
3104 kiblnd_check_sends(conn);
3105 kiblnd_conn_decref(conn);
3110 kiblnd_disconnect_conn (kib_conn_t *conn)
3112 LASSERT (!cfs_in_interrupt());
3113 LASSERT (current == kiblnd_data.kib_connd);
3114 LASSERT (conn->ibc_state == IBLND_CONN_CLOSING);
3116 rdma_disconnect(conn->ibc_cmid);
3117 kiblnd_finalise_conn(conn);
3119 kiblnd_peer_notify(conn->ibc_peer);
3123 kiblnd_connd (void *arg)
3125 cfs_waitlink_t wait;
3126 unsigned long flags;
3132 unsigned long deadline = jiffies;
3134 cfs_daemonize ("kiblnd_connd");
3135 cfs_block_allsigs ();
3137 cfs_waitlink_init (&wait);
3138 kiblnd_data.kib_connd = current;
3140 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3142 while (!kiblnd_data.kib_shutdown) {
3146 if (!cfs_list_empty (&kiblnd_data.kib_connd_zombies)) {
3147 conn = cfs_list_entry(kiblnd_data. \
3148 kib_connd_zombies.next,
3149 kib_conn_t, ibc_list);
3150 cfs_list_del(&conn->ibc_list);
3152 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock,
3156 kiblnd_destroy_conn(conn);
3158 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3161 if (!cfs_list_empty(&kiblnd_data.kib_connd_conns)) {
3162 conn = cfs_list_entry(kiblnd_data.kib_connd_conns.next,
3163 kib_conn_t, ibc_list);
3164 cfs_list_del(&conn->ibc_list);
3166 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock,
3170 kiblnd_disconnect_conn(conn);
3171 kiblnd_conn_decref(conn);
3173 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3176 /* careful with the jiffy wrap... */
3177 timeout = (int)(deadline - jiffies);
3181 int chunk = kiblnd_data.kib_peer_hash_size;
3183 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
3186 /* Time to check for RDMA timeouts on a few more
3187 * peers: I do checks every 'p' seconds on a
3188 * proportion of the peer table and I need to check
3189 * every connection 'n' times within a timeout
3190 * interval, to ensure I detect a timeout on any
3191 * connection within (n+1)/n times the timeout
3194 if (*kiblnd_tunables.kib_timeout > n * p)
3195 chunk = (chunk * n * p) /
3196 *kiblnd_tunables.kib_timeout;
3200 for (i = 0; i < chunk; i++) {
3201 kiblnd_check_conns(peer_index);
3202 peer_index = (peer_index + 1) %
3203 kiblnd_data.kib_peer_hash_size;
3206 deadline += p * CFS_HZ;
3207 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3213 /* Nothing to do for 'timeout' */
3214 cfs_set_current_state(CFS_TASK_INTERRUPTIBLE);
3215 cfs_waitq_add(&kiblnd_data.kib_connd_waitq, &wait);
3216 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
3218 cfs_waitq_timedwait(&wait, CFS_TASK_INTERRUPTIBLE, timeout);
3220 cfs_set_current_state(CFS_TASK_RUNNING);
3221 cfs_waitq_del(&kiblnd_data.kib_connd_waitq, &wait);
3222 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3225 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
3227 kiblnd_thread_fini();
3232 kiblnd_qp_event(struct ib_event *event, void *arg)
3234 kib_conn_t *conn = arg;
3236 switch (event->event) {
3237 case IB_EVENT_COMM_EST:
3238 CDEBUG(D_NET, "%s established\n",
3239 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3243 CERROR("%s: Async QP event type %d\n",
3244 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3250 kiblnd_complete (struct ib_wc *wc)
3252 switch (kiblnd_wreqid2type(wc->wr_id)) {
3256 case IBLND_WID_RDMA:
3257 /* We only get RDMA completion notification if it fails. All
3258 * subsequent work items, including the final SEND will fail
3259 * too. However we can't print out any more info about the
3260 * failing RDMA because 'tx' might be back on the idle list or
3261 * even reused already if we didn't manage to post all our work
3263 CNETERR("RDMA (tx: %p) failed: %d\n",
3264 kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3268 kiblnd_tx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3272 kiblnd_rx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status,
3279 kiblnd_cq_completion(struct ib_cq *cq, void *arg)
3281 /* NB I'm not allowed to schedule this conn once its refcount has
3282 * reached 0. Since fundamentally I'm racing with scheduler threads
3283 * consuming my CQ I could be called after all completions have
3284 * occurred. But in this case, ibc_nrx == 0 && ibc_nsends_posted == 0
3285 * and this CQ is about to be destroyed so I NOOP. */
3286 kib_conn_t *conn = (kib_conn_t *)arg;
3287 struct kib_sched_info *sched = conn->ibc_sched;
3288 unsigned long flags;
3290 LASSERT(cq == conn->ibc_cq);
3292 spin_lock_irqsave(&sched->ibs_lock, flags);
3294 conn->ibc_ready = 1;
3296 if (!conn->ibc_scheduled &&
3297 (conn->ibc_nrx > 0 ||
3298 conn->ibc_nsends_posted > 0)) {
3299 kiblnd_conn_addref(conn); /* +1 ref for sched_conns */
3300 conn->ibc_scheduled = 1;
3301 cfs_list_add_tail(&conn->ibc_sched_list, &sched->ibs_conns);
3303 if (cfs_waitq_active(&sched->ibs_waitq))
3304 cfs_waitq_signal(&sched->ibs_waitq);
3307 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3311 kiblnd_cq_event(struct ib_event *event, void *arg)
3313 kib_conn_t *conn = arg;
3315 CERROR("%s: async CQ event type %d\n",
3316 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3320 kiblnd_scheduler(void *arg)
3322 long id = (long)arg;
3323 struct kib_sched_info *sched;
3325 cfs_waitlink_t wait;
3326 unsigned long flags;
3333 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
3334 KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
3336 cfs_daemonize(name);
3337 cfs_block_allsigs();
3339 cfs_waitlink_init(&wait);
3341 sched = kiblnd_data.kib_scheds[KIB_THREAD_CPT(id)];
3343 rc = cfs_cpt_bind(lnet_cpt_table(), sched->ibs_cpt);
3345 CWARN("Failed to bind %s on CPT %d, please verify whether "
3346 "all CPUs are healthy and reload modules if necessary, "
3347 "otherwise your system might under risk of low "
3348 "performance\n", name, sched->ibs_cpt);
3351 spin_lock_irqsave(&sched->ibs_lock, flags);
3353 while (!kiblnd_data.kib_shutdown) {
3354 if (busy_loops++ >= IBLND_RESCHED) {
3355 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3360 spin_lock_irqsave(&sched->ibs_lock, flags);
3365 if (!cfs_list_empty(&sched->ibs_conns)) {
3366 conn = cfs_list_entry(sched->ibs_conns.next,
3367 kib_conn_t, ibc_sched_list);
3368 /* take over kib_sched_conns' ref on conn... */
3369 LASSERT(conn->ibc_scheduled);
3370 cfs_list_del(&conn->ibc_sched_list);
3371 conn->ibc_ready = 0;
3373 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3375 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3377 rc = ib_req_notify_cq(conn->ibc_cq,
3380 CWARN("%s: ib_req_notify_cq failed: %d, "
3381 "closing connection\n",
3382 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
3383 kiblnd_close_conn(conn, -EIO);
3384 kiblnd_conn_decref(conn);
3385 spin_lock_irqsave(&sched->ibs_lock,
3390 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3394 CWARN("%s: ib_poll_cq failed: %d, "
3395 "closing connection\n",
3396 libcfs_nid2str(conn->ibc_peer->ibp_nid),
3398 kiblnd_close_conn(conn, -EIO);
3399 kiblnd_conn_decref(conn);
3400 spin_lock_irqsave(&sched->ibs_lock, flags);
3404 spin_lock_irqsave(&sched->ibs_lock, flags);
3406 if (rc != 0 || conn->ibc_ready) {
3407 /* There may be another completion waiting; get
3408 * another scheduler to check while I handle
3410 /* +1 ref for sched_conns */
3411 kiblnd_conn_addref(conn);
3412 cfs_list_add_tail(&conn->ibc_sched_list,
3414 if (cfs_waitq_active(&sched->ibs_waitq))
3415 cfs_waitq_signal(&sched->ibs_waitq);
3417 conn->ibc_scheduled = 0;
3421 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3422 kiblnd_complete(&wc);
3424 spin_lock_irqsave(&sched->ibs_lock, flags);
3427 kiblnd_conn_decref(conn); /* ...drop my ref from above */
3434 cfs_set_current_state(CFS_TASK_INTERRUPTIBLE);
3435 cfs_waitq_add_exclusive(&sched->ibs_waitq, &wait);
3436 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3438 cfs_waitq_wait(&wait, CFS_TASK_INTERRUPTIBLE);
3441 cfs_waitq_del(&sched->ibs_waitq, &wait);
3442 cfs_set_current_state(CFS_TASK_RUNNING);
3443 spin_lock_irqsave(&sched->ibs_lock, flags);
3446 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3448 kiblnd_thread_fini();
3453 kiblnd_failover_thread(void *arg)
3455 rwlock_t *glock = &kiblnd_data.kib_global_lock;
3457 cfs_waitlink_t wait;
3458 unsigned long flags;
3461 LASSERT (*kiblnd_tunables.kib_dev_failover != 0);
3463 cfs_daemonize ("kiblnd_failover");
3464 cfs_block_allsigs ();
3466 cfs_waitlink_init(&wait);
3467 write_lock_irqsave(glock, flags);
3469 while (!kiblnd_data.kib_shutdown) {
3470 int do_failover = 0;
3473 cfs_list_for_each_entry(dev, &kiblnd_data.kib_failed_devs,
3475 if (cfs_time_before(cfs_time_current(),
3476 dev->ibd_next_failover))
3483 cfs_list_del_init(&dev->ibd_fail_list);
3484 dev->ibd_failover = 1;
3485 write_unlock_irqrestore(glock, flags);
3487 rc = kiblnd_dev_failover(dev);
3489 write_lock_irqsave(glock, flags);
3491 LASSERT (dev->ibd_failover);
3492 dev->ibd_failover = 0;
3493 if (rc >= 0) { /* Device is OK or failover succeed */
3494 dev->ibd_next_failover = cfs_time_shift(3);
3498 /* failed to failover, retry later */
3499 dev->ibd_next_failover =
3500 cfs_time_shift(min(dev->ibd_failed_failover, 10));
3501 if (kiblnd_dev_can_failover(dev)) {
3502 cfs_list_add_tail(&dev->ibd_fail_list,
3503 &kiblnd_data.kib_failed_devs);
3509 /* long sleep if no more pending failover */
3510 long_sleep = cfs_list_empty(&kiblnd_data.kib_failed_devs);
3512 cfs_set_current_state(CFS_TASK_INTERRUPTIBLE);
3513 cfs_waitq_add(&kiblnd_data.kib_failover_waitq, &wait);
3514 write_unlock_irqrestore(glock, flags);
3516 rc = schedule_timeout(long_sleep ? cfs_time_seconds(10) :
3517 cfs_time_seconds(1));
3518 cfs_set_current_state(CFS_TASK_RUNNING);
3519 cfs_waitq_del(&kiblnd_data.kib_failover_waitq, &wait);
3520 write_lock_irqsave(glock, flags);
3522 if (!long_sleep || rc != 0)
3525 /* have a long sleep, routine check all active devices,
3526 * we need checking like this because if there is not active
3527 * connection on the dev and no SEND from local, we may listen
3528 * on wrong HCA for ever while there is a bonding failover */
3529 cfs_list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3530 if (kiblnd_dev_can_failover(dev)) {
3531 cfs_list_add_tail(&dev->ibd_fail_list,
3532 &kiblnd_data.kib_failed_devs);
3537 write_unlock_irqrestore(glock, flags);
3539 kiblnd_thread_fini();