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, 2014, 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>
43 static void kiblnd_peer_alive(kib_peer_t *peer);
44 static void kiblnd_peer_connect_failed(kib_peer_t *peer, int active, int error);
45 static void kiblnd_check_sends(kib_conn_t *conn);
46 static void kiblnd_init_tx_msg(lnet_ni_t *ni, kib_tx_t *tx,
47 int type, int body_nob);
48 static int kiblnd_init_rdma(kib_conn_t *conn, kib_tx_t *tx, int type,
49 int resid, kib_rdma_desc_t *dstrd, __u64 dstcookie);
50 static void kiblnd_queue_tx_locked(kib_tx_t *tx, kib_conn_t *conn);
51 static void kiblnd_queue_tx(kib_tx_t *tx, kib_conn_t *conn);
54 kiblnd_tx_done (lnet_ni_t *ni, kib_tx_t *tx)
56 lnet_msg_t *lntmsg[2];
57 kib_net_t *net = ni->ni_data;
61 LASSERT (net != NULL);
62 LASSERT (!in_interrupt());
63 LASSERT (!tx->tx_queued); /* mustn't be queued for sending */
64 LASSERT (tx->tx_sending == 0); /* mustn't be awaiting sent callback */
65 LASSERT (!tx->tx_waiting); /* mustn't be awaiting peer response */
66 LASSERT (tx->tx_pool != NULL);
68 kiblnd_unmap_tx(ni, tx);
70 /* tx may have up to 2 lnet msgs to finalise */
71 lntmsg[0] = tx->tx_lntmsg[0]; tx->tx_lntmsg[0] = NULL;
72 lntmsg[1] = tx->tx_lntmsg[1]; tx->tx_lntmsg[1] = NULL;
75 if (tx->tx_conn != NULL) {
76 LASSERT (ni == tx->tx_conn->ibc_peer->ibp_ni);
78 kiblnd_conn_decref(tx->tx_conn);
85 kiblnd_pool_free_node(&tx->tx_pool->tpo_pool, &tx->tx_list);
87 /* delay finalize until my descs have been freed */
88 for (i = 0; i < 2; i++) {
89 if (lntmsg[i] == NULL)
92 lnet_finalize(ni, lntmsg[i], rc);
97 kiblnd_txlist_done(lnet_ni_t *ni, struct list_head *txlist, int status)
101 while (!list_empty(txlist)) {
102 tx = list_entry(txlist->next, kib_tx_t, tx_list);
104 list_del(&tx->tx_list);
107 tx->tx_status = status;
108 kiblnd_tx_done(ni, tx);
113 kiblnd_get_idle_tx(lnet_ni_t *ni, lnet_nid_t target)
115 kib_net_t *net = (kib_net_t *)ni->ni_data;
116 struct list_head *node;
118 kib_tx_poolset_t *tps;
120 tps = net->ibn_tx_ps[lnet_cpt_of_nid(target)];
121 node = kiblnd_pool_alloc_node(&tps->tps_poolset);
124 tx = container_of(node, kib_tx_t, tx_list);
126 LASSERT (tx->tx_nwrq == 0);
127 LASSERT (!tx->tx_queued);
128 LASSERT (tx->tx_sending == 0);
129 LASSERT (!tx->tx_waiting);
130 LASSERT (tx->tx_status == 0);
131 LASSERT (tx->tx_conn == NULL);
132 LASSERT (tx->tx_lntmsg[0] == NULL);
133 LASSERT (tx->tx_lntmsg[1] == NULL);
134 LASSERT (tx->tx_u.pmr == NULL);
135 LASSERT (tx->tx_nfrags == 0);
141 kiblnd_drop_rx(kib_rx_t *rx)
143 kib_conn_t *conn = rx->rx_conn;
144 struct kib_sched_info *sched = conn->ibc_sched;
147 spin_lock_irqsave(&sched->ibs_lock, flags);
148 LASSERT(conn->ibc_nrx > 0);
150 spin_unlock_irqrestore(&sched->ibs_lock, flags);
152 kiblnd_conn_decref(conn);
156 kiblnd_post_rx (kib_rx_t *rx, int credit)
158 kib_conn_t *conn = rx->rx_conn;
159 kib_net_t *net = conn->ibc_peer->ibp_ni->ni_data;
160 struct ib_recv_wr *bad_wrq = NULL;
164 LASSERT (net != NULL);
165 LASSERT (!in_interrupt());
166 LASSERT (credit == IBLND_POSTRX_NO_CREDIT ||
167 credit == IBLND_POSTRX_PEER_CREDIT ||
168 credit == IBLND_POSTRX_RSRVD_CREDIT);
170 mr = kiblnd_find_dma_mr(conn->ibc_hdev, rx->rx_msgaddr, IBLND_MSG_SIZE);
171 LASSERT (mr != NULL);
173 rx->rx_sge.lkey = mr->lkey;
174 rx->rx_sge.addr = rx->rx_msgaddr;
175 rx->rx_sge.length = IBLND_MSG_SIZE;
177 rx->rx_wrq.next = NULL;
178 rx->rx_wrq.sg_list = &rx->rx_sge;
179 rx->rx_wrq.num_sge = 1;
180 rx->rx_wrq.wr_id = kiblnd_ptr2wreqid(rx, IBLND_WID_RX);
182 LASSERT (conn->ibc_state >= IBLND_CONN_INIT);
183 LASSERT (rx->rx_nob >= 0); /* not posted */
185 if (conn->ibc_state > IBLND_CONN_ESTABLISHED) {
186 kiblnd_drop_rx(rx); /* No more posts for this rx */
190 rx->rx_nob = -1; /* flag posted */
192 /* NB: need an extra reference after ib_post_recv because we don't
193 * own this rx (and rx::rx_conn) anymore, LU-5678.
195 kiblnd_conn_addref(conn);
196 rc = ib_post_recv(conn->ibc_cmid->qp, &rx->rx_wrq, &bad_wrq);
197 if (unlikely(rc != 0)) {
198 CERROR("Can't post rx for %s: %d, bad_wrq: %p\n",
199 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc, bad_wrq);
203 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) /* Initial post */
206 if (unlikely(rc != 0)) {
207 kiblnd_close_conn(conn, rc);
208 kiblnd_drop_rx(rx); /* No more posts for this rx */
212 if (credit == IBLND_POSTRX_NO_CREDIT)
215 spin_lock(&conn->ibc_lock);
216 if (credit == IBLND_POSTRX_PEER_CREDIT)
217 conn->ibc_outstanding_credits++;
219 conn->ibc_reserved_credits++;
220 spin_unlock(&conn->ibc_lock);
222 kiblnd_check_sends(conn);
224 kiblnd_conn_decref(conn);
229 kiblnd_find_waiting_tx_locked(kib_conn_t *conn, int txtype, __u64 cookie)
231 struct list_head *tmp;
233 list_for_each(tmp, &conn->ibc_active_txs) {
234 kib_tx_t *tx = list_entry(tmp, kib_tx_t, tx_list);
236 LASSERT(!tx->tx_queued);
237 LASSERT(tx->tx_sending != 0 || tx->tx_waiting);
239 if (tx->tx_cookie != cookie)
242 if (tx->tx_waiting &&
243 tx->tx_msg->ibm_type == txtype)
246 CWARN("Bad completion: %swaiting, type %x (wanted %x)\n",
247 tx->tx_waiting ? "" : "NOT ",
248 tx->tx_msg->ibm_type, txtype);
254 kiblnd_handle_completion(kib_conn_t *conn, int txtype, int status, __u64 cookie)
257 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
260 spin_lock(&conn->ibc_lock);
262 tx = kiblnd_find_waiting_tx_locked(conn, txtype, cookie);
264 spin_unlock(&conn->ibc_lock);
266 CWARN("Unmatched completion type %x cookie "LPX64" from %s\n",
267 txtype, cookie, libcfs_nid2str(conn->ibc_peer->ibp_nid));
268 kiblnd_close_conn(conn, -EPROTO);
272 if (tx->tx_status == 0) { /* success so far */
273 if (status < 0) { /* failed? */
274 tx->tx_status = status;
275 } else if (txtype == IBLND_MSG_GET_REQ) {
276 lnet_set_reply_msg_len(ni, tx->tx_lntmsg[1], status);
282 idle = !tx->tx_queued && (tx->tx_sending == 0);
284 list_del(&tx->tx_list);
286 spin_unlock(&conn->ibc_lock);
289 kiblnd_tx_done(ni, tx);
293 kiblnd_send_completion(kib_conn_t *conn, int type, int status, __u64 cookie)
295 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
296 kib_tx_t *tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
299 CERROR("Can't get tx for completion %x for %s\n",
300 type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
304 tx->tx_msg->ibm_u.completion.ibcm_status = status;
305 tx->tx_msg->ibm_u.completion.ibcm_cookie = cookie;
306 kiblnd_init_tx_msg(ni, tx, type, sizeof(kib_completion_msg_t));
308 kiblnd_queue_tx(tx, conn);
312 kiblnd_handle_rx (kib_rx_t *rx)
314 kib_msg_t *msg = rx->rx_msg;
315 kib_conn_t *conn = rx->rx_conn;
316 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
317 int credits = msg->ibm_credits;
323 LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
325 CDEBUG (D_NET, "Received %x[%d] from %s\n",
326 msg->ibm_type, credits,
327 libcfs_nid2str(conn->ibc_peer->ibp_nid));
330 /* Have I received credits that will let me send? */
331 spin_lock(&conn->ibc_lock);
333 if (conn->ibc_credits + credits >
334 IBLND_MSG_QUEUE_SIZE(conn->ibc_version)) {
335 rc2 = conn->ibc_credits;
336 spin_unlock(&conn->ibc_lock);
338 CERROR("Bad credits from %s: %d + %d > %d\n",
339 libcfs_nid2str(conn->ibc_peer->ibp_nid),
341 IBLND_MSG_QUEUE_SIZE(conn->ibc_version));
343 kiblnd_close_conn(conn, -EPROTO);
344 kiblnd_post_rx(rx, IBLND_POSTRX_NO_CREDIT);
348 conn->ibc_credits += credits;
350 /* This ensures the credit taken by NOOP can be returned */
351 if (msg->ibm_type == IBLND_MSG_NOOP &&
352 !IBLND_OOB_CAPABLE(conn->ibc_version)) /* v1 only */
353 conn->ibc_outstanding_credits++;
355 spin_unlock(&conn->ibc_lock);
356 kiblnd_check_sends(conn);
359 switch (msg->ibm_type) {
361 CERROR("Bad IBLND message type %x from %s\n",
362 msg->ibm_type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
363 post_credit = IBLND_POSTRX_NO_CREDIT;
368 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
369 post_credit = IBLND_POSTRX_NO_CREDIT;
373 if (credits != 0) /* credit already posted */
374 post_credit = IBLND_POSTRX_NO_CREDIT;
375 else /* a keepalive NOOP */
376 post_credit = IBLND_POSTRX_PEER_CREDIT;
379 case IBLND_MSG_IMMEDIATE:
380 post_credit = IBLND_POSTRX_DONT_POST;
381 rc = lnet_parse(ni, &msg->ibm_u.immediate.ibim_hdr,
382 msg->ibm_srcnid, rx, 0);
383 if (rc < 0) /* repost on error */
384 post_credit = IBLND_POSTRX_PEER_CREDIT;
387 case IBLND_MSG_PUT_REQ:
388 post_credit = IBLND_POSTRX_DONT_POST;
389 rc = lnet_parse(ni, &msg->ibm_u.putreq.ibprm_hdr,
390 msg->ibm_srcnid, rx, 1);
391 if (rc < 0) /* repost on error */
392 post_credit = IBLND_POSTRX_PEER_CREDIT;
395 case IBLND_MSG_PUT_NAK:
396 CWARN ("PUT_NACK from %s\n",
397 libcfs_nid2str(conn->ibc_peer->ibp_nid));
398 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
399 kiblnd_handle_completion(conn, IBLND_MSG_PUT_REQ,
400 msg->ibm_u.completion.ibcm_status,
401 msg->ibm_u.completion.ibcm_cookie);
404 case IBLND_MSG_PUT_ACK:
405 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
407 spin_lock(&conn->ibc_lock);
408 tx = kiblnd_find_waiting_tx_locked(conn, IBLND_MSG_PUT_REQ,
409 msg->ibm_u.putack.ibpam_src_cookie);
411 list_del(&tx->tx_list);
412 spin_unlock(&conn->ibc_lock);
415 CERROR("Unmatched PUT_ACK from %s\n",
416 libcfs_nid2str(conn->ibc_peer->ibp_nid));
421 LASSERT (tx->tx_waiting);
422 /* CAVEAT EMPTOR: I could be racing with tx_complete, but...
423 * (a) I can overwrite tx_msg since my peer has received it!
424 * (b) tx_waiting set tells tx_complete() it's not done. */
426 tx->tx_nwrq = 0; /* overwrite PUT_REQ */
428 rc2 = kiblnd_init_rdma(conn, tx, IBLND_MSG_PUT_DONE,
429 kiblnd_rd_size(&msg->ibm_u.putack.ibpam_rd),
430 &msg->ibm_u.putack.ibpam_rd,
431 msg->ibm_u.putack.ibpam_dst_cookie);
433 CERROR("Can't setup rdma for PUT to %s: %d\n",
434 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc2);
436 spin_lock(&conn->ibc_lock);
437 tx->tx_waiting = 0; /* clear waiting and queue atomically */
438 kiblnd_queue_tx_locked(tx, conn);
439 spin_unlock(&conn->ibc_lock);
442 case IBLND_MSG_PUT_DONE:
443 post_credit = IBLND_POSTRX_PEER_CREDIT;
444 kiblnd_handle_completion(conn, IBLND_MSG_PUT_ACK,
445 msg->ibm_u.completion.ibcm_status,
446 msg->ibm_u.completion.ibcm_cookie);
449 case IBLND_MSG_GET_REQ:
450 post_credit = IBLND_POSTRX_DONT_POST;
451 rc = lnet_parse(ni, &msg->ibm_u.get.ibgm_hdr,
452 msg->ibm_srcnid, rx, 1);
453 if (rc < 0) /* repost on error */
454 post_credit = IBLND_POSTRX_PEER_CREDIT;
457 case IBLND_MSG_GET_DONE:
458 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
459 kiblnd_handle_completion(conn, IBLND_MSG_GET_REQ,
460 msg->ibm_u.completion.ibcm_status,
461 msg->ibm_u.completion.ibcm_cookie);
465 if (rc < 0) /* protocol error */
466 kiblnd_close_conn(conn, rc);
468 if (post_credit != IBLND_POSTRX_DONT_POST)
469 kiblnd_post_rx(rx, post_credit);
473 kiblnd_rx_complete (kib_rx_t *rx, int status, int nob)
475 kib_msg_t *msg = rx->rx_msg;
476 kib_conn_t *conn = rx->rx_conn;
477 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
478 kib_net_t *net = ni->ni_data;
482 LASSERT (net != NULL);
483 LASSERT (rx->rx_nob < 0); /* was posted */
484 rx->rx_nob = 0; /* isn't now */
486 if (conn->ibc_state > IBLND_CONN_ESTABLISHED)
489 if (status != IB_WC_SUCCESS) {
490 CNETERR("Rx from %s failed: %d\n",
491 libcfs_nid2str(conn->ibc_peer->ibp_nid), status);
498 rc = kiblnd_unpack_msg(msg, rx->rx_nob);
500 CERROR ("Error %d unpacking rx from %s\n",
501 rc, libcfs_nid2str(conn->ibc_peer->ibp_nid));
505 if (msg->ibm_srcnid != conn->ibc_peer->ibp_nid ||
506 msg->ibm_dstnid != ni->ni_nid ||
507 msg->ibm_srcstamp != conn->ibc_incarnation ||
508 msg->ibm_dststamp != net->ibn_incarnation) {
509 CERROR ("Stale rx from %s\n",
510 libcfs_nid2str(conn->ibc_peer->ibp_nid));
515 /* set time last known alive */
516 kiblnd_peer_alive(conn->ibc_peer);
518 /* racing with connection establishment/teardown! */
520 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
521 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
524 write_lock_irqsave(g_lock, flags);
525 /* must check holding global lock to eliminate race */
526 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
527 list_add_tail(&rx->rx_list, &conn->ibc_early_rxs);
528 write_unlock_irqrestore(g_lock, flags);
531 write_unlock_irqrestore(g_lock, flags);
533 kiblnd_handle_rx(rx);
537 CDEBUG(D_NET, "rx %p conn %p\n", rx, conn);
538 kiblnd_close_conn(conn, err);
540 kiblnd_drop_rx(rx); /* Don't re-post rx. */
544 kiblnd_kvaddr_to_page (unsigned long vaddr)
548 if (is_vmalloc_addr((void *)vaddr)) {
549 page = vmalloc_to_page ((void *)vaddr);
550 LASSERT (page != NULL);
553 #ifdef CONFIG_HIGHMEM
554 if (vaddr >= PKMAP_BASE &&
555 vaddr < (PKMAP_BASE + LAST_PKMAP * PAGE_SIZE)) {
556 /* No highmem pages only used for bulk (kiov) I/O */
557 CERROR("find page for address in highmem\n");
561 page = virt_to_page (vaddr);
562 LASSERT (page != NULL);
567 kiblnd_fmr_map_tx(kib_net_t *net, kib_tx_t *tx, kib_rdma_desc_t *rd, int nob)
570 __u64 *pages = tx->tx_pages;
571 kib_fmr_poolset_t *fps;
578 LASSERT(tx->tx_pool != NULL);
579 LASSERT(tx->tx_pool->tpo_pool.po_owner != NULL);
581 hdev = tx->tx_pool->tpo_hdev;
583 for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
584 for (size = 0; size < rd->rd_frags[i].rf_nob;
585 size += hdev->ibh_page_size) {
586 pages[npages ++] = (rd->rd_frags[i].rf_addr &
587 hdev->ibh_page_mask) + size;
591 cpt = tx->tx_pool->tpo_pool.po_owner->ps_cpt;
593 fps = net->ibn_fmr_ps[cpt];
594 rc = kiblnd_fmr_pool_map(fps, pages, npages, 0, &tx->tx_u.fmr);
596 CERROR ("Can't map %d pages: %d\n", npages, rc);
600 /* If rd is not tx_rd, it's going to get sent to a peer, who will need
602 rd->rd_key = (rd != tx->tx_rd) ? tx->tx_u.fmr.fmr_pfmr->fmr->rkey :
603 tx->tx_u.fmr.fmr_pfmr->fmr->lkey;
604 rd->rd_frags[0].rf_addr &= ~hdev->ibh_page_mask;
605 rd->rd_frags[0].rf_nob = nob;
612 kiblnd_pmr_map_tx(kib_net_t *net, kib_tx_t *tx, kib_rdma_desc_t *rd, int nob)
615 kib_pmr_poolset_t *pps;
620 LASSERT(tx->tx_pool != NULL);
621 LASSERT(tx->tx_pool->tpo_pool.po_owner != NULL);
623 hdev = tx->tx_pool->tpo_hdev;
625 iova = rd->rd_frags[0].rf_addr & ~hdev->ibh_page_mask;
627 cpt = tx->tx_pool->tpo_pool.po_owner->ps_cpt;
629 pps = net->ibn_pmr_ps[cpt];
630 rc = kiblnd_pmr_pool_map(pps, hdev, rd, &iova, &tx->tx_u.pmr);
632 CERROR("Failed to create MR by phybuf: %d\n", rc);
636 /* If rd is not tx_rd, it's going to get sent to a peer, who will need
638 rd->rd_key = (rd != tx->tx_rd) ? tx->tx_u.pmr->pmr_mr->rkey :
639 tx->tx_u.pmr->pmr_mr->lkey;
641 rd->rd_frags[0].rf_addr = iova;
642 rd->rd_frags[0].rf_nob = nob;
648 kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx)
650 kib_net_t *net = ni->ni_data;
652 LASSERT(net != NULL);
654 if (net->ibn_fmr_ps != NULL && tx->tx_u.fmr.fmr_pfmr != NULL) {
655 kiblnd_fmr_pool_unmap(&tx->tx_u.fmr, tx->tx_status);
656 tx->tx_u.fmr.fmr_pfmr = NULL;
658 } else if (net->ibn_pmr_ps != NULL && tx->tx_u.pmr != NULL) {
659 kiblnd_pmr_pool_unmap(tx->tx_u.pmr);
663 if (tx->tx_nfrags != 0) {
664 kiblnd_dma_unmap_sg(tx->tx_pool->tpo_hdev->ibh_ibdev,
665 tx->tx_frags, tx->tx_nfrags, tx->tx_dmadir);
671 kiblnd_map_tx(lnet_ni_t *ni, kib_tx_t *tx,
672 kib_rdma_desc_t *rd, int nfrags)
674 kib_hca_dev_t *hdev = tx->tx_pool->tpo_hdev;
675 kib_net_t *net = ni->ni_data;
676 struct ib_mr *mr = NULL;
680 /* If rd is not tx_rd, it's going to get sent to a peer and I'm the
682 tx->tx_dmadir = (rd != tx->tx_rd) ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
683 tx->tx_nfrags = nfrags;
686 kiblnd_dma_map_sg(hdev->ibh_ibdev,
687 tx->tx_frags, tx->tx_nfrags, tx->tx_dmadir);
689 for (i = 0, nob = 0; i < rd->rd_nfrags; i++) {
690 rd->rd_frags[i].rf_nob = kiblnd_sg_dma_len(
691 hdev->ibh_ibdev, &tx->tx_frags[i]);
692 rd->rd_frags[i].rf_addr = kiblnd_sg_dma_address(
693 hdev->ibh_ibdev, &tx->tx_frags[i]);
694 nob += rd->rd_frags[i].rf_nob;
697 /* looking for pre-mapping MR */
698 mr = kiblnd_find_rd_dma_mr(hdev, rd);
700 /* found pre-mapping MR */
701 rd->rd_key = (rd != tx->tx_rd) ? mr->rkey : mr->lkey;
705 if (net->ibn_fmr_ps != NULL)
706 return kiblnd_fmr_map_tx(net, tx, rd, nob);
707 else if (net->ibn_pmr_ps != NULL)
708 return kiblnd_pmr_map_tx(net, tx, rd, nob);
715 kiblnd_setup_rd_iov(lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd,
716 unsigned int niov, struct iovec *iov, int offset, int nob)
718 kib_net_t *net = ni->ni_data;
720 struct scatterlist *sg;
727 LASSERT (net != NULL);
729 while (offset >= iov->iov_len) {
730 offset -= iov->iov_len;
740 vaddr = ((unsigned long)iov->iov_base) + offset;
741 page_offset = vaddr & (PAGE_SIZE - 1);
742 page = kiblnd_kvaddr_to_page(vaddr);
744 CERROR ("Can't find page\n");
748 fragnob = min((int)(iov->iov_len - offset), nob);
749 fragnob = min(fragnob, (int)PAGE_SIZE - page_offset);
751 sg_set_page(sg, page, fragnob, page_offset);
754 if (offset + fragnob < iov->iov_len) {
764 return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags);
768 kiblnd_setup_rd_kiov (lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd,
769 int nkiov, lnet_kiov_t *kiov, int offset, int nob)
771 kib_net_t *net = ni->ni_data;
772 struct scatterlist *sg;
775 CDEBUG(D_NET, "niov %d offset %d nob %d\n", nkiov, offset, nob);
779 LASSERT (net != NULL);
781 while (offset >= kiov->kiov_len) {
782 offset -= kiov->kiov_len;
792 fragnob = min((int)(kiov->kiov_len - offset), nob);
794 sg_set_page(sg, kiov->kiov_page, fragnob,
795 kiov->kiov_offset + offset);
804 return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags);
808 kiblnd_post_tx_locked (kib_conn_t *conn, kib_tx_t *tx, int credit)
809 __must_hold(&conn->ibc_lock)
811 kib_msg_t *msg = tx->tx_msg;
812 kib_peer_t *peer = conn->ibc_peer;
813 int ver = conn->ibc_version;
816 struct ib_send_wr *bad_wrq;
818 LASSERT (tx->tx_queued);
819 /* We rely on this for QP sizing */
820 LASSERT (tx->tx_nwrq > 0);
821 LASSERT (tx->tx_nwrq <= 1 + IBLND_RDMA_FRAGS(ver));
823 LASSERT (credit == 0 || credit == 1);
824 LASSERT (conn->ibc_outstanding_credits >= 0);
825 LASSERT (conn->ibc_outstanding_credits <= IBLND_MSG_QUEUE_SIZE(ver));
826 LASSERT (conn->ibc_credits >= 0);
827 LASSERT (conn->ibc_credits <= IBLND_MSG_QUEUE_SIZE(ver));
829 if (conn->ibc_nsends_posted == IBLND_CONCURRENT_SENDS(ver)) {
830 /* tx completions outstanding... */
831 CDEBUG(D_NET, "%s: posted enough\n",
832 libcfs_nid2str(peer->ibp_nid));
836 if (credit != 0 && conn->ibc_credits == 0) { /* no credits */
837 CDEBUG(D_NET, "%s: no credits\n",
838 libcfs_nid2str(peer->ibp_nid));
842 if (credit != 0 && !IBLND_OOB_CAPABLE(ver) &&
843 conn->ibc_credits == 1 && /* last credit reserved */
844 msg->ibm_type != IBLND_MSG_NOOP) { /* for NOOP */
845 CDEBUG(D_NET, "%s: not using last credit\n",
846 libcfs_nid2str(peer->ibp_nid));
850 /* NB don't drop ibc_lock before bumping tx_sending */
851 list_del(&tx->tx_list);
854 if (msg->ibm_type == IBLND_MSG_NOOP &&
855 (!kiblnd_need_noop(conn) || /* redundant NOOP */
856 (IBLND_OOB_CAPABLE(ver) && /* posted enough NOOP */
857 conn->ibc_noops_posted == IBLND_OOB_MSGS(ver)))) {
858 /* OK to drop when posted enough NOOPs, since
859 * kiblnd_check_sends will queue NOOP again when
860 * posted NOOPs complete */
861 spin_unlock(&conn->ibc_lock);
862 kiblnd_tx_done(peer->ibp_ni, tx);
863 spin_lock(&conn->ibc_lock);
864 CDEBUG(D_NET, "%s(%d): redundant or enough NOOP\n",
865 libcfs_nid2str(peer->ibp_nid),
866 conn->ibc_noops_posted);
870 kiblnd_pack_msg(peer->ibp_ni, msg, ver, conn->ibc_outstanding_credits,
871 peer->ibp_nid, conn->ibc_incarnation);
873 conn->ibc_credits -= credit;
874 conn->ibc_outstanding_credits = 0;
875 conn->ibc_nsends_posted++;
876 if (msg->ibm_type == IBLND_MSG_NOOP)
877 conn->ibc_noops_posted++;
879 /* CAVEAT EMPTOR! This tx could be the PUT_DONE of an RDMA
880 * PUT. If so, it was first queued here as a PUT_REQ, sent and
881 * stashed on ibc_active_txs, matched by an incoming PUT_ACK,
882 * and then re-queued here. It's (just) possible that
883 * tx_sending is non-zero if we've not done the tx_complete()
884 * from the first send; hence the ++ rather than = below. */
886 list_add(&tx->tx_list, &conn->ibc_active_txs);
888 /* I'm still holding ibc_lock! */
889 if (conn->ibc_state != IBLND_CONN_ESTABLISHED) {
891 } else if (tx->tx_pool->tpo_pool.po_failed ||
892 conn->ibc_hdev != tx->tx_pool->tpo_hdev) {
893 /* close_conn will launch failover */
896 rc = ib_post_send(conn->ibc_cmid->qp,
897 tx->tx_wrq, &bad_wrq);
900 conn->ibc_last_send = jiffies;
905 /* NB credits are transferred in the actual
906 * message, which can only be the last work item */
907 conn->ibc_credits += credit;
908 conn->ibc_outstanding_credits += msg->ibm_credits;
909 conn->ibc_nsends_posted--;
910 if (msg->ibm_type == IBLND_MSG_NOOP)
911 conn->ibc_noops_posted--;
917 done = (tx->tx_sending == 0);
919 list_del(&tx->tx_list);
921 spin_unlock(&conn->ibc_lock);
923 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
924 CERROR("Error %d posting transmit to %s\n",
925 rc, libcfs_nid2str(peer->ibp_nid));
927 CDEBUG(D_NET, "Error %d posting transmit to %s\n",
928 rc, libcfs_nid2str(peer->ibp_nid));
930 kiblnd_close_conn(conn, rc);
933 kiblnd_tx_done(peer->ibp_ni, tx);
935 spin_lock(&conn->ibc_lock);
941 kiblnd_check_sends (kib_conn_t *conn)
943 int ver = conn->ibc_version;
944 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
947 /* Don't send anything until after the connection is established */
948 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
949 CDEBUG(D_NET, "%s too soon\n",
950 libcfs_nid2str(conn->ibc_peer->ibp_nid));
954 spin_lock(&conn->ibc_lock);
956 LASSERT (conn->ibc_nsends_posted <= IBLND_CONCURRENT_SENDS(ver));
957 LASSERT (!IBLND_OOB_CAPABLE(ver) ||
958 conn->ibc_noops_posted <= IBLND_OOB_MSGS(ver));
959 LASSERT (conn->ibc_reserved_credits >= 0);
961 while (conn->ibc_reserved_credits > 0 &&
962 !list_empty(&conn->ibc_tx_queue_rsrvd)) {
963 tx = list_entry(conn->ibc_tx_queue_rsrvd.next,
965 list_del(&tx->tx_list);
966 list_add_tail(&tx->tx_list, &conn->ibc_tx_queue);
967 conn->ibc_reserved_credits--;
970 if (kiblnd_need_noop(conn)) {
971 spin_unlock(&conn->ibc_lock);
973 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
975 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_NOOP, 0);
977 spin_lock(&conn->ibc_lock);
979 kiblnd_queue_tx_locked(tx, conn);
982 kiblnd_conn_addref(conn); /* 1 ref for me.... (see b21911) */
987 if (!list_empty(&conn->ibc_tx_queue_nocred)) {
989 tx = list_entry(conn->ibc_tx_queue_nocred.next,
991 } else if (!list_empty(&conn->ibc_tx_noops)) {
992 LASSERT (!IBLND_OOB_CAPABLE(ver));
994 tx = list_entry(conn->ibc_tx_noops.next,
996 } else if (!list_empty(&conn->ibc_tx_queue)) {
998 tx = list_entry(conn->ibc_tx_queue.next,
1003 if (kiblnd_post_tx_locked(conn, tx, credit) != 0)
1007 spin_unlock(&conn->ibc_lock);
1009 kiblnd_conn_decref(conn); /* ...until here */
1013 kiblnd_tx_complete (kib_tx_t *tx, int status)
1015 int failed = (status != IB_WC_SUCCESS);
1016 kib_conn_t *conn = tx->tx_conn;
1019 LASSERT (tx->tx_sending > 0);
1022 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
1023 CNETERR("Tx -> %s cookie "LPX64
1024 " sending %d waiting %d: failed %d\n",
1025 libcfs_nid2str(conn->ibc_peer->ibp_nid),
1026 tx->tx_cookie, tx->tx_sending, tx->tx_waiting,
1029 kiblnd_close_conn(conn, -EIO);
1031 kiblnd_peer_alive(conn->ibc_peer);
1034 spin_lock(&conn->ibc_lock);
1036 /* I could be racing with rdma completion. Whoever makes 'tx' idle
1037 * gets to free it, which also drops its ref on 'conn'. */
1040 conn->ibc_nsends_posted--;
1041 if (tx->tx_msg->ibm_type == IBLND_MSG_NOOP)
1042 conn->ibc_noops_posted--;
1045 tx->tx_waiting = 0; /* don't wait for peer */
1046 tx->tx_status = -EIO;
1049 idle = (tx->tx_sending == 0) && /* This is the final callback */
1050 !tx->tx_waiting && /* Not waiting for peer */
1051 !tx->tx_queued; /* Not re-queued (PUT_DONE) */
1053 list_del(&tx->tx_list);
1055 kiblnd_conn_addref(conn); /* 1 ref for me.... */
1057 spin_unlock(&conn->ibc_lock);
1060 kiblnd_tx_done(conn->ibc_peer->ibp_ni, tx);
1062 kiblnd_check_sends(conn);
1064 kiblnd_conn_decref(conn); /* ...until here */
1068 kiblnd_init_tx_msg (lnet_ni_t *ni, kib_tx_t *tx, int type, int body_nob)
1070 kib_hca_dev_t *hdev = tx->tx_pool->tpo_hdev;
1071 struct ib_sge *sge = &tx->tx_sge[tx->tx_nwrq];
1072 struct ib_send_wr *wrq = &tx->tx_wrq[tx->tx_nwrq];
1073 int nob = offsetof (kib_msg_t, ibm_u) + body_nob;
1076 LASSERT (tx->tx_nwrq >= 0);
1077 LASSERT (tx->tx_nwrq < IBLND_MAX_RDMA_FRAGS + 1);
1078 LASSERT (nob <= IBLND_MSG_SIZE);
1080 kiblnd_init_msg(tx->tx_msg, type, body_nob);
1082 mr = kiblnd_find_dma_mr(hdev, tx->tx_msgaddr, nob);
1083 LASSERT (mr != NULL);
1085 sge->lkey = mr->lkey;
1086 sge->addr = tx->tx_msgaddr;
1089 memset(wrq, 0, sizeof(*wrq));
1092 wrq->wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_TX);
1095 wrq->opcode = IB_WR_SEND;
1096 wrq->send_flags = IB_SEND_SIGNALED;
1102 kiblnd_init_rdma(kib_conn_t *conn, kib_tx_t *tx, int type,
1103 int resid, kib_rdma_desc_t *dstrd, __u64 dstcookie)
1105 kib_msg_t *ibmsg = tx->tx_msg;
1106 kib_rdma_desc_t *srcrd = tx->tx_rd;
1107 struct ib_sge *sge = &tx->tx_sge[0];
1108 struct ib_send_wr *wrq = &tx->tx_wrq[0];
1114 LASSERT (!in_interrupt());
1115 LASSERT (tx->tx_nwrq == 0);
1116 LASSERT (type == IBLND_MSG_GET_DONE ||
1117 type == IBLND_MSG_PUT_DONE);
1119 srcidx = dstidx = 0;
1122 if (srcidx >= srcrd->rd_nfrags) {
1123 CERROR("Src buffer exhausted: %d frags\n", srcidx);
1128 if (dstidx == dstrd->rd_nfrags) {
1129 CERROR("Dst buffer exhausted: %d frags\n", dstidx);
1134 if (tx->tx_nwrq == IBLND_RDMA_FRAGS(conn->ibc_version)) {
1135 CERROR("RDMA too fragmented for %s (%d): "
1136 "%d/%d src %d/%d dst frags\n",
1137 libcfs_nid2str(conn->ibc_peer->ibp_nid),
1138 IBLND_RDMA_FRAGS(conn->ibc_version),
1139 srcidx, srcrd->rd_nfrags,
1140 dstidx, dstrd->rd_nfrags);
1145 wrknob = MIN(MIN(kiblnd_rd_frag_size(srcrd, srcidx),
1146 kiblnd_rd_frag_size(dstrd, dstidx)), resid);
1148 sge = &tx->tx_sge[tx->tx_nwrq];
1149 sge->addr = kiblnd_rd_frag_addr(srcrd, srcidx);
1150 sge->lkey = kiblnd_rd_frag_key(srcrd, srcidx);
1151 sge->length = wrknob;
1153 wrq = &tx->tx_wrq[tx->tx_nwrq];
1155 wrq->next = wrq + 1;
1156 wrq->wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_RDMA);
1159 wrq->opcode = IB_WR_RDMA_WRITE;
1160 wrq->send_flags = 0;
1162 wrq->wr.rdma.remote_addr = kiblnd_rd_frag_addr(dstrd, dstidx);
1163 wrq->wr.rdma.rkey = kiblnd_rd_frag_key(dstrd, dstidx);
1165 srcidx = kiblnd_rd_consume_frag(srcrd, srcidx, wrknob);
1166 dstidx = kiblnd_rd_consume_frag(dstrd, dstidx, wrknob);
1175 if (rc < 0) /* no RDMA if completing with failure */
1178 ibmsg->ibm_u.completion.ibcm_status = rc;
1179 ibmsg->ibm_u.completion.ibcm_cookie = dstcookie;
1180 kiblnd_init_tx_msg(conn->ibc_peer->ibp_ni, tx,
1181 type, sizeof (kib_completion_msg_t));
1187 kiblnd_queue_tx_locked(kib_tx_t *tx, kib_conn_t *conn)
1189 struct list_head *q;
1191 LASSERT(tx->tx_nwrq > 0); /* work items set up */
1192 LASSERT(!tx->tx_queued); /* not queued for sending already */
1193 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1196 tx->tx_deadline = jiffies +
1197 msecs_to_jiffies(*kiblnd_tunables.kib_timeout *
1200 if (tx->tx_conn == NULL) {
1201 kiblnd_conn_addref(conn);
1203 LASSERT (tx->tx_msg->ibm_type != IBLND_MSG_PUT_DONE);
1205 /* PUT_DONE first attached to conn as a PUT_REQ */
1206 LASSERT (tx->tx_conn == conn);
1207 LASSERT (tx->tx_msg->ibm_type == IBLND_MSG_PUT_DONE);
1210 switch (tx->tx_msg->ibm_type) {
1214 case IBLND_MSG_PUT_REQ:
1215 case IBLND_MSG_GET_REQ:
1216 q = &conn->ibc_tx_queue_rsrvd;
1219 case IBLND_MSG_PUT_NAK:
1220 case IBLND_MSG_PUT_ACK:
1221 case IBLND_MSG_PUT_DONE:
1222 case IBLND_MSG_GET_DONE:
1223 q = &conn->ibc_tx_queue_nocred;
1226 case IBLND_MSG_NOOP:
1227 if (IBLND_OOB_CAPABLE(conn->ibc_version))
1228 q = &conn->ibc_tx_queue_nocred;
1230 q = &conn->ibc_tx_noops;
1233 case IBLND_MSG_IMMEDIATE:
1234 q = &conn->ibc_tx_queue;
1238 list_add_tail(&tx->tx_list, q);
1242 kiblnd_queue_tx (kib_tx_t *tx, kib_conn_t *conn)
1244 spin_lock(&conn->ibc_lock);
1245 kiblnd_queue_tx_locked(tx, conn);
1246 spin_unlock(&conn->ibc_lock);
1248 kiblnd_check_sends(conn);
1251 static int kiblnd_resolve_addr(struct rdma_cm_id *cmid,
1252 struct sockaddr_in *srcaddr,
1253 struct sockaddr_in *dstaddr,
1256 unsigned short port;
1259 #ifdef HAVE_OFED_RDMA_SET_REUSEADDR
1260 /* allow the port to be reused */
1261 rc = rdma_set_reuseaddr(cmid, 1);
1263 CERROR("Unable to set reuse on cmid: %d\n", rc);
1268 /* look for a free privileged port */
1269 for (port = PROT_SOCK-1; port > 0; port--) {
1270 srcaddr->sin_port = htons(port);
1271 rc = rdma_resolve_addr(cmid,
1272 (struct sockaddr *)srcaddr,
1273 (struct sockaddr *)dstaddr,
1276 CDEBUG(D_NET, "bound to port %hu\n", port);
1278 } else if (rc == -EADDRINUSE || rc == -EADDRNOTAVAIL) {
1279 CDEBUG(D_NET, "bind to port %hu failed: %d\n",
1286 CERROR("Failed to bind to a free privileged port\n");
1287 #ifndef HAVE_OFED_RDMA_SET_REUSEADDR
1288 CERROR("You may need IB verbs that supports rdma_set_reuseaddr()\n");
1294 kiblnd_connect_peer (kib_peer_t *peer)
1296 struct rdma_cm_id *cmid;
1298 kib_net_t *net = peer->ibp_ni->ni_data;
1299 struct sockaddr_in srcaddr;
1300 struct sockaddr_in dstaddr;
1303 LASSERT (net != NULL);
1304 LASSERT (peer->ibp_connecting > 0);
1306 cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, peer, RDMA_PS_TCP,
1310 CERROR("Can't create CMID for %s: %ld\n",
1311 libcfs_nid2str(peer->ibp_nid), PTR_ERR(cmid));
1317 memset(&srcaddr, 0, sizeof(srcaddr));
1318 srcaddr.sin_family = AF_INET;
1319 srcaddr.sin_addr.s_addr = htonl(dev->ibd_ifip);
1321 memset(&dstaddr, 0, sizeof(dstaddr));
1322 dstaddr.sin_family = AF_INET;
1323 dstaddr.sin_port = htons(*kiblnd_tunables.kib_service);
1324 dstaddr.sin_addr.s_addr = htonl(LNET_NIDADDR(peer->ibp_nid));
1326 kiblnd_peer_addref(peer); /* cmid's ref */
1328 if (*kiblnd_tunables.kib_use_priv_port) {
1329 rc = kiblnd_resolve_addr(cmid, &srcaddr, &dstaddr,
1330 *kiblnd_tunables.kib_timeout * 1000);
1332 rc = rdma_resolve_addr(cmid,
1333 (struct sockaddr *)&srcaddr,
1334 (struct sockaddr *)&dstaddr,
1335 *kiblnd_tunables.kib_timeout * 1000);
1338 /* Can't initiate address resolution: */
1339 CERROR("Can't resolve addr for %s: %d\n",
1340 libcfs_nid2str(peer->ibp_nid), rc);
1344 LASSERT (cmid->device != NULL);
1345 CDEBUG(D_NET, "%s: connection bound to %s:%u.%u.%u.%u:%s\n",
1346 libcfs_nid2str(peer->ibp_nid), dev->ibd_ifname,
1347 HIPQUAD(dev->ibd_ifip), cmid->device->name);
1352 kiblnd_peer_decref(peer); /* cmid's ref */
1353 rdma_destroy_id(cmid);
1355 kiblnd_peer_connect_failed(peer, 1, rc);
1359 kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid)
1364 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
1365 unsigned long flags;
1368 /* If I get here, I've committed to send, so I complete the tx with
1369 * failure on any problems */
1371 LASSERT (tx == NULL || tx->tx_conn == NULL); /* only set when assigned a conn */
1372 LASSERT (tx == NULL || tx->tx_nwrq > 0); /* work items have been set up */
1374 /* First time, just use a read lock since I expect to find my peer
1376 read_lock_irqsave(g_lock, flags);
1378 peer = kiblnd_find_peer_locked(nid);
1379 if (peer != NULL && !list_empty(&peer->ibp_conns)) {
1380 /* Found a peer with an established connection */
1381 conn = kiblnd_get_conn_locked(peer);
1382 kiblnd_conn_addref(conn); /* 1 ref for me... */
1384 read_unlock_irqrestore(g_lock, flags);
1387 kiblnd_queue_tx(tx, conn);
1388 kiblnd_conn_decref(conn); /* ...to here */
1392 read_unlock(g_lock);
1393 /* Re-try with a write lock */
1396 peer = kiblnd_find_peer_locked(nid);
1398 if (list_empty(&peer->ibp_conns)) {
1399 /* found a peer, but it's still connecting... */
1400 LASSERT (peer->ibp_connecting != 0 ||
1401 peer->ibp_accepting != 0);
1403 list_add_tail(&tx->tx_list,
1404 &peer->ibp_tx_queue);
1405 write_unlock_irqrestore(g_lock, flags);
1407 conn = kiblnd_get_conn_locked(peer);
1408 kiblnd_conn_addref(conn); /* 1 ref for me... */
1410 write_unlock_irqrestore(g_lock, flags);
1413 kiblnd_queue_tx(tx, conn);
1414 kiblnd_conn_decref(conn); /* ...to here */
1419 write_unlock_irqrestore(g_lock, flags);
1421 /* Allocate a peer ready to add to the peer table and retry */
1422 rc = kiblnd_create_peer(ni, &peer, nid);
1424 CERROR("Can't create peer %s\n", libcfs_nid2str(nid));
1426 tx->tx_status = -EHOSTUNREACH;
1428 kiblnd_tx_done(ni, tx);
1433 write_lock_irqsave(g_lock, flags);
1435 peer2 = kiblnd_find_peer_locked(nid);
1436 if (peer2 != NULL) {
1437 if (list_empty(&peer2->ibp_conns)) {
1438 /* found a peer, but it's still connecting... */
1439 LASSERT (peer2->ibp_connecting != 0 ||
1440 peer2->ibp_accepting != 0);
1442 list_add_tail(&tx->tx_list,
1443 &peer2->ibp_tx_queue);
1444 write_unlock_irqrestore(g_lock, flags);
1446 conn = kiblnd_get_conn_locked(peer2);
1447 kiblnd_conn_addref(conn); /* 1 ref for me... */
1449 write_unlock_irqrestore(g_lock, flags);
1452 kiblnd_queue_tx(tx, conn);
1453 kiblnd_conn_decref(conn); /* ...to here */
1456 kiblnd_peer_decref(peer);
1460 /* Brand new peer */
1461 LASSERT (peer->ibp_connecting == 0);
1462 peer->ibp_connecting = 1;
1464 /* always called with a ref on ni, which prevents ni being shutdown */
1465 LASSERT (((kib_net_t *)ni->ni_data)->ibn_shutdown == 0);
1468 list_add_tail(&tx->tx_list, &peer->ibp_tx_queue);
1470 kiblnd_peer_addref(peer);
1471 list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid));
1473 write_unlock_irqrestore(g_lock, flags);
1475 kiblnd_connect_peer(peer);
1476 kiblnd_peer_decref(peer);
1480 kiblnd_send (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg)
1482 lnet_hdr_t *hdr = &lntmsg->msg_hdr;
1483 int type = lntmsg->msg_type;
1484 lnet_process_id_t target = lntmsg->msg_target;
1485 int target_is_router = lntmsg->msg_target_is_router;
1486 int routing = lntmsg->msg_routing;
1487 unsigned int payload_niov = lntmsg->msg_niov;
1488 struct iovec *payload_iov = lntmsg->msg_iov;
1489 lnet_kiov_t *payload_kiov = lntmsg->msg_kiov;
1490 unsigned int payload_offset = lntmsg->msg_offset;
1491 unsigned int payload_nob = lntmsg->msg_len;
1497 /* NB 'private' is different depending on what we're sending.... */
1499 CDEBUG(D_NET, "sending %d bytes in %d frags to %s\n",
1500 payload_nob, payload_niov, libcfs_id2str(target));
1502 LASSERT (payload_nob == 0 || payload_niov > 0);
1503 LASSERT (payload_niov <= LNET_MAX_IOV);
1505 /* Thread context */
1506 LASSERT (!in_interrupt());
1507 /* payload is either all vaddrs or all pages */
1508 LASSERT (!(payload_kiov != NULL && payload_iov != NULL));
1516 LASSERT (payload_nob == 0);
1520 if (routing || target_is_router)
1521 break; /* send IMMEDIATE */
1523 /* is the REPLY message too small for RDMA? */
1524 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[lntmsg->msg_md->md_length]);
1525 if (nob <= IBLND_MSG_SIZE)
1526 break; /* send IMMEDIATE */
1528 tx = kiblnd_get_idle_tx(ni, target.nid);
1530 CERROR("Can't allocate txd for GET to %s\n",
1531 libcfs_nid2str(target.nid));
1537 if ((lntmsg->msg_md->md_options & LNET_MD_KIOV) == 0)
1538 rc = kiblnd_setup_rd_iov(ni, tx,
1539 &ibmsg->ibm_u.get.ibgm_rd,
1540 lntmsg->msg_md->md_niov,
1541 lntmsg->msg_md->md_iov.iov,
1542 0, lntmsg->msg_md->md_length);
1544 rc = kiblnd_setup_rd_kiov(ni, tx,
1545 &ibmsg->ibm_u.get.ibgm_rd,
1546 lntmsg->msg_md->md_niov,
1547 lntmsg->msg_md->md_iov.kiov,
1548 0, lntmsg->msg_md->md_length);
1550 CERROR("Can't setup GET sink for %s: %d\n",
1551 libcfs_nid2str(target.nid), rc);
1552 kiblnd_tx_done(ni, tx);
1556 nob = offsetof(kib_get_msg_t, ibgm_rd.rd_frags[tx->tx_nfrags]);
1557 ibmsg->ibm_u.get.ibgm_cookie = tx->tx_cookie;
1558 ibmsg->ibm_u.get.ibgm_hdr = *hdr;
1560 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_GET_REQ, nob);
1562 tx->tx_lntmsg[1] = lnet_create_reply_msg(ni, lntmsg);
1563 if (tx->tx_lntmsg[1] == NULL) {
1564 CERROR("Can't create reply for GET -> %s\n",
1565 libcfs_nid2str(target.nid));
1566 kiblnd_tx_done(ni, tx);
1570 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg[0,1] on completion */
1571 tx->tx_waiting = 1; /* waiting for GET_DONE */
1572 kiblnd_launch_tx(ni, tx, target.nid);
1575 case LNET_MSG_REPLY:
1577 /* Is the payload small enough not to need RDMA? */
1578 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob]);
1579 if (nob <= IBLND_MSG_SIZE)
1580 break; /* send IMMEDIATE */
1582 tx = kiblnd_get_idle_tx(ni, target.nid);
1584 CERROR("Can't allocate %s txd for %s\n",
1585 type == LNET_MSG_PUT ? "PUT" : "REPLY",
1586 libcfs_nid2str(target.nid));
1590 if (payload_kiov == NULL)
1591 rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1592 payload_niov, payload_iov,
1593 payload_offset, payload_nob);
1595 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1596 payload_niov, payload_kiov,
1597 payload_offset, payload_nob);
1599 CERROR("Can't setup PUT src for %s: %d\n",
1600 libcfs_nid2str(target.nid), rc);
1601 kiblnd_tx_done(ni, tx);
1606 ibmsg->ibm_u.putreq.ibprm_hdr = *hdr;
1607 ibmsg->ibm_u.putreq.ibprm_cookie = tx->tx_cookie;
1608 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_REQ, sizeof(kib_putreq_msg_t));
1610 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1611 tx->tx_waiting = 1; /* waiting for PUT_{ACK,NAK} */
1612 kiblnd_launch_tx(ni, tx, target.nid);
1616 /* send IMMEDIATE */
1618 LASSERT (offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob])
1621 tx = kiblnd_get_idle_tx(ni, target.nid);
1623 CERROR ("Can't send %d to %s: tx descs exhausted\n",
1624 type, libcfs_nid2str(target.nid));
1629 ibmsg->ibm_u.immediate.ibim_hdr = *hdr;
1631 if (payload_kiov != NULL)
1632 lnet_copy_kiov2flat(IBLND_MSG_SIZE, ibmsg,
1633 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1634 payload_niov, payload_kiov,
1635 payload_offset, payload_nob);
1637 lnet_copy_iov2flat(IBLND_MSG_SIZE, ibmsg,
1638 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1639 payload_niov, payload_iov,
1640 payload_offset, payload_nob);
1642 nob = offsetof(kib_immediate_msg_t, ibim_payload[payload_nob]);
1643 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_IMMEDIATE, nob);
1645 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1646 kiblnd_launch_tx(ni, tx, target.nid);
1651 kiblnd_reply (lnet_ni_t *ni, kib_rx_t *rx, lnet_msg_t *lntmsg)
1653 lnet_process_id_t target = lntmsg->msg_target;
1654 unsigned int niov = lntmsg->msg_niov;
1655 struct iovec *iov = lntmsg->msg_iov;
1656 lnet_kiov_t *kiov = lntmsg->msg_kiov;
1657 unsigned int offset = lntmsg->msg_offset;
1658 unsigned int nob = lntmsg->msg_len;
1662 tx = kiblnd_get_idle_tx(ni, rx->rx_conn->ibc_peer->ibp_nid);
1664 CERROR("Can't get tx for REPLY to %s\n",
1665 libcfs_nid2str(target.nid));
1671 else if (kiov == NULL)
1672 rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1673 niov, iov, offset, nob);
1675 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1676 niov, kiov, offset, nob);
1679 CERROR("Can't setup GET src for %s: %d\n",
1680 libcfs_nid2str(target.nid), rc);
1684 rc = kiblnd_init_rdma(rx->rx_conn, tx,
1685 IBLND_MSG_GET_DONE, nob,
1686 &rx->rx_msg->ibm_u.get.ibgm_rd,
1687 rx->rx_msg->ibm_u.get.ibgm_cookie);
1689 CERROR("Can't setup rdma for GET from %s: %d\n",
1690 libcfs_nid2str(target.nid), rc);
1695 /* No RDMA: local completion may happen now! */
1696 lnet_finalize(ni, lntmsg, 0);
1698 /* RDMA: lnet_finalize(lntmsg) when it
1700 tx->tx_lntmsg[0] = lntmsg;
1703 kiblnd_queue_tx(tx, rx->rx_conn);
1707 kiblnd_tx_done(ni, tx);
1709 lnet_finalize(ni, lntmsg, -EIO);
1713 kiblnd_recv (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed,
1714 unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov,
1715 unsigned int offset, unsigned int mlen, unsigned int rlen)
1717 kib_rx_t *rx = private;
1718 kib_msg_t *rxmsg = rx->rx_msg;
1719 kib_conn_t *conn = rx->rx_conn;
1723 int post_credit = IBLND_POSTRX_PEER_CREDIT;
1726 LASSERT (mlen <= rlen);
1727 LASSERT (!in_interrupt());
1728 /* Either all pages or all vaddrs */
1729 LASSERT (!(kiov != NULL && iov != NULL));
1731 switch (rxmsg->ibm_type) {
1735 case IBLND_MSG_IMMEDIATE:
1736 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[rlen]);
1737 if (nob > rx->rx_nob) {
1738 CERROR ("Immediate message from %s too big: %d(%d)\n",
1739 libcfs_nid2str(rxmsg->ibm_u.immediate.ibim_hdr.src_nid),
1746 lnet_copy_flat2kiov(niov, kiov, offset,
1747 IBLND_MSG_SIZE, rxmsg,
1748 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1751 lnet_copy_flat2iov(niov, iov, offset,
1752 IBLND_MSG_SIZE, rxmsg,
1753 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1755 lnet_finalize (ni, lntmsg, 0);
1758 case IBLND_MSG_PUT_REQ:
1760 lnet_finalize(ni, lntmsg, 0);
1761 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK, 0,
1762 rxmsg->ibm_u.putreq.ibprm_cookie);
1766 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
1768 CERROR("Can't allocate tx for %s\n",
1769 libcfs_nid2str(conn->ibc_peer->ibp_nid));
1770 /* Not replying will break the connection */
1777 rc = kiblnd_setup_rd_iov(ni, tx,
1778 &txmsg->ibm_u.putack.ibpam_rd,
1779 niov, iov, offset, mlen);
1781 rc = kiblnd_setup_rd_kiov(ni, tx,
1782 &txmsg->ibm_u.putack.ibpam_rd,
1783 niov, kiov, offset, mlen);
1785 CERROR("Can't setup PUT sink for %s: %d\n",
1786 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
1787 kiblnd_tx_done(ni, tx);
1788 /* tell peer it's over */
1789 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK, rc,
1790 rxmsg->ibm_u.putreq.ibprm_cookie);
1794 nob = offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[tx->tx_nfrags]);
1795 txmsg->ibm_u.putack.ibpam_src_cookie = rxmsg->ibm_u.putreq.ibprm_cookie;
1796 txmsg->ibm_u.putack.ibpam_dst_cookie = tx->tx_cookie;
1798 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_ACK, nob);
1800 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1801 tx->tx_waiting = 1; /* waiting for PUT_DONE */
1802 kiblnd_queue_tx(tx, conn);
1804 /* reposted buffer reserved for PUT_DONE */
1805 post_credit = IBLND_POSTRX_NO_CREDIT;
1808 case IBLND_MSG_GET_REQ:
1809 if (lntmsg != NULL) {
1810 /* Optimized GET; RDMA lntmsg's payload */
1811 kiblnd_reply(ni, rx, lntmsg);
1813 /* GET didn't match anything */
1814 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_GET_DONE,
1816 rxmsg->ibm_u.get.ibgm_cookie);
1821 kiblnd_post_rx(rx, post_credit);
1826 kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name)
1828 struct task_struct *task = kthread_run(fn, arg, name);
1831 return PTR_ERR(task);
1833 atomic_inc(&kiblnd_data.kib_nthreads);
1838 kiblnd_thread_fini (void)
1840 atomic_dec (&kiblnd_data.kib_nthreads);
1844 kiblnd_peer_alive (kib_peer_t *peer)
1846 /* This is racy, but everyone's only writing cfs_time_current() */
1847 peer->ibp_last_alive = cfs_time_current();
1852 kiblnd_peer_notify (kib_peer_t *peer)
1855 cfs_time_t last_alive = 0;
1856 unsigned long flags;
1858 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1860 if (list_empty(&peer->ibp_conns) &&
1861 peer->ibp_accepting == 0 &&
1862 peer->ibp_connecting == 0 &&
1863 peer->ibp_error != 0) {
1864 error = peer->ibp_error;
1865 peer->ibp_error = 0;
1867 last_alive = peer->ibp_last_alive;
1870 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1873 lnet_notify(peer->ibp_ni,
1874 peer->ibp_nid, 0, last_alive);
1878 kiblnd_close_conn_locked (kib_conn_t *conn, int error)
1880 /* This just does the immediate housekeeping. 'error' is zero for a
1881 * normal shutdown which can happen only after the connection has been
1882 * established. If the connection is established, schedule the
1883 * connection to be finished off by the connd. Otherwise the connd is
1884 * already dealing with it (either to set it up or tear it down).
1885 * Caller holds kib_global_lock exclusively in irq context */
1886 kib_peer_t *peer = conn->ibc_peer;
1888 unsigned long flags;
1890 LASSERT (error != 0 || conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1892 if (error != 0 && conn->ibc_comms_error == 0)
1893 conn->ibc_comms_error = error;
1895 if (conn->ibc_state != IBLND_CONN_ESTABLISHED)
1896 return; /* already being handled */
1899 list_empty(&conn->ibc_tx_noops) &&
1900 list_empty(&conn->ibc_tx_queue) &&
1901 list_empty(&conn->ibc_tx_queue_rsrvd) &&
1902 list_empty(&conn->ibc_tx_queue_nocred) &&
1903 list_empty(&conn->ibc_active_txs)) {
1904 CDEBUG(D_NET, "closing conn to %s\n",
1905 libcfs_nid2str(peer->ibp_nid));
1907 CNETERR("Closing conn to %s: error %d%s%s%s%s%s\n",
1908 libcfs_nid2str(peer->ibp_nid), error,
1909 list_empty(&conn->ibc_tx_queue) ? "" : "(sending)",
1910 list_empty(&conn->ibc_tx_noops) ? "" : "(sending_noops)",
1911 list_empty(&conn->ibc_tx_queue_rsrvd) ?
1912 "" : "(sending_rsrvd)",
1913 list_empty(&conn->ibc_tx_queue_nocred) ?
1914 "" : "(sending_nocred)",
1915 list_empty(&conn->ibc_active_txs) ? "" : "(waiting)");
1918 dev = ((kib_net_t *)peer->ibp_ni->ni_data)->ibn_dev;
1919 list_del(&conn->ibc_list);
1920 /* connd (see below) takes over ibc_list's ref */
1922 if (list_empty(&peer->ibp_conns) && /* no more conns */
1923 kiblnd_peer_active(peer)) { /* still in peer table */
1924 kiblnd_unlink_peer_locked(peer);
1926 /* set/clear error on last conn */
1927 peer->ibp_error = conn->ibc_comms_error;
1930 kiblnd_set_conn_state(conn, IBLND_CONN_CLOSING);
1933 kiblnd_dev_can_failover(dev)) {
1934 list_add_tail(&dev->ibd_fail_list,
1935 &kiblnd_data.kib_failed_devs);
1936 wake_up(&kiblnd_data.kib_failover_waitq);
1939 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
1941 list_add_tail(&conn->ibc_list, &kiblnd_data.kib_connd_conns);
1942 wake_up(&kiblnd_data.kib_connd_waitq);
1944 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
1948 kiblnd_close_conn(kib_conn_t *conn, int error)
1950 unsigned long flags;
1952 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1954 kiblnd_close_conn_locked(conn, error);
1956 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1960 kiblnd_handle_early_rxs(kib_conn_t *conn)
1962 unsigned long flags;
1965 LASSERT(!in_interrupt());
1966 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1968 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1969 while (!list_empty(&conn->ibc_early_rxs)) {
1970 rx = list_entry(conn->ibc_early_rxs.next,
1972 list_del(&rx->rx_list);
1973 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1975 kiblnd_handle_rx(rx);
1977 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1979 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1983 kiblnd_abort_txs(kib_conn_t *conn, struct list_head *txs)
1985 struct list_head zombies = LIST_HEAD_INIT(zombies);
1986 struct list_head *tmp;
1987 struct list_head *nxt;
1990 spin_lock(&conn->ibc_lock);
1992 list_for_each_safe(tmp, nxt, txs) {
1993 tx = list_entry(tmp, kib_tx_t, tx_list);
1995 if (txs == &conn->ibc_active_txs) {
1996 LASSERT(!tx->tx_queued);
1997 LASSERT(tx->tx_waiting ||
1998 tx->tx_sending != 0);
2000 LASSERT(tx->tx_queued);
2003 tx->tx_status = -ECONNABORTED;
2006 if (tx->tx_sending == 0) {
2008 list_del(&tx->tx_list);
2009 list_add(&tx->tx_list, &zombies);
2013 spin_unlock(&conn->ibc_lock);
2015 kiblnd_txlist_done(conn->ibc_peer->ibp_ni, &zombies, -ECONNABORTED);
2019 kiblnd_finalise_conn (kib_conn_t *conn)
2021 LASSERT (!in_interrupt());
2022 LASSERT (conn->ibc_state > IBLND_CONN_INIT);
2024 kiblnd_set_conn_state(conn, IBLND_CONN_DISCONNECTED);
2026 /* abort_receives moves QP state to IB_QPS_ERR. This is only required
2027 * for connections that didn't get as far as being connected, because
2028 * rdma_disconnect() does this for free. */
2029 kiblnd_abort_receives(conn);
2031 /* Complete all tx descs not waiting for sends to complete.
2032 * NB we should be safe from RDMA now that the QP has changed state */
2034 kiblnd_abort_txs(conn, &conn->ibc_tx_noops);
2035 kiblnd_abort_txs(conn, &conn->ibc_tx_queue);
2036 kiblnd_abort_txs(conn, &conn->ibc_tx_queue_rsrvd);
2037 kiblnd_abort_txs(conn, &conn->ibc_tx_queue_nocred);
2038 kiblnd_abort_txs(conn, &conn->ibc_active_txs);
2040 kiblnd_handle_early_rxs(conn);
2044 kiblnd_peer_connect_failed(kib_peer_t *peer, int active, int error)
2046 struct list_head zombies = LIST_HEAD_INIT(zombies);
2047 unsigned long flags;
2049 LASSERT (error != 0);
2050 LASSERT (!in_interrupt());
2052 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2055 LASSERT (peer->ibp_connecting > 0);
2056 peer->ibp_connecting--;
2058 LASSERT (peer->ibp_accepting > 0);
2059 peer->ibp_accepting--;
2062 if (peer->ibp_connecting != 0 ||
2063 peer->ibp_accepting != 0) {
2064 /* another connection attempt under way... */
2065 write_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2070 if (list_empty(&peer->ibp_conns)) {
2071 /* Take peer's blocked transmits to complete with error */
2072 list_add(&zombies, &peer->ibp_tx_queue);
2073 list_del_init(&peer->ibp_tx_queue);
2075 if (kiblnd_peer_active(peer))
2076 kiblnd_unlink_peer_locked(peer);
2078 peer->ibp_error = error;
2080 /* Can't have blocked transmits if there are connections */
2081 LASSERT(list_empty(&peer->ibp_tx_queue));
2084 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2086 kiblnd_peer_notify(peer);
2088 if (list_empty(&zombies))
2091 CNETERR("Deleting messages for %s: connection failed\n",
2092 libcfs_nid2str(peer->ibp_nid));
2094 kiblnd_txlist_done(peer->ibp_ni, &zombies, -EHOSTUNREACH);
2098 kiblnd_connreq_done(kib_conn_t *conn, int status)
2100 kib_peer_t *peer = conn->ibc_peer;
2102 struct list_head txs;
2103 unsigned long flags;
2106 active = (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2108 CDEBUG(D_NET,"%s: active(%d), version(%x), status(%d)\n",
2109 libcfs_nid2str(peer->ibp_nid), active,
2110 conn->ibc_version, status);
2112 LASSERT (!in_interrupt());
2113 LASSERT ((conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT &&
2114 peer->ibp_connecting > 0) ||
2115 (conn->ibc_state == IBLND_CONN_PASSIVE_WAIT &&
2116 peer->ibp_accepting > 0));
2118 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
2119 conn->ibc_connvars = NULL;
2122 /* failed to establish connection */
2123 kiblnd_peer_connect_failed(peer, active, status);
2124 kiblnd_finalise_conn(conn);
2128 /* connection established */
2129 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2131 conn->ibc_last_send = jiffies;
2132 kiblnd_set_conn_state(conn, IBLND_CONN_ESTABLISHED);
2133 kiblnd_peer_alive(peer);
2135 /* Add conn to peer's list and nuke any dangling conns from a different
2136 * peer instance... */
2137 kiblnd_conn_addref(conn); /* +1 ref for ibc_list */
2138 list_add(&conn->ibc_list, &peer->ibp_conns);
2140 peer->ibp_connecting--;
2142 peer->ibp_accepting--;
2144 if (peer->ibp_version == 0) {
2145 peer->ibp_version = conn->ibc_version;
2146 peer->ibp_incarnation = conn->ibc_incarnation;
2149 if (peer->ibp_version != conn->ibc_version ||
2150 peer->ibp_incarnation != conn->ibc_incarnation) {
2151 kiblnd_close_stale_conns_locked(peer, conn->ibc_version,
2152 conn->ibc_incarnation);
2153 peer->ibp_version = conn->ibc_version;
2154 peer->ibp_incarnation = conn->ibc_incarnation;
2157 /* grab pending txs while I have the lock */
2158 list_add(&txs, &peer->ibp_tx_queue);
2159 list_del_init(&peer->ibp_tx_queue);
2161 if (!kiblnd_peer_active(peer) || /* peer has been deleted */
2162 conn->ibc_comms_error != 0) { /* error has happened already */
2163 lnet_ni_t *ni = peer->ibp_ni;
2165 /* start to shut down connection */
2166 kiblnd_close_conn_locked(conn, -ECONNABORTED);
2167 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2169 kiblnd_txlist_done(ni, &txs, -ECONNABORTED);
2174 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2176 /* Schedule blocked txs */
2177 spin_lock(&conn->ibc_lock);
2178 while (!list_empty(&txs)) {
2179 tx = list_entry(txs.next, kib_tx_t, tx_list);
2180 list_del(&tx->tx_list);
2182 kiblnd_queue_tx_locked(tx, conn);
2184 spin_unlock(&conn->ibc_lock);
2186 kiblnd_check_sends(conn);
2188 /* schedule blocked rxs */
2189 kiblnd_handle_early_rxs(conn);
2193 kiblnd_reject(struct rdma_cm_id *cmid, kib_rej_t *rej)
2197 rc = rdma_reject(cmid, rej, sizeof(*rej));
2200 CWARN("Error %d sending reject\n", rc);
2204 kiblnd_passive_connect (struct rdma_cm_id *cmid, void *priv, int priv_nob)
2206 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2207 kib_msg_t *reqmsg = priv;
2213 lnet_ni_t *ni = NULL;
2214 kib_net_t *net = NULL;
2216 struct rdma_conn_param cp;
2218 int version = IBLND_MSG_VERSION;
2219 unsigned long flags;
2221 struct sockaddr_in *peer_addr;
2222 LASSERT (!in_interrupt());
2224 /* cmid inherits 'context' from the corresponding listener id */
2225 ibdev = (kib_dev_t *)cmid->context;
2226 LASSERT (ibdev != NULL);
2228 memset(&rej, 0, sizeof(rej));
2229 rej.ibr_magic = IBLND_MSG_MAGIC;
2230 rej.ibr_why = IBLND_REJECT_FATAL;
2231 rej.ibr_cp.ibcp_max_msg_size = IBLND_MSG_SIZE;
2233 peer_addr = (struct sockaddr_in *)&(cmid->route.addr.dst_addr);
2234 if (*kiblnd_tunables.kib_require_priv_port &&
2235 ntohs(peer_addr->sin_port) >= PROT_SOCK) {
2236 __u32 ip = ntohl(peer_addr->sin_addr.s_addr);
2237 CERROR("Peer's port (%u.%u.%u.%u:%hu) is not privileged\n",
2238 HIPQUAD(ip), ntohs(peer_addr->sin_port));
2242 if (priv_nob < offsetof(kib_msg_t, ibm_type)) {
2243 CERROR("Short connection request\n");
2247 /* Future protocol version compatibility support! If the
2248 * o2iblnd-specific protocol changes, or when LNET unifies
2249 * protocols over all LNDs, the initial connection will
2250 * negotiate a protocol version. I trap this here to avoid
2251 * console errors; the reject tells the peer which protocol I
2253 if (reqmsg->ibm_magic == LNET_PROTO_MAGIC ||
2254 reqmsg->ibm_magic == __swab32(LNET_PROTO_MAGIC))
2256 if (reqmsg->ibm_magic == IBLND_MSG_MAGIC &&
2257 reqmsg->ibm_version != IBLND_MSG_VERSION &&
2258 reqmsg->ibm_version != IBLND_MSG_VERSION_1)
2260 if (reqmsg->ibm_magic == __swab32(IBLND_MSG_MAGIC) &&
2261 reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION) &&
2262 reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION_1))
2265 rc = kiblnd_unpack_msg(reqmsg, priv_nob);
2267 CERROR("Can't parse connection request: %d\n", rc);
2271 nid = reqmsg->ibm_srcnid;
2272 ni = lnet_net2ni(LNET_NIDNET(reqmsg->ibm_dstnid));
2275 net = (kib_net_t *)ni->ni_data;
2276 rej.ibr_incarnation = net->ibn_incarnation;
2279 if (ni == NULL || /* no matching net */
2280 ni->ni_nid != reqmsg->ibm_dstnid || /* right NET, wrong NID! */
2281 net->ibn_dev != ibdev) { /* wrong device */
2282 CERROR("Can't accept %s on %s (%s:%d:%u.%u.%u.%u): "
2283 "bad dst nid %s\n", libcfs_nid2str(nid),
2284 ni == NULL ? "NA" : libcfs_nid2str(ni->ni_nid),
2285 ibdev->ibd_ifname, ibdev->ibd_nnets,
2286 HIPQUAD(ibdev->ibd_ifip),
2287 libcfs_nid2str(reqmsg->ibm_dstnid));
2292 /* check time stamp as soon as possible */
2293 if (reqmsg->ibm_dststamp != 0 &&
2294 reqmsg->ibm_dststamp != net->ibn_incarnation) {
2295 CWARN("Stale connection request\n");
2296 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2300 /* I can accept peer's version */
2301 version = reqmsg->ibm_version;
2303 if (reqmsg->ibm_type != IBLND_MSG_CONNREQ) {
2304 CERROR("Unexpected connreq msg type: %x from %s\n",
2305 reqmsg->ibm_type, libcfs_nid2str(nid));
2309 if (reqmsg->ibm_u.connparams.ibcp_queue_depth !=
2310 IBLND_MSG_QUEUE_SIZE(version)) {
2311 CERROR("Can't accept %s: incompatible queue depth %d (%d wanted)\n",
2312 libcfs_nid2str(nid), reqmsg->ibm_u.connparams.ibcp_queue_depth,
2313 IBLND_MSG_QUEUE_SIZE(version));
2315 if (version == IBLND_MSG_VERSION)
2316 rej.ibr_why = IBLND_REJECT_MSG_QUEUE_SIZE;
2321 if (reqmsg->ibm_u.connparams.ibcp_max_frags !=
2322 IBLND_RDMA_FRAGS(version)) {
2323 CERROR("Can't accept %s(version %x): "
2324 "incompatible max_frags %d (%d wanted)\n",
2325 libcfs_nid2str(nid), version,
2326 reqmsg->ibm_u.connparams.ibcp_max_frags,
2327 IBLND_RDMA_FRAGS(version));
2329 if (version == IBLND_MSG_VERSION)
2330 rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2336 if (reqmsg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2337 CERROR("Can't accept %s: message size %d too big (%d max)\n",
2338 libcfs_nid2str(nid),
2339 reqmsg->ibm_u.connparams.ibcp_max_msg_size,
2344 /* assume 'nid' is a new peer; create */
2345 rc = kiblnd_create_peer(ni, &peer, nid);
2347 CERROR("Can't create peer for %s\n", libcfs_nid2str(nid));
2348 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2352 write_lock_irqsave(g_lock, flags);
2354 peer2 = kiblnd_find_peer_locked(nid);
2355 if (peer2 != NULL) {
2356 if (peer2->ibp_version == 0) {
2357 peer2->ibp_version = version;
2358 peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2361 /* not the guy I've talked with */
2362 if (peer2->ibp_incarnation != reqmsg->ibm_srcstamp ||
2363 peer2->ibp_version != version) {
2364 kiblnd_close_peer_conns_locked(peer2, -ESTALE);
2365 write_unlock_irqrestore(g_lock, flags);
2367 CWARN("Conn stale %s [old ver: %x, new ver: %x]\n",
2368 libcfs_nid2str(nid), peer2->ibp_version, version);
2370 kiblnd_peer_decref(peer);
2371 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2375 /* tie-break connection race in favour of the higher NID */
2376 if (peer2->ibp_connecting != 0 &&
2378 write_unlock_irqrestore(g_lock, flags);
2380 CWARN("Conn race %s\n", libcfs_nid2str(peer2->ibp_nid));
2382 kiblnd_peer_decref(peer);
2383 rej.ibr_why = IBLND_REJECT_CONN_RACE;
2387 peer2->ibp_accepting++;
2388 kiblnd_peer_addref(peer2);
2390 write_unlock_irqrestore(g_lock, flags);
2391 kiblnd_peer_decref(peer);
2394 /* Brand new peer */
2395 LASSERT (peer->ibp_accepting == 0);
2396 LASSERT (peer->ibp_version == 0 &&
2397 peer->ibp_incarnation == 0);
2399 peer->ibp_accepting = 1;
2400 peer->ibp_version = version;
2401 peer->ibp_incarnation = reqmsg->ibm_srcstamp;
2403 /* I have a ref on ni that prevents it being shutdown */
2404 LASSERT (net->ibn_shutdown == 0);
2406 kiblnd_peer_addref(peer);
2407 list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid));
2409 write_unlock_irqrestore(g_lock, flags);
2412 conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_PASSIVE_WAIT, version);
2414 kiblnd_peer_connect_failed(peer, 0, -ENOMEM);
2415 kiblnd_peer_decref(peer);
2416 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2420 /* conn now "owns" cmid, so I return success from here on to ensure the
2421 * CM callback doesn't destroy cmid. */
2423 conn->ibc_incarnation = reqmsg->ibm_srcstamp;
2424 conn->ibc_credits = IBLND_MSG_QUEUE_SIZE(version);
2425 conn->ibc_reserved_credits = IBLND_MSG_QUEUE_SIZE(version);
2426 LASSERT (conn->ibc_credits + conn->ibc_reserved_credits + IBLND_OOB_MSGS(version)
2427 <= IBLND_RX_MSGS(version));
2429 ackmsg = &conn->ibc_connvars->cv_msg;
2430 memset(ackmsg, 0, sizeof(*ackmsg));
2432 kiblnd_init_msg(ackmsg, IBLND_MSG_CONNACK,
2433 sizeof(ackmsg->ibm_u.connparams));
2434 ackmsg->ibm_u.connparams.ibcp_queue_depth = IBLND_MSG_QUEUE_SIZE(version);
2435 ackmsg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2436 ackmsg->ibm_u.connparams.ibcp_max_frags = IBLND_RDMA_FRAGS(version);
2438 kiblnd_pack_msg(ni, ackmsg, version, 0, nid, reqmsg->ibm_srcstamp);
2440 memset(&cp, 0, sizeof(cp));
2441 cp.private_data = ackmsg;
2442 cp.private_data_len = ackmsg->ibm_nob;
2443 cp.responder_resources = 0; /* No atomic ops or RDMA reads */
2444 cp.initiator_depth = 0;
2445 cp.flow_control = 1;
2446 cp.retry_count = *kiblnd_tunables.kib_retry_count;
2447 cp.rnr_retry_count = *kiblnd_tunables.kib_rnr_retry_count;
2449 CDEBUG(D_NET, "Accept %s\n", libcfs_nid2str(nid));
2451 rc = rdma_accept(cmid, &cp);
2453 CERROR("Can't accept %s: %d\n", libcfs_nid2str(nid), rc);
2454 rej.ibr_version = version;
2455 rej.ibr_why = IBLND_REJECT_FATAL;
2457 kiblnd_reject(cmid, &rej);
2458 kiblnd_connreq_done(conn, rc);
2459 kiblnd_conn_decref(conn);
2469 rej.ibr_version = version;
2470 rej.ibr_cp.ibcp_queue_depth = IBLND_MSG_QUEUE_SIZE(version);
2471 rej.ibr_cp.ibcp_max_frags = IBLND_RDMA_FRAGS(version);
2472 kiblnd_reject(cmid, &rej);
2474 return -ECONNREFUSED;
2478 kiblnd_reconnect (kib_conn_t *conn, int version,
2479 __u64 incarnation, int why, kib_connparams_t *cp)
2481 kib_peer_t *peer = conn->ibc_peer;
2484 unsigned long flags;
2486 LASSERT (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2487 LASSERT (peer->ibp_connecting > 0); /* 'conn' at least */
2489 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2491 /* retry connection if it's still needed and no other connection
2492 * attempts (active or passive) are in progress
2493 * NB: reconnect is still needed even when ibp_tx_queue is
2494 * empty if ibp_version != version because reconnect may be
2495 * initiated by kiblnd_query() */
2496 if ((!list_empty(&peer->ibp_tx_queue) ||
2497 peer->ibp_version != version) &&
2498 peer->ibp_connecting == 1 &&
2499 peer->ibp_accepting == 0) {
2501 peer->ibp_connecting++;
2503 peer->ibp_version = version;
2504 peer->ibp_incarnation = incarnation;
2507 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2517 case IBLND_REJECT_CONN_STALE:
2521 case IBLND_REJECT_CONN_RACE:
2522 reason = "conn race";
2525 case IBLND_REJECT_CONN_UNCOMPAT:
2526 reason = "version negotiation";
2530 CNETERR("%s: retrying (%s), %x, %x, "
2531 "queue_dep: %d, max_frag: %d, msg_size: %d\n",
2532 libcfs_nid2str(peer->ibp_nid),
2533 reason, IBLND_MSG_VERSION, version,
2534 cp != NULL? cp->ibcp_queue_depth :IBLND_MSG_QUEUE_SIZE(version),
2535 cp != NULL? cp->ibcp_max_frags : IBLND_RDMA_FRAGS(version),
2536 cp != NULL? cp->ibcp_max_msg_size: IBLND_MSG_SIZE);
2538 kiblnd_connect_peer(peer);
2542 kiblnd_rejected (kib_conn_t *conn, int reason, void *priv, int priv_nob)
2544 kib_peer_t *peer = conn->ibc_peer;
2546 LASSERT (!in_interrupt());
2547 LASSERT (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2550 case IB_CM_REJ_STALE_CONN:
2551 kiblnd_reconnect(conn, IBLND_MSG_VERSION, 0,
2552 IBLND_REJECT_CONN_STALE, NULL);
2555 case IB_CM_REJ_INVALID_SERVICE_ID:
2556 CNETERR("%s rejected: no listener at %d\n",
2557 libcfs_nid2str(peer->ibp_nid),
2558 *kiblnd_tunables.kib_service);
2561 case IB_CM_REJ_CONSUMER_DEFINED:
2562 if (priv_nob >= offsetof(kib_rej_t, ibr_padding)) {
2563 kib_rej_t *rej = priv;
2564 kib_connparams_t *cp = NULL;
2566 __u64 incarnation = -1;
2568 /* NB. default incarnation is -1 because:
2569 * a) V1 will ignore dst incarnation in connreq.
2570 * b) V2 will provide incarnation while rejecting me,
2571 * -1 will be overwrote.
2573 * if I try to connect to a V1 peer with V2 protocol,
2574 * it rejected me then upgrade to V2, I have no idea
2575 * about the upgrading and try to reconnect with V1,
2576 * in this case upgraded V2 can find out I'm trying to
2577 * talk to the old guy and reject me(incarnation is -1).
2580 if (rej->ibr_magic == __swab32(IBLND_MSG_MAGIC) ||
2581 rej->ibr_magic == __swab32(LNET_PROTO_MAGIC)) {
2582 __swab32s(&rej->ibr_magic);
2583 __swab16s(&rej->ibr_version);
2587 if (priv_nob >= sizeof(kib_rej_t) &&
2588 rej->ibr_version > IBLND_MSG_VERSION_1) {
2589 /* priv_nob is always 148 in current version
2590 * of OFED, so we still need to check version.
2591 * (define of IB_CM_REJ_PRIVATE_DATA_SIZE) */
2595 __swab64s(&rej->ibr_incarnation);
2596 __swab16s(&cp->ibcp_queue_depth);
2597 __swab16s(&cp->ibcp_max_frags);
2598 __swab32s(&cp->ibcp_max_msg_size);
2601 incarnation = rej->ibr_incarnation;
2604 if (rej->ibr_magic != IBLND_MSG_MAGIC &&
2605 rej->ibr_magic != LNET_PROTO_MAGIC) {
2606 CERROR("%s rejected: consumer defined fatal error\n",
2607 libcfs_nid2str(peer->ibp_nid));
2611 if (rej->ibr_version != IBLND_MSG_VERSION &&
2612 rej->ibr_version != IBLND_MSG_VERSION_1) {
2613 CERROR("%s rejected: o2iblnd version %x error\n",
2614 libcfs_nid2str(peer->ibp_nid),
2619 if (rej->ibr_why == IBLND_REJECT_FATAL &&
2620 rej->ibr_version == IBLND_MSG_VERSION_1) {
2621 CDEBUG(D_NET, "rejected by old version peer %s: %x\n",
2622 libcfs_nid2str(peer->ibp_nid), rej->ibr_version);
2624 if (conn->ibc_version != IBLND_MSG_VERSION_1)
2625 rej->ibr_why = IBLND_REJECT_CONN_UNCOMPAT;
2628 switch (rej->ibr_why) {
2629 case IBLND_REJECT_CONN_RACE:
2630 case IBLND_REJECT_CONN_STALE:
2631 case IBLND_REJECT_CONN_UNCOMPAT:
2632 kiblnd_reconnect(conn, rej->ibr_version,
2633 incarnation, rej->ibr_why, cp);
2636 case IBLND_REJECT_MSG_QUEUE_SIZE:
2637 CERROR("%s rejected: incompatible message queue depth %d, %d\n",
2638 libcfs_nid2str(peer->ibp_nid),
2639 cp != NULL ? cp->ibcp_queue_depth :
2640 IBLND_MSG_QUEUE_SIZE(rej->ibr_version),
2641 IBLND_MSG_QUEUE_SIZE(conn->ibc_version));
2644 case IBLND_REJECT_RDMA_FRAGS:
2645 CERROR("%s rejected: incompatible # of RDMA fragments %d, %d\n",
2646 libcfs_nid2str(peer->ibp_nid),
2647 cp != NULL ? cp->ibcp_max_frags :
2648 IBLND_RDMA_FRAGS(rej->ibr_version),
2649 IBLND_RDMA_FRAGS(conn->ibc_version));
2652 case IBLND_REJECT_NO_RESOURCES:
2653 CERROR("%s rejected: o2iblnd no resources\n",
2654 libcfs_nid2str(peer->ibp_nid));
2657 case IBLND_REJECT_FATAL:
2658 CERROR("%s rejected: o2iblnd fatal error\n",
2659 libcfs_nid2str(peer->ibp_nid));
2663 CERROR("%s rejected: o2iblnd reason %d\n",
2664 libcfs_nid2str(peer->ibp_nid),
2672 CNETERR("%s rejected: reason %d, size %d\n",
2673 libcfs_nid2str(peer->ibp_nid), reason, priv_nob);
2677 kiblnd_connreq_done(conn, -ECONNREFUSED);
2681 kiblnd_check_connreply (kib_conn_t *conn, void *priv, int priv_nob)
2683 kib_peer_t *peer = conn->ibc_peer;
2684 lnet_ni_t *ni = peer->ibp_ni;
2685 kib_net_t *net = ni->ni_data;
2686 kib_msg_t *msg = priv;
2687 int ver = conn->ibc_version;
2688 int rc = kiblnd_unpack_msg(msg, priv_nob);
2689 unsigned long flags;
2691 LASSERT (net != NULL);
2694 CERROR("Can't unpack connack from %s: %d\n",
2695 libcfs_nid2str(peer->ibp_nid), rc);
2699 if (msg->ibm_type != IBLND_MSG_CONNACK) {
2700 CERROR("Unexpected message %d from %s\n",
2701 msg->ibm_type, libcfs_nid2str(peer->ibp_nid));
2706 if (ver != msg->ibm_version) {
2707 CERROR("%s replied version %x is different with "
2708 "requested version %x\n",
2709 libcfs_nid2str(peer->ibp_nid), msg->ibm_version, ver);
2714 if (msg->ibm_u.connparams.ibcp_queue_depth !=
2715 IBLND_MSG_QUEUE_SIZE(ver)) {
2716 CERROR("%s has incompatible queue depth %d(%d wanted)\n",
2717 libcfs_nid2str(peer->ibp_nid),
2718 msg->ibm_u.connparams.ibcp_queue_depth,
2719 IBLND_MSG_QUEUE_SIZE(ver));
2724 if (msg->ibm_u.connparams.ibcp_max_frags !=
2725 IBLND_RDMA_FRAGS(ver)) {
2726 CERROR("%s has incompatible max_frags %d (%d wanted)\n",
2727 libcfs_nid2str(peer->ibp_nid),
2728 msg->ibm_u.connparams.ibcp_max_frags,
2729 IBLND_RDMA_FRAGS(ver));
2734 if (msg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2735 CERROR("%s max message size %d too big (%d max)\n",
2736 libcfs_nid2str(peer->ibp_nid),
2737 msg->ibm_u.connparams.ibcp_max_msg_size,
2743 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2744 if (msg->ibm_dstnid == ni->ni_nid &&
2745 msg->ibm_dststamp == net->ibn_incarnation)
2749 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2752 CERROR("Bad connection reply from %s, rc = %d, "
2753 "version: %x max_frags: %d\n",
2754 libcfs_nid2str(peer->ibp_nid), rc,
2755 msg->ibm_version, msg->ibm_u.connparams.ibcp_max_frags);
2759 conn->ibc_incarnation = msg->ibm_srcstamp;
2761 conn->ibc_reserved_credits = IBLND_MSG_QUEUE_SIZE(ver);
2762 LASSERT (conn->ibc_credits + conn->ibc_reserved_credits + IBLND_OOB_MSGS(ver)
2763 <= IBLND_RX_MSGS(ver));
2765 kiblnd_connreq_done(conn, 0);
2769 /* NB My QP has already established itself, so I handle anything going
2770 * wrong here by setting ibc_comms_error.
2771 * kiblnd_connreq_done(0) moves the conn state to ESTABLISHED, but then
2772 * immediately tears it down. */
2775 conn->ibc_comms_error = rc;
2776 kiblnd_connreq_done(conn, 0);
2780 kiblnd_active_connect (struct rdma_cm_id *cmid)
2782 kib_peer_t *peer = (kib_peer_t *)cmid->context;
2785 struct rdma_conn_param cp;
2788 unsigned long flags;
2791 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2793 incarnation = peer->ibp_incarnation;
2794 version = (peer->ibp_version == 0) ? IBLND_MSG_VERSION :
2797 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2799 conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_ACTIVE_CONNECT, version);
2801 kiblnd_peer_connect_failed(peer, 1, -ENOMEM);
2802 kiblnd_peer_decref(peer); /* lose cmid's ref */
2806 /* conn "owns" cmid now, so I return success from here on to ensure the
2807 * CM callback doesn't destroy cmid. conn also takes over cmid's ref
2810 msg = &conn->ibc_connvars->cv_msg;
2812 memset(msg, 0, sizeof(*msg));
2813 kiblnd_init_msg(msg, IBLND_MSG_CONNREQ, sizeof(msg->ibm_u.connparams));
2814 msg->ibm_u.connparams.ibcp_queue_depth = IBLND_MSG_QUEUE_SIZE(version);
2815 msg->ibm_u.connparams.ibcp_max_frags = IBLND_RDMA_FRAGS(version);
2816 msg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2818 kiblnd_pack_msg(peer->ibp_ni, msg, version,
2819 0, peer->ibp_nid, incarnation);
2821 memset(&cp, 0, sizeof(cp));
2822 cp.private_data = msg;
2823 cp.private_data_len = msg->ibm_nob;
2824 cp.responder_resources = 0; /* No atomic ops or RDMA reads */
2825 cp.initiator_depth = 0;
2826 cp.flow_control = 1;
2827 cp.retry_count = *kiblnd_tunables.kib_retry_count;
2828 cp.rnr_retry_count = *kiblnd_tunables.kib_rnr_retry_count;
2830 LASSERT(cmid->context == (void *)conn);
2831 LASSERT(conn->ibc_cmid == cmid);
2833 rc = rdma_connect(cmid, &cp);
2835 CERROR("Can't connect to %s: %d\n",
2836 libcfs_nid2str(peer->ibp_nid), rc);
2837 kiblnd_connreq_done(conn, rc);
2838 kiblnd_conn_decref(conn);
2845 kiblnd_cm_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2851 switch (event->event) {
2853 CERROR("Unexpected event: %d, status: %d\n",
2854 event->event, event->status);
2857 case RDMA_CM_EVENT_CONNECT_REQUEST:
2858 /* destroy cmid on failure */
2859 rc = kiblnd_passive_connect(cmid,
2860 (void *)KIBLND_CONN_PARAM(event),
2861 KIBLND_CONN_PARAM_LEN(event));
2862 CDEBUG(D_NET, "connreq: %d\n", rc);
2865 case RDMA_CM_EVENT_ADDR_ERROR:
2866 peer = (kib_peer_t *)cmid->context;
2867 CNETERR("%s: ADDR ERROR %d\n",
2868 libcfs_nid2str(peer->ibp_nid), event->status);
2869 kiblnd_peer_connect_failed(peer, 1, -EHOSTUNREACH);
2870 kiblnd_peer_decref(peer);
2871 return -EHOSTUNREACH; /* rc != 0 destroys cmid */
2873 case RDMA_CM_EVENT_ADDR_RESOLVED:
2874 peer = (kib_peer_t *)cmid->context;
2876 CDEBUG(D_NET,"%s Addr resolved: %d\n",
2877 libcfs_nid2str(peer->ibp_nid), event->status);
2879 if (event->status != 0) {
2880 CNETERR("Can't resolve address for %s: %d\n",
2881 libcfs_nid2str(peer->ibp_nid), event->status);
2884 rc = rdma_resolve_route(
2885 cmid, *kiblnd_tunables.kib_timeout * 1000);
2888 /* Can't initiate route resolution */
2889 CERROR("Can't resolve route for %s: %d\n",
2890 libcfs_nid2str(peer->ibp_nid), rc);
2892 kiblnd_peer_connect_failed(peer, 1, rc);
2893 kiblnd_peer_decref(peer);
2894 return rc; /* rc != 0 destroys cmid */
2896 case RDMA_CM_EVENT_ROUTE_ERROR:
2897 peer = (kib_peer_t *)cmid->context;
2898 CNETERR("%s: ROUTE ERROR %d\n",
2899 libcfs_nid2str(peer->ibp_nid), event->status);
2900 kiblnd_peer_connect_failed(peer, 1, -EHOSTUNREACH);
2901 kiblnd_peer_decref(peer);
2902 return -EHOSTUNREACH; /* rc != 0 destroys cmid */
2904 case RDMA_CM_EVENT_ROUTE_RESOLVED:
2905 peer = (kib_peer_t *)cmid->context;
2906 CDEBUG(D_NET,"%s Route resolved: %d\n",
2907 libcfs_nid2str(peer->ibp_nid), event->status);
2909 if (event->status == 0)
2910 return kiblnd_active_connect(cmid);
2912 CNETERR("Can't resolve route for %s: %d\n",
2913 libcfs_nid2str(peer->ibp_nid), event->status);
2914 kiblnd_peer_connect_failed(peer, 1, event->status);
2915 kiblnd_peer_decref(peer);
2916 return event->status; /* rc != 0 destroys cmid */
2918 case RDMA_CM_EVENT_UNREACHABLE:
2919 conn = (kib_conn_t *)cmid->context;
2920 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
2921 conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
2922 CNETERR("%s: UNREACHABLE %d\n",
2923 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
2924 kiblnd_connreq_done(conn, -ENETDOWN);
2925 kiblnd_conn_decref(conn);
2928 case RDMA_CM_EVENT_CONNECT_ERROR:
2929 conn = (kib_conn_t *)cmid->context;
2930 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
2931 conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
2932 CNETERR("%s: CONNECT ERROR %d\n",
2933 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
2934 kiblnd_connreq_done(conn, -ENOTCONN);
2935 kiblnd_conn_decref(conn);
2938 case RDMA_CM_EVENT_REJECTED:
2939 conn = (kib_conn_t *)cmid->context;
2940 switch (conn->ibc_state) {
2944 case IBLND_CONN_PASSIVE_WAIT:
2945 CERROR ("%s: REJECTED %d\n",
2946 libcfs_nid2str(conn->ibc_peer->ibp_nid),
2948 kiblnd_connreq_done(conn, -ECONNRESET);
2951 case IBLND_CONN_ACTIVE_CONNECT:
2952 kiblnd_rejected(conn, event->status,
2953 (void *)KIBLND_CONN_PARAM(event),
2954 KIBLND_CONN_PARAM_LEN(event));
2957 kiblnd_conn_decref(conn);
2960 case RDMA_CM_EVENT_ESTABLISHED:
2961 conn = (kib_conn_t *)cmid->context;
2962 switch (conn->ibc_state) {
2966 case IBLND_CONN_PASSIVE_WAIT:
2967 CDEBUG(D_NET, "ESTABLISHED (passive): %s\n",
2968 libcfs_nid2str(conn->ibc_peer->ibp_nid));
2969 kiblnd_connreq_done(conn, 0);
2972 case IBLND_CONN_ACTIVE_CONNECT:
2973 CDEBUG(D_NET, "ESTABLISHED(active): %s\n",
2974 libcfs_nid2str(conn->ibc_peer->ibp_nid));
2975 kiblnd_check_connreply(conn,
2976 (void *)KIBLND_CONN_PARAM(event),
2977 KIBLND_CONN_PARAM_LEN(event));
2980 /* net keeps its ref on conn! */
2983 #ifdef HAVE_OFED_RDMA_CMEV_TIMEWAIT_EXIT
2984 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
2985 CDEBUG(D_NET, "Ignore TIMEWAIT_EXIT event\n");
2988 case RDMA_CM_EVENT_DISCONNECTED:
2989 conn = (kib_conn_t *)cmid->context;
2990 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
2991 CERROR("%s DISCONNECTED\n",
2992 libcfs_nid2str(conn->ibc_peer->ibp_nid));
2993 kiblnd_connreq_done(conn, -ECONNRESET);
2995 kiblnd_close_conn(conn, 0);
2997 kiblnd_conn_decref(conn);
2998 cmid->context = NULL;
3001 case RDMA_CM_EVENT_DEVICE_REMOVAL:
3002 LCONSOLE_ERROR_MSG(0x131,
3003 "Received notification of device removal\n"
3004 "Please shutdown LNET to allow this to proceed\n");
3005 /* Can't remove network from underneath LNET for now, so I have
3009 #ifdef HAVE_OFED_RDMA_CMEV_ADDRCHANGE
3010 case RDMA_CM_EVENT_ADDR_CHANGE:
3011 LCONSOLE_INFO("Physical link changed (eg hca/port)\n");
3018 kiblnd_check_txs_locked(kib_conn_t *conn, struct list_head *txs)
3021 struct list_head *ttmp;
3023 list_for_each(ttmp, txs) {
3024 tx = list_entry(ttmp, kib_tx_t, tx_list);
3026 if (txs != &conn->ibc_active_txs) {
3027 LASSERT(tx->tx_queued);
3029 LASSERT(!tx->tx_queued);
3030 LASSERT(tx->tx_waiting || tx->tx_sending != 0);
3033 if (cfs_time_aftereq(jiffies, tx->tx_deadline)) {
3034 CERROR("Timed out tx: %s, %lu seconds\n",
3035 kiblnd_queue2str(conn, txs),
3036 cfs_duration_sec(jiffies - tx->tx_deadline));
3045 kiblnd_conn_timed_out_locked(kib_conn_t *conn)
3047 return kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue) ||
3048 kiblnd_check_txs_locked(conn, &conn->ibc_tx_noops) ||
3049 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_rsrvd) ||
3050 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_nocred) ||
3051 kiblnd_check_txs_locked(conn, &conn->ibc_active_txs);
3055 kiblnd_check_conns (int idx)
3057 struct list_head closes = LIST_HEAD_INIT(closes);
3058 struct list_head checksends = LIST_HEAD_INIT(checksends);
3059 struct list_head *peers = &kiblnd_data.kib_peers[idx];
3060 struct list_head *ptmp;
3063 struct list_head *ctmp;
3064 unsigned long flags;
3066 /* NB. We expect to have a look at all the peers and not find any
3067 * RDMAs to time out, so we just use a shared lock while we
3069 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3071 list_for_each(ptmp, peers) {
3072 peer = list_entry(ptmp, kib_peer_t, ibp_list);
3074 list_for_each(ctmp, &peer->ibp_conns) {
3078 conn = list_entry(ctmp, kib_conn_t, ibc_list);
3080 LASSERT(conn->ibc_state == IBLND_CONN_ESTABLISHED);
3082 spin_lock(&conn->ibc_lock);
3084 sendnoop = kiblnd_need_noop(conn);
3085 timedout = kiblnd_conn_timed_out_locked(conn);
3086 if (!sendnoop && !timedout) {
3087 spin_unlock(&conn->ibc_lock);
3092 CERROR("Timed out RDMA with %s (%lu): "
3093 "c: %u, oc: %u, rc: %u\n",
3094 libcfs_nid2str(peer->ibp_nid),
3095 cfs_duration_sec(cfs_time_current() -
3096 peer->ibp_last_alive),
3098 conn->ibc_outstanding_credits,
3099 conn->ibc_reserved_credits);
3100 list_add(&conn->ibc_connd_list, &closes);
3102 list_add(&conn->ibc_connd_list, &checksends);
3104 /* +ref for 'closes' or 'checksends' */
3105 kiblnd_conn_addref(conn);
3107 spin_unlock(&conn->ibc_lock);
3111 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3113 /* Handle timeout by closing the whole
3114 * connection. We can only be sure RDMA activity
3115 * has ceased once the QP has been modified. */
3116 while (!list_empty(&closes)) {
3117 conn = list_entry(closes.next,
3118 kib_conn_t, ibc_connd_list);
3119 list_del(&conn->ibc_connd_list);
3120 kiblnd_close_conn(conn, -ETIMEDOUT);
3121 kiblnd_conn_decref(conn);
3124 /* In case we have enough credits to return via a
3125 * NOOP, but there were no non-blocking tx descs
3126 * free to do it last time... */
3127 while (!list_empty(&checksends)) {
3128 conn = list_entry(checksends.next,
3129 kib_conn_t, ibc_connd_list);
3130 list_del(&conn->ibc_connd_list);
3131 kiblnd_check_sends(conn);
3132 kiblnd_conn_decref(conn);
3137 kiblnd_disconnect_conn (kib_conn_t *conn)
3139 LASSERT (!in_interrupt());
3140 LASSERT (current == kiblnd_data.kib_connd);
3141 LASSERT (conn->ibc_state == IBLND_CONN_CLOSING);
3143 rdma_disconnect(conn->ibc_cmid);
3144 kiblnd_finalise_conn(conn);
3146 kiblnd_peer_notify(conn->ibc_peer);
3150 kiblnd_connd (void *arg)
3153 unsigned long flags;
3159 unsigned long deadline = jiffies;
3161 cfs_block_allsigs ();
3163 init_waitqueue_entry_current (&wait);
3164 kiblnd_data.kib_connd = current;
3166 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3168 while (!kiblnd_data.kib_shutdown) {
3172 if (!list_empty(&kiblnd_data.kib_connd_zombies)) {
3173 conn = list_entry(kiblnd_data. \
3174 kib_connd_zombies.next,
3175 kib_conn_t, ibc_list);
3176 list_del(&conn->ibc_list);
3178 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock,
3182 kiblnd_destroy_conn(conn);
3184 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3187 if (!list_empty(&kiblnd_data.kib_connd_conns)) {
3188 conn = list_entry(kiblnd_data.kib_connd_conns.next,
3189 kib_conn_t, ibc_list);
3190 list_del(&conn->ibc_list);
3192 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock,
3196 kiblnd_disconnect_conn(conn);
3197 kiblnd_conn_decref(conn);
3199 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3202 /* careful with the jiffy wrap... */
3203 timeout = (int)(deadline - jiffies);
3207 int chunk = kiblnd_data.kib_peer_hash_size;
3209 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
3212 /* Time to check for RDMA timeouts on a few more
3213 * peers: I do checks every 'p' seconds on a
3214 * proportion of the peer table and I need to check
3215 * every connection 'n' times within a timeout
3216 * interval, to ensure I detect a timeout on any
3217 * connection within (n+1)/n times the timeout
3220 if (*kiblnd_tunables.kib_timeout > n * p)
3221 chunk = (chunk * n * p) /
3222 *kiblnd_tunables.kib_timeout;
3226 for (i = 0; i < chunk; i++) {
3227 kiblnd_check_conns(peer_index);
3228 peer_index = (peer_index + 1) %
3229 kiblnd_data.kib_peer_hash_size;
3232 deadline += msecs_to_jiffies(p * MSEC_PER_SEC);
3233 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3239 /* Nothing to do for 'timeout' */
3240 set_current_state(TASK_INTERRUPTIBLE);
3241 add_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3242 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
3244 waitq_timedwait(&wait, TASK_INTERRUPTIBLE, timeout);
3246 set_current_state(TASK_RUNNING);
3247 remove_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3248 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3251 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
3253 kiblnd_thread_fini();
3258 kiblnd_qp_event(struct ib_event *event, void *arg)
3260 kib_conn_t *conn = arg;
3262 switch (event->event) {
3263 case IB_EVENT_COMM_EST:
3264 CDEBUG(D_NET, "%s established\n",
3265 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3269 CERROR("%s: Async QP event type %d\n",
3270 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3276 kiblnd_complete (struct ib_wc *wc)
3278 switch (kiblnd_wreqid2type(wc->wr_id)) {
3282 case IBLND_WID_RDMA:
3283 /* We only get RDMA completion notification if it fails. All
3284 * subsequent work items, including the final SEND will fail
3285 * too. However we can't print out any more info about the
3286 * failing RDMA because 'tx' might be back on the idle list or
3287 * even reused already if we didn't manage to post all our work
3289 CNETERR("RDMA (tx: %p) failed: %d\n",
3290 kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3294 kiblnd_tx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3298 kiblnd_rx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status,
3305 kiblnd_cq_completion(struct ib_cq *cq, void *arg)
3307 /* NB I'm not allowed to schedule this conn once its refcount has
3308 * reached 0. Since fundamentally I'm racing with scheduler threads
3309 * consuming my CQ I could be called after all completions have
3310 * occurred. But in this case, ibc_nrx == 0 && ibc_nsends_posted == 0
3311 * and this CQ is about to be destroyed so I NOOP. */
3312 kib_conn_t *conn = (kib_conn_t *)arg;
3313 struct kib_sched_info *sched = conn->ibc_sched;
3314 unsigned long flags;
3316 LASSERT(cq == conn->ibc_cq);
3318 spin_lock_irqsave(&sched->ibs_lock, flags);
3320 conn->ibc_ready = 1;
3322 if (!conn->ibc_scheduled &&
3323 (conn->ibc_nrx > 0 ||
3324 conn->ibc_nsends_posted > 0)) {
3325 kiblnd_conn_addref(conn); /* +1 ref for sched_conns */
3326 conn->ibc_scheduled = 1;
3327 list_add_tail(&conn->ibc_sched_list, &sched->ibs_conns);
3329 if (waitqueue_active(&sched->ibs_waitq))
3330 wake_up(&sched->ibs_waitq);
3333 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3337 kiblnd_cq_event(struct ib_event *event, void *arg)
3339 kib_conn_t *conn = arg;
3341 CERROR("%s: async CQ event type %d\n",
3342 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3346 kiblnd_scheduler(void *arg)
3348 long id = (long)arg;
3349 struct kib_sched_info *sched;
3352 unsigned long flags;
3358 cfs_block_allsigs();
3360 init_waitqueue_entry_current(&wait);
3362 sched = kiblnd_data.kib_scheds[KIB_THREAD_CPT(id)];
3364 rc = cfs_cpt_bind(lnet_cpt_table(), sched->ibs_cpt);
3366 CWARN("Failed to bind on CPT %d, please verify whether "
3367 "all CPUs are healthy and reload modules if necessary, "
3368 "otherwise your system might under risk of low "
3369 "performance\n", sched->ibs_cpt);
3372 spin_lock_irqsave(&sched->ibs_lock, flags);
3374 while (!kiblnd_data.kib_shutdown) {
3375 if (busy_loops++ >= IBLND_RESCHED) {
3376 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3381 spin_lock_irqsave(&sched->ibs_lock, flags);
3386 if (!list_empty(&sched->ibs_conns)) {
3387 conn = list_entry(sched->ibs_conns.next,
3388 kib_conn_t, ibc_sched_list);
3389 /* take over kib_sched_conns' ref on conn... */
3390 LASSERT(conn->ibc_scheduled);
3391 list_del(&conn->ibc_sched_list);
3392 conn->ibc_ready = 0;
3394 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3396 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3398 rc = ib_req_notify_cq(conn->ibc_cq,
3401 CWARN("%s: ib_req_notify_cq failed: %d, "
3402 "closing connection\n",
3403 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
3404 kiblnd_close_conn(conn, -EIO);
3405 kiblnd_conn_decref(conn);
3406 spin_lock_irqsave(&sched->ibs_lock,
3411 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3415 CWARN("%s: ib_poll_cq failed: %d, "
3416 "closing connection\n",
3417 libcfs_nid2str(conn->ibc_peer->ibp_nid),
3419 kiblnd_close_conn(conn, -EIO);
3420 kiblnd_conn_decref(conn);
3421 spin_lock_irqsave(&sched->ibs_lock, flags);
3425 spin_lock_irqsave(&sched->ibs_lock, flags);
3427 if (rc != 0 || conn->ibc_ready) {
3428 /* There may be another completion waiting; get
3429 * another scheduler to check while I handle
3431 /* +1 ref for sched_conns */
3432 kiblnd_conn_addref(conn);
3433 list_add_tail(&conn->ibc_sched_list,
3435 if (waitqueue_active(&sched->ibs_waitq))
3436 wake_up(&sched->ibs_waitq);
3438 conn->ibc_scheduled = 0;
3442 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3443 kiblnd_complete(&wc);
3445 spin_lock_irqsave(&sched->ibs_lock, flags);
3448 kiblnd_conn_decref(conn); /* ...drop my ref from above */
3455 set_current_state(TASK_INTERRUPTIBLE);
3456 add_wait_queue_exclusive(&sched->ibs_waitq, &wait);
3457 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3459 waitq_wait(&wait, TASK_INTERRUPTIBLE);
3462 remove_wait_queue(&sched->ibs_waitq, &wait);
3463 set_current_state(TASK_RUNNING);
3464 spin_lock_irqsave(&sched->ibs_lock, flags);
3467 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3469 kiblnd_thread_fini();
3474 kiblnd_failover_thread(void *arg)
3476 rwlock_t *glock = &kiblnd_data.kib_global_lock;
3479 unsigned long flags;
3482 LASSERT (*kiblnd_tunables.kib_dev_failover != 0);
3484 cfs_block_allsigs ();
3486 init_waitqueue_entry_current(&wait);
3487 write_lock_irqsave(glock, flags);
3489 while (!kiblnd_data.kib_shutdown) {
3490 int do_failover = 0;
3493 list_for_each_entry(dev, &kiblnd_data.kib_failed_devs,
3495 if (cfs_time_before(cfs_time_current(),
3496 dev->ibd_next_failover))
3503 list_del_init(&dev->ibd_fail_list);
3504 dev->ibd_failover = 1;
3505 write_unlock_irqrestore(glock, flags);
3507 rc = kiblnd_dev_failover(dev);
3509 write_lock_irqsave(glock, flags);
3511 LASSERT (dev->ibd_failover);
3512 dev->ibd_failover = 0;
3513 if (rc >= 0) { /* Device is OK or failover succeed */
3514 dev->ibd_next_failover = cfs_time_shift(3);
3518 /* failed to failover, retry later */
3519 dev->ibd_next_failover =
3520 cfs_time_shift(min(dev->ibd_failed_failover, 10));
3521 if (kiblnd_dev_can_failover(dev)) {
3522 list_add_tail(&dev->ibd_fail_list,
3523 &kiblnd_data.kib_failed_devs);
3529 /* long sleep if no more pending failover */
3530 long_sleep = list_empty(&kiblnd_data.kib_failed_devs);
3532 set_current_state(TASK_INTERRUPTIBLE);
3533 add_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3534 write_unlock_irqrestore(glock, flags);
3536 rc = schedule_timeout(long_sleep ? cfs_time_seconds(10) :
3537 cfs_time_seconds(1));
3538 set_current_state(TASK_RUNNING);
3539 remove_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3540 write_lock_irqsave(glock, flags);
3542 if (!long_sleep || rc != 0)
3545 /* have a long sleep, routine check all active devices,
3546 * we need checking like this because if there is not active
3547 * connection on the dev and no SEND from local, we may listen
3548 * on wrong HCA for ever while there is a bonding failover */
3549 list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3550 if (kiblnd_dev_can_failover(dev)) {
3551 list_add_tail(&dev->ibd_fail_list,
3552 &kiblnd_data.kib_failed_devs);
3557 write_unlock_irqrestore(glock, flags);
3559 kiblnd_thread_fini();