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 #define MAX_CONN_RACES_BEFORE_ABORT 20
45 static void kiblnd_peer_alive(kib_peer_t *peer);
46 static void kiblnd_peer_connect_failed(kib_peer_t *peer, int active, int error);
47 static void kiblnd_init_tx_msg(lnet_ni_t *ni, kib_tx_t *tx,
48 int type, int body_nob);
49 static int kiblnd_init_rdma(kib_conn_t *conn, kib_tx_t *tx, int type,
50 int resid, kib_rdma_desc_t *dstrd, __u64 dstcookie);
51 static void kiblnd_queue_tx_locked(kib_tx_t *tx, kib_conn_t *conn);
52 static void kiblnd_queue_tx(kib_tx_t *tx, kib_conn_t *conn);
53 static void kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx);
54 static void kiblnd_check_sends_locked(kib_conn_t *conn);
57 kiblnd_tx_done (lnet_ni_t *ni, kib_tx_t *tx)
59 lnet_msg_t *lntmsg[2];
60 kib_net_t *net = ni->ni_data;
64 LASSERT (net != NULL);
65 LASSERT (!in_interrupt());
66 LASSERT (!tx->tx_queued); /* mustn't be queued for sending */
67 LASSERT (tx->tx_sending == 0); /* mustn't be awaiting sent callback */
68 LASSERT (!tx->tx_waiting); /* mustn't be awaiting peer response */
69 LASSERT (tx->tx_pool != NULL);
71 kiblnd_unmap_tx(ni, tx);
73 /* tx may have up to 2 lnet msgs to finalise */
74 lntmsg[0] = tx->tx_lntmsg[0]; tx->tx_lntmsg[0] = NULL;
75 lntmsg[1] = tx->tx_lntmsg[1]; tx->tx_lntmsg[1] = NULL;
78 if (tx->tx_conn != NULL) {
79 LASSERT (ni == tx->tx_conn->ibc_peer->ibp_ni);
81 kiblnd_conn_decref(tx->tx_conn);
88 kiblnd_pool_free_node(&tx->tx_pool->tpo_pool, &tx->tx_list);
90 /* delay finalize until my descs have been freed */
91 for (i = 0; i < 2; i++) {
92 if (lntmsg[i] == NULL)
95 lnet_finalize(ni, lntmsg[i], rc);
100 kiblnd_txlist_done(lnet_ni_t *ni, struct list_head *txlist, int status)
104 while (!list_empty(txlist)) {
105 tx = list_entry(txlist->next, kib_tx_t, tx_list);
107 list_del(&tx->tx_list);
110 tx->tx_status = status;
111 kiblnd_tx_done(ni, tx);
116 kiblnd_get_idle_tx(lnet_ni_t *ni, lnet_nid_t target)
118 kib_net_t *net = (kib_net_t *)ni->ni_data;
119 struct list_head *node;
121 kib_tx_poolset_t *tps;
123 tps = net->ibn_tx_ps[lnet_cpt_of_nid(target)];
124 node = kiblnd_pool_alloc_node(&tps->tps_poolset);
127 tx = container_of(node, kib_tx_t, tx_list);
129 LASSERT (tx->tx_nwrq == 0);
130 LASSERT (!tx->tx_queued);
131 LASSERT (tx->tx_sending == 0);
132 LASSERT (!tx->tx_waiting);
133 LASSERT (tx->tx_status == 0);
134 LASSERT (tx->tx_conn == NULL);
135 LASSERT (tx->tx_lntmsg[0] == NULL);
136 LASSERT (tx->tx_lntmsg[1] == NULL);
137 LASSERT (tx->tx_nfrags == 0);
143 kiblnd_drop_rx(kib_rx_t *rx)
145 kib_conn_t *conn = rx->rx_conn;
146 struct kib_sched_info *sched = conn->ibc_sched;
149 spin_lock_irqsave(&sched->ibs_lock, flags);
150 LASSERT(conn->ibc_nrx > 0);
152 spin_unlock_irqrestore(&sched->ibs_lock, flags);
154 kiblnd_conn_decref(conn);
158 kiblnd_post_rx (kib_rx_t *rx, int credit)
160 kib_conn_t *conn = rx->rx_conn;
161 kib_net_t *net = conn->ibc_peer->ibp_ni->ni_data;
162 struct ib_recv_wr *bad_wrq = NULL;
163 struct ib_mr *mr = conn->ibc_hdev->ibh_mrs;
166 LASSERT (net != NULL);
167 LASSERT (!in_interrupt());
168 LASSERT (credit == IBLND_POSTRX_NO_CREDIT ||
169 credit == IBLND_POSTRX_PEER_CREDIT ||
170 credit == IBLND_POSTRX_RSRVD_CREDIT);
173 rx->rx_sge.lkey = mr->lkey;
174 rx->rx_sge.addr = rx->rx_msgaddr;
175 rx->rx_sge.length = IBLND_MSG_SIZE;
177 rx->rx_wrq.next = NULL;
178 rx->rx_wrq.sg_list = &rx->rx_sge;
179 rx->rx_wrq.num_sge = 1;
180 rx->rx_wrq.wr_id = kiblnd_ptr2wreqid(rx, IBLND_WID_RX);
182 LASSERT (conn->ibc_state >= IBLND_CONN_INIT);
183 LASSERT (rx->rx_nob >= 0); /* not posted */
185 if (conn->ibc_state > IBLND_CONN_ESTABLISHED) {
186 kiblnd_drop_rx(rx); /* No more posts for this rx */
190 rx->rx_nob = -1; /* flag posted */
192 /* NB: need an extra reference after ib_post_recv because we don't
193 * own this rx (and rx::rx_conn) anymore, LU-5678.
195 kiblnd_conn_addref(conn);
196 rc = ib_post_recv(conn->ibc_cmid->qp, &rx->rx_wrq, &bad_wrq);
197 if (unlikely(rc != 0)) {
198 CERROR("Can't post rx for %s: %d, bad_wrq: %p\n",
199 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc, bad_wrq);
203 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) /* Initial post */
206 if (unlikely(rc != 0)) {
207 kiblnd_close_conn(conn, rc);
208 kiblnd_drop_rx(rx); /* No more posts for this rx */
212 if (credit == IBLND_POSTRX_NO_CREDIT)
215 spin_lock(&conn->ibc_lock);
216 if (credit == IBLND_POSTRX_PEER_CREDIT)
217 conn->ibc_outstanding_credits++;
219 conn->ibc_reserved_credits++;
220 kiblnd_check_sends_locked(conn);
221 spin_unlock(&conn->ibc_lock);
224 kiblnd_conn_decref(conn);
229 kiblnd_find_waiting_tx_locked(kib_conn_t *conn, int txtype, __u64 cookie)
231 struct list_head *tmp;
233 list_for_each(tmp, &conn->ibc_active_txs) {
234 kib_tx_t *tx = list_entry(tmp, kib_tx_t, tx_list);
236 LASSERT(!tx->tx_queued);
237 LASSERT(tx->tx_sending != 0 || tx->tx_waiting);
239 if (tx->tx_cookie != cookie)
242 if (tx->tx_waiting &&
243 tx->tx_msg->ibm_type == txtype)
246 CWARN("Bad completion: %swaiting, type %x (wanted %x)\n",
247 tx->tx_waiting ? "" : "NOT ",
248 tx->tx_msg->ibm_type, txtype);
254 kiblnd_handle_completion(kib_conn_t *conn, int txtype, int status, __u64 cookie)
257 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
260 spin_lock(&conn->ibc_lock);
262 tx = kiblnd_find_waiting_tx_locked(conn, txtype, cookie);
264 spin_unlock(&conn->ibc_lock);
266 CWARN("Unmatched completion type %x cookie %#llx from %s\n",
267 txtype, cookie, libcfs_nid2str(conn->ibc_peer->ibp_nid));
268 kiblnd_close_conn(conn, -EPROTO);
272 if (tx->tx_status == 0) { /* success so far */
273 if (status < 0) { /* failed? */
274 tx->tx_status = status;
275 } else if (txtype == IBLND_MSG_GET_REQ) {
276 lnet_set_reply_msg_len(ni, tx->tx_lntmsg[1], status);
282 idle = !tx->tx_queued && (tx->tx_sending == 0);
284 list_del(&tx->tx_list);
286 spin_unlock(&conn->ibc_lock);
289 kiblnd_tx_done(ni, tx);
293 kiblnd_send_completion(kib_conn_t *conn, int type, int status, __u64 cookie)
295 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
296 kib_tx_t *tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
299 CERROR("Can't get tx for completion %x for %s\n",
300 type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
304 tx->tx_msg->ibm_u.completion.ibcm_status = status;
305 tx->tx_msg->ibm_u.completion.ibcm_cookie = cookie;
306 kiblnd_init_tx_msg(ni, tx, type, sizeof(kib_completion_msg_t));
308 kiblnd_queue_tx(tx, conn);
312 kiblnd_handle_rx (kib_rx_t *rx)
314 kib_msg_t *msg = rx->rx_msg;
315 kib_conn_t *conn = rx->rx_conn;
316 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
317 int credits = msg->ibm_credits;
323 LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
325 CDEBUG (D_NET, "Received %x[%d] from %s\n",
326 msg->ibm_type, credits,
327 libcfs_nid2str(conn->ibc_peer->ibp_nid));
330 /* Have I received credits that will let me send? */
331 spin_lock(&conn->ibc_lock);
333 if (conn->ibc_credits + credits >
334 conn->ibc_queue_depth) {
335 rc2 = conn->ibc_credits;
336 spin_unlock(&conn->ibc_lock);
338 CERROR("Bad credits from %s: %d + %d > %d\n",
339 libcfs_nid2str(conn->ibc_peer->ibp_nid),
341 conn->ibc_queue_depth);
343 kiblnd_close_conn(conn, -EPROTO);
344 kiblnd_post_rx(rx, IBLND_POSTRX_NO_CREDIT);
348 conn->ibc_credits += credits;
350 /* This ensures the credit taken by NOOP can be returned */
351 if (msg->ibm_type == IBLND_MSG_NOOP &&
352 !IBLND_OOB_CAPABLE(conn->ibc_version)) /* v1 only */
353 conn->ibc_outstanding_credits++;
355 kiblnd_check_sends_locked(conn);
356 spin_unlock(&conn->ibc_lock);
359 switch (msg->ibm_type) {
361 CERROR("Bad IBLND message type %x from %s\n",
362 msg->ibm_type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
363 post_credit = IBLND_POSTRX_NO_CREDIT;
368 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
369 post_credit = IBLND_POSTRX_NO_CREDIT;
373 if (credits != 0) /* credit already posted */
374 post_credit = IBLND_POSTRX_NO_CREDIT;
375 else /* a keepalive NOOP */
376 post_credit = IBLND_POSTRX_PEER_CREDIT;
379 case IBLND_MSG_IMMEDIATE:
380 post_credit = IBLND_POSTRX_DONT_POST;
381 rc = lnet_parse(ni, &msg->ibm_u.immediate.ibim_hdr,
382 msg->ibm_srcnid, rx, 0);
383 if (rc < 0) /* repost on error */
384 post_credit = IBLND_POSTRX_PEER_CREDIT;
387 case IBLND_MSG_PUT_REQ:
388 post_credit = IBLND_POSTRX_DONT_POST;
389 rc = lnet_parse(ni, &msg->ibm_u.putreq.ibprm_hdr,
390 msg->ibm_srcnid, rx, 1);
391 if (rc < 0) /* repost on error */
392 post_credit = IBLND_POSTRX_PEER_CREDIT;
395 case IBLND_MSG_PUT_NAK:
396 CWARN ("PUT_NACK from %s\n",
397 libcfs_nid2str(conn->ibc_peer->ibp_nid));
398 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
399 kiblnd_handle_completion(conn, IBLND_MSG_PUT_REQ,
400 msg->ibm_u.completion.ibcm_status,
401 msg->ibm_u.completion.ibcm_cookie);
404 case IBLND_MSG_PUT_ACK:
405 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
407 spin_lock(&conn->ibc_lock);
408 tx = kiblnd_find_waiting_tx_locked(conn, IBLND_MSG_PUT_REQ,
409 msg->ibm_u.putack.ibpam_src_cookie);
411 list_del(&tx->tx_list);
412 spin_unlock(&conn->ibc_lock);
415 CERROR("Unmatched PUT_ACK from %s\n",
416 libcfs_nid2str(conn->ibc_peer->ibp_nid));
421 LASSERT (tx->tx_waiting);
422 /* CAVEAT EMPTOR: I could be racing with tx_complete, but...
423 * (a) I can overwrite tx_msg since my peer has received it!
424 * (b) tx_waiting set tells tx_complete() it's not done. */
426 tx->tx_nwrq = 0; /* overwrite PUT_REQ */
428 rc2 = kiblnd_init_rdma(conn, tx, IBLND_MSG_PUT_DONE,
429 kiblnd_rd_size(&msg->ibm_u.putack.ibpam_rd),
430 &msg->ibm_u.putack.ibpam_rd,
431 msg->ibm_u.putack.ibpam_dst_cookie);
433 CERROR("Can't setup rdma for PUT to %s: %d\n",
434 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc2);
436 spin_lock(&conn->ibc_lock);
437 tx->tx_waiting = 0; /* clear waiting and queue atomically */
438 kiblnd_queue_tx_locked(tx, conn);
439 spin_unlock(&conn->ibc_lock);
442 case IBLND_MSG_PUT_DONE:
443 post_credit = IBLND_POSTRX_PEER_CREDIT;
444 kiblnd_handle_completion(conn, IBLND_MSG_PUT_ACK,
445 msg->ibm_u.completion.ibcm_status,
446 msg->ibm_u.completion.ibcm_cookie);
449 case IBLND_MSG_GET_REQ:
450 post_credit = IBLND_POSTRX_DONT_POST;
451 rc = lnet_parse(ni, &msg->ibm_u.get.ibgm_hdr,
452 msg->ibm_srcnid, rx, 1);
453 if (rc < 0) /* repost on error */
454 post_credit = IBLND_POSTRX_PEER_CREDIT;
457 case IBLND_MSG_GET_DONE:
458 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
459 kiblnd_handle_completion(conn, IBLND_MSG_GET_REQ,
460 msg->ibm_u.completion.ibcm_status,
461 msg->ibm_u.completion.ibcm_cookie);
465 if (rc < 0) /* protocol error */
466 kiblnd_close_conn(conn, rc);
468 if (post_credit != IBLND_POSTRX_DONT_POST)
469 kiblnd_post_rx(rx, post_credit);
473 kiblnd_rx_complete (kib_rx_t *rx, int status, int nob)
475 kib_msg_t *msg = rx->rx_msg;
476 kib_conn_t *conn = rx->rx_conn;
477 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
478 kib_net_t *net = ni->ni_data;
482 LASSERT (net != NULL);
483 LASSERT (rx->rx_nob < 0); /* was posted */
484 rx->rx_nob = 0; /* isn't now */
486 if (conn->ibc_state > IBLND_CONN_ESTABLISHED)
489 if (status != IB_WC_SUCCESS) {
490 CNETERR("Rx from %s failed: %d\n",
491 libcfs_nid2str(conn->ibc_peer->ibp_nid), status);
498 rc = kiblnd_unpack_msg(msg, rx->rx_nob);
500 CERROR ("Error %d unpacking rx from %s\n",
501 rc, libcfs_nid2str(conn->ibc_peer->ibp_nid));
505 if (msg->ibm_srcnid != conn->ibc_peer->ibp_nid ||
506 msg->ibm_dstnid != ni->ni_nid ||
507 msg->ibm_srcstamp != conn->ibc_incarnation ||
508 msg->ibm_dststamp != net->ibn_incarnation) {
509 CERROR ("Stale rx from %s\n",
510 libcfs_nid2str(conn->ibc_peer->ibp_nid));
515 /* set time last known alive */
516 kiblnd_peer_alive(conn->ibc_peer);
518 /* racing with connection establishment/teardown! */
520 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
521 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
524 write_lock_irqsave(g_lock, flags);
525 /* must check holding global lock to eliminate race */
526 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
527 list_add_tail(&rx->rx_list, &conn->ibc_early_rxs);
528 write_unlock_irqrestore(g_lock, flags);
531 write_unlock_irqrestore(g_lock, flags);
533 kiblnd_handle_rx(rx);
537 CDEBUG(D_NET, "rx %p conn %p\n", rx, conn);
538 kiblnd_close_conn(conn, err);
540 kiblnd_drop_rx(rx); /* Don't re-post rx. */
544 kiblnd_kvaddr_to_page (unsigned long vaddr)
548 if (is_vmalloc_addr((void *)vaddr)) {
549 page = vmalloc_to_page ((void *)vaddr);
550 LASSERT (page != NULL);
553 #ifdef CONFIG_HIGHMEM
554 if (vaddr >= PKMAP_BASE &&
555 vaddr < (PKMAP_BASE + LAST_PKMAP * PAGE_SIZE)) {
556 /* No highmem pages only used for bulk (kiov) I/O */
557 CERROR("find page for address in highmem\n");
561 page = virt_to_page (vaddr);
562 LASSERT (page != NULL);
567 kiblnd_fmr_map_tx(kib_net_t *net, kib_tx_t *tx, kib_rdma_desc_t *rd, __u32 nob)
570 kib_fmr_poolset_t *fps;
574 LASSERT(tx->tx_pool != NULL);
575 LASSERT(tx->tx_pool->tpo_pool.po_owner != NULL);
577 hdev = tx->tx_pool->tpo_hdev;
578 cpt = tx->tx_pool->tpo_pool.po_owner->ps_cpt;
580 fps = net->ibn_fmr_ps[cpt];
581 rc = kiblnd_fmr_pool_map(fps, tx, rd, nob, 0, &tx->fmr);
583 CERROR("Can't map %u pages: %d\n", nob, rc);
587 /* If rd is not tx_rd, it's going to get sent to a peer, who will need
589 rd->rd_key = tx->fmr.fmr_key;
590 rd->rd_frags[0].rf_addr &= ~hdev->ibh_page_mask;
591 rd->rd_frags[0].rf_nob = nob;
598 kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx)
600 kib_net_t *net = ni->ni_data;
602 LASSERT(net != NULL);
604 if (net->ibn_fmr_ps != NULL)
605 kiblnd_fmr_pool_unmap(&tx->fmr, tx->tx_status);
607 if (tx->tx_nfrags != 0) {
608 kiblnd_dma_unmap_sg(tx->tx_pool->tpo_hdev->ibh_ibdev,
609 tx->tx_frags, tx->tx_nfrags, tx->tx_dmadir);
615 kiblnd_map_tx(lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd, int nfrags)
617 kib_net_t *net = ni->ni_data;
618 kib_hca_dev_t *hdev = net->ibn_dev->ibd_hdev;
619 struct ib_mr *mr = NULL;
623 /* If rd is not tx_rd, it's going to get sent to a peer and I'm the
625 tx->tx_dmadir = (rd != tx->tx_rd) ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
626 tx->tx_nfrags = nfrags;
628 rd->rd_nfrags = kiblnd_dma_map_sg(hdev->ibh_ibdev, tx->tx_frags,
629 tx->tx_nfrags, tx->tx_dmadir);
631 for (i = 0, nob = 0; i < rd->rd_nfrags; i++) {
632 rd->rd_frags[i].rf_nob = kiblnd_sg_dma_len(
633 hdev->ibh_ibdev, &tx->tx_frags[i]);
634 rd->rd_frags[i].rf_addr = kiblnd_sg_dma_address(
635 hdev->ibh_ibdev, &tx->tx_frags[i]);
636 nob += rd->rd_frags[i].rf_nob;
639 mr = kiblnd_find_rd_dma_mr(ni, rd,
640 (tx->tx_conn != NULL) ?
641 tx->tx_conn->ibc_max_frags : -1);
643 /* found pre-mapping MR */
644 rd->rd_key = (rd != tx->tx_rd) ? mr->rkey : mr->lkey;
648 if (net->ibn_fmr_ps != NULL)
649 return kiblnd_fmr_map_tx(net, tx, rd, nob);
656 kiblnd_setup_rd_iov(lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd,
657 unsigned int niov, struct kvec *iov, int offset, int nob)
659 kib_net_t *net = ni->ni_data;
661 struct scatterlist *sg;
668 LASSERT (net != NULL);
670 while (offset >= iov->iov_len) {
671 offset -= iov->iov_len;
681 vaddr = ((unsigned long)iov->iov_base) + offset;
682 page_offset = vaddr & (PAGE_SIZE - 1);
683 page = kiblnd_kvaddr_to_page(vaddr);
685 CERROR ("Can't find page\n");
689 fragnob = min((int)(iov->iov_len - offset), nob);
690 fragnob = min(fragnob, (int)PAGE_SIZE - page_offset);
692 sg_set_page(sg, page, fragnob, page_offset);
695 CERROR("lacking enough sg entries to map tx\n");
699 if (offset + fragnob < iov->iov_len) {
709 return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags);
713 kiblnd_setup_rd_kiov (lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd,
714 int nkiov, lnet_kiov_t *kiov, int offset, int nob)
716 kib_net_t *net = ni->ni_data;
717 struct scatterlist *sg;
720 CDEBUG(D_NET, "niov %d offset %d nob %d\n", nkiov, offset, nob);
724 LASSERT (net != NULL);
726 while (offset >= kiov->kiov_len) {
727 offset -= kiov->kiov_len;
737 fragnob = min((int)(kiov->kiov_len - offset), nob);
739 sg_set_page(sg, kiov->kiov_page, fragnob,
740 kiov->kiov_offset + offset);
743 CERROR("lacking enough sg entries to map tx\n");
753 return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags);
757 kiblnd_post_tx_locked (kib_conn_t *conn, kib_tx_t *tx, int credit)
758 __must_hold(&conn->ibc_lock)
760 kib_msg_t *msg = tx->tx_msg;
761 kib_peer_t *peer = conn->ibc_peer;
762 struct lnet_ni *ni = peer->ibp_ni;
763 int ver = conn->ibc_version;
767 LASSERT(tx->tx_queued);
768 /* We rely on this for QP sizing */
769 LASSERT(tx->tx_nwrq > 0);
770 LASSERT(tx->tx_nwrq <= 1 + conn->ibc_max_frags);
772 LASSERT(credit == 0 || credit == 1);
773 LASSERT(conn->ibc_outstanding_credits >= 0);
774 LASSERT(conn->ibc_outstanding_credits <= conn->ibc_queue_depth);
775 LASSERT(conn->ibc_credits >= 0);
776 LASSERT(conn->ibc_credits <= conn->ibc_queue_depth);
778 if (conn->ibc_nsends_posted ==
779 kiblnd_concurrent_sends(ver, ni)) {
780 /* tx completions outstanding... */
781 CDEBUG(D_NET, "%s: posted enough\n",
782 libcfs_nid2str(peer->ibp_nid));
786 if (credit != 0 && conn->ibc_credits == 0) { /* no credits */
787 CDEBUG(D_NET, "%s: no credits\n",
788 libcfs_nid2str(peer->ibp_nid));
792 if (credit != 0 && !IBLND_OOB_CAPABLE(ver) &&
793 conn->ibc_credits == 1 && /* last credit reserved */
794 msg->ibm_type != IBLND_MSG_NOOP) { /* for NOOP */
795 CDEBUG(D_NET, "%s: not using last credit\n",
796 libcfs_nid2str(peer->ibp_nid));
800 /* NB don't drop ibc_lock before bumping tx_sending */
801 list_del(&tx->tx_list);
804 if (msg->ibm_type == IBLND_MSG_NOOP &&
805 (!kiblnd_need_noop(conn) || /* redundant NOOP */
806 (IBLND_OOB_CAPABLE(ver) && /* posted enough NOOP */
807 conn->ibc_noops_posted == IBLND_OOB_MSGS(ver)))) {
808 /* OK to drop when posted enough NOOPs, since
809 * kiblnd_check_sends_locked will queue NOOP again when
810 * posted NOOPs complete */
811 spin_unlock(&conn->ibc_lock);
812 kiblnd_tx_done(peer->ibp_ni, tx);
813 spin_lock(&conn->ibc_lock);
814 CDEBUG(D_NET, "%s(%d): redundant or enough NOOP\n",
815 libcfs_nid2str(peer->ibp_nid),
816 conn->ibc_noops_posted);
820 kiblnd_pack_msg(peer->ibp_ni, msg, ver, conn->ibc_outstanding_credits,
821 peer->ibp_nid, conn->ibc_incarnation);
823 conn->ibc_credits -= credit;
824 conn->ibc_outstanding_credits = 0;
825 conn->ibc_nsends_posted++;
826 if (msg->ibm_type == IBLND_MSG_NOOP)
827 conn->ibc_noops_posted++;
829 /* CAVEAT EMPTOR! This tx could be the PUT_DONE of an RDMA
830 * PUT. If so, it was first queued here as a PUT_REQ, sent and
831 * stashed on ibc_active_txs, matched by an incoming PUT_ACK,
832 * and then re-queued here. It's (just) possible that
833 * tx_sending is non-zero if we've not done the tx_complete()
834 * from the first send; hence the ++ rather than = below. */
836 list_add(&tx->tx_list, &conn->ibc_active_txs);
838 /* I'm still holding ibc_lock! */
839 if (conn->ibc_state != IBLND_CONN_ESTABLISHED) {
841 } else if (tx->tx_pool->tpo_pool.po_failed ||
842 conn->ibc_hdev != tx->tx_pool->tpo_hdev) {
843 /* close_conn will launch failover */
846 struct kib_fast_reg_descriptor *frd = tx->fmr.fmr_frd;
847 struct ib_send_wr *bad = &tx->tx_wrq[tx->tx_nwrq - 1].wr;
848 struct ib_send_wr *wr = &tx->tx_wrq[0].wr;
851 if (!frd->frd_valid) {
852 wr = &frd->frd_inv_wr.wr;
853 wr->next = &frd->frd_fastreg_wr.wr;
855 wr = &frd->frd_fastreg_wr.wr;
857 frd->frd_fastreg_wr.wr.next = &tx->tx_wrq[0].wr;
860 LASSERTF(bad->wr_id == kiblnd_ptr2wreqid(tx, IBLND_WID_TX),
861 "bad wr_id %#llx, opc %d, flags %d, peer: %s\n",
862 bad->wr_id, bad->opcode, bad->send_flags,
863 libcfs_nid2str(conn->ibc_peer->ibp_nid));
866 rc = ib_post_send(conn->ibc_cmid->qp, wr, &bad);
869 conn->ibc_last_send = jiffies;
874 /* NB credits are transferred in the actual
875 * 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 == 0);
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(kib_conn_t *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 <=
924 kiblnd_concurrent_sends(ver, ni));
925 LASSERT (!IBLND_OOB_CAPABLE(ver) ||
926 conn->ibc_noops_posted <= IBLND_OOB_MSGS(ver));
927 LASSERT (conn->ibc_reserved_credits >= 0);
929 while (conn->ibc_reserved_credits > 0 &&
930 !list_empty(&conn->ibc_tx_queue_rsrvd)) {
931 tx = list_entry(conn->ibc_tx_queue_rsrvd.next,
933 list_del(&tx->tx_list);
934 list_add_tail(&tx->tx_list, &conn->ibc_tx_queue);
935 conn->ibc_reserved_credits--;
938 if (kiblnd_need_noop(conn)) {
939 spin_unlock(&conn->ibc_lock);
941 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
943 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_NOOP, 0);
945 spin_lock(&conn->ibc_lock);
947 kiblnd_queue_tx_locked(tx, conn);
953 if (!list_empty(&conn->ibc_tx_queue_nocred)) {
955 tx = list_entry(conn->ibc_tx_queue_nocred.next,
957 } else if (!list_empty(&conn->ibc_tx_noops)) {
958 LASSERT (!IBLND_OOB_CAPABLE(ver));
960 tx = list_entry(conn->ibc_tx_noops.next,
962 } else if (!list_empty(&conn->ibc_tx_queue)) {
964 tx = list_entry(conn->ibc_tx_queue.next,
969 if (kiblnd_post_tx_locked(conn, tx, credit) != 0)
975 kiblnd_tx_complete (kib_tx_t *tx, int status)
977 int failed = (status != IB_WC_SUCCESS);
978 kib_conn_t *conn = tx->tx_conn;
981 LASSERT (tx->tx_sending > 0);
984 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
985 CNETERR("Tx -> %s cookie %#llx"
986 " sending %d waiting %d: failed %d\n",
987 libcfs_nid2str(conn->ibc_peer->ibp_nid),
988 tx->tx_cookie, tx->tx_sending, tx->tx_waiting,
991 kiblnd_close_conn(conn, -EIO);
993 kiblnd_peer_alive(conn->ibc_peer);
996 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 == 0) && /* 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, kib_tx_t *tx, int type, int body_nob)
1027 kib_hca_dev_t *hdev = tx->tx_pool->tpo_hdev;
1028 struct ib_sge *sge = &tx->tx_sge[tx->tx_nwrq];
1029 struct ib_rdma_wr *wrq;
1030 int nob = offsetof(kib_msg_t, 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);
1036 LASSERT(mr != NULL);
1038 kiblnd_init_msg(tx->tx_msg, type, body_nob);
1040 sge->lkey = mr->lkey;
1041 sge->addr = tx->tx_msgaddr;
1044 wrq = &tx->tx_wrq[tx->tx_nwrq];
1045 memset(wrq, 0, sizeof(*wrq));
1047 wrq->wr.next = NULL;
1048 wrq->wr.wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_TX);
1049 wrq->wr.sg_list = sge;
1050 wrq->wr.num_sge = 1;
1051 wrq->wr.opcode = IB_WR_SEND;
1052 wrq->wr.send_flags = IB_SEND_SIGNALED;
1058 kiblnd_init_rdma(kib_conn_t *conn, kib_tx_t *tx, int type,
1059 int resid, kib_rdma_desc_t *dstrd, __u64 dstcookie)
1061 kib_msg_t *ibmsg = tx->tx_msg;
1062 kib_rdma_desc_t *srcrd = tx->tx_rd;
1063 struct ib_sge *sge = &tx->tx_sge[0];
1064 struct ib_rdma_wr *wrq;
1070 LASSERT (!in_interrupt());
1071 LASSERT (tx->tx_nwrq == 0);
1072 LASSERT (type == IBLND_MSG_GET_DONE ||
1073 type == IBLND_MSG_PUT_DONE);
1075 srcidx = dstidx = 0;
1078 if (srcidx >= srcrd->rd_nfrags) {
1079 CERROR("Src buffer exhausted: %d frags\n", srcidx);
1084 if (dstidx == dstrd->rd_nfrags) {
1085 CERROR("Dst buffer exhausted: %d frags\n", dstidx);
1090 if (tx->tx_nwrq >= conn->ibc_max_frags) {
1091 CERROR("RDMA has too many fragments for peer %s (%d), "
1092 "src idx/frags: %d/%d dst idx/frags: %d/%d\n",
1093 libcfs_nid2str(conn->ibc_peer->ibp_nid),
1094 conn->ibc_max_frags,
1095 srcidx, srcrd->rd_nfrags,
1096 dstidx, dstrd->rd_nfrags);
1101 wrknob = MIN(MIN(kiblnd_rd_frag_size(srcrd, srcidx),
1102 kiblnd_rd_frag_size(dstrd, dstidx)), resid);
1104 sge = &tx->tx_sge[tx->tx_nwrq];
1105 sge->addr = kiblnd_rd_frag_addr(srcrd, srcidx);
1106 sge->lkey = kiblnd_rd_frag_key(srcrd, srcidx);
1107 sge->length = wrknob;
1109 wrq = &tx->tx_wrq[tx->tx_nwrq];
1111 wrq->wr.next = &(wrq + 1)->wr;
1112 wrq->wr.wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_RDMA);
1113 wrq->wr.sg_list = sge;
1114 wrq->wr.num_sge = 1;
1115 wrq->wr.opcode = IB_WR_RDMA_WRITE;
1116 wrq->wr.send_flags = 0;
1118 #ifdef HAVE_IB_RDMA_WR
1119 wrq->remote_addr = kiblnd_rd_frag_addr(dstrd, dstidx);
1120 wrq->rkey = kiblnd_rd_frag_key(dstrd, dstidx);
1122 wrq->wr.wr.rdma.remote_addr = kiblnd_rd_frag_addr(dstrd, dstidx);
1123 wrq->wr.wr.rdma.rkey = kiblnd_rd_frag_key(dstrd, dstidx);
1126 srcidx = kiblnd_rd_consume_frag(srcrd, srcidx, wrknob);
1127 dstidx = kiblnd_rd_consume_frag(dstrd, dstidx, wrknob);
1136 if (rc < 0) /* no RDMA if completing with failure */
1139 ibmsg->ibm_u.completion.ibcm_status = rc;
1140 ibmsg->ibm_u.completion.ibcm_cookie = dstcookie;
1141 kiblnd_init_tx_msg(conn->ibc_peer->ibp_ni, tx,
1142 type, sizeof (kib_completion_msg_t));
1148 kiblnd_queue_tx_locked(kib_tx_t *tx, kib_conn_t *conn)
1150 struct list_head *q;
1152 LASSERT(tx->tx_nwrq > 0); /* work items set up */
1153 LASSERT(!tx->tx_queued); /* not queued for sending already */
1154 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1157 tx->tx_deadline = jiffies +
1158 msecs_to_jiffies(*kiblnd_tunables.kib_timeout *
1161 if (tx->tx_conn == NULL) {
1162 kiblnd_conn_addref(conn);
1164 LASSERT (tx->tx_msg->ibm_type != IBLND_MSG_PUT_DONE);
1166 /* PUT_DONE first attached to conn as a PUT_REQ */
1167 LASSERT (tx->tx_conn == conn);
1168 LASSERT (tx->tx_msg->ibm_type == IBLND_MSG_PUT_DONE);
1171 switch (tx->tx_msg->ibm_type) {
1175 case IBLND_MSG_PUT_REQ:
1176 case IBLND_MSG_GET_REQ:
1177 q = &conn->ibc_tx_queue_rsrvd;
1180 case IBLND_MSG_PUT_NAK:
1181 case IBLND_MSG_PUT_ACK:
1182 case IBLND_MSG_PUT_DONE:
1183 case IBLND_MSG_GET_DONE:
1184 q = &conn->ibc_tx_queue_nocred;
1187 case IBLND_MSG_NOOP:
1188 if (IBLND_OOB_CAPABLE(conn->ibc_version))
1189 q = &conn->ibc_tx_queue_nocred;
1191 q = &conn->ibc_tx_noops;
1194 case IBLND_MSG_IMMEDIATE:
1195 q = &conn->ibc_tx_queue;
1199 list_add_tail(&tx->tx_list, q);
1203 kiblnd_queue_tx (kib_tx_t *tx, kib_conn_t *conn)
1205 spin_lock(&conn->ibc_lock);
1206 kiblnd_queue_tx_locked(tx, conn);
1207 kiblnd_check_sends_locked(conn);
1208 spin_unlock(&conn->ibc_lock);
1211 static int kiblnd_resolve_addr(struct rdma_cm_id *cmid,
1212 struct sockaddr_in *srcaddr,
1213 struct sockaddr_in *dstaddr,
1216 unsigned short port;
1219 /* allow the port to be reused */
1220 rc = rdma_set_reuseaddr(cmid, 1);
1222 CERROR("Unable to set reuse on cmid: %d\n", rc);
1226 /* look for a free privileged port */
1227 for (port = PROT_SOCK-1; port > 0; port--) {
1228 srcaddr->sin_port = htons(port);
1229 rc = rdma_resolve_addr(cmid,
1230 (struct sockaddr *)srcaddr,
1231 (struct sockaddr *)dstaddr,
1234 CDEBUG(D_NET, "bound to port %hu\n", port);
1236 } else if (rc == -EADDRINUSE || rc == -EADDRNOTAVAIL) {
1237 CDEBUG(D_NET, "bind to port %hu failed: %d\n",
1244 CERROR("Failed to bind to a free privileged port\n");
1249 kiblnd_connect_peer (kib_peer_t *peer)
1251 struct rdma_cm_id *cmid;
1253 kib_net_t *net = peer->ibp_ni->ni_data;
1254 struct sockaddr_in srcaddr;
1255 struct sockaddr_in dstaddr;
1258 LASSERT (net != NULL);
1259 LASSERT (peer->ibp_connecting > 0);
1260 LASSERT(!peer->ibp_reconnecting);
1262 cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, peer, RDMA_PS_TCP,
1266 CERROR("Can't create CMID for %s: %ld\n",
1267 libcfs_nid2str(peer->ibp_nid), PTR_ERR(cmid));
1273 memset(&srcaddr, 0, sizeof(srcaddr));
1274 srcaddr.sin_family = AF_INET;
1275 srcaddr.sin_addr.s_addr = htonl(dev->ibd_ifip);
1277 memset(&dstaddr, 0, sizeof(dstaddr));
1278 dstaddr.sin_family = AF_INET;
1279 dstaddr.sin_port = htons(*kiblnd_tunables.kib_service);
1280 dstaddr.sin_addr.s_addr = htonl(LNET_NIDADDR(peer->ibp_nid));
1282 kiblnd_peer_addref(peer); /* cmid's ref */
1284 if (*kiblnd_tunables.kib_use_priv_port) {
1285 rc = kiblnd_resolve_addr(cmid, &srcaddr, &dstaddr,
1286 *kiblnd_tunables.kib_timeout * 1000);
1288 rc = rdma_resolve_addr(cmid,
1289 (struct sockaddr *)&srcaddr,
1290 (struct sockaddr *)&dstaddr,
1291 *kiblnd_tunables.kib_timeout * 1000);
1294 /* Can't initiate address resolution: */
1295 CERROR("Can't resolve addr for %s: %d\n",
1296 libcfs_nid2str(peer->ibp_nid), rc);
1300 LASSERT (cmid->device != NULL);
1301 CDEBUG(D_NET, "%s: connection bound to %s:%pI4h:%s\n",
1302 libcfs_nid2str(peer->ibp_nid), dev->ibd_ifname,
1303 &dev->ibd_ifip, cmid->device->name);
1308 kiblnd_peer_connect_failed(peer, 1, rc);
1309 kiblnd_peer_decref(peer); /* cmid's ref */
1310 rdma_destroy_id(cmid);
1313 kiblnd_peer_connect_failed(peer, 1, rc);
1317 kiblnd_reconnect_peer(kib_peer_t *peer)
1319 rwlock_t *glock = &kiblnd_data.kib_global_lock;
1320 char *reason = NULL;
1321 struct list_head txs;
1322 unsigned long flags;
1324 INIT_LIST_HEAD(&txs);
1326 write_lock_irqsave(glock, flags);
1327 if (peer->ibp_reconnecting == 0) {
1328 if (peer->ibp_accepting)
1329 reason = "accepting";
1330 else if (peer->ibp_connecting)
1331 reason = "connecting";
1332 else if (!list_empty(&peer->ibp_conns))
1333 reason = "connected";
1334 else /* connected then closed */
1340 LASSERT(!peer->ibp_accepting && !peer->ibp_connecting &&
1341 list_empty(&peer->ibp_conns));
1342 peer->ibp_reconnecting = 0;
1344 if (!kiblnd_peer_active(peer)) {
1345 list_splice_init(&peer->ibp_tx_queue, &txs);
1346 reason = "unlinked";
1350 peer->ibp_connecting++;
1351 peer->ibp_reconnected++;
1352 write_unlock_irqrestore(glock, flags);
1354 kiblnd_connect_peer(peer);
1358 write_unlock_irqrestore(glock, flags);
1360 CWARN("Abort reconnection of %s: %s\n",
1361 libcfs_nid2str(peer->ibp_nid), reason);
1362 kiblnd_txlist_done(peer->ibp_ni, &txs, -ECONNABORTED);
1367 kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid)
1372 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
1373 unsigned long flags;
1376 /* If I get here, I've committed to send, so I complete the tx with
1377 * failure on any problems */
1379 LASSERT (tx == NULL || tx->tx_conn == NULL); /* only set when assigned a conn */
1380 LASSERT (tx == NULL || tx->tx_nwrq > 0); /* work items have been set up */
1382 /* 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 != NULL && !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);
1443 if (peer2 != NULL) {
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 == 0);
1468 peer->ibp_connecting = 1;
1470 /* always called with a ref on ni, which prevents ni being shutdown */
1471 LASSERT (((kib_net_t *)ni->ni_data)->ibn_shutdown == 0);
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;
1499 kib_rdma_desc_t *rd;
1504 /* NB 'private' is different depending on what we're sending.... */
1506 CDEBUG(D_NET, "sending %d bytes in %d frags to %s\n",
1507 payload_nob, payload_niov, libcfs_id2str(target));
1509 LASSERT (payload_nob == 0 || payload_niov > 0);
1510 LASSERT (payload_niov <= LNET_MAX_IOV);
1512 /* Thread context */
1513 LASSERT (!in_interrupt());
1514 /* payload is either all vaddrs or all pages */
1515 LASSERT (!(payload_kiov != NULL && payload_iov != NULL));
1523 LASSERT (payload_nob == 0);
1527 if (routing || target_is_router)
1528 break; /* send IMMEDIATE */
1530 /* is the REPLY message too small for RDMA? */
1531 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[lntmsg->msg_md->md_length]);
1532 if (nob <= IBLND_MSG_SIZE)
1533 break; /* send IMMEDIATE */
1535 tx = kiblnd_get_idle_tx(ni, target.nid);
1537 CERROR("Can't allocate txd for GET to %s\n",
1538 libcfs_nid2str(target.nid));
1543 rd = &ibmsg->ibm_u.get.ibgm_rd;
1544 if ((lntmsg->msg_md->md_options & LNET_MD_KIOV) == 0)
1545 rc = kiblnd_setup_rd_iov(ni, tx, rd,
1546 lntmsg->msg_md->md_niov,
1547 lntmsg->msg_md->md_iov.iov,
1548 0, lntmsg->msg_md->md_length);
1550 rc = kiblnd_setup_rd_kiov(ni, tx, rd,
1551 lntmsg->msg_md->md_niov,
1552 lntmsg->msg_md->md_iov.kiov,
1553 0, lntmsg->msg_md->md_length);
1555 CERROR("Can't setup GET sink for %s: %d\n",
1556 libcfs_nid2str(target.nid), rc);
1557 kiblnd_tx_done(ni, tx);
1561 nob = offsetof(kib_get_msg_t, ibgm_rd.rd_frags[rd->rd_nfrags]);
1562 ibmsg->ibm_u.get.ibgm_cookie = tx->tx_cookie;
1563 ibmsg->ibm_u.get.ibgm_hdr = *hdr;
1565 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_GET_REQ, nob);
1567 tx->tx_lntmsg[1] = lnet_create_reply_msg(ni, lntmsg);
1568 if (tx->tx_lntmsg[1] == NULL) {
1569 CERROR("Can't create reply for GET -> %s\n",
1570 libcfs_nid2str(target.nid));
1571 kiblnd_tx_done(ni, tx);
1575 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg[0,1] on completion */
1576 tx->tx_waiting = 1; /* waiting for GET_DONE */
1577 kiblnd_launch_tx(ni, tx, target.nid);
1580 case LNET_MSG_REPLY:
1582 /* Is the payload small enough not to need RDMA? */
1583 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob]);
1584 if (nob <= IBLND_MSG_SIZE)
1585 break; /* send IMMEDIATE */
1587 tx = kiblnd_get_idle_tx(ni, target.nid);
1589 CERROR("Can't allocate %s txd for %s\n",
1590 type == LNET_MSG_PUT ? "PUT" : "REPLY",
1591 libcfs_nid2str(target.nid));
1595 if (payload_kiov == NULL)
1596 rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1597 payload_niov, payload_iov,
1598 payload_offset, payload_nob);
1600 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1601 payload_niov, payload_kiov,
1602 payload_offset, payload_nob);
1604 CERROR("Can't setup PUT src for %s: %d\n",
1605 libcfs_nid2str(target.nid), rc);
1606 kiblnd_tx_done(ni, tx);
1611 ibmsg->ibm_u.putreq.ibprm_hdr = *hdr;
1612 ibmsg->ibm_u.putreq.ibprm_cookie = tx->tx_cookie;
1613 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_REQ, sizeof(kib_putreq_msg_t));
1615 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1616 tx->tx_waiting = 1; /* waiting for PUT_{ACK,NAK} */
1617 kiblnd_launch_tx(ni, tx, target.nid);
1621 /* send IMMEDIATE */
1623 LASSERT (offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob])
1626 tx = kiblnd_get_idle_tx(ni, target.nid);
1628 CERROR ("Can't send %d to %s: tx descs exhausted\n",
1629 type, libcfs_nid2str(target.nid));
1634 ibmsg->ibm_u.immediate.ibim_hdr = *hdr;
1636 if (payload_kiov != NULL)
1637 lnet_copy_kiov2flat(IBLND_MSG_SIZE, ibmsg,
1638 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1639 payload_niov, payload_kiov,
1640 payload_offset, payload_nob);
1642 lnet_copy_iov2flat(IBLND_MSG_SIZE, ibmsg,
1643 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1644 payload_niov, payload_iov,
1645 payload_offset, payload_nob);
1647 nob = offsetof(kib_immediate_msg_t, ibim_payload[payload_nob]);
1648 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_IMMEDIATE, nob);
1650 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1651 kiblnd_launch_tx(ni, tx, target.nid);
1656 kiblnd_reply (lnet_ni_t *ni, kib_rx_t *rx, lnet_msg_t *lntmsg)
1658 lnet_process_id_t target = lntmsg->msg_target;
1659 unsigned int niov = lntmsg->msg_niov;
1660 struct kvec *iov = lntmsg->msg_iov;
1661 lnet_kiov_t *kiov = lntmsg->msg_kiov;
1662 unsigned int offset = lntmsg->msg_offset;
1663 unsigned int nob = lntmsg->msg_len;
1667 tx = kiblnd_get_idle_tx(ni, rx->rx_conn->ibc_peer->ibp_nid);
1669 CERROR("Can't get tx for REPLY to %s\n",
1670 libcfs_nid2str(target.nid));
1676 else if (kiov == NULL)
1677 rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1678 niov, iov, offset, nob);
1680 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1681 niov, kiov, offset, nob);
1684 CERROR("Can't setup GET src for %s: %d\n",
1685 libcfs_nid2str(target.nid), rc);
1689 rc = kiblnd_init_rdma(rx->rx_conn, tx,
1690 IBLND_MSG_GET_DONE, nob,
1691 &rx->rx_msg->ibm_u.get.ibgm_rd,
1692 rx->rx_msg->ibm_u.get.ibgm_cookie);
1694 CERROR("Can't setup rdma for GET from %s: %d\n",
1695 libcfs_nid2str(target.nid), rc);
1700 /* No RDMA: local completion may happen now! */
1701 lnet_finalize(ni, lntmsg, 0);
1703 /* RDMA: lnet_finalize(lntmsg) when it
1705 tx->tx_lntmsg[0] = lntmsg;
1708 kiblnd_queue_tx(tx, rx->rx_conn);
1712 kiblnd_tx_done(ni, tx);
1714 lnet_finalize(ni, lntmsg, -EIO);
1718 kiblnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed,
1719 unsigned int niov, struct kvec *iov, lnet_kiov_t *kiov,
1720 unsigned int offset, unsigned int mlen, unsigned int rlen)
1722 kib_rx_t *rx = private;
1723 kib_msg_t *rxmsg = rx->rx_msg;
1724 kib_conn_t *conn = rx->rx_conn;
1727 int post_credit = IBLND_POSTRX_PEER_CREDIT;
1730 LASSERT (mlen <= rlen);
1731 LASSERT (!in_interrupt());
1732 /* Either all pages or all vaddrs */
1733 LASSERT (!(kiov != NULL && iov != NULL));
1735 switch (rxmsg->ibm_type) {
1739 case IBLND_MSG_IMMEDIATE:
1740 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[rlen]);
1741 if (nob > rx->rx_nob) {
1742 CERROR ("Immediate message from %s too big: %d(%d)\n",
1743 libcfs_nid2str(rxmsg->ibm_u.immediate.ibim_hdr.src_nid),
1750 lnet_copy_flat2kiov(niov, kiov, offset,
1751 IBLND_MSG_SIZE, rxmsg,
1752 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1755 lnet_copy_flat2iov(niov, iov, offset,
1756 IBLND_MSG_SIZE, rxmsg,
1757 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1759 lnet_finalize (ni, lntmsg, 0);
1762 case IBLND_MSG_PUT_REQ: {
1764 kib_rdma_desc_t *rd;
1767 lnet_finalize(ni, lntmsg, 0);
1768 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK, 0,
1769 rxmsg->ibm_u.putreq.ibprm_cookie);
1773 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
1775 CERROR("Can't allocate tx for %s\n",
1776 libcfs_nid2str(conn->ibc_peer->ibp_nid));
1777 /* Not replying will break the connection */
1783 rd = &txmsg->ibm_u.putack.ibpam_rd;
1785 rc = kiblnd_setup_rd_iov(ni, tx, rd,
1786 niov, iov, offset, mlen);
1788 rc = kiblnd_setup_rd_kiov(ni, tx, rd,
1789 niov, kiov, offset, mlen);
1791 CERROR("Can't setup PUT sink for %s: %d\n",
1792 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
1793 kiblnd_tx_done(ni, tx);
1794 /* tell peer it's over */
1795 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK, rc,
1796 rxmsg->ibm_u.putreq.ibprm_cookie);
1800 nob = offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[rd->rd_nfrags]);
1801 txmsg->ibm_u.putack.ibpam_src_cookie = rxmsg->ibm_u.putreq.ibprm_cookie;
1802 txmsg->ibm_u.putack.ibpam_dst_cookie = tx->tx_cookie;
1804 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_ACK, nob);
1806 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1807 tx->tx_waiting = 1; /* waiting for PUT_DONE */
1808 kiblnd_queue_tx(tx, conn);
1810 /* reposted buffer reserved for PUT_DONE */
1811 post_credit = IBLND_POSTRX_NO_CREDIT;
1815 case IBLND_MSG_GET_REQ:
1816 if (lntmsg != NULL) {
1817 /* Optimized GET; RDMA lntmsg's payload */
1818 kiblnd_reply(ni, rx, lntmsg);
1820 /* GET didn't match anything */
1821 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_GET_DONE,
1823 rxmsg->ibm_u.get.ibgm_cookie);
1828 kiblnd_post_rx(rx, post_credit);
1833 kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name)
1835 struct task_struct *task = kthread_run(fn, arg, name);
1838 return PTR_ERR(task);
1840 atomic_inc(&kiblnd_data.kib_nthreads);
1845 kiblnd_thread_fini (void)
1847 atomic_dec (&kiblnd_data.kib_nthreads);
1851 kiblnd_peer_alive (kib_peer_t *peer)
1853 /* This is racy, but everyone's only writing cfs_time_current() */
1854 peer->ibp_last_alive = cfs_time_current();
1859 kiblnd_peer_notify (kib_peer_t *peer)
1862 cfs_time_t last_alive = 0;
1863 unsigned long flags;
1865 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1867 if (kiblnd_peer_idle(peer) && peer->ibp_error != 0) {
1868 error = peer->ibp_error;
1869 peer->ibp_error = 0;
1871 last_alive = peer->ibp_last_alive;
1874 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1877 lnet_notify(peer->ibp_ni,
1878 peer->ibp_nid, 0, last_alive);
1882 kiblnd_close_conn_locked (kib_conn_t *conn, int error)
1884 /* This just does the immediate housekeeping. 'error' is zero for a
1885 * normal shutdown which can happen only after the connection has been
1886 * established. If the connection is established, schedule the
1887 * connection to be finished off by the connd. Otherwise the connd is
1888 * already dealing with it (either to set it up or tear it down).
1889 * Caller holds kib_global_lock exclusively in irq context */
1890 kib_peer_t *peer = conn->ibc_peer;
1892 unsigned long flags;
1894 LASSERT (error != 0 || conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1896 if (error != 0 && conn->ibc_comms_error == 0)
1897 conn->ibc_comms_error = error;
1899 if (conn->ibc_state != IBLND_CONN_ESTABLISHED)
1900 return; /* already being handled */
1903 list_empty(&conn->ibc_tx_noops) &&
1904 list_empty(&conn->ibc_tx_queue) &&
1905 list_empty(&conn->ibc_tx_queue_rsrvd) &&
1906 list_empty(&conn->ibc_tx_queue_nocred) &&
1907 list_empty(&conn->ibc_active_txs)) {
1908 CDEBUG(D_NET, "closing conn to %s\n",
1909 libcfs_nid2str(peer->ibp_nid));
1911 CNETERR("Closing conn to %s: error %d%s%s%s%s%s\n",
1912 libcfs_nid2str(peer->ibp_nid), error,
1913 list_empty(&conn->ibc_tx_queue) ? "" : "(sending)",
1914 list_empty(&conn->ibc_tx_noops) ? "" : "(sending_noops)",
1915 list_empty(&conn->ibc_tx_queue_rsrvd) ?
1916 "" : "(sending_rsrvd)",
1917 list_empty(&conn->ibc_tx_queue_nocred) ?
1918 "" : "(sending_nocred)",
1919 list_empty(&conn->ibc_active_txs) ? "" : "(waiting)");
1922 dev = ((kib_net_t *)peer->ibp_ni->ni_data)->ibn_dev;
1923 list_del(&conn->ibc_list);
1924 /* connd (see below) takes over ibc_list's ref */
1926 if (list_empty(&peer->ibp_conns) && /* no more conns */
1927 kiblnd_peer_active(peer)) { /* still in peer table */
1928 kiblnd_unlink_peer_locked(peer);
1930 /* set/clear error on last conn */
1931 peer->ibp_error = conn->ibc_comms_error;
1934 kiblnd_set_conn_state(conn, IBLND_CONN_CLOSING);
1937 kiblnd_dev_can_failover(dev)) {
1938 list_add_tail(&dev->ibd_fail_list,
1939 &kiblnd_data.kib_failed_devs);
1940 wake_up(&kiblnd_data.kib_failover_waitq);
1943 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
1945 list_add_tail(&conn->ibc_list, &kiblnd_data.kib_connd_conns);
1946 wake_up(&kiblnd_data.kib_connd_waitq);
1948 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
1952 kiblnd_close_conn(kib_conn_t *conn, int error)
1954 unsigned long flags;
1956 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1958 kiblnd_close_conn_locked(conn, error);
1960 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1964 kiblnd_handle_early_rxs(kib_conn_t *conn)
1966 unsigned long flags;
1969 LASSERT(!in_interrupt());
1970 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1972 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1973 while (!list_empty(&conn->ibc_early_rxs)) {
1974 rx = list_entry(conn->ibc_early_rxs.next,
1976 list_del(&rx->rx_list);
1977 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1979 kiblnd_handle_rx(rx);
1981 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1983 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1987 kiblnd_abort_txs(kib_conn_t *conn, struct list_head *txs)
1989 struct list_head zombies = LIST_HEAD_INIT(zombies);
1990 struct list_head *tmp;
1991 struct list_head *nxt;
1994 spin_lock(&conn->ibc_lock);
1996 list_for_each_safe(tmp, nxt, txs) {
1997 tx = list_entry(tmp, kib_tx_t, tx_list);
1999 if (txs == &conn->ibc_active_txs) {
2000 LASSERT(!tx->tx_queued);
2001 LASSERT(tx->tx_waiting ||
2002 tx->tx_sending != 0);
2004 LASSERT(tx->tx_queued);
2007 tx->tx_status = -ECONNABORTED;
2010 if (tx->tx_sending == 0) {
2012 list_del(&tx->tx_list);
2013 list_add(&tx->tx_list, &zombies);
2017 spin_unlock(&conn->ibc_lock);
2019 kiblnd_txlist_done(conn->ibc_peer->ibp_ni, &zombies, -ECONNABORTED);
2023 kiblnd_finalise_conn (kib_conn_t *conn)
2025 LASSERT (!in_interrupt());
2026 LASSERT (conn->ibc_state > IBLND_CONN_INIT);
2028 kiblnd_set_conn_state(conn, IBLND_CONN_DISCONNECTED);
2030 /* abort_receives moves QP state to IB_QPS_ERR. This is only required
2031 * for connections that didn't get as far as being connected, because
2032 * rdma_disconnect() does this for free. */
2033 kiblnd_abort_receives(conn);
2035 /* Complete all tx descs not waiting for sends to complete.
2036 * NB we should be safe from RDMA now that the QP has changed state */
2038 kiblnd_abort_txs(conn, &conn->ibc_tx_noops);
2039 kiblnd_abort_txs(conn, &conn->ibc_tx_queue);
2040 kiblnd_abort_txs(conn, &conn->ibc_tx_queue_rsrvd);
2041 kiblnd_abort_txs(conn, &conn->ibc_tx_queue_nocred);
2042 kiblnd_abort_txs(conn, &conn->ibc_active_txs);
2044 kiblnd_handle_early_rxs(conn);
2048 kiblnd_peer_connect_failed(kib_peer_t *peer, int active, int error)
2050 struct list_head zombies = LIST_HEAD_INIT(zombies);
2051 unsigned long flags;
2053 LASSERT (error != 0);
2054 LASSERT (!in_interrupt());
2056 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2059 LASSERT (peer->ibp_connecting > 0);
2060 peer->ibp_connecting--;
2062 LASSERT (peer->ibp_accepting > 0);
2063 peer->ibp_accepting--;
2066 if (kiblnd_peer_connecting(peer)) {
2067 /* another connection attempt under way... */
2068 write_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2073 peer->ibp_reconnected = 0;
2074 if (list_empty(&peer->ibp_conns)) {
2075 /* Take peer's blocked transmits to complete with error */
2076 list_add(&zombies, &peer->ibp_tx_queue);
2077 list_del_init(&peer->ibp_tx_queue);
2079 if (kiblnd_peer_active(peer))
2080 kiblnd_unlink_peer_locked(peer);
2082 peer->ibp_error = error;
2084 /* Can't have blocked transmits if there are connections */
2085 LASSERT(list_empty(&peer->ibp_tx_queue));
2088 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2090 kiblnd_peer_notify(peer);
2092 if (list_empty(&zombies))
2095 CNETERR("Deleting messages for %s: connection failed\n",
2096 libcfs_nid2str(peer->ibp_nid));
2098 kiblnd_txlist_done(peer->ibp_ni, &zombies, -EHOSTUNREACH);
2102 kiblnd_connreq_done(kib_conn_t *conn, int status)
2104 kib_peer_t *peer = conn->ibc_peer;
2106 struct list_head txs;
2107 unsigned long flags;
2110 active = (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2112 CDEBUG(D_NET,"%s: active(%d), version(%x), status(%d)\n",
2113 libcfs_nid2str(peer->ibp_nid), active,
2114 conn->ibc_version, status);
2116 LASSERT (!in_interrupt());
2117 LASSERT ((conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT &&
2118 peer->ibp_connecting > 0) ||
2119 (conn->ibc_state == IBLND_CONN_PASSIVE_WAIT &&
2120 peer->ibp_accepting > 0));
2122 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
2123 conn->ibc_connvars = NULL;
2126 /* failed to establish connection */
2127 kiblnd_peer_connect_failed(peer, active, status);
2128 kiblnd_finalise_conn(conn);
2132 /* connection established */
2133 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2135 conn->ibc_last_send = jiffies;
2136 kiblnd_set_conn_state(conn, IBLND_CONN_ESTABLISHED);
2137 kiblnd_peer_alive(peer);
2139 /* Add conn to peer's list and nuke any dangling conns from a different
2140 * peer instance... */
2141 kiblnd_conn_addref(conn); /* +1 ref for ibc_list */
2142 list_add(&conn->ibc_list, &peer->ibp_conns);
2143 peer->ibp_reconnected = 0;
2145 peer->ibp_connecting--;
2147 peer->ibp_accepting--;
2149 if (peer->ibp_version == 0) {
2150 peer->ibp_version = conn->ibc_version;
2151 peer->ibp_incarnation = conn->ibc_incarnation;
2154 if (peer->ibp_version != conn->ibc_version ||
2155 peer->ibp_incarnation != conn->ibc_incarnation) {
2156 kiblnd_close_stale_conns_locked(peer, conn->ibc_version,
2157 conn->ibc_incarnation);
2158 peer->ibp_version = conn->ibc_version;
2159 peer->ibp_incarnation = conn->ibc_incarnation;
2162 /* grab pending txs while I have the lock */
2163 list_add(&txs, &peer->ibp_tx_queue);
2164 list_del_init(&peer->ibp_tx_queue);
2166 if (!kiblnd_peer_active(peer) || /* peer has been deleted */
2167 conn->ibc_comms_error != 0) { /* error has happened already */
2168 lnet_ni_t *ni = peer->ibp_ni;
2170 /* start to shut down connection */
2171 kiblnd_close_conn_locked(conn, -ECONNABORTED);
2172 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2174 kiblnd_txlist_done(ni, &txs, -ECONNABORTED);
2179 /* +1 ref for myself, this connection is visible to other threads
2180 * now, refcount of peer:ibp_conns can be released by connection
2181 * close from either a different thread, or the calling of
2182 * kiblnd_check_sends_locked() below. See bz21911 for details.
2184 kiblnd_conn_addref(conn);
2185 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2187 /* Schedule blocked txs */
2188 spin_lock(&conn->ibc_lock);
2189 while (!list_empty(&txs)) {
2190 tx = list_entry(txs.next, kib_tx_t, tx_list);
2191 list_del(&tx->tx_list);
2193 kiblnd_queue_tx_locked(tx, conn);
2195 kiblnd_check_sends_locked(conn);
2196 spin_unlock(&conn->ibc_lock);
2198 /* schedule blocked rxs */
2199 kiblnd_handle_early_rxs(conn);
2200 kiblnd_conn_decref(conn);
2204 kiblnd_reject(struct rdma_cm_id *cmid, kib_rej_t *rej)
2208 rc = rdma_reject(cmid, rej, sizeof(*rej));
2211 CWARN("Error %d sending reject\n", rc);
2215 kiblnd_passive_connect(struct rdma_cm_id *cmid, void *priv, int priv_nob)
2217 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2218 kib_msg_t *reqmsg = priv;
2224 lnet_ni_t *ni = NULL;
2225 kib_net_t *net = NULL;
2227 struct rdma_conn_param cp;
2229 int version = IBLND_MSG_VERSION;
2230 unsigned long flags;
2232 struct sockaddr_in *peer_addr;
2233 LASSERT (!in_interrupt());
2235 /* cmid inherits 'context' from the corresponding listener id */
2236 ibdev = (kib_dev_t *)cmid->context;
2237 LASSERT (ibdev != NULL);
2239 memset(&rej, 0, sizeof(rej));
2240 rej.ibr_magic = IBLND_MSG_MAGIC;
2241 rej.ibr_why = IBLND_REJECT_FATAL;
2242 rej.ibr_cp.ibcp_max_msg_size = IBLND_MSG_SIZE;
2244 peer_addr = (struct sockaddr_in *)&(cmid->route.addr.dst_addr);
2245 if (*kiblnd_tunables.kib_require_priv_port &&
2246 ntohs(peer_addr->sin_port) >= PROT_SOCK) {
2247 __u32 ip = ntohl(peer_addr->sin_addr.s_addr);
2248 CERROR("Peer's port (%pI4h:%hu) is not privileged\n",
2249 &ip, ntohs(peer_addr->sin_port));
2253 if (priv_nob < offsetof(kib_msg_t, ibm_type)) {
2254 CERROR("Short connection request\n");
2258 /* Future protocol version compatibility support! If the
2259 * o2iblnd-specific protocol changes, or when LNET unifies
2260 * protocols over all LNDs, the initial connection will
2261 * negotiate a protocol version. I trap this here to avoid
2262 * console errors; the reject tells the peer which protocol I
2264 if (reqmsg->ibm_magic == LNET_PROTO_MAGIC ||
2265 reqmsg->ibm_magic == __swab32(LNET_PROTO_MAGIC))
2267 if (reqmsg->ibm_magic == IBLND_MSG_MAGIC &&
2268 reqmsg->ibm_version != IBLND_MSG_VERSION &&
2269 reqmsg->ibm_version != IBLND_MSG_VERSION_1)
2271 if (reqmsg->ibm_magic == __swab32(IBLND_MSG_MAGIC) &&
2272 reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION) &&
2273 reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION_1))
2276 rc = kiblnd_unpack_msg(reqmsg, priv_nob);
2278 CERROR("Can't parse connection request: %d\n", rc);
2282 nid = reqmsg->ibm_srcnid;
2283 ni = lnet_net2ni(LNET_NIDNET(reqmsg->ibm_dstnid));
2286 net = (kib_net_t *)ni->ni_data;
2287 rej.ibr_incarnation = net->ibn_incarnation;
2290 if (ni == NULL || /* no matching net */
2291 ni->ni_nid != reqmsg->ibm_dstnid || /* right NET, wrong NID! */
2292 net->ibn_dev != ibdev) { /* wrong device */
2293 CERROR("Can't accept conn from %s on %s (%s:%d:%pI4h): "
2294 "bad dst nid %s\n", libcfs_nid2str(nid),
2295 ni == NULL ? "NA" : libcfs_nid2str(ni->ni_nid),
2296 ibdev->ibd_ifname, ibdev->ibd_nnets,
2298 libcfs_nid2str(reqmsg->ibm_dstnid));
2303 /* check time stamp as soon as possible */
2304 if (reqmsg->ibm_dststamp != 0 &&
2305 reqmsg->ibm_dststamp != net->ibn_incarnation) {
2306 CWARN("Stale connection request\n");
2307 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2311 /* I can accept peer's version */
2312 version = reqmsg->ibm_version;
2314 if (reqmsg->ibm_type != IBLND_MSG_CONNREQ) {
2315 CERROR("Unexpected connreq msg type: %x from %s\n",
2316 reqmsg->ibm_type, libcfs_nid2str(nid));
2320 if (reqmsg->ibm_u.connparams.ibcp_queue_depth >
2321 kiblnd_msg_queue_size(version, ni)) {
2322 CERROR("Can't accept conn from %s, queue depth too large: "
2323 " %d (<=%d wanted)\n",
2324 libcfs_nid2str(nid),
2325 reqmsg->ibm_u.connparams.ibcp_queue_depth,
2326 kiblnd_msg_queue_size(version, ni));
2328 if (version == IBLND_MSG_VERSION)
2329 rej.ibr_why = IBLND_REJECT_MSG_QUEUE_SIZE;
2334 if (reqmsg->ibm_u.connparams.ibcp_max_frags >
2335 kiblnd_rdma_frags(version, ni)) {
2336 CWARN("Can't accept conn from %s (version %x): "
2337 "max_frags %d too large (%d wanted)\n",
2338 libcfs_nid2str(nid), version,
2339 reqmsg->ibm_u.connparams.ibcp_max_frags,
2340 kiblnd_rdma_frags(version, ni));
2342 if (version >= IBLND_MSG_VERSION)
2343 rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2346 } else if (reqmsg->ibm_u.connparams.ibcp_max_frags <
2347 kiblnd_rdma_frags(version, ni) &&
2348 net->ibn_fmr_ps == NULL) {
2349 CWARN("Can't accept conn from %s (version %x): "
2350 "max_frags %d incompatible without FMR pool "
2352 libcfs_nid2str(nid), version,
2353 reqmsg->ibm_u.connparams.ibcp_max_frags,
2354 kiblnd_rdma_frags(version, ni));
2356 if (version == IBLND_MSG_VERSION)
2357 rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2362 if (reqmsg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2363 CERROR("Can't accept %s: message size %d too big (%d max)\n",
2364 libcfs_nid2str(nid),
2365 reqmsg->ibm_u.connparams.ibcp_max_msg_size,
2370 /* assume 'nid' is a new peer; create */
2371 rc = kiblnd_create_peer(ni, &peer, nid);
2373 CERROR("Can't create peer for %s\n", libcfs_nid2str(nid));
2374 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2378 /* We have validated the peer's parameters so use those */
2379 peer->ibp_max_frags = reqmsg->ibm_u.connparams.ibcp_max_frags;
2380 peer->ibp_queue_depth = reqmsg->ibm_u.connparams.ibcp_queue_depth;
2382 write_lock_irqsave(g_lock, flags);
2384 peer2 = kiblnd_find_peer_locked(nid);
2385 if (peer2 != NULL) {
2386 if (peer2->ibp_version == 0) {
2387 peer2->ibp_version = version;
2388 peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2391 /* not the guy I've talked with */
2392 if (peer2->ibp_incarnation != reqmsg->ibm_srcstamp ||
2393 peer2->ibp_version != version) {
2394 kiblnd_close_peer_conns_locked(peer2, -ESTALE);
2396 if (kiblnd_peer_active(peer2)) {
2397 peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2398 peer2->ibp_version = version;
2400 write_unlock_irqrestore(g_lock, flags);
2402 CWARN("Conn stale %s version %x/%x incarnation %llu/%llu\n",
2403 libcfs_nid2str(nid), peer2->ibp_version, version,
2404 peer2->ibp_incarnation, reqmsg->ibm_srcstamp);
2406 kiblnd_peer_decref(peer);
2407 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2411 /* Tie-break connection race in favour of the higher NID.
2412 * If we keep running into a race condition multiple times,
2413 * we have to assume that the connection attempt with the
2414 * higher NID is stuck in a connecting state and will never
2415 * recover. As such, we pass through this if-block and let
2416 * the lower NID connection win so we can move forward.
2418 if (peer2->ibp_connecting != 0 &&
2419 nid < ni->ni_nid && peer2->ibp_races <
2420 MAX_CONN_RACES_BEFORE_ABORT) {
2422 write_unlock_irqrestore(g_lock, flags);
2424 CDEBUG(D_NET, "Conn race %s\n",
2425 libcfs_nid2str(peer2->ibp_nid));
2427 kiblnd_peer_decref(peer);
2428 rej.ibr_why = IBLND_REJECT_CONN_RACE;
2431 if (peer2->ibp_races >= MAX_CONN_RACES_BEFORE_ABORT)
2432 CNETERR("Conn race %s: unresolved after %d attempts, letting lower NID win\n",
2433 libcfs_nid2str(peer2->ibp_nid),
2434 MAX_CONN_RACES_BEFORE_ABORT);
2436 * passive connection is allowed even this peer is waiting for
2439 peer2->ibp_reconnecting = 0;
2440 peer2->ibp_races = 0;
2441 peer2->ibp_accepting++;
2442 kiblnd_peer_addref(peer2);
2444 /* Race with kiblnd_launch_tx (active connect) to create peer
2445 * so copy validated parameters since we now know what the
2446 * peer's limits are */
2447 peer2->ibp_max_frags = peer->ibp_max_frags;
2448 peer2->ibp_queue_depth = peer->ibp_queue_depth;
2450 write_unlock_irqrestore(g_lock, flags);
2451 kiblnd_peer_decref(peer);
2454 /* Brand new peer */
2455 LASSERT (peer->ibp_accepting == 0);
2456 LASSERT (peer->ibp_version == 0 &&
2457 peer->ibp_incarnation == 0);
2459 peer->ibp_accepting = 1;
2460 peer->ibp_version = version;
2461 peer->ibp_incarnation = reqmsg->ibm_srcstamp;
2463 /* I have a ref on ni that prevents it being shutdown */
2464 LASSERT (net->ibn_shutdown == 0);
2466 kiblnd_peer_addref(peer);
2467 list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid));
2469 write_unlock_irqrestore(g_lock, flags);
2472 conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_PASSIVE_WAIT, version);
2474 kiblnd_peer_connect_failed(peer, 0, -ENOMEM);
2475 kiblnd_peer_decref(peer);
2476 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2480 /* conn now "owns" cmid, so I return success from here on to ensure the
2481 * CM callback doesn't destroy cmid. */
2482 conn->ibc_incarnation = reqmsg->ibm_srcstamp;
2483 conn->ibc_credits = conn->ibc_queue_depth;
2484 conn->ibc_reserved_credits = conn->ibc_queue_depth;
2485 LASSERT(conn->ibc_credits + conn->ibc_reserved_credits +
2486 IBLND_OOB_MSGS(version) <= IBLND_RX_MSGS(conn));
2488 ackmsg = &conn->ibc_connvars->cv_msg;
2489 memset(ackmsg, 0, sizeof(*ackmsg));
2491 kiblnd_init_msg(ackmsg, IBLND_MSG_CONNACK,
2492 sizeof(ackmsg->ibm_u.connparams));
2493 ackmsg->ibm_u.connparams.ibcp_queue_depth = conn->ibc_queue_depth;
2494 ackmsg->ibm_u.connparams.ibcp_max_frags = conn->ibc_max_frags;
2495 ackmsg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2497 kiblnd_pack_msg(ni, ackmsg, version, 0, nid, reqmsg->ibm_srcstamp);
2499 memset(&cp, 0, sizeof(cp));
2500 cp.private_data = ackmsg;
2501 cp.private_data_len = ackmsg->ibm_nob;
2502 cp.responder_resources = 0; /* No atomic ops or RDMA reads */
2503 cp.initiator_depth = 0;
2504 cp.flow_control = 1;
2505 cp.retry_count = *kiblnd_tunables.kib_retry_count;
2506 cp.rnr_retry_count = *kiblnd_tunables.kib_rnr_retry_count;
2508 CDEBUG(D_NET, "Accept %s\n", libcfs_nid2str(nid));
2510 rc = rdma_accept(cmid, &cp);
2512 CERROR("Can't accept %s: %d\n", libcfs_nid2str(nid), rc);
2513 rej.ibr_version = version;
2514 rej.ibr_why = IBLND_REJECT_FATAL;
2516 kiblnd_reject(cmid, &rej);
2517 kiblnd_connreq_done(conn, rc);
2518 kiblnd_conn_decref(conn);
2526 rej.ibr_cp.ibcp_queue_depth =
2527 kiblnd_msg_queue_size(version, ni);
2528 rej.ibr_cp.ibcp_max_frags = kiblnd_rdma_frags(version, ni);
2532 rej.ibr_version = version;
2533 kiblnd_reject(cmid, &rej);
2535 return -ECONNREFUSED;
2539 kiblnd_check_reconnect(kib_conn_t *conn, int version,
2540 __u64 incarnation, int why, kib_connparams_t *cp)
2542 rwlock_t *glock = &kiblnd_data.kib_global_lock;
2543 kib_peer_t *peer = conn->ibc_peer;
2545 int msg_size = IBLND_MSG_SIZE;
2549 unsigned long flags;
2551 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2552 LASSERT(peer->ibp_connecting > 0); /* 'conn' at least */
2553 LASSERT(!peer->ibp_reconnecting);
2556 msg_size = cp->ibcp_max_msg_size;
2557 frag_num = cp->ibcp_max_frags;
2558 queue_dep = cp->ibcp_queue_depth;
2561 write_lock_irqsave(glock, flags);
2562 /* retry connection if it's still needed and no other connection
2563 * attempts (active or passive) are in progress
2564 * NB: reconnect is still needed even when ibp_tx_queue is
2565 * empty if ibp_version != version because reconnect may be
2566 * initiated by kiblnd_query() */
2567 reconnect = (!list_empty(&peer->ibp_tx_queue) ||
2568 peer->ibp_version != version) &&
2569 peer->ibp_connecting == 1 &&
2570 peer->ibp_accepting == 0;
2581 case IBLND_REJECT_RDMA_FRAGS: {
2582 struct lnet_ioctl_config_lnd_tunables *tunables;
2585 reason = "can't negotiate max frags";
2588 tunables = peer->ibp_ni->ni_lnd_tunables;
2589 if (!tunables->lt_tun_u.lt_o2ib.lnd_map_on_demand) {
2590 reason = "map_on_demand must be enabled";
2593 if (conn->ibc_max_frags <= frag_num) {
2594 reason = "unsupported max frags";
2598 peer->ibp_max_frags = frag_num;
2599 reason = "rdma fragments";
2602 case IBLND_REJECT_MSG_QUEUE_SIZE:
2604 reason = "can't negotiate queue depth";
2607 if (conn->ibc_queue_depth <= queue_dep) {
2608 reason = "unsupported queue depth";
2612 peer->ibp_queue_depth = queue_dep;
2613 reason = "queue depth";
2616 case IBLND_REJECT_CONN_STALE:
2620 case IBLND_REJECT_CONN_RACE:
2621 reason = "conn race";
2624 case IBLND_REJECT_CONN_UNCOMPAT:
2625 reason = "version negotiation";
2629 conn->ibc_reconnect = 1;
2630 peer->ibp_reconnecting = 1;
2631 peer->ibp_version = version;
2632 if (incarnation != 0)
2633 peer->ibp_incarnation = incarnation;
2635 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2637 CNETERR("%s: %s (%s), %x, %x, msg_size: %d, queue_depth: %d/%d, max_frags: %d/%d\n",
2638 libcfs_nid2str(peer->ibp_nid),
2639 reconnect ? "reconnect" : "don't reconnect",
2640 reason, IBLND_MSG_VERSION, version, msg_size,
2641 conn->ibc_queue_depth, queue_dep,
2642 conn->ibc_max_frags, frag_num);
2644 * if conn::ibc_reconnect is TRUE, connd will reconnect to the peer
2645 * while destroying the zombie
2650 kiblnd_rejected (kib_conn_t *conn, int reason, void *priv, int priv_nob)
2652 kib_peer_t *peer = conn->ibc_peer;
2654 LASSERT (!in_interrupt());
2655 LASSERT (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2658 case IB_CM_REJ_STALE_CONN:
2659 kiblnd_check_reconnect(conn, IBLND_MSG_VERSION, 0,
2660 IBLND_REJECT_CONN_STALE, NULL);
2663 case IB_CM_REJ_INVALID_SERVICE_ID:
2664 CNETERR("%s rejected: no listener at %d\n",
2665 libcfs_nid2str(peer->ibp_nid),
2666 *kiblnd_tunables.kib_service);
2669 case IB_CM_REJ_CONSUMER_DEFINED:
2670 if (priv_nob >= offsetof(kib_rej_t, ibr_padding)) {
2671 kib_rej_t *rej = priv;
2672 kib_connparams_t *cp = NULL;
2674 __u64 incarnation = -1;
2676 /* NB. default incarnation is -1 because:
2677 * a) V1 will ignore dst incarnation in connreq.
2678 * b) V2 will provide incarnation while rejecting me,
2679 * -1 will be overwrote.
2681 * if I try to connect to a V1 peer with V2 protocol,
2682 * it rejected me then upgrade to V2, I have no idea
2683 * about the upgrading and try to reconnect with V1,
2684 * in this case upgraded V2 can find out I'm trying to
2685 * talk to the old guy and reject me(incarnation is -1).
2688 if (rej->ibr_magic == __swab32(IBLND_MSG_MAGIC) ||
2689 rej->ibr_magic == __swab32(LNET_PROTO_MAGIC)) {
2690 __swab32s(&rej->ibr_magic);
2691 __swab16s(&rej->ibr_version);
2695 if (priv_nob >= sizeof(kib_rej_t) &&
2696 rej->ibr_version > IBLND_MSG_VERSION_1) {
2697 /* priv_nob is always 148 in current version
2698 * of OFED, so we still need to check version.
2699 * (define of IB_CM_REJ_PRIVATE_DATA_SIZE) */
2703 __swab64s(&rej->ibr_incarnation);
2704 __swab16s(&cp->ibcp_queue_depth);
2705 __swab16s(&cp->ibcp_max_frags);
2706 __swab32s(&cp->ibcp_max_msg_size);
2709 incarnation = rej->ibr_incarnation;
2712 if (rej->ibr_magic != IBLND_MSG_MAGIC &&
2713 rej->ibr_magic != LNET_PROTO_MAGIC) {
2714 CERROR("%s rejected: consumer defined fatal error\n",
2715 libcfs_nid2str(peer->ibp_nid));
2719 if (rej->ibr_version != IBLND_MSG_VERSION &&
2720 rej->ibr_version != IBLND_MSG_VERSION_1) {
2721 CERROR("%s rejected: o2iblnd version %x error\n",
2722 libcfs_nid2str(peer->ibp_nid),
2727 if (rej->ibr_why == IBLND_REJECT_FATAL &&
2728 rej->ibr_version == IBLND_MSG_VERSION_1) {
2729 CDEBUG(D_NET, "rejected by old version peer %s: %x\n",
2730 libcfs_nid2str(peer->ibp_nid), rej->ibr_version);
2732 if (conn->ibc_version != IBLND_MSG_VERSION_1)
2733 rej->ibr_why = IBLND_REJECT_CONN_UNCOMPAT;
2736 switch (rej->ibr_why) {
2737 case IBLND_REJECT_CONN_RACE:
2738 case IBLND_REJECT_CONN_STALE:
2739 case IBLND_REJECT_CONN_UNCOMPAT:
2740 case IBLND_REJECT_MSG_QUEUE_SIZE:
2741 case IBLND_REJECT_RDMA_FRAGS:
2742 kiblnd_check_reconnect(conn, rej->ibr_version,
2743 incarnation, rej->ibr_why, cp);
2746 case IBLND_REJECT_NO_RESOURCES:
2747 CERROR("%s rejected: o2iblnd no resources\n",
2748 libcfs_nid2str(peer->ibp_nid));
2751 case IBLND_REJECT_FATAL:
2752 CERROR("%s rejected: o2iblnd fatal error\n",
2753 libcfs_nid2str(peer->ibp_nid));
2757 CERROR("%s rejected: o2iblnd reason %d\n",
2758 libcfs_nid2str(peer->ibp_nid),
2766 CNETERR("%s rejected: reason %d, size %d\n",
2767 libcfs_nid2str(peer->ibp_nid), reason, priv_nob);
2771 kiblnd_connreq_done(conn, -ECONNREFUSED);
2775 kiblnd_check_connreply (kib_conn_t *conn, void *priv, int priv_nob)
2777 kib_peer_t *peer = conn->ibc_peer;
2778 lnet_ni_t *ni = peer->ibp_ni;
2779 kib_net_t *net = ni->ni_data;
2780 kib_msg_t *msg = priv;
2781 int ver = conn->ibc_version;
2782 int rc = kiblnd_unpack_msg(msg, priv_nob);
2783 unsigned long flags;
2785 LASSERT (net != NULL);
2788 CERROR("Can't unpack connack from %s: %d\n",
2789 libcfs_nid2str(peer->ibp_nid), rc);
2793 if (msg->ibm_type != IBLND_MSG_CONNACK) {
2794 CERROR("Unexpected message %d from %s\n",
2795 msg->ibm_type, libcfs_nid2str(peer->ibp_nid));
2800 if (ver != msg->ibm_version) {
2801 CERROR("%s replied version %x is different with "
2802 "requested version %x\n",
2803 libcfs_nid2str(peer->ibp_nid), msg->ibm_version, ver);
2808 if (msg->ibm_u.connparams.ibcp_queue_depth >
2809 conn->ibc_queue_depth) {
2810 CERROR("%s has incompatible queue depth %d (<=%d wanted)\n",
2811 libcfs_nid2str(peer->ibp_nid),
2812 msg->ibm_u.connparams.ibcp_queue_depth,
2813 conn->ibc_queue_depth);
2818 if (msg->ibm_u.connparams.ibcp_max_frags >
2819 conn->ibc_max_frags) {
2820 CERROR("%s has incompatible max_frags %d (<=%d wanted)\n",
2821 libcfs_nid2str(peer->ibp_nid),
2822 msg->ibm_u.connparams.ibcp_max_frags,
2823 conn->ibc_max_frags);
2828 if (msg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2829 CERROR("%s max message size %d too big (%d max)\n",
2830 libcfs_nid2str(peer->ibp_nid),
2831 msg->ibm_u.connparams.ibcp_max_msg_size,
2837 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2838 if (msg->ibm_dstnid == ni->ni_nid &&
2839 msg->ibm_dststamp == net->ibn_incarnation)
2843 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2846 CERROR("Bad connection reply from %s, rc = %d, "
2847 "version: %x max_frags: %d\n",
2848 libcfs_nid2str(peer->ibp_nid), rc,
2849 msg->ibm_version, msg->ibm_u.connparams.ibcp_max_frags);
2853 conn->ibc_incarnation = msg->ibm_srcstamp;
2854 conn->ibc_credits = msg->ibm_u.connparams.ibcp_queue_depth;
2855 conn->ibc_reserved_credits = msg->ibm_u.connparams.ibcp_queue_depth;
2856 conn->ibc_queue_depth = msg->ibm_u.connparams.ibcp_queue_depth;
2857 conn->ibc_max_frags = msg->ibm_u.connparams.ibcp_max_frags;
2858 LASSERT(conn->ibc_credits + conn->ibc_reserved_credits +
2859 IBLND_OOB_MSGS(ver) <= IBLND_RX_MSGS(conn));
2861 kiblnd_connreq_done(conn, 0);
2865 /* NB My QP has already established itself, so I handle anything going
2866 * wrong here by setting ibc_comms_error.
2867 * kiblnd_connreq_done(0) moves the conn state to ESTABLISHED, but then
2868 * immediately tears it down. */
2871 conn->ibc_comms_error = rc;
2872 kiblnd_connreq_done(conn, 0);
2876 kiblnd_active_connect (struct rdma_cm_id *cmid)
2878 kib_peer_t *peer = (kib_peer_t *)cmid->context;
2881 struct rdma_conn_param cp;
2884 unsigned long flags;
2887 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2889 incarnation = peer->ibp_incarnation;
2890 version = (peer->ibp_version == 0) ? IBLND_MSG_VERSION :
2893 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2895 conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_ACTIVE_CONNECT,
2898 kiblnd_peer_connect_failed(peer, 1, -ENOMEM);
2899 kiblnd_peer_decref(peer); /* lose cmid's ref */
2903 /* conn "owns" cmid now, so I return success from here on to ensure the
2904 * CM callback doesn't destroy cmid. conn also takes over cmid's ref
2907 msg = &conn->ibc_connvars->cv_msg;
2909 memset(msg, 0, sizeof(*msg));
2910 kiblnd_init_msg(msg, IBLND_MSG_CONNREQ, sizeof(msg->ibm_u.connparams));
2911 msg->ibm_u.connparams.ibcp_queue_depth = conn->ibc_queue_depth;
2912 msg->ibm_u.connparams.ibcp_max_frags = conn->ibc_max_frags;
2913 msg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2915 kiblnd_pack_msg(peer->ibp_ni, msg, version,
2916 0, peer->ibp_nid, incarnation);
2918 memset(&cp, 0, sizeof(cp));
2919 cp.private_data = msg;
2920 cp.private_data_len = msg->ibm_nob;
2921 cp.responder_resources = 0; /* No atomic ops or RDMA reads */
2922 cp.initiator_depth = 0;
2923 cp.flow_control = 1;
2924 cp.retry_count = *kiblnd_tunables.kib_retry_count;
2925 cp.rnr_retry_count = *kiblnd_tunables.kib_rnr_retry_count;
2927 LASSERT(cmid->context == (void *)conn);
2928 LASSERT(conn->ibc_cmid == cmid);
2930 rc = rdma_connect(cmid, &cp);
2932 CERROR("Can't connect to %s: %d\n",
2933 libcfs_nid2str(peer->ibp_nid), rc);
2934 kiblnd_connreq_done(conn, rc);
2935 kiblnd_conn_decref(conn);
2942 kiblnd_cm_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2948 switch (event->event) {
2950 CERROR("Unexpected event: %d, status: %d\n",
2951 event->event, event->status);
2954 case RDMA_CM_EVENT_CONNECT_REQUEST:
2955 /* destroy cmid on failure */
2956 rc = kiblnd_passive_connect(cmid,
2957 (void *)KIBLND_CONN_PARAM(event),
2958 KIBLND_CONN_PARAM_LEN(event));
2959 CDEBUG(D_NET, "connreq: %d\n", rc);
2962 case RDMA_CM_EVENT_ADDR_ERROR:
2963 peer = (kib_peer_t *)cmid->context;
2964 CNETERR("%s: ADDR ERROR %d\n",
2965 libcfs_nid2str(peer->ibp_nid), event->status);
2966 kiblnd_peer_connect_failed(peer, 1, -EHOSTUNREACH);
2967 kiblnd_peer_decref(peer);
2968 return -EHOSTUNREACH; /* rc != 0 destroys cmid */
2970 case RDMA_CM_EVENT_ADDR_RESOLVED:
2971 peer = (kib_peer_t *)cmid->context;
2973 CDEBUG(D_NET,"%s Addr resolved: %d\n",
2974 libcfs_nid2str(peer->ibp_nid), event->status);
2976 if (event->status != 0) {
2977 CNETERR("Can't resolve address for %s: %d\n",
2978 libcfs_nid2str(peer->ibp_nid), event->status);
2981 rc = rdma_resolve_route(
2982 cmid, *kiblnd_tunables.kib_timeout * 1000);
2985 /* Can't initiate route resolution */
2986 CERROR("Can't resolve route for %s: %d\n",
2987 libcfs_nid2str(peer->ibp_nid), rc);
2989 kiblnd_peer_connect_failed(peer, 1, rc);
2990 kiblnd_peer_decref(peer);
2991 return rc; /* rc != 0 destroys cmid */
2993 case RDMA_CM_EVENT_ROUTE_ERROR:
2994 peer = (kib_peer_t *)cmid->context;
2995 CNETERR("%s: ROUTE ERROR %d\n",
2996 libcfs_nid2str(peer->ibp_nid), event->status);
2997 kiblnd_peer_connect_failed(peer, 1, -EHOSTUNREACH);
2998 kiblnd_peer_decref(peer);
2999 return -EHOSTUNREACH; /* rc != 0 destroys cmid */
3001 case RDMA_CM_EVENT_ROUTE_RESOLVED:
3002 peer = (kib_peer_t *)cmid->context;
3003 CDEBUG(D_NET,"%s Route resolved: %d\n",
3004 libcfs_nid2str(peer->ibp_nid), event->status);
3006 if (event->status == 0)
3007 return kiblnd_active_connect(cmid);
3009 CNETERR("Can't resolve route for %s: %d\n",
3010 libcfs_nid2str(peer->ibp_nid), event->status);
3011 kiblnd_peer_connect_failed(peer, 1, event->status);
3012 kiblnd_peer_decref(peer);
3013 return event->status; /* rc != 0 destroys cmid */
3015 case RDMA_CM_EVENT_UNREACHABLE:
3016 conn = (kib_conn_t *)cmid->context;
3017 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
3018 conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
3019 CNETERR("%s: UNREACHABLE %d\n",
3020 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
3021 kiblnd_connreq_done(conn, -ENETDOWN);
3022 kiblnd_conn_decref(conn);
3025 case RDMA_CM_EVENT_CONNECT_ERROR:
3026 conn = (kib_conn_t *)cmid->context;
3027 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
3028 conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
3029 CNETERR("%s: CONNECT ERROR %d\n",
3030 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
3031 kiblnd_connreq_done(conn, -ENOTCONN);
3032 kiblnd_conn_decref(conn);
3035 case RDMA_CM_EVENT_REJECTED:
3036 conn = (kib_conn_t *)cmid->context;
3037 switch (conn->ibc_state) {
3041 case IBLND_CONN_PASSIVE_WAIT:
3042 CERROR ("%s: REJECTED %d\n",
3043 libcfs_nid2str(conn->ibc_peer->ibp_nid),
3045 kiblnd_connreq_done(conn, -ECONNRESET);
3048 case IBLND_CONN_ACTIVE_CONNECT:
3049 kiblnd_rejected(conn, event->status,
3050 (void *)KIBLND_CONN_PARAM(event),
3051 KIBLND_CONN_PARAM_LEN(event));
3054 kiblnd_conn_decref(conn);
3057 case RDMA_CM_EVENT_ESTABLISHED:
3058 conn = (kib_conn_t *)cmid->context;
3059 switch (conn->ibc_state) {
3063 case IBLND_CONN_PASSIVE_WAIT:
3064 CDEBUG(D_NET, "ESTABLISHED (passive): %s\n",
3065 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3066 kiblnd_connreq_done(conn, 0);
3069 case IBLND_CONN_ACTIVE_CONNECT:
3070 CDEBUG(D_NET, "ESTABLISHED(active): %s\n",
3071 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3072 kiblnd_check_connreply(conn,
3073 (void *)KIBLND_CONN_PARAM(event),
3074 KIBLND_CONN_PARAM_LEN(event));
3077 /* net keeps its ref on conn! */
3080 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
3081 CDEBUG(D_NET, "Ignore TIMEWAIT_EXIT event\n");
3084 case RDMA_CM_EVENT_DISCONNECTED:
3085 conn = (kib_conn_t *)cmid->context;
3086 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
3087 CERROR("%s DISCONNECTED\n",
3088 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3089 kiblnd_connreq_done(conn, -ECONNRESET);
3091 kiblnd_close_conn(conn, 0);
3093 kiblnd_conn_decref(conn);
3094 cmid->context = NULL;
3097 case RDMA_CM_EVENT_DEVICE_REMOVAL:
3098 LCONSOLE_ERROR_MSG(0x131,
3099 "Received notification of device removal\n"
3100 "Please shutdown LNET to allow this to proceed\n");
3101 /* Can't remove network from underneath LNET for now, so I have
3105 case RDMA_CM_EVENT_ADDR_CHANGE:
3106 LCONSOLE_INFO("Physical link changed (eg hca/port)\n");
3112 kiblnd_check_txs_locked(kib_conn_t *conn, struct list_head *txs)
3115 struct list_head *ttmp;
3117 list_for_each(ttmp, txs) {
3118 tx = list_entry(ttmp, kib_tx_t, tx_list);
3120 if (txs != &conn->ibc_active_txs) {
3121 LASSERT(tx->tx_queued);
3123 LASSERT(!tx->tx_queued);
3124 LASSERT(tx->tx_waiting || tx->tx_sending != 0);
3127 if (cfs_time_aftereq(jiffies, tx->tx_deadline)) {
3128 CERROR("Timed out tx: %s, %lu seconds\n",
3129 kiblnd_queue2str(conn, txs),
3130 cfs_duration_sec(jiffies - tx->tx_deadline));
3139 kiblnd_conn_timed_out_locked(kib_conn_t *conn)
3141 return kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue) ||
3142 kiblnd_check_txs_locked(conn, &conn->ibc_tx_noops) ||
3143 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_rsrvd) ||
3144 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_nocred) ||
3145 kiblnd_check_txs_locked(conn, &conn->ibc_active_txs);
3149 kiblnd_check_conns (int idx)
3151 struct list_head closes = LIST_HEAD_INIT(closes);
3152 struct list_head checksends = LIST_HEAD_INIT(checksends);
3153 struct list_head *peers = &kiblnd_data.kib_peers[idx];
3154 struct list_head *ptmp;
3157 struct list_head *ctmp;
3158 unsigned long flags;
3160 /* NB. We expect to have a look at all the peers and not find any
3161 * RDMAs to time out, so we just use a shared lock while we
3163 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3165 list_for_each(ptmp, peers) {
3166 peer = list_entry(ptmp, kib_peer_t, ibp_list);
3168 list_for_each(ctmp, &peer->ibp_conns) {
3172 conn = list_entry(ctmp, kib_conn_t, ibc_list);
3174 LASSERT(conn->ibc_state == IBLND_CONN_ESTABLISHED);
3176 spin_lock(&conn->ibc_lock);
3178 sendnoop = kiblnd_need_noop(conn);
3179 timedout = kiblnd_conn_timed_out_locked(conn);
3180 if (!sendnoop && !timedout) {
3181 spin_unlock(&conn->ibc_lock);
3186 CERROR("Timed out RDMA with %s (%lu): "
3187 "c: %u, oc: %u, rc: %u\n",
3188 libcfs_nid2str(peer->ibp_nid),
3189 cfs_duration_sec(cfs_time_current() -
3190 peer->ibp_last_alive),
3192 conn->ibc_outstanding_credits,
3193 conn->ibc_reserved_credits);
3194 list_add(&conn->ibc_connd_list, &closes);
3196 list_add(&conn->ibc_connd_list, &checksends);
3198 /* +ref for 'closes' or 'checksends' */
3199 kiblnd_conn_addref(conn);
3201 spin_unlock(&conn->ibc_lock);
3205 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3207 /* Handle timeout by closing the whole
3208 * connection. We can only be sure RDMA activity
3209 * has ceased once the QP has been modified. */
3210 while (!list_empty(&closes)) {
3211 conn = list_entry(closes.next,
3212 kib_conn_t, ibc_connd_list);
3213 list_del(&conn->ibc_connd_list);
3214 kiblnd_close_conn(conn, -ETIMEDOUT);
3215 kiblnd_conn_decref(conn);
3218 /* In case we have enough credits to return via a
3219 * NOOP, but there were no non-blocking tx descs
3220 * free to do it last time... */
3221 while (!list_empty(&checksends)) {
3222 conn = list_entry(checksends.next,
3223 kib_conn_t, ibc_connd_list);
3224 list_del(&conn->ibc_connd_list);
3226 spin_lock(&conn->ibc_lock);
3227 kiblnd_check_sends_locked(conn);
3228 spin_unlock(&conn->ibc_lock);
3230 kiblnd_conn_decref(conn);
3235 kiblnd_disconnect_conn (kib_conn_t *conn)
3237 LASSERT (!in_interrupt());
3238 LASSERT (current == kiblnd_data.kib_connd);
3239 LASSERT (conn->ibc_state == IBLND_CONN_CLOSING);
3241 rdma_disconnect(conn->ibc_cmid);
3242 kiblnd_finalise_conn(conn);
3244 kiblnd_peer_notify(conn->ibc_peer);
3248 * High-water for reconnection to the same peer, reconnection attempt should
3249 * be delayed after trying more than KIB_RECONN_HIGH_RACE.
3251 #define KIB_RECONN_HIGH_RACE 10
3253 * Allow connd to take a break and handle other things after consecutive
3254 * reconnection attemps.
3256 #define KIB_RECONN_BREAK 100
3259 kiblnd_connd (void *arg)
3261 spinlock_t *lock= &kiblnd_data.kib_connd_lock;
3263 unsigned long flags;
3269 unsigned long deadline = jiffies;
3271 cfs_block_allsigs();
3273 init_waitqueue_entry(&wait, current);
3274 kiblnd_data.kib_connd = current;
3276 spin_lock_irqsave(lock, flags);
3278 while (!kiblnd_data.kib_shutdown) {
3283 if (!list_empty(&kiblnd_data.kib_connd_zombies)) {
3284 kib_peer_t *peer = NULL;
3286 conn = list_entry(kiblnd_data.kib_connd_zombies.next,
3287 kib_conn_t, ibc_list);
3288 list_del(&conn->ibc_list);
3289 if (conn->ibc_reconnect) {
3290 peer = conn->ibc_peer;
3291 kiblnd_peer_addref(peer);
3294 spin_unlock_irqrestore(lock, flags);
3297 kiblnd_destroy_conn(conn, !peer);
3299 spin_lock_irqsave(lock, flags);
3303 conn->ibc_peer = peer;
3304 if (peer->ibp_reconnected < KIB_RECONN_HIGH_RACE)
3305 list_add_tail(&conn->ibc_list,
3306 &kiblnd_data.kib_reconn_list);
3308 list_add_tail(&conn->ibc_list,
3309 &kiblnd_data.kib_reconn_wait);
3312 if (!list_empty(&kiblnd_data.kib_connd_conns)) {
3313 conn = list_entry(kiblnd_data.kib_connd_conns.next,
3314 kib_conn_t, ibc_list);
3315 list_del(&conn->ibc_list);
3317 spin_unlock_irqrestore(lock, flags);
3320 kiblnd_disconnect_conn(conn);
3321 kiblnd_conn_decref(conn);
3323 spin_lock_irqsave(lock, flags);
3326 while (reconn < KIB_RECONN_BREAK) {
3327 if (kiblnd_data.kib_reconn_sec != get_seconds()) {
3328 kiblnd_data.kib_reconn_sec = get_seconds();
3329 list_splice_init(&kiblnd_data.kib_reconn_wait,
3330 &kiblnd_data.kib_reconn_list);
3333 if (list_empty(&kiblnd_data.kib_reconn_list))
3336 conn = list_entry(kiblnd_data.kib_reconn_list.next,
3337 kib_conn_t, ibc_list);
3338 list_del(&conn->ibc_list);
3340 spin_unlock_irqrestore(lock, flags);
3343 reconn += kiblnd_reconnect_peer(conn->ibc_peer);
3344 kiblnd_peer_decref(conn->ibc_peer);
3345 LIBCFS_FREE(conn, sizeof(*conn));
3347 spin_lock_irqsave(lock, flags);
3350 /* careful with the jiffy wrap... */
3351 timeout = (int)(deadline - jiffies);
3355 int chunk = kiblnd_data.kib_peer_hash_size;
3357 spin_unlock_irqrestore(lock, flags);
3360 /* Time to check for RDMA timeouts on a few more
3361 * peers: I do checks every 'p' seconds on a
3362 * proportion of the peer table and I need to check
3363 * every connection 'n' times within a timeout
3364 * interval, to ensure I detect a timeout on any
3365 * connection within (n+1)/n times the timeout
3368 if (*kiblnd_tunables.kib_timeout > n * p)
3369 chunk = (chunk * n * p) /
3370 *kiblnd_tunables.kib_timeout;
3374 for (i = 0; i < chunk; i++) {
3375 kiblnd_check_conns(peer_index);
3376 peer_index = (peer_index + 1) %
3377 kiblnd_data.kib_peer_hash_size;
3380 deadline += msecs_to_jiffies(p * MSEC_PER_SEC);
3381 spin_lock_irqsave(lock, flags);
3387 /* Nothing to do for 'timeout' */
3388 set_current_state(TASK_INTERRUPTIBLE);
3389 add_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3390 spin_unlock_irqrestore(lock, flags);
3392 schedule_timeout(timeout);
3394 set_current_state(TASK_RUNNING);
3395 remove_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3396 spin_lock_irqsave(lock, flags);
3399 spin_unlock_irqrestore(lock, flags);
3401 kiblnd_thread_fini();
3406 kiblnd_qp_event(struct ib_event *event, void *arg)
3408 kib_conn_t *conn = arg;
3410 switch (event->event) {
3411 case IB_EVENT_COMM_EST:
3412 CDEBUG(D_NET, "%s established\n",
3413 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3414 /* We received a packet but connection isn't established
3415 * probably handshake packet was lost, so free to
3416 * force make connection established */
3417 rdma_notify(conn->ibc_cmid, IB_EVENT_COMM_EST);
3421 CERROR("%s: Async QP event type %d\n",
3422 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3428 kiblnd_complete (struct ib_wc *wc)
3430 switch (kiblnd_wreqid2type(wc->wr_id)) {
3435 if (wc->status != IB_WC_SUCCESS &&
3436 wc->status != IB_WC_WR_FLUSH_ERR)
3437 CNETERR("FastReg failed: %d\n", wc->status);
3440 case IBLND_WID_RDMA:
3441 /* We only get RDMA completion notification if it fails. All
3442 * subsequent work items, including the final SEND will fail
3443 * too. However we can't print out any more info about the
3444 * failing RDMA because 'tx' might be back on the idle list or
3445 * even reused already if we didn't manage to post all our work
3447 CNETERR("RDMA (tx: %p) failed: %d\n",
3448 kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3452 kiblnd_tx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3456 kiblnd_rx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status,
3463 kiblnd_cq_completion(struct ib_cq *cq, void *arg)
3465 /* NB I'm not allowed to schedule this conn once its refcount has
3466 * reached 0. Since fundamentally I'm racing with scheduler threads
3467 * consuming my CQ I could be called after all completions have
3468 * occurred. But in this case, ibc_nrx == 0 && ibc_nsends_posted == 0
3469 * and this CQ is about to be destroyed so I NOOP. */
3470 kib_conn_t *conn = (kib_conn_t *)arg;
3471 struct kib_sched_info *sched = conn->ibc_sched;
3472 unsigned long flags;
3474 LASSERT(cq == conn->ibc_cq);
3476 spin_lock_irqsave(&sched->ibs_lock, flags);
3478 conn->ibc_ready = 1;
3480 if (!conn->ibc_scheduled &&
3481 (conn->ibc_nrx > 0 ||
3482 conn->ibc_nsends_posted > 0)) {
3483 kiblnd_conn_addref(conn); /* +1 ref for sched_conns */
3484 conn->ibc_scheduled = 1;
3485 list_add_tail(&conn->ibc_sched_list, &sched->ibs_conns);
3487 if (waitqueue_active(&sched->ibs_waitq))
3488 wake_up(&sched->ibs_waitq);
3491 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3495 kiblnd_cq_event(struct ib_event *event, void *arg)
3497 kib_conn_t *conn = arg;
3499 CERROR("%s: async CQ event type %d\n",
3500 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3504 kiblnd_scheduler(void *arg)
3506 long id = (long)arg;
3507 struct kib_sched_info *sched;
3510 unsigned long flags;
3516 cfs_block_allsigs();
3518 init_waitqueue_entry(&wait, current);
3520 sched = kiblnd_data.kib_scheds[KIB_THREAD_CPT(id)];
3522 rc = cfs_cpt_bind(lnet_cpt_table(), sched->ibs_cpt);
3524 CWARN("Failed to bind on CPT %d, please verify whether "
3525 "all CPUs are healthy and reload modules if necessary, "
3526 "otherwise your system might under risk of low "
3527 "performance\n", sched->ibs_cpt);
3530 spin_lock_irqsave(&sched->ibs_lock, flags);
3532 while (!kiblnd_data.kib_shutdown) {
3533 if (busy_loops++ >= IBLND_RESCHED) {
3534 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3539 spin_lock_irqsave(&sched->ibs_lock, flags);
3544 if (!list_empty(&sched->ibs_conns)) {
3545 conn = list_entry(sched->ibs_conns.next,
3546 kib_conn_t, ibc_sched_list);
3547 /* take over kib_sched_conns' ref on conn... */
3548 LASSERT(conn->ibc_scheduled);
3549 list_del(&conn->ibc_sched_list);
3550 conn->ibc_ready = 0;
3552 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3554 wc.wr_id = IBLND_WID_INVAL;
3556 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3558 rc = ib_req_notify_cq(conn->ibc_cq,
3561 CWARN("%s: ib_req_notify_cq failed: %d, "
3562 "closing connection\n",
3563 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
3564 kiblnd_close_conn(conn, -EIO);
3565 kiblnd_conn_decref(conn);
3566 spin_lock_irqsave(&sched->ibs_lock,
3571 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3574 if (unlikely(rc > 0 && wc.wr_id == IBLND_WID_INVAL)) {
3576 "ib_poll_cq (rc: %d) returned invalid "
3577 "wr_id, opcode %d, status: %d, "
3578 "vendor_err: %d, conn: %s status: %d\n"
3579 "please upgrade firmware and OFED or "
3580 "contact vendor.\n", rc,
3581 wc.opcode, wc.status, wc.vendor_err,
3582 libcfs_nid2str(conn->ibc_peer->ibp_nid),
3588 CWARN("%s: ib_poll_cq failed: %d, "
3589 "closing connection\n",
3590 libcfs_nid2str(conn->ibc_peer->ibp_nid),
3592 kiblnd_close_conn(conn, -EIO);
3593 kiblnd_conn_decref(conn);
3594 spin_lock_irqsave(&sched->ibs_lock, flags);
3598 spin_lock_irqsave(&sched->ibs_lock, flags);
3600 if (rc != 0 || conn->ibc_ready) {
3601 /* There may be another completion waiting; get
3602 * another scheduler to check while I handle
3604 /* +1 ref for sched_conns */
3605 kiblnd_conn_addref(conn);
3606 list_add_tail(&conn->ibc_sched_list,
3608 if (waitqueue_active(&sched->ibs_waitq))
3609 wake_up(&sched->ibs_waitq);
3611 conn->ibc_scheduled = 0;
3615 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3616 kiblnd_complete(&wc);
3618 spin_lock_irqsave(&sched->ibs_lock, flags);
3621 kiblnd_conn_decref(conn); /* ...drop my ref from above */
3628 set_current_state(TASK_INTERRUPTIBLE);
3629 add_wait_queue_exclusive(&sched->ibs_waitq, &wait);
3630 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3635 remove_wait_queue(&sched->ibs_waitq, &wait);
3636 set_current_state(TASK_RUNNING);
3637 spin_lock_irqsave(&sched->ibs_lock, flags);
3640 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3642 kiblnd_thread_fini();
3647 kiblnd_failover_thread(void *arg)
3649 rwlock_t *glock = &kiblnd_data.kib_global_lock;
3652 unsigned long flags;
3655 LASSERT(*kiblnd_tunables.kib_dev_failover != 0);
3657 cfs_block_allsigs();
3659 init_waitqueue_entry(&wait, current);
3660 write_lock_irqsave(glock, flags);
3662 while (!kiblnd_data.kib_shutdown) {
3663 int do_failover = 0;
3666 list_for_each_entry(dev, &kiblnd_data.kib_failed_devs,
3668 if (cfs_time_before(cfs_time_current(),
3669 dev->ibd_next_failover))
3676 list_del_init(&dev->ibd_fail_list);
3677 dev->ibd_failover = 1;
3678 write_unlock_irqrestore(glock, flags);
3680 rc = kiblnd_dev_failover(dev);
3682 write_lock_irqsave(glock, flags);
3684 LASSERT (dev->ibd_failover);
3685 dev->ibd_failover = 0;
3686 if (rc >= 0) { /* Device is OK or failover succeed */
3687 dev->ibd_next_failover = cfs_time_shift(3);
3691 /* failed to failover, retry later */
3692 dev->ibd_next_failover =
3693 cfs_time_shift(min(dev->ibd_failed_failover, 10));
3694 if (kiblnd_dev_can_failover(dev)) {
3695 list_add_tail(&dev->ibd_fail_list,
3696 &kiblnd_data.kib_failed_devs);
3702 /* long sleep if no more pending failover */
3703 long_sleep = list_empty(&kiblnd_data.kib_failed_devs);
3705 set_current_state(TASK_INTERRUPTIBLE);
3706 add_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3707 write_unlock_irqrestore(glock, flags);
3709 rc = schedule_timeout(long_sleep ? cfs_time_seconds(10) :
3710 cfs_time_seconds(1));
3711 set_current_state(TASK_RUNNING);
3712 remove_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3713 write_lock_irqsave(glock, flags);
3715 if (!long_sleep || rc != 0)
3718 /* have a long sleep, routine check all active devices,
3719 * we need checking like this because if there is not active
3720 * connection on the dev and no SEND from local, we may listen
3721 * on wrong HCA for ever while there is a bonding failover */
3722 list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3723 if (kiblnd_dev_can_failover(dev)) {
3724 list_add_tail(&dev->ibd_fail_list,
3725 &kiblnd_data.kib_failed_devs);
3730 write_unlock_irqrestore(glock, flags);
3732 kiblnd_thread_fini();