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.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2015, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lnet/klnds/o2iblnd/o2iblnd_cb.c
34 * Author: Eric Barton <eric@bartonsoftware.com>
39 #define MAX_CONN_RACES_BEFORE_ABORT 20
41 static void kiblnd_peer_alive(struct kib_peer *peer);
42 static void kiblnd_peer_connect_failed(struct kib_peer *peer, int active, int error);
43 static void kiblnd_init_tx_msg(lnet_ni_t *ni, struct kib_tx *tx,
44 int type, int body_nob);
45 static int kiblnd_init_rdma(struct kib_conn *conn, struct kib_tx *tx, int type,
46 int resid, struct kib_rdma_desc *dstrd,
48 static void kiblnd_queue_tx_locked(struct kib_tx *tx, struct kib_conn *conn);
49 static void kiblnd_queue_tx(struct kib_tx *tx, struct kib_conn *conn);
50 static void kiblnd_unmap_tx(lnet_ni_t *ni, struct kib_tx *tx);
51 static void kiblnd_check_sends_locked(struct kib_conn *conn);
54 kiblnd_tx_done(lnet_ni_t *ni, struct kib_tx *tx)
56 lnet_msg_t *lntmsg[2];
57 struct kib_net *net = ni->ni_data;
62 LASSERT(!in_interrupt());
63 LASSERT(!tx->tx_queued); /* mustn't be queued for sending */
64 LASSERT(!tx->tx_sending); /* mustn't be awaiting sent callback */
65 LASSERT(!tx->tx_waiting); /* mustn't be awaiting peer response */
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;
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++) {
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, struct kib_tx, tx_list);
104 list_del(&tx->tx_list);
107 tx->tx_status = status;
108 kiblnd_tx_done(ni, tx);
112 static struct kib_tx *
113 kiblnd_get_idle_tx(lnet_ni_t *ni, lnet_nid_t target)
115 struct kib_net *net = (struct kib_net *)ni->ni_data;
116 struct list_head *node;
118 struct kib_tx_poolset *tps;
120 tps = net->ibn_tx_ps[lnet_cpt_of_nid(target)];
121 node = kiblnd_pool_alloc_node(&tps->tps_poolset);
124 tx = list_entry(node, struct kib_tx, tx_list);
126 LASSERT(!tx->tx_nwrq);
127 LASSERT(!tx->tx_queued);
128 LASSERT(!tx->tx_sending);
129 LASSERT(!tx->tx_waiting);
130 LASSERT(!tx->tx_status);
131 LASSERT(!tx->tx_conn);
132 LASSERT(!tx->tx_lntmsg[0]);
133 LASSERT(!tx->tx_lntmsg[1]);
134 LASSERT(!tx->tx_nfrags);
140 kiblnd_drop_rx(struct kib_rx *rx)
142 struct kib_conn *conn = rx->rx_conn;
143 struct kib_sched_info *sched = conn->ibc_sched;
146 spin_lock_irqsave(&sched->ibs_lock, flags);
147 LASSERT(conn->ibc_nrx > 0);
149 spin_unlock_irqrestore(&sched->ibs_lock, flags);
151 kiblnd_conn_decref(conn);
155 kiblnd_post_rx(struct kib_rx *rx, int credit)
157 struct kib_conn *conn = rx->rx_conn;
158 struct kib_net *net = conn->ibc_peer->ibp_ni->ni_data;
159 struct ib_recv_wr *bad_wrq = NULL;
160 struct ib_mr *mr = conn->ibc_hdev->ibh_mrs;
164 LASSERT(!in_interrupt());
165 LASSERT(credit == IBLND_POSTRX_NO_CREDIT ||
166 credit == IBLND_POSTRX_PEER_CREDIT ||
167 credit == IBLND_POSTRX_RSRVD_CREDIT);
170 rx->rx_sge.lkey = mr->lkey;
171 rx->rx_sge.addr = rx->rx_msgaddr;
172 rx->rx_sge.length = IBLND_MSG_SIZE;
174 rx->rx_wrq.next = NULL;
175 rx->rx_wrq.sg_list = &rx->rx_sge;
176 rx->rx_wrq.num_sge = 1;
177 rx->rx_wrq.wr_id = kiblnd_ptr2wreqid(rx, IBLND_WID_RX);
179 LASSERT(conn->ibc_state >= IBLND_CONN_INIT);
180 LASSERT(rx->rx_nob >= 0); /* not posted */
182 if (conn->ibc_state > IBLND_CONN_ESTABLISHED) {
183 kiblnd_drop_rx(rx); /* No more posts for this rx */
187 rx->rx_nob = -1; /* flag posted */
189 /* NB: need an extra reference after ib_post_recv because we don't
190 * own this rx (and rx::rx_conn) anymore, LU-5678.
192 kiblnd_conn_addref(conn);
193 rc = ib_post_recv(conn->ibc_cmid->qp, &rx->rx_wrq, &bad_wrq);
195 CERROR("Can't post rx for %s: %d, bad_wrq: %p\n",
196 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc, bad_wrq);
200 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) /* Initial post */
204 kiblnd_close_conn(conn, rc);
205 kiblnd_drop_rx(rx); /* No more posts for this rx */
209 if (credit == IBLND_POSTRX_NO_CREDIT)
212 spin_lock(&conn->ibc_lock);
213 if (credit == IBLND_POSTRX_PEER_CREDIT)
214 conn->ibc_outstanding_credits++;
216 conn->ibc_reserved_credits++;
217 kiblnd_check_sends_locked(conn);
218 spin_unlock(&conn->ibc_lock);
221 kiblnd_conn_decref(conn);
225 static struct kib_tx *
226 kiblnd_find_waiting_tx_locked(struct kib_conn *conn, int txtype, __u64 cookie)
228 struct list_head *tmp;
230 list_for_each(tmp, &conn->ibc_active_txs) {
231 struct kib_tx *tx = list_entry(tmp, struct kib_tx, tx_list);
233 LASSERT(!tx->tx_queued);
234 LASSERT(tx->tx_sending || tx->tx_waiting);
236 if (tx->tx_cookie != cookie)
239 if (tx->tx_waiting &&
240 tx->tx_msg->ibm_type == txtype)
243 CWARN("Bad completion: %swaiting, type %x (wanted %x)\n",
244 tx->tx_waiting ? "" : "NOT ",
245 tx->tx_msg->ibm_type, txtype);
251 kiblnd_handle_completion(struct kib_conn *conn, int txtype, int status, __u64 cookie)
254 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
257 spin_lock(&conn->ibc_lock);
259 tx = kiblnd_find_waiting_tx_locked(conn, txtype, cookie);
261 spin_unlock(&conn->ibc_lock);
263 CWARN("Unmatched completion type %x cookie %#llx from %s\n",
264 txtype, cookie, libcfs_nid2str(conn->ibc_peer->ibp_nid));
265 kiblnd_close_conn(conn, -EPROTO);
269 if (!tx->tx_status) { /* success so far */
270 if (status < 0) /* failed? */
271 tx->tx_status = status;
272 else if (txtype == IBLND_MSG_GET_REQ)
273 lnet_set_reply_msg_len(ni, tx->tx_lntmsg[1], status);
278 idle = !tx->tx_queued && !tx->tx_sending;
280 list_del(&tx->tx_list);
282 spin_unlock(&conn->ibc_lock);
285 kiblnd_tx_done(ni, tx);
289 kiblnd_send_completion(struct kib_conn *conn, int type, int status, __u64 cookie)
291 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
292 struct kib_tx *tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
295 CERROR("Can't get tx for completion %x for %s\n",
296 type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
300 tx->tx_msg->ibm_u.completion.ibcm_status = status;
301 tx->tx_msg->ibm_u.completion.ibcm_cookie = cookie;
302 kiblnd_init_tx_msg(ni, tx, type, sizeof(struct kib_completion_msg));
304 kiblnd_queue_tx(tx, conn);
308 kiblnd_handle_rx(struct kib_rx *rx)
310 struct kib_msg *msg = rx->rx_msg;
311 struct kib_conn *conn = rx->rx_conn;
312 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
313 int credits = msg->ibm_credits;
319 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
321 CDEBUG(D_NET, "Received %x[%d] from %s\n",
322 msg->ibm_type, credits,
323 libcfs_nid2str(conn->ibc_peer->ibp_nid));
326 /* Have I received credits that will let me send? */
327 spin_lock(&conn->ibc_lock);
329 if (conn->ibc_credits + credits >
330 conn->ibc_queue_depth) {
331 rc2 = conn->ibc_credits;
332 spin_unlock(&conn->ibc_lock);
334 CERROR("Bad credits from %s: %d + %d > %d\n",
335 libcfs_nid2str(conn->ibc_peer->ibp_nid),
336 rc2, credits, conn->ibc_queue_depth);
338 kiblnd_close_conn(conn, -EPROTO);
339 kiblnd_post_rx(rx, IBLND_POSTRX_NO_CREDIT);
343 conn->ibc_credits += credits;
345 /* This ensures the credit taken by NOOP can be returned */
346 if (msg->ibm_type == IBLND_MSG_NOOP &&
347 !IBLND_OOB_CAPABLE(conn->ibc_version)) /* v1 only */
348 conn->ibc_outstanding_credits++;
350 kiblnd_check_sends_locked(conn);
351 spin_unlock(&conn->ibc_lock);
354 switch (msg->ibm_type) {
356 CERROR("Bad IBLND message type %x from %s\n",
357 msg->ibm_type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
358 post_credit = IBLND_POSTRX_NO_CREDIT;
363 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
364 post_credit = IBLND_POSTRX_NO_CREDIT;
368 if (credits) /* credit already posted */
369 post_credit = IBLND_POSTRX_NO_CREDIT;
370 else /* a keepalive NOOP */
371 post_credit = IBLND_POSTRX_PEER_CREDIT;
374 case IBLND_MSG_IMMEDIATE:
375 post_credit = IBLND_POSTRX_DONT_POST;
376 rc = lnet_parse(ni, &msg->ibm_u.immediate.ibim_hdr,
377 msg->ibm_srcnid, rx, 0);
378 if (rc < 0) /* repost on error */
379 post_credit = IBLND_POSTRX_PEER_CREDIT;
382 case IBLND_MSG_PUT_REQ:
383 post_credit = IBLND_POSTRX_DONT_POST;
384 rc = lnet_parse(ni, &msg->ibm_u.putreq.ibprm_hdr,
385 msg->ibm_srcnid, rx, 1);
386 if (rc < 0) /* repost on error */
387 post_credit = IBLND_POSTRX_PEER_CREDIT;
390 case IBLND_MSG_PUT_NAK:
391 CWARN("PUT_NACK from %s\n",
392 libcfs_nid2str(conn->ibc_peer->ibp_nid));
393 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
394 kiblnd_handle_completion(conn, IBLND_MSG_PUT_REQ,
395 msg->ibm_u.completion.ibcm_status,
396 msg->ibm_u.completion.ibcm_cookie);
399 case IBLND_MSG_PUT_ACK:
400 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
402 spin_lock(&conn->ibc_lock);
403 tx = kiblnd_find_waiting_tx_locked(conn, IBLND_MSG_PUT_REQ,
404 msg->ibm_u.putack.ibpam_src_cookie);
406 list_del(&tx->tx_list);
407 spin_unlock(&conn->ibc_lock);
410 CERROR("Unmatched PUT_ACK from %s\n",
411 libcfs_nid2str(conn->ibc_peer->ibp_nid));
416 LASSERT(tx->tx_waiting);
418 * CAVEAT EMPTOR: I could be racing with tx_complete, but...
419 * (a) I can overwrite tx_msg since my peer has received it!
420 * (b) tx_waiting set tells tx_complete() it's not done.
422 tx->tx_nwrq = 0; /* overwrite PUT_REQ */
424 rc2 = kiblnd_init_rdma(conn, tx, IBLND_MSG_PUT_DONE,
425 kiblnd_rd_size(&msg->ibm_u.putack.ibpam_rd),
426 &msg->ibm_u.putack.ibpam_rd,
427 msg->ibm_u.putack.ibpam_dst_cookie);
429 CERROR("Can't setup rdma for PUT to %s: %d\n",
430 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc2);
432 spin_lock(&conn->ibc_lock);
433 tx->tx_waiting = 0; /* clear waiting and queue atomically */
434 kiblnd_queue_tx_locked(tx, conn);
435 spin_unlock(&conn->ibc_lock);
438 case IBLND_MSG_PUT_DONE:
439 post_credit = IBLND_POSTRX_PEER_CREDIT;
440 kiblnd_handle_completion(conn, IBLND_MSG_PUT_ACK,
441 msg->ibm_u.completion.ibcm_status,
442 msg->ibm_u.completion.ibcm_cookie);
445 case IBLND_MSG_GET_REQ:
446 post_credit = IBLND_POSTRX_DONT_POST;
447 rc = lnet_parse(ni, &msg->ibm_u.get.ibgm_hdr,
448 msg->ibm_srcnid, rx, 1);
449 if (rc < 0) /* repost on error */
450 post_credit = IBLND_POSTRX_PEER_CREDIT;
453 case IBLND_MSG_GET_DONE:
454 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
455 kiblnd_handle_completion(conn, IBLND_MSG_GET_REQ,
456 msg->ibm_u.completion.ibcm_status,
457 msg->ibm_u.completion.ibcm_cookie);
461 if (rc < 0) /* protocol error */
462 kiblnd_close_conn(conn, rc);
464 if (post_credit != IBLND_POSTRX_DONT_POST)
465 kiblnd_post_rx(rx, post_credit);
469 kiblnd_rx_complete(struct kib_rx *rx, int status, int nob)
471 struct kib_msg *msg = rx->rx_msg;
472 struct kib_conn *conn = rx->rx_conn;
473 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
474 struct kib_net *net = ni->ni_data;
479 LASSERT(rx->rx_nob < 0); /* was posted */
480 rx->rx_nob = 0; /* isn't now */
482 if (conn->ibc_state > IBLND_CONN_ESTABLISHED)
485 if (status != IB_WC_SUCCESS) {
486 CNETERR("Rx from %s failed: %d\n",
487 libcfs_nid2str(conn->ibc_peer->ibp_nid), status);
494 rc = kiblnd_unpack_msg(msg, rx->rx_nob);
496 CERROR("Error %d unpacking rx from %s\n",
497 rc, libcfs_nid2str(conn->ibc_peer->ibp_nid));
501 if (msg->ibm_srcnid != conn->ibc_peer->ibp_nid ||
502 msg->ibm_dstnid != ni->ni_nid ||
503 msg->ibm_srcstamp != conn->ibc_incarnation ||
504 msg->ibm_dststamp != net->ibn_incarnation) {
505 CERROR("Stale rx from %s\n",
506 libcfs_nid2str(conn->ibc_peer->ibp_nid));
511 /* set time last known alive */
512 kiblnd_peer_alive(conn->ibc_peer);
514 /* racing with connection establishment/teardown! */
516 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
517 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
520 write_lock_irqsave(g_lock, flags);
521 /* must check holding global lock to eliminate race */
522 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
523 list_add_tail(&rx->rx_list, &conn->ibc_early_rxs);
524 write_unlock_irqrestore(g_lock, flags);
527 write_unlock_irqrestore(g_lock, flags);
529 kiblnd_handle_rx(rx);
533 CDEBUG(D_NET, "rx %p conn %p\n", rx, conn);
534 kiblnd_close_conn(conn, err);
536 kiblnd_drop_rx(rx); /* Don't re-post rx. */
540 kiblnd_kvaddr_to_page(unsigned long vaddr)
544 if (is_vmalloc_addr((void *)vaddr)) {
545 page = vmalloc_to_page((void *)vaddr);
549 #ifdef CONFIG_HIGHMEM
550 if (vaddr >= PKMAP_BASE &&
551 vaddr < (PKMAP_BASE + LAST_PKMAP * PAGE_SIZE)) {
552 /* No highmem pages only used for bulk (kiov) I/O */
553 CERROR("find page for address in highmem\n");
557 page = virt_to_page(vaddr);
563 kiblnd_fmr_map_tx(struct kib_net *net, struct kib_tx *tx, struct kib_rdma_desc *rd, __u32 nob)
565 struct kib_hca_dev *hdev;
566 struct kib_fmr_poolset *fps;
570 LASSERT(tx->tx_pool);
571 LASSERT(tx->tx_pool->tpo_pool.po_owner);
573 hdev = tx->tx_pool->tpo_hdev;
574 cpt = tx->tx_pool->tpo_pool.po_owner->ps_cpt;
576 fps = net->ibn_fmr_ps[cpt];
577 rc = kiblnd_fmr_pool_map(fps, tx, rd, nob, 0, &tx->fmr);
579 CERROR("Can't map %u bytes: %d\n", nob, rc);
584 * If rd is not tx_rd, it's going to get sent to a peer, who will need
587 rd->rd_key = tx->fmr.fmr_key;
588 rd->rd_frags[0].rf_addr &= ~hdev->ibh_page_mask;
589 rd->rd_frags[0].rf_nob = nob;
595 static void kiblnd_unmap_tx(lnet_ni_t *ni, struct kib_tx *tx)
597 struct kib_net *net = ni->ni_data;
602 kiblnd_fmr_pool_unmap(&tx->fmr, tx->tx_status);
605 kiblnd_dma_unmap_sg(tx->tx_pool->tpo_hdev->ibh_ibdev,
606 tx->tx_frags, tx->tx_nfrags, tx->tx_dmadir);
611 static int kiblnd_map_tx(lnet_ni_t *ni, struct kib_tx *tx, struct kib_rdma_desc *rd,
614 struct kib_net *net = ni->ni_data;
615 struct kib_hca_dev *hdev = net->ibn_dev->ibd_hdev;
616 struct ib_mr *mr = NULL;
621 * If rd is not tx_rd, it's going to get sent to a peer and I'm the
624 tx->tx_dmadir = (rd != tx->tx_rd) ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
625 tx->tx_nfrags = nfrags;
627 rd->rd_nfrags = kiblnd_dma_map_sg(hdev->ibh_ibdev, tx->tx_frags,
628 tx->tx_nfrags, tx->tx_dmadir);
630 for (i = 0, nob = 0; i < rd->rd_nfrags; i++) {
631 rd->rd_frags[i].rf_nob = kiblnd_sg_dma_len(
632 hdev->ibh_ibdev, &tx->tx_frags[i]);
633 rd->rd_frags[i].rf_addr = kiblnd_sg_dma_address(
634 hdev->ibh_ibdev, &tx->tx_frags[i]);
635 nob += rd->rd_frags[i].rf_nob;
638 mr = kiblnd_find_rd_dma_mr(ni, rd, tx->tx_conn ?
639 tx->tx_conn->ibc_max_frags : -1);
641 /* found pre-mapping MR */
642 rd->rd_key = (rd != tx->tx_rd) ? mr->rkey : mr->lkey;
647 return kiblnd_fmr_map_tx(net, tx, rd, nob);
653 kiblnd_setup_rd_iov(lnet_ni_t *ni, struct kib_tx *tx, struct kib_rdma_desc *rd,
654 unsigned int niov, const struct kvec *iov, int offset, int nob)
656 struct kib_net *net = ni->ni_data;
658 struct scatterlist *sg;
667 while (offset >= iov->iov_len) {
668 offset -= iov->iov_len;
678 vaddr = ((unsigned long)iov->iov_base) + offset;
679 page_offset = vaddr & (PAGE_SIZE - 1);
680 page = kiblnd_kvaddr_to_page(vaddr);
682 CERROR("Can't find page\n");
686 fragnob = min((int)(iov->iov_len - offset), nob);
687 fragnob = min(fragnob, (int)PAGE_SIZE - page_offset);
689 sg_set_page(sg, page, fragnob, page_offset);
692 CERROR("lacking enough sg entries to map tx\n");
696 if (offset + fragnob < iov->iov_len) {
706 return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags);
710 kiblnd_setup_rd_kiov(lnet_ni_t *ni, struct kib_tx *tx, struct kib_rdma_desc *rd,
711 int nkiov, const lnet_kiov_t *kiov, int offset, int nob)
713 struct kib_net *net = ni->ni_data;
714 struct scatterlist *sg;
717 CDEBUG(D_NET, "niov %d offset %d nob %d\n", nkiov, offset, nob);
723 while (offset >= kiov->bv_len) {
724 offset -= kiov->bv_len;
734 fragnob = min((int)(kiov->bv_len - offset), nob);
736 sg_set_page(sg, kiov->bv_page, fragnob,
737 kiov->bv_offset + offset);
740 CERROR("lacking enough sg entries to map tx\n");
750 return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags);
754 kiblnd_post_tx_locked(struct kib_conn *conn, struct kib_tx *tx, int credit)
755 __must_hold(&conn->ibc_lock)
757 struct kib_msg *msg = tx->tx_msg;
758 struct kib_peer *peer = conn->ibc_peer;
759 struct lnet_ni *ni = peer->ibp_ni;
760 int ver = conn->ibc_version;
764 LASSERT(tx->tx_queued);
765 /* We rely on this for QP sizing */
766 LASSERT(tx->tx_nwrq > 0);
768 LASSERT(!credit || credit == 1);
769 LASSERT(conn->ibc_outstanding_credits >= 0);
770 LASSERT(conn->ibc_outstanding_credits <= conn->ibc_queue_depth);
771 LASSERT(conn->ibc_credits >= 0);
772 LASSERT(conn->ibc_credits <= conn->ibc_queue_depth);
774 if (conn->ibc_nsends_posted == kiblnd_concurrent_sends(ver, ni)) {
775 /* tx completions outstanding... */
776 CDEBUG(D_NET, "%s: posted enough\n",
777 libcfs_nid2str(peer->ibp_nid));
781 if (credit && !conn->ibc_credits) { /* no credits */
782 CDEBUG(D_NET, "%s: no credits\n",
783 libcfs_nid2str(peer->ibp_nid));
787 if (credit && !IBLND_OOB_CAPABLE(ver) &&
788 conn->ibc_credits == 1 && /* last credit reserved */
789 msg->ibm_type != IBLND_MSG_NOOP) { /* for NOOP */
790 CDEBUG(D_NET, "%s: not using last credit\n",
791 libcfs_nid2str(peer->ibp_nid));
795 /* NB don't drop ibc_lock before bumping tx_sending */
796 list_del(&tx->tx_list);
799 if (msg->ibm_type == IBLND_MSG_NOOP &&
800 (!kiblnd_need_noop(conn) || /* redundant NOOP */
801 (IBLND_OOB_CAPABLE(ver) && /* posted enough NOOP */
802 conn->ibc_noops_posted == IBLND_OOB_MSGS(ver)))) {
804 * OK to drop when posted enough NOOPs, since
805 * kiblnd_check_sends_locked will queue NOOP again when
806 * posted NOOPs complete
808 spin_unlock(&conn->ibc_lock);
809 kiblnd_tx_done(peer->ibp_ni, tx);
810 spin_lock(&conn->ibc_lock);
811 CDEBUG(D_NET, "%s(%d): redundant or enough NOOP\n",
812 libcfs_nid2str(peer->ibp_nid),
813 conn->ibc_noops_posted);
817 kiblnd_pack_msg(peer->ibp_ni, msg, ver, conn->ibc_outstanding_credits,
818 peer->ibp_nid, conn->ibc_incarnation);
820 conn->ibc_credits -= credit;
821 conn->ibc_outstanding_credits = 0;
822 conn->ibc_nsends_posted++;
823 if (msg->ibm_type == IBLND_MSG_NOOP)
824 conn->ibc_noops_posted++;
827 * CAVEAT EMPTOR! This tx could be the PUT_DONE of an RDMA
828 * PUT. If so, it was first queued here as a PUT_REQ, sent and
829 * stashed on ibc_active_txs, matched by an incoming PUT_ACK,
830 * and then re-queued here. It's (just) possible that
831 * tx_sending is non-zero if we've not done the tx_complete()
832 * from the first send; hence the ++ rather than = below.
835 list_add(&tx->tx_list, &conn->ibc_active_txs);
837 /* I'm still holding ibc_lock! */
838 if (conn->ibc_state != IBLND_CONN_ESTABLISHED) {
840 } else if (tx->tx_pool->tpo_pool.po_failed ||
841 conn->ibc_hdev != tx->tx_pool->tpo_hdev) {
842 /* close_conn will launch failover */
845 struct kib_fast_reg_descriptor *frd = tx->fmr.fmr_frd;
846 struct ib_send_wr *bad = &tx->tx_wrq[tx->tx_nwrq - 1].wr;
847 struct ib_send_wr *wrq = &tx->tx_wrq[0].wr;
850 if (!frd->frd_valid) {
851 wrq = &frd->frd_inv_wr;
852 wrq->next = &frd->frd_fastreg_wr.wr;
854 wrq = &frd->frd_fastreg_wr.wr;
856 frd->frd_fastreg_wr.wr.next = &tx->tx_wrq[0].wr;
859 LASSERTF(bad->wr_id == kiblnd_ptr2wreqid(tx, IBLND_WID_TX),
860 "bad wr_id %llx, opc %d, flags %d, peer: %s\n",
861 bad->wr_id, bad->opcode, bad->send_flags,
862 libcfs_nid2str(conn->ibc_peer->ibp_nid));
864 rc = ib_post_send(conn->ibc_cmid->qp, wrq, &bad);
867 conn->ibc_last_send = jiffies;
873 * NB credits are transferred in the actual
874 * message, which can only be the last work item
876 conn->ibc_credits += credit;
877 conn->ibc_outstanding_credits += msg->ibm_credits;
878 conn->ibc_nsends_posted--;
879 if (msg->ibm_type == IBLND_MSG_NOOP)
880 conn->ibc_noops_posted--;
886 done = !tx->tx_sending;
888 list_del(&tx->tx_list);
890 spin_unlock(&conn->ibc_lock);
892 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
893 CERROR("Error %d posting transmit to %s\n",
894 rc, libcfs_nid2str(peer->ibp_nid));
896 CDEBUG(D_NET, "Error %d posting transmit to %s\n",
897 rc, libcfs_nid2str(peer->ibp_nid));
899 kiblnd_close_conn(conn, rc);
902 kiblnd_tx_done(peer->ibp_ni, tx);
904 spin_lock(&conn->ibc_lock);
910 kiblnd_check_sends_locked(struct kib_conn *conn)
912 int ver = conn->ibc_version;
913 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
916 /* Don't send anything until after the connection is established */
917 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
918 CDEBUG(D_NET, "%s too soon\n",
919 libcfs_nid2str(conn->ibc_peer->ibp_nid));
923 LASSERT(conn->ibc_nsends_posted <= kiblnd_concurrent_sends(ver, ni));
924 LASSERT(!IBLND_OOB_CAPABLE(ver) ||
925 conn->ibc_noops_posted <= IBLND_OOB_MSGS(ver));
926 LASSERT(conn->ibc_reserved_credits >= 0);
928 while (conn->ibc_reserved_credits > 0 &&
929 !list_empty(&conn->ibc_tx_queue_rsrvd)) {
930 tx = list_entry(conn->ibc_tx_queue_rsrvd.next,
931 struct kib_tx, tx_list);
932 list_del(&tx->tx_list);
933 list_add_tail(&tx->tx_list, &conn->ibc_tx_queue);
934 conn->ibc_reserved_credits--;
937 if (kiblnd_need_noop(conn)) {
938 spin_unlock(&conn->ibc_lock);
940 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
942 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_NOOP, 0);
944 spin_lock(&conn->ibc_lock);
946 kiblnd_queue_tx_locked(tx, conn);
952 if (!list_empty(&conn->ibc_tx_queue_nocred)) {
954 tx = list_entry(conn->ibc_tx_queue_nocred.next,
955 struct kib_tx, tx_list);
956 } else if (!list_empty(&conn->ibc_tx_noops)) {
957 LASSERT(!IBLND_OOB_CAPABLE(ver));
959 tx = list_entry(conn->ibc_tx_noops.next,
960 struct kib_tx, tx_list);
961 } else if (!list_empty(&conn->ibc_tx_queue)) {
963 tx = list_entry(conn->ibc_tx_queue.next,
964 struct kib_tx, tx_list);
969 if (kiblnd_post_tx_locked(conn, tx, credit))
975 kiblnd_tx_complete(struct kib_tx *tx, int status)
977 int failed = (status != IB_WC_SUCCESS);
978 struct kib_conn *conn = tx->tx_conn;
981 LASSERT(tx->tx_sending > 0);
984 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
985 CNETERR("Tx -> %s cookie %#llx sending %d waiting %d: failed %d\n",
986 libcfs_nid2str(conn->ibc_peer->ibp_nid),
987 tx->tx_cookie, tx->tx_sending, tx->tx_waiting,
990 kiblnd_close_conn(conn, -EIO);
992 kiblnd_peer_alive(conn->ibc_peer);
995 spin_lock(&conn->ibc_lock);
998 * I could be racing with rdma completion. Whoever makes 'tx' idle
999 * gets to free it, which also drops its ref on 'conn'.
1002 conn->ibc_nsends_posted--;
1003 if (tx->tx_msg->ibm_type == IBLND_MSG_NOOP)
1004 conn->ibc_noops_posted--;
1007 tx->tx_waiting = 0; /* don't wait for peer */
1008 tx->tx_status = -EIO;
1011 idle = !tx->tx_sending && /* This is the final callback */
1012 !tx->tx_waiting && /* Not waiting for peer */
1013 !tx->tx_queued; /* Not re-queued (PUT_DONE) */
1015 list_del(&tx->tx_list);
1017 kiblnd_check_sends_locked(conn);
1018 spin_unlock(&conn->ibc_lock);
1021 kiblnd_tx_done(conn->ibc_peer->ibp_ni, tx);
1025 kiblnd_init_tx_msg(lnet_ni_t *ni, struct kib_tx *tx, int type, int body_nob)
1027 struct kib_hca_dev *hdev = tx->tx_pool->tpo_hdev;
1028 struct ib_sge *sge = &tx->tx_sge[tx->tx_nwrq];
1029 struct ib_rdma_wr *wrq = &tx->tx_wrq[tx->tx_nwrq];
1030 int nob = offsetof(struct kib_msg, ibm_u) + body_nob;
1031 struct ib_mr *mr = hdev->ibh_mrs;
1033 LASSERT(tx->tx_nwrq >= 0);
1034 LASSERT(tx->tx_nwrq < IBLND_MAX_RDMA_FRAGS + 1);
1035 LASSERT(nob <= IBLND_MSG_SIZE);
1038 kiblnd_init_msg(tx->tx_msg, type, body_nob);
1040 sge->lkey = mr->lkey;
1041 sge->addr = tx->tx_msgaddr;
1044 memset(wrq, 0, sizeof(*wrq));
1046 wrq->wr.next = NULL;
1047 wrq->wr.wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_TX);
1048 wrq->wr.sg_list = sge;
1049 wrq->wr.num_sge = 1;
1050 wrq->wr.opcode = IB_WR_SEND;
1051 wrq->wr.send_flags = IB_SEND_SIGNALED;
1057 kiblnd_init_rdma(struct kib_conn *conn, struct kib_tx *tx, int type,
1058 int resid, struct kib_rdma_desc *dstrd, __u64 dstcookie)
1060 struct kib_msg *ibmsg = tx->tx_msg;
1061 struct kib_rdma_desc *srcrd = tx->tx_rd;
1062 struct ib_sge *sge = &tx->tx_sge[0];
1063 struct ib_rdma_wr *wrq, *next;
1069 LASSERT(!in_interrupt());
1070 LASSERT(!tx->tx_nwrq);
1071 LASSERT(type == IBLND_MSG_GET_DONE ||
1072 type == IBLND_MSG_PUT_DONE);
1074 if (kiblnd_rd_size(srcrd) > conn->ibc_max_frags << PAGE_SHIFT) {
1075 CERROR("RDMA is too large for peer %s (%d), src size: %d dst size: %d\n",
1076 libcfs_nid2str(conn->ibc_peer->ibp_nid),
1077 conn->ibc_max_frags << PAGE_SHIFT,
1078 kiblnd_rd_size(srcrd), kiblnd_rd_size(dstrd));
1084 if (srcidx >= srcrd->rd_nfrags) {
1085 CERROR("Src buffer exhausted: %d frags\n", srcidx);
1090 if (dstidx == dstrd->rd_nfrags) {
1091 CERROR("Dst buffer exhausted: %d frags\n", dstidx);
1096 if (tx->tx_nwrq >= IBLND_MAX_RDMA_FRAGS) {
1097 CERROR("RDMA has too many fragments for peer %s (%d), src idx/frags: %d/%d dst idx/frags: %d/%d\n",
1098 libcfs_nid2str(conn->ibc_peer->ibp_nid),
1099 IBLND_MAX_RDMA_FRAGS,
1100 srcidx, srcrd->rd_nfrags,
1101 dstidx, dstrd->rd_nfrags);
1106 wrknob = min(min(kiblnd_rd_frag_size(srcrd, srcidx),
1107 kiblnd_rd_frag_size(dstrd, dstidx)),
1110 sge = &tx->tx_sge[tx->tx_nwrq];
1111 sge->addr = kiblnd_rd_frag_addr(srcrd, srcidx);
1112 sge->lkey = kiblnd_rd_frag_key(srcrd, srcidx);
1113 sge->length = wrknob;
1115 wrq = &tx->tx_wrq[tx->tx_nwrq];
1118 wrq->wr.next = &next->wr;
1119 wrq->wr.wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_RDMA);
1120 wrq->wr.sg_list = sge;
1121 wrq->wr.num_sge = 1;
1122 wrq->wr.opcode = IB_WR_RDMA_WRITE;
1123 wrq->wr.send_flags = 0;
1125 wrq->remote_addr = kiblnd_rd_frag_addr(dstrd, dstidx);
1126 wrq->rkey = kiblnd_rd_frag_key(dstrd, dstidx);
1128 srcidx = kiblnd_rd_consume_frag(srcrd, srcidx, wrknob);
1129 dstidx = kiblnd_rd_consume_frag(dstrd, dstidx, wrknob);
1138 if (rc < 0) /* no RDMA if completing with failure */
1141 ibmsg->ibm_u.completion.ibcm_status = rc;
1142 ibmsg->ibm_u.completion.ibcm_cookie = dstcookie;
1143 kiblnd_init_tx_msg(conn->ibc_peer->ibp_ni, tx,
1144 type, sizeof(struct kib_completion_msg));
1150 kiblnd_queue_tx_locked(struct kib_tx *tx, struct kib_conn *conn)
1152 struct list_head *q;
1154 LASSERT(tx->tx_nwrq > 0); /* work items set up */
1155 LASSERT(!tx->tx_queued); /* not queued for sending already */
1156 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1159 tx->tx_deadline = jiffies +
1160 msecs_to_jiffies(*kiblnd_tunables.kib_timeout *
1164 kiblnd_conn_addref(conn);
1166 LASSERT(tx->tx_msg->ibm_type != IBLND_MSG_PUT_DONE);
1168 /* PUT_DONE first attached to conn as a PUT_REQ */
1169 LASSERT(tx->tx_conn == conn);
1170 LASSERT(tx->tx_msg->ibm_type == IBLND_MSG_PUT_DONE);
1173 switch (tx->tx_msg->ibm_type) {
1177 case IBLND_MSG_PUT_REQ:
1178 case IBLND_MSG_GET_REQ:
1179 q = &conn->ibc_tx_queue_rsrvd;
1182 case IBLND_MSG_PUT_NAK:
1183 case IBLND_MSG_PUT_ACK:
1184 case IBLND_MSG_PUT_DONE:
1185 case IBLND_MSG_GET_DONE:
1186 q = &conn->ibc_tx_queue_nocred;
1189 case IBLND_MSG_NOOP:
1190 if (IBLND_OOB_CAPABLE(conn->ibc_version))
1191 q = &conn->ibc_tx_queue_nocred;
1193 q = &conn->ibc_tx_noops;
1196 case IBLND_MSG_IMMEDIATE:
1197 q = &conn->ibc_tx_queue;
1201 list_add_tail(&tx->tx_list, q);
1205 kiblnd_queue_tx(struct kib_tx *tx, struct kib_conn *conn)
1207 spin_lock(&conn->ibc_lock);
1208 kiblnd_queue_tx_locked(tx, conn);
1209 kiblnd_check_sends_locked(conn);
1210 spin_unlock(&conn->ibc_lock);
1213 static int kiblnd_resolve_addr(struct rdma_cm_id *cmid,
1214 struct sockaddr_in *srcaddr,
1215 struct sockaddr_in *dstaddr,
1218 unsigned short port;
1221 /* allow the port to be reused */
1222 rc = rdma_set_reuseaddr(cmid, 1);
1224 CERROR("Unable to set reuse on cmid: %d\n", rc);
1228 /* look for a free privileged port */
1229 for (port = PROT_SOCK - 1; port > 0; port--) {
1230 srcaddr->sin_port = htons(port);
1231 rc = rdma_resolve_addr(cmid,
1232 (struct sockaddr *)srcaddr,
1233 (struct sockaddr *)dstaddr,
1236 CDEBUG(D_NET, "bound to port %hu\n", port);
1238 } else if (rc == -EADDRINUSE || rc == -EADDRNOTAVAIL) {
1239 CDEBUG(D_NET, "bind to port %hu failed: %d\n",
1246 CERROR("Failed to bind to a free privileged port\n");
1251 kiblnd_connect_peer(struct kib_peer *peer)
1253 struct rdma_cm_id *cmid;
1254 struct kib_dev *dev;
1255 struct kib_net *net = peer->ibp_ni->ni_data;
1256 struct sockaddr_in srcaddr;
1257 struct sockaddr_in dstaddr;
1261 LASSERT(peer->ibp_connecting > 0);
1262 LASSERT(!peer->ibp_reconnecting);
1264 cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, peer, RDMA_PS_TCP,
1268 CERROR("Can't create CMID for %s: %ld\n",
1269 libcfs_nid2str(peer->ibp_nid), PTR_ERR(cmid));
1275 memset(&srcaddr, 0, sizeof(srcaddr));
1276 srcaddr.sin_family = AF_INET;
1277 srcaddr.sin_addr.s_addr = htonl(dev->ibd_ifip);
1279 memset(&dstaddr, 0, sizeof(dstaddr));
1280 dstaddr.sin_family = AF_INET;
1281 dstaddr.sin_port = htons(*kiblnd_tunables.kib_service);
1282 dstaddr.sin_addr.s_addr = htonl(LNET_NIDADDR(peer->ibp_nid));
1284 kiblnd_peer_addref(peer); /* cmid's ref */
1286 if (*kiblnd_tunables.kib_use_priv_port) {
1287 rc = kiblnd_resolve_addr(cmid, &srcaddr, &dstaddr,
1288 *kiblnd_tunables.kib_timeout * 1000);
1290 rc = rdma_resolve_addr(cmid,
1291 (struct sockaddr *)&srcaddr,
1292 (struct sockaddr *)&dstaddr,
1293 *kiblnd_tunables.kib_timeout * 1000);
1296 /* Can't initiate address resolution: */
1297 CERROR("Can't resolve addr for %s: %d\n",
1298 libcfs_nid2str(peer->ibp_nid), rc);
1305 kiblnd_peer_connect_failed(peer, 1, rc);
1306 kiblnd_peer_decref(peer); /* cmid's ref */
1307 rdma_destroy_id(cmid);
1310 kiblnd_peer_connect_failed(peer, 1, rc);
1314 kiblnd_reconnect_peer(struct kib_peer *peer)
1316 rwlock_t *glock = &kiblnd_data.kib_global_lock;
1317 char *reason = NULL;
1318 struct list_head txs;
1319 unsigned long flags;
1321 INIT_LIST_HEAD(&txs);
1323 write_lock_irqsave(glock, flags);
1324 if (!peer->ibp_reconnecting) {
1325 if (peer->ibp_accepting)
1326 reason = "accepting";
1327 else if (peer->ibp_connecting)
1328 reason = "connecting";
1329 else if (!list_empty(&peer->ibp_conns))
1330 reason = "connected";
1331 else /* connected then closed */
1337 LASSERT(!peer->ibp_accepting && !peer->ibp_connecting &&
1338 list_empty(&peer->ibp_conns));
1339 peer->ibp_reconnecting = 0;
1341 if (!kiblnd_peer_active(peer)) {
1342 list_splice_init(&peer->ibp_tx_queue, &txs);
1343 reason = "unlinked";
1347 peer->ibp_connecting++;
1348 peer->ibp_reconnected++;
1349 write_unlock_irqrestore(glock, flags);
1351 kiblnd_connect_peer(peer);
1355 write_unlock_irqrestore(glock, flags);
1357 CWARN("Abort reconnection of %s: %s\n",
1358 libcfs_nid2str(peer->ibp_nid), reason);
1359 kiblnd_txlist_done(peer->ibp_ni, &txs, -ECONNABORTED);
1364 kiblnd_launch_tx(lnet_ni_t *ni, struct kib_tx *tx, lnet_nid_t nid)
1366 struct kib_peer *peer;
1367 struct kib_peer *peer2;
1368 struct kib_conn *conn;
1369 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
1370 unsigned long flags;
1374 * If I get here, I've committed to send, so I complete the tx with
1375 * failure on any problems
1377 LASSERT(!tx || !tx->tx_conn); /* only set when assigned a conn */
1378 LASSERT(!tx || tx->tx_nwrq > 0); /* work items have been set up */
1381 * First time, just use a read lock since I expect to find my peer
1384 read_lock_irqsave(g_lock, flags);
1386 peer = kiblnd_find_peer_locked(nid);
1387 if (peer && !list_empty(&peer->ibp_conns)) {
1388 /* Found a peer with an established connection */
1389 conn = kiblnd_get_conn_locked(peer);
1390 kiblnd_conn_addref(conn); /* 1 ref for me... */
1392 read_unlock_irqrestore(g_lock, flags);
1395 kiblnd_queue_tx(tx, conn);
1396 kiblnd_conn_decref(conn); /* ...to here */
1400 read_unlock(g_lock);
1401 /* Re-try with a write lock */
1404 peer = kiblnd_find_peer_locked(nid);
1406 if (list_empty(&peer->ibp_conns)) {
1407 /* found a peer, but it's still connecting... */
1408 LASSERT(kiblnd_peer_connecting(peer));
1410 list_add_tail(&tx->tx_list,
1411 &peer->ibp_tx_queue);
1412 write_unlock_irqrestore(g_lock, flags);
1414 conn = kiblnd_get_conn_locked(peer);
1415 kiblnd_conn_addref(conn); /* 1 ref for me... */
1417 write_unlock_irqrestore(g_lock, flags);
1420 kiblnd_queue_tx(tx, conn);
1421 kiblnd_conn_decref(conn); /* ...to here */
1426 write_unlock_irqrestore(g_lock, flags);
1428 /* Allocate a peer ready to add to the peer table and retry */
1429 rc = kiblnd_create_peer(ni, &peer, nid);
1431 CERROR("Can't create peer %s\n", libcfs_nid2str(nid));
1433 tx->tx_status = -EHOSTUNREACH;
1435 kiblnd_tx_done(ni, tx);
1440 write_lock_irqsave(g_lock, flags);
1442 peer2 = kiblnd_find_peer_locked(nid);
1444 if (list_empty(&peer2->ibp_conns)) {
1445 /* found a peer, but it's still connecting... */
1446 LASSERT(kiblnd_peer_connecting(peer2));
1448 list_add_tail(&tx->tx_list,
1449 &peer2->ibp_tx_queue);
1450 write_unlock_irqrestore(g_lock, flags);
1452 conn = kiblnd_get_conn_locked(peer2);
1453 kiblnd_conn_addref(conn); /* 1 ref for me... */
1455 write_unlock_irqrestore(g_lock, flags);
1458 kiblnd_queue_tx(tx, conn);
1459 kiblnd_conn_decref(conn); /* ...to here */
1462 kiblnd_peer_decref(peer);
1466 /* Brand new peer */
1467 LASSERT(!peer->ibp_connecting);
1468 peer->ibp_connecting = 1;
1470 /* always called with a ref on ni, which prevents ni being shutdown */
1471 LASSERT(!((struct kib_net *)ni->ni_data)->ibn_shutdown);
1474 list_add_tail(&tx->tx_list, &peer->ibp_tx_queue);
1476 kiblnd_peer_addref(peer);
1477 list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid));
1479 write_unlock_irqrestore(g_lock, flags);
1481 kiblnd_connect_peer(peer);
1482 kiblnd_peer_decref(peer);
1486 kiblnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg)
1488 lnet_hdr_t *hdr = &lntmsg->msg_hdr;
1489 int type = lntmsg->msg_type;
1490 lnet_process_id_t target = lntmsg->msg_target;
1491 int target_is_router = lntmsg->msg_target_is_router;
1492 int routing = lntmsg->msg_routing;
1493 unsigned int payload_niov = lntmsg->msg_niov;
1494 struct kvec *payload_iov = lntmsg->msg_iov;
1495 lnet_kiov_t *payload_kiov = lntmsg->msg_kiov;
1496 unsigned int payload_offset = lntmsg->msg_offset;
1497 unsigned int payload_nob = lntmsg->msg_len;
1498 struct iov_iter from;
1499 struct kib_msg *ibmsg;
1500 struct kib_rdma_desc *rd;
1505 /* NB 'private' is different depending on what we're sending.... */
1507 CDEBUG(D_NET, "sending %d bytes in %d frags to %s\n",
1508 payload_nob, payload_niov, libcfs_id2str(target));
1510 LASSERT(!payload_nob || payload_niov > 0);
1511 LASSERT(payload_niov <= LNET_MAX_IOV);
1513 /* Thread context */
1514 LASSERT(!in_interrupt());
1515 /* payload is either all vaddrs or all pages */
1516 LASSERT(!(payload_kiov && payload_iov));
1519 iov_iter_bvec(&from, ITER_BVEC | WRITE,
1520 payload_kiov, payload_niov,
1521 payload_nob + payload_offset);
1523 iov_iter_kvec(&from, ITER_KVEC | WRITE,
1524 payload_iov, payload_niov,
1525 payload_nob + payload_offset);
1527 iov_iter_advance(&from, payload_offset);
1535 LASSERT(!payload_nob);
1539 if (routing || target_is_router)
1540 break; /* send IMMEDIATE */
1542 /* is the REPLY message too small for RDMA? */
1543 nob = offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[lntmsg->msg_md->md_length]);
1544 if (nob <= IBLND_MSG_SIZE)
1545 break; /* send IMMEDIATE */
1547 tx = kiblnd_get_idle_tx(ni, target.nid);
1549 CERROR("Can't allocate txd for GET to %s\n",
1550 libcfs_nid2str(target.nid));
1555 rd = &ibmsg->ibm_u.get.ibgm_rd;
1556 if (!(lntmsg->msg_md->md_options & LNET_MD_KIOV))
1557 rc = kiblnd_setup_rd_iov(ni, tx, rd,
1558 lntmsg->msg_md->md_niov,
1559 lntmsg->msg_md->md_iov.iov,
1560 0, lntmsg->msg_md->md_length);
1562 rc = kiblnd_setup_rd_kiov(ni, tx, rd,
1563 lntmsg->msg_md->md_niov,
1564 lntmsg->msg_md->md_iov.kiov,
1565 0, lntmsg->msg_md->md_length);
1567 CERROR("Can't setup GET sink for %s: %d\n",
1568 libcfs_nid2str(target.nid), rc);
1569 kiblnd_tx_done(ni, tx);
1573 nob = offsetof(struct kib_get_msg, ibgm_rd.rd_frags[rd->rd_nfrags]);
1574 ibmsg->ibm_u.get.ibgm_cookie = tx->tx_cookie;
1575 ibmsg->ibm_u.get.ibgm_hdr = *hdr;
1577 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_GET_REQ, nob);
1579 tx->tx_lntmsg[1] = lnet_create_reply_msg(ni, lntmsg);
1580 if (!tx->tx_lntmsg[1]) {
1581 CERROR("Can't create reply for GET -> %s\n",
1582 libcfs_nid2str(target.nid));
1583 kiblnd_tx_done(ni, tx);
1587 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg[0,1] on completion */
1588 tx->tx_waiting = 1; /* waiting for GET_DONE */
1589 kiblnd_launch_tx(ni, tx, target.nid);
1592 case LNET_MSG_REPLY:
1594 /* Is the payload small enough not to need RDMA? */
1595 nob = offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[payload_nob]);
1596 if (nob <= IBLND_MSG_SIZE)
1597 break; /* send IMMEDIATE */
1599 tx = kiblnd_get_idle_tx(ni, target.nid);
1601 CERROR("Can't allocate %s txd for %s\n",
1602 type == LNET_MSG_PUT ? "PUT" : "REPLY",
1603 libcfs_nid2str(target.nid));
1608 rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1609 payload_niov, payload_iov,
1610 payload_offset, payload_nob);
1612 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1613 payload_niov, payload_kiov,
1614 payload_offset, payload_nob);
1616 CERROR("Can't setup PUT src for %s: %d\n",
1617 libcfs_nid2str(target.nid), rc);
1618 kiblnd_tx_done(ni, tx);
1623 ibmsg->ibm_u.putreq.ibprm_hdr = *hdr;
1624 ibmsg->ibm_u.putreq.ibprm_cookie = tx->tx_cookie;
1625 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_REQ, sizeof(struct kib_putreq_msg));
1627 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1628 tx->tx_waiting = 1; /* waiting for PUT_{ACK,NAK} */
1629 kiblnd_launch_tx(ni, tx, target.nid);
1633 /* send IMMEDIATE */
1635 LASSERT(offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[payload_nob])
1638 tx = kiblnd_get_idle_tx(ni, target.nid);
1640 CERROR("Can't send %d to %s: tx descs exhausted\n",
1641 type, libcfs_nid2str(target.nid));
1646 ibmsg->ibm_u.immediate.ibim_hdr = *hdr;
1648 rc = copy_from_iter(&ibmsg->ibm_u.immediate.ibim_payload, payload_nob,
1650 if (rc != payload_nob) {
1651 kiblnd_pool_free_node(&tx->tx_pool->tpo_pool, &tx->tx_list);
1655 nob = offsetof(struct kib_immediate_msg, ibim_payload[payload_nob]);
1656 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_IMMEDIATE, nob);
1658 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1659 kiblnd_launch_tx(ni, tx, target.nid);
1664 kiblnd_reply(lnet_ni_t *ni, struct kib_rx *rx, lnet_msg_t *lntmsg)
1666 lnet_process_id_t target = lntmsg->msg_target;
1667 unsigned int niov = lntmsg->msg_niov;
1668 struct kvec *iov = lntmsg->msg_iov;
1669 lnet_kiov_t *kiov = lntmsg->msg_kiov;
1670 unsigned int offset = lntmsg->msg_offset;
1671 unsigned int nob = lntmsg->msg_len;
1675 tx = kiblnd_get_idle_tx(ni, rx->rx_conn->ibc_peer->ibp_nid);
1677 CERROR("Can't get tx for REPLY to %s\n",
1678 libcfs_nid2str(target.nid));
1685 rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1686 niov, iov, offset, nob);
1688 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1689 niov, kiov, offset, nob);
1692 CERROR("Can't setup GET src for %s: %d\n",
1693 libcfs_nid2str(target.nid), rc);
1697 rc = kiblnd_init_rdma(rx->rx_conn, tx,
1698 IBLND_MSG_GET_DONE, nob,
1699 &rx->rx_msg->ibm_u.get.ibgm_rd,
1700 rx->rx_msg->ibm_u.get.ibgm_cookie);
1702 CERROR("Can't setup rdma for GET from %s: %d\n",
1703 libcfs_nid2str(target.nid), rc);
1708 /* No RDMA: local completion may happen now! */
1709 lnet_finalize(ni, lntmsg, 0);
1711 /* RDMA: lnet_finalize(lntmsg) when it completes */
1712 tx->tx_lntmsg[0] = lntmsg;
1715 kiblnd_queue_tx(tx, rx->rx_conn);
1719 kiblnd_tx_done(ni, tx);
1721 lnet_finalize(ni, lntmsg, -EIO);
1725 kiblnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed,
1726 struct iov_iter *to, unsigned int rlen)
1728 struct kib_rx *rx = private;
1729 struct kib_msg *rxmsg = rx->rx_msg;
1730 struct kib_conn *conn = rx->rx_conn;
1733 int post_credit = IBLND_POSTRX_PEER_CREDIT;
1736 LASSERT(iov_iter_count(to) <= rlen);
1737 LASSERT(!in_interrupt());
1738 /* Either all pages or all vaddrs */
1740 switch (rxmsg->ibm_type) {
1744 case IBLND_MSG_IMMEDIATE:
1745 nob = offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[rlen]);
1746 if (nob > rx->rx_nob) {
1747 CERROR("Immediate message from %s too big: %d(%d)\n",
1748 libcfs_nid2str(rxmsg->ibm_u.immediate.ibim_hdr.src_nid),
1754 rc = copy_to_iter(&rxmsg->ibm_u.immediate.ibim_payload, rlen,
1762 lnet_finalize(ni, lntmsg, 0);
1765 case IBLND_MSG_PUT_REQ: {
1766 struct kib_msg *txmsg;
1767 struct kib_rdma_desc *rd;
1769 if (!iov_iter_count(to)) {
1770 lnet_finalize(ni, lntmsg, 0);
1771 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK, 0,
1772 rxmsg->ibm_u.putreq.ibprm_cookie);
1776 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
1778 CERROR("Can't allocate tx for %s\n",
1779 libcfs_nid2str(conn->ibc_peer->ibp_nid));
1780 /* Not replying will break the connection */
1786 rd = &txmsg->ibm_u.putack.ibpam_rd;
1787 if (!(to->type & ITER_BVEC))
1788 rc = kiblnd_setup_rd_iov(ni, tx, rd,
1789 to->nr_segs, to->kvec,
1791 iov_iter_count(to));
1793 rc = kiblnd_setup_rd_kiov(ni, tx, rd,
1794 to->nr_segs, to->bvec,
1796 iov_iter_count(to));
1798 CERROR("Can't setup PUT sink for %s: %d\n",
1799 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
1800 kiblnd_tx_done(ni, tx);
1801 /* tell peer it's over */
1802 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK, rc,
1803 rxmsg->ibm_u.putreq.ibprm_cookie);
1807 nob = offsetof(struct kib_putack_msg, ibpam_rd.rd_frags[rd->rd_nfrags]);
1808 txmsg->ibm_u.putack.ibpam_src_cookie = rxmsg->ibm_u.putreq.ibprm_cookie;
1809 txmsg->ibm_u.putack.ibpam_dst_cookie = tx->tx_cookie;
1811 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_ACK, nob);
1813 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1814 tx->tx_waiting = 1; /* waiting for PUT_DONE */
1815 kiblnd_queue_tx(tx, conn);
1817 /* reposted buffer reserved for PUT_DONE */
1818 post_credit = IBLND_POSTRX_NO_CREDIT;
1822 case IBLND_MSG_GET_REQ:
1824 /* Optimized GET; RDMA lntmsg's payload */
1825 kiblnd_reply(ni, rx, lntmsg);
1827 /* GET didn't match anything */
1828 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_GET_DONE,
1830 rxmsg->ibm_u.get.ibgm_cookie);
1835 kiblnd_post_rx(rx, post_credit);
1840 kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name)
1842 struct task_struct *task = kthread_run(fn, arg, "%s", name);
1845 return PTR_ERR(task);
1847 atomic_inc(&kiblnd_data.kib_nthreads);
1852 kiblnd_thread_fini(void)
1854 atomic_dec(&kiblnd_data.kib_nthreads);
1858 kiblnd_peer_alive(struct kib_peer *peer)
1860 /* This is racy, but everyone's only writing cfs_time_current() */
1861 peer->ibp_last_alive = cfs_time_current();
1866 kiblnd_peer_notify(struct kib_peer *peer)
1869 unsigned long last_alive = 0;
1870 unsigned long flags;
1872 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1874 if (kiblnd_peer_idle(peer) && peer->ibp_error) {
1875 error = peer->ibp_error;
1876 peer->ibp_error = 0;
1878 last_alive = peer->ibp_last_alive;
1881 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1884 lnet_notify(peer->ibp_ni,
1885 peer->ibp_nid, 0, last_alive);
1889 kiblnd_close_conn_locked(struct kib_conn *conn, int error)
1892 * This just does the immediate housekeeping. 'error' is zero for a
1893 * normal shutdown which can happen only after the connection has been
1894 * established. If the connection is established, schedule the
1895 * connection to be finished off by the connd. Otherwise the connd is
1896 * already dealing with it (either to set it up or tear it down).
1897 * Caller holds kib_global_lock exclusively in irq context
1899 struct kib_peer *peer = conn->ibc_peer;
1900 struct kib_dev *dev;
1901 unsigned long flags;
1903 LASSERT(error || conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1905 if (error && !conn->ibc_comms_error)
1906 conn->ibc_comms_error = error;
1908 if (conn->ibc_state != IBLND_CONN_ESTABLISHED)
1909 return; /* already being handled */
1912 list_empty(&conn->ibc_tx_noops) &&
1913 list_empty(&conn->ibc_tx_queue) &&
1914 list_empty(&conn->ibc_tx_queue_rsrvd) &&
1915 list_empty(&conn->ibc_tx_queue_nocred) &&
1916 list_empty(&conn->ibc_active_txs)) {
1917 CDEBUG(D_NET, "closing conn to %s\n",
1918 libcfs_nid2str(peer->ibp_nid));
1920 CNETERR("Closing conn to %s: error %d%s%s%s%s%s\n",
1921 libcfs_nid2str(peer->ibp_nid), error,
1922 list_empty(&conn->ibc_tx_queue) ? "" : "(sending)",
1923 list_empty(&conn->ibc_tx_noops) ? "" : "(sending_noops)",
1924 list_empty(&conn->ibc_tx_queue_rsrvd) ? "" : "(sending_rsrvd)",
1925 list_empty(&conn->ibc_tx_queue_nocred) ? "" : "(sending_nocred)",
1926 list_empty(&conn->ibc_active_txs) ? "" : "(waiting)");
1929 dev = ((struct kib_net *)peer->ibp_ni->ni_data)->ibn_dev;
1930 list_del(&conn->ibc_list);
1931 /* connd (see below) takes over ibc_list's ref */
1933 if (list_empty(&peer->ibp_conns) && /* no more conns */
1934 kiblnd_peer_active(peer)) { /* still in peer table */
1935 kiblnd_unlink_peer_locked(peer);
1937 /* set/clear error on last conn */
1938 peer->ibp_error = conn->ibc_comms_error;
1941 kiblnd_set_conn_state(conn, IBLND_CONN_CLOSING);
1944 kiblnd_dev_can_failover(dev)) {
1945 list_add_tail(&dev->ibd_fail_list,
1946 &kiblnd_data.kib_failed_devs);
1947 wake_up(&kiblnd_data.kib_failover_waitq);
1950 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
1952 list_add_tail(&conn->ibc_list, &kiblnd_data.kib_connd_conns);
1953 wake_up(&kiblnd_data.kib_connd_waitq);
1955 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
1959 kiblnd_close_conn(struct kib_conn *conn, int error)
1961 unsigned long flags;
1963 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1965 kiblnd_close_conn_locked(conn, error);
1967 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1971 kiblnd_handle_early_rxs(struct kib_conn *conn)
1973 unsigned long flags;
1977 LASSERT(!in_interrupt());
1978 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1980 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1981 list_for_each_entry_safe(rx, tmp, &conn->ibc_early_rxs, rx_list) {
1982 list_del(&rx->rx_list);
1983 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1985 kiblnd_handle_rx(rx);
1987 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1989 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1993 kiblnd_abort_txs(struct kib_conn *conn, struct list_head *txs)
1996 struct list_head *tmp;
1997 struct list_head *nxt;
2000 spin_lock(&conn->ibc_lock);
2002 list_for_each_safe(tmp, nxt, txs) {
2003 tx = list_entry(tmp, struct kib_tx, tx_list);
2005 if (txs == &conn->ibc_active_txs) {
2006 LASSERT(!tx->tx_queued);
2007 LASSERT(tx->tx_waiting || tx->tx_sending);
2009 LASSERT(tx->tx_queued);
2012 tx->tx_status = -ECONNABORTED;
2015 if (!tx->tx_sending) {
2017 list_del(&tx->tx_list);
2018 list_add(&tx->tx_list, &zombies);
2022 spin_unlock(&conn->ibc_lock);
2024 kiblnd_txlist_done(conn->ibc_peer->ibp_ni, &zombies, -ECONNABORTED);
2028 kiblnd_finalise_conn(struct kib_conn *conn)
2030 LASSERT(!in_interrupt());
2031 LASSERT(conn->ibc_state > IBLND_CONN_INIT);
2033 kiblnd_set_conn_state(conn, IBLND_CONN_DISCONNECTED);
2036 * abort_receives moves QP state to IB_QPS_ERR. This is only required
2037 * for connections that didn't get as far as being connected, because
2038 * rdma_disconnect() does this for free.
2040 kiblnd_abort_receives(conn);
2043 * Complete all tx descs not waiting for sends to complete.
2044 * NB we should be safe from RDMA now that the QP has changed state
2046 kiblnd_abort_txs(conn, &conn->ibc_tx_noops);
2047 kiblnd_abort_txs(conn, &conn->ibc_tx_queue);
2048 kiblnd_abort_txs(conn, &conn->ibc_tx_queue_rsrvd);
2049 kiblnd_abort_txs(conn, &conn->ibc_tx_queue_nocred);
2050 kiblnd_abort_txs(conn, &conn->ibc_active_txs);
2052 kiblnd_handle_early_rxs(conn);
2056 kiblnd_peer_connect_failed(struct kib_peer *peer, int active, int error)
2059 unsigned long flags;
2062 LASSERT(!in_interrupt());
2064 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2067 LASSERT(peer->ibp_connecting > 0);
2068 peer->ibp_connecting--;
2070 LASSERT(peer->ibp_accepting > 0);
2071 peer->ibp_accepting--;
2074 if (kiblnd_peer_connecting(peer)) {
2075 /* another connection attempt under way... */
2076 write_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2081 peer->ibp_reconnected = 0;
2082 if (list_empty(&peer->ibp_conns)) {
2083 /* Take peer's blocked transmits to complete with error */
2084 list_add(&zombies, &peer->ibp_tx_queue);
2085 list_del_init(&peer->ibp_tx_queue);
2087 if (kiblnd_peer_active(peer))
2088 kiblnd_unlink_peer_locked(peer);
2090 peer->ibp_error = error;
2092 /* Can't have blocked transmits if there are connections */
2093 LASSERT(list_empty(&peer->ibp_tx_queue));
2096 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2098 kiblnd_peer_notify(peer);
2100 if (list_empty(&zombies))
2103 CNETERR("Deleting messages for %s: connection failed\n",
2104 libcfs_nid2str(peer->ibp_nid));
2106 kiblnd_txlist_done(peer->ibp_ni, &zombies, -EHOSTUNREACH);
2110 kiblnd_connreq_done(struct kib_conn *conn, int status)
2112 struct kib_peer *peer = conn->ibc_peer;
2115 struct list_head txs;
2116 unsigned long flags;
2119 active = (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2121 CDEBUG(D_NET, "%s: active(%d), version(%x), status(%d)\n",
2122 libcfs_nid2str(peer->ibp_nid), active,
2123 conn->ibc_version, status);
2125 LASSERT(!in_interrupt());
2126 LASSERT((conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT &&
2127 peer->ibp_connecting > 0) ||
2128 (conn->ibc_state == IBLND_CONN_PASSIVE_WAIT &&
2129 peer->ibp_accepting > 0));
2131 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
2132 conn->ibc_connvars = NULL;
2135 /* failed to establish connection */
2136 kiblnd_peer_connect_failed(peer, active, status);
2137 kiblnd_finalise_conn(conn);
2141 /* connection established */
2142 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2144 conn->ibc_last_send = jiffies;
2145 kiblnd_set_conn_state(conn, IBLND_CONN_ESTABLISHED);
2146 kiblnd_peer_alive(peer);
2149 * Add conn to peer's list and nuke any dangling conns from a different
2152 kiblnd_conn_addref(conn); /* +1 ref for ibc_list */
2153 list_add(&conn->ibc_list, &peer->ibp_conns);
2154 peer->ibp_reconnected = 0;
2156 peer->ibp_connecting--;
2158 peer->ibp_accepting--;
2160 if (!peer->ibp_version) {
2161 peer->ibp_version = conn->ibc_version;
2162 peer->ibp_incarnation = conn->ibc_incarnation;
2165 if (peer->ibp_version != conn->ibc_version ||
2166 peer->ibp_incarnation != conn->ibc_incarnation) {
2167 kiblnd_close_stale_conns_locked(peer, conn->ibc_version,
2168 conn->ibc_incarnation);
2169 peer->ibp_version = conn->ibc_version;
2170 peer->ibp_incarnation = conn->ibc_incarnation;
2173 /* grab pending txs while I have the lock */
2174 list_add(&txs, &peer->ibp_tx_queue);
2175 list_del_init(&peer->ibp_tx_queue);
2177 if (!kiblnd_peer_active(peer) || /* peer has been deleted */
2178 conn->ibc_comms_error) { /* error has happened already */
2179 lnet_ni_t *ni = peer->ibp_ni;
2181 /* start to shut down connection */
2182 kiblnd_close_conn_locked(conn, -ECONNABORTED);
2183 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2185 kiblnd_txlist_done(ni, &txs, -ECONNABORTED);
2191 * +1 ref for myself, this connection is visible to other threads
2192 * now, refcount of peer:ibp_conns can be released by connection
2193 * close from either a different thread, or the calling of
2194 * kiblnd_check_sends_locked() below. See bz21911 for details.
2196 kiblnd_conn_addref(conn);
2197 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2199 /* Schedule blocked txs */
2200 spin_lock(&conn->ibc_lock);
2201 list_for_each_entry_safe(tx, tmp, &txs, tx_list) {
2202 list_del(&tx->tx_list);
2204 kiblnd_queue_tx_locked(tx, conn);
2206 kiblnd_check_sends_locked(conn);
2207 spin_unlock(&conn->ibc_lock);
2209 /* schedule blocked rxs */
2210 kiblnd_handle_early_rxs(conn);
2212 kiblnd_conn_decref(conn);
2216 kiblnd_reject(struct rdma_cm_id *cmid, struct kib_rej *rej)
2220 rc = rdma_reject(cmid, rej, sizeof(*rej));
2223 CWARN("Error %d sending reject\n", rc);
2227 kiblnd_passive_connect(struct rdma_cm_id *cmid, void *priv, int priv_nob)
2229 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2230 struct kib_msg *reqmsg = priv;
2231 struct kib_msg *ackmsg;
2232 struct kib_dev *ibdev;
2233 struct kib_peer *peer;
2234 struct kib_peer *peer2;
2235 struct kib_conn *conn;
2236 lnet_ni_t *ni = NULL;
2237 struct kib_net *net = NULL;
2239 struct rdma_conn_param cp;
2241 int version = IBLND_MSG_VERSION;
2242 unsigned long flags;
2245 struct sockaddr_in *peer_addr;
2247 LASSERT(!in_interrupt());
2249 /* cmid inherits 'context' from the corresponding listener id */
2250 ibdev = (struct kib_dev *)cmid->context;
2253 memset(&rej, 0, sizeof(rej));
2254 rej.ibr_magic = IBLND_MSG_MAGIC;
2255 rej.ibr_why = IBLND_REJECT_FATAL;
2256 rej.ibr_cp.ibcp_max_msg_size = IBLND_MSG_SIZE;
2258 peer_addr = (struct sockaddr_in *)&cmid->route.addr.dst_addr;
2259 if (*kiblnd_tunables.kib_require_priv_port &&
2260 ntohs(peer_addr->sin_port) >= PROT_SOCK) {
2261 __u32 ip = ntohl(peer_addr->sin_addr.s_addr);
2263 CERROR("Peer's port (%pI4h:%hu) is not privileged\n",
2264 &ip, ntohs(peer_addr->sin_port));
2268 if (priv_nob < offsetof(struct kib_msg, ibm_type)) {
2269 CERROR("Short connection request\n");
2274 * Future protocol version compatibility support! If the
2275 * o2iblnd-specific protocol changes, or when LNET unifies
2276 * protocols over all LNDs, the initial connection will
2277 * negotiate a protocol version. I trap this here to avoid
2278 * console errors; the reject tells the peer which protocol I
2281 if (reqmsg->ibm_magic == LNET_PROTO_MAGIC ||
2282 reqmsg->ibm_magic == __swab32(LNET_PROTO_MAGIC))
2284 if (reqmsg->ibm_magic == IBLND_MSG_MAGIC &&
2285 reqmsg->ibm_version != IBLND_MSG_VERSION &&
2286 reqmsg->ibm_version != IBLND_MSG_VERSION_1)
2288 if (reqmsg->ibm_magic == __swab32(IBLND_MSG_MAGIC) &&
2289 reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION) &&
2290 reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION_1))
2293 rc = kiblnd_unpack_msg(reqmsg, priv_nob);
2295 CERROR("Can't parse connection request: %d\n", rc);
2299 nid = reqmsg->ibm_srcnid;
2300 ni = lnet_net2ni(LNET_NIDNET(reqmsg->ibm_dstnid));
2303 net = (struct kib_net *)ni->ni_data;
2304 rej.ibr_incarnation = net->ibn_incarnation;
2307 if (!ni || /* no matching net */
2308 ni->ni_nid != reqmsg->ibm_dstnid || /* right NET, wrong NID! */
2309 net->ibn_dev != ibdev) { /* wrong device */
2310 CERROR("Can't accept conn from %s on %s (%s:%d:%pI4h): bad dst nid %s\n",
2311 libcfs_nid2str(nid),
2312 !ni ? "NA" : libcfs_nid2str(ni->ni_nid),
2313 ibdev->ibd_ifname, ibdev->ibd_nnets,
2315 libcfs_nid2str(reqmsg->ibm_dstnid));
2320 /* check time stamp as soon as possible */
2321 if (reqmsg->ibm_dststamp &&
2322 reqmsg->ibm_dststamp != net->ibn_incarnation) {
2323 CWARN("Stale connection request\n");
2324 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2328 /* I can accept peer's version */
2329 version = reqmsg->ibm_version;
2331 if (reqmsg->ibm_type != IBLND_MSG_CONNREQ) {
2332 CERROR("Unexpected connreq msg type: %x from %s\n",
2333 reqmsg->ibm_type, libcfs_nid2str(nid));
2337 if (reqmsg->ibm_u.connparams.ibcp_queue_depth >
2338 kiblnd_msg_queue_size(version, ni)) {
2339 CERROR("Can't accept conn from %s, queue depth too large: %d (<=%d wanted)\n",
2340 libcfs_nid2str(nid),
2341 reqmsg->ibm_u.connparams.ibcp_queue_depth,
2342 kiblnd_msg_queue_size(version, ni));
2344 if (version == IBLND_MSG_VERSION)
2345 rej.ibr_why = IBLND_REJECT_MSG_QUEUE_SIZE;
2350 max_frags = reqmsg->ibm_u.connparams.ibcp_max_frags >> IBLND_FRAG_SHIFT;
2351 if (max_frags > kiblnd_rdma_frags(version, ni)) {
2352 CWARN("Can't accept conn from %s (version %x): max message size %d is too large (%d wanted)\n",
2353 libcfs_nid2str(nid), version, max_frags,
2354 kiblnd_rdma_frags(version, ni));
2356 if (version >= IBLND_MSG_VERSION)
2357 rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2360 } else if (max_frags < kiblnd_rdma_frags(version, ni) &&
2362 CWARN("Can't accept conn from %s (version %x): max message size %d incompatible without FMR pool (%d wanted)\n",
2363 libcfs_nid2str(nid), version, max_frags,
2364 kiblnd_rdma_frags(version, ni));
2366 if (version == IBLND_MSG_VERSION)
2367 rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2372 if (reqmsg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2373 CERROR("Can't accept %s: message size %d too big (%d max)\n",
2374 libcfs_nid2str(nid),
2375 reqmsg->ibm_u.connparams.ibcp_max_msg_size,
2380 /* assume 'nid' is a new peer; create */
2381 rc = kiblnd_create_peer(ni, &peer, nid);
2383 CERROR("Can't create peer for %s\n", libcfs_nid2str(nid));
2384 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2388 /* We have validated the peer's parameters so use those */
2389 peer->ibp_max_frags = max_frags;
2390 peer->ibp_queue_depth = reqmsg->ibm_u.connparams.ibcp_queue_depth;
2392 write_lock_irqsave(g_lock, flags);
2394 peer2 = kiblnd_find_peer_locked(nid);
2396 if (!peer2->ibp_version) {
2397 peer2->ibp_version = version;
2398 peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2401 /* not the guy I've talked with */
2402 if (peer2->ibp_incarnation != reqmsg->ibm_srcstamp ||
2403 peer2->ibp_version != version) {
2404 kiblnd_close_peer_conns_locked(peer2, -ESTALE);
2406 if (kiblnd_peer_active(peer2)) {
2407 peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2408 peer2->ibp_version = version;
2410 write_unlock_irqrestore(g_lock, flags);
2412 CWARN("Conn stale %s version %x/%x incarnation %llu/%llu\n",
2413 libcfs_nid2str(nid), peer2->ibp_version, version,
2414 peer2->ibp_incarnation, reqmsg->ibm_srcstamp);
2416 kiblnd_peer_decref(peer);
2417 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2422 * Tie-break connection race in favour of the higher NID.
2423 * If we keep running into a race condition multiple times,
2424 * we have to assume that the connection attempt with the
2425 * higher NID is stuck in a connecting state and will never
2426 * recover. As such, we pass through this if-block and let
2427 * the lower NID connection win so we can move forward.
2429 if (peer2->ibp_connecting &&
2430 nid < ni->ni_nid && peer2->ibp_races <
2431 MAX_CONN_RACES_BEFORE_ABORT) {
2433 write_unlock_irqrestore(g_lock, flags);
2435 CDEBUG(D_NET, "Conn race %s\n",
2436 libcfs_nid2str(peer2->ibp_nid));
2438 kiblnd_peer_decref(peer);
2439 rej.ibr_why = IBLND_REJECT_CONN_RACE;
2442 if (peer2->ibp_races >= MAX_CONN_RACES_BEFORE_ABORT)
2443 CNETERR("Conn race %s: unresolved after %d attempts, letting lower NID win\n",
2444 libcfs_nid2str(peer2->ibp_nid),
2445 MAX_CONN_RACES_BEFORE_ABORT);
2447 * passive connection is allowed even this peer is waiting for
2450 peer2->ibp_reconnecting = 0;
2451 peer2->ibp_races = 0;
2452 peer2->ibp_accepting++;
2453 kiblnd_peer_addref(peer2);
2456 * Race with kiblnd_launch_tx (active connect) to create peer
2457 * so copy validated parameters since we now know what the
2460 peer2->ibp_max_frags = peer->ibp_max_frags;
2461 peer2->ibp_queue_depth = peer->ibp_queue_depth;
2463 write_unlock_irqrestore(g_lock, flags);
2464 kiblnd_peer_decref(peer);
2467 /* Brand new peer */
2468 LASSERT(!peer->ibp_accepting);
2469 LASSERT(!peer->ibp_version &&
2470 !peer->ibp_incarnation);
2472 peer->ibp_accepting = 1;
2473 peer->ibp_version = version;
2474 peer->ibp_incarnation = reqmsg->ibm_srcstamp;
2476 /* I have a ref on ni that prevents it being shutdown */
2477 LASSERT(!net->ibn_shutdown);
2479 kiblnd_peer_addref(peer);
2480 list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid));
2482 write_unlock_irqrestore(g_lock, flags);
2485 conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_PASSIVE_WAIT,
2488 kiblnd_peer_connect_failed(peer, 0, -ENOMEM);
2489 kiblnd_peer_decref(peer);
2490 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2495 * conn now "owns" cmid, so I return success from here on to ensure the
2496 * CM callback doesn't destroy cmid.
2498 conn->ibc_incarnation = reqmsg->ibm_srcstamp;
2499 conn->ibc_credits = conn->ibc_queue_depth;
2500 conn->ibc_reserved_credits = conn->ibc_queue_depth;
2501 LASSERT(conn->ibc_credits + conn->ibc_reserved_credits +
2502 IBLND_OOB_MSGS(version) <= IBLND_RX_MSGS(conn));
2504 ackmsg = &conn->ibc_connvars->cv_msg;
2505 memset(ackmsg, 0, sizeof(*ackmsg));
2507 kiblnd_init_msg(ackmsg, IBLND_MSG_CONNACK,
2508 sizeof(ackmsg->ibm_u.connparams));
2509 ackmsg->ibm_u.connparams.ibcp_queue_depth = conn->ibc_queue_depth;
2510 ackmsg->ibm_u.connparams.ibcp_max_frags = conn->ibc_max_frags << IBLND_FRAG_SHIFT;
2511 ackmsg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2513 kiblnd_pack_msg(ni, ackmsg, version, 0, nid, reqmsg->ibm_srcstamp);
2515 memset(&cp, 0, sizeof(cp));
2516 cp.private_data = ackmsg;
2517 cp.private_data_len = ackmsg->ibm_nob;
2518 cp.responder_resources = 0; /* No atomic ops or RDMA reads */
2519 cp.initiator_depth = 0;
2520 cp.flow_control = 1;
2521 cp.retry_count = *kiblnd_tunables.kib_retry_count;
2522 cp.rnr_retry_count = *kiblnd_tunables.kib_rnr_retry_count;
2524 CDEBUG(D_NET, "Accept %s\n", libcfs_nid2str(nid));
2526 rc = rdma_accept(cmid, &cp);
2528 CERROR("Can't accept %s: %d\n", libcfs_nid2str(nid), rc);
2529 rej.ibr_version = version;
2530 rej.ibr_why = IBLND_REJECT_FATAL;
2532 kiblnd_reject(cmid, &rej);
2533 kiblnd_connreq_done(conn, rc);
2534 kiblnd_conn_decref(conn);
2542 rej.ibr_cp.ibcp_queue_depth = kiblnd_msg_queue_size(version, ni);
2543 rej.ibr_cp.ibcp_max_frags = kiblnd_rdma_frags(version, ni);
2547 rej.ibr_version = version;
2548 kiblnd_reject(cmid, &rej);
2550 return -ECONNREFUSED;
2554 kiblnd_check_reconnect(struct kib_conn *conn, int version,
2555 __u64 incarnation, int why, struct kib_connparams *cp)
2557 rwlock_t *glock = &kiblnd_data.kib_global_lock;
2558 struct kib_peer *peer = conn->ibc_peer;
2560 int msg_size = IBLND_MSG_SIZE;
2564 unsigned long flags;
2566 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2567 LASSERT(peer->ibp_connecting > 0); /* 'conn' at least */
2568 LASSERT(!peer->ibp_reconnecting);
2571 msg_size = cp->ibcp_max_msg_size;
2572 frag_num = cp->ibcp_max_frags << IBLND_FRAG_SHIFT;
2573 queue_dep = cp->ibcp_queue_depth;
2576 write_lock_irqsave(glock, flags);
2578 * retry connection if it's still needed and no other connection
2579 * attempts (active or passive) are in progress
2580 * NB: reconnect is still needed even when ibp_tx_queue is
2581 * empty if ibp_version != version because reconnect may be
2582 * initiated by kiblnd_query()
2584 reconnect = (!list_empty(&peer->ibp_tx_queue) ||
2585 peer->ibp_version != version) &&
2586 peer->ibp_connecting == 1 &&
2587 !peer->ibp_accepting;
2598 case IBLND_REJECT_RDMA_FRAGS: {
2599 struct lnet_ioctl_config_lnd_tunables *tunables;
2602 reason = "can't negotiate max frags";
2605 tunables = peer->ibp_ni->ni_lnd_tunables;
2606 if (!tunables->lt_tun_u.lt_o2ib.lnd_map_on_demand) {
2607 reason = "map_on_demand must be enabled";
2610 if (conn->ibc_max_frags <= frag_num) {
2611 reason = "unsupported max frags";
2615 peer->ibp_max_frags = frag_num;
2616 reason = "rdma fragments";
2619 case IBLND_REJECT_MSG_QUEUE_SIZE:
2621 reason = "can't negotiate queue depth";
2624 if (conn->ibc_queue_depth <= queue_dep) {
2625 reason = "unsupported queue depth";
2629 peer->ibp_queue_depth = queue_dep;
2630 reason = "queue depth";
2633 case IBLND_REJECT_CONN_STALE:
2637 case IBLND_REJECT_CONN_RACE:
2638 reason = "conn race";
2641 case IBLND_REJECT_CONN_UNCOMPAT:
2642 reason = "version negotiation";
2646 conn->ibc_reconnect = 1;
2647 peer->ibp_reconnecting = 1;
2648 peer->ibp_version = version;
2650 peer->ibp_incarnation = incarnation;
2652 write_unlock_irqrestore(glock, flags);
2654 CNETERR("%s: %s (%s), %x, %x, msg_size: %d, queue_depth: %d/%d, max_frags: %d/%d\n",
2655 libcfs_nid2str(peer->ibp_nid),
2656 reconnect ? "reconnect" : "don't reconnect",
2657 reason, IBLND_MSG_VERSION, version, msg_size,
2658 conn->ibc_queue_depth, queue_dep,
2659 conn->ibc_max_frags, frag_num);
2661 * if conn::ibc_reconnect is TRUE, connd will reconnect to the peer
2662 * while destroying the zombie
2667 kiblnd_rejected(struct kib_conn *conn, int reason, void *priv, int priv_nob)
2669 struct kib_peer *peer = conn->ibc_peer;
2671 LASSERT(!in_interrupt());
2672 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2675 case IB_CM_REJ_STALE_CONN:
2676 kiblnd_check_reconnect(conn, IBLND_MSG_VERSION, 0,
2677 IBLND_REJECT_CONN_STALE, NULL);
2680 case IB_CM_REJ_INVALID_SERVICE_ID:
2681 CNETERR("%s rejected: no listener at %d\n",
2682 libcfs_nid2str(peer->ibp_nid),
2683 *kiblnd_tunables.kib_service);
2686 case IB_CM_REJ_CONSUMER_DEFINED:
2687 if (priv_nob >= offsetof(struct kib_rej, ibr_padding)) {
2688 struct kib_rej *rej = priv;
2689 struct kib_connparams *cp = NULL;
2691 __u64 incarnation = -1;
2693 /* NB. default incarnation is -1 because:
2694 * a) V1 will ignore dst incarnation in connreq.
2695 * b) V2 will provide incarnation while rejecting me,
2696 * -1 will be overwrote.
2698 * if I try to connect to a V1 peer with V2 protocol,
2699 * it rejected me then upgrade to V2, I have no idea
2700 * about the upgrading and try to reconnect with V1,
2701 * in this case upgraded V2 can find out I'm trying to
2702 * talk to the old guy and reject me(incarnation is -1).
2705 if (rej->ibr_magic == __swab32(IBLND_MSG_MAGIC) ||
2706 rej->ibr_magic == __swab32(LNET_PROTO_MAGIC)) {
2707 __swab32s(&rej->ibr_magic);
2708 __swab16s(&rej->ibr_version);
2712 if (priv_nob >= sizeof(struct kib_rej) &&
2713 rej->ibr_version > IBLND_MSG_VERSION_1) {
2715 * priv_nob is always 148 in current version
2716 * of OFED, so we still need to check version.
2717 * (define of IB_CM_REJ_PRIVATE_DATA_SIZE)
2722 __swab64s(&rej->ibr_incarnation);
2723 __swab16s(&cp->ibcp_queue_depth);
2724 __swab16s(&cp->ibcp_max_frags);
2725 __swab32s(&cp->ibcp_max_msg_size);
2728 incarnation = rej->ibr_incarnation;
2731 if (rej->ibr_magic != IBLND_MSG_MAGIC &&
2732 rej->ibr_magic != LNET_PROTO_MAGIC) {
2733 CERROR("%s rejected: consumer defined fatal error\n",
2734 libcfs_nid2str(peer->ibp_nid));
2738 if (rej->ibr_version != IBLND_MSG_VERSION &&
2739 rej->ibr_version != IBLND_MSG_VERSION_1) {
2740 CERROR("%s rejected: o2iblnd version %x error\n",
2741 libcfs_nid2str(peer->ibp_nid),
2746 if (rej->ibr_why == IBLND_REJECT_FATAL &&
2747 rej->ibr_version == IBLND_MSG_VERSION_1) {
2748 CDEBUG(D_NET, "rejected by old version peer %s: %x\n",
2749 libcfs_nid2str(peer->ibp_nid), rej->ibr_version);
2751 if (conn->ibc_version != IBLND_MSG_VERSION_1)
2752 rej->ibr_why = IBLND_REJECT_CONN_UNCOMPAT;
2755 switch (rej->ibr_why) {
2756 case IBLND_REJECT_CONN_RACE:
2757 case IBLND_REJECT_CONN_STALE:
2758 case IBLND_REJECT_CONN_UNCOMPAT:
2759 case IBLND_REJECT_MSG_QUEUE_SIZE:
2760 case IBLND_REJECT_RDMA_FRAGS:
2761 kiblnd_check_reconnect(conn, rej->ibr_version,
2766 case IBLND_REJECT_NO_RESOURCES:
2767 CERROR("%s rejected: o2iblnd no resources\n",
2768 libcfs_nid2str(peer->ibp_nid));
2771 case IBLND_REJECT_FATAL:
2772 CERROR("%s rejected: o2iblnd fatal error\n",
2773 libcfs_nid2str(peer->ibp_nid));
2777 CERROR("%s rejected: o2iblnd reason %d\n",
2778 libcfs_nid2str(peer->ibp_nid),
2786 CNETERR("%s rejected: reason %d, size %d\n",
2787 libcfs_nid2str(peer->ibp_nid), reason, priv_nob);
2791 kiblnd_connreq_done(conn, -ECONNREFUSED);
2795 kiblnd_check_connreply(struct kib_conn *conn, void *priv, int priv_nob)
2797 struct kib_peer *peer = conn->ibc_peer;
2798 lnet_ni_t *ni = peer->ibp_ni;
2799 struct kib_net *net = ni->ni_data;
2800 struct kib_msg *msg = priv;
2801 int ver = conn->ibc_version;
2802 int rc = kiblnd_unpack_msg(msg, priv_nob);
2803 unsigned long flags;
2808 CERROR("Can't unpack connack from %s: %d\n",
2809 libcfs_nid2str(peer->ibp_nid), rc);
2813 if (msg->ibm_type != IBLND_MSG_CONNACK) {
2814 CERROR("Unexpected message %d from %s\n",
2815 msg->ibm_type, libcfs_nid2str(peer->ibp_nid));
2820 if (ver != msg->ibm_version) {
2821 CERROR("%s replied version %x is different with requested version %x\n",
2822 libcfs_nid2str(peer->ibp_nid), msg->ibm_version, ver);
2827 if (msg->ibm_u.connparams.ibcp_queue_depth >
2828 conn->ibc_queue_depth) {
2829 CERROR("%s has incompatible queue depth %d (<=%d wanted)\n",
2830 libcfs_nid2str(peer->ibp_nid),
2831 msg->ibm_u.connparams.ibcp_queue_depth,
2832 conn->ibc_queue_depth);
2837 if ((msg->ibm_u.connparams.ibcp_max_frags >> IBLND_FRAG_SHIFT) >
2838 conn->ibc_max_frags) {
2839 CERROR("%s has incompatible max_frags %d (<=%d wanted)\n",
2840 libcfs_nid2str(peer->ibp_nid),
2841 msg->ibm_u.connparams.ibcp_max_frags >> IBLND_FRAG_SHIFT,
2842 conn->ibc_max_frags);
2847 if (msg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2848 CERROR("%s max message size %d too big (%d max)\n",
2849 libcfs_nid2str(peer->ibp_nid),
2850 msg->ibm_u.connparams.ibcp_max_msg_size,
2856 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2857 if (msg->ibm_dstnid == ni->ni_nid &&
2858 msg->ibm_dststamp == net->ibn_incarnation)
2862 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2865 CERROR("Bad connection reply from %s, rc = %d, version: %x max_frags: %d\n",
2866 libcfs_nid2str(peer->ibp_nid), rc,
2867 msg->ibm_version, msg->ibm_u.connparams.ibcp_max_frags);
2871 conn->ibc_incarnation = msg->ibm_srcstamp;
2872 conn->ibc_credits = msg->ibm_u.connparams.ibcp_queue_depth;
2873 conn->ibc_reserved_credits = msg->ibm_u.connparams.ibcp_queue_depth;
2874 conn->ibc_queue_depth = msg->ibm_u.connparams.ibcp_queue_depth;
2875 conn->ibc_max_frags = msg->ibm_u.connparams.ibcp_max_frags >> IBLND_FRAG_SHIFT;
2876 LASSERT(conn->ibc_credits + conn->ibc_reserved_credits +
2877 IBLND_OOB_MSGS(ver) <= IBLND_RX_MSGS(conn));
2879 kiblnd_connreq_done(conn, 0);
2884 * NB My QP has already established itself, so I handle anything going
2885 * wrong here by setting ibc_comms_error.
2886 * kiblnd_connreq_done(0) moves the conn state to ESTABLISHED, but then
2887 * immediately tears it down.
2890 conn->ibc_comms_error = rc;
2891 kiblnd_connreq_done(conn, 0);
2895 kiblnd_active_connect(struct rdma_cm_id *cmid)
2897 struct kib_peer *peer = (struct kib_peer *)cmid->context;
2898 struct kib_conn *conn;
2899 struct kib_msg *msg;
2900 struct rdma_conn_param cp;
2903 unsigned long flags;
2906 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2908 incarnation = peer->ibp_incarnation;
2909 version = !peer->ibp_version ? IBLND_MSG_VERSION :
2912 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2914 conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_ACTIVE_CONNECT,
2917 kiblnd_peer_connect_failed(peer, 1, -ENOMEM);
2918 kiblnd_peer_decref(peer); /* lose cmid's ref */
2923 * conn "owns" cmid now, so I return success from here on to ensure the
2924 * CM callback doesn't destroy cmid. conn also takes over cmid's ref
2927 msg = &conn->ibc_connvars->cv_msg;
2929 memset(msg, 0, sizeof(*msg));
2930 kiblnd_init_msg(msg, IBLND_MSG_CONNREQ, sizeof(msg->ibm_u.connparams));
2931 msg->ibm_u.connparams.ibcp_queue_depth = conn->ibc_queue_depth;
2932 msg->ibm_u.connparams.ibcp_max_frags = conn->ibc_max_frags << IBLND_FRAG_SHIFT;
2933 msg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2935 kiblnd_pack_msg(peer->ibp_ni, msg, version,
2936 0, peer->ibp_nid, incarnation);
2938 memset(&cp, 0, sizeof(cp));
2939 cp.private_data = msg;
2940 cp.private_data_len = msg->ibm_nob;
2941 cp.responder_resources = 0; /* No atomic ops or RDMA reads */
2942 cp.initiator_depth = 0;
2943 cp.flow_control = 1;
2944 cp.retry_count = *kiblnd_tunables.kib_retry_count;
2945 cp.rnr_retry_count = *kiblnd_tunables.kib_rnr_retry_count;
2947 LASSERT(cmid->context == (void *)conn);
2948 LASSERT(conn->ibc_cmid == cmid);
2950 rc = rdma_connect(cmid, &cp);
2952 CERROR("Can't connect to %s: %d\n",
2953 libcfs_nid2str(peer->ibp_nid), rc);
2954 kiblnd_connreq_done(conn, rc);
2955 kiblnd_conn_decref(conn);
2962 kiblnd_cm_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2964 struct kib_peer *peer;
2965 struct kib_conn *conn;
2968 switch (event->event) {
2970 CERROR("Unexpected event: %d, status: %d\n",
2971 event->event, event->status);
2974 case RDMA_CM_EVENT_CONNECT_REQUEST:
2975 /* destroy cmid on failure */
2976 rc = kiblnd_passive_connect(cmid,
2977 (void *)KIBLND_CONN_PARAM(event),
2978 KIBLND_CONN_PARAM_LEN(event));
2979 CDEBUG(D_NET, "connreq: %d\n", rc);
2982 case RDMA_CM_EVENT_ADDR_ERROR:
2983 peer = (struct kib_peer *)cmid->context;
2984 CNETERR("%s: ADDR ERROR %d\n",
2985 libcfs_nid2str(peer->ibp_nid), event->status);
2986 kiblnd_peer_connect_failed(peer, 1, -EHOSTUNREACH);
2987 kiblnd_peer_decref(peer);
2988 return -EHOSTUNREACH; /* rc destroys cmid */
2990 case RDMA_CM_EVENT_ADDR_RESOLVED:
2991 peer = (struct kib_peer *)cmid->context;
2993 CDEBUG(D_NET, "%s Addr resolved: %d\n",
2994 libcfs_nid2str(peer->ibp_nid), event->status);
2996 if (event->status) {
2997 CNETERR("Can't resolve address for %s: %d\n",
2998 libcfs_nid2str(peer->ibp_nid), event->status);
3001 rc = rdma_resolve_route(
3002 cmid, *kiblnd_tunables.kib_timeout * 1000);
3004 struct kib_net *net = peer->ibp_ni->ni_data;
3005 struct kib_dev *dev = net->ibn_dev;
3007 CDEBUG(D_NET, "%s: connection bound to "\
3009 libcfs_nid2str(peer->ibp_nid),
3011 &dev->ibd_ifip, cmid->device->name);
3016 /* Can't initiate route resolution */
3017 CERROR("Can't resolve route for %s: %d\n",
3018 libcfs_nid2str(peer->ibp_nid), rc);
3020 kiblnd_peer_connect_failed(peer, 1, rc);
3021 kiblnd_peer_decref(peer);
3022 return rc; /* rc destroys cmid */
3024 case RDMA_CM_EVENT_ROUTE_ERROR:
3025 peer = (struct kib_peer *)cmid->context;
3026 CNETERR("%s: ROUTE ERROR %d\n",
3027 libcfs_nid2str(peer->ibp_nid), event->status);
3028 kiblnd_peer_connect_failed(peer, 1, -EHOSTUNREACH);
3029 kiblnd_peer_decref(peer);
3030 return -EHOSTUNREACH; /* rc destroys cmid */
3032 case RDMA_CM_EVENT_ROUTE_RESOLVED:
3033 peer = (struct kib_peer *)cmid->context;
3034 CDEBUG(D_NET, "%s Route resolved: %d\n",
3035 libcfs_nid2str(peer->ibp_nid), event->status);
3038 return kiblnd_active_connect(cmid);
3040 CNETERR("Can't resolve route for %s: %d\n",
3041 libcfs_nid2str(peer->ibp_nid), event->status);
3042 kiblnd_peer_connect_failed(peer, 1, event->status);
3043 kiblnd_peer_decref(peer);
3044 return event->status; /* rc destroys cmid */
3046 case RDMA_CM_EVENT_UNREACHABLE:
3047 conn = (struct kib_conn *)cmid->context;
3048 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
3049 conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
3050 CNETERR("%s: UNREACHABLE %d\n",
3051 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
3052 kiblnd_connreq_done(conn, -ENETDOWN);
3053 kiblnd_conn_decref(conn);
3056 case RDMA_CM_EVENT_CONNECT_ERROR:
3057 conn = (struct kib_conn *)cmid->context;
3058 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
3059 conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
3060 CNETERR("%s: CONNECT ERROR %d\n",
3061 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
3062 kiblnd_connreq_done(conn, -ENOTCONN);
3063 kiblnd_conn_decref(conn);
3066 case RDMA_CM_EVENT_REJECTED:
3067 conn = (struct kib_conn *)cmid->context;
3068 switch (conn->ibc_state) {
3072 case IBLND_CONN_PASSIVE_WAIT:
3073 CERROR("%s: REJECTED %d\n",
3074 libcfs_nid2str(conn->ibc_peer->ibp_nid),
3076 kiblnd_connreq_done(conn, -ECONNRESET);
3079 case IBLND_CONN_ACTIVE_CONNECT:
3080 kiblnd_rejected(conn, event->status,
3081 (void *)KIBLND_CONN_PARAM(event),
3082 KIBLND_CONN_PARAM_LEN(event));
3085 kiblnd_conn_decref(conn);
3088 case RDMA_CM_EVENT_ESTABLISHED:
3089 conn = (struct kib_conn *)cmid->context;
3090 switch (conn->ibc_state) {
3094 case IBLND_CONN_PASSIVE_WAIT:
3095 CDEBUG(D_NET, "ESTABLISHED (passive): %s\n",
3096 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3097 kiblnd_connreq_done(conn, 0);
3100 case IBLND_CONN_ACTIVE_CONNECT:
3101 CDEBUG(D_NET, "ESTABLISHED(active): %s\n",
3102 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3103 kiblnd_check_connreply(conn,
3104 (void *)KIBLND_CONN_PARAM(event),
3105 KIBLND_CONN_PARAM_LEN(event));
3108 /* net keeps its ref on conn! */
3111 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
3112 CDEBUG(D_NET, "Ignore TIMEWAIT_EXIT event\n");
3114 case RDMA_CM_EVENT_DISCONNECTED:
3115 conn = (struct kib_conn *)cmid->context;
3116 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
3117 CERROR("%s DISCONNECTED\n",
3118 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3119 kiblnd_connreq_done(conn, -ECONNRESET);
3121 kiblnd_close_conn(conn, 0);
3123 kiblnd_conn_decref(conn);
3124 cmid->context = NULL;
3127 case RDMA_CM_EVENT_DEVICE_REMOVAL:
3128 LCONSOLE_ERROR_MSG(0x131,
3129 "Received notification of device removal\n"
3130 "Please shutdown LNET to allow this to proceed\n");
3132 * Can't remove network from underneath LNET for now, so I have
3137 case RDMA_CM_EVENT_ADDR_CHANGE:
3138 LCONSOLE_INFO("Physical link changed (eg hca/port)\n");
3144 kiblnd_check_txs_locked(struct kib_conn *conn, struct list_head *txs)
3147 struct list_head *ttmp;
3149 list_for_each(ttmp, txs) {
3150 tx = list_entry(ttmp, struct kib_tx, tx_list);
3152 if (txs != &conn->ibc_active_txs) {
3153 LASSERT(tx->tx_queued);
3155 LASSERT(!tx->tx_queued);
3156 LASSERT(tx->tx_waiting || tx->tx_sending);
3159 if (cfs_time_aftereq(jiffies, tx->tx_deadline)) {
3160 CERROR("Timed out tx: %s, %lu seconds\n",
3161 kiblnd_queue2str(conn, txs),
3162 cfs_duration_sec(jiffies - tx->tx_deadline));
3171 kiblnd_conn_timed_out_locked(struct kib_conn *conn)
3173 return kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue) ||
3174 kiblnd_check_txs_locked(conn, &conn->ibc_tx_noops) ||
3175 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_rsrvd) ||
3176 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_nocred) ||
3177 kiblnd_check_txs_locked(conn, &conn->ibc_active_txs);
3181 kiblnd_check_conns(int idx)
3184 LIST_HEAD(checksends);
3185 struct list_head *peers = &kiblnd_data.kib_peers[idx];
3186 struct list_head *ptmp;
3187 struct kib_peer *peer;
3188 struct kib_conn *conn;
3189 struct kib_conn *temp;
3190 struct kib_conn *tmp;
3191 struct list_head *ctmp;
3192 unsigned long flags;
3195 * NB. We expect to have a look at all the peers and not find any
3196 * RDMAs to time out, so we just use a shared lock while we
3199 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3201 list_for_each(ptmp, peers) {
3202 peer = list_entry(ptmp, struct kib_peer, ibp_list);
3204 list_for_each(ctmp, &peer->ibp_conns) {
3208 conn = list_entry(ctmp, struct kib_conn, ibc_list);
3210 LASSERT(conn->ibc_state == IBLND_CONN_ESTABLISHED);
3212 spin_lock(&conn->ibc_lock);
3214 sendnoop = kiblnd_need_noop(conn);
3215 timedout = kiblnd_conn_timed_out_locked(conn);
3216 if (!sendnoop && !timedout) {
3217 spin_unlock(&conn->ibc_lock);
3222 CERROR("Timed out RDMA with %s (%lu): c: %u, oc: %u, rc: %u\n",
3223 libcfs_nid2str(peer->ibp_nid),
3224 cfs_duration_sec(cfs_time_current() -
3225 peer->ibp_last_alive),
3227 conn->ibc_outstanding_credits,
3228 conn->ibc_reserved_credits);
3229 list_add(&conn->ibc_connd_list, &closes);
3231 list_add(&conn->ibc_connd_list, &checksends);
3233 /* +ref for 'closes' or 'checksends' */
3234 kiblnd_conn_addref(conn);
3236 spin_unlock(&conn->ibc_lock);
3240 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3243 * Handle timeout by closing the whole
3244 * connection. We can only be sure RDMA activity
3245 * has ceased once the QP has been modified.
3247 list_for_each_entry_safe(conn, tmp, &closes, ibc_connd_list) {
3248 list_del(&conn->ibc_connd_list);
3249 kiblnd_close_conn(conn, -ETIMEDOUT);
3250 kiblnd_conn_decref(conn);
3254 * In case we have enough credits to return via a
3255 * NOOP, but there were no non-blocking tx descs
3256 * free to do it last time...
3258 list_for_each_entry_safe(conn, temp, &checksends, ibc_connd_list) {
3259 list_del(&conn->ibc_connd_list);
3261 spin_lock(&conn->ibc_lock);
3262 kiblnd_check_sends_locked(conn);
3263 spin_unlock(&conn->ibc_lock);
3265 kiblnd_conn_decref(conn);
3270 kiblnd_disconnect_conn(struct kib_conn *conn)
3272 LASSERT(!in_interrupt());
3273 LASSERT(current == kiblnd_data.kib_connd);
3274 LASSERT(conn->ibc_state == IBLND_CONN_CLOSING);
3276 rdma_disconnect(conn->ibc_cmid);
3277 kiblnd_finalise_conn(conn);
3279 kiblnd_peer_notify(conn->ibc_peer);
3283 * High-water for reconnection to the same peer, reconnection attempt should
3284 * be delayed after trying more than KIB_RECONN_HIGH_RACE.
3286 #define KIB_RECONN_HIGH_RACE 10
3288 * Allow connd to take a break and handle other things after consecutive
3289 * reconnection attemps.
3291 #define KIB_RECONN_BREAK 100
3294 kiblnd_connd(void *arg)
3296 spinlock_t *lock= &kiblnd_data.kib_connd_lock;
3298 unsigned long flags;
3299 struct kib_conn *conn;
3304 unsigned long deadline = jiffies;
3306 cfs_block_allsigs();
3308 init_waitqueue_entry(&wait, current);
3309 kiblnd_data.kib_connd = current;
3311 spin_lock_irqsave(lock, flags);
3313 while (!kiblnd_data.kib_shutdown) {
3318 if (!list_empty(&kiblnd_data.kib_connd_zombies)) {
3319 struct kib_peer *peer = NULL;
3321 conn = list_entry(kiblnd_data.kib_connd_zombies.next,
3322 struct kib_conn, ibc_list);
3323 list_del(&conn->ibc_list);
3324 if (conn->ibc_reconnect) {
3325 peer = conn->ibc_peer;
3326 kiblnd_peer_addref(peer);
3329 spin_unlock_irqrestore(lock, flags);
3332 kiblnd_destroy_conn(conn);
3334 spin_lock_irqsave(lock, flags);
3340 conn->ibc_peer = peer;
3341 if (peer->ibp_reconnected < KIB_RECONN_HIGH_RACE)
3342 list_add_tail(&conn->ibc_list,
3343 &kiblnd_data.kib_reconn_list);
3345 list_add_tail(&conn->ibc_list,
3346 &kiblnd_data.kib_reconn_wait);
3349 if (!list_empty(&kiblnd_data.kib_connd_conns)) {
3350 conn = list_entry(kiblnd_data.kib_connd_conns.next,
3351 struct kib_conn, ibc_list);
3352 list_del(&conn->ibc_list);
3354 spin_unlock_irqrestore(lock, flags);
3357 kiblnd_disconnect_conn(conn);
3358 kiblnd_conn_decref(conn);
3360 spin_lock_irqsave(lock, flags);
3363 while (reconn < KIB_RECONN_BREAK) {
3364 if (kiblnd_data.kib_reconn_sec !=
3365 ktime_get_real_seconds()) {
3366 kiblnd_data.kib_reconn_sec = ktime_get_real_seconds();
3367 list_splice_init(&kiblnd_data.kib_reconn_wait,
3368 &kiblnd_data.kib_reconn_list);
3371 if (list_empty(&kiblnd_data.kib_reconn_list))
3374 conn = list_entry(kiblnd_data.kib_reconn_list.next,
3375 struct kib_conn, ibc_list);
3376 list_del(&conn->ibc_list);
3378 spin_unlock_irqrestore(lock, flags);
3381 reconn += kiblnd_reconnect_peer(conn->ibc_peer);
3382 kiblnd_peer_decref(conn->ibc_peer);
3383 LIBCFS_FREE(conn, sizeof(*conn));
3385 spin_lock_irqsave(lock, flags);
3388 /* careful with the jiffy wrap... */
3389 timeout = (int)(deadline - jiffies);
3393 int chunk = kiblnd_data.kib_peer_hash_size;
3395 spin_unlock_irqrestore(lock, flags);
3399 * Time to check for RDMA timeouts on a few more
3400 * peers: I do checks every 'p' seconds on a
3401 * proportion of the peer table and I need to check
3402 * every connection 'n' times within a timeout
3403 * interval, to ensure I detect a timeout on any
3404 * connection within (n+1)/n times the timeout
3407 if (*kiblnd_tunables.kib_timeout > n * p)
3408 chunk = (chunk * n * p) /
3409 *kiblnd_tunables.kib_timeout;
3413 for (i = 0; i < chunk; i++) {
3414 kiblnd_check_conns(peer_index);
3415 peer_index = (peer_index + 1) %
3416 kiblnd_data.kib_peer_hash_size;
3419 deadline += msecs_to_jiffies(p * MSEC_PER_SEC);
3420 spin_lock_irqsave(lock, flags);
3426 /* Nothing to do for 'timeout' */
3427 set_current_state(TASK_INTERRUPTIBLE);
3428 add_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3429 spin_unlock_irqrestore(lock, flags);
3431 schedule_timeout(timeout);
3433 remove_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3434 spin_lock_irqsave(lock, flags);
3437 spin_unlock_irqrestore(lock, flags);
3439 kiblnd_thread_fini();
3444 kiblnd_qp_event(struct ib_event *event, void *arg)
3446 struct kib_conn *conn = arg;
3448 switch (event->event) {
3449 case IB_EVENT_COMM_EST:
3450 CDEBUG(D_NET, "%s established\n",
3451 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3453 * We received a packet but connection isn't established
3454 * probably handshake packet was lost, so free to
3455 * force make connection established
3457 rdma_notify(conn->ibc_cmid, IB_EVENT_COMM_EST);
3461 CERROR("%s: Async QP event type %d\n",
3462 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3468 kiblnd_complete(struct ib_wc *wc)
3470 switch (kiblnd_wreqid2type(wc->wr_id)) {
3475 if (wc->status != IB_WC_SUCCESS &&
3476 wc->status != IB_WC_WR_FLUSH_ERR)
3477 CNETERR("FastReg failed: %d\n", wc->status);
3480 case IBLND_WID_RDMA:
3482 * We only get RDMA completion notification if it fails. All
3483 * subsequent work items, including the final SEND will fail
3484 * too. However we can't print out any more info about the
3485 * failing RDMA because 'tx' might be back on the idle list or
3486 * even reused already if we didn't manage to post all our work
3489 CNETERR("RDMA (tx: %p) failed: %d\n",
3490 kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3494 kiblnd_tx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3498 kiblnd_rx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status,
3505 kiblnd_cq_completion(struct ib_cq *cq, void *arg)
3508 * NB I'm not allowed to schedule this conn once its refcount has
3509 * reached 0. Since fundamentally I'm racing with scheduler threads
3510 * consuming my CQ I could be called after all completions have
3511 * occurred. But in this case, !ibc_nrx && !ibc_nsends_posted
3512 * and this CQ is about to be destroyed so I NOOP.
3514 struct kib_conn *conn = arg;
3515 struct kib_sched_info *sched = conn->ibc_sched;
3516 unsigned long flags;
3518 LASSERT(cq == conn->ibc_cq);
3520 spin_lock_irqsave(&sched->ibs_lock, flags);
3522 conn->ibc_ready = 1;
3524 if (!conn->ibc_scheduled &&
3525 (conn->ibc_nrx > 0 ||
3526 conn->ibc_nsends_posted > 0)) {
3527 kiblnd_conn_addref(conn); /* +1 ref for sched_conns */
3528 conn->ibc_scheduled = 1;
3529 list_add_tail(&conn->ibc_sched_list, &sched->ibs_conns);
3531 if (waitqueue_active(&sched->ibs_waitq))
3532 wake_up(&sched->ibs_waitq);
3535 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3539 kiblnd_cq_event(struct ib_event *event, void *arg)
3541 struct kib_conn *conn = arg;
3543 CERROR("%s: async CQ event type %d\n",
3544 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3548 kiblnd_scheduler(void *arg)
3550 long id = (long)arg;
3551 struct kib_sched_info *sched;
3552 struct kib_conn *conn;
3554 unsigned long flags;
3560 cfs_block_allsigs();
3562 init_waitqueue_entry(&wait, current);
3564 sched = kiblnd_data.kib_scheds[KIB_THREAD_CPT(id)];
3566 rc = cfs_cpt_bind(lnet_cpt_table(), sched->ibs_cpt);
3568 CWARN("Failed to bind on CPT %d, please verify whether all CPUs are healthy and reload modules if necessary, otherwise your system might under risk of low performance\n",
3572 spin_lock_irqsave(&sched->ibs_lock, flags);
3574 while (!kiblnd_data.kib_shutdown) {
3575 if (busy_loops++ >= IBLND_RESCHED) {
3576 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3581 spin_lock_irqsave(&sched->ibs_lock, flags);
3586 if (!list_empty(&sched->ibs_conns)) {
3587 conn = list_entry(sched->ibs_conns.next, struct kib_conn,
3589 /* take over kib_sched_conns' ref on conn... */
3590 LASSERT(conn->ibc_scheduled);
3591 list_del(&conn->ibc_sched_list);
3592 conn->ibc_ready = 0;
3594 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3596 wc.wr_id = IBLND_WID_INVAL;
3598 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3600 rc = ib_req_notify_cq(conn->ibc_cq,
3603 CWARN("%s: ib_req_notify_cq failed: %d, closing connection\n",
3604 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
3605 kiblnd_close_conn(conn, -EIO);
3606 kiblnd_conn_decref(conn);
3607 spin_lock_irqsave(&sched->ibs_lock,
3612 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3615 if (unlikely(rc > 0 && wc.wr_id == IBLND_WID_INVAL)) {
3616 LCONSOLE_ERROR("ib_poll_cq (rc: %d) returned invalid wr_id, opcode %d, status: %d, vendor_err: %d, conn: %s status: %d\nplease upgrade firmware and OFED or contact vendor.\n",
3617 rc, wc.opcode, wc.status,
3619 libcfs_nid2str(conn->ibc_peer->ibp_nid),
3625 CWARN("%s: ib_poll_cq failed: %d, closing connection\n",
3626 libcfs_nid2str(conn->ibc_peer->ibp_nid),
3628 kiblnd_close_conn(conn, -EIO);
3629 kiblnd_conn_decref(conn);
3630 spin_lock_irqsave(&sched->ibs_lock, flags);
3634 spin_lock_irqsave(&sched->ibs_lock, flags);
3636 if (rc || conn->ibc_ready) {
3638 * There may be another completion waiting; get
3639 * another scheduler to check while I handle
3642 /* +1 ref for sched_conns */
3643 kiblnd_conn_addref(conn);
3644 list_add_tail(&conn->ibc_sched_list,
3646 if (waitqueue_active(&sched->ibs_waitq))
3647 wake_up(&sched->ibs_waitq);
3649 conn->ibc_scheduled = 0;
3653 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3654 kiblnd_complete(&wc);
3656 spin_lock_irqsave(&sched->ibs_lock, flags);
3659 kiblnd_conn_decref(conn); /* ...drop my ref from above */
3666 set_current_state(TASK_INTERRUPTIBLE);
3667 add_wait_queue_exclusive(&sched->ibs_waitq, &wait);
3668 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3673 remove_wait_queue(&sched->ibs_waitq, &wait);
3674 spin_lock_irqsave(&sched->ibs_lock, flags);
3677 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3679 kiblnd_thread_fini();
3684 kiblnd_failover_thread(void *arg)
3686 rwlock_t *glock = &kiblnd_data.kib_global_lock;
3687 struct kib_dev *dev;
3689 unsigned long flags;
3692 LASSERT(*kiblnd_tunables.kib_dev_failover);
3694 cfs_block_allsigs();
3696 init_waitqueue_entry(&wait, current);
3697 write_lock_irqsave(glock, flags);
3699 while (!kiblnd_data.kib_shutdown) {
3700 int do_failover = 0;
3703 list_for_each_entry(dev, &kiblnd_data.kib_failed_devs,
3705 if (time_before(cfs_time_current(),
3706 dev->ibd_next_failover))
3713 list_del_init(&dev->ibd_fail_list);
3714 dev->ibd_failover = 1;
3715 write_unlock_irqrestore(glock, flags);
3717 rc = kiblnd_dev_failover(dev);
3719 write_lock_irqsave(glock, flags);
3721 LASSERT(dev->ibd_failover);
3722 dev->ibd_failover = 0;
3723 if (rc >= 0) { /* Device is OK or failover succeed */
3724 dev->ibd_next_failover = cfs_time_shift(3);
3728 /* failed to failover, retry later */
3729 dev->ibd_next_failover =
3730 cfs_time_shift(min(dev->ibd_failed_failover, 10));
3731 if (kiblnd_dev_can_failover(dev)) {
3732 list_add_tail(&dev->ibd_fail_list,
3733 &kiblnd_data.kib_failed_devs);
3739 /* long sleep if no more pending failover */
3740 long_sleep = list_empty(&kiblnd_data.kib_failed_devs);
3742 set_current_state(TASK_INTERRUPTIBLE);
3743 add_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3744 write_unlock_irqrestore(glock, flags);
3746 rc = schedule_timeout(long_sleep ? cfs_time_seconds(10) :
3747 cfs_time_seconds(1));
3748 remove_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3749 write_lock_irqsave(glock, flags);
3751 if (!long_sleep || rc)
3755 * have a long sleep, routine check all active devices,
3756 * we need checking like this because if there is not active
3757 * connection on the dev and no SEND from local, we may listen
3758 * on wrong HCA for ever while there is a bonding failover
3760 list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3761 if (kiblnd_dev_can_failover(dev)) {
3762 list_add_tail(&dev->ibd_fail_list,
3763 &kiblnd_data.kib_failed_devs);
3768 write_unlock_irqrestore(glock, flags);
3770 kiblnd_thread_fini();