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, 2015, 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);
52 static void kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx);
55 kiblnd_tx_done (lnet_ni_t *ni, kib_tx_t *tx)
57 lnet_msg_t *lntmsg[2];
58 kib_net_t *net = ni->ni_data;
62 LASSERT (net != NULL);
63 LASSERT (!in_interrupt());
64 LASSERT (!tx->tx_queued); /* mustn't be queued for sending */
65 LASSERT (tx->tx_sending == 0); /* mustn't be awaiting sent callback */
66 LASSERT (!tx->tx_waiting); /* mustn't be awaiting peer response */
67 LASSERT (tx->tx_pool != NULL);
69 kiblnd_unmap_tx(ni, tx);
71 /* tx may have up to 2 lnet msgs to finalise */
72 lntmsg[0] = tx->tx_lntmsg[0]; tx->tx_lntmsg[0] = NULL;
73 lntmsg[1] = tx->tx_lntmsg[1]; tx->tx_lntmsg[1] = NULL;
76 if (tx->tx_conn != NULL) {
77 LASSERT (ni == tx->tx_conn->ibc_peer->ibp_ni);
79 kiblnd_conn_decref(tx->tx_conn);
86 kiblnd_pool_free_node(&tx->tx_pool->tpo_pool, &tx->tx_list);
88 /* delay finalize until my descs have been freed */
89 for (i = 0; i < 2; i++) {
90 if (lntmsg[i] == NULL)
93 lnet_finalize(ni, lntmsg[i], rc);
98 kiblnd_txlist_done(lnet_ni_t *ni, struct list_head *txlist, int status)
102 while (!list_empty(txlist)) {
103 tx = list_entry(txlist->next, kib_tx_t, tx_list);
105 list_del(&tx->tx_list);
108 tx->tx_status = status;
109 kiblnd_tx_done(ni, tx);
114 kiblnd_get_idle_tx(lnet_ni_t *ni, lnet_nid_t target)
116 kib_net_t *net = (kib_net_t *)ni->ni_data;
117 struct list_head *node;
119 kib_tx_poolset_t *tps;
121 tps = net->ibn_tx_ps[lnet_cpt_of_nid(target)];
122 node = kiblnd_pool_alloc_node(&tps->tps_poolset);
125 tx = container_of(node, kib_tx_t, tx_list);
127 LASSERT (tx->tx_nwrq == 0);
128 LASSERT (!tx->tx_queued);
129 LASSERT (tx->tx_sending == 0);
130 LASSERT (!tx->tx_waiting);
131 LASSERT (tx->tx_status == 0);
132 LASSERT (tx->tx_conn == NULL);
133 LASSERT (tx->tx_lntmsg[0] == NULL);
134 LASSERT (tx->tx_lntmsg[1] == 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;
161 struct ib_mr *mr = conn->ibc_hdev->ibh_mrs;
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);
171 rx->rx_sge.lkey = mr->lkey;
172 rx->rx_sge.addr = rx->rx_msgaddr;
173 rx->rx_sge.length = IBLND_MSG_SIZE;
175 rx->rx_wrq.next = NULL;
176 rx->rx_wrq.sg_list = &rx->rx_sge;
177 rx->rx_wrq.num_sge = 1;
178 rx->rx_wrq.wr_id = kiblnd_ptr2wreqid(rx, IBLND_WID_RX);
180 LASSERT (conn->ibc_state >= IBLND_CONN_INIT);
181 LASSERT (rx->rx_nob >= 0); /* not posted */
183 if (conn->ibc_state > IBLND_CONN_ESTABLISHED) {
184 kiblnd_drop_rx(rx); /* No more posts for this rx */
188 rx->rx_nob = -1; /* flag posted */
190 /* NB: need an extra reference after ib_post_recv because we don't
191 * own this rx (and rx::rx_conn) anymore, LU-5678.
193 kiblnd_conn_addref(conn);
194 rc = ib_post_recv(conn->ibc_cmid->qp, &rx->rx_wrq, &bad_wrq);
195 if (unlikely(rc != 0)) {
196 CERROR("Can't post rx for %s: %d, bad_wrq: %p\n",
197 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc, bad_wrq);
201 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) /* Initial post */
204 if (unlikely(rc != 0)) {
205 kiblnd_close_conn(conn, rc);
206 kiblnd_drop_rx(rx); /* No more posts for this rx */
210 if (credit == IBLND_POSTRX_NO_CREDIT)
213 spin_lock(&conn->ibc_lock);
214 if (credit == IBLND_POSTRX_PEER_CREDIT)
215 conn->ibc_outstanding_credits++;
217 conn->ibc_reserved_credits++;
218 spin_unlock(&conn->ibc_lock);
220 kiblnd_check_sends(conn);
222 kiblnd_conn_decref(conn);
227 kiblnd_find_waiting_tx_locked(kib_conn_t *conn, int txtype, __u64 cookie)
229 struct list_head *tmp;
231 list_for_each(tmp, &conn->ibc_active_txs) {
232 kib_tx_t *tx = list_entry(tmp, kib_tx_t, tx_list);
234 LASSERT(!tx->tx_queued);
235 LASSERT(tx->tx_sending != 0 || tx->tx_waiting);
237 if (tx->tx_cookie != cookie)
240 if (tx->tx_waiting &&
241 tx->tx_msg->ibm_type == txtype)
244 CWARN("Bad completion: %swaiting, type %x (wanted %x)\n",
245 tx->tx_waiting ? "" : "NOT ",
246 tx->tx_msg->ibm_type, txtype);
252 kiblnd_handle_completion(kib_conn_t *conn, int txtype, int status, __u64 cookie)
255 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
258 spin_lock(&conn->ibc_lock);
260 tx = kiblnd_find_waiting_tx_locked(conn, txtype, cookie);
262 spin_unlock(&conn->ibc_lock);
264 CWARN("Unmatched completion type %x cookie "LPX64" from %s\n",
265 txtype, cookie, libcfs_nid2str(conn->ibc_peer->ibp_nid));
266 kiblnd_close_conn(conn, -EPROTO);
270 if (tx->tx_status == 0) { /* success so far */
271 if (status < 0) { /* failed? */
272 tx->tx_status = status;
273 } else if (txtype == IBLND_MSG_GET_REQ) {
274 lnet_set_reply_msg_len(ni, tx->tx_lntmsg[1], status);
280 idle = !tx->tx_queued && (tx->tx_sending == 0);
282 list_del(&tx->tx_list);
284 spin_unlock(&conn->ibc_lock);
287 kiblnd_tx_done(ni, tx);
291 kiblnd_send_completion(kib_conn_t *conn, int type, int status, __u64 cookie)
293 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
294 kib_tx_t *tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
297 CERROR("Can't get tx for completion %x for %s\n",
298 type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
302 tx->tx_msg->ibm_u.completion.ibcm_status = status;
303 tx->tx_msg->ibm_u.completion.ibcm_cookie = cookie;
304 kiblnd_init_tx_msg(ni, tx, type, sizeof(kib_completion_msg_t));
306 kiblnd_queue_tx(tx, conn);
310 kiblnd_handle_rx (kib_rx_t *rx)
312 kib_msg_t *msg = rx->rx_msg;
313 kib_conn_t *conn = rx->rx_conn;
314 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
315 int credits = msg->ibm_credits;
321 LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
323 CDEBUG (D_NET, "Received %x[%d] from %s\n",
324 msg->ibm_type, credits,
325 libcfs_nid2str(conn->ibc_peer->ibp_nid));
328 /* Have I received credits that will let me send? */
329 spin_lock(&conn->ibc_lock);
331 if (conn->ibc_credits + credits >
332 conn->ibc_queue_depth) {
333 rc2 = conn->ibc_credits;
334 spin_unlock(&conn->ibc_lock);
336 CERROR("Bad credits from %s: %d + %d > %d\n",
337 libcfs_nid2str(conn->ibc_peer->ibp_nid),
339 conn->ibc_queue_depth);
341 kiblnd_close_conn(conn, -EPROTO);
342 kiblnd_post_rx(rx, IBLND_POSTRX_NO_CREDIT);
346 conn->ibc_credits += credits;
348 /* This ensures the credit taken by NOOP can be returned */
349 if (msg->ibm_type == IBLND_MSG_NOOP &&
350 !IBLND_OOB_CAPABLE(conn->ibc_version)) /* v1 only */
351 conn->ibc_outstanding_credits++;
353 spin_unlock(&conn->ibc_lock);
354 kiblnd_check_sends(conn);
357 switch (msg->ibm_type) {
359 CERROR("Bad IBLND message type %x from %s\n",
360 msg->ibm_type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
361 post_credit = IBLND_POSTRX_NO_CREDIT;
366 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
367 post_credit = IBLND_POSTRX_NO_CREDIT;
371 if (credits != 0) /* credit already posted */
372 post_credit = IBLND_POSTRX_NO_CREDIT;
373 else /* a keepalive NOOP */
374 post_credit = IBLND_POSTRX_PEER_CREDIT;
377 case IBLND_MSG_IMMEDIATE:
378 post_credit = IBLND_POSTRX_DONT_POST;
379 rc = lnet_parse(ni, &msg->ibm_u.immediate.ibim_hdr,
380 msg->ibm_srcnid, rx, 0);
381 if (rc < 0) /* repost on error */
382 post_credit = IBLND_POSTRX_PEER_CREDIT;
385 case IBLND_MSG_PUT_REQ:
386 post_credit = IBLND_POSTRX_DONT_POST;
387 rc = lnet_parse(ni, &msg->ibm_u.putreq.ibprm_hdr,
388 msg->ibm_srcnid, rx, 1);
389 if (rc < 0) /* repost on error */
390 post_credit = IBLND_POSTRX_PEER_CREDIT;
393 case IBLND_MSG_PUT_NAK:
394 CWARN ("PUT_NACK from %s\n",
395 libcfs_nid2str(conn->ibc_peer->ibp_nid));
396 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
397 kiblnd_handle_completion(conn, IBLND_MSG_PUT_REQ,
398 msg->ibm_u.completion.ibcm_status,
399 msg->ibm_u.completion.ibcm_cookie);
402 case IBLND_MSG_PUT_ACK:
403 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
405 spin_lock(&conn->ibc_lock);
406 tx = kiblnd_find_waiting_tx_locked(conn, IBLND_MSG_PUT_REQ,
407 msg->ibm_u.putack.ibpam_src_cookie);
409 list_del(&tx->tx_list);
410 spin_unlock(&conn->ibc_lock);
413 CERROR("Unmatched PUT_ACK from %s\n",
414 libcfs_nid2str(conn->ibc_peer->ibp_nid));
419 LASSERT (tx->tx_waiting);
420 /* CAVEAT EMPTOR: I could be racing with tx_complete, but...
421 * (a) I can overwrite tx_msg since my peer has received it!
422 * (b) tx_waiting set tells tx_complete() it's not done. */
424 tx->tx_nwrq = 0; /* overwrite PUT_REQ */
426 rc2 = kiblnd_init_rdma(conn, tx, IBLND_MSG_PUT_DONE,
427 kiblnd_rd_size(&msg->ibm_u.putack.ibpam_rd),
428 &msg->ibm_u.putack.ibpam_rd,
429 msg->ibm_u.putack.ibpam_dst_cookie);
431 CERROR("Can't setup rdma for PUT to %s: %d\n",
432 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc2);
434 spin_lock(&conn->ibc_lock);
435 tx->tx_waiting = 0; /* clear waiting and queue atomically */
436 kiblnd_queue_tx_locked(tx, conn);
437 spin_unlock(&conn->ibc_lock);
440 case IBLND_MSG_PUT_DONE:
441 post_credit = IBLND_POSTRX_PEER_CREDIT;
442 kiblnd_handle_completion(conn, IBLND_MSG_PUT_ACK,
443 msg->ibm_u.completion.ibcm_status,
444 msg->ibm_u.completion.ibcm_cookie);
447 case IBLND_MSG_GET_REQ:
448 post_credit = IBLND_POSTRX_DONT_POST;
449 rc = lnet_parse(ni, &msg->ibm_u.get.ibgm_hdr,
450 msg->ibm_srcnid, rx, 1);
451 if (rc < 0) /* repost on error */
452 post_credit = IBLND_POSTRX_PEER_CREDIT;
455 case IBLND_MSG_GET_DONE:
456 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
457 kiblnd_handle_completion(conn, IBLND_MSG_GET_REQ,
458 msg->ibm_u.completion.ibcm_status,
459 msg->ibm_u.completion.ibcm_cookie);
463 if (rc < 0) /* protocol error */
464 kiblnd_close_conn(conn, rc);
466 if (post_credit != IBLND_POSTRX_DONT_POST)
467 kiblnd_post_rx(rx, post_credit);
471 kiblnd_rx_complete (kib_rx_t *rx, int status, int nob)
473 kib_msg_t *msg = rx->rx_msg;
474 kib_conn_t *conn = rx->rx_conn;
475 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
476 kib_net_t *net = ni->ni_data;
480 LASSERT (net != NULL);
481 LASSERT (rx->rx_nob < 0); /* was posted */
482 rx->rx_nob = 0; /* isn't now */
484 if (conn->ibc_state > IBLND_CONN_ESTABLISHED)
487 if (status != IB_WC_SUCCESS) {
488 CNETERR("Rx from %s failed: %d\n",
489 libcfs_nid2str(conn->ibc_peer->ibp_nid), status);
496 rc = kiblnd_unpack_msg(msg, rx->rx_nob);
498 CERROR ("Error %d unpacking rx from %s\n",
499 rc, libcfs_nid2str(conn->ibc_peer->ibp_nid));
503 if (msg->ibm_srcnid != conn->ibc_peer->ibp_nid ||
504 msg->ibm_dstnid != ni->ni_nid ||
505 msg->ibm_srcstamp != conn->ibc_incarnation ||
506 msg->ibm_dststamp != net->ibn_incarnation) {
507 CERROR ("Stale rx from %s\n",
508 libcfs_nid2str(conn->ibc_peer->ibp_nid));
513 /* set time last known alive */
514 kiblnd_peer_alive(conn->ibc_peer);
516 /* racing with connection establishment/teardown! */
518 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
519 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
522 write_lock_irqsave(g_lock, flags);
523 /* must check holding global lock to eliminate race */
524 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
525 list_add_tail(&rx->rx_list, &conn->ibc_early_rxs);
526 write_unlock_irqrestore(g_lock, flags);
529 write_unlock_irqrestore(g_lock, flags);
531 kiblnd_handle_rx(rx);
535 CDEBUG(D_NET, "rx %p conn %p\n", rx, conn);
536 kiblnd_close_conn(conn, err);
538 kiblnd_drop_rx(rx); /* Don't re-post rx. */
542 kiblnd_kvaddr_to_page (unsigned long vaddr)
546 if (is_vmalloc_addr((void *)vaddr)) {
547 page = vmalloc_to_page ((void *)vaddr);
548 LASSERT (page != NULL);
551 #ifdef CONFIG_HIGHMEM
552 if (vaddr >= PKMAP_BASE &&
553 vaddr < (PKMAP_BASE + LAST_PKMAP * PAGE_SIZE)) {
554 /* No highmem pages only used for bulk (kiov) I/O */
555 CERROR("find page for address in highmem\n");
559 page = virt_to_page (vaddr);
560 LASSERT (page != NULL);
565 kiblnd_fmr_map_tx(kib_net_t *net, kib_tx_t *tx, kib_rdma_desc_t *rd, int nob)
568 __u64 *pages = tx->tx_pages;
569 kib_fmr_poolset_t *fps;
576 LASSERT(tx->tx_pool != NULL);
577 LASSERT(tx->tx_pool->tpo_pool.po_owner != NULL);
579 hdev = tx->tx_pool->tpo_hdev;
581 for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
582 for (size = 0; size < rd->rd_frags[i].rf_nob;
583 size += hdev->ibh_page_size) {
584 pages[npages ++] = (rd->rd_frags[i].rf_addr &
585 hdev->ibh_page_mask) + size;
589 cpt = tx->tx_pool->tpo_pool.po_owner->ps_cpt;
591 fps = net->ibn_fmr_ps[cpt];
592 rc = kiblnd_fmr_pool_map(fps, pages, npages, 0, &tx->fmr);
594 CERROR ("Can't map %d pages: %d\n", npages, rc);
598 /* If rd is not tx_rd, it's going to get sent to a peer, who will need
600 rd->rd_key = (rd != tx->tx_rd) ? tx->fmr.fmr_pfmr->fmr->rkey :
601 tx->fmr.fmr_pfmr->fmr->lkey;
602 rd->rd_frags[0].rf_addr &= ~hdev->ibh_page_mask;
603 rd->rd_frags[0].rf_nob = nob;
610 kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx)
612 kib_net_t *net = ni->ni_data;
614 LASSERT(net != NULL);
616 if (net->ibn_fmr_ps != NULL && tx->fmr.fmr_pfmr != NULL) {
617 kiblnd_fmr_pool_unmap(&tx->fmr, tx->tx_status);
618 tx->fmr.fmr_pfmr = NULL;
621 if (tx->tx_nfrags != 0) {
622 kiblnd_dma_unmap_sg(tx->tx_pool->tpo_hdev->ibh_ibdev,
623 tx->tx_frags, tx->tx_nfrags, tx->tx_dmadir);
629 kiblnd_map_tx(lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd, int nfrags)
631 kib_hca_dev_t *hdev = tx->tx_pool->tpo_hdev;
632 kib_net_t *net = ni->ni_data;
633 struct ib_mr *mr = NULL;
637 /* If rd is not tx_rd, it's going to get sent to a peer and I'm the
639 tx->tx_dmadir = (rd != tx->tx_rd) ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
640 tx->tx_nfrags = nfrags;
642 rd->rd_nfrags = kiblnd_dma_map_sg(hdev->ibh_ibdev, tx->tx_frags,
643 tx->tx_nfrags, tx->tx_dmadir);
645 for (i = 0, nob = 0; i < rd->rd_nfrags; i++) {
646 rd->rd_frags[i].rf_nob = kiblnd_sg_dma_len(
647 hdev->ibh_ibdev, &tx->tx_frags[i]);
648 rd->rd_frags[i].rf_addr = kiblnd_sg_dma_address(
649 hdev->ibh_ibdev, &tx->tx_frags[i]);
650 nob += rd->rd_frags[i].rf_nob;
653 mr = kiblnd_find_rd_dma_mr(hdev, rd,
654 (tx->tx_conn != NULL) ?
655 tx->tx_conn->ibc_max_frags : -1);
657 /* found pre-mapping MR */
658 rd->rd_key = (rd != tx->tx_rd) ? mr->rkey : mr->lkey;
662 if (net->ibn_fmr_ps != NULL)
663 return kiblnd_fmr_map_tx(net, tx, rd, nob);
670 kiblnd_setup_rd_iov(lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd,
671 unsigned int niov, struct iovec *iov, int offset, int nob)
673 kib_net_t *net = ni->ni_data;
675 struct scatterlist *sg;
682 LASSERT (net != NULL);
684 while (offset >= iov->iov_len) {
685 offset -= iov->iov_len;
695 vaddr = ((unsigned long)iov->iov_base) + offset;
696 page_offset = vaddr & (PAGE_SIZE - 1);
697 page = kiblnd_kvaddr_to_page(vaddr);
699 CERROR ("Can't find page\n");
703 fragnob = min((int)(iov->iov_len - offset), nob);
704 fragnob = min(fragnob, (int)PAGE_SIZE - page_offset);
706 sg_set_page(sg, page, fragnob, page_offset);
709 if (offset + fragnob < iov->iov_len) {
719 return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags);
723 kiblnd_setup_rd_kiov (lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd,
724 int nkiov, lnet_kiov_t *kiov, int offset, int nob)
726 kib_net_t *net = ni->ni_data;
727 struct scatterlist *sg;
730 CDEBUG(D_NET, "niov %d offset %d nob %d\n", nkiov, offset, nob);
734 LASSERT (net != NULL);
736 while (offset >= kiov->kiov_len) {
737 offset -= kiov->kiov_len;
747 fragnob = min((int)(kiov->kiov_len - offset), nob);
749 sg_set_page(sg, kiov->kiov_page, fragnob,
750 kiov->kiov_offset + offset);
759 return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags);
763 kiblnd_post_tx_locked (kib_conn_t *conn, kib_tx_t *tx, int credit)
764 __must_hold(&conn->ibc_lock)
766 kib_msg_t *msg = tx->tx_msg;
767 kib_peer_t *peer = conn->ibc_peer;
768 int ver = conn->ibc_version;
771 struct ib_send_wr *bad_wrq;
773 LASSERT(tx->tx_queued);
774 /* We rely on this for QP sizing */
775 LASSERT(tx->tx_nwrq > 0);
776 LASSERT(tx->tx_nwrq <= 1 + conn->ibc_max_frags);
778 LASSERT(credit == 0 || credit == 1);
779 LASSERT(conn->ibc_outstanding_credits >= 0);
780 LASSERT(conn->ibc_outstanding_credits <= conn->ibc_queue_depth);
781 LASSERT(conn->ibc_credits >= 0);
782 LASSERT(conn->ibc_credits <= conn->ibc_queue_depth);
784 if (conn->ibc_nsends_posted == IBLND_CONCURRENT_SENDS(ver)) {
785 /* tx completions outstanding... */
786 CDEBUG(D_NET, "%s: posted enough\n",
787 libcfs_nid2str(peer->ibp_nid));
791 if (credit != 0 && conn->ibc_credits == 0) { /* no credits */
792 CDEBUG(D_NET, "%s: no credits\n",
793 libcfs_nid2str(peer->ibp_nid));
797 if (credit != 0 && !IBLND_OOB_CAPABLE(ver) &&
798 conn->ibc_credits == 1 && /* last credit reserved */
799 msg->ibm_type != IBLND_MSG_NOOP) { /* for NOOP */
800 CDEBUG(D_NET, "%s: not using last credit\n",
801 libcfs_nid2str(peer->ibp_nid));
805 /* NB don't drop ibc_lock before bumping tx_sending */
806 list_del(&tx->tx_list);
809 if (msg->ibm_type == IBLND_MSG_NOOP &&
810 (!kiblnd_need_noop(conn) || /* redundant NOOP */
811 (IBLND_OOB_CAPABLE(ver) && /* posted enough NOOP */
812 conn->ibc_noops_posted == IBLND_OOB_MSGS(ver)))) {
813 /* OK to drop when posted enough NOOPs, since
814 * kiblnd_check_sends will queue NOOP again when
815 * posted NOOPs complete */
816 spin_unlock(&conn->ibc_lock);
817 kiblnd_tx_done(peer->ibp_ni, tx);
818 spin_lock(&conn->ibc_lock);
819 CDEBUG(D_NET, "%s(%d): redundant or enough NOOP\n",
820 libcfs_nid2str(peer->ibp_nid),
821 conn->ibc_noops_posted);
825 kiblnd_pack_msg(peer->ibp_ni, msg, ver, conn->ibc_outstanding_credits,
826 peer->ibp_nid, conn->ibc_incarnation);
828 conn->ibc_credits -= credit;
829 conn->ibc_outstanding_credits = 0;
830 conn->ibc_nsends_posted++;
831 if (msg->ibm_type == IBLND_MSG_NOOP)
832 conn->ibc_noops_posted++;
834 /* CAVEAT EMPTOR! This tx could be the PUT_DONE of an RDMA
835 * PUT. If so, it was first queued here as a PUT_REQ, sent and
836 * stashed on ibc_active_txs, matched by an incoming PUT_ACK,
837 * and then re-queued here. It's (just) possible that
838 * tx_sending is non-zero if we've not done the tx_complete()
839 * from the first send; hence the ++ rather than = below. */
841 list_add(&tx->tx_list, &conn->ibc_active_txs);
843 /* I'm still holding ibc_lock! */
844 if (conn->ibc_state != IBLND_CONN_ESTABLISHED) {
846 } else if (tx->tx_pool->tpo_pool.po_failed ||
847 conn->ibc_hdev != tx->tx_pool->tpo_hdev) {
848 /* close_conn will launch failover */
851 rc = ib_post_send(conn->ibc_cmid->qp,
852 tx->tx_wrq, &bad_wrq);
855 conn->ibc_last_send = jiffies;
860 /* NB credits are transferred in the actual
861 * message, which can only be the last work item */
862 conn->ibc_credits += credit;
863 conn->ibc_outstanding_credits += msg->ibm_credits;
864 conn->ibc_nsends_posted--;
865 if (msg->ibm_type == IBLND_MSG_NOOP)
866 conn->ibc_noops_posted--;
872 done = (tx->tx_sending == 0);
874 list_del(&tx->tx_list);
876 spin_unlock(&conn->ibc_lock);
878 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
879 CERROR("Error %d posting transmit to %s\n",
880 rc, libcfs_nid2str(peer->ibp_nid));
882 CDEBUG(D_NET, "Error %d posting transmit to %s\n",
883 rc, libcfs_nid2str(peer->ibp_nid));
885 kiblnd_close_conn(conn, rc);
888 kiblnd_tx_done(peer->ibp_ni, tx);
890 spin_lock(&conn->ibc_lock);
896 kiblnd_check_sends (kib_conn_t *conn)
898 int ver = conn->ibc_version;
899 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
902 /* Don't send anything until after the connection is established */
903 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
904 CDEBUG(D_NET, "%s too soon\n",
905 libcfs_nid2str(conn->ibc_peer->ibp_nid));
909 spin_lock(&conn->ibc_lock);
911 LASSERT (conn->ibc_nsends_posted <= IBLND_CONCURRENT_SENDS(ver));
912 LASSERT (!IBLND_OOB_CAPABLE(ver) ||
913 conn->ibc_noops_posted <= IBLND_OOB_MSGS(ver));
914 LASSERT (conn->ibc_reserved_credits >= 0);
916 while (conn->ibc_reserved_credits > 0 &&
917 !list_empty(&conn->ibc_tx_queue_rsrvd)) {
918 tx = list_entry(conn->ibc_tx_queue_rsrvd.next,
920 list_del(&tx->tx_list);
921 list_add_tail(&tx->tx_list, &conn->ibc_tx_queue);
922 conn->ibc_reserved_credits--;
925 if (kiblnd_need_noop(conn)) {
926 spin_unlock(&conn->ibc_lock);
928 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
930 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_NOOP, 0);
932 spin_lock(&conn->ibc_lock);
934 kiblnd_queue_tx_locked(tx, conn);
937 kiblnd_conn_addref(conn); /* 1 ref for me.... (see b21911) */
942 if (!list_empty(&conn->ibc_tx_queue_nocred)) {
944 tx = list_entry(conn->ibc_tx_queue_nocred.next,
946 } else if (!list_empty(&conn->ibc_tx_noops)) {
947 LASSERT (!IBLND_OOB_CAPABLE(ver));
949 tx = list_entry(conn->ibc_tx_noops.next,
951 } else if (!list_empty(&conn->ibc_tx_queue)) {
953 tx = list_entry(conn->ibc_tx_queue.next,
958 if (kiblnd_post_tx_locked(conn, tx, credit) != 0)
962 spin_unlock(&conn->ibc_lock);
964 kiblnd_conn_decref(conn); /* ...until here */
968 kiblnd_tx_complete (kib_tx_t *tx, int status)
970 int failed = (status != IB_WC_SUCCESS);
971 kib_conn_t *conn = tx->tx_conn;
974 LASSERT (tx->tx_sending > 0);
977 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
978 CNETERR("Tx -> %s cookie "LPX64
979 " sending %d waiting %d: failed %d\n",
980 libcfs_nid2str(conn->ibc_peer->ibp_nid),
981 tx->tx_cookie, tx->tx_sending, tx->tx_waiting,
984 kiblnd_close_conn(conn, -EIO);
986 kiblnd_peer_alive(conn->ibc_peer);
989 spin_lock(&conn->ibc_lock);
991 /* I could be racing with rdma completion. Whoever makes 'tx' idle
992 * gets to free it, which also drops its ref on 'conn'. */
995 conn->ibc_nsends_posted--;
996 if (tx->tx_msg->ibm_type == IBLND_MSG_NOOP)
997 conn->ibc_noops_posted--;
1000 tx->tx_waiting = 0; /* don't wait for peer */
1001 tx->tx_status = -EIO;
1004 idle = (tx->tx_sending == 0) && /* This is the final callback */
1005 !tx->tx_waiting && /* Not waiting for peer */
1006 !tx->tx_queued; /* Not re-queued (PUT_DONE) */
1008 list_del(&tx->tx_list);
1010 kiblnd_conn_addref(conn); /* 1 ref for me.... */
1012 spin_unlock(&conn->ibc_lock);
1015 kiblnd_tx_done(conn->ibc_peer->ibp_ni, tx);
1017 kiblnd_check_sends(conn);
1019 kiblnd_conn_decref(conn); /* ...until here */
1023 kiblnd_init_tx_msg (lnet_ni_t *ni, kib_tx_t *tx, int type, int body_nob)
1025 kib_hca_dev_t *hdev = tx->tx_pool->tpo_hdev;
1026 struct ib_sge *sge = &tx->tx_sge[tx->tx_nwrq];
1027 struct ib_send_wr *wrq = &tx->tx_wrq[tx->tx_nwrq];
1028 int nob = offsetof (kib_msg_t, ibm_u) + body_nob;
1029 struct ib_mr *mr = hdev->ibh_mrs;
1031 LASSERT(tx->tx_nwrq >= 0);
1032 LASSERT(tx->tx_nwrq < IBLND_MAX_RDMA_FRAGS + 1);
1033 LASSERT(nob <= IBLND_MSG_SIZE);
1034 LASSERT(mr != NULL);
1036 kiblnd_init_msg(tx->tx_msg, type, body_nob);
1038 sge->lkey = mr->lkey;
1039 sge->addr = tx->tx_msgaddr;
1042 memset(wrq, 0, sizeof(*wrq));
1045 wrq->wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_TX);
1048 wrq->opcode = IB_WR_SEND;
1049 wrq->send_flags = IB_SEND_SIGNALED;
1055 kiblnd_init_rdma(kib_conn_t *conn, kib_tx_t *tx, int type,
1056 int resid, kib_rdma_desc_t *dstrd, __u64 dstcookie)
1058 kib_msg_t *ibmsg = tx->tx_msg;
1059 kib_rdma_desc_t *srcrd = tx->tx_rd;
1060 struct ib_sge *sge = &tx->tx_sge[0];
1061 struct ib_send_wr *wrq = &tx->tx_wrq[0];
1067 LASSERT (!in_interrupt());
1068 LASSERT (tx->tx_nwrq == 0);
1069 LASSERT (type == IBLND_MSG_GET_DONE ||
1070 type == IBLND_MSG_PUT_DONE);
1072 srcidx = dstidx = 0;
1075 if (srcidx >= srcrd->rd_nfrags) {
1076 CERROR("Src buffer exhausted: %d frags\n", srcidx);
1081 if (dstidx == dstrd->rd_nfrags) {
1082 CERROR("Dst buffer exhausted: %d frags\n", dstidx);
1087 if (tx->tx_nwrq >= conn->ibc_max_frags) {
1088 CERROR("RDMA has too many fragments for peer %s (%d), "
1089 "src idx/frags: %d/%d dst idx/frags: %d/%d\n",
1090 libcfs_nid2str(conn->ibc_peer->ibp_nid),
1091 conn->ibc_max_frags,
1092 srcidx, srcrd->rd_nfrags,
1093 dstidx, dstrd->rd_nfrags);
1098 wrknob = MIN(MIN(kiblnd_rd_frag_size(srcrd, srcidx),
1099 kiblnd_rd_frag_size(dstrd, dstidx)), resid);
1101 sge = &tx->tx_sge[tx->tx_nwrq];
1102 sge->addr = kiblnd_rd_frag_addr(srcrd, srcidx);
1103 sge->lkey = kiblnd_rd_frag_key(srcrd, srcidx);
1104 sge->length = wrknob;
1106 wrq = &tx->tx_wrq[tx->tx_nwrq];
1108 wrq->next = wrq + 1;
1109 wrq->wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_RDMA);
1112 wrq->opcode = IB_WR_RDMA_WRITE;
1113 wrq->send_flags = 0;
1115 wrq->wr.rdma.remote_addr = kiblnd_rd_frag_addr(dstrd, dstidx);
1116 wrq->wr.rdma.rkey = kiblnd_rd_frag_key(dstrd, dstidx);
1118 srcidx = kiblnd_rd_consume_frag(srcrd, srcidx, wrknob);
1119 dstidx = kiblnd_rd_consume_frag(dstrd, dstidx, wrknob);
1128 if (rc < 0) /* no RDMA if completing with failure */
1131 ibmsg->ibm_u.completion.ibcm_status = rc;
1132 ibmsg->ibm_u.completion.ibcm_cookie = dstcookie;
1133 kiblnd_init_tx_msg(conn->ibc_peer->ibp_ni, tx,
1134 type, sizeof (kib_completion_msg_t));
1140 kiblnd_queue_tx_locked(kib_tx_t *tx, kib_conn_t *conn)
1142 struct list_head *q;
1144 LASSERT(tx->tx_nwrq > 0); /* work items set up */
1145 LASSERT(!tx->tx_queued); /* not queued for sending already */
1146 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1149 tx->tx_deadline = jiffies +
1150 msecs_to_jiffies(*kiblnd_tunables.kib_timeout *
1153 if (tx->tx_conn == NULL) {
1154 kiblnd_conn_addref(conn);
1156 LASSERT (tx->tx_msg->ibm_type != IBLND_MSG_PUT_DONE);
1158 /* PUT_DONE first attached to conn as a PUT_REQ */
1159 LASSERT (tx->tx_conn == conn);
1160 LASSERT (tx->tx_msg->ibm_type == IBLND_MSG_PUT_DONE);
1163 switch (tx->tx_msg->ibm_type) {
1167 case IBLND_MSG_PUT_REQ:
1168 case IBLND_MSG_GET_REQ:
1169 q = &conn->ibc_tx_queue_rsrvd;
1172 case IBLND_MSG_PUT_NAK:
1173 case IBLND_MSG_PUT_ACK:
1174 case IBLND_MSG_PUT_DONE:
1175 case IBLND_MSG_GET_DONE:
1176 q = &conn->ibc_tx_queue_nocred;
1179 case IBLND_MSG_NOOP:
1180 if (IBLND_OOB_CAPABLE(conn->ibc_version))
1181 q = &conn->ibc_tx_queue_nocred;
1183 q = &conn->ibc_tx_noops;
1186 case IBLND_MSG_IMMEDIATE:
1187 q = &conn->ibc_tx_queue;
1191 list_add_tail(&tx->tx_list, q);
1195 kiblnd_queue_tx (kib_tx_t *tx, kib_conn_t *conn)
1197 spin_lock(&conn->ibc_lock);
1198 kiblnd_queue_tx_locked(tx, conn);
1199 spin_unlock(&conn->ibc_lock);
1201 kiblnd_check_sends(conn);
1204 static int kiblnd_resolve_addr(struct rdma_cm_id *cmid,
1205 struct sockaddr_in *srcaddr,
1206 struct sockaddr_in *dstaddr,
1209 unsigned short port;
1212 /* allow the port to be reused */
1213 rc = rdma_set_reuseaddr(cmid, 1);
1215 CERROR("Unable to set reuse on cmid: %d\n", rc);
1219 /* look for a free privileged port */
1220 for (port = PROT_SOCK-1; port > 0; port--) {
1221 srcaddr->sin_port = htons(port);
1222 rc = rdma_resolve_addr(cmid,
1223 (struct sockaddr *)srcaddr,
1224 (struct sockaddr *)dstaddr,
1227 CDEBUG(D_NET, "bound to port %hu\n", port);
1229 } else if (rc == -EADDRINUSE || rc == -EADDRNOTAVAIL) {
1230 CDEBUG(D_NET, "bind to port %hu failed: %d\n",
1237 CERROR("Failed to bind to a free privileged port\n");
1242 kiblnd_connect_peer (kib_peer_t *peer)
1244 struct rdma_cm_id *cmid;
1246 kib_net_t *net = peer->ibp_ni->ni_data;
1247 struct sockaddr_in srcaddr;
1248 struct sockaddr_in dstaddr;
1251 LASSERT (net != NULL);
1252 LASSERT (peer->ibp_connecting > 0);
1254 cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, peer, RDMA_PS_TCP,
1258 CERROR("Can't create CMID for %s: %ld\n",
1259 libcfs_nid2str(peer->ibp_nid), PTR_ERR(cmid));
1265 memset(&srcaddr, 0, sizeof(srcaddr));
1266 srcaddr.sin_family = AF_INET;
1267 srcaddr.sin_addr.s_addr = htonl(dev->ibd_ifip);
1269 memset(&dstaddr, 0, sizeof(dstaddr));
1270 dstaddr.sin_family = AF_INET;
1271 dstaddr.sin_port = htons(*kiblnd_tunables.kib_service);
1272 dstaddr.sin_addr.s_addr = htonl(LNET_NIDADDR(peer->ibp_nid));
1274 kiblnd_peer_addref(peer); /* cmid's ref */
1276 if (*kiblnd_tunables.kib_use_priv_port) {
1277 rc = kiblnd_resolve_addr(cmid, &srcaddr, &dstaddr,
1278 *kiblnd_tunables.kib_timeout * 1000);
1280 rc = rdma_resolve_addr(cmid,
1281 (struct sockaddr *)&srcaddr,
1282 (struct sockaddr *)&dstaddr,
1283 *kiblnd_tunables.kib_timeout * 1000);
1286 /* Can't initiate address resolution: */
1287 CERROR("Can't resolve addr for %s: %d\n",
1288 libcfs_nid2str(peer->ibp_nid), rc);
1292 LASSERT (cmid->device != NULL);
1293 CDEBUG(D_NET, "%s: connection bound to %s:%pI4h:%s\n",
1294 libcfs_nid2str(peer->ibp_nid), dev->ibd_ifname,
1295 &dev->ibd_ifip, cmid->device->name);
1300 kiblnd_peer_decref(peer); /* cmid's ref */
1301 rdma_destroy_id(cmid);
1303 kiblnd_peer_connect_failed(peer, 1, rc);
1307 kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid)
1312 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
1313 unsigned long flags;
1316 /* If I get here, I've committed to send, so I complete the tx with
1317 * failure on any problems */
1319 LASSERT (tx == NULL || tx->tx_conn == NULL); /* only set when assigned a conn */
1320 LASSERT (tx == NULL || tx->tx_nwrq > 0); /* work items have been set up */
1322 /* First time, just use a read lock since I expect to find my peer
1324 read_lock_irqsave(g_lock, flags);
1326 peer = kiblnd_find_peer_locked(nid);
1327 if (peer != NULL && !list_empty(&peer->ibp_conns)) {
1328 /* Found a peer with an established connection */
1329 conn = kiblnd_get_conn_locked(peer);
1330 kiblnd_conn_addref(conn); /* 1 ref for me... */
1332 read_unlock_irqrestore(g_lock, flags);
1335 kiblnd_queue_tx(tx, conn);
1336 kiblnd_conn_decref(conn); /* ...to here */
1340 read_unlock(g_lock);
1341 /* Re-try with a write lock */
1344 peer = kiblnd_find_peer_locked(nid);
1346 if (list_empty(&peer->ibp_conns)) {
1347 /* found a peer, but it's still connecting... */
1348 LASSERT (peer->ibp_connecting != 0 ||
1349 peer->ibp_accepting != 0);
1351 list_add_tail(&tx->tx_list,
1352 &peer->ibp_tx_queue);
1353 write_unlock_irqrestore(g_lock, flags);
1355 conn = kiblnd_get_conn_locked(peer);
1356 kiblnd_conn_addref(conn); /* 1 ref for me... */
1358 write_unlock_irqrestore(g_lock, flags);
1361 kiblnd_queue_tx(tx, conn);
1362 kiblnd_conn_decref(conn); /* ...to here */
1367 write_unlock_irqrestore(g_lock, flags);
1369 /* Allocate a peer ready to add to the peer table and retry */
1370 rc = kiblnd_create_peer(ni, &peer, nid);
1372 CERROR("Can't create peer %s\n", libcfs_nid2str(nid));
1374 tx->tx_status = -EHOSTUNREACH;
1376 kiblnd_tx_done(ni, tx);
1381 write_lock_irqsave(g_lock, flags);
1383 peer2 = kiblnd_find_peer_locked(nid);
1384 if (peer2 != NULL) {
1385 if (list_empty(&peer2->ibp_conns)) {
1386 /* found a peer, but it's still connecting... */
1387 LASSERT (peer2->ibp_connecting != 0 ||
1388 peer2->ibp_accepting != 0);
1390 list_add_tail(&tx->tx_list,
1391 &peer2->ibp_tx_queue);
1392 write_unlock_irqrestore(g_lock, flags);
1394 conn = kiblnd_get_conn_locked(peer2);
1395 kiblnd_conn_addref(conn); /* 1 ref for me... */
1397 write_unlock_irqrestore(g_lock, flags);
1400 kiblnd_queue_tx(tx, conn);
1401 kiblnd_conn_decref(conn); /* ...to here */
1404 kiblnd_peer_decref(peer);
1408 /* Brand new peer */
1409 LASSERT (peer->ibp_connecting == 0);
1410 peer->ibp_connecting = 1;
1412 /* always called with a ref on ni, which prevents ni being shutdown */
1413 LASSERT (((kib_net_t *)ni->ni_data)->ibn_shutdown == 0);
1416 list_add_tail(&tx->tx_list, &peer->ibp_tx_queue);
1418 kiblnd_peer_addref(peer);
1419 list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid));
1421 write_unlock_irqrestore(g_lock, flags);
1423 kiblnd_connect_peer(peer);
1424 kiblnd_peer_decref(peer);
1428 kiblnd_send (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg)
1430 lnet_hdr_t *hdr = &lntmsg->msg_hdr;
1431 int type = lntmsg->msg_type;
1432 lnet_process_id_t target = lntmsg->msg_target;
1433 int target_is_router = lntmsg->msg_target_is_router;
1434 int routing = lntmsg->msg_routing;
1435 unsigned int payload_niov = lntmsg->msg_niov;
1436 struct iovec *payload_iov = lntmsg->msg_iov;
1437 lnet_kiov_t *payload_kiov = lntmsg->msg_kiov;
1438 unsigned int payload_offset = lntmsg->msg_offset;
1439 unsigned int payload_nob = lntmsg->msg_len;
1441 kib_rdma_desc_t *rd;
1446 /* NB 'private' is different depending on what we're sending.... */
1448 CDEBUG(D_NET, "sending %d bytes in %d frags to %s\n",
1449 payload_nob, payload_niov, libcfs_id2str(target));
1451 LASSERT (payload_nob == 0 || payload_niov > 0);
1452 LASSERT (payload_niov <= LNET_MAX_IOV);
1454 /* Thread context */
1455 LASSERT (!in_interrupt());
1456 /* payload is either all vaddrs or all pages */
1457 LASSERT (!(payload_kiov != NULL && payload_iov != NULL));
1465 LASSERT (payload_nob == 0);
1469 if (routing || target_is_router)
1470 break; /* send IMMEDIATE */
1472 /* is the REPLY message too small for RDMA? */
1473 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[lntmsg->msg_md->md_length]);
1474 if (nob <= IBLND_MSG_SIZE)
1475 break; /* send IMMEDIATE */
1477 tx = kiblnd_get_idle_tx(ni, target.nid);
1479 CERROR("Can't allocate txd for GET to %s\n",
1480 libcfs_nid2str(target.nid));
1485 rd = &ibmsg->ibm_u.get.ibgm_rd;
1486 if ((lntmsg->msg_md->md_options & LNET_MD_KIOV) == 0)
1487 rc = kiblnd_setup_rd_iov(ni, tx, rd,
1488 lntmsg->msg_md->md_niov,
1489 lntmsg->msg_md->md_iov.iov,
1490 0, lntmsg->msg_md->md_length);
1492 rc = kiblnd_setup_rd_kiov(ni, tx, rd,
1493 lntmsg->msg_md->md_niov,
1494 lntmsg->msg_md->md_iov.kiov,
1495 0, lntmsg->msg_md->md_length);
1497 CERROR("Can't setup GET sink for %s: %d\n",
1498 libcfs_nid2str(target.nid), rc);
1499 kiblnd_tx_done(ni, tx);
1503 nob = offsetof(kib_get_msg_t, ibgm_rd.rd_frags[rd->rd_nfrags]);
1504 ibmsg->ibm_u.get.ibgm_cookie = tx->tx_cookie;
1505 ibmsg->ibm_u.get.ibgm_hdr = *hdr;
1507 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_GET_REQ, nob);
1509 tx->tx_lntmsg[1] = lnet_create_reply_msg(ni, lntmsg);
1510 if (tx->tx_lntmsg[1] == NULL) {
1511 CERROR("Can't create reply for GET -> %s\n",
1512 libcfs_nid2str(target.nid));
1513 kiblnd_tx_done(ni, tx);
1517 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg[0,1] on completion */
1518 tx->tx_waiting = 1; /* waiting for GET_DONE */
1519 kiblnd_launch_tx(ni, tx, target.nid);
1522 case LNET_MSG_REPLY:
1524 /* Is the payload small enough not to need RDMA? */
1525 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob]);
1526 if (nob <= IBLND_MSG_SIZE)
1527 break; /* send IMMEDIATE */
1529 tx = kiblnd_get_idle_tx(ni, target.nid);
1531 CERROR("Can't allocate %s txd for %s\n",
1532 type == LNET_MSG_PUT ? "PUT" : "REPLY",
1533 libcfs_nid2str(target.nid));
1537 if (payload_kiov == NULL)
1538 rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1539 payload_niov, payload_iov,
1540 payload_offset, payload_nob);
1542 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1543 payload_niov, payload_kiov,
1544 payload_offset, payload_nob);
1546 CERROR("Can't setup PUT src for %s: %d\n",
1547 libcfs_nid2str(target.nid), rc);
1548 kiblnd_tx_done(ni, tx);
1553 ibmsg->ibm_u.putreq.ibprm_hdr = *hdr;
1554 ibmsg->ibm_u.putreq.ibprm_cookie = tx->tx_cookie;
1555 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_REQ, sizeof(kib_putreq_msg_t));
1557 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1558 tx->tx_waiting = 1; /* waiting for PUT_{ACK,NAK} */
1559 kiblnd_launch_tx(ni, tx, target.nid);
1563 /* send IMMEDIATE */
1565 LASSERT (offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob])
1568 tx = kiblnd_get_idle_tx(ni, target.nid);
1570 CERROR ("Can't send %d to %s: tx descs exhausted\n",
1571 type, libcfs_nid2str(target.nid));
1576 ibmsg->ibm_u.immediate.ibim_hdr = *hdr;
1578 if (payload_kiov != NULL)
1579 lnet_copy_kiov2flat(IBLND_MSG_SIZE, ibmsg,
1580 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1581 payload_niov, payload_kiov,
1582 payload_offset, payload_nob);
1584 lnet_copy_iov2flat(IBLND_MSG_SIZE, ibmsg,
1585 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1586 payload_niov, payload_iov,
1587 payload_offset, payload_nob);
1589 nob = offsetof(kib_immediate_msg_t, ibim_payload[payload_nob]);
1590 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_IMMEDIATE, nob);
1592 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1593 kiblnd_launch_tx(ni, tx, target.nid);
1598 kiblnd_reply (lnet_ni_t *ni, kib_rx_t *rx, lnet_msg_t *lntmsg)
1600 lnet_process_id_t target = lntmsg->msg_target;
1601 unsigned int niov = lntmsg->msg_niov;
1602 struct iovec *iov = lntmsg->msg_iov;
1603 lnet_kiov_t *kiov = lntmsg->msg_kiov;
1604 unsigned int offset = lntmsg->msg_offset;
1605 unsigned int nob = lntmsg->msg_len;
1609 tx = kiblnd_get_idle_tx(ni, rx->rx_conn->ibc_peer->ibp_nid);
1611 CERROR("Can't get tx for REPLY to %s\n",
1612 libcfs_nid2str(target.nid));
1618 else if (kiov == NULL)
1619 rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1620 niov, iov, offset, nob);
1622 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1623 niov, kiov, offset, nob);
1626 CERROR("Can't setup GET src for %s: %d\n",
1627 libcfs_nid2str(target.nid), rc);
1631 rc = kiblnd_init_rdma(rx->rx_conn, tx,
1632 IBLND_MSG_GET_DONE, nob,
1633 &rx->rx_msg->ibm_u.get.ibgm_rd,
1634 rx->rx_msg->ibm_u.get.ibgm_cookie);
1636 CERROR("Can't setup rdma for GET from %s: %d\n",
1637 libcfs_nid2str(target.nid), rc);
1642 /* No RDMA: local completion may happen now! */
1643 lnet_finalize(ni, lntmsg, 0);
1645 /* RDMA: lnet_finalize(lntmsg) when it
1647 tx->tx_lntmsg[0] = lntmsg;
1650 kiblnd_queue_tx(tx, rx->rx_conn);
1654 kiblnd_tx_done(ni, tx);
1656 lnet_finalize(ni, lntmsg, -EIO);
1660 kiblnd_recv (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed,
1661 unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov,
1662 unsigned int offset, unsigned int mlen, unsigned int rlen)
1664 kib_rx_t *rx = private;
1665 kib_msg_t *rxmsg = rx->rx_msg;
1666 kib_conn_t *conn = rx->rx_conn;
1669 int post_credit = IBLND_POSTRX_PEER_CREDIT;
1672 LASSERT (mlen <= rlen);
1673 LASSERT (!in_interrupt());
1674 /* Either all pages or all vaddrs */
1675 LASSERT (!(kiov != NULL && iov != NULL));
1677 switch (rxmsg->ibm_type) {
1681 case IBLND_MSG_IMMEDIATE:
1682 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[rlen]);
1683 if (nob > rx->rx_nob) {
1684 CERROR ("Immediate message from %s too big: %d(%d)\n",
1685 libcfs_nid2str(rxmsg->ibm_u.immediate.ibim_hdr.src_nid),
1692 lnet_copy_flat2kiov(niov, kiov, offset,
1693 IBLND_MSG_SIZE, rxmsg,
1694 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1697 lnet_copy_flat2iov(niov, iov, offset,
1698 IBLND_MSG_SIZE, rxmsg,
1699 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1701 lnet_finalize (ni, lntmsg, 0);
1704 case IBLND_MSG_PUT_REQ: {
1706 kib_rdma_desc_t *rd;
1709 lnet_finalize(ni, lntmsg, 0);
1710 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK, 0,
1711 rxmsg->ibm_u.putreq.ibprm_cookie);
1715 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
1717 CERROR("Can't allocate tx for %s\n",
1718 libcfs_nid2str(conn->ibc_peer->ibp_nid));
1719 /* Not replying will break the connection */
1725 rd = &txmsg->ibm_u.putack.ibpam_rd;
1727 rc = kiblnd_setup_rd_iov(ni, tx, rd,
1728 niov, iov, offset, mlen);
1730 rc = kiblnd_setup_rd_kiov(ni, tx, rd,
1731 niov, kiov, offset, mlen);
1733 CERROR("Can't setup PUT sink for %s: %d\n",
1734 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
1735 kiblnd_tx_done(ni, tx);
1736 /* tell peer it's over */
1737 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK, rc,
1738 rxmsg->ibm_u.putreq.ibprm_cookie);
1742 nob = offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[rd->rd_nfrags]);
1743 txmsg->ibm_u.putack.ibpam_src_cookie = rxmsg->ibm_u.putreq.ibprm_cookie;
1744 txmsg->ibm_u.putack.ibpam_dst_cookie = tx->tx_cookie;
1746 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_ACK, nob);
1748 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1749 tx->tx_waiting = 1; /* waiting for PUT_DONE */
1750 kiblnd_queue_tx(tx, conn);
1752 /* reposted buffer reserved for PUT_DONE */
1753 post_credit = IBLND_POSTRX_NO_CREDIT;
1757 case IBLND_MSG_GET_REQ:
1758 if (lntmsg != NULL) {
1759 /* Optimized GET; RDMA lntmsg's payload */
1760 kiblnd_reply(ni, rx, lntmsg);
1762 /* GET didn't match anything */
1763 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_GET_DONE,
1765 rxmsg->ibm_u.get.ibgm_cookie);
1770 kiblnd_post_rx(rx, post_credit);
1775 kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name)
1777 struct task_struct *task = kthread_run(fn, arg, name);
1780 return PTR_ERR(task);
1782 atomic_inc(&kiblnd_data.kib_nthreads);
1787 kiblnd_thread_fini (void)
1789 atomic_dec (&kiblnd_data.kib_nthreads);
1793 kiblnd_peer_alive (kib_peer_t *peer)
1795 /* This is racy, but everyone's only writing cfs_time_current() */
1796 peer->ibp_last_alive = cfs_time_current();
1801 kiblnd_peer_notify (kib_peer_t *peer)
1804 cfs_time_t last_alive = 0;
1805 unsigned long flags;
1807 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1809 if (list_empty(&peer->ibp_conns) &&
1810 peer->ibp_accepting == 0 &&
1811 peer->ibp_connecting == 0 &&
1812 peer->ibp_error != 0) {
1813 error = peer->ibp_error;
1814 peer->ibp_error = 0;
1816 last_alive = peer->ibp_last_alive;
1819 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1822 lnet_notify(peer->ibp_ni,
1823 peer->ibp_nid, 0, last_alive);
1827 kiblnd_close_conn_locked (kib_conn_t *conn, int error)
1829 /* This just does the immediate housekeeping. 'error' is zero for a
1830 * normal shutdown which can happen only after the connection has been
1831 * established. If the connection is established, schedule the
1832 * connection to be finished off by the connd. Otherwise the connd is
1833 * already dealing with it (either to set it up or tear it down).
1834 * Caller holds kib_global_lock exclusively in irq context */
1835 kib_peer_t *peer = conn->ibc_peer;
1837 unsigned long flags;
1839 LASSERT (error != 0 || conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1841 if (error != 0 && conn->ibc_comms_error == 0)
1842 conn->ibc_comms_error = error;
1844 if (conn->ibc_state != IBLND_CONN_ESTABLISHED)
1845 return; /* already being handled */
1848 list_empty(&conn->ibc_tx_noops) &&
1849 list_empty(&conn->ibc_tx_queue) &&
1850 list_empty(&conn->ibc_tx_queue_rsrvd) &&
1851 list_empty(&conn->ibc_tx_queue_nocred) &&
1852 list_empty(&conn->ibc_active_txs)) {
1853 CDEBUG(D_NET, "closing conn to %s\n",
1854 libcfs_nid2str(peer->ibp_nid));
1856 CNETERR("Closing conn to %s: error %d%s%s%s%s%s\n",
1857 libcfs_nid2str(peer->ibp_nid), error,
1858 list_empty(&conn->ibc_tx_queue) ? "" : "(sending)",
1859 list_empty(&conn->ibc_tx_noops) ? "" : "(sending_noops)",
1860 list_empty(&conn->ibc_tx_queue_rsrvd) ?
1861 "" : "(sending_rsrvd)",
1862 list_empty(&conn->ibc_tx_queue_nocred) ?
1863 "" : "(sending_nocred)",
1864 list_empty(&conn->ibc_active_txs) ? "" : "(waiting)");
1867 dev = ((kib_net_t *)peer->ibp_ni->ni_data)->ibn_dev;
1868 list_del(&conn->ibc_list);
1869 /* connd (see below) takes over ibc_list's ref */
1871 if (list_empty(&peer->ibp_conns) && /* no more conns */
1872 kiblnd_peer_active(peer)) { /* still in peer table */
1873 kiblnd_unlink_peer_locked(peer);
1875 /* set/clear error on last conn */
1876 peer->ibp_error = conn->ibc_comms_error;
1879 kiblnd_set_conn_state(conn, IBLND_CONN_CLOSING);
1882 kiblnd_dev_can_failover(dev)) {
1883 list_add_tail(&dev->ibd_fail_list,
1884 &kiblnd_data.kib_failed_devs);
1885 wake_up(&kiblnd_data.kib_failover_waitq);
1888 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
1890 list_add_tail(&conn->ibc_list, &kiblnd_data.kib_connd_conns);
1891 wake_up(&kiblnd_data.kib_connd_waitq);
1893 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
1897 kiblnd_close_conn(kib_conn_t *conn, int error)
1899 unsigned long flags;
1901 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1903 kiblnd_close_conn_locked(conn, error);
1905 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1909 kiblnd_handle_early_rxs(kib_conn_t *conn)
1911 unsigned long flags;
1914 LASSERT(!in_interrupt());
1915 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1917 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1918 while (!list_empty(&conn->ibc_early_rxs)) {
1919 rx = list_entry(conn->ibc_early_rxs.next,
1921 list_del(&rx->rx_list);
1922 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1924 kiblnd_handle_rx(rx);
1926 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1928 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1932 kiblnd_abort_txs(kib_conn_t *conn, struct list_head *txs)
1934 struct list_head zombies = LIST_HEAD_INIT(zombies);
1935 struct list_head *tmp;
1936 struct list_head *nxt;
1939 spin_lock(&conn->ibc_lock);
1941 list_for_each_safe(tmp, nxt, txs) {
1942 tx = list_entry(tmp, kib_tx_t, tx_list);
1944 if (txs == &conn->ibc_active_txs) {
1945 LASSERT(!tx->tx_queued);
1946 LASSERT(tx->tx_waiting ||
1947 tx->tx_sending != 0);
1949 LASSERT(tx->tx_queued);
1952 tx->tx_status = -ECONNABORTED;
1955 if (tx->tx_sending == 0) {
1957 list_del(&tx->tx_list);
1958 list_add(&tx->tx_list, &zombies);
1962 spin_unlock(&conn->ibc_lock);
1964 kiblnd_txlist_done(conn->ibc_peer->ibp_ni, &zombies, -ECONNABORTED);
1968 kiblnd_finalise_conn (kib_conn_t *conn)
1970 LASSERT (!in_interrupt());
1971 LASSERT (conn->ibc_state > IBLND_CONN_INIT);
1973 kiblnd_set_conn_state(conn, IBLND_CONN_DISCONNECTED);
1975 /* abort_receives moves QP state to IB_QPS_ERR. This is only required
1976 * for connections that didn't get as far as being connected, because
1977 * rdma_disconnect() does this for free. */
1978 kiblnd_abort_receives(conn);
1980 /* Complete all tx descs not waiting for sends to complete.
1981 * NB we should be safe from RDMA now that the QP has changed state */
1983 kiblnd_abort_txs(conn, &conn->ibc_tx_noops);
1984 kiblnd_abort_txs(conn, &conn->ibc_tx_queue);
1985 kiblnd_abort_txs(conn, &conn->ibc_tx_queue_rsrvd);
1986 kiblnd_abort_txs(conn, &conn->ibc_tx_queue_nocred);
1987 kiblnd_abort_txs(conn, &conn->ibc_active_txs);
1989 kiblnd_handle_early_rxs(conn);
1993 kiblnd_peer_connect_failed(kib_peer_t *peer, int active, int error)
1995 struct list_head zombies = LIST_HEAD_INIT(zombies);
1996 unsigned long flags;
1998 LASSERT (error != 0);
1999 LASSERT (!in_interrupt());
2001 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2004 LASSERT (peer->ibp_connecting > 0);
2005 peer->ibp_connecting--;
2007 LASSERT (peer->ibp_accepting > 0);
2008 peer->ibp_accepting--;
2011 if (peer->ibp_connecting != 0 ||
2012 peer->ibp_accepting != 0) {
2013 /* another connection attempt under way... */
2014 write_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2019 if (list_empty(&peer->ibp_conns)) {
2020 /* Take peer's blocked transmits to complete with error */
2021 list_add(&zombies, &peer->ibp_tx_queue);
2022 list_del_init(&peer->ibp_tx_queue);
2024 if (kiblnd_peer_active(peer))
2025 kiblnd_unlink_peer_locked(peer);
2027 peer->ibp_error = error;
2029 /* Can't have blocked transmits if there are connections */
2030 LASSERT(list_empty(&peer->ibp_tx_queue));
2033 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2035 kiblnd_peer_notify(peer);
2037 if (list_empty(&zombies))
2040 CNETERR("Deleting messages for %s: connection failed\n",
2041 libcfs_nid2str(peer->ibp_nid));
2043 kiblnd_txlist_done(peer->ibp_ni, &zombies, -EHOSTUNREACH);
2047 kiblnd_connreq_done(kib_conn_t *conn, int status)
2049 kib_peer_t *peer = conn->ibc_peer;
2051 struct list_head txs;
2052 unsigned long flags;
2055 active = (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2057 CDEBUG(D_NET,"%s: active(%d), version(%x), status(%d)\n",
2058 libcfs_nid2str(peer->ibp_nid), active,
2059 conn->ibc_version, status);
2061 LASSERT (!in_interrupt());
2062 LASSERT ((conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT &&
2063 peer->ibp_connecting > 0) ||
2064 (conn->ibc_state == IBLND_CONN_PASSIVE_WAIT &&
2065 peer->ibp_accepting > 0));
2067 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
2068 conn->ibc_connvars = NULL;
2071 /* failed to establish connection */
2072 kiblnd_peer_connect_failed(peer, active, status);
2073 kiblnd_finalise_conn(conn);
2077 /* connection established */
2078 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2080 conn->ibc_last_send = jiffies;
2081 kiblnd_set_conn_state(conn, IBLND_CONN_ESTABLISHED);
2082 kiblnd_peer_alive(peer);
2084 /* Add conn to peer's list and nuke any dangling conns from a different
2085 * peer instance... */
2086 kiblnd_conn_addref(conn); /* +1 ref for ibc_list */
2087 list_add(&conn->ibc_list, &peer->ibp_conns);
2089 peer->ibp_connecting--;
2091 peer->ibp_accepting--;
2093 if (peer->ibp_version == 0) {
2094 peer->ibp_version = conn->ibc_version;
2095 peer->ibp_incarnation = conn->ibc_incarnation;
2098 if (peer->ibp_version != conn->ibc_version ||
2099 peer->ibp_incarnation != conn->ibc_incarnation) {
2100 kiblnd_close_stale_conns_locked(peer, conn->ibc_version,
2101 conn->ibc_incarnation);
2102 peer->ibp_version = conn->ibc_version;
2103 peer->ibp_incarnation = conn->ibc_incarnation;
2106 /* grab pending txs while I have the lock */
2107 list_add(&txs, &peer->ibp_tx_queue);
2108 list_del_init(&peer->ibp_tx_queue);
2110 if (!kiblnd_peer_active(peer) || /* peer has been deleted */
2111 conn->ibc_comms_error != 0) { /* error has happened already */
2112 lnet_ni_t *ni = peer->ibp_ni;
2114 /* start to shut down connection */
2115 kiblnd_close_conn_locked(conn, -ECONNABORTED);
2116 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2118 kiblnd_txlist_done(ni, &txs, -ECONNABORTED);
2123 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2125 /* Schedule blocked txs */
2126 spin_lock(&conn->ibc_lock);
2127 while (!list_empty(&txs)) {
2128 tx = list_entry(txs.next, kib_tx_t, tx_list);
2129 list_del(&tx->tx_list);
2131 kiblnd_queue_tx_locked(tx, conn);
2133 spin_unlock(&conn->ibc_lock);
2135 kiblnd_check_sends(conn);
2137 /* schedule blocked rxs */
2138 kiblnd_handle_early_rxs(conn);
2142 kiblnd_reject(struct rdma_cm_id *cmid, kib_rej_t *rej)
2146 rc = rdma_reject(cmid, rej, sizeof(*rej));
2149 CWARN("Error %d sending reject\n", rc);
2153 kiblnd_passive_connect(struct rdma_cm_id *cmid, void *priv, int priv_nob)
2155 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2156 kib_msg_t *reqmsg = priv;
2162 lnet_ni_t *ni = NULL;
2163 kib_net_t *net = NULL;
2165 struct rdma_conn_param cp;
2167 int version = IBLND_MSG_VERSION;
2168 unsigned long flags;
2170 struct sockaddr_in *peer_addr;
2171 LASSERT (!in_interrupt());
2173 /* cmid inherits 'context' from the corresponding listener id */
2174 ibdev = (kib_dev_t *)cmid->context;
2175 LASSERT (ibdev != NULL);
2177 memset(&rej, 0, sizeof(rej));
2178 rej.ibr_magic = IBLND_MSG_MAGIC;
2179 rej.ibr_why = IBLND_REJECT_FATAL;
2180 rej.ibr_cp.ibcp_max_msg_size = IBLND_MSG_SIZE;
2182 peer_addr = (struct sockaddr_in *)&(cmid->route.addr.dst_addr);
2183 if (*kiblnd_tunables.kib_require_priv_port &&
2184 ntohs(peer_addr->sin_port) >= PROT_SOCK) {
2185 __u32 ip = ntohl(peer_addr->sin_addr.s_addr);
2186 CERROR("Peer's port (%pI4h:%hu) is not privileged\n",
2187 &ip, ntohs(peer_addr->sin_port));
2191 if (priv_nob < offsetof(kib_msg_t, ibm_type)) {
2192 CERROR("Short connection request\n");
2196 /* Future protocol version compatibility support! If the
2197 * o2iblnd-specific protocol changes, or when LNET unifies
2198 * protocols over all LNDs, the initial connection will
2199 * negotiate a protocol version. I trap this here to avoid
2200 * console errors; the reject tells the peer which protocol I
2202 if (reqmsg->ibm_magic == LNET_PROTO_MAGIC ||
2203 reqmsg->ibm_magic == __swab32(LNET_PROTO_MAGIC))
2205 if (reqmsg->ibm_magic == IBLND_MSG_MAGIC &&
2206 reqmsg->ibm_version != IBLND_MSG_VERSION &&
2207 reqmsg->ibm_version != IBLND_MSG_VERSION_1)
2209 if (reqmsg->ibm_magic == __swab32(IBLND_MSG_MAGIC) &&
2210 reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION) &&
2211 reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION_1))
2214 rc = kiblnd_unpack_msg(reqmsg, priv_nob);
2216 CERROR("Can't parse connection request: %d\n", rc);
2220 nid = reqmsg->ibm_srcnid;
2221 ni = lnet_net2ni(LNET_NIDNET(reqmsg->ibm_dstnid));
2224 net = (kib_net_t *)ni->ni_data;
2225 rej.ibr_incarnation = net->ibn_incarnation;
2228 if (ni == NULL || /* no matching net */
2229 ni->ni_nid != reqmsg->ibm_dstnid || /* right NET, wrong NID! */
2230 net->ibn_dev != ibdev) { /* wrong device */
2231 CERROR("Can't accept conn from %s on %s (%s:%d:%pI4h): "
2232 "bad dst nid %s\n", libcfs_nid2str(nid),
2233 ni == NULL ? "NA" : libcfs_nid2str(ni->ni_nid),
2234 ibdev->ibd_ifname, ibdev->ibd_nnets,
2236 libcfs_nid2str(reqmsg->ibm_dstnid));
2241 /* check time stamp as soon as possible */
2242 if (reqmsg->ibm_dststamp != 0 &&
2243 reqmsg->ibm_dststamp != net->ibn_incarnation) {
2244 CWARN("Stale connection request\n");
2245 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2249 /* I can accept peer's version */
2250 version = reqmsg->ibm_version;
2252 if (reqmsg->ibm_type != IBLND_MSG_CONNREQ) {
2253 CERROR("Unexpected connreq msg type: %x from %s\n",
2254 reqmsg->ibm_type, libcfs_nid2str(nid));
2258 if (reqmsg->ibm_u.connparams.ibcp_queue_depth >
2259 IBLND_MSG_QUEUE_SIZE(version)) {
2260 CERROR("Can't accept conn from %s, queue depth too large: "
2261 " %d (<=%d wanted)\n",
2262 libcfs_nid2str(nid),
2263 reqmsg->ibm_u.connparams.ibcp_queue_depth,
2264 IBLND_MSG_QUEUE_SIZE(version));
2266 if (version == IBLND_MSG_VERSION)
2267 rej.ibr_why = IBLND_REJECT_MSG_QUEUE_SIZE;
2272 if (reqmsg->ibm_u.connparams.ibcp_max_frags >
2273 IBLND_RDMA_FRAGS(version)) {
2274 CWARN("Can't accept conn from %s (version %x): "
2275 "max_frags %d too large (%d wanted)\n",
2276 libcfs_nid2str(nid), version,
2277 reqmsg->ibm_u.connparams.ibcp_max_frags,
2278 IBLND_RDMA_FRAGS(version));
2280 if (version >= IBLND_MSG_VERSION)
2281 rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2284 } else if (reqmsg->ibm_u.connparams.ibcp_max_frags <
2285 IBLND_RDMA_FRAGS(version) && net->ibn_fmr_ps == NULL) {
2286 CWARN("Can't accept conn from %s (version %x): "
2287 "max_frags %d incompatible without FMR pool "
2289 libcfs_nid2str(nid), version,
2290 reqmsg->ibm_u.connparams.ibcp_max_frags,
2291 IBLND_RDMA_FRAGS(version));
2293 if (version >= IBLND_MSG_VERSION)
2294 rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2299 if (reqmsg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2300 CERROR("Can't accept %s: message size %d too big (%d max)\n",
2301 libcfs_nid2str(nid),
2302 reqmsg->ibm_u.connparams.ibcp_max_msg_size,
2307 /* assume 'nid' is a new peer; create */
2308 rc = kiblnd_create_peer(ni, &peer, nid);
2310 CERROR("Can't create peer for %s\n", libcfs_nid2str(nid));
2311 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2315 /* We have validated the peer's parameters so use those */
2316 peer->ibp_max_frags = reqmsg->ibm_u.connparams.ibcp_max_frags;
2317 peer->ibp_queue_depth = reqmsg->ibm_u.connparams.ibcp_queue_depth;
2319 write_lock_irqsave(g_lock, flags);
2321 peer2 = kiblnd_find_peer_locked(nid);
2322 if (peer2 != NULL) {
2323 if (peer2->ibp_version == 0) {
2324 peer2->ibp_version = version;
2325 peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2328 /* not the guy I've talked with */
2329 if (peer2->ibp_incarnation != reqmsg->ibm_srcstamp ||
2330 peer2->ibp_version != version) {
2331 kiblnd_close_peer_conns_locked(peer2, -ESTALE);
2332 write_unlock_irqrestore(g_lock, flags);
2334 CWARN("Conn stale %s [old ver: %x, new ver: %x]\n",
2335 libcfs_nid2str(nid), peer2->ibp_version, version);
2337 kiblnd_peer_decref(peer);
2338 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2342 /* tie-break connection race in favour of the higher NID */
2343 if (peer2->ibp_connecting != 0 &&
2345 write_unlock_irqrestore(g_lock, flags);
2347 CWARN("Conn race %s\n", libcfs_nid2str(peer2->ibp_nid));
2349 kiblnd_peer_decref(peer);
2350 rej.ibr_why = IBLND_REJECT_CONN_RACE;
2354 peer2->ibp_accepting++;
2355 kiblnd_peer_addref(peer2);
2357 /* Race with kiblnd_launch_tx (active connect) to create peer
2358 * so copy validated parameters since we now know what the
2359 * peer's limits are */
2360 peer2->ibp_max_frags = peer->ibp_max_frags;
2361 peer2->ibp_queue_depth = peer->ibp_queue_depth;
2363 write_unlock_irqrestore(g_lock, flags);
2364 kiblnd_peer_decref(peer);
2367 /* Brand new peer */
2368 LASSERT (peer->ibp_accepting == 0);
2369 LASSERT (peer->ibp_version == 0 &&
2370 peer->ibp_incarnation == 0);
2372 peer->ibp_accepting = 1;
2373 peer->ibp_version = version;
2374 peer->ibp_incarnation = reqmsg->ibm_srcstamp;
2376 /* I have a ref on ni that prevents it being shutdown */
2377 LASSERT (net->ibn_shutdown == 0);
2379 kiblnd_peer_addref(peer);
2380 list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid));
2382 write_unlock_irqrestore(g_lock, flags);
2385 conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_PASSIVE_WAIT, version);
2387 kiblnd_peer_connect_failed(peer, 0, -ENOMEM);
2388 kiblnd_peer_decref(peer);
2389 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2393 /* conn now "owns" cmid, so I return success from here on to ensure the
2394 * CM callback doesn't destroy cmid. */
2395 conn->ibc_incarnation = reqmsg->ibm_srcstamp;
2396 conn->ibc_credits = conn->ibc_queue_depth;
2397 conn->ibc_reserved_credits = conn->ibc_queue_depth;
2398 LASSERT(conn->ibc_credits + conn->ibc_reserved_credits +
2399 IBLND_OOB_MSGS(version) <= IBLND_RX_MSGS(conn));
2401 ackmsg = &conn->ibc_connvars->cv_msg;
2402 memset(ackmsg, 0, sizeof(*ackmsg));
2404 kiblnd_init_msg(ackmsg, IBLND_MSG_CONNACK,
2405 sizeof(ackmsg->ibm_u.connparams));
2406 ackmsg->ibm_u.connparams.ibcp_queue_depth = conn->ibc_queue_depth;
2407 ackmsg->ibm_u.connparams.ibcp_max_frags = conn->ibc_max_frags;
2408 ackmsg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2410 kiblnd_pack_msg(ni, ackmsg, version, 0, nid, reqmsg->ibm_srcstamp);
2412 memset(&cp, 0, sizeof(cp));
2413 cp.private_data = ackmsg;
2414 cp.private_data_len = ackmsg->ibm_nob;
2415 cp.responder_resources = 0; /* No atomic ops or RDMA reads */
2416 cp.initiator_depth = 0;
2417 cp.flow_control = 1;
2418 cp.retry_count = *kiblnd_tunables.kib_retry_count;
2419 cp.rnr_retry_count = *kiblnd_tunables.kib_rnr_retry_count;
2421 CDEBUG(D_NET, "Accept %s\n", libcfs_nid2str(nid));
2423 rc = rdma_accept(cmid, &cp);
2425 CERROR("Can't accept %s: %d\n", libcfs_nid2str(nid), rc);
2426 rej.ibr_version = version;
2427 rej.ibr_why = IBLND_REJECT_FATAL;
2429 kiblnd_reject(cmid, &rej);
2430 kiblnd_connreq_done(conn, rc);
2431 kiblnd_conn_decref(conn);
2441 rej.ibr_version = version;
2442 rej.ibr_cp.ibcp_queue_depth = IBLND_MSG_QUEUE_SIZE(version);
2443 rej.ibr_cp.ibcp_max_frags = IBLND_RDMA_FRAGS(version);
2444 kiblnd_reject(cmid, &rej);
2446 return -ECONNREFUSED;
2450 kiblnd_reconnect (kib_conn_t *conn, int version,
2451 __u64 incarnation, int why, kib_connparams_t *cp)
2453 kib_peer_t *peer = conn->ibc_peer;
2456 unsigned long flags;
2458 LASSERT (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2459 LASSERT (peer->ibp_connecting > 0); /* 'conn' at least */
2461 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2463 /* retry connection if it's still needed and no other connection
2464 * attempts (active or passive) are in progress
2465 * NB: reconnect is still needed even when ibp_tx_queue is
2466 * empty if ibp_version != version because reconnect may be
2467 * initiated by kiblnd_query() */
2468 if ((!list_empty(&peer->ibp_tx_queue) ||
2469 peer->ibp_version != version) &&
2470 peer->ibp_connecting == 1 &&
2471 peer->ibp_accepting == 0) {
2472 if (why == IBLND_REJECT_CONN_RACE) {
2473 /* don't reconnect immediately, intensive reconnecting
2474 * may consume a lot of memory. kiblnd_destroy_conn
2475 * will reconnect after releasing all resources of
2476 * this connection */
2477 conn->ibc_conn_race = 1;
2481 peer->ibp_connecting++;
2482 peer->ibp_version = version;
2483 peer->ibp_incarnation = incarnation;
2486 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2496 case IBLND_REJECT_RDMA_FRAGS:
2500 if (conn->ibc_max_frags <= cp->ibcp_max_frags) {
2501 CNETERR("Unsupported max frags, peer supports %d\n",
2502 cp->ibcp_max_frags);
2504 } else if (*kiblnd_tunables.kib_map_on_demand == 0) {
2505 CNETERR("map_on_demand must be enabled to support "
2506 "map_on_demand peers\n");
2510 peer->ibp_max_frags = cp->ibcp_max_frags;
2511 reason = "rdma fragments";
2514 case IBLND_REJECT_MSG_QUEUE_SIZE:
2518 if (conn->ibc_queue_depth <= cp->ibcp_queue_depth) {
2519 CNETERR("Unsupported queue depth, peer supports %d\n",
2520 cp->ibcp_queue_depth);
2524 peer->ibp_queue_depth = cp->ibcp_queue_depth;
2525 reason = "queue depth";
2528 case IBLND_REJECT_CONN_STALE:
2532 case IBLND_REJECT_CONN_UNCOMPAT:
2533 reason = "version negotiation";
2537 CNETERR("%s: retrying (%s), %x, %x, "
2538 "queue_depth: %d, max_frags: %d, msg_size: %d\n",
2539 libcfs_nid2str(peer->ibp_nid),
2540 reason, IBLND_MSG_VERSION, version,
2541 conn->ibc_queue_depth, conn->ibc_max_frags,
2542 cp != NULL ? cp->ibcp_max_msg_size : IBLND_MSG_SIZE);
2544 kiblnd_connect_peer(peer);
2548 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2549 peer->ibp_connecting--;
2550 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2556 kiblnd_rejected (kib_conn_t *conn, int reason, void *priv, int priv_nob)
2558 kib_peer_t *peer = conn->ibc_peer;
2560 LASSERT (!in_interrupt());
2561 LASSERT (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2564 case IB_CM_REJ_STALE_CONN:
2565 kiblnd_reconnect(conn, IBLND_MSG_VERSION, 0,
2566 IBLND_REJECT_CONN_STALE, NULL);
2569 case IB_CM_REJ_INVALID_SERVICE_ID:
2570 CNETERR("%s rejected: no listener at %d\n",
2571 libcfs_nid2str(peer->ibp_nid),
2572 *kiblnd_tunables.kib_service);
2575 case IB_CM_REJ_CONSUMER_DEFINED:
2576 if (priv_nob >= offsetof(kib_rej_t, ibr_padding)) {
2577 kib_rej_t *rej = priv;
2578 kib_connparams_t *cp = NULL;
2580 __u64 incarnation = -1;
2582 /* NB. default incarnation is -1 because:
2583 * a) V1 will ignore dst incarnation in connreq.
2584 * b) V2 will provide incarnation while rejecting me,
2585 * -1 will be overwrote.
2587 * if I try to connect to a V1 peer with V2 protocol,
2588 * it rejected me then upgrade to V2, I have no idea
2589 * about the upgrading and try to reconnect with V1,
2590 * in this case upgraded V2 can find out I'm trying to
2591 * talk to the old guy and reject me(incarnation is -1).
2594 if (rej->ibr_magic == __swab32(IBLND_MSG_MAGIC) ||
2595 rej->ibr_magic == __swab32(LNET_PROTO_MAGIC)) {
2596 __swab32s(&rej->ibr_magic);
2597 __swab16s(&rej->ibr_version);
2601 if (priv_nob >= sizeof(kib_rej_t) &&
2602 rej->ibr_version > IBLND_MSG_VERSION_1) {
2603 /* priv_nob is always 148 in current version
2604 * of OFED, so we still need to check version.
2605 * (define of IB_CM_REJ_PRIVATE_DATA_SIZE) */
2609 __swab64s(&rej->ibr_incarnation);
2610 __swab16s(&cp->ibcp_queue_depth);
2611 __swab16s(&cp->ibcp_max_frags);
2612 __swab32s(&cp->ibcp_max_msg_size);
2615 incarnation = rej->ibr_incarnation;
2618 if (rej->ibr_magic != IBLND_MSG_MAGIC &&
2619 rej->ibr_magic != LNET_PROTO_MAGIC) {
2620 CERROR("%s rejected: consumer defined fatal error\n",
2621 libcfs_nid2str(peer->ibp_nid));
2625 if (rej->ibr_version != IBLND_MSG_VERSION &&
2626 rej->ibr_version != IBLND_MSG_VERSION_1) {
2627 CERROR("%s rejected: o2iblnd version %x error\n",
2628 libcfs_nid2str(peer->ibp_nid),
2633 if (rej->ibr_why == IBLND_REJECT_FATAL &&
2634 rej->ibr_version == IBLND_MSG_VERSION_1) {
2635 CDEBUG(D_NET, "rejected by old version peer %s: %x\n",
2636 libcfs_nid2str(peer->ibp_nid), rej->ibr_version);
2638 if (conn->ibc_version != IBLND_MSG_VERSION_1)
2639 rej->ibr_why = IBLND_REJECT_CONN_UNCOMPAT;
2642 switch (rej->ibr_why) {
2643 case IBLND_REJECT_CONN_RACE:
2644 case IBLND_REJECT_CONN_STALE:
2645 case IBLND_REJECT_CONN_UNCOMPAT:
2646 case IBLND_REJECT_MSG_QUEUE_SIZE:
2647 case IBLND_REJECT_RDMA_FRAGS:
2648 kiblnd_reconnect(conn, rej->ibr_version,
2649 incarnation, rej->ibr_why, cp);
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 conn->ibc_queue_depth) {
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 conn->ibc_queue_depth);
2724 if (msg->ibm_u.connparams.ibcp_max_frags >
2725 conn->ibc_max_frags) {
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 conn->ibc_max_frags);
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;
2760 conn->ibc_credits = msg->ibm_u.connparams.ibcp_queue_depth;
2761 conn->ibc_reserved_credits = msg->ibm_u.connparams.ibcp_queue_depth;
2762 conn->ibc_queue_depth = msg->ibm_u.connparams.ibcp_queue_depth;
2763 conn->ibc_max_frags = msg->ibm_u.connparams.ibcp_max_frags;
2764 LASSERT(conn->ibc_credits + conn->ibc_reserved_credits +
2765 IBLND_OOB_MSGS(ver) <= IBLND_RX_MSGS(conn));
2767 kiblnd_connreq_done(conn, 0);
2771 /* NB My QP has already established itself, so I handle anything going
2772 * wrong here by setting ibc_comms_error.
2773 * kiblnd_connreq_done(0) moves the conn state to ESTABLISHED, but then
2774 * immediately tears it down. */
2777 conn->ibc_comms_error = rc;
2778 kiblnd_connreq_done(conn, 0);
2782 kiblnd_active_connect (struct rdma_cm_id *cmid)
2784 kib_peer_t *peer = (kib_peer_t *)cmid->context;
2787 struct rdma_conn_param cp;
2790 unsigned long flags;
2793 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2795 incarnation = peer->ibp_incarnation;
2796 version = (peer->ibp_version == 0) ? IBLND_MSG_VERSION :
2799 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2801 conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_ACTIVE_CONNECT,
2804 kiblnd_peer_connect_failed(peer, 1, -ENOMEM);
2805 kiblnd_peer_decref(peer); /* lose cmid's ref */
2809 /* conn "owns" cmid now, so I return success from here on to ensure the
2810 * CM callback doesn't destroy cmid. conn also takes over cmid's ref
2813 msg = &conn->ibc_connvars->cv_msg;
2815 memset(msg, 0, sizeof(*msg));
2816 kiblnd_init_msg(msg, IBLND_MSG_CONNREQ, sizeof(msg->ibm_u.connparams));
2817 msg->ibm_u.connparams.ibcp_queue_depth = conn->ibc_queue_depth;
2818 msg->ibm_u.connparams.ibcp_max_frags = conn->ibc_max_frags;
2819 msg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2821 kiblnd_pack_msg(peer->ibp_ni, msg, version,
2822 0, peer->ibp_nid, incarnation);
2824 memset(&cp, 0, sizeof(cp));
2825 cp.private_data = msg;
2826 cp.private_data_len = msg->ibm_nob;
2827 cp.responder_resources = 0; /* No atomic ops or RDMA reads */
2828 cp.initiator_depth = 0;
2829 cp.flow_control = 1;
2830 cp.retry_count = *kiblnd_tunables.kib_retry_count;
2831 cp.rnr_retry_count = *kiblnd_tunables.kib_rnr_retry_count;
2833 LASSERT(cmid->context == (void *)conn);
2834 LASSERT(conn->ibc_cmid == cmid);
2836 rc = rdma_connect(cmid, &cp);
2838 CERROR("Can't connect to %s: %d\n",
2839 libcfs_nid2str(peer->ibp_nid), rc);
2840 kiblnd_connreq_done(conn, rc);
2841 kiblnd_conn_decref(conn);
2848 kiblnd_cm_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2854 switch (event->event) {
2856 CERROR("Unexpected event: %d, status: %d\n",
2857 event->event, event->status);
2860 case RDMA_CM_EVENT_CONNECT_REQUEST:
2861 /* destroy cmid on failure */
2862 rc = kiblnd_passive_connect(cmid,
2863 (void *)KIBLND_CONN_PARAM(event),
2864 KIBLND_CONN_PARAM_LEN(event));
2865 CDEBUG(D_NET, "connreq: %d\n", rc);
2868 case RDMA_CM_EVENT_ADDR_ERROR:
2869 peer = (kib_peer_t *)cmid->context;
2870 CNETERR("%s: ADDR 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_ADDR_RESOLVED:
2877 peer = (kib_peer_t *)cmid->context;
2879 CDEBUG(D_NET,"%s Addr resolved: %d\n",
2880 libcfs_nid2str(peer->ibp_nid), event->status);
2882 if (event->status != 0) {
2883 CNETERR("Can't resolve address for %s: %d\n",
2884 libcfs_nid2str(peer->ibp_nid), event->status);
2887 rc = rdma_resolve_route(
2888 cmid, *kiblnd_tunables.kib_timeout * 1000);
2891 /* Can't initiate route resolution */
2892 CERROR("Can't resolve route for %s: %d\n",
2893 libcfs_nid2str(peer->ibp_nid), rc);
2895 kiblnd_peer_connect_failed(peer, 1, rc);
2896 kiblnd_peer_decref(peer);
2897 return rc; /* rc != 0 destroys cmid */
2899 case RDMA_CM_EVENT_ROUTE_ERROR:
2900 peer = (kib_peer_t *)cmid->context;
2901 CNETERR("%s: ROUTE ERROR %d\n",
2902 libcfs_nid2str(peer->ibp_nid), event->status);
2903 kiblnd_peer_connect_failed(peer, 1, -EHOSTUNREACH);
2904 kiblnd_peer_decref(peer);
2905 return -EHOSTUNREACH; /* rc != 0 destroys cmid */
2907 case RDMA_CM_EVENT_ROUTE_RESOLVED:
2908 peer = (kib_peer_t *)cmid->context;
2909 CDEBUG(D_NET,"%s Route resolved: %d\n",
2910 libcfs_nid2str(peer->ibp_nid), event->status);
2912 if (event->status == 0)
2913 return kiblnd_active_connect(cmid);
2915 CNETERR("Can't resolve route for %s: %d\n",
2916 libcfs_nid2str(peer->ibp_nid), event->status);
2917 kiblnd_peer_connect_failed(peer, 1, event->status);
2918 kiblnd_peer_decref(peer);
2919 return event->status; /* rc != 0 destroys cmid */
2921 case RDMA_CM_EVENT_UNREACHABLE:
2922 conn = (kib_conn_t *)cmid->context;
2923 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
2924 conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
2925 CNETERR("%s: UNREACHABLE %d\n",
2926 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
2927 kiblnd_connreq_done(conn, -ENETDOWN);
2928 kiblnd_conn_decref(conn);
2931 case RDMA_CM_EVENT_CONNECT_ERROR:
2932 conn = (kib_conn_t *)cmid->context;
2933 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
2934 conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
2935 CNETERR("%s: CONNECT ERROR %d\n",
2936 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
2937 kiblnd_connreq_done(conn, -ENOTCONN);
2938 kiblnd_conn_decref(conn);
2941 case RDMA_CM_EVENT_REJECTED:
2942 conn = (kib_conn_t *)cmid->context;
2943 switch (conn->ibc_state) {
2947 case IBLND_CONN_PASSIVE_WAIT:
2948 CERROR ("%s: REJECTED %d\n",
2949 libcfs_nid2str(conn->ibc_peer->ibp_nid),
2951 kiblnd_connreq_done(conn, -ECONNRESET);
2954 case IBLND_CONN_ACTIVE_CONNECT:
2955 kiblnd_rejected(conn, event->status,
2956 (void *)KIBLND_CONN_PARAM(event),
2957 KIBLND_CONN_PARAM_LEN(event));
2960 kiblnd_conn_decref(conn);
2963 case RDMA_CM_EVENT_ESTABLISHED:
2964 conn = (kib_conn_t *)cmid->context;
2965 switch (conn->ibc_state) {
2969 case IBLND_CONN_PASSIVE_WAIT:
2970 CDEBUG(D_NET, "ESTABLISHED (passive): %s\n",
2971 libcfs_nid2str(conn->ibc_peer->ibp_nid));
2972 kiblnd_connreq_done(conn, 0);
2975 case IBLND_CONN_ACTIVE_CONNECT:
2976 CDEBUG(D_NET, "ESTABLISHED(active): %s\n",
2977 libcfs_nid2str(conn->ibc_peer->ibp_nid));
2978 kiblnd_check_connreply(conn,
2979 (void *)KIBLND_CONN_PARAM(event),
2980 KIBLND_CONN_PARAM_LEN(event));
2983 /* net keeps its ref on conn! */
2986 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
2987 CDEBUG(D_NET, "Ignore TIMEWAIT_EXIT event\n");
2990 case RDMA_CM_EVENT_DISCONNECTED:
2991 conn = (kib_conn_t *)cmid->context;
2992 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
2993 CERROR("%s DISCONNECTED\n",
2994 libcfs_nid2str(conn->ibc_peer->ibp_nid));
2995 kiblnd_connreq_done(conn, -ECONNRESET);
2997 kiblnd_close_conn(conn, 0);
2999 kiblnd_conn_decref(conn);
3000 cmid->context = NULL;
3003 case RDMA_CM_EVENT_DEVICE_REMOVAL:
3004 LCONSOLE_ERROR_MSG(0x131,
3005 "Received notification of device removal\n"
3006 "Please shutdown LNET to allow this to proceed\n");
3007 /* Can't remove network from underneath LNET for now, so I have
3011 case RDMA_CM_EVENT_ADDR_CHANGE:
3012 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(&wait, current);
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 schedule_timeout(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(&wait, current);
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);
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(&wait, current);
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();