4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2012, 2015, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lnet/klnds/o2iblnd/o2iblnd_cb.c
38 * Author: Eric Barton <eric@bartonsoftware.com>
43 static void kiblnd_peer_alive(kib_peer_t *peer);
44 static void kiblnd_peer_connect_failed(kib_peer_t *peer, int active, int error);
45 static void kiblnd_check_sends(kib_conn_t *conn);
46 static void kiblnd_init_tx_msg(lnet_ni_t *ni, kib_tx_t *tx,
47 int type, int body_nob);
48 static int kiblnd_init_rdma(kib_conn_t *conn, kib_tx_t *tx, int type,
49 int resid, kib_rdma_desc_t *dstrd, __u64 dstcookie);
50 static void kiblnd_queue_tx_locked(kib_tx_t *tx, kib_conn_t *conn);
51 static void kiblnd_queue_tx(kib_tx_t *tx, kib_conn_t *conn);
52 static void kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx);
55 kiblnd_tx_done (lnet_ni_t *ni, kib_tx_t *tx)
57 lnet_msg_t *lntmsg[2];
58 kib_net_t *net = ni->ni_data;
62 LASSERT (net != NULL);
63 LASSERT (!in_interrupt());
64 LASSERT (!tx->tx_queued); /* mustn't be queued for sending */
65 LASSERT (tx->tx_sending == 0); /* mustn't be awaiting sent callback */
66 LASSERT (!tx->tx_waiting); /* mustn't be awaiting peer response */
67 LASSERT (tx->tx_pool != NULL);
69 kiblnd_unmap_tx(ni, tx);
71 /* tx may have up to 2 lnet msgs to finalise */
72 lntmsg[0] = tx->tx_lntmsg[0]; tx->tx_lntmsg[0] = NULL;
73 lntmsg[1] = tx->tx_lntmsg[1]; tx->tx_lntmsg[1] = NULL;
76 if (tx->tx_conn != NULL) {
77 LASSERT (ni == tx->tx_conn->ibc_peer->ibp_ni);
79 kiblnd_conn_decref(tx->tx_conn);
86 kiblnd_pool_free_node(&tx->tx_pool->tpo_pool, &tx->tx_list);
88 /* delay finalize until my descs have been freed */
89 for (i = 0; i < 2; i++) {
90 if (lntmsg[i] == NULL)
93 lnet_finalize(ni, lntmsg[i], rc);
98 kiblnd_txlist_done(lnet_ni_t *ni, struct list_head *txlist, int status)
102 while (!list_empty(txlist)) {
103 tx = list_entry(txlist->next, kib_tx_t, tx_list);
105 list_del(&tx->tx_list);
108 tx->tx_status = status;
109 kiblnd_tx_done(ni, tx);
114 kiblnd_get_idle_tx(lnet_ni_t *ni, lnet_nid_t target)
116 kib_net_t *net = (kib_net_t *)ni->ni_data;
117 struct list_head *node;
119 kib_tx_poolset_t *tps;
121 tps = net->ibn_tx_ps[lnet_cpt_of_nid(target)];
122 node = kiblnd_pool_alloc_node(&tps->tps_poolset);
125 tx = container_of(node, kib_tx_t, tx_list);
127 LASSERT (tx->tx_nwrq == 0);
128 LASSERT (!tx->tx_queued);
129 LASSERT (tx->tx_sending == 0);
130 LASSERT (!tx->tx_waiting);
131 LASSERT (tx->tx_status == 0);
132 LASSERT (tx->tx_conn == NULL);
133 LASSERT (tx->tx_lntmsg[0] == NULL);
134 LASSERT (tx->tx_lntmsg[1] == NULL);
135 LASSERT (tx->tx_nfrags == 0);
141 kiblnd_drop_rx(kib_rx_t *rx)
143 kib_conn_t *conn = rx->rx_conn;
144 struct kib_sched_info *sched = conn->ibc_sched;
147 spin_lock_irqsave(&sched->ibs_lock, flags);
148 LASSERT(conn->ibc_nrx > 0);
150 spin_unlock_irqrestore(&sched->ibs_lock, flags);
152 kiblnd_conn_decref(conn);
156 kiblnd_post_rx (kib_rx_t *rx, int credit)
158 kib_conn_t *conn = rx->rx_conn;
159 kib_net_t *net = conn->ibc_peer->ibp_ni->ni_data;
160 struct ib_recv_wr *bad_wrq = NULL;
161 struct ib_mr *mr = conn->ibc_hdev->ibh_mrs;
164 LASSERT (net != NULL);
165 LASSERT (!in_interrupt());
166 LASSERT (credit == IBLND_POSTRX_NO_CREDIT ||
167 credit == IBLND_POSTRX_PEER_CREDIT ||
168 credit == IBLND_POSTRX_RSRVD_CREDIT);
171 rx->rx_sge.lkey = mr->lkey;
172 rx->rx_sge.addr = rx->rx_msgaddr;
173 rx->rx_sge.length = IBLND_MSG_SIZE;
175 rx->rx_wrq.next = NULL;
176 rx->rx_wrq.sg_list = &rx->rx_sge;
177 rx->rx_wrq.num_sge = 1;
178 rx->rx_wrq.wr_id = kiblnd_ptr2wreqid(rx, IBLND_WID_RX);
180 LASSERT (conn->ibc_state >= IBLND_CONN_INIT);
181 LASSERT (rx->rx_nob >= 0); /* not posted */
183 if (conn->ibc_state > IBLND_CONN_ESTABLISHED) {
184 kiblnd_drop_rx(rx); /* No more posts for this rx */
188 rx->rx_nob = -1; /* flag posted */
190 /* NB: need an extra reference after ib_post_recv because we don't
191 * own this rx (and rx::rx_conn) anymore, LU-5678.
193 kiblnd_conn_addref(conn);
194 rc = ib_post_recv(conn->ibc_cmid->qp, &rx->rx_wrq, &bad_wrq);
195 if (unlikely(rc != 0)) {
196 CERROR("Can't post rx for %s: %d, bad_wrq: %p\n",
197 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc, bad_wrq);
201 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) /* Initial post */
204 if (unlikely(rc != 0)) {
205 kiblnd_close_conn(conn, rc);
206 kiblnd_drop_rx(rx); /* No more posts for this rx */
210 if (credit == IBLND_POSTRX_NO_CREDIT)
213 spin_lock(&conn->ibc_lock);
214 if (credit == IBLND_POSTRX_PEER_CREDIT)
215 conn->ibc_outstanding_credits++;
217 conn->ibc_reserved_credits++;
218 spin_unlock(&conn->ibc_lock);
220 kiblnd_check_sends(conn);
222 kiblnd_conn_decref(conn);
227 kiblnd_find_waiting_tx_locked(kib_conn_t *conn, int txtype, __u64 cookie)
229 struct list_head *tmp;
231 list_for_each(tmp, &conn->ibc_active_txs) {
232 kib_tx_t *tx = list_entry(tmp, kib_tx_t, tx_list);
234 LASSERT(!tx->tx_queued);
235 LASSERT(tx->tx_sending != 0 || tx->tx_waiting);
237 if (tx->tx_cookie != cookie)
240 if (tx->tx_waiting &&
241 tx->tx_msg->ibm_type == txtype)
244 CWARN("Bad completion: %swaiting, type %x (wanted %x)\n",
245 tx->tx_waiting ? "" : "NOT ",
246 tx->tx_msg->ibm_type, txtype);
252 kiblnd_handle_completion(kib_conn_t *conn, int txtype, int status, __u64 cookie)
255 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
258 spin_lock(&conn->ibc_lock);
260 tx = kiblnd_find_waiting_tx_locked(conn, txtype, cookie);
262 spin_unlock(&conn->ibc_lock);
264 CWARN("Unmatched completion type %x cookie "LPX64" from %s\n",
265 txtype, cookie, libcfs_nid2str(conn->ibc_peer->ibp_nid));
266 kiblnd_close_conn(conn, -EPROTO);
270 if (tx->tx_status == 0) { /* success so far */
271 if (status < 0) { /* failed? */
272 tx->tx_status = status;
273 } else if (txtype == IBLND_MSG_GET_REQ) {
274 lnet_set_reply_msg_len(ni, tx->tx_lntmsg[1], status);
280 idle = !tx->tx_queued && (tx->tx_sending == 0);
282 list_del(&tx->tx_list);
284 spin_unlock(&conn->ibc_lock);
287 kiblnd_tx_done(ni, tx);
291 kiblnd_send_completion(kib_conn_t *conn, int type, int status, __u64 cookie)
293 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
294 kib_tx_t *tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
297 CERROR("Can't get tx for completion %x for %s\n",
298 type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
302 tx->tx_msg->ibm_u.completion.ibcm_status = status;
303 tx->tx_msg->ibm_u.completion.ibcm_cookie = cookie;
304 kiblnd_init_tx_msg(ni, tx, type, sizeof(kib_completion_msg_t));
306 kiblnd_queue_tx(tx, conn);
310 kiblnd_handle_rx (kib_rx_t *rx)
312 kib_msg_t *msg = rx->rx_msg;
313 kib_conn_t *conn = rx->rx_conn;
314 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
315 int credits = msg->ibm_credits;
321 LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
323 CDEBUG (D_NET, "Received %x[%d] from %s\n",
324 msg->ibm_type, credits,
325 libcfs_nid2str(conn->ibc_peer->ibp_nid));
328 /* Have I received credits that will let me send? */
329 spin_lock(&conn->ibc_lock);
331 if (conn->ibc_credits + credits >
332 conn->ibc_queue_depth) {
333 rc2 = conn->ibc_credits;
334 spin_unlock(&conn->ibc_lock);
336 CERROR("Bad credits from %s: %d + %d > %d\n",
337 libcfs_nid2str(conn->ibc_peer->ibp_nid),
339 conn->ibc_queue_depth);
341 kiblnd_close_conn(conn, -EPROTO);
342 kiblnd_post_rx(rx, IBLND_POSTRX_NO_CREDIT);
346 conn->ibc_credits += credits;
348 /* This ensures the credit taken by NOOP can be returned */
349 if (msg->ibm_type == IBLND_MSG_NOOP &&
350 !IBLND_OOB_CAPABLE(conn->ibc_version)) /* v1 only */
351 conn->ibc_outstanding_credits++;
353 spin_unlock(&conn->ibc_lock);
354 kiblnd_check_sends(conn);
357 switch (msg->ibm_type) {
359 CERROR("Bad IBLND message type %x from %s\n",
360 msg->ibm_type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
361 post_credit = IBLND_POSTRX_NO_CREDIT;
366 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
367 post_credit = IBLND_POSTRX_NO_CREDIT;
371 if (credits != 0) /* credit already posted */
372 post_credit = IBLND_POSTRX_NO_CREDIT;
373 else /* a keepalive NOOP */
374 post_credit = IBLND_POSTRX_PEER_CREDIT;
377 case IBLND_MSG_IMMEDIATE:
378 post_credit = IBLND_POSTRX_DONT_POST;
379 rc = lnet_parse(ni, &msg->ibm_u.immediate.ibim_hdr,
380 msg->ibm_srcnid, rx, 0);
381 if (rc < 0) /* repost on error */
382 post_credit = IBLND_POSTRX_PEER_CREDIT;
385 case IBLND_MSG_PUT_REQ:
386 post_credit = IBLND_POSTRX_DONT_POST;
387 rc = lnet_parse(ni, &msg->ibm_u.putreq.ibprm_hdr,
388 msg->ibm_srcnid, rx, 1);
389 if (rc < 0) /* repost on error */
390 post_credit = IBLND_POSTRX_PEER_CREDIT;
393 case IBLND_MSG_PUT_NAK:
394 CWARN ("PUT_NACK from %s\n",
395 libcfs_nid2str(conn->ibc_peer->ibp_nid));
396 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
397 kiblnd_handle_completion(conn, IBLND_MSG_PUT_REQ,
398 msg->ibm_u.completion.ibcm_status,
399 msg->ibm_u.completion.ibcm_cookie);
402 case IBLND_MSG_PUT_ACK:
403 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
405 spin_lock(&conn->ibc_lock);
406 tx = kiblnd_find_waiting_tx_locked(conn, IBLND_MSG_PUT_REQ,
407 msg->ibm_u.putack.ibpam_src_cookie);
409 list_del(&tx->tx_list);
410 spin_unlock(&conn->ibc_lock);
413 CERROR("Unmatched PUT_ACK from %s\n",
414 libcfs_nid2str(conn->ibc_peer->ibp_nid));
419 LASSERT (tx->tx_waiting);
420 /* CAVEAT EMPTOR: I could be racing with tx_complete, but...
421 * (a) I can overwrite tx_msg since my peer has received it!
422 * (b) tx_waiting set tells tx_complete() it's not done. */
424 tx->tx_nwrq = 0; /* overwrite PUT_REQ */
426 rc2 = kiblnd_init_rdma(conn, tx, IBLND_MSG_PUT_DONE,
427 kiblnd_rd_size(&msg->ibm_u.putack.ibpam_rd),
428 &msg->ibm_u.putack.ibpam_rd,
429 msg->ibm_u.putack.ibpam_dst_cookie);
431 CERROR("Can't setup rdma for PUT to %s: %d\n",
432 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc2);
434 spin_lock(&conn->ibc_lock);
435 tx->tx_waiting = 0; /* clear waiting and queue atomically */
436 kiblnd_queue_tx_locked(tx, conn);
437 spin_unlock(&conn->ibc_lock);
440 case IBLND_MSG_PUT_DONE:
441 post_credit = IBLND_POSTRX_PEER_CREDIT;
442 kiblnd_handle_completion(conn, IBLND_MSG_PUT_ACK,
443 msg->ibm_u.completion.ibcm_status,
444 msg->ibm_u.completion.ibcm_cookie);
447 case IBLND_MSG_GET_REQ:
448 post_credit = IBLND_POSTRX_DONT_POST;
449 rc = lnet_parse(ni, &msg->ibm_u.get.ibgm_hdr,
450 msg->ibm_srcnid, rx, 1);
451 if (rc < 0) /* repost on error */
452 post_credit = IBLND_POSTRX_PEER_CREDIT;
455 case IBLND_MSG_GET_DONE:
456 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
457 kiblnd_handle_completion(conn, IBLND_MSG_GET_REQ,
458 msg->ibm_u.completion.ibcm_status,
459 msg->ibm_u.completion.ibcm_cookie);
463 if (rc < 0) /* protocol error */
464 kiblnd_close_conn(conn, rc);
466 if (post_credit != IBLND_POSTRX_DONT_POST)
467 kiblnd_post_rx(rx, post_credit);
471 kiblnd_rx_complete (kib_rx_t *rx, int status, int nob)
473 kib_msg_t *msg = rx->rx_msg;
474 kib_conn_t *conn = rx->rx_conn;
475 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
476 kib_net_t *net = ni->ni_data;
480 LASSERT (net != NULL);
481 LASSERT (rx->rx_nob < 0); /* was posted */
482 rx->rx_nob = 0; /* isn't now */
484 if (conn->ibc_state > IBLND_CONN_ESTABLISHED)
487 if (status != IB_WC_SUCCESS) {
488 CNETERR("Rx from %s failed: %d\n",
489 libcfs_nid2str(conn->ibc_peer->ibp_nid), status);
496 rc = kiblnd_unpack_msg(msg, rx->rx_nob);
498 CERROR ("Error %d unpacking rx from %s\n",
499 rc, libcfs_nid2str(conn->ibc_peer->ibp_nid));
503 if (msg->ibm_srcnid != conn->ibc_peer->ibp_nid ||
504 msg->ibm_dstnid != ni->ni_nid ||
505 msg->ibm_srcstamp != conn->ibc_incarnation ||
506 msg->ibm_dststamp != net->ibn_incarnation) {
507 CERROR ("Stale rx from %s\n",
508 libcfs_nid2str(conn->ibc_peer->ibp_nid));
513 /* set time last known alive */
514 kiblnd_peer_alive(conn->ibc_peer);
516 /* racing with connection establishment/teardown! */
518 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
519 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
522 write_lock_irqsave(g_lock, flags);
523 /* must check holding global lock to eliminate race */
524 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
525 list_add_tail(&rx->rx_list, &conn->ibc_early_rxs);
526 write_unlock_irqrestore(g_lock, flags);
529 write_unlock_irqrestore(g_lock, flags);
531 kiblnd_handle_rx(rx);
535 CDEBUG(D_NET, "rx %p conn %p\n", rx, conn);
536 kiblnd_close_conn(conn, err);
538 kiblnd_drop_rx(rx); /* Don't re-post rx. */
542 kiblnd_kvaddr_to_page (unsigned long vaddr)
546 if (is_vmalloc_addr((void *)vaddr)) {
547 page = vmalloc_to_page ((void *)vaddr);
548 LASSERT (page != NULL);
551 #ifdef CONFIG_HIGHMEM
552 if (vaddr >= PKMAP_BASE &&
553 vaddr < (PKMAP_BASE + LAST_PKMAP * PAGE_SIZE)) {
554 /* No highmem pages only used for bulk (kiov) I/O */
555 CERROR("find page for address in highmem\n");
559 page = virt_to_page (vaddr);
560 LASSERT (page != NULL);
565 kiblnd_fmr_map_tx(kib_net_t *net, kib_tx_t *tx, kib_rdma_desc_t *rd, __u32 nob)
568 __u64 *pages = tx->tx_pages;
569 kib_fmr_poolset_t *fps;
576 LASSERT(tx->tx_pool != NULL);
577 LASSERT(tx->tx_pool->tpo_pool.po_owner != NULL);
579 hdev = tx->tx_pool->tpo_hdev;
581 for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
582 for (size = 0; size < rd->rd_frags[i].rf_nob;
583 size += hdev->ibh_page_size) {
584 pages[npages ++] = (rd->rd_frags[i].rf_addr &
585 hdev->ibh_page_mask) + size;
589 cpt = tx->tx_pool->tpo_pool.po_owner->ps_cpt;
591 fps = net->ibn_fmr_ps[cpt];
592 rc = kiblnd_fmr_pool_map(fps, pages, npages, nob, 0, (rd != tx->tx_rd),
595 CERROR("Can't map %d pages: %d\n", npages, rc);
599 /* If rd is not tx_rd, it's going to get sent to a peer, who will need
601 rd->rd_key = tx->fmr.fmr_key;
602 rd->rd_frags[0].rf_addr &= ~hdev->ibh_page_mask;
603 rd->rd_frags[0].rf_nob = nob;
610 kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx)
612 kib_net_t *net = ni->ni_data;
614 LASSERT(net != NULL);
616 if (net->ibn_fmr_ps != NULL)
617 kiblnd_fmr_pool_unmap(&tx->fmr, tx->tx_status);
619 if (tx->tx_nfrags != 0) {
620 kiblnd_dma_unmap_sg(tx->tx_pool->tpo_hdev->ibh_ibdev,
621 tx->tx_frags, tx->tx_nfrags, tx->tx_dmadir);
627 kiblnd_map_tx(lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd, int nfrags)
629 kib_net_t *net = ni->ni_data;
630 kib_hca_dev_t *hdev = net->ibn_dev->ibd_hdev;
631 struct ib_mr *mr = NULL;
635 /* If rd is not tx_rd, it's going to get sent to a peer and I'm the
637 tx->tx_dmadir = (rd != tx->tx_rd) ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
638 tx->tx_nfrags = nfrags;
640 rd->rd_nfrags = kiblnd_dma_map_sg(hdev->ibh_ibdev, tx->tx_frags,
641 tx->tx_nfrags, tx->tx_dmadir);
643 for (i = 0, nob = 0; i < rd->rd_nfrags; i++) {
644 rd->rd_frags[i].rf_nob = kiblnd_sg_dma_len(
645 hdev->ibh_ibdev, &tx->tx_frags[i]);
646 rd->rd_frags[i].rf_addr = kiblnd_sg_dma_address(
647 hdev->ibh_ibdev, &tx->tx_frags[i]);
648 nob += rd->rd_frags[i].rf_nob;
651 mr = kiblnd_find_rd_dma_mr(ni, rd,
652 (tx->tx_conn != NULL) ?
653 tx->tx_conn->ibc_max_frags : -1);
655 /* found pre-mapping MR */
656 rd->rd_key = (rd != tx->tx_rd) ? mr->rkey : mr->lkey;
660 if (net->ibn_fmr_ps != NULL)
661 return kiblnd_fmr_map_tx(net, tx, rd, nob);
668 kiblnd_setup_rd_iov(lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd,
669 unsigned int niov, struct kvec *iov, int offset, int nob)
671 kib_net_t *net = ni->ni_data;
673 struct scatterlist *sg;
680 LASSERT (net != NULL);
682 while (offset >= iov->iov_len) {
683 offset -= iov->iov_len;
693 vaddr = ((unsigned long)iov->iov_base) + offset;
694 page_offset = vaddr & (PAGE_SIZE - 1);
695 page = kiblnd_kvaddr_to_page(vaddr);
697 CERROR ("Can't find page\n");
701 fragnob = min((int)(iov->iov_len - offset), nob);
702 fragnob = min(fragnob, (int)PAGE_SIZE - page_offset);
704 sg_set_page(sg, page, fragnob, page_offset);
707 if (offset + fragnob < iov->iov_len) {
717 return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags);
721 kiblnd_setup_rd_kiov (lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd,
722 int nkiov, lnet_kiov_t *kiov, int offset, int nob)
724 kib_net_t *net = ni->ni_data;
725 struct scatterlist *sg;
728 CDEBUG(D_NET, "niov %d offset %d nob %d\n", nkiov, offset, nob);
732 LASSERT (net != NULL);
734 while (offset >= kiov->kiov_len) {
735 offset -= kiov->kiov_len;
745 fragnob = min((int)(kiov->kiov_len - offset), nob);
747 sg_set_page(sg, kiov->kiov_page, fragnob,
748 kiov->kiov_offset + offset);
757 return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags);
761 kiblnd_post_tx_locked (kib_conn_t *conn, kib_tx_t *tx, int credit)
762 __must_hold(&conn->ibc_lock)
764 kib_msg_t *msg = tx->tx_msg;
765 kib_peer_t *peer = conn->ibc_peer;
766 struct lnet_ni *ni = peer->ibp_ni;
767 int ver = conn->ibc_version;
771 LASSERT(tx->tx_queued);
772 /* We rely on this for QP sizing */
773 LASSERT(tx->tx_nwrq > 0);
774 LASSERT(tx->tx_nwrq <= 1 + conn->ibc_max_frags);
776 LASSERT(credit == 0 || credit == 1);
777 LASSERT(conn->ibc_outstanding_credits >= 0);
778 LASSERT(conn->ibc_outstanding_credits <= conn->ibc_queue_depth);
779 LASSERT(conn->ibc_credits >= 0);
780 LASSERT(conn->ibc_credits <= conn->ibc_queue_depth);
782 if (conn->ibc_nsends_posted ==
783 kiblnd_concurrent_sends(ver, ni)) {
784 /* tx completions outstanding... */
785 CDEBUG(D_NET, "%s: posted enough\n",
786 libcfs_nid2str(peer->ibp_nid));
790 if (credit != 0 && conn->ibc_credits == 0) { /* no credits */
791 CDEBUG(D_NET, "%s: no credits\n",
792 libcfs_nid2str(peer->ibp_nid));
796 if (credit != 0 && !IBLND_OOB_CAPABLE(ver) &&
797 conn->ibc_credits == 1 && /* last credit reserved */
798 msg->ibm_type != IBLND_MSG_NOOP) { /* for NOOP */
799 CDEBUG(D_NET, "%s: not using last credit\n",
800 libcfs_nid2str(peer->ibp_nid));
804 /* NB don't drop ibc_lock before bumping tx_sending */
805 list_del(&tx->tx_list);
808 if (msg->ibm_type == IBLND_MSG_NOOP &&
809 (!kiblnd_need_noop(conn) || /* redundant NOOP */
810 (IBLND_OOB_CAPABLE(ver) && /* posted enough NOOP */
811 conn->ibc_noops_posted == IBLND_OOB_MSGS(ver)))) {
812 /* OK to drop when posted enough NOOPs, since
813 * kiblnd_check_sends will queue NOOP again when
814 * posted NOOPs complete */
815 spin_unlock(&conn->ibc_lock);
816 kiblnd_tx_done(peer->ibp_ni, tx);
817 spin_lock(&conn->ibc_lock);
818 CDEBUG(D_NET, "%s(%d): redundant or enough NOOP\n",
819 libcfs_nid2str(peer->ibp_nid),
820 conn->ibc_noops_posted);
824 kiblnd_pack_msg(peer->ibp_ni, msg, ver, conn->ibc_outstanding_credits,
825 peer->ibp_nid, conn->ibc_incarnation);
827 conn->ibc_credits -= credit;
828 conn->ibc_outstanding_credits = 0;
829 conn->ibc_nsends_posted++;
830 if (msg->ibm_type == IBLND_MSG_NOOP)
831 conn->ibc_noops_posted++;
833 /* CAVEAT EMPTOR! This tx could be the PUT_DONE of an RDMA
834 * PUT. If so, it was first queued here as a PUT_REQ, sent and
835 * stashed on ibc_active_txs, matched by an incoming PUT_ACK,
836 * and then re-queued here. It's (just) possible that
837 * tx_sending is non-zero if we've not done the tx_complete()
838 * from the first send; hence the ++ rather than = below. */
840 list_add(&tx->tx_list, &conn->ibc_active_txs);
842 /* I'm still holding ibc_lock! */
843 if (conn->ibc_state != IBLND_CONN_ESTABLISHED) {
845 } else if (tx->tx_pool->tpo_pool.po_failed ||
846 conn->ibc_hdev != tx->tx_pool->tpo_hdev) {
847 /* close_conn will launch failover */
850 struct kib_fast_reg_descriptor *frd = tx->fmr.fmr_frd;
851 struct ib_send_wr *bad = &tx->tx_wrq[tx->tx_nwrq - 1];
852 struct ib_send_wr *wrq = tx->tx_wrq;
855 if (!frd->frd_valid) {
856 wrq = &frd->frd_inv_wr;
857 wrq->next = &frd->frd_fastreg_wr;
859 wrq = &frd->frd_fastreg_wr;
861 frd->frd_fastreg_wr.next = tx->tx_wrq;
864 LASSERTF(bad->wr_id == kiblnd_ptr2wreqid(tx, IBLND_WID_TX),
865 "bad wr_id "LPX64", opc %d, flags %d, peer: %s\n",
866 bad->wr_id, bad->opcode, bad->send_flags,
867 libcfs_nid2str(conn->ibc_peer->ibp_nid));
869 rc = ib_post_send(conn->ibc_cmid->qp, wrq, &bad);
872 conn->ibc_last_send = jiffies;
877 /* NB credits are transferred in the actual
878 * message, which can only be the last work item */
879 conn->ibc_credits += credit;
880 conn->ibc_outstanding_credits += msg->ibm_credits;
881 conn->ibc_nsends_posted--;
882 if (msg->ibm_type == IBLND_MSG_NOOP)
883 conn->ibc_noops_posted--;
889 done = (tx->tx_sending == 0);
891 list_del(&tx->tx_list);
893 spin_unlock(&conn->ibc_lock);
895 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
896 CERROR("Error %d posting transmit to %s\n",
897 rc, libcfs_nid2str(peer->ibp_nid));
899 CDEBUG(D_NET, "Error %d posting transmit to %s\n",
900 rc, libcfs_nid2str(peer->ibp_nid));
902 kiblnd_close_conn(conn, rc);
905 kiblnd_tx_done(peer->ibp_ni, tx);
907 spin_lock(&conn->ibc_lock);
913 kiblnd_check_sends (kib_conn_t *conn)
915 int ver = conn->ibc_version;
916 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
919 /* Don't send anything until after the connection is established */
920 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
921 CDEBUG(D_NET, "%s too soon\n",
922 libcfs_nid2str(conn->ibc_peer->ibp_nid));
926 spin_lock(&conn->ibc_lock);
928 LASSERT(conn->ibc_nsends_posted <=
929 kiblnd_concurrent_sends(ver, ni));
930 LASSERT (!IBLND_OOB_CAPABLE(ver) ||
931 conn->ibc_noops_posted <= IBLND_OOB_MSGS(ver));
932 LASSERT (conn->ibc_reserved_credits >= 0);
934 while (conn->ibc_reserved_credits > 0 &&
935 !list_empty(&conn->ibc_tx_queue_rsrvd)) {
936 tx = list_entry(conn->ibc_tx_queue_rsrvd.next,
938 list_del(&tx->tx_list);
939 list_add_tail(&tx->tx_list, &conn->ibc_tx_queue);
940 conn->ibc_reserved_credits--;
943 if (kiblnd_need_noop(conn)) {
944 spin_unlock(&conn->ibc_lock);
946 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
948 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_NOOP, 0);
950 spin_lock(&conn->ibc_lock);
952 kiblnd_queue_tx_locked(tx, conn);
958 if (!list_empty(&conn->ibc_tx_queue_nocred)) {
960 tx = list_entry(conn->ibc_tx_queue_nocred.next,
962 } else if (!list_empty(&conn->ibc_tx_noops)) {
963 LASSERT (!IBLND_OOB_CAPABLE(ver));
965 tx = list_entry(conn->ibc_tx_noops.next,
967 } else if (!list_empty(&conn->ibc_tx_queue)) {
969 tx = list_entry(conn->ibc_tx_queue.next,
974 if (kiblnd_post_tx_locked(conn, tx, credit) != 0)
978 spin_unlock(&conn->ibc_lock);
982 kiblnd_tx_complete (kib_tx_t *tx, int status)
984 int failed = (status != IB_WC_SUCCESS);
985 kib_conn_t *conn = tx->tx_conn;
988 LASSERT (tx->tx_sending > 0);
991 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
992 CNETERR("Tx -> %s cookie "LPX64
993 " sending %d waiting %d: failed %d\n",
994 libcfs_nid2str(conn->ibc_peer->ibp_nid),
995 tx->tx_cookie, tx->tx_sending, tx->tx_waiting,
998 kiblnd_close_conn(conn, -EIO);
1000 kiblnd_peer_alive(conn->ibc_peer);
1003 spin_lock(&conn->ibc_lock);
1005 /* I could be racing with rdma completion. Whoever makes 'tx' idle
1006 * gets to free it, which also drops its ref on 'conn'. */
1009 conn->ibc_nsends_posted--;
1010 if (tx->tx_msg->ibm_type == IBLND_MSG_NOOP)
1011 conn->ibc_noops_posted--;
1014 tx->tx_waiting = 0; /* don't wait for peer */
1015 tx->tx_status = -EIO;
1018 idle = (tx->tx_sending == 0) && /* This is the final callback */
1019 !tx->tx_waiting && /* Not waiting for peer */
1020 !tx->tx_queued; /* Not re-queued (PUT_DONE) */
1022 list_del(&tx->tx_list);
1024 kiblnd_conn_addref(conn); /* 1 ref for me.... */
1026 spin_unlock(&conn->ibc_lock);
1029 kiblnd_tx_done(conn->ibc_peer->ibp_ni, tx);
1031 kiblnd_check_sends(conn);
1033 kiblnd_conn_decref(conn); /* ...until here */
1037 kiblnd_init_tx_msg (lnet_ni_t *ni, kib_tx_t *tx, int type, int body_nob)
1039 kib_hca_dev_t *hdev = tx->tx_pool->tpo_hdev;
1040 struct ib_sge *sge = &tx->tx_sge[tx->tx_nwrq];
1041 struct ib_send_wr *wrq = &tx->tx_wrq[tx->tx_nwrq];
1042 int nob = offsetof (kib_msg_t, ibm_u) + body_nob;
1043 struct ib_mr *mr = hdev->ibh_mrs;
1045 LASSERT(tx->tx_nwrq >= 0);
1046 LASSERT(tx->tx_nwrq < IBLND_MAX_RDMA_FRAGS + 1);
1047 LASSERT(nob <= IBLND_MSG_SIZE);
1048 LASSERT(mr != NULL);
1050 kiblnd_init_msg(tx->tx_msg, type, body_nob);
1052 sge->lkey = mr->lkey;
1053 sge->addr = tx->tx_msgaddr;
1056 memset(wrq, 0, sizeof(*wrq));
1059 wrq->wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_TX);
1062 wrq->opcode = IB_WR_SEND;
1063 wrq->send_flags = IB_SEND_SIGNALED;
1069 kiblnd_init_rdma(kib_conn_t *conn, kib_tx_t *tx, int type,
1070 int resid, kib_rdma_desc_t *dstrd, __u64 dstcookie)
1072 kib_msg_t *ibmsg = tx->tx_msg;
1073 kib_rdma_desc_t *srcrd = tx->tx_rd;
1074 struct ib_sge *sge = &tx->tx_sge[0];
1075 struct ib_send_wr *wrq = &tx->tx_wrq[0];
1081 LASSERT (!in_interrupt());
1082 LASSERT (tx->tx_nwrq == 0);
1083 LASSERT (type == IBLND_MSG_GET_DONE ||
1084 type == IBLND_MSG_PUT_DONE);
1086 srcidx = dstidx = 0;
1089 if (srcidx >= srcrd->rd_nfrags) {
1090 CERROR("Src buffer exhausted: %d frags\n", srcidx);
1095 if (dstidx == dstrd->rd_nfrags) {
1096 CERROR("Dst buffer exhausted: %d frags\n", dstidx);
1101 if (tx->tx_nwrq >= conn->ibc_max_frags) {
1102 CERROR("RDMA has too many fragments for peer %s (%d), "
1103 "src idx/frags: %d/%d dst idx/frags: %d/%d\n",
1104 libcfs_nid2str(conn->ibc_peer->ibp_nid),
1105 conn->ibc_max_frags,
1106 srcidx, srcrd->rd_nfrags,
1107 dstidx, dstrd->rd_nfrags);
1112 wrknob = MIN(MIN(kiblnd_rd_frag_size(srcrd, srcidx),
1113 kiblnd_rd_frag_size(dstrd, dstidx)), resid);
1115 sge = &tx->tx_sge[tx->tx_nwrq];
1116 sge->addr = kiblnd_rd_frag_addr(srcrd, srcidx);
1117 sge->lkey = kiblnd_rd_frag_key(srcrd, srcidx);
1118 sge->length = wrknob;
1120 wrq = &tx->tx_wrq[tx->tx_nwrq];
1122 wrq->next = wrq + 1;
1123 wrq->wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_RDMA);
1126 wrq->opcode = IB_WR_RDMA_WRITE;
1127 wrq->send_flags = 0;
1129 wrq->wr.rdma.remote_addr = kiblnd_rd_frag_addr(dstrd, dstidx);
1130 wrq->wr.rdma.rkey = kiblnd_rd_frag_key(dstrd, dstidx);
1132 srcidx = kiblnd_rd_consume_frag(srcrd, srcidx, wrknob);
1133 dstidx = kiblnd_rd_consume_frag(dstrd, dstidx, wrknob);
1142 if (rc < 0) /* no RDMA if completing with failure */
1145 ibmsg->ibm_u.completion.ibcm_status = rc;
1146 ibmsg->ibm_u.completion.ibcm_cookie = dstcookie;
1147 kiblnd_init_tx_msg(conn->ibc_peer->ibp_ni, tx,
1148 type, sizeof (kib_completion_msg_t));
1154 kiblnd_queue_tx_locked(kib_tx_t *tx, kib_conn_t *conn)
1156 struct list_head *q;
1158 LASSERT(tx->tx_nwrq > 0); /* work items set up */
1159 LASSERT(!tx->tx_queued); /* not queued for sending already */
1160 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1163 tx->tx_deadline = jiffies +
1164 msecs_to_jiffies(*kiblnd_tunables.kib_timeout *
1167 if (tx->tx_conn == NULL) {
1168 kiblnd_conn_addref(conn);
1170 LASSERT (tx->tx_msg->ibm_type != IBLND_MSG_PUT_DONE);
1172 /* PUT_DONE first attached to conn as a PUT_REQ */
1173 LASSERT (tx->tx_conn == conn);
1174 LASSERT (tx->tx_msg->ibm_type == IBLND_MSG_PUT_DONE);
1177 switch (tx->tx_msg->ibm_type) {
1181 case IBLND_MSG_PUT_REQ:
1182 case IBLND_MSG_GET_REQ:
1183 q = &conn->ibc_tx_queue_rsrvd;
1186 case IBLND_MSG_PUT_NAK:
1187 case IBLND_MSG_PUT_ACK:
1188 case IBLND_MSG_PUT_DONE:
1189 case IBLND_MSG_GET_DONE:
1190 q = &conn->ibc_tx_queue_nocred;
1193 case IBLND_MSG_NOOP:
1194 if (IBLND_OOB_CAPABLE(conn->ibc_version))
1195 q = &conn->ibc_tx_queue_nocred;
1197 q = &conn->ibc_tx_noops;
1200 case IBLND_MSG_IMMEDIATE:
1201 q = &conn->ibc_tx_queue;
1205 list_add_tail(&tx->tx_list, q);
1209 kiblnd_queue_tx (kib_tx_t *tx, kib_conn_t *conn)
1211 spin_lock(&conn->ibc_lock);
1212 kiblnd_queue_tx_locked(tx, conn);
1213 spin_unlock(&conn->ibc_lock);
1215 kiblnd_check_sends(conn);
1218 static int kiblnd_resolve_addr(struct rdma_cm_id *cmid,
1219 struct sockaddr_in *srcaddr,
1220 struct sockaddr_in *dstaddr,
1223 unsigned short port;
1226 /* allow the port to be reused */
1227 rc = rdma_set_reuseaddr(cmid, 1);
1229 CERROR("Unable to set reuse on cmid: %d\n", rc);
1233 /* look for a free privileged port */
1234 for (port = PROT_SOCK-1; port > 0; port--) {
1235 srcaddr->sin_port = htons(port);
1236 rc = rdma_resolve_addr(cmid,
1237 (struct sockaddr *)srcaddr,
1238 (struct sockaddr *)dstaddr,
1241 CDEBUG(D_NET, "bound to port %hu\n", port);
1243 } else if (rc == -EADDRINUSE || rc == -EADDRNOTAVAIL) {
1244 CDEBUG(D_NET, "bind to port %hu failed: %d\n",
1251 CERROR("Failed to bind to a free privileged port\n");
1256 kiblnd_connect_peer (kib_peer_t *peer)
1258 struct rdma_cm_id *cmid;
1260 kib_net_t *net = peer->ibp_ni->ni_data;
1261 struct sockaddr_in srcaddr;
1262 struct sockaddr_in dstaddr;
1265 LASSERT (net != NULL);
1266 LASSERT (peer->ibp_connecting > 0);
1267 LASSERT(!peer->ibp_reconnecting);
1269 cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, peer, RDMA_PS_TCP,
1273 CERROR("Can't create CMID for %s: %ld\n",
1274 libcfs_nid2str(peer->ibp_nid), PTR_ERR(cmid));
1280 memset(&srcaddr, 0, sizeof(srcaddr));
1281 srcaddr.sin_family = AF_INET;
1282 srcaddr.sin_addr.s_addr = htonl(dev->ibd_ifip);
1284 memset(&dstaddr, 0, sizeof(dstaddr));
1285 dstaddr.sin_family = AF_INET;
1286 dstaddr.sin_port = htons(*kiblnd_tunables.kib_service);
1287 dstaddr.sin_addr.s_addr = htonl(LNET_NIDADDR(peer->ibp_nid));
1289 kiblnd_peer_addref(peer); /* cmid's ref */
1291 if (*kiblnd_tunables.kib_use_priv_port) {
1292 rc = kiblnd_resolve_addr(cmid, &srcaddr, &dstaddr,
1293 *kiblnd_tunables.kib_timeout * 1000);
1295 rc = rdma_resolve_addr(cmid,
1296 (struct sockaddr *)&srcaddr,
1297 (struct sockaddr *)&dstaddr,
1298 *kiblnd_tunables.kib_timeout * 1000);
1301 /* Can't initiate address resolution: */
1302 CERROR("Can't resolve addr for %s: %d\n",
1303 libcfs_nid2str(peer->ibp_nid), rc);
1307 LASSERT (cmid->device != NULL);
1308 CDEBUG(D_NET, "%s: connection bound to %s:%pI4h:%s\n",
1309 libcfs_nid2str(peer->ibp_nid), dev->ibd_ifname,
1310 &dev->ibd_ifip, cmid->device->name);
1315 kiblnd_peer_connect_failed(peer, 1, rc);
1316 kiblnd_peer_decref(peer); /* cmid's ref */
1317 rdma_destroy_id(cmid);
1320 kiblnd_peer_connect_failed(peer, 1, rc);
1324 kiblnd_reconnect_peer(kib_peer_t *peer)
1326 rwlock_t *glock = &kiblnd_data.kib_global_lock;
1327 char *reason = NULL;
1328 struct list_head txs;
1329 unsigned long flags;
1331 INIT_LIST_HEAD(&txs);
1333 write_lock_irqsave(glock, flags);
1334 if (peer->ibp_reconnecting == 0) {
1335 if (peer->ibp_accepting)
1336 reason = "accepting";
1337 else if (peer->ibp_connecting)
1338 reason = "connecting";
1339 else if (!list_empty(&peer->ibp_conns))
1340 reason = "connected";
1341 else /* connected then closed */
1347 LASSERT(!peer->ibp_accepting && !peer->ibp_connecting &&
1348 list_empty(&peer->ibp_conns));
1349 peer->ibp_reconnecting = 0;
1351 if (!kiblnd_peer_active(peer)) {
1352 list_splice_init(&peer->ibp_tx_queue, &txs);
1353 reason = "unlinked";
1357 peer->ibp_connecting++;
1358 peer->ibp_reconnected++;
1359 write_unlock_irqrestore(glock, flags);
1361 kiblnd_connect_peer(peer);
1365 write_unlock_irqrestore(glock, flags);
1367 CWARN("Abort reconnection of %s: %s\n",
1368 libcfs_nid2str(peer->ibp_nid), reason);
1369 kiblnd_txlist_done(peer->ibp_ni, &txs, -ECONNABORTED);
1374 kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid)
1379 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
1380 unsigned long flags;
1383 /* If I get here, I've committed to send, so I complete the tx with
1384 * failure on any problems */
1386 LASSERT (tx == NULL || tx->tx_conn == NULL); /* only set when assigned a conn */
1387 LASSERT (tx == NULL || tx->tx_nwrq > 0); /* work items have been set up */
1389 /* First time, just use a read lock since I expect to find my peer
1391 read_lock_irqsave(g_lock, flags);
1393 peer = kiblnd_find_peer_locked(nid);
1394 if (peer != NULL && !list_empty(&peer->ibp_conns)) {
1395 /* Found a peer with an established connection */
1396 conn = kiblnd_get_conn_locked(peer);
1397 kiblnd_conn_addref(conn); /* 1 ref for me... */
1399 read_unlock_irqrestore(g_lock, flags);
1402 kiblnd_queue_tx(tx, conn);
1403 kiblnd_conn_decref(conn); /* ...to here */
1407 read_unlock(g_lock);
1408 /* Re-try with a write lock */
1411 peer = kiblnd_find_peer_locked(nid);
1413 if (list_empty(&peer->ibp_conns)) {
1414 /* found a peer, but it's still connecting... */
1415 LASSERT(kiblnd_peer_connecting(peer));
1417 list_add_tail(&tx->tx_list,
1418 &peer->ibp_tx_queue);
1419 write_unlock_irqrestore(g_lock, flags);
1421 conn = kiblnd_get_conn_locked(peer);
1422 kiblnd_conn_addref(conn); /* 1 ref for me... */
1424 write_unlock_irqrestore(g_lock, flags);
1427 kiblnd_queue_tx(tx, conn);
1428 kiblnd_conn_decref(conn); /* ...to here */
1433 write_unlock_irqrestore(g_lock, flags);
1435 /* Allocate a peer ready to add to the peer table and retry */
1436 rc = kiblnd_create_peer(ni, &peer, nid);
1438 CERROR("Can't create peer %s\n", libcfs_nid2str(nid));
1440 tx->tx_status = -EHOSTUNREACH;
1442 kiblnd_tx_done(ni, tx);
1447 write_lock_irqsave(g_lock, flags);
1449 peer2 = kiblnd_find_peer_locked(nid);
1450 if (peer2 != NULL) {
1451 if (list_empty(&peer2->ibp_conns)) {
1452 /* found a peer, but it's still connecting... */
1453 LASSERT(kiblnd_peer_connecting(peer2));
1455 list_add_tail(&tx->tx_list,
1456 &peer2->ibp_tx_queue);
1457 write_unlock_irqrestore(g_lock, flags);
1459 conn = kiblnd_get_conn_locked(peer2);
1460 kiblnd_conn_addref(conn); /* 1 ref for me... */
1462 write_unlock_irqrestore(g_lock, flags);
1465 kiblnd_queue_tx(tx, conn);
1466 kiblnd_conn_decref(conn); /* ...to here */
1469 kiblnd_peer_decref(peer);
1473 /* Brand new peer */
1474 LASSERT (peer->ibp_connecting == 0);
1475 peer->ibp_connecting = 1;
1477 /* always called with a ref on ni, which prevents ni being shutdown */
1478 LASSERT (((kib_net_t *)ni->ni_data)->ibn_shutdown == 0);
1481 list_add_tail(&tx->tx_list, &peer->ibp_tx_queue);
1483 kiblnd_peer_addref(peer);
1484 list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid));
1486 write_unlock_irqrestore(g_lock, flags);
1488 kiblnd_connect_peer(peer);
1489 kiblnd_peer_decref(peer);
1493 kiblnd_send (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg)
1495 lnet_hdr_t *hdr = &lntmsg->msg_hdr;
1496 int type = lntmsg->msg_type;
1497 lnet_process_id_t target = lntmsg->msg_target;
1498 int target_is_router = lntmsg->msg_target_is_router;
1499 int routing = lntmsg->msg_routing;
1500 unsigned int payload_niov = lntmsg->msg_niov;
1501 struct kvec *payload_iov = lntmsg->msg_iov;
1502 lnet_kiov_t *payload_kiov = lntmsg->msg_kiov;
1503 unsigned int payload_offset = lntmsg->msg_offset;
1504 unsigned int payload_nob = lntmsg->msg_len;
1506 kib_rdma_desc_t *rd;
1511 /* NB 'private' is different depending on what we're sending.... */
1513 CDEBUG(D_NET, "sending %d bytes in %d frags to %s\n",
1514 payload_nob, payload_niov, libcfs_id2str(target));
1516 LASSERT (payload_nob == 0 || payload_niov > 0);
1517 LASSERT (payload_niov <= LNET_MAX_IOV);
1519 /* Thread context */
1520 LASSERT (!in_interrupt());
1521 /* payload is either all vaddrs or all pages */
1522 LASSERT (!(payload_kiov != NULL && payload_iov != NULL));
1530 LASSERT (payload_nob == 0);
1534 if (routing || target_is_router)
1535 break; /* send IMMEDIATE */
1537 /* is the REPLY message too small for RDMA? */
1538 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[lntmsg->msg_md->md_length]);
1539 if (nob <= IBLND_MSG_SIZE)
1540 break; /* send IMMEDIATE */
1542 tx = kiblnd_get_idle_tx(ni, target.nid);
1544 CERROR("Can't allocate txd for GET to %s\n",
1545 libcfs_nid2str(target.nid));
1550 rd = &ibmsg->ibm_u.get.ibgm_rd;
1551 if ((lntmsg->msg_md->md_options & LNET_MD_KIOV) == 0)
1552 rc = kiblnd_setup_rd_iov(ni, tx, rd,
1553 lntmsg->msg_md->md_niov,
1554 lntmsg->msg_md->md_iov.iov,
1555 0, lntmsg->msg_md->md_length);
1557 rc = kiblnd_setup_rd_kiov(ni, tx, rd,
1558 lntmsg->msg_md->md_niov,
1559 lntmsg->msg_md->md_iov.kiov,
1560 0, lntmsg->msg_md->md_length);
1562 CERROR("Can't setup GET sink for %s: %d\n",
1563 libcfs_nid2str(target.nid), rc);
1564 kiblnd_tx_done(ni, tx);
1568 nob = offsetof(kib_get_msg_t, ibgm_rd.rd_frags[rd->rd_nfrags]);
1569 ibmsg->ibm_u.get.ibgm_cookie = tx->tx_cookie;
1570 ibmsg->ibm_u.get.ibgm_hdr = *hdr;
1572 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_GET_REQ, nob);
1574 tx->tx_lntmsg[1] = lnet_create_reply_msg(ni, lntmsg);
1575 if (tx->tx_lntmsg[1] == NULL) {
1576 CERROR("Can't create reply for GET -> %s\n",
1577 libcfs_nid2str(target.nid));
1578 kiblnd_tx_done(ni, tx);
1582 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg[0,1] on completion */
1583 tx->tx_waiting = 1; /* waiting for GET_DONE */
1584 kiblnd_launch_tx(ni, tx, target.nid);
1587 case LNET_MSG_REPLY:
1589 /* Is the payload small enough not to need RDMA? */
1590 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob]);
1591 if (nob <= IBLND_MSG_SIZE)
1592 break; /* send IMMEDIATE */
1594 tx = kiblnd_get_idle_tx(ni, target.nid);
1596 CERROR("Can't allocate %s txd for %s\n",
1597 type == LNET_MSG_PUT ? "PUT" : "REPLY",
1598 libcfs_nid2str(target.nid));
1602 if (payload_kiov == NULL)
1603 rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1604 payload_niov, payload_iov,
1605 payload_offset, payload_nob);
1607 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1608 payload_niov, payload_kiov,
1609 payload_offset, payload_nob);
1611 CERROR("Can't setup PUT src for %s: %d\n",
1612 libcfs_nid2str(target.nid), rc);
1613 kiblnd_tx_done(ni, tx);
1618 ibmsg->ibm_u.putreq.ibprm_hdr = *hdr;
1619 ibmsg->ibm_u.putreq.ibprm_cookie = tx->tx_cookie;
1620 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_REQ, sizeof(kib_putreq_msg_t));
1622 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1623 tx->tx_waiting = 1; /* waiting for PUT_{ACK,NAK} */
1624 kiblnd_launch_tx(ni, tx, target.nid);
1628 /* send IMMEDIATE */
1630 LASSERT (offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob])
1633 tx = kiblnd_get_idle_tx(ni, target.nid);
1635 CERROR ("Can't send %d to %s: tx descs exhausted\n",
1636 type, libcfs_nid2str(target.nid));
1641 ibmsg->ibm_u.immediate.ibim_hdr = *hdr;
1643 if (payload_kiov != NULL)
1644 lnet_copy_kiov2flat(IBLND_MSG_SIZE, ibmsg,
1645 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1646 payload_niov, payload_kiov,
1647 payload_offset, payload_nob);
1649 lnet_copy_iov2flat(IBLND_MSG_SIZE, ibmsg,
1650 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1651 payload_niov, payload_iov,
1652 payload_offset, payload_nob);
1654 nob = offsetof(kib_immediate_msg_t, ibim_payload[payload_nob]);
1655 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_IMMEDIATE, nob);
1657 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1658 kiblnd_launch_tx(ni, tx, target.nid);
1663 kiblnd_reply (lnet_ni_t *ni, kib_rx_t *rx, lnet_msg_t *lntmsg)
1665 lnet_process_id_t target = lntmsg->msg_target;
1666 unsigned int niov = lntmsg->msg_niov;
1667 struct kvec *iov = lntmsg->msg_iov;
1668 lnet_kiov_t *kiov = lntmsg->msg_kiov;
1669 unsigned int offset = lntmsg->msg_offset;
1670 unsigned int nob = lntmsg->msg_len;
1674 tx = kiblnd_get_idle_tx(ni, rx->rx_conn->ibc_peer->ibp_nid);
1676 CERROR("Can't get tx for REPLY to %s\n",
1677 libcfs_nid2str(target.nid));
1683 else if (kiov == NULL)
1684 rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1685 niov, iov, offset, nob);
1687 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1688 niov, kiov, offset, nob);
1691 CERROR("Can't setup GET src for %s: %d\n",
1692 libcfs_nid2str(target.nid), rc);
1696 rc = kiblnd_init_rdma(rx->rx_conn, tx,
1697 IBLND_MSG_GET_DONE, nob,
1698 &rx->rx_msg->ibm_u.get.ibgm_rd,
1699 rx->rx_msg->ibm_u.get.ibgm_cookie);
1701 CERROR("Can't setup rdma for GET from %s: %d\n",
1702 libcfs_nid2str(target.nid), rc);
1707 /* No RDMA: local completion may happen now! */
1708 lnet_finalize(ni, lntmsg, 0);
1710 /* RDMA: lnet_finalize(lntmsg) when it
1712 tx->tx_lntmsg[0] = lntmsg;
1715 kiblnd_queue_tx(tx, rx->rx_conn);
1719 kiblnd_tx_done(ni, tx);
1721 lnet_finalize(ni, lntmsg, -EIO);
1725 kiblnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed,
1726 unsigned int niov, struct kvec *iov, lnet_kiov_t *kiov,
1727 unsigned int offset, unsigned int mlen, unsigned int rlen)
1729 kib_rx_t *rx = private;
1730 kib_msg_t *rxmsg = rx->rx_msg;
1731 kib_conn_t *conn = rx->rx_conn;
1734 int post_credit = IBLND_POSTRX_PEER_CREDIT;
1737 LASSERT (mlen <= rlen);
1738 LASSERT (!in_interrupt());
1739 /* Either all pages or all vaddrs */
1740 LASSERT (!(kiov != NULL && iov != NULL));
1742 switch (rxmsg->ibm_type) {
1746 case IBLND_MSG_IMMEDIATE:
1747 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[rlen]);
1748 if (nob > rx->rx_nob) {
1749 CERROR ("Immediate message from %s too big: %d(%d)\n",
1750 libcfs_nid2str(rxmsg->ibm_u.immediate.ibim_hdr.src_nid),
1757 lnet_copy_flat2kiov(niov, kiov, offset,
1758 IBLND_MSG_SIZE, rxmsg,
1759 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1762 lnet_copy_flat2iov(niov, iov, offset,
1763 IBLND_MSG_SIZE, rxmsg,
1764 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1766 lnet_finalize (ni, lntmsg, 0);
1769 case IBLND_MSG_PUT_REQ: {
1771 kib_rdma_desc_t *rd;
1774 lnet_finalize(ni, lntmsg, 0);
1775 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK, 0,
1776 rxmsg->ibm_u.putreq.ibprm_cookie);
1780 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
1782 CERROR("Can't allocate tx for %s\n",
1783 libcfs_nid2str(conn->ibc_peer->ibp_nid));
1784 /* Not replying will break the connection */
1790 rd = &txmsg->ibm_u.putack.ibpam_rd;
1792 rc = kiblnd_setup_rd_iov(ni, tx, rd,
1793 niov, iov, offset, mlen);
1795 rc = kiblnd_setup_rd_kiov(ni, tx, rd,
1796 niov, kiov, offset, mlen);
1798 CERROR("Can't setup PUT sink for %s: %d\n",
1799 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
1800 kiblnd_tx_done(ni, tx);
1801 /* tell peer it's over */
1802 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK, rc,
1803 rxmsg->ibm_u.putreq.ibprm_cookie);
1807 nob = offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[rd->rd_nfrags]);
1808 txmsg->ibm_u.putack.ibpam_src_cookie = rxmsg->ibm_u.putreq.ibprm_cookie;
1809 txmsg->ibm_u.putack.ibpam_dst_cookie = tx->tx_cookie;
1811 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_ACK, nob);
1813 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1814 tx->tx_waiting = 1; /* waiting for PUT_DONE */
1815 kiblnd_queue_tx(tx, conn);
1817 /* reposted buffer reserved for PUT_DONE */
1818 post_credit = IBLND_POSTRX_NO_CREDIT;
1822 case IBLND_MSG_GET_REQ:
1823 if (lntmsg != NULL) {
1824 /* Optimized GET; RDMA lntmsg's payload */
1825 kiblnd_reply(ni, rx, lntmsg);
1827 /* GET didn't match anything */
1828 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_GET_DONE,
1830 rxmsg->ibm_u.get.ibgm_cookie);
1835 kiblnd_post_rx(rx, post_credit);
1840 kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name)
1842 struct task_struct *task = kthread_run(fn, arg, name);
1845 return PTR_ERR(task);
1847 atomic_inc(&kiblnd_data.kib_nthreads);
1852 kiblnd_thread_fini (void)
1854 atomic_dec (&kiblnd_data.kib_nthreads);
1858 kiblnd_peer_alive (kib_peer_t *peer)
1860 /* This is racy, but everyone's only writing cfs_time_current() */
1861 peer->ibp_last_alive = cfs_time_current();
1866 kiblnd_peer_notify (kib_peer_t *peer)
1869 cfs_time_t last_alive = 0;
1870 unsigned long flags;
1872 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1874 if (kiblnd_peer_idle(peer) && peer->ibp_error != 0) {
1875 error = peer->ibp_error;
1876 peer->ibp_error = 0;
1878 last_alive = peer->ibp_last_alive;
1881 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1884 lnet_notify(peer->ibp_ni,
1885 peer->ibp_nid, 0, last_alive);
1889 kiblnd_close_conn_locked (kib_conn_t *conn, int error)
1891 /* This just does the immediate housekeeping. 'error' is zero for a
1892 * normal shutdown which can happen only after the connection has been
1893 * established. If the connection is established, schedule the
1894 * connection to be finished off by the connd. Otherwise the connd is
1895 * already dealing with it (either to set it up or tear it down).
1896 * Caller holds kib_global_lock exclusively in irq context */
1897 kib_peer_t *peer = conn->ibc_peer;
1899 unsigned long flags;
1901 LASSERT (error != 0 || conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1903 if (error != 0 && conn->ibc_comms_error == 0)
1904 conn->ibc_comms_error = error;
1906 if (conn->ibc_state != IBLND_CONN_ESTABLISHED)
1907 return; /* already being handled */
1910 list_empty(&conn->ibc_tx_noops) &&
1911 list_empty(&conn->ibc_tx_queue) &&
1912 list_empty(&conn->ibc_tx_queue_rsrvd) &&
1913 list_empty(&conn->ibc_tx_queue_nocred) &&
1914 list_empty(&conn->ibc_active_txs)) {
1915 CDEBUG(D_NET, "closing conn to %s\n",
1916 libcfs_nid2str(peer->ibp_nid));
1918 CNETERR("Closing conn to %s: error %d%s%s%s%s%s\n",
1919 libcfs_nid2str(peer->ibp_nid), error,
1920 list_empty(&conn->ibc_tx_queue) ? "" : "(sending)",
1921 list_empty(&conn->ibc_tx_noops) ? "" : "(sending_noops)",
1922 list_empty(&conn->ibc_tx_queue_rsrvd) ?
1923 "" : "(sending_rsrvd)",
1924 list_empty(&conn->ibc_tx_queue_nocred) ?
1925 "" : "(sending_nocred)",
1926 list_empty(&conn->ibc_active_txs) ? "" : "(waiting)");
1929 dev = ((kib_net_t *)peer->ibp_ni->ni_data)->ibn_dev;
1930 list_del(&conn->ibc_list);
1931 /* connd (see below) takes over ibc_list's ref */
1933 if (list_empty(&peer->ibp_conns) && /* no more conns */
1934 kiblnd_peer_active(peer)) { /* still in peer table */
1935 kiblnd_unlink_peer_locked(peer);
1937 /* set/clear error on last conn */
1938 peer->ibp_error = conn->ibc_comms_error;
1941 kiblnd_set_conn_state(conn, IBLND_CONN_CLOSING);
1944 kiblnd_dev_can_failover(dev)) {
1945 list_add_tail(&dev->ibd_fail_list,
1946 &kiblnd_data.kib_failed_devs);
1947 wake_up(&kiblnd_data.kib_failover_waitq);
1950 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
1952 list_add_tail(&conn->ibc_list, &kiblnd_data.kib_connd_conns);
1953 wake_up(&kiblnd_data.kib_connd_waitq);
1955 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
1959 kiblnd_close_conn(kib_conn_t *conn, int error)
1961 unsigned long flags;
1963 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1965 kiblnd_close_conn_locked(conn, error);
1967 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1971 kiblnd_handle_early_rxs(kib_conn_t *conn)
1973 unsigned long flags;
1976 LASSERT(!in_interrupt());
1977 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1979 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1980 while (!list_empty(&conn->ibc_early_rxs)) {
1981 rx = list_entry(conn->ibc_early_rxs.next,
1983 list_del(&rx->rx_list);
1984 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1986 kiblnd_handle_rx(rx);
1988 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1990 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1994 kiblnd_abort_txs(kib_conn_t *conn, struct list_head *txs)
1996 struct list_head zombies = LIST_HEAD_INIT(zombies);
1997 struct list_head *tmp;
1998 struct list_head *nxt;
2001 spin_lock(&conn->ibc_lock);
2003 list_for_each_safe(tmp, nxt, txs) {
2004 tx = list_entry(tmp, kib_tx_t, tx_list);
2006 if (txs == &conn->ibc_active_txs) {
2007 LASSERT(!tx->tx_queued);
2008 LASSERT(tx->tx_waiting ||
2009 tx->tx_sending != 0);
2011 LASSERT(tx->tx_queued);
2014 tx->tx_status = -ECONNABORTED;
2017 if (tx->tx_sending == 0) {
2019 list_del(&tx->tx_list);
2020 list_add(&tx->tx_list, &zombies);
2024 spin_unlock(&conn->ibc_lock);
2026 kiblnd_txlist_done(conn->ibc_peer->ibp_ni, &zombies, -ECONNABORTED);
2030 kiblnd_finalise_conn (kib_conn_t *conn)
2032 LASSERT (!in_interrupt());
2033 LASSERT (conn->ibc_state > IBLND_CONN_INIT);
2035 kiblnd_set_conn_state(conn, IBLND_CONN_DISCONNECTED);
2037 /* abort_receives moves QP state to IB_QPS_ERR. This is only required
2038 * for connections that didn't get as far as being connected, because
2039 * rdma_disconnect() does this for free. */
2040 kiblnd_abort_receives(conn);
2042 /* Complete all tx descs not waiting for sends to complete.
2043 * NB we should be safe from RDMA now that the QP has changed state */
2045 kiblnd_abort_txs(conn, &conn->ibc_tx_noops);
2046 kiblnd_abort_txs(conn, &conn->ibc_tx_queue);
2047 kiblnd_abort_txs(conn, &conn->ibc_tx_queue_rsrvd);
2048 kiblnd_abort_txs(conn, &conn->ibc_tx_queue_nocred);
2049 kiblnd_abort_txs(conn, &conn->ibc_active_txs);
2051 kiblnd_handle_early_rxs(conn);
2055 kiblnd_peer_connect_failed(kib_peer_t *peer, int active, int error)
2057 struct list_head zombies = LIST_HEAD_INIT(zombies);
2058 unsigned long flags;
2060 LASSERT (error != 0);
2061 LASSERT (!in_interrupt());
2063 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2066 LASSERT (peer->ibp_connecting > 0);
2067 peer->ibp_connecting--;
2069 LASSERT (peer->ibp_accepting > 0);
2070 peer->ibp_accepting--;
2073 if (kiblnd_peer_connecting(peer)) {
2074 /* another connection attempt under way... */
2075 write_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2080 peer->ibp_reconnected = 0;
2081 if (list_empty(&peer->ibp_conns)) {
2082 /* Take peer's blocked transmits to complete with error */
2083 list_add(&zombies, &peer->ibp_tx_queue);
2084 list_del_init(&peer->ibp_tx_queue);
2086 if (kiblnd_peer_active(peer))
2087 kiblnd_unlink_peer_locked(peer);
2089 peer->ibp_error = error;
2091 /* Can't have blocked transmits if there are connections */
2092 LASSERT(list_empty(&peer->ibp_tx_queue));
2095 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2097 kiblnd_peer_notify(peer);
2099 if (list_empty(&zombies))
2102 CNETERR("Deleting messages for %s: connection failed\n",
2103 libcfs_nid2str(peer->ibp_nid));
2105 kiblnd_txlist_done(peer->ibp_ni, &zombies, -EHOSTUNREACH);
2109 kiblnd_connreq_done(kib_conn_t *conn, int status)
2111 kib_peer_t *peer = conn->ibc_peer;
2113 struct list_head txs;
2114 unsigned long flags;
2117 active = (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2119 CDEBUG(D_NET,"%s: active(%d), version(%x), status(%d)\n",
2120 libcfs_nid2str(peer->ibp_nid), active,
2121 conn->ibc_version, status);
2123 LASSERT (!in_interrupt());
2124 LASSERT ((conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT &&
2125 peer->ibp_connecting > 0) ||
2126 (conn->ibc_state == IBLND_CONN_PASSIVE_WAIT &&
2127 peer->ibp_accepting > 0));
2129 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
2130 conn->ibc_connvars = NULL;
2133 /* failed to establish connection */
2134 kiblnd_peer_connect_failed(peer, active, status);
2135 kiblnd_finalise_conn(conn);
2139 /* connection established */
2140 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2142 conn->ibc_last_send = jiffies;
2143 kiblnd_set_conn_state(conn, IBLND_CONN_ESTABLISHED);
2144 kiblnd_peer_alive(peer);
2146 /* Add conn to peer's list and nuke any dangling conns from a different
2147 * peer instance... */
2148 kiblnd_conn_addref(conn); /* +1 ref for ibc_list */
2149 list_add(&conn->ibc_list, &peer->ibp_conns);
2150 peer->ibp_reconnected = 0;
2152 peer->ibp_connecting--;
2154 peer->ibp_accepting--;
2156 if (peer->ibp_version == 0) {
2157 peer->ibp_version = conn->ibc_version;
2158 peer->ibp_incarnation = conn->ibc_incarnation;
2161 if (peer->ibp_version != conn->ibc_version ||
2162 peer->ibp_incarnation != conn->ibc_incarnation) {
2163 kiblnd_close_stale_conns_locked(peer, conn->ibc_version,
2164 conn->ibc_incarnation);
2165 peer->ibp_version = conn->ibc_version;
2166 peer->ibp_incarnation = conn->ibc_incarnation;
2169 /* grab pending txs while I have the lock */
2170 list_add(&txs, &peer->ibp_tx_queue);
2171 list_del_init(&peer->ibp_tx_queue);
2173 if (!kiblnd_peer_active(peer) || /* peer has been deleted */
2174 conn->ibc_comms_error != 0) { /* error has happened already */
2175 lnet_ni_t *ni = peer->ibp_ni;
2177 /* start to shut down connection */
2178 kiblnd_close_conn_locked(conn, -ECONNABORTED);
2179 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2181 kiblnd_txlist_done(ni, &txs, -ECONNABORTED);
2186 /* refcount taken by cmid is not reliable after I released the glock
2187 * because this connection is visible to other threads now, another
2188 * thread can find and close this connection right after I released
2189 * the glock, if kiblnd_cm_callback for RDMA_CM_EVENT_DISCONNECTED is
2190 * called, it can release the connection refcount taken by cmid.
2191 * It means the connection could be destroyed before I finish my
2194 kiblnd_conn_addref(conn);
2195 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2197 /* Schedule blocked txs */
2198 spin_lock(&conn->ibc_lock);
2199 while (!list_empty(&txs)) {
2200 tx = list_entry(txs.next, kib_tx_t, tx_list);
2201 list_del(&tx->tx_list);
2203 kiblnd_queue_tx_locked(tx, conn);
2205 spin_unlock(&conn->ibc_lock);
2207 kiblnd_check_sends(conn);
2209 /* schedule blocked rxs */
2210 kiblnd_handle_early_rxs(conn);
2212 kiblnd_conn_decref(conn);
2216 kiblnd_reject(struct rdma_cm_id *cmid, kib_rej_t *rej)
2220 rc = rdma_reject(cmid, rej, sizeof(*rej));
2223 CWARN("Error %d sending reject\n", rc);
2227 kiblnd_passive_connect(struct rdma_cm_id *cmid, void *priv, int priv_nob)
2229 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2230 kib_msg_t *reqmsg = priv;
2236 lnet_ni_t *ni = NULL;
2237 kib_net_t *net = NULL;
2239 struct rdma_conn_param cp;
2241 int version = IBLND_MSG_VERSION;
2242 unsigned long flags;
2244 struct sockaddr_in *peer_addr;
2245 LASSERT (!in_interrupt());
2247 /* cmid inherits 'context' from the corresponding listener id */
2248 ibdev = (kib_dev_t *)cmid->context;
2249 LASSERT (ibdev != NULL);
2251 memset(&rej, 0, sizeof(rej));
2252 rej.ibr_magic = IBLND_MSG_MAGIC;
2253 rej.ibr_why = IBLND_REJECT_FATAL;
2254 rej.ibr_cp.ibcp_max_msg_size = IBLND_MSG_SIZE;
2256 peer_addr = (struct sockaddr_in *)&(cmid->route.addr.dst_addr);
2257 if (*kiblnd_tunables.kib_require_priv_port &&
2258 ntohs(peer_addr->sin_port) >= PROT_SOCK) {
2259 __u32 ip = ntohl(peer_addr->sin_addr.s_addr);
2260 CERROR("Peer's port (%pI4h:%hu) is not privileged\n",
2261 &ip, ntohs(peer_addr->sin_port));
2265 if (priv_nob < offsetof(kib_msg_t, ibm_type)) {
2266 CERROR("Short connection request\n");
2270 /* Future protocol version compatibility support! If the
2271 * o2iblnd-specific protocol changes, or when LNET unifies
2272 * protocols over all LNDs, the initial connection will
2273 * negotiate a protocol version. I trap this here to avoid
2274 * console errors; the reject tells the peer which protocol I
2276 if (reqmsg->ibm_magic == LNET_PROTO_MAGIC ||
2277 reqmsg->ibm_magic == __swab32(LNET_PROTO_MAGIC))
2279 if (reqmsg->ibm_magic == IBLND_MSG_MAGIC &&
2280 reqmsg->ibm_version != IBLND_MSG_VERSION &&
2281 reqmsg->ibm_version != IBLND_MSG_VERSION_1)
2283 if (reqmsg->ibm_magic == __swab32(IBLND_MSG_MAGIC) &&
2284 reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION) &&
2285 reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION_1))
2288 rc = kiblnd_unpack_msg(reqmsg, priv_nob);
2290 CERROR("Can't parse connection request: %d\n", rc);
2294 nid = reqmsg->ibm_srcnid;
2295 ni = lnet_net2ni(LNET_NIDNET(reqmsg->ibm_dstnid));
2298 net = (kib_net_t *)ni->ni_data;
2299 rej.ibr_incarnation = net->ibn_incarnation;
2302 if (ni == NULL || /* no matching net */
2303 ni->ni_nid != reqmsg->ibm_dstnid || /* right NET, wrong NID! */
2304 net->ibn_dev != ibdev) { /* wrong device */
2305 CERROR("Can't accept conn from %s on %s (%s:%d:%pI4h): "
2306 "bad dst nid %s\n", libcfs_nid2str(nid),
2307 ni == NULL ? "NA" : libcfs_nid2str(ni->ni_nid),
2308 ibdev->ibd_ifname, ibdev->ibd_nnets,
2310 libcfs_nid2str(reqmsg->ibm_dstnid));
2315 /* check time stamp as soon as possible */
2316 if (reqmsg->ibm_dststamp != 0 &&
2317 reqmsg->ibm_dststamp != net->ibn_incarnation) {
2318 CWARN("Stale connection request\n");
2319 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2323 /* I can accept peer's version */
2324 version = reqmsg->ibm_version;
2326 if (reqmsg->ibm_type != IBLND_MSG_CONNREQ) {
2327 CERROR("Unexpected connreq msg type: %x from %s\n",
2328 reqmsg->ibm_type, libcfs_nid2str(nid));
2332 if (reqmsg->ibm_u.connparams.ibcp_queue_depth >
2333 kiblnd_msg_queue_size(version, ni)) {
2334 CERROR("Can't accept conn from %s, queue depth too large: "
2335 " %d (<=%d wanted)\n",
2336 libcfs_nid2str(nid),
2337 reqmsg->ibm_u.connparams.ibcp_queue_depth,
2338 kiblnd_msg_queue_size(version, ni));
2340 if (version == IBLND_MSG_VERSION)
2341 rej.ibr_why = IBLND_REJECT_MSG_QUEUE_SIZE;
2346 if (reqmsg->ibm_u.connparams.ibcp_max_frags >
2347 kiblnd_rdma_frags(version, ni)) {
2348 CWARN("Can't accept conn from %s (version %x): "
2349 "max_frags %d too large (%d wanted)\n",
2350 libcfs_nid2str(nid), version,
2351 reqmsg->ibm_u.connparams.ibcp_max_frags,
2352 kiblnd_rdma_frags(version, ni));
2354 if (version >= IBLND_MSG_VERSION)
2355 rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2358 } else if (reqmsg->ibm_u.connparams.ibcp_max_frags <
2359 kiblnd_rdma_frags(version, ni) &&
2360 net->ibn_fmr_ps == NULL) {
2361 CWARN("Can't accept conn from %s (version %x): "
2362 "max_frags %d incompatible without FMR pool "
2364 libcfs_nid2str(nid), version,
2365 reqmsg->ibm_u.connparams.ibcp_max_frags,
2366 kiblnd_rdma_frags(version, ni));
2368 if (version == IBLND_MSG_VERSION)
2369 rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2374 if (reqmsg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2375 CERROR("Can't accept %s: message size %d too big (%d max)\n",
2376 libcfs_nid2str(nid),
2377 reqmsg->ibm_u.connparams.ibcp_max_msg_size,
2382 /* assume 'nid' is a new peer; create */
2383 rc = kiblnd_create_peer(ni, &peer, nid);
2385 CERROR("Can't create peer for %s\n", libcfs_nid2str(nid));
2386 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2390 /* We have validated the peer's parameters so use those */
2391 peer->ibp_max_frags = reqmsg->ibm_u.connparams.ibcp_max_frags;
2392 peer->ibp_queue_depth = reqmsg->ibm_u.connparams.ibcp_queue_depth;
2394 write_lock_irqsave(g_lock, flags);
2396 peer2 = kiblnd_find_peer_locked(nid);
2397 if (peer2 != NULL) {
2398 if (peer2->ibp_version == 0) {
2399 peer2->ibp_version = version;
2400 peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2403 /* not the guy I've talked with */
2404 if (peer2->ibp_incarnation != reqmsg->ibm_srcstamp ||
2405 peer2->ibp_version != version) {
2406 kiblnd_close_peer_conns_locked(peer2, -ESTALE);
2408 if (kiblnd_peer_active(peer2)) {
2409 peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2410 peer2->ibp_version = version;
2412 write_unlock_irqrestore(g_lock, flags);
2414 CWARN("Conn stale %s version %x/%x incarnation "LPU64"/"LPU64"\n",
2415 libcfs_nid2str(nid), peer2->ibp_version, version,
2416 peer2->ibp_incarnation, reqmsg->ibm_srcstamp);
2418 kiblnd_peer_decref(peer);
2419 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2423 /* tie-break connection race in favour of the higher NID */
2424 if (peer2->ibp_connecting != 0 &&
2426 write_unlock_irqrestore(g_lock, flags);
2428 CWARN("Conn race %s\n", libcfs_nid2str(peer2->ibp_nid));
2430 kiblnd_peer_decref(peer);
2431 rej.ibr_why = IBLND_REJECT_CONN_RACE;
2435 * passive connection is allowed even this peer is waiting for
2438 peer2->ibp_reconnecting = 0;
2439 peer2->ibp_accepting++;
2440 kiblnd_peer_addref(peer2);
2442 /* Race with kiblnd_launch_tx (active connect) to create peer
2443 * so copy validated parameters since we now know what the
2444 * peer's limits are */
2445 peer2->ibp_max_frags = peer->ibp_max_frags;
2446 peer2->ibp_queue_depth = peer->ibp_queue_depth;
2448 write_unlock_irqrestore(g_lock, flags);
2449 kiblnd_peer_decref(peer);
2452 /* Brand new peer */
2453 LASSERT (peer->ibp_accepting == 0);
2454 LASSERT (peer->ibp_version == 0 &&
2455 peer->ibp_incarnation == 0);
2457 peer->ibp_accepting = 1;
2458 peer->ibp_version = version;
2459 peer->ibp_incarnation = reqmsg->ibm_srcstamp;
2461 /* I have a ref on ni that prevents it being shutdown */
2462 LASSERT (net->ibn_shutdown == 0);
2464 kiblnd_peer_addref(peer);
2465 list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid));
2467 write_unlock_irqrestore(g_lock, flags);
2470 conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_PASSIVE_WAIT, version);
2472 kiblnd_peer_connect_failed(peer, 0, -ENOMEM);
2473 kiblnd_peer_decref(peer);
2474 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2478 /* conn now "owns" cmid, so I return success from here on to ensure the
2479 * CM callback doesn't destroy cmid. */
2480 conn->ibc_incarnation = reqmsg->ibm_srcstamp;
2481 conn->ibc_credits = conn->ibc_queue_depth;
2482 conn->ibc_reserved_credits = conn->ibc_queue_depth;
2483 LASSERT(conn->ibc_credits + conn->ibc_reserved_credits +
2484 IBLND_OOB_MSGS(version) <= IBLND_RX_MSGS(conn));
2486 ackmsg = &conn->ibc_connvars->cv_msg;
2487 memset(ackmsg, 0, sizeof(*ackmsg));
2489 kiblnd_init_msg(ackmsg, IBLND_MSG_CONNACK,
2490 sizeof(ackmsg->ibm_u.connparams));
2491 ackmsg->ibm_u.connparams.ibcp_queue_depth = conn->ibc_queue_depth;
2492 ackmsg->ibm_u.connparams.ibcp_max_frags = conn->ibc_max_frags;
2493 ackmsg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2495 kiblnd_pack_msg(ni, ackmsg, version, 0, nid, reqmsg->ibm_srcstamp);
2497 memset(&cp, 0, sizeof(cp));
2498 cp.private_data = ackmsg;
2499 cp.private_data_len = ackmsg->ibm_nob;
2500 cp.responder_resources = 0; /* No atomic ops or RDMA reads */
2501 cp.initiator_depth = 0;
2502 cp.flow_control = 1;
2503 cp.retry_count = *kiblnd_tunables.kib_retry_count;
2504 cp.rnr_retry_count = *kiblnd_tunables.kib_rnr_retry_count;
2506 CDEBUG(D_NET, "Accept %s\n", libcfs_nid2str(nid));
2508 rc = rdma_accept(cmid, &cp);
2510 CERROR("Can't accept %s: %d\n", libcfs_nid2str(nid), rc);
2511 rej.ibr_version = version;
2512 rej.ibr_why = IBLND_REJECT_FATAL;
2514 kiblnd_reject(cmid, &rej);
2515 kiblnd_connreq_done(conn, rc);
2516 kiblnd_conn_decref(conn);
2526 rej.ibr_version = version;
2527 rej.ibr_cp.ibcp_queue_depth = kiblnd_msg_queue_size(version, ni);
2528 rej.ibr_cp.ibcp_max_frags = kiblnd_rdma_frags(version, ni);
2529 kiblnd_reject(cmid, &rej);
2531 return -ECONNREFUSED;
2535 kiblnd_check_reconnect(kib_conn_t *conn, int version,
2536 __u64 incarnation, int why, kib_connparams_t *cp)
2538 rwlock_t *glock = &kiblnd_data.kib_global_lock;
2539 kib_peer_t *peer = conn->ibc_peer;
2541 int msg_size = IBLND_MSG_SIZE;
2545 unsigned long flags;
2547 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2548 LASSERT(peer->ibp_connecting > 0); /* 'conn' at least */
2549 LASSERT(!peer->ibp_reconnecting);
2552 msg_size = cp->ibcp_max_msg_size;
2553 frag_num = cp->ibcp_max_frags;
2554 queue_dep = cp->ibcp_queue_depth;
2557 write_lock_irqsave(glock, flags);
2558 /* retry connection if it's still needed and no other connection
2559 * attempts (active or passive) are in progress
2560 * NB: reconnect is still needed even when ibp_tx_queue is
2561 * empty if ibp_version != version because reconnect may be
2562 * initiated by kiblnd_query() */
2563 reconnect = (!list_empty(&peer->ibp_tx_queue) ||
2564 peer->ibp_version != version) &&
2565 peer->ibp_connecting == 1 &&
2566 peer->ibp_accepting == 0;
2577 case IBLND_REJECT_RDMA_FRAGS: {
2578 struct lnet_ioctl_config_lnd_tunables *tunables;
2581 reason = "can't negotiate max frags";
2584 tunables = peer->ibp_ni->ni_lnd_tunables;
2585 if (!tunables->lt_tun_u.lt_o2ib.lnd_map_on_demand) {
2586 reason = "map_on_demand must be enabled";
2589 if (conn->ibc_max_frags <= frag_num) {
2590 reason = "unsupported max frags";
2594 peer->ibp_max_frags = frag_num;
2595 reason = "rdma fragments";
2598 case IBLND_REJECT_MSG_QUEUE_SIZE:
2600 reason = "can't negotiate queue depth";
2603 if (conn->ibc_queue_depth <= queue_dep) {
2604 reason = "unsupported queue depth";
2608 peer->ibp_queue_depth = queue_dep;
2609 reason = "queue depth";
2612 case IBLND_REJECT_CONN_STALE:
2616 case IBLND_REJECT_CONN_RACE:
2617 reason = "conn race";
2620 case IBLND_REJECT_CONN_UNCOMPAT:
2621 reason = "version negotiation";
2625 conn->ibc_reconnect = 1;
2626 peer->ibp_reconnecting = 1;
2627 peer->ibp_version = version;
2628 if (incarnation != 0)
2629 peer->ibp_incarnation = incarnation;
2631 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2633 CNETERR("%s: %s (%s), %x, %x, msg_size: %d, queue_depth: %d/%d, max_frags: %d/%d\n",
2634 libcfs_nid2str(peer->ibp_nid),
2635 reconnect ? "reconnect" : "don't reconnect",
2636 reason, IBLND_MSG_VERSION, version, msg_size,
2637 conn->ibc_queue_depth, queue_dep,
2638 conn->ibc_max_frags, frag_num);
2640 * if conn::ibc_reconnect is TRUE, connd will reconnect to the peer
2641 * while destroying the zombie
2646 kiblnd_rejected (kib_conn_t *conn, int reason, void *priv, int priv_nob)
2648 kib_peer_t *peer = conn->ibc_peer;
2650 LASSERT (!in_interrupt());
2651 LASSERT (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2654 case IB_CM_REJ_STALE_CONN:
2655 kiblnd_check_reconnect(conn, IBLND_MSG_VERSION, 0,
2656 IBLND_REJECT_CONN_STALE, NULL);
2659 case IB_CM_REJ_INVALID_SERVICE_ID:
2660 CNETERR("%s rejected: no listener at %d\n",
2661 libcfs_nid2str(peer->ibp_nid),
2662 *kiblnd_tunables.kib_service);
2665 case IB_CM_REJ_CONSUMER_DEFINED:
2666 if (priv_nob >= offsetof(kib_rej_t, ibr_padding)) {
2667 kib_rej_t *rej = priv;
2668 kib_connparams_t *cp = NULL;
2670 __u64 incarnation = -1;
2672 /* NB. default incarnation is -1 because:
2673 * a) V1 will ignore dst incarnation in connreq.
2674 * b) V2 will provide incarnation while rejecting me,
2675 * -1 will be overwrote.
2677 * if I try to connect to a V1 peer with V2 protocol,
2678 * it rejected me then upgrade to V2, I have no idea
2679 * about the upgrading and try to reconnect with V1,
2680 * in this case upgraded V2 can find out I'm trying to
2681 * talk to the old guy and reject me(incarnation is -1).
2684 if (rej->ibr_magic == __swab32(IBLND_MSG_MAGIC) ||
2685 rej->ibr_magic == __swab32(LNET_PROTO_MAGIC)) {
2686 __swab32s(&rej->ibr_magic);
2687 __swab16s(&rej->ibr_version);
2691 if (priv_nob >= sizeof(kib_rej_t) &&
2692 rej->ibr_version > IBLND_MSG_VERSION_1) {
2693 /* priv_nob is always 148 in current version
2694 * of OFED, so we still need to check version.
2695 * (define of IB_CM_REJ_PRIVATE_DATA_SIZE) */
2699 __swab64s(&rej->ibr_incarnation);
2700 __swab16s(&cp->ibcp_queue_depth);
2701 __swab16s(&cp->ibcp_max_frags);
2702 __swab32s(&cp->ibcp_max_msg_size);
2705 incarnation = rej->ibr_incarnation;
2708 if (rej->ibr_magic != IBLND_MSG_MAGIC &&
2709 rej->ibr_magic != LNET_PROTO_MAGIC) {
2710 CERROR("%s rejected: consumer defined fatal error\n",
2711 libcfs_nid2str(peer->ibp_nid));
2715 if (rej->ibr_version != IBLND_MSG_VERSION &&
2716 rej->ibr_version != IBLND_MSG_VERSION_1) {
2717 CERROR("%s rejected: o2iblnd version %x error\n",
2718 libcfs_nid2str(peer->ibp_nid),
2723 if (rej->ibr_why == IBLND_REJECT_FATAL &&
2724 rej->ibr_version == IBLND_MSG_VERSION_1) {
2725 CDEBUG(D_NET, "rejected by old version peer %s: %x\n",
2726 libcfs_nid2str(peer->ibp_nid), rej->ibr_version);
2728 if (conn->ibc_version != IBLND_MSG_VERSION_1)
2729 rej->ibr_why = IBLND_REJECT_CONN_UNCOMPAT;
2732 switch (rej->ibr_why) {
2733 case IBLND_REJECT_CONN_RACE:
2734 case IBLND_REJECT_CONN_STALE:
2735 case IBLND_REJECT_CONN_UNCOMPAT:
2736 case IBLND_REJECT_MSG_QUEUE_SIZE:
2737 case IBLND_REJECT_RDMA_FRAGS:
2738 kiblnd_check_reconnect(conn, rej->ibr_version,
2739 incarnation, rej->ibr_why, cp);
2742 case IBLND_REJECT_NO_RESOURCES:
2743 CERROR("%s rejected: o2iblnd no resources\n",
2744 libcfs_nid2str(peer->ibp_nid));
2747 case IBLND_REJECT_FATAL:
2748 CERROR("%s rejected: o2iblnd fatal error\n",
2749 libcfs_nid2str(peer->ibp_nid));
2753 CERROR("%s rejected: o2iblnd reason %d\n",
2754 libcfs_nid2str(peer->ibp_nid),
2762 CNETERR("%s rejected: reason %d, size %d\n",
2763 libcfs_nid2str(peer->ibp_nid), reason, priv_nob);
2767 kiblnd_connreq_done(conn, -ECONNREFUSED);
2771 kiblnd_check_connreply (kib_conn_t *conn, void *priv, int priv_nob)
2773 kib_peer_t *peer = conn->ibc_peer;
2774 lnet_ni_t *ni = peer->ibp_ni;
2775 kib_net_t *net = ni->ni_data;
2776 kib_msg_t *msg = priv;
2777 int ver = conn->ibc_version;
2778 int rc = kiblnd_unpack_msg(msg, priv_nob);
2779 unsigned long flags;
2781 LASSERT (net != NULL);
2784 CERROR("Can't unpack connack from %s: %d\n",
2785 libcfs_nid2str(peer->ibp_nid), rc);
2789 if (msg->ibm_type != IBLND_MSG_CONNACK) {
2790 CERROR("Unexpected message %d from %s\n",
2791 msg->ibm_type, libcfs_nid2str(peer->ibp_nid));
2796 if (ver != msg->ibm_version) {
2797 CERROR("%s replied version %x is different with "
2798 "requested version %x\n",
2799 libcfs_nid2str(peer->ibp_nid), msg->ibm_version, ver);
2804 if (msg->ibm_u.connparams.ibcp_queue_depth >
2805 conn->ibc_queue_depth) {
2806 CERROR("%s has incompatible queue depth %d (<=%d wanted)\n",
2807 libcfs_nid2str(peer->ibp_nid),
2808 msg->ibm_u.connparams.ibcp_queue_depth,
2809 conn->ibc_queue_depth);
2814 if (msg->ibm_u.connparams.ibcp_max_frags >
2815 conn->ibc_max_frags) {
2816 CERROR("%s has incompatible max_frags %d (<=%d wanted)\n",
2817 libcfs_nid2str(peer->ibp_nid),
2818 msg->ibm_u.connparams.ibcp_max_frags,
2819 conn->ibc_max_frags);
2824 if (msg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2825 CERROR("%s max message size %d too big (%d max)\n",
2826 libcfs_nid2str(peer->ibp_nid),
2827 msg->ibm_u.connparams.ibcp_max_msg_size,
2833 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2834 if (msg->ibm_dstnid == ni->ni_nid &&
2835 msg->ibm_dststamp == net->ibn_incarnation)
2839 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2842 CERROR("Bad connection reply from %s, rc = %d, "
2843 "version: %x max_frags: %d\n",
2844 libcfs_nid2str(peer->ibp_nid), rc,
2845 msg->ibm_version, msg->ibm_u.connparams.ibcp_max_frags);
2849 conn->ibc_incarnation = msg->ibm_srcstamp;
2850 conn->ibc_credits = msg->ibm_u.connparams.ibcp_queue_depth;
2851 conn->ibc_reserved_credits = msg->ibm_u.connparams.ibcp_queue_depth;
2852 conn->ibc_queue_depth = msg->ibm_u.connparams.ibcp_queue_depth;
2853 conn->ibc_max_frags = msg->ibm_u.connparams.ibcp_max_frags;
2854 LASSERT(conn->ibc_credits + conn->ibc_reserved_credits +
2855 IBLND_OOB_MSGS(ver) <= IBLND_RX_MSGS(conn));
2857 kiblnd_connreq_done(conn, 0);
2861 /* NB My QP has already established itself, so I handle anything going
2862 * wrong here by setting ibc_comms_error.
2863 * kiblnd_connreq_done(0) moves the conn state to ESTABLISHED, but then
2864 * immediately tears it down. */
2867 conn->ibc_comms_error = rc;
2868 kiblnd_connreq_done(conn, 0);
2872 kiblnd_active_connect (struct rdma_cm_id *cmid)
2874 kib_peer_t *peer = (kib_peer_t *)cmid->context;
2877 struct rdma_conn_param cp;
2880 unsigned long flags;
2883 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2885 incarnation = peer->ibp_incarnation;
2886 version = (peer->ibp_version == 0) ? IBLND_MSG_VERSION :
2889 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2891 conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_ACTIVE_CONNECT,
2894 kiblnd_peer_connect_failed(peer, 1, -ENOMEM);
2895 kiblnd_peer_decref(peer); /* lose cmid's ref */
2899 /* conn "owns" cmid now, so I return success from here on to ensure the
2900 * CM callback doesn't destroy cmid. conn also takes over cmid's ref
2903 msg = &conn->ibc_connvars->cv_msg;
2905 memset(msg, 0, sizeof(*msg));
2906 kiblnd_init_msg(msg, IBLND_MSG_CONNREQ, sizeof(msg->ibm_u.connparams));
2907 msg->ibm_u.connparams.ibcp_queue_depth = conn->ibc_queue_depth;
2908 msg->ibm_u.connparams.ibcp_max_frags = conn->ibc_max_frags;
2909 msg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2911 kiblnd_pack_msg(peer->ibp_ni, msg, version,
2912 0, peer->ibp_nid, incarnation);
2914 memset(&cp, 0, sizeof(cp));
2915 cp.private_data = msg;
2916 cp.private_data_len = msg->ibm_nob;
2917 cp.responder_resources = 0; /* No atomic ops or RDMA reads */
2918 cp.initiator_depth = 0;
2919 cp.flow_control = 1;
2920 cp.retry_count = *kiblnd_tunables.kib_retry_count;
2921 cp.rnr_retry_count = *kiblnd_tunables.kib_rnr_retry_count;
2923 LASSERT(cmid->context == (void *)conn);
2924 LASSERT(conn->ibc_cmid == cmid);
2926 rc = rdma_connect(cmid, &cp);
2928 CERROR("Can't connect to %s: %d\n",
2929 libcfs_nid2str(peer->ibp_nid), rc);
2930 kiblnd_connreq_done(conn, rc);
2931 kiblnd_conn_decref(conn);
2938 kiblnd_cm_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2944 switch (event->event) {
2946 CERROR("Unexpected event: %d, status: %d\n",
2947 event->event, event->status);
2950 case RDMA_CM_EVENT_CONNECT_REQUEST:
2951 /* destroy cmid on failure */
2952 rc = kiblnd_passive_connect(cmid,
2953 (void *)KIBLND_CONN_PARAM(event),
2954 KIBLND_CONN_PARAM_LEN(event));
2955 CDEBUG(D_NET, "connreq: %d\n", rc);
2958 case RDMA_CM_EVENT_ADDR_ERROR:
2959 peer = (kib_peer_t *)cmid->context;
2960 CNETERR("%s: ADDR ERROR %d\n",
2961 libcfs_nid2str(peer->ibp_nid), event->status);
2962 kiblnd_peer_connect_failed(peer, 1, -EHOSTUNREACH);
2963 kiblnd_peer_decref(peer);
2964 return -EHOSTUNREACH; /* rc != 0 destroys cmid */
2966 case RDMA_CM_EVENT_ADDR_RESOLVED:
2967 peer = (kib_peer_t *)cmid->context;
2969 CDEBUG(D_NET,"%s Addr resolved: %d\n",
2970 libcfs_nid2str(peer->ibp_nid), event->status);
2972 if (event->status != 0) {
2973 CNETERR("Can't resolve address for %s: %d\n",
2974 libcfs_nid2str(peer->ibp_nid), event->status);
2977 rc = rdma_resolve_route(
2978 cmid, *kiblnd_tunables.kib_timeout * 1000);
2981 /* Can't initiate route resolution */
2982 CERROR("Can't resolve route for %s: %d\n",
2983 libcfs_nid2str(peer->ibp_nid), rc);
2985 kiblnd_peer_connect_failed(peer, 1, rc);
2986 kiblnd_peer_decref(peer);
2987 return rc; /* rc != 0 destroys cmid */
2989 case RDMA_CM_EVENT_ROUTE_ERROR:
2990 peer = (kib_peer_t *)cmid->context;
2991 CNETERR("%s: ROUTE ERROR %d\n",
2992 libcfs_nid2str(peer->ibp_nid), event->status);
2993 kiblnd_peer_connect_failed(peer, 1, -EHOSTUNREACH);
2994 kiblnd_peer_decref(peer);
2995 return -EHOSTUNREACH; /* rc != 0 destroys cmid */
2997 case RDMA_CM_EVENT_ROUTE_RESOLVED:
2998 peer = (kib_peer_t *)cmid->context;
2999 CDEBUG(D_NET,"%s Route resolved: %d\n",
3000 libcfs_nid2str(peer->ibp_nid), event->status);
3002 if (event->status == 0)
3003 return kiblnd_active_connect(cmid);
3005 CNETERR("Can't resolve route for %s: %d\n",
3006 libcfs_nid2str(peer->ibp_nid), event->status);
3007 kiblnd_peer_connect_failed(peer, 1, event->status);
3008 kiblnd_peer_decref(peer);
3009 return event->status; /* rc != 0 destroys cmid */
3011 case RDMA_CM_EVENT_UNREACHABLE:
3012 conn = (kib_conn_t *)cmid->context;
3013 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
3014 conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
3015 CNETERR("%s: UNREACHABLE %d\n",
3016 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
3017 kiblnd_connreq_done(conn, -ENETDOWN);
3018 kiblnd_conn_decref(conn);
3021 case RDMA_CM_EVENT_CONNECT_ERROR:
3022 conn = (kib_conn_t *)cmid->context;
3023 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
3024 conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
3025 CNETERR("%s: CONNECT ERROR %d\n",
3026 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
3027 kiblnd_connreq_done(conn, -ENOTCONN);
3028 kiblnd_conn_decref(conn);
3031 case RDMA_CM_EVENT_REJECTED:
3032 conn = (kib_conn_t *)cmid->context;
3033 switch (conn->ibc_state) {
3037 case IBLND_CONN_PASSIVE_WAIT:
3038 CERROR ("%s: REJECTED %d\n",
3039 libcfs_nid2str(conn->ibc_peer->ibp_nid),
3041 kiblnd_connreq_done(conn, -ECONNRESET);
3044 case IBLND_CONN_ACTIVE_CONNECT:
3045 kiblnd_rejected(conn, event->status,
3046 (void *)KIBLND_CONN_PARAM(event),
3047 KIBLND_CONN_PARAM_LEN(event));
3050 kiblnd_conn_decref(conn);
3053 case RDMA_CM_EVENT_ESTABLISHED:
3054 conn = (kib_conn_t *)cmid->context;
3055 switch (conn->ibc_state) {
3059 case IBLND_CONN_PASSIVE_WAIT:
3060 CDEBUG(D_NET, "ESTABLISHED (passive): %s\n",
3061 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3062 kiblnd_connreq_done(conn, 0);
3065 case IBLND_CONN_ACTIVE_CONNECT:
3066 CDEBUG(D_NET, "ESTABLISHED(active): %s\n",
3067 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3068 kiblnd_check_connreply(conn,
3069 (void *)KIBLND_CONN_PARAM(event),
3070 KIBLND_CONN_PARAM_LEN(event));
3073 /* net keeps its ref on conn! */
3076 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
3077 CDEBUG(D_NET, "Ignore TIMEWAIT_EXIT event\n");
3080 case RDMA_CM_EVENT_DISCONNECTED:
3081 conn = (kib_conn_t *)cmid->context;
3082 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
3083 CERROR("%s DISCONNECTED\n",
3084 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3085 kiblnd_connreq_done(conn, -ECONNRESET);
3087 kiblnd_close_conn(conn, 0);
3089 kiblnd_conn_decref(conn);
3090 cmid->context = NULL;
3093 case RDMA_CM_EVENT_DEVICE_REMOVAL:
3094 LCONSOLE_ERROR_MSG(0x131,
3095 "Received notification of device removal\n"
3096 "Please shutdown LNET to allow this to proceed\n");
3097 /* Can't remove network from underneath LNET for now, so I have
3101 case RDMA_CM_EVENT_ADDR_CHANGE:
3102 LCONSOLE_INFO("Physical link changed (eg hca/port)\n");
3108 kiblnd_check_txs_locked(kib_conn_t *conn, struct list_head *txs)
3111 struct list_head *ttmp;
3113 list_for_each(ttmp, txs) {
3114 tx = list_entry(ttmp, kib_tx_t, tx_list);
3116 if (txs != &conn->ibc_active_txs) {
3117 LASSERT(tx->tx_queued);
3119 LASSERT(!tx->tx_queued);
3120 LASSERT(tx->tx_waiting || tx->tx_sending != 0);
3123 if (cfs_time_aftereq(jiffies, tx->tx_deadline)) {
3124 CERROR("Timed out tx: %s, %lu seconds\n",
3125 kiblnd_queue2str(conn, txs),
3126 cfs_duration_sec(jiffies - tx->tx_deadline));
3135 kiblnd_conn_timed_out_locked(kib_conn_t *conn)
3137 return kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue) ||
3138 kiblnd_check_txs_locked(conn, &conn->ibc_tx_noops) ||
3139 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_rsrvd) ||
3140 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_nocred) ||
3141 kiblnd_check_txs_locked(conn, &conn->ibc_active_txs);
3145 kiblnd_check_conns (int idx)
3147 struct list_head closes = LIST_HEAD_INIT(closes);
3148 struct list_head checksends = LIST_HEAD_INIT(checksends);
3149 struct list_head *peers = &kiblnd_data.kib_peers[idx];
3150 struct list_head *ptmp;
3153 struct list_head *ctmp;
3154 unsigned long flags;
3156 /* NB. We expect to have a look at all the peers and not find any
3157 * RDMAs to time out, so we just use a shared lock while we
3159 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3161 list_for_each(ptmp, peers) {
3162 peer = list_entry(ptmp, kib_peer_t, ibp_list);
3164 list_for_each(ctmp, &peer->ibp_conns) {
3168 conn = list_entry(ctmp, kib_conn_t, ibc_list);
3170 LASSERT(conn->ibc_state == IBLND_CONN_ESTABLISHED);
3172 spin_lock(&conn->ibc_lock);
3174 sendnoop = kiblnd_need_noop(conn);
3175 timedout = kiblnd_conn_timed_out_locked(conn);
3176 if (!sendnoop && !timedout) {
3177 spin_unlock(&conn->ibc_lock);
3182 CERROR("Timed out RDMA with %s (%lu): "
3183 "c: %u, oc: %u, rc: %u\n",
3184 libcfs_nid2str(peer->ibp_nid),
3185 cfs_duration_sec(cfs_time_current() -
3186 peer->ibp_last_alive),
3188 conn->ibc_outstanding_credits,
3189 conn->ibc_reserved_credits);
3190 list_add(&conn->ibc_connd_list, &closes);
3192 list_add(&conn->ibc_connd_list, &checksends);
3194 /* +ref for 'closes' or 'checksends' */
3195 kiblnd_conn_addref(conn);
3197 spin_unlock(&conn->ibc_lock);
3201 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3203 /* Handle timeout by closing the whole
3204 * connection. We can only be sure RDMA activity
3205 * has ceased once the QP has been modified. */
3206 while (!list_empty(&closes)) {
3207 conn = list_entry(closes.next,
3208 kib_conn_t, ibc_connd_list);
3209 list_del(&conn->ibc_connd_list);
3210 kiblnd_close_conn(conn, -ETIMEDOUT);
3211 kiblnd_conn_decref(conn);
3214 /* In case we have enough credits to return via a
3215 * NOOP, but there were no non-blocking tx descs
3216 * free to do it last time... */
3217 while (!list_empty(&checksends)) {
3218 conn = list_entry(checksends.next,
3219 kib_conn_t, ibc_connd_list);
3220 list_del(&conn->ibc_connd_list);
3221 kiblnd_check_sends(conn);
3222 kiblnd_conn_decref(conn);
3227 kiblnd_disconnect_conn (kib_conn_t *conn)
3229 LASSERT (!in_interrupt());
3230 LASSERT (current == kiblnd_data.kib_connd);
3231 LASSERT (conn->ibc_state == IBLND_CONN_CLOSING);
3233 rdma_disconnect(conn->ibc_cmid);
3234 kiblnd_finalise_conn(conn);
3236 kiblnd_peer_notify(conn->ibc_peer);
3240 * High-water for reconnection to the same peer, reconnection attempt should
3241 * be delayed after trying more than KIB_RECONN_HIGH_RACE.
3243 #define KIB_RECONN_HIGH_RACE 10
3245 * Allow connd to take a break and handle other things after consecutive
3246 * reconnection attemps.
3248 #define KIB_RECONN_BREAK 100
3251 kiblnd_connd (void *arg)
3253 spinlock_t *lock= &kiblnd_data.kib_connd_lock;
3255 unsigned long flags;
3261 unsigned long deadline = jiffies;
3263 cfs_block_allsigs();
3265 init_waitqueue_entry(&wait, current);
3266 kiblnd_data.kib_connd = current;
3268 spin_lock_irqsave(lock, flags);
3270 while (!kiblnd_data.kib_shutdown) {
3275 if (!list_empty(&kiblnd_data.kib_connd_zombies)) {
3276 kib_peer_t *peer = NULL;
3278 conn = list_entry(kiblnd_data.kib_connd_zombies.next,
3279 kib_conn_t, ibc_list);
3280 list_del(&conn->ibc_list);
3281 if (conn->ibc_reconnect) {
3282 peer = conn->ibc_peer;
3283 kiblnd_peer_addref(peer);
3286 spin_unlock_irqrestore(lock, flags);
3289 kiblnd_destroy_conn(conn, !peer);
3291 spin_lock_irqsave(lock, flags);
3295 conn->ibc_peer = peer;
3296 if (peer->ibp_reconnected < KIB_RECONN_HIGH_RACE)
3297 list_add_tail(&conn->ibc_list,
3298 &kiblnd_data.kib_reconn_list);
3300 list_add_tail(&conn->ibc_list,
3301 &kiblnd_data.kib_reconn_wait);
3304 if (!list_empty(&kiblnd_data.kib_connd_conns)) {
3305 conn = list_entry(kiblnd_data.kib_connd_conns.next,
3306 kib_conn_t, ibc_list);
3307 list_del(&conn->ibc_list);
3309 spin_unlock_irqrestore(lock, flags);
3312 kiblnd_disconnect_conn(conn);
3313 kiblnd_conn_decref(conn);
3315 spin_lock_irqsave(lock, flags);
3318 while (reconn < KIB_RECONN_BREAK) {
3319 if (kiblnd_data.kib_reconn_sec != get_seconds()) {
3320 kiblnd_data.kib_reconn_sec = get_seconds();
3321 list_splice_init(&kiblnd_data.kib_reconn_wait,
3322 &kiblnd_data.kib_reconn_list);
3325 if (list_empty(&kiblnd_data.kib_reconn_list))
3328 conn = list_entry(kiblnd_data.kib_reconn_list.next,
3329 kib_conn_t, ibc_list);
3330 list_del(&conn->ibc_list);
3332 spin_unlock_irqrestore(lock, flags);
3335 reconn += kiblnd_reconnect_peer(conn->ibc_peer);
3336 kiblnd_peer_decref(conn->ibc_peer);
3337 LIBCFS_FREE(conn, sizeof(*conn));
3339 spin_lock_irqsave(lock, flags);
3342 /* careful with the jiffy wrap... */
3343 timeout = (int)(deadline - jiffies);
3347 int chunk = kiblnd_data.kib_peer_hash_size;
3349 spin_unlock_irqrestore(lock, flags);
3352 /* Time to check for RDMA timeouts on a few more
3353 * peers: I do checks every 'p' seconds on a
3354 * proportion of the peer table and I need to check
3355 * every connection 'n' times within a timeout
3356 * interval, to ensure I detect a timeout on any
3357 * connection within (n+1)/n times the timeout
3360 if (*kiblnd_tunables.kib_timeout > n * p)
3361 chunk = (chunk * n * p) /
3362 *kiblnd_tunables.kib_timeout;
3366 for (i = 0; i < chunk; i++) {
3367 kiblnd_check_conns(peer_index);
3368 peer_index = (peer_index + 1) %
3369 kiblnd_data.kib_peer_hash_size;
3372 deadline += msecs_to_jiffies(p * MSEC_PER_SEC);
3373 spin_lock_irqsave(lock, flags);
3379 /* Nothing to do for 'timeout' */
3380 set_current_state(TASK_INTERRUPTIBLE);
3381 add_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3382 spin_unlock_irqrestore(lock, flags);
3384 schedule_timeout(timeout);
3386 set_current_state(TASK_RUNNING);
3387 remove_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3388 spin_lock_irqsave(lock, flags);
3391 spin_unlock_irqrestore(lock, flags);
3393 kiblnd_thread_fini();
3398 kiblnd_qp_event(struct ib_event *event, void *arg)
3400 kib_conn_t *conn = arg;
3402 switch (event->event) {
3403 case IB_EVENT_COMM_EST:
3404 CDEBUG(D_NET, "%s established\n",
3405 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3409 CERROR("%s: Async QP event type %d\n",
3410 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3416 kiblnd_complete (struct ib_wc *wc)
3418 switch (kiblnd_wreqid2type(wc->wr_id)) {
3423 if (wc->status != IB_WC_SUCCESS &&
3424 wc->status != IB_WC_WR_FLUSH_ERR)
3425 CNETERR("FastReg failed: %d\n", wc->status);
3428 case IBLND_WID_RDMA:
3429 /* We only get RDMA completion notification if it fails. All
3430 * subsequent work items, including the final SEND will fail
3431 * too. However we can't print out any more info about the
3432 * failing RDMA because 'tx' might be back on the idle list or
3433 * even reused already if we didn't manage to post all our work
3435 CNETERR("RDMA (tx: %p) failed: %d\n",
3436 kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3440 kiblnd_tx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3444 kiblnd_rx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status,
3451 kiblnd_cq_completion(struct ib_cq *cq, void *arg)
3453 /* NB I'm not allowed to schedule this conn once its refcount has
3454 * reached 0. Since fundamentally I'm racing with scheduler threads
3455 * consuming my CQ I could be called after all completions have
3456 * occurred. But in this case, ibc_nrx == 0 && ibc_nsends_posted == 0
3457 * and this CQ is about to be destroyed so I NOOP. */
3458 kib_conn_t *conn = (kib_conn_t *)arg;
3459 struct kib_sched_info *sched = conn->ibc_sched;
3460 unsigned long flags;
3462 LASSERT(cq == conn->ibc_cq);
3464 spin_lock_irqsave(&sched->ibs_lock, flags);
3466 conn->ibc_ready = 1;
3468 if (!conn->ibc_scheduled &&
3469 (conn->ibc_nrx > 0 ||
3470 conn->ibc_nsends_posted > 0)) {
3471 kiblnd_conn_addref(conn); /* +1 ref for sched_conns */
3472 conn->ibc_scheduled = 1;
3473 list_add_tail(&conn->ibc_sched_list, &sched->ibs_conns);
3475 if (waitqueue_active(&sched->ibs_waitq))
3476 wake_up(&sched->ibs_waitq);
3479 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3483 kiblnd_cq_event(struct ib_event *event, void *arg)
3485 kib_conn_t *conn = arg;
3487 CERROR("%s: async CQ event type %d\n",
3488 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3492 kiblnd_scheduler(void *arg)
3494 long id = (long)arg;
3495 struct kib_sched_info *sched;
3498 unsigned long flags;
3504 cfs_block_allsigs();
3506 init_waitqueue_entry(&wait, current);
3508 sched = kiblnd_data.kib_scheds[KIB_THREAD_CPT(id)];
3510 rc = cfs_cpt_bind(lnet_cpt_table(), sched->ibs_cpt);
3512 CWARN("Failed to bind on CPT %d, please verify whether "
3513 "all CPUs are healthy and reload modules if necessary, "
3514 "otherwise your system might under risk of low "
3515 "performance\n", sched->ibs_cpt);
3518 spin_lock_irqsave(&sched->ibs_lock, flags);
3520 while (!kiblnd_data.kib_shutdown) {
3521 if (busy_loops++ >= IBLND_RESCHED) {
3522 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3527 spin_lock_irqsave(&sched->ibs_lock, flags);
3532 if (!list_empty(&sched->ibs_conns)) {
3533 conn = list_entry(sched->ibs_conns.next,
3534 kib_conn_t, ibc_sched_list);
3535 /* take over kib_sched_conns' ref on conn... */
3536 LASSERT(conn->ibc_scheduled);
3537 list_del(&conn->ibc_sched_list);
3538 conn->ibc_ready = 0;
3540 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3542 wc.wr_id = IBLND_WID_INVAL;
3544 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3546 rc = ib_req_notify_cq(conn->ibc_cq,
3549 CWARN("%s: ib_req_notify_cq failed: %d, "
3550 "closing connection\n",
3551 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
3552 kiblnd_close_conn(conn, -EIO);
3553 kiblnd_conn_decref(conn);
3554 spin_lock_irqsave(&sched->ibs_lock,
3559 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3562 if (unlikely(rc > 0 && wc.wr_id == IBLND_WID_INVAL)) {
3564 "ib_poll_cq (rc: %d) returned invalid "
3565 "wr_id, opcode %d, status: %d, "
3566 "vendor_err: %d, conn: %s status: %d\n"
3567 "please upgrade firmware and OFED or "
3568 "contact vendor.\n", rc,
3569 wc.opcode, wc.status, wc.vendor_err,
3570 libcfs_nid2str(conn->ibc_peer->ibp_nid),
3576 CWARN("%s: ib_poll_cq failed: %d, "
3577 "closing connection\n",
3578 libcfs_nid2str(conn->ibc_peer->ibp_nid),
3580 kiblnd_close_conn(conn, -EIO);
3581 kiblnd_conn_decref(conn);
3582 spin_lock_irqsave(&sched->ibs_lock, flags);
3586 spin_lock_irqsave(&sched->ibs_lock, flags);
3588 if (rc != 0 || conn->ibc_ready) {
3589 /* There may be another completion waiting; get
3590 * another scheduler to check while I handle
3592 /* +1 ref for sched_conns */
3593 kiblnd_conn_addref(conn);
3594 list_add_tail(&conn->ibc_sched_list,
3596 if (waitqueue_active(&sched->ibs_waitq))
3597 wake_up(&sched->ibs_waitq);
3599 conn->ibc_scheduled = 0;
3603 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3604 kiblnd_complete(&wc);
3606 spin_lock_irqsave(&sched->ibs_lock, flags);
3609 kiblnd_conn_decref(conn); /* ...drop my ref from above */
3616 set_current_state(TASK_INTERRUPTIBLE);
3617 add_wait_queue_exclusive(&sched->ibs_waitq, &wait);
3618 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3623 remove_wait_queue(&sched->ibs_waitq, &wait);
3624 set_current_state(TASK_RUNNING);
3625 spin_lock_irqsave(&sched->ibs_lock, flags);
3628 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3630 kiblnd_thread_fini();
3635 kiblnd_failover_thread(void *arg)
3637 rwlock_t *glock = &kiblnd_data.kib_global_lock;
3640 unsigned long flags;
3643 LASSERT(*kiblnd_tunables.kib_dev_failover != 0);
3645 cfs_block_allsigs();
3647 init_waitqueue_entry(&wait, current);
3648 write_lock_irqsave(glock, flags);
3650 while (!kiblnd_data.kib_shutdown) {
3651 int do_failover = 0;
3654 list_for_each_entry(dev, &kiblnd_data.kib_failed_devs,
3656 if (cfs_time_before(cfs_time_current(),
3657 dev->ibd_next_failover))
3664 list_del_init(&dev->ibd_fail_list);
3665 dev->ibd_failover = 1;
3666 write_unlock_irqrestore(glock, flags);
3668 rc = kiblnd_dev_failover(dev);
3670 write_lock_irqsave(glock, flags);
3672 LASSERT (dev->ibd_failover);
3673 dev->ibd_failover = 0;
3674 if (rc >= 0) { /* Device is OK or failover succeed */
3675 dev->ibd_next_failover = cfs_time_shift(3);
3679 /* failed to failover, retry later */
3680 dev->ibd_next_failover =
3681 cfs_time_shift(min(dev->ibd_failed_failover, 10));
3682 if (kiblnd_dev_can_failover(dev)) {
3683 list_add_tail(&dev->ibd_fail_list,
3684 &kiblnd_data.kib_failed_devs);
3690 /* long sleep if no more pending failover */
3691 long_sleep = list_empty(&kiblnd_data.kib_failed_devs);
3693 set_current_state(TASK_INTERRUPTIBLE);
3694 add_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3695 write_unlock_irqrestore(glock, flags);
3697 rc = schedule_timeout(long_sleep ? cfs_time_seconds(10) :
3698 cfs_time_seconds(1));
3699 set_current_state(TASK_RUNNING);
3700 remove_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3701 write_lock_irqsave(glock, flags);
3703 if (!long_sleep || rc != 0)
3706 /* have a long sleep, routine check all active devices,
3707 * we need checking like this because if there is not active
3708 * connection on the dev and no SEND from local, we may listen
3709 * on wrong HCA for ever while there is a bonding failover */
3710 list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3711 if (kiblnd_dev_can_failover(dev)) {
3712 list_add_tail(&dev->ibd_fail_list,
3713 &kiblnd_data.kib_failed_devs);
3718 write_unlock_irqrestore(glock, flags);
3720 kiblnd_thread_fini();