4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lnet/klnds/o2iblnd/o2iblnd_cb.c
34 * Author: Eric Barton <eric@bartonsoftware.com>
39 #define MAX_CONN_RACES_BEFORE_ABORT 20
41 static void kiblnd_peer_alive(kib_peer_ni_t *peer_ni);
42 static void kiblnd_peer_connect_failed(kib_peer_ni_t *peer_ni, int active, int error);
43 static void kiblnd_init_tx_msg(struct lnet_ni *ni, kib_tx_t *tx,
44 int type, int body_nob);
45 static int kiblnd_init_rdma(kib_conn_t *conn, kib_tx_t *tx, int type,
46 int resid, kib_rdma_desc_t *dstrd, __u64 dstcookie);
47 static void kiblnd_queue_tx_locked(kib_tx_t *tx, kib_conn_t *conn);
48 static void kiblnd_queue_tx(kib_tx_t *tx, kib_conn_t *conn);
50 static void kiblnd_unmap_tx(kib_tx_t *tx);
51 static void kiblnd_check_sends_locked(kib_conn_t *conn);
54 kiblnd_tx_done(kib_tx_t *tx)
56 struct lnet_msg *lntmsg[2];
60 LASSERT (!in_interrupt());
61 LASSERT (!tx->tx_queued); /* mustn't be queued for sending */
62 LASSERT (tx->tx_sending == 0); /* mustn't be awaiting sent callback */
63 LASSERT (!tx->tx_waiting); /* mustn't be awaiting peer_ni response */
64 LASSERT (tx->tx_pool != NULL);
68 /* tx may have up to 2 lnet msgs to finalise */
69 lntmsg[0] = tx->tx_lntmsg[0]; tx->tx_lntmsg[0] = NULL;
70 lntmsg[1] = tx->tx_lntmsg[1]; tx->tx_lntmsg[1] = NULL;
73 if (tx->tx_conn != NULL) {
74 kiblnd_conn_decref(tx->tx_conn);
81 kiblnd_pool_free_node(&tx->tx_pool->tpo_pool, &tx->tx_list);
83 /* delay finalize until my descs have been freed */
84 for (i = 0; i < 2; i++) {
85 if (lntmsg[i] == NULL)
88 lnet_finalize(lntmsg[i], rc);
93 kiblnd_txlist_done(struct list_head *txlist, int status)
97 while (!list_empty(txlist)) {
98 tx = list_entry(txlist->next, kib_tx_t, tx_list);
100 list_del(&tx->tx_list);
103 tx->tx_status = status;
109 kiblnd_get_idle_tx(struct lnet_ni *ni, lnet_nid_t target)
111 kib_net_t *net = (kib_net_t *)ni->ni_data;
112 struct list_head *node;
114 kib_tx_poolset_t *tps;
116 tps = net->ibn_tx_ps[lnet_cpt_of_nid(target, ni)];
117 node = kiblnd_pool_alloc_node(&tps->tps_poolset);
120 tx = container_of(node, kib_tx_t, tx_list);
122 LASSERT (tx->tx_nwrq == 0);
123 LASSERT (!tx->tx_queued);
124 LASSERT (tx->tx_sending == 0);
125 LASSERT (!tx->tx_waiting);
126 LASSERT (tx->tx_status == 0);
127 LASSERT (tx->tx_conn == NULL);
128 LASSERT (tx->tx_lntmsg[0] == NULL);
129 LASSERT (tx->tx_lntmsg[1] == NULL);
130 LASSERT (tx->tx_nfrags == 0);
136 kiblnd_drop_rx(kib_rx_t *rx)
138 kib_conn_t *conn = rx->rx_conn;
139 struct kib_sched_info *sched = conn->ibc_sched;
142 spin_lock_irqsave(&sched->ibs_lock, flags);
143 LASSERT(conn->ibc_nrx > 0);
145 spin_unlock_irqrestore(&sched->ibs_lock, flags);
147 kiblnd_conn_decref(conn);
151 kiblnd_post_rx (kib_rx_t *rx, int credit)
153 kib_conn_t *conn = rx->rx_conn;
154 kib_net_t *net = conn->ibc_peer->ibp_ni->ni_data;
155 struct ib_recv_wr *bad_wrq = NULL;
156 struct ib_mr *mr = conn->ibc_hdev->ibh_mrs;
159 LASSERT (net != NULL);
160 LASSERT (!in_interrupt());
161 LASSERT (credit == IBLND_POSTRX_NO_CREDIT ||
162 credit == IBLND_POSTRX_PEER_CREDIT ||
163 credit == IBLND_POSTRX_RSRVD_CREDIT);
166 rx->rx_sge.lkey = mr->lkey;
167 rx->rx_sge.addr = rx->rx_msgaddr;
168 rx->rx_sge.length = IBLND_MSG_SIZE;
170 rx->rx_wrq.next = NULL;
171 rx->rx_wrq.sg_list = &rx->rx_sge;
172 rx->rx_wrq.num_sge = 1;
173 rx->rx_wrq.wr_id = kiblnd_ptr2wreqid(rx, IBLND_WID_RX);
175 LASSERT (conn->ibc_state >= IBLND_CONN_INIT);
176 LASSERT (rx->rx_nob >= 0); /* not posted */
178 if (conn->ibc_state > IBLND_CONN_ESTABLISHED) {
179 kiblnd_drop_rx(rx); /* No more posts for this rx */
183 rx->rx_nob = -1; /* flag posted */
185 /* NB: need an extra reference after ib_post_recv because we don't
186 * own this rx (and rx::rx_conn) anymore, LU-5678.
188 kiblnd_conn_addref(conn);
189 rc = ib_post_recv(conn->ibc_cmid->qp, &rx->rx_wrq, &bad_wrq);
190 if (unlikely(rc != 0)) {
191 CERROR("Can't post rx for %s: %d, bad_wrq: %p\n",
192 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc, bad_wrq);
196 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) /* Initial post */
199 if (unlikely(rc != 0)) {
200 kiblnd_close_conn(conn, rc);
201 kiblnd_drop_rx(rx); /* No more posts for this rx */
205 if (credit == IBLND_POSTRX_NO_CREDIT)
208 spin_lock(&conn->ibc_lock);
209 if (credit == IBLND_POSTRX_PEER_CREDIT)
210 conn->ibc_outstanding_credits++;
212 conn->ibc_reserved_credits++;
213 kiblnd_check_sends_locked(conn);
214 spin_unlock(&conn->ibc_lock);
217 kiblnd_conn_decref(conn);
222 kiblnd_find_waiting_tx_locked(kib_conn_t *conn, int txtype, __u64 cookie)
224 struct list_head *tmp;
226 list_for_each(tmp, &conn->ibc_active_txs) {
227 kib_tx_t *tx = list_entry(tmp, kib_tx_t, tx_list);
229 LASSERT(!tx->tx_queued);
230 LASSERT(tx->tx_sending != 0 || tx->tx_waiting);
232 if (tx->tx_cookie != cookie)
235 if (tx->tx_waiting &&
236 tx->tx_msg->ibm_type == txtype)
239 CWARN("Bad completion: %swaiting, type %x (wanted %x)\n",
240 tx->tx_waiting ? "" : "NOT ",
241 tx->tx_msg->ibm_type, txtype);
247 kiblnd_handle_completion(kib_conn_t *conn, int txtype, int status, __u64 cookie)
250 struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
253 spin_lock(&conn->ibc_lock);
255 tx = kiblnd_find_waiting_tx_locked(conn, txtype, cookie);
257 spin_unlock(&conn->ibc_lock);
259 CWARN("Unmatched completion type %x cookie %#llx from %s\n",
260 txtype, cookie, libcfs_nid2str(conn->ibc_peer->ibp_nid));
261 kiblnd_close_conn(conn, -EPROTO);
265 if (tx->tx_status == 0) { /* success so far */
266 if (status < 0) { /* failed? */
267 tx->tx_status = status;
268 } else if (txtype == IBLND_MSG_GET_REQ) {
269 lnet_set_reply_msg_len(ni, tx->tx_lntmsg[1], status);
275 idle = !tx->tx_queued && (tx->tx_sending == 0);
277 list_del(&tx->tx_list);
279 spin_unlock(&conn->ibc_lock);
286 kiblnd_send_completion(kib_conn_t *conn, int type, int status, __u64 cookie)
288 struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
289 kib_tx_t *tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
292 CERROR("Can't get tx for completion %x for %s\n",
293 type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
297 tx->tx_msg->ibm_u.completion.ibcm_status = status;
298 tx->tx_msg->ibm_u.completion.ibcm_cookie = cookie;
299 kiblnd_init_tx_msg(ni, tx, type, sizeof(kib_completion_msg_t));
301 kiblnd_queue_tx(tx, conn);
305 kiblnd_handle_rx (kib_rx_t *rx)
307 kib_msg_t *msg = rx->rx_msg;
308 kib_conn_t *conn = rx->rx_conn;
309 struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
310 int credits = msg->ibm_credits;
316 LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
318 CDEBUG (D_NET, "Received %x[%d] from %s\n",
319 msg->ibm_type, credits,
320 libcfs_nid2str(conn->ibc_peer->ibp_nid));
323 /* Have I received credits that will let me send? */
324 spin_lock(&conn->ibc_lock);
326 if (conn->ibc_credits + credits >
327 conn->ibc_queue_depth) {
328 rc2 = conn->ibc_credits;
329 spin_unlock(&conn->ibc_lock);
331 CERROR("Bad credits from %s: %d + %d > %d\n",
332 libcfs_nid2str(conn->ibc_peer->ibp_nid),
334 conn->ibc_queue_depth);
336 kiblnd_close_conn(conn, -EPROTO);
337 kiblnd_post_rx(rx, IBLND_POSTRX_NO_CREDIT);
341 conn->ibc_credits += credits;
343 /* This ensures the credit taken by NOOP can be returned */
344 if (msg->ibm_type == IBLND_MSG_NOOP &&
345 !IBLND_OOB_CAPABLE(conn->ibc_version)) /* v1 only */
346 conn->ibc_outstanding_credits++;
348 kiblnd_check_sends_locked(conn);
349 spin_unlock(&conn->ibc_lock);
352 switch (msg->ibm_type) {
354 CERROR("Bad IBLND message type %x from %s\n",
355 msg->ibm_type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
356 post_credit = IBLND_POSTRX_NO_CREDIT;
361 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
362 post_credit = IBLND_POSTRX_NO_CREDIT;
366 if (credits != 0) /* credit already posted */
367 post_credit = IBLND_POSTRX_NO_CREDIT;
368 else /* a keepalive NOOP */
369 post_credit = IBLND_POSTRX_PEER_CREDIT;
372 case IBLND_MSG_IMMEDIATE:
373 post_credit = IBLND_POSTRX_DONT_POST;
374 rc = lnet_parse(ni, &msg->ibm_u.immediate.ibim_hdr,
375 msg->ibm_srcnid, rx, 0);
376 if (rc < 0) /* repost on error */
377 post_credit = IBLND_POSTRX_PEER_CREDIT;
380 case IBLND_MSG_PUT_REQ:
381 post_credit = IBLND_POSTRX_DONT_POST;
382 rc = lnet_parse(ni, &msg->ibm_u.putreq.ibprm_hdr,
383 msg->ibm_srcnid, rx, 1);
384 if (rc < 0) /* repost on error */
385 post_credit = IBLND_POSTRX_PEER_CREDIT;
388 case IBLND_MSG_PUT_NAK:
389 CWARN ("PUT_NACK from %s\n",
390 libcfs_nid2str(conn->ibc_peer->ibp_nid));
391 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
392 kiblnd_handle_completion(conn, IBLND_MSG_PUT_REQ,
393 msg->ibm_u.completion.ibcm_status,
394 msg->ibm_u.completion.ibcm_cookie);
397 case IBLND_MSG_PUT_ACK:
398 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
400 spin_lock(&conn->ibc_lock);
401 tx = kiblnd_find_waiting_tx_locked(conn, IBLND_MSG_PUT_REQ,
402 msg->ibm_u.putack.ibpam_src_cookie);
404 list_del(&tx->tx_list);
405 spin_unlock(&conn->ibc_lock);
408 CERROR("Unmatched PUT_ACK from %s\n",
409 libcfs_nid2str(conn->ibc_peer->ibp_nid));
414 LASSERT (tx->tx_waiting);
415 /* CAVEAT EMPTOR: I could be racing with tx_complete, but...
416 * (a) I can overwrite tx_msg since my peer_ni has received it!
417 * (b) tx_waiting set tells tx_complete() it's not done. */
419 tx->tx_nwrq = 0; /* overwrite PUT_REQ */
421 rc2 = kiblnd_init_rdma(conn, tx, IBLND_MSG_PUT_DONE,
422 kiblnd_rd_size(&msg->ibm_u.putack.ibpam_rd),
423 &msg->ibm_u.putack.ibpam_rd,
424 msg->ibm_u.putack.ibpam_dst_cookie);
426 CERROR("Can't setup rdma for PUT to %s: %d\n",
427 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc2);
429 spin_lock(&conn->ibc_lock);
430 tx->tx_waiting = 0; /* clear waiting and queue atomically */
431 kiblnd_queue_tx_locked(tx, conn);
432 spin_unlock(&conn->ibc_lock);
435 case IBLND_MSG_PUT_DONE:
436 post_credit = IBLND_POSTRX_PEER_CREDIT;
437 kiblnd_handle_completion(conn, IBLND_MSG_PUT_ACK,
438 msg->ibm_u.completion.ibcm_status,
439 msg->ibm_u.completion.ibcm_cookie);
442 case IBLND_MSG_GET_REQ:
443 post_credit = IBLND_POSTRX_DONT_POST;
444 rc = lnet_parse(ni, &msg->ibm_u.get.ibgm_hdr,
445 msg->ibm_srcnid, rx, 1);
446 if (rc < 0) /* repost on error */
447 post_credit = IBLND_POSTRX_PEER_CREDIT;
450 case IBLND_MSG_GET_DONE:
451 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
452 kiblnd_handle_completion(conn, IBLND_MSG_GET_REQ,
453 msg->ibm_u.completion.ibcm_status,
454 msg->ibm_u.completion.ibcm_cookie);
458 if (rc < 0) /* protocol error */
459 kiblnd_close_conn(conn, rc);
461 if (post_credit != IBLND_POSTRX_DONT_POST)
462 kiblnd_post_rx(rx, post_credit);
466 kiblnd_rx_complete (kib_rx_t *rx, int status, int nob)
468 kib_msg_t *msg = rx->rx_msg;
469 kib_conn_t *conn = rx->rx_conn;
470 struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
471 kib_net_t *net = ni->ni_data;
475 LASSERT (net != NULL);
476 LASSERT (rx->rx_nob < 0); /* was posted */
477 rx->rx_nob = 0; /* isn't now */
479 if (conn->ibc_state > IBLND_CONN_ESTABLISHED)
482 if (status != IB_WC_SUCCESS) {
483 CNETERR("Rx from %s failed: %d\n",
484 libcfs_nid2str(conn->ibc_peer->ibp_nid), status);
491 rc = kiblnd_unpack_msg(msg, rx->rx_nob);
493 CERROR ("Error %d unpacking rx from %s\n",
494 rc, libcfs_nid2str(conn->ibc_peer->ibp_nid));
498 if (msg->ibm_srcnid != conn->ibc_peer->ibp_nid ||
499 msg->ibm_dstnid != ni->ni_nid ||
500 msg->ibm_srcstamp != conn->ibc_incarnation ||
501 msg->ibm_dststamp != net->ibn_incarnation) {
502 CERROR ("Stale rx from %s\n",
503 libcfs_nid2str(conn->ibc_peer->ibp_nid));
508 /* set time last known alive */
509 kiblnd_peer_alive(conn->ibc_peer);
511 /* racing with connection establishment/teardown! */
513 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
514 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
517 write_lock_irqsave(g_lock, flags);
518 /* must check holding global lock to eliminate race */
519 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
520 list_add_tail(&rx->rx_list, &conn->ibc_early_rxs);
521 write_unlock_irqrestore(g_lock, flags);
524 write_unlock_irqrestore(g_lock, flags);
526 kiblnd_handle_rx(rx);
530 CDEBUG(D_NET, "rx %p conn %p\n", rx, conn);
531 kiblnd_close_conn(conn, err);
533 kiblnd_drop_rx(rx); /* Don't re-post rx. */
537 kiblnd_kvaddr_to_page (unsigned long vaddr)
541 if (is_vmalloc_addr((void *)vaddr)) {
542 page = vmalloc_to_page ((void *)vaddr);
543 LASSERT (page != NULL);
546 #ifdef CONFIG_HIGHMEM
547 if (vaddr >= PKMAP_BASE &&
548 vaddr < (PKMAP_BASE + LAST_PKMAP * PAGE_SIZE)) {
549 /* No highmem pages only used for bulk (kiov) I/O */
550 CERROR("find page for address in highmem\n");
554 page = virt_to_page (vaddr);
555 LASSERT (page != NULL);
560 kiblnd_fmr_map_tx(kib_net_t *net, kib_tx_t *tx, kib_rdma_desc_t *rd, __u32 nob)
563 kib_fmr_poolset_t *fps;
567 LASSERT(tx->tx_pool != NULL);
568 LASSERT(tx->tx_pool->tpo_pool.po_owner != NULL);
570 hdev = tx->tx_pool->tpo_hdev;
571 cpt = tx->tx_pool->tpo_pool.po_owner->ps_cpt;
573 fps = net->ibn_fmr_ps[cpt];
574 rc = kiblnd_fmr_pool_map(fps, tx, rd, nob, 0, &tx->fmr);
576 CERROR("Can't map %u pages: %d\n", nob, rc);
580 /* If rd is not tx_rd, it's going to get sent to a peer_ni, who will need
582 rd->rd_key = tx->fmr.fmr_key;
583 rd->rd_frags[0].rf_addr &= ~hdev->ibh_page_mask;
584 rd->rd_frags[0].rf_nob = nob;
591 kiblnd_unmap_tx(kib_tx_t *tx)
593 if (tx->fmr.fmr_pfmr != NULL)
594 kiblnd_fmr_pool_unmap(&tx->fmr, tx->tx_status);
596 if (tx->tx_nfrags != 0) {
597 kiblnd_dma_unmap_sg(tx->tx_pool->tpo_hdev->ibh_ibdev,
598 tx->tx_frags, tx->tx_nfrags, tx->tx_dmadir);
604 kiblnd_map_tx(struct lnet_ni *ni, kib_tx_t *tx, kib_rdma_desc_t *rd, int nfrags)
606 kib_net_t *net = ni->ni_data;
607 kib_hca_dev_t *hdev = net->ibn_dev->ibd_hdev;
608 struct ib_mr *mr = NULL;
612 /* If rd is not tx_rd, it's going to get sent to a peer_ni and I'm the
614 tx->tx_dmadir = (rd != tx->tx_rd) ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
615 tx->tx_nfrags = nfrags;
617 rd->rd_nfrags = kiblnd_dma_map_sg(hdev->ibh_ibdev, tx->tx_frags,
618 tx->tx_nfrags, tx->tx_dmadir);
620 for (i = 0, nob = 0; i < rd->rd_nfrags; i++) {
621 rd->rd_frags[i].rf_nob = kiblnd_sg_dma_len(
622 hdev->ibh_ibdev, &tx->tx_frags[i]);
623 rd->rd_frags[i].rf_addr = kiblnd_sg_dma_address(
624 hdev->ibh_ibdev, &tx->tx_frags[i]);
625 nob += rd->rd_frags[i].rf_nob;
628 mr = kiblnd_find_rd_dma_mr(ni, rd,
629 (tx->tx_conn != NULL) ?
630 tx->tx_conn->ibc_max_frags : -1);
632 /* found pre-mapping MR */
633 rd->rd_key = (rd != tx->tx_rd) ? mr->rkey : mr->lkey;
637 if (net->ibn_fmr_ps != NULL)
638 return kiblnd_fmr_map_tx(net, tx, rd, nob);
645 kiblnd_setup_rd_iov(struct lnet_ni *ni, kib_tx_t *tx, kib_rdma_desc_t *rd,
646 unsigned int niov, struct kvec *iov, int offset, int nob)
648 kib_net_t *net = ni->ni_data;
650 struct scatterlist *sg;
657 LASSERT (net != NULL);
659 while (offset >= iov->iov_len) {
660 offset -= iov->iov_len;
670 vaddr = ((unsigned long)iov->iov_base) + offset;
671 page_offset = vaddr & (PAGE_SIZE - 1);
672 page = kiblnd_kvaddr_to_page(vaddr);
674 CERROR ("Can't find page\n");
678 fragnob = min((int)(iov->iov_len - offset), nob);
679 fragnob = min(fragnob, (int)PAGE_SIZE - page_offset);
681 sg_set_page(sg, page, fragnob, page_offset);
684 CERROR("lacking enough sg entries to map tx\n");
688 if (offset + fragnob < iov->iov_len) {
698 return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags);
702 kiblnd_setup_rd_kiov(struct lnet_ni *ni, kib_tx_t *tx, kib_rdma_desc_t *rd,
703 int nkiov, lnet_kiov_t *kiov, int offset, int nob)
705 kib_net_t *net = ni->ni_data;
706 struct scatterlist *sg;
709 CDEBUG(D_NET, "niov %d offset %d nob %d\n", nkiov, offset, nob);
713 LASSERT (net != NULL);
715 while (offset >= kiov->kiov_len) {
716 offset -= kiov->kiov_len;
726 fragnob = min((int)(kiov->kiov_len - offset), nob);
728 sg_set_page(sg, kiov->kiov_page, fragnob,
729 kiov->kiov_offset + offset);
732 CERROR("lacking enough sg entries to map tx\n");
742 return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags);
746 kiblnd_post_tx_locked (kib_conn_t *conn, kib_tx_t *tx, int credit)
747 __must_hold(&conn->ibc_lock)
749 kib_msg_t *msg = tx->tx_msg;
750 kib_peer_ni_t *peer_ni = conn->ibc_peer;
751 struct lnet_ni *ni = peer_ni->ibp_ni;
752 int ver = conn->ibc_version;
756 LASSERT(tx->tx_queued);
757 /* We rely on this for QP sizing */
758 LASSERT(tx->tx_nwrq > 0);
759 LASSERT(tx->tx_nwrq <= 1 + conn->ibc_max_frags);
761 LASSERT(credit == 0 || credit == 1);
762 LASSERT(conn->ibc_outstanding_credits >= 0);
763 LASSERT(conn->ibc_outstanding_credits <= conn->ibc_queue_depth);
764 LASSERT(conn->ibc_credits >= 0);
765 LASSERT(conn->ibc_credits <= conn->ibc_queue_depth);
767 if (conn->ibc_nsends_posted ==
768 kiblnd_concurrent_sends(ver, ni)) {
769 /* tx completions outstanding... */
770 CDEBUG(D_NET, "%s: posted enough\n",
771 libcfs_nid2str(peer_ni->ibp_nid));
775 if (credit != 0 && conn->ibc_credits == 0) { /* no credits */
776 CDEBUG(D_NET, "%s: no credits\n",
777 libcfs_nid2str(peer_ni->ibp_nid));
781 if (credit != 0 && !IBLND_OOB_CAPABLE(ver) &&
782 conn->ibc_credits == 1 && /* last credit reserved */
783 msg->ibm_type != IBLND_MSG_NOOP) { /* for NOOP */
784 CDEBUG(D_NET, "%s: not using last credit\n",
785 libcfs_nid2str(peer_ni->ibp_nid));
789 /* NB don't drop ibc_lock before bumping tx_sending */
790 list_del(&tx->tx_list);
793 if (msg->ibm_type == IBLND_MSG_NOOP &&
794 (!kiblnd_need_noop(conn) || /* redundant NOOP */
795 (IBLND_OOB_CAPABLE(ver) && /* posted enough NOOP */
796 conn->ibc_noops_posted == IBLND_OOB_MSGS(ver)))) {
797 /* OK to drop when posted enough NOOPs, since
798 * kiblnd_check_sends_locked will queue NOOP again when
799 * posted NOOPs complete */
800 spin_unlock(&conn->ibc_lock);
802 spin_lock(&conn->ibc_lock);
803 CDEBUG(D_NET, "%s(%d): redundant or enough NOOP\n",
804 libcfs_nid2str(peer_ni->ibp_nid),
805 conn->ibc_noops_posted);
809 kiblnd_pack_msg(peer_ni->ibp_ni, msg, ver, conn->ibc_outstanding_credits,
810 peer_ni->ibp_nid, conn->ibc_incarnation);
812 conn->ibc_credits -= credit;
813 conn->ibc_outstanding_credits = 0;
814 conn->ibc_nsends_posted++;
815 if (msg->ibm_type == IBLND_MSG_NOOP)
816 conn->ibc_noops_posted++;
818 /* CAVEAT EMPTOR! This tx could be the PUT_DONE of an RDMA
819 * PUT. If so, it was first queued here as a PUT_REQ, sent and
820 * stashed on ibc_active_txs, matched by an incoming PUT_ACK,
821 * and then re-queued here. It's (just) possible that
822 * tx_sending is non-zero if we've not done the tx_complete()
823 * from the first send; hence the ++ rather than = below. */
825 list_add(&tx->tx_list, &conn->ibc_active_txs);
827 /* I'm still holding ibc_lock! */
828 if (conn->ibc_state != IBLND_CONN_ESTABLISHED) {
830 } else if (tx->tx_pool->tpo_pool.po_failed ||
831 conn->ibc_hdev != tx->tx_pool->tpo_hdev) {
832 /* close_conn will launch failover */
835 struct kib_fast_reg_descriptor *frd = tx->fmr.fmr_frd;
836 struct ib_send_wr *bad = &tx->tx_wrq[tx->tx_nwrq - 1].wr;
837 struct ib_send_wr *wr = &tx->tx_wrq[0].wr;
840 if (!frd->frd_valid) {
841 wr = &frd->frd_inv_wr.wr;
842 wr->next = &frd->frd_fastreg_wr.wr;
844 wr = &frd->frd_fastreg_wr.wr;
846 frd->frd_fastreg_wr.wr.next = &tx->tx_wrq[0].wr;
849 LASSERTF(bad->wr_id == kiblnd_ptr2wreqid(tx, IBLND_WID_TX),
850 "bad wr_id %#llx, opc %d, flags %d, peer_ni: %s\n",
851 bad->wr_id, bad->opcode, bad->send_flags,
852 libcfs_nid2str(conn->ibc_peer->ibp_nid));
855 rc = ib_post_send(conn->ibc_cmid->qp, wr, &bad);
858 conn->ibc_last_send = jiffies;
863 /* NB credits are transferred in the actual
864 * message, which can only be the last work item */
865 conn->ibc_credits += credit;
866 conn->ibc_outstanding_credits += msg->ibm_credits;
867 conn->ibc_nsends_posted--;
868 if (msg->ibm_type == IBLND_MSG_NOOP)
869 conn->ibc_noops_posted--;
875 done = (tx->tx_sending == 0);
877 list_del(&tx->tx_list);
879 spin_unlock(&conn->ibc_lock);
881 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
882 CERROR("Error %d posting transmit to %s\n",
883 rc, libcfs_nid2str(peer_ni->ibp_nid));
885 CDEBUG(D_NET, "Error %d posting transmit to %s\n",
886 rc, libcfs_nid2str(peer_ni->ibp_nid));
888 kiblnd_close_conn(conn, rc);
893 spin_lock(&conn->ibc_lock);
899 kiblnd_check_sends_locked(kib_conn_t *conn)
901 int ver = conn->ibc_version;
902 struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
905 /* Don't send anything until after the connection is established */
906 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
907 CDEBUG(D_NET, "%s too soon\n",
908 libcfs_nid2str(conn->ibc_peer->ibp_nid));
912 LASSERT(conn->ibc_nsends_posted <=
913 kiblnd_concurrent_sends(ver, ni));
914 LASSERT (!IBLND_OOB_CAPABLE(ver) ||
915 conn->ibc_noops_posted <= IBLND_OOB_MSGS(ver));
916 LASSERT (conn->ibc_reserved_credits >= 0);
918 while (conn->ibc_reserved_credits > 0 &&
919 !list_empty(&conn->ibc_tx_queue_rsrvd)) {
920 tx = list_entry(conn->ibc_tx_queue_rsrvd.next,
922 list_del(&tx->tx_list);
923 list_add_tail(&tx->tx_list, &conn->ibc_tx_queue);
924 conn->ibc_reserved_credits--;
927 if (kiblnd_need_noop(conn)) {
928 spin_unlock(&conn->ibc_lock);
930 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
932 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_NOOP, 0);
934 spin_lock(&conn->ibc_lock);
936 kiblnd_queue_tx_locked(tx, conn);
942 if (!list_empty(&conn->ibc_tx_queue_nocred)) {
944 tx = list_entry(conn->ibc_tx_queue_nocred.next,
946 } else if (!list_empty(&conn->ibc_tx_noops)) {
947 LASSERT (!IBLND_OOB_CAPABLE(ver));
949 tx = list_entry(conn->ibc_tx_noops.next,
951 } else if (!list_empty(&conn->ibc_tx_queue)) {
953 tx = list_entry(conn->ibc_tx_queue.next,
958 if (kiblnd_post_tx_locked(conn, tx, credit) != 0)
964 kiblnd_tx_complete (kib_tx_t *tx, int status)
966 int failed = (status != IB_WC_SUCCESS);
967 kib_conn_t *conn = tx->tx_conn;
970 LASSERT (tx->tx_sending > 0);
973 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
974 CNETERR("Tx -> %s cookie %#llx"
975 " sending %d waiting %d: failed %d\n",
976 libcfs_nid2str(conn->ibc_peer->ibp_nid),
977 tx->tx_cookie, tx->tx_sending, tx->tx_waiting,
980 kiblnd_close_conn(conn, -EIO);
982 kiblnd_peer_alive(conn->ibc_peer);
985 spin_lock(&conn->ibc_lock);
987 /* I could be racing with rdma completion. Whoever makes 'tx' idle
988 * gets to free it, which also drops its ref on 'conn'. */
991 conn->ibc_nsends_posted--;
992 if (tx->tx_msg->ibm_type == IBLND_MSG_NOOP)
993 conn->ibc_noops_posted--;
996 tx->tx_waiting = 0; /* don't wait for peer_ni */
997 tx->tx_status = -EIO;
1000 idle = (tx->tx_sending == 0) && /* This is the final callback */
1001 !tx->tx_waiting && /* Not waiting for peer_ni */
1002 !tx->tx_queued; /* Not re-queued (PUT_DONE) */
1004 list_del(&tx->tx_list);
1006 kiblnd_check_sends_locked(conn);
1007 spin_unlock(&conn->ibc_lock);
1014 kiblnd_init_tx_msg(struct lnet_ni *ni, kib_tx_t *tx, int type, int body_nob)
1016 kib_hca_dev_t *hdev = tx->tx_pool->tpo_hdev;
1017 struct ib_sge *sge = &tx->tx_sge[tx->tx_nwrq];
1018 struct ib_rdma_wr *wrq;
1019 int nob = offsetof(kib_msg_t, ibm_u) + body_nob;
1020 struct ib_mr *mr = hdev->ibh_mrs;
1022 LASSERT(tx->tx_nwrq >= 0);
1023 LASSERT(tx->tx_nwrq < IBLND_MAX_RDMA_FRAGS + 1);
1024 LASSERT(nob <= IBLND_MSG_SIZE);
1025 LASSERT(mr != NULL);
1027 kiblnd_init_msg(tx->tx_msg, type, body_nob);
1029 sge->lkey = mr->lkey;
1030 sge->addr = tx->tx_msgaddr;
1033 wrq = &tx->tx_wrq[tx->tx_nwrq];
1034 memset(wrq, 0, sizeof(*wrq));
1036 wrq->wr.next = NULL;
1037 wrq->wr.wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_TX);
1038 wrq->wr.sg_list = sge;
1039 wrq->wr.num_sge = 1;
1040 wrq->wr.opcode = IB_WR_SEND;
1041 wrq->wr.send_flags = IB_SEND_SIGNALED;
1047 kiblnd_init_rdma(kib_conn_t *conn, kib_tx_t *tx, int type,
1048 int resid, kib_rdma_desc_t *dstrd, __u64 dstcookie)
1050 kib_msg_t *ibmsg = tx->tx_msg;
1051 kib_rdma_desc_t *srcrd = tx->tx_rd;
1052 struct ib_sge *sge = &tx->tx_sge[0];
1053 struct ib_rdma_wr *wrq;
1059 LASSERT (!in_interrupt());
1060 LASSERT (tx->tx_nwrq == 0);
1061 LASSERT (type == IBLND_MSG_GET_DONE ||
1062 type == IBLND_MSG_PUT_DONE);
1064 srcidx = dstidx = 0;
1067 if (srcidx >= srcrd->rd_nfrags) {
1068 CERROR("Src buffer exhausted: %d frags\n", srcidx);
1073 if (dstidx == dstrd->rd_nfrags) {
1074 CERROR("Dst buffer exhausted: %d frags\n", dstidx);
1079 if (tx->tx_nwrq >= conn->ibc_max_frags) {
1080 CERROR("RDMA has too many fragments for peer_ni %s (%d), "
1081 "src idx/frags: %d/%d dst idx/frags: %d/%d\n",
1082 libcfs_nid2str(conn->ibc_peer->ibp_nid),
1083 conn->ibc_max_frags,
1084 srcidx, srcrd->rd_nfrags,
1085 dstidx, dstrd->rd_nfrags);
1090 wrknob = MIN(MIN(kiblnd_rd_frag_size(srcrd, srcidx),
1091 kiblnd_rd_frag_size(dstrd, dstidx)), resid);
1093 sge = &tx->tx_sge[tx->tx_nwrq];
1094 sge->addr = kiblnd_rd_frag_addr(srcrd, srcidx);
1095 sge->lkey = kiblnd_rd_frag_key(srcrd, srcidx);
1096 sge->length = wrknob;
1098 wrq = &tx->tx_wrq[tx->tx_nwrq];
1100 wrq->wr.next = &(wrq + 1)->wr;
1101 wrq->wr.wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_RDMA);
1102 wrq->wr.sg_list = sge;
1103 wrq->wr.num_sge = 1;
1104 wrq->wr.opcode = IB_WR_RDMA_WRITE;
1105 wrq->wr.send_flags = 0;
1107 #ifdef HAVE_IB_RDMA_WR
1108 wrq->remote_addr = kiblnd_rd_frag_addr(dstrd, dstidx);
1109 wrq->rkey = kiblnd_rd_frag_key(dstrd, dstidx);
1111 wrq->wr.wr.rdma.remote_addr = kiblnd_rd_frag_addr(dstrd, dstidx);
1112 wrq->wr.wr.rdma.rkey = kiblnd_rd_frag_key(dstrd, dstidx);
1115 srcidx = kiblnd_rd_consume_frag(srcrd, srcidx, wrknob);
1116 dstidx = kiblnd_rd_consume_frag(dstrd, dstidx, wrknob);
1125 if (rc < 0) /* no RDMA if completing with failure */
1128 ibmsg->ibm_u.completion.ibcm_status = rc;
1129 ibmsg->ibm_u.completion.ibcm_cookie = dstcookie;
1130 kiblnd_init_tx_msg(conn->ibc_peer->ibp_ni, tx,
1131 type, sizeof (kib_completion_msg_t));
1137 kiblnd_queue_tx_locked(kib_tx_t *tx, kib_conn_t *conn)
1139 struct list_head *q;
1141 LASSERT(tx->tx_nwrq > 0); /* work items set up */
1142 LASSERT(!tx->tx_queued); /* not queued for sending already */
1143 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1146 tx->tx_deadline = jiffies +
1147 msecs_to_jiffies(*kiblnd_tunables.kib_timeout *
1150 if (tx->tx_conn == NULL) {
1151 kiblnd_conn_addref(conn);
1153 LASSERT (tx->tx_msg->ibm_type != IBLND_MSG_PUT_DONE);
1155 /* PUT_DONE first attached to conn as a PUT_REQ */
1156 LASSERT (tx->tx_conn == conn);
1157 LASSERT (tx->tx_msg->ibm_type == IBLND_MSG_PUT_DONE);
1160 switch (tx->tx_msg->ibm_type) {
1164 case IBLND_MSG_PUT_REQ:
1165 case IBLND_MSG_GET_REQ:
1166 q = &conn->ibc_tx_queue_rsrvd;
1169 case IBLND_MSG_PUT_NAK:
1170 case IBLND_MSG_PUT_ACK:
1171 case IBLND_MSG_PUT_DONE:
1172 case IBLND_MSG_GET_DONE:
1173 q = &conn->ibc_tx_queue_nocred;
1176 case IBLND_MSG_NOOP:
1177 if (IBLND_OOB_CAPABLE(conn->ibc_version))
1178 q = &conn->ibc_tx_queue_nocred;
1180 q = &conn->ibc_tx_noops;
1183 case IBLND_MSG_IMMEDIATE:
1184 q = &conn->ibc_tx_queue;
1188 list_add_tail(&tx->tx_list, q);
1192 kiblnd_queue_tx (kib_tx_t *tx, kib_conn_t *conn)
1194 spin_lock(&conn->ibc_lock);
1195 kiblnd_queue_tx_locked(tx, conn);
1196 kiblnd_check_sends_locked(conn);
1197 spin_unlock(&conn->ibc_lock);
1200 static int kiblnd_resolve_addr(struct rdma_cm_id *cmid,
1201 struct sockaddr_in *srcaddr,
1202 struct sockaddr_in *dstaddr,
1205 unsigned short port;
1208 /* allow the port to be reused */
1209 rc = rdma_set_reuseaddr(cmid, 1);
1211 CERROR("Unable to set reuse on cmid: %d\n", rc);
1215 /* look for a free privileged port */
1216 for (port = PROT_SOCK-1; port > 0; port--) {
1217 srcaddr->sin_port = htons(port);
1218 rc = rdma_resolve_addr(cmid,
1219 (struct sockaddr *)srcaddr,
1220 (struct sockaddr *)dstaddr,
1223 CDEBUG(D_NET, "bound to port %hu\n", port);
1225 } else if (rc == -EADDRINUSE || rc == -EADDRNOTAVAIL) {
1226 CDEBUG(D_NET, "bind to port %hu failed: %d\n",
1233 CERROR("Failed to bind to a free privileged port\n");
1238 kiblnd_connect_peer (kib_peer_ni_t *peer_ni)
1240 struct rdma_cm_id *cmid;
1242 kib_net_t *net = peer_ni->ibp_ni->ni_data;
1243 struct sockaddr_in srcaddr;
1244 struct sockaddr_in dstaddr;
1247 LASSERT (net != NULL);
1248 LASSERT (peer_ni->ibp_connecting > 0);
1249 LASSERT(!peer_ni->ibp_reconnecting);
1251 cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, peer_ni, RDMA_PS_TCP,
1255 CERROR("Can't create CMID for %s: %ld\n",
1256 libcfs_nid2str(peer_ni->ibp_nid), PTR_ERR(cmid));
1262 memset(&srcaddr, 0, sizeof(srcaddr));
1263 srcaddr.sin_family = AF_INET;
1264 srcaddr.sin_addr.s_addr = htonl(dev->ibd_ifip);
1266 memset(&dstaddr, 0, sizeof(dstaddr));
1267 dstaddr.sin_family = AF_INET;
1268 dstaddr.sin_port = htons(*kiblnd_tunables.kib_service);
1269 dstaddr.sin_addr.s_addr = htonl(LNET_NIDADDR(peer_ni->ibp_nid));
1271 kiblnd_peer_addref(peer_ni); /* cmid's ref */
1273 if (*kiblnd_tunables.kib_use_priv_port) {
1274 rc = kiblnd_resolve_addr(cmid, &srcaddr, &dstaddr,
1275 *kiblnd_tunables.kib_timeout * 1000);
1277 rc = rdma_resolve_addr(cmid,
1278 (struct sockaddr *)&srcaddr,
1279 (struct sockaddr *)&dstaddr,
1280 *kiblnd_tunables.kib_timeout * 1000);
1283 /* Can't initiate address resolution: */
1284 CERROR("Can't resolve addr for %s: %d\n",
1285 libcfs_nid2str(peer_ni->ibp_nid), rc);
1289 LASSERT (cmid->device != NULL);
1290 CDEBUG(D_NET, "%s: connection bound to %s:%pI4h:%s\n",
1291 libcfs_nid2str(peer_ni->ibp_nid), dev->ibd_ifname,
1292 &dev->ibd_ifip, cmid->device->name);
1297 kiblnd_peer_connect_failed(peer_ni, 1, rc);
1298 kiblnd_peer_decref(peer_ni); /* cmid's ref */
1299 rdma_destroy_id(cmid);
1302 kiblnd_peer_connect_failed(peer_ni, 1, rc);
1306 kiblnd_reconnect_peer(kib_peer_ni_t *peer_ni)
1308 rwlock_t *glock = &kiblnd_data.kib_global_lock;
1309 char *reason = NULL;
1310 struct list_head txs;
1311 unsigned long flags;
1313 INIT_LIST_HEAD(&txs);
1315 write_lock_irqsave(glock, flags);
1316 if (peer_ni->ibp_reconnecting == 0) {
1317 if (peer_ni->ibp_accepting)
1318 reason = "accepting";
1319 else if (peer_ni->ibp_connecting)
1320 reason = "connecting";
1321 else if (!list_empty(&peer_ni->ibp_conns))
1322 reason = "connected";
1323 else /* connected then closed */
1329 LASSERT(!peer_ni->ibp_accepting && !peer_ni->ibp_connecting &&
1330 list_empty(&peer_ni->ibp_conns));
1331 peer_ni->ibp_reconnecting = 0;
1333 if (!kiblnd_peer_active(peer_ni)) {
1334 list_splice_init(&peer_ni->ibp_tx_queue, &txs);
1335 reason = "unlinked";
1339 peer_ni->ibp_connecting++;
1340 peer_ni->ibp_reconnected++;
1342 write_unlock_irqrestore(glock, flags);
1344 kiblnd_connect_peer(peer_ni);
1348 write_unlock_irqrestore(glock, flags);
1350 CWARN("Abort reconnection of %s: %s\n",
1351 libcfs_nid2str(peer_ni->ibp_nid), reason);
1352 kiblnd_txlist_done(&txs, -ECONNABORTED);
1357 kiblnd_launch_tx(struct lnet_ni *ni, kib_tx_t *tx, lnet_nid_t nid)
1359 kib_peer_ni_t *peer_ni;
1360 kib_peer_ni_t *peer2;
1362 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
1363 unsigned long flags;
1366 /* If I get here, I've committed to send, so I complete the tx with
1367 * failure on any problems */
1369 LASSERT (tx == NULL || tx->tx_conn == NULL); /* only set when assigned a conn */
1370 LASSERT (tx == NULL || tx->tx_nwrq > 0); /* work items have been set up */
1372 /* First time, just use a read lock since I expect to find my peer_ni
1374 read_lock_irqsave(g_lock, flags);
1376 peer_ni = kiblnd_find_peer_locked(ni, nid);
1377 if (peer_ni != NULL && !list_empty(&peer_ni->ibp_conns)) {
1378 /* Found a peer_ni with an established connection */
1379 conn = kiblnd_get_conn_locked(peer_ni);
1380 kiblnd_conn_addref(conn); /* 1 ref for me... */
1382 read_unlock_irqrestore(g_lock, flags);
1385 kiblnd_queue_tx(tx, conn);
1386 kiblnd_conn_decref(conn); /* ...to here */
1390 read_unlock(g_lock);
1391 /* Re-try with a write lock */
1394 peer_ni = kiblnd_find_peer_locked(ni, nid);
1395 if (peer_ni != NULL) {
1396 if (list_empty(&peer_ni->ibp_conns)) {
1397 /* found a peer_ni, but it's still connecting... */
1398 LASSERT(kiblnd_peer_connecting(peer_ni));
1400 list_add_tail(&tx->tx_list,
1401 &peer_ni->ibp_tx_queue);
1402 write_unlock_irqrestore(g_lock, flags);
1404 conn = kiblnd_get_conn_locked(peer_ni);
1405 kiblnd_conn_addref(conn); /* 1 ref for me... */
1407 write_unlock_irqrestore(g_lock, flags);
1410 kiblnd_queue_tx(tx, conn);
1411 kiblnd_conn_decref(conn); /* ...to here */
1416 write_unlock_irqrestore(g_lock, flags);
1418 /* Allocate a peer_ni ready to add to the peer_ni table and retry */
1419 rc = kiblnd_create_peer(ni, &peer_ni, nid);
1421 CERROR("Can't create peer_ni %s\n", libcfs_nid2str(nid));
1423 tx->tx_status = -EHOSTUNREACH;
1430 write_lock_irqsave(g_lock, flags);
1432 peer2 = kiblnd_find_peer_locked(ni, nid);
1433 if (peer2 != NULL) {
1434 if (list_empty(&peer2->ibp_conns)) {
1435 /* found a peer_ni, but it's still connecting... */
1436 LASSERT(kiblnd_peer_connecting(peer2));
1438 list_add_tail(&tx->tx_list,
1439 &peer2->ibp_tx_queue);
1440 write_unlock_irqrestore(g_lock, flags);
1442 conn = kiblnd_get_conn_locked(peer2);
1443 kiblnd_conn_addref(conn); /* 1 ref for me... */
1445 write_unlock_irqrestore(g_lock, flags);
1448 kiblnd_queue_tx(tx, conn);
1449 kiblnd_conn_decref(conn); /* ...to here */
1452 kiblnd_peer_decref(peer_ni);
1456 /* Brand new peer_ni */
1457 LASSERT (peer_ni->ibp_connecting == 0);
1458 peer_ni->ibp_connecting = 1;
1460 /* always called with a ref on ni, which prevents ni being shutdown */
1461 LASSERT (((kib_net_t *)ni->ni_data)->ibn_shutdown == 0);
1464 list_add_tail(&tx->tx_list, &peer_ni->ibp_tx_queue);
1466 kiblnd_peer_addref(peer_ni);
1467 list_add_tail(&peer_ni->ibp_list, kiblnd_nid2peerlist(nid));
1469 write_unlock_irqrestore(g_lock, flags);
1471 kiblnd_connect_peer(peer_ni);
1472 kiblnd_peer_decref(peer_ni);
1476 kiblnd_send(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg)
1478 struct lnet_hdr *hdr = &lntmsg->msg_hdr;
1479 int type = lntmsg->msg_type;
1480 struct lnet_process_id target = lntmsg->msg_target;
1481 int target_is_router = lntmsg->msg_target_is_router;
1482 int routing = lntmsg->msg_routing;
1483 unsigned int payload_niov = lntmsg->msg_niov;
1484 struct kvec *payload_iov = lntmsg->msg_iov;
1485 lnet_kiov_t *payload_kiov = lntmsg->msg_kiov;
1486 unsigned int payload_offset = lntmsg->msg_offset;
1487 unsigned int payload_nob = lntmsg->msg_len;
1489 kib_rdma_desc_t *rd;
1494 /* NB 'private' is different depending on what we're sending.... */
1496 CDEBUG(D_NET, "sending %d bytes in %d frags to %s\n",
1497 payload_nob, payload_niov, libcfs_id2str(target));
1499 LASSERT (payload_nob == 0 || payload_niov > 0);
1500 LASSERT (payload_niov <= LNET_MAX_IOV);
1502 /* Thread context */
1503 LASSERT (!in_interrupt());
1504 /* payload is either all vaddrs or all pages */
1505 LASSERT (!(payload_kiov != NULL && payload_iov != NULL));
1513 LASSERT (payload_nob == 0);
1517 if (routing || target_is_router)
1518 break; /* send IMMEDIATE */
1520 /* is the REPLY message too small for RDMA? */
1521 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[lntmsg->msg_md->md_length]);
1522 if (nob <= IBLND_MSG_SIZE)
1523 break; /* send IMMEDIATE */
1525 tx = kiblnd_get_idle_tx(ni, target.nid);
1527 CERROR("Can't allocate txd for GET to %s\n",
1528 libcfs_nid2str(target.nid));
1533 rd = &ibmsg->ibm_u.get.ibgm_rd;
1534 if ((lntmsg->msg_md->md_options & LNET_MD_KIOV) == 0)
1535 rc = kiblnd_setup_rd_iov(ni, tx, rd,
1536 lntmsg->msg_md->md_niov,
1537 lntmsg->msg_md->md_iov.iov,
1538 0, lntmsg->msg_md->md_length);
1540 rc = kiblnd_setup_rd_kiov(ni, tx, rd,
1541 lntmsg->msg_md->md_niov,
1542 lntmsg->msg_md->md_iov.kiov,
1543 0, lntmsg->msg_md->md_length);
1545 CERROR("Can't setup GET sink for %s: %d\n",
1546 libcfs_nid2str(target.nid), rc);
1551 nob = offsetof(kib_get_msg_t, ibgm_rd.rd_frags[rd->rd_nfrags]);
1552 ibmsg->ibm_u.get.ibgm_cookie = tx->tx_cookie;
1553 ibmsg->ibm_u.get.ibgm_hdr = *hdr;
1555 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_GET_REQ, nob);
1557 tx->tx_lntmsg[1] = lnet_create_reply_msg(ni, lntmsg);
1558 if (tx->tx_lntmsg[1] == NULL) {
1559 CERROR("Can't create reply for GET -> %s\n",
1560 libcfs_nid2str(target.nid));
1565 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg[0,1] on completion */
1566 tx->tx_waiting = 1; /* waiting for GET_DONE */
1567 kiblnd_launch_tx(ni, tx, target.nid);
1570 case LNET_MSG_REPLY:
1572 /* Is the payload small enough not to need RDMA? */
1573 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob]);
1574 if (nob <= IBLND_MSG_SIZE)
1575 break; /* send IMMEDIATE */
1577 tx = kiblnd_get_idle_tx(ni, target.nid);
1579 CERROR("Can't allocate %s txd for %s\n",
1580 type == LNET_MSG_PUT ? "PUT" : "REPLY",
1581 libcfs_nid2str(target.nid));
1585 if (payload_kiov == NULL)
1586 rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1587 payload_niov, payload_iov,
1588 payload_offset, payload_nob);
1590 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1591 payload_niov, payload_kiov,
1592 payload_offset, payload_nob);
1594 CERROR("Can't setup PUT src for %s: %d\n",
1595 libcfs_nid2str(target.nid), rc);
1601 ibmsg->ibm_u.putreq.ibprm_hdr = *hdr;
1602 ibmsg->ibm_u.putreq.ibprm_cookie = tx->tx_cookie;
1603 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_REQ, sizeof(kib_putreq_msg_t));
1605 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1606 tx->tx_waiting = 1; /* waiting for PUT_{ACK,NAK} */
1607 kiblnd_launch_tx(ni, tx, target.nid);
1611 /* send IMMEDIATE */
1613 LASSERT (offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob])
1616 tx = kiblnd_get_idle_tx(ni, target.nid);
1618 CERROR ("Can't send %d to %s: tx descs exhausted\n",
1619 type, libcfs_nid2str(target.nid));
1624 ibmsg->ibm_u.immediate.ibim_hdr = *hdr;
1626 if (payload_kiov != NULL)
1627 lnet_copy_kiov2flat(IBLND_MSG_SIZE, ibmsg,
1628 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1629 payload_niov, payload_kiov,
1630 payload_offset, payload_nob);
1632 lnet_copy_iov2flat(IBLND_MSG_SIZE, ibmsg,
1633 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1634 payload_niov, payload_iov,
1635 payload_offset, payload_nob);
1637 nob = offsetof(kib_immediate_msg_t, ibim_payload[payload_nob]);
1638 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_IMMEDIATE, nob);
1640 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1641 kiblnd_launch_tx(ni, tx, target.nid);
1646 kiblnd_reply(struct lnet_ni *ni, kib_rx_t *rx, struct lnet_msg *lntmsg)
1648 struct lnet_process_id target = lntmsg->msg_target;
1649 unsigned int niov = lntmsg->msg_niov;
1650 struct kvec *iov = lntmsg->msg_iov;
1651 lnet_kiov_t *kiov = lntmsg->msg_kiov;
1652 unsigned int offset = lntmsg->msg_offset;
1653 unsigned int nob = lntmsg->msg_len;
1657 tx = kiblnd_get_idle_tx(ni, rx->rx_conn->ibc_peer->ibp_nid);
1659 CERROR("Can't get tx for REPLY to %s\n",
1660 libcfs_nid2str(target.nid));
1666 else if (kiov == NULL)
1667 rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1668 niov, iov, offset, nob);
1670 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1671 niov, kiov, offset, nob);
1674 CERROR("Can't setup GET src for %s: %d\n",
1675 libcfs_nid2str(target.nid), rc);
1679 rc = kiblnd_init_rdma(rx->rx_conn, tx,
1680 IBLND_MSG_GET_DONE, nob,
1681 &rx->rx_msg->ibm_u.get.ibgm_rd,
1682 rx->rx_msg->ibm_u.get.ibgm_cookie);
1684 CERROR("Can't setup rdma for GET from %s: %d\n",
1685 libcfs_nid2str(target.nid), rc);
1690 /* No RDMA: local completion may happen now! */
1691 lnet_finalize(lntmsg, 0);
1693 /* RDMA: lnet_finalize(lntmsg) when it
1695 tx->tx_lntmsg[0] = lntmsg;
1698 kiblnd_queue_tx(tx, rx->rx_conn);
1704 lnet_finalize(lntmsg, -EIO);
1708 kiblnd_recv(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg,
1709 int delayed, unsigned int niov, struct kvec *iov, lnet_kiov_t *kiov,
1710 unsigned int offset, unsigned int mlen, unsigned int rlen)
1712 kib_rx_t *rx = private;
1713 kib_msg_t *rxmsg = rx->rx_msg;
1714 kib_conn_t *conn = rx->rx_conn;
1718 int post_credit = IBLND_POSTRX_PEER_CREDIT;
1721 LASSERT (mlen <= rlen);
1722 LASSERT (!in_interrupt());
1723 /* Either all pages or all vaddrs */
1724 LASSERT (!(kiov != NULL && iov != NULL));
1726 switch (rxmsg->ibm_type) {
1730 case IBLND_MSG_IMMEDIATE:
1731 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[rlen]);
1732 if (nob > rx->rx_nob) {
1733 CERROR ("Immediate message from %s too big: %d(%d)\n",
1734 libcfs_nid2str(rxmsg->ibm_u.immediate.ibim_hdr.src_nid),
1741 lnet_copy_flat2kiov(niov, kiov, offset,
1742 IBLND_MSG_SIZE, rxmsg,
1743 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1746 lnet_copy_flat2iov(niov, iov, offset,
1747 IBLND_MSG_SIZE, rxmsg,
1748 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1750 lnet_finalize(lntmsg, 0);
1753 case IBLND_MSG_PUT_REQ: {
1755 kib_rdma_desc_t *rd;
1756 ibprm_cookie = rxmsg->ibm_u.putreq.ibprm_cookie;
1759 lnet_finalize(lntmsg, 0);
1760 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK,
1765 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
1767 CERROR("Can't allocate tx for %s\n",
1768 libcfs_nid2str(conn->ibc_peer->ibp_nid));
1769 /* Not replying will break the connection */
1775 rd = &txmsg->ibm_u.putack.ibpam_rd;
1777 rc = kiblnd_setup_rd_iov(ni, tx, rd,
1778 niov, iov, offset, mlen);
1780 rc = kiblnd_setup_rd_kiov(ni, tx, rd,
1781 niov, kiov, offset, mlen);
1783 CERROR("Can't setup PUT sink for %s: %d\n",
1784 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
1786 /* tell peer_ni it's over */
1787 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK,
1792 nob = offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[rd->rd_nfrags]);
1793 txmsg->ibm_u.putack.ibpam_src_cookie = ibprm_cookie;
1794 txmsg->ibm_u.putack.ibpam_dst_cookie = tx->tx_cookie;
1796 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_ACK, nob);
1798 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1799 tx->tx_waiting = 1; /* waiting for PUT_DONE */
1800 kiblnd_queue_tx(tx, conn);
1802 /* reposted buffer reserved for PUT_DONE */
1803 post_credit = IBLND_POSTRX_NO_CREDIT;
1807 case IBLND_MSG_GET_REQ:
1808 if (lntmsg != NULL) {
1809 /* Optimized GET; RDMA lntmsg's payload */
1810 kiblnd_reply(ni, rx, lntmsg);
1812 /* GET didn't match anything */
1813 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_GET_DONE,
1815 rxmsg->ibm_u.get.ibgm_cookie);
1820 kiblnd_post_rx(rx, post_credit);
1825 kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name)
1827 struct task_struct *task = kthread_run(fn, arg, name);
1830 return PTR_ERR(task);
1832 atomic_inc(&kiblnd_data.kib_nthreads);
1837 kiblnd_thread_fini (void)
1839 atomic_dec (&kiblnd_data.kib_nthreads);
1843 kiblnd_peer_alive (kib_peer_ni_t *peer_ni)
1845 /* This is racy, but everyone's only writing cfs_time_current() */
1846 peer_ni->ibp_last_alive = cfs_time_current();
1851 kiblnd_peer_notify (kib_peer_ni_t *peer_ni)
1854 cfs_time_t last_alive = 0;
1855 unsigned long flags;
1857 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1859 if (kiblnd_peer_idle(peer_ni) && peer_ni->ibp_error != 0) {
1860 error = peer_ni->ibp_error;
1861 peer_ni->ibp_error = 0;
1863 last_alive = peer_ni->ibp_last_alive;
1866 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1869 lnet_notify(peer_ni->ibp_ni,
1870 peer_ni->ibp_nid, 0, last_alive);
1874 kiblnd_close_conn_locked (kib_conn_t *conn, int error)
1876 /* This just does the immediate housekeeping. 'error' is zero for a
1877 * normal shutdown which can happen only after the connection has been
1878 * established. If the connection is established, schedule the
1879 * connection to be finished off by the connd. Otherwise the connd is
1880 * already dealing with it (either to set it up or tear it down).
1881 * Caller holds kib_global_lock exclusively in irq context */
1882 kib_peer_ni_t *peer_ni = conn->ibc_peer;
1884 unsigned long flags;
1886 LASSERT (error != 0 || conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1888 if (error != 0 && conn->ibc_comms_error == 0)
1889 conn->ibc_comms_error = error;
1891 if (conn->ibc_state != IBLND_CONN_ESTABLISHED)
1892 return; /* already being handled */
1895 list_empty(&conn->ibc_tx_noops) &&
1896 list_empty(&conn->ibc_tx_queue) &&
1897 list_empty(&conn->ibc_tx_queue_rsrvd) &&
1898 list_empty(&conn->ibc_tx_queue_nocred) &&
1899 list_empty(&conn->ibc_active_txs)) {
1900 CDEBUG(D_NET, "closing conn to %s\n",
1901 libcfs_nid2str(peer_ni->ibp_nid));
1903 CNETERR("Closing conn to %s: error %d%s%s%s%s%s\n",
1904 libcfs_nid2str(peer_ni->ibp_nid), error,
1905 list_empty(&conn->ibc_tx_queue) ? "" : "(sending)",
1906 list_empty(&conn->ibc_tx_noops) ? "" : "(sending_noops)",
1907 list_empty(&conn->ibc_tx_queue_rsrvd) ?
1908 "" : "(sending_rsrvd)",
1909 list_empty(&conn->ibc_tx_queue_nocred) ?
1910 "" : "(sending_nocred)",
1911 list_empty(&conn->ibc_active_txs) ? "" : "(waiting)");
1914 dev = ((kib_net_t *)peer_ni->ibp_ni->ni_data)->ibn_dev;
1915 list_del(&conn->ibc_list);
1916 /* connd (see below) takes over ibc_list's ref */
1918 if (list_empty(&peer_ni->ibp_conns) && /* no more conns */
1919 kiblnd_peer_active(peer_ni)) { /* still in peer_ni table */
1920 kiblnd_unlink_peer_locked(peer_ni);
1922 /* set/clear error on last conn */
1923 peer_ni->ibp_error = conn->ibc_comms_error;
1926 kiblnd_set_conn_state(conn, IBLND_CONN_CLOSING);
1929 kiblnd_dev_can_failover(dev)) {
1930 list_add_tail(&dev->ibd_fail_list,
1931 &kiblnd_data.kib_failed_devs);
1932 wake_up(&kiblnd_data.kib_failover_waitq);
1935 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
1937 list_add_tail(&conn->ibc_list, &kiblnd_data.kib_connd_conns);
1938 wake_up(&kiblnd_data.kib_connd_waitq);
1940 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
1944 kiblnd_close_conn(kib_conn_t *conn, int error)
1946 unsigned long flags;
1948 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1950 kiblnd_close_conn_locked(conn, error);
1952 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1956 kiblnd_handle_early_rxs(kib_conn_t *conn)
1958 unsigned long flags;
1961 LASSERT(!in_interrupt());
1962 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1964 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1965 while (!list_empty(&conn->ibc_early_rxs)) {
1966 rx = list_entry(conn->ibc_early_rxs.next,
1968 list_del(&rx->rx_list);
1969 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1971 kiblnd_handle_rx(rx);
1973 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1975 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1979 kiblnd_abort_txs(kib_conn_t *conn, struct list_head *txs)
1981 struct list_head zombies = LIST_HEAD_INIT(zombies);
1982 struct list_head *tmp;
1983 struct list_head *nxt;
1986 spin_lock(&conn->ibc_lock);
1988 list_for_each_safe(tmp, nxt, txs) {
1989 tx = list_entry(tmp, kib_tx_t, tx_list);
1991 if (txs == &conn->ibc_active_txs) {
1992 LASSERT(!tx->tx_queued);
1993 LASSERT(tx->tx_waiting ||
1994 tx->tx_sending != 0);
1996 LASSERT(tx->tx_queued);
1999 tx->tx_status = -ECONNABORTED;
2002 if (tx->tx_sending == 0) {
2004 list_del(&tx->tx_list);
2005 list_add(&tx->tx_list, &zombies);
2009 spin_unlock(&conn->ibc_lock);
2011 kiblnd_txlist_done(&zombies, -ECONNABORTED);
2015 kiblnd_finalise_conn (kib_conn_t *conn)
2017 LASSERT (!in_interrupt());
2018 LASSERT (conn->ibc_state > IBLND_CONN_INIT);
2020 kiblnd_set_conn_state(conn, IBLND_CONN_DISCONNECTED);
2022 /* abort_receives moves QP state to IB_QPS_ERR. This is only required
2023 * for connections that didn't get as far as being connected, because
2024 * rdma_disconnect() does this for free. */
2025 kiblnd_abort_receives(conn);
2027 /* Complete all tx descs not waiting for sends to complete.
2028 * NB we should be safe from RDMA now that the QP has changed state */
2030 kiblnd_abort_txs(conn, &conn->ibc_tx_noops);
2031 kiblnd_abort_txs(conn, &conn->ibc_tx_queue);
2032 kiblnd_abort_txs(conn, &conn->ibc_tx_queue_rsrvd);
2033 kiblnd_abort_txs(conn, &conn->ibc_tx_queue_nocred);
2034 kiblnd_abort_txs(conn, &conn->ibc_active_txs);
2036 kiblnd_handle_early_rxs(conn);
2040 kiblnd_peer_connect_failed(kib_peer_ni_t *peer_ni, int active, int error)
2042 struct list_head zombies = LIST_HEAD_INIT(zombies);
2043 unsigned long flags;
2045 LASSERT (error != 0);
2046 LASSERT (!in_interrupt());
2048 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2051 LASSERT(peer_ni->ibp_connecting > 0);
2052 peer_ni->ibp_connecting--;
2054 LASSERT (peer_ni->ibp_accepting > 0);
2055 peer_ni->ibp_accepting--;
2058 if (kiblnd_peer_connecting(peer_ni)) {
2059 /* another connection attempt under way... */
2060 write_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2065 peer_ni->ibp_reconnected = 0;
2066 if (list_empty(&peer_ni->ibp_conns)) {
2067 /* Take peer_ni's blocked transmits to complete with error */
2068 list_add(&zombies, &peer_ni->ibp_tx_queue);
2069 list_del_init(&peer_ni->ibp_tx_queue);
2071 if (kiblnd_peer_active(peer_ni))
2072 kiblnd_unlink_peer_locked(peer_ni);
2074 peer_ni->ibp_error = error;
2076 /* Can't have blocked transmits if there are connections */
2077 LASSERT(list_empty(&peer_ni->ibp_tx_queue));
2080 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2082 kiblnd_peer_notify(peer_ni);
2084 if (list_empty(&zombies))
2087 CNETERR("Deleting messages for %s: connection failed\n",
2088 libcfs_nid2str(peer_ni->ibp_nid));
2090 kiblnd_txlist_done(&zombies, -EHOSTUNREACH);
2094 kiblnd_connreq_done(kib_conn_t *conn, int status)
2096 kib_peer_ni_t *peer_ni = conn->ibc_peer;
2098 struct list_head txs;
2099 unsigned long flags;
2102 active = (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2104 CDEBUG(D_NET,"%s: active(%d), version(%x), status(%d)\n",
2105 libcfs_nid2str(peer_ni->ibp_nid), active,
2106 conn->ibc_version, status);
2108 LASSERT (!in_interrupt());
2109 LASSERT ((conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT &&
2110 peer_ni->ibp_connecting > 0) ||
2111 (conn->ibc_state == IBLND_CONN_PASSIVE_WAIT &&
2112 peer_ni->ibp_accepting > 0));
2114 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
2115 conn->ibc_connvars = NULL;
2118 /* failed to establish connection */
2119 kiblnd_peer_connect_failed(peer_ni, active, status);
2120 kiblnd_finalise_conn(conn);
2124 /* connection established */
2125 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2127 conn->ibc_last_send = jiffies;
2128 kiblnd_set_conn_state(conn, IBLND_CONN_ESTABLISHED);
2129 kiblnd_peer_alive(peer_ni);
2131 /* Add conn to peer_ni's list and nuke any dangling conns from a different
2132 * peer_ni instance... */
2133 kiblnd_conn_addref(conn); /* +1 ref for ibc_list */
2134 list_add(&conn->ibc_list, &peer_ni->ibp_conns);
2135 peer_ni->ibp_reconnected = 0;
2137 peer_ni->ibp_connecting--;
2139 peer_ni->ibp_accepting--;
2141 if (peer_ni->ibp_version == 0) {
2142 peer_ni->ibp_version = conn->ibc_version;
2143 peer_ni->ibp_incarnation = conn->ibc_incarnation;
2146 if (peer_ni->ibp_version != conn->ibc_version ||
2147 peer_ni->ibp_incarnation != conn->ibc_incarnation) {
2148 kiblnd_close_stale_conns_locked(peer_ni, conn->ibc_version,
2149 conn->ibc_incarnation);
2150 peer_ni->ibp_version = conn->ibc_version;
2151 peer_ni->ibp_incarnation = conn->ibc_incarnation;
2154 /* grab pending txs while I have the lock */
2155 list_add(&txs, &peer_ni->ibp_tx_queue);
2156 list_del_init(&peer_ni->ibp_tx_queue);
2158 if (!kiblnd_peer_active(peer_ni) || /* peer_ni has been deleted */
2159 conn->ibc_comms_error != 0) { /* error has happened already */
2161 /* start to shut down connection */
2162 kiblnd_close_conn_locked(conn, -ECONNABORTED);
2163 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2165 kiblnd_txlist_done(&txs, -ECONNABORTED);
2170 /* +1 ref for myself, this connection is visible to other threads
2171 * now, refcount of peer:ibp_conns can be released by connection
2172 * close from either a different thread, or the calling of
2173 * kiblnd_check_sends_locked() below. See bz21911 for details.
2175 kiblnd_conn_addref(conn);
2176 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2178 /* Schedule blocked txs */
2179 spin_lock(&conn->ibc_lock);
2180 while (!list_empty(&txs)) {
2181 tx = list_entry(txs.next, kib_tx_t, tx_list);
2182 list_del(&tx->tx_list);
2184 kiblnd_queue_tx_locked(tx, conn);
2186 kiblnd_check_sends_locked(conn);
2187 spin_unlock(&conn->ibc_lock);
2189 /* schedule blocked rxs */
2190 kiblnd_handle_early_rxs(conn);
2191 kiblnd_conn_decref(conn);
2195 kiblnd_reject(struct rdma_cm_id *cmid, kib_rej_t *rej)
2199 rc = rdma_reject(cmid, rej, sizeof(*rej));
2202 CWARN("Error %d sending reject\n", rc);
2206 kiblnd_passive_connect(struct rdma_cm_id *cmid, void *priv, int priv_nob)
2208 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2209 kib_msg_t *reqmsg = priv;
2212 kib_peer_ni_t *peer_ni;
2213 kib_peer_ni_t *peer2;
2215 struct lnet_ni *ni = NULL;
2216 kib_net_t *net = NULL;
2218 struct rdma_conn_param cp;
2220 int version = IBLND_MSG_VERSION;
2221 unsigned long flags;
2223 struct sockaddr_in *peer_addr;
2224 LASSERT (!in_interrupt());
2226 /* cmid inherits 'context' from the corresponding listener id */
2227 ibdev = (kib_dev_t *)cmid->context;
2228 LASSERT (ibdev != NULL);
2230 memset(&rej, 0, sizeof(rej));
2231 rej.ibr_magic = IBLND_MSG_MAGIC;
2232 rej.ibr_why = IBLND_REJECT_FATAL;
2233 rej.ibr_cp.ibcp_max_msg_size = IBLND_MSG_SIZE;
2235 peer_addr = (struct sockaddr_in *)&(cmid->route.addr.dst_addr);
2236 if (*kiblnd_tunables.kib_require_priv_port &&
2237 ntohs(peer_addr->sin_port) >= PROT_SOCK) {
2238 __u32 ip = ntohl(peer_addr->sin_addr.s_addr);
2239 CERROR("peer_ni's port (%pI4h:%hu) is not privileged\n",
2240 &ip, ntohs(peer_addr->sin_port));
2244 if (priv_nob < offsetof(kib_msg_t, ibm_type)) {
2245 CERROR("Short connection request\n");
2249 /* Future protocol version compatibility support! If the
2250 * o2iblnd-specific protocol changes, or when LNET unifies
2251 * protocols over all LNDs, the initial connection will
2252 * negotiate a protocol version. I trap this here to avoid
2253 * console errors; the reject tells the peer_ni which protocol I
2255 if (reqmsg->ibm_magic == LNET_PROTO_MAGIC ||
2256 reqmsg->ibm_magic == __swab32(LNET_PROTO_MAGIC))
2258 if (reqmsg->ibm_magic == IBLND_MSG_MAGIC &&
2259 reqmsg->ibm_version != IBLND_MSG_VERSION &&
2260 reqmsg->ibm_version != IBLND_MSG_VERSION_1)
2262 if (reqmsg->ibm_magic == __swab32(IBLND_MSG_MAGIC) &&
2263 reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION) &&
2264 reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION_1))
2267 rc = kiblnd_unpack_msg(reqmsg, priv_nob);
2269 CERROR("Can't parse connection request: %d\n", rc);
2273 nid = reqmsg->ibm_srcnid;
2274 ni = lnet_nid2ni_addref(reqmsg->ibm_dstnid);
2277 net = (kib_net_t *)ni->ni_data;
2278 rej.ibr_incarnation = net->ibn_incarnation;
2281 if (ni == NULL || /* no matching net */
2282 ni->ni_nid != reqmsg->ibm_dstnid || /* right NET, wrong NID! */
2283 net->ibn_dev != ibdev) { /* wrong device */
2284 CERROR("Can't accept conn from %s on %s (%s:%d:%pI4h): "
2285 "bad dst nid %s\n", libcfs_nid2str(nid),
2286 ni == NULL ? "NA" : libcfs_nid2str(ni->ni_nid),
2287 ibdev->ibd_ifname, ibdev->ibd_nnets,
2289 libcfs_nid2str(reqmsg->ibm_dstnid));
2294 /* check time stamp as soon as possible */
2295 if (reqmsg->ibm_dststamp != 0 &&
2296 reqmsg->ibm_dststamp != net->ibn_incarnation) {
2297 CWARN("Stale connection request\n");
2298 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2302 /* I can accept peer_ni's version */
2303 version = reqmsg->ibm_version;
2305 if (reqmsg->ibm_type != IBLND_MSG_CONNREQ) {
2306 CERROR("Unexpected connreq msg type: %x from %s\n",
2307 reqmsg->ibm_type, libcfs_nid2str(nid));
2311 if (reqmsg->ibm_u.connparams.ibcp_queue_depth >
2312 kiblnd_msg_queue_size(version, ni)) {
2313 CERROR("Can't accept conn from %s, queue depth too large: "
2314 " %d (<=%d wanted)\n",
2315 libcfs_nid2str(nid),
2316 reqmsg->ibm_u.connparams.ibcp_queue_depth,
2317 kiblnd_msg_queue_size(version, ni));
2319 if (version == IBLND_MSG_VERSION)
2320 rej.ibr_why = IBLND_REJECT_MSG_QUEUE_SIZE;
2325 if (reqmsg->ibm_u.connparams.ibcp_max_frags >
2326 kiblnd_rdma_frags(version, ni)) {
2327 CWARN("Can't accept conn from %s (version %x): "
2328 "max_frags %d too large (%d wanted)\n",
2329 libcfs_nid2str(nid), version,
2330 reqmsg->ibm_u.connparams.ibcp_max_frags,
2331 kiblnd_rdma_frags(version, ni));
2333 if (version >= IBLND_MSG_VERSION)
2334 rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2337 } else if (reqmsg->ibm_u.connparams.ibcp_max_frags <
2338 kiblnd_rdma_frags(version, ni) &&
2339 net->ibn_fmr_ps == NULL) {
2340 CWARN("Can't accept conn from %s (version %x): "
2341 "max_frags %d incompatible without FMR pool "
2343 libcfs_nid2str(nid), version,
2344 reqmsg->ibm_u.connparams.ibcp_max_frags,
2345 kiblnd_rdma_frags(version, ni));
2347 if (version == IBLND_MSG_VERSION)
2348 rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2353 if (reqmsg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2354 CERROR("Can't accept %s: message size %d too big (%d max)\n",
2355 libcfs_nid2str(nid),
2356 reqmsg->ibm_u.connparams.ibcp_max_msg_size,
2361 /* assume 'nid' is a new peer_ni; create */
2362 rc = kiblnd_create_peer(ni, &peer_ni, nid);
2364 CERROR("Can't create peer_ni for %s\n", libcfs_nid2str(nid));
2365 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2369 /* We have validated the peer's parameters so use those */
2370 peer_ni->ibp_max_frags = reqmsg->ibm_u.connparams.ibcp_max_frags;
2371 peer_ni->ibp_queue_depth = reqmsg->ibm_u.connparams.ibcp_queue_depth;
2373 write_lock_irqsave(g_lock, flags);
2375 peer2 = kiblnd_find_peer_locked(ni, nid);
2376 if (peer2 != NULL) {
2377 if (peer2->ibp_version == 0) {
2378 peer2->ibp_version = version;
2379 peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2382 /* not the guy I've talked with */
2383 if (peer2->ibp_incarnation != reqmsg->ibm_srcstamp ||
2384 peer2->ibp_version != version) {
2385 kiblnd_close_peer_conns_locked(peer2, -ESTALE);
2387 if (kiblnd_peer_active(peer2)) {
2388 peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2389 peer2->ibp_version = version;
2391 write_unlock_irqrestore(g_lock, flags);
2393 CWARN("Conn stale %s version %x/%x incarnation %llu/%llu\n",
2394 libcfs_nid2str(nid), peer2->ibp_version, version,
2395 peer2->ibp_incarnation, reqmsg->ibm_srcstamp);
2397 kiblnd_peer_decref(peer_ni);
2398 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2402 /* Tie-break connection race in favour of the higher NID.
2403 * If we keep running into a race condition multiple times,
2404 * we have to assume that the connection attempt with the
2405 * higher NID is stuck in a connecting state and will never
2406 * recover. As such, we pass through this if-block and let
2407 * the lower NID connection win so we can move forward.
2409 if (peer2->ibp_connecting != 0 &&
2410 nid < ni->ni_nid && peer2->ibp_races <
2411 MAX_CONN_RACES_BEFORE_ABORT) {
2413 write_unlock_irqrestore(g_lock, flags);
2415 CDEBUG(D_NET, "Conn race %s\n",
2416 libcfs_nid2str(peer2->ibp_nid));
2418 kiblnd_peer_decref(peer_ni);
2419 rej.ibr_why = IBLND_REJECT_CONN_RACE;
2422 if (peer2->ibp_races >= MAX_CONN_RACES_BEFORE_ABORT)
2423 CNETERR("Conn race %s: unresolved after %d attempts, letting lower NID win\n",
2424 libcfs_nid2str(peer2->ibp_nid),
2425 MAX_CONN_RACES_BEFORE_ABORT);
2427 * passive connection is allowed even this peer_ni is waiting for
2430 peer2->ibp_reconnecting = 0;
2431 peer2->ibp_races = 0;
2432 peer2->ibp_accepting++;
2433 kiblnd_peer_addref(peer2);
2435 /* Race with kiblnd_launch_tx (active connect) to create peer_ni
2436 * so copy validated parameters since we now know what the
2437 * peer_ni's limits are */
2438 peer2->ibp_max_frags = peer_ni->ibp_max_frags;
2439 peer2->ibp_queue_depth = peer_ni->ibp_queue_depth;
2441 write_unlock_irqrestore(g_lock, flags);
2442 kiblnd_peer_decref(peer_ni);
2445 /* Brand new peer_ni */
2446 LASSERT (peer_ni->ibp_accepting == 0);
2447 LASSERT (peer_ni->ibp_version == 0 &&
2448 peer_ni->ibp_incarnation == 0);
2450 peer_ni->ibp_accepting = 1;
2451 peer_ni->ibp_version = version;
2452 peer_ni->ibp_incarnation = reqmsg->ibm_srcstamp;
2454 /* I have a ref on ni that prevents it being shutdown */
2455 LASSERT (net->ibn_shutdown == 0);
2457 kiblnd_peer_addref(peer_ni);
2458 list_add_tail(&peer_ni->ibp_list, kiblnd_nid2peerlist(nid));
2460 write_unlock_irqrestore(g_lock, flags);
2463 conn = kiblnd_create_conn(peer_ni, cmid, IBLND_CONN_PASSIVE_WAIT, version);
2465 kiblnd_peer_connect_failed(peer_ni, 0, -ENOMEM);
2466 kiblnd_peer_decref(peer_ni);
2467 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2471 /* conn now "owns" cmid, so I return success from here on to ensure the
2472 * CM callback doesn't destroy cmid. */
2473 conn->ibc_incarnation = reqmsg->ibm_srcstamp;
2474 conn->ibc_credits = conn->ibc_queue_depth;
2475 conn->ibc_reserved_credits = conn->ibc_queue_depth;
2476 LASSERT(conn->ibc_credits + conn->ibc_reserved_credits +
2477 IBLND_OOB_MSGS(version) <= IBLND_RX_MSGS(conn));
2479 ackmsg = &conn->ibc_connvars->cv_msg;
2480 memset(ackmsg, 0, sizeof(*ackmsg));
2482 kiblnd_init_msg(ackmsg, IBLND_MSG_CONNACK,
2483 sizeof(ackmsg->ibm_u.connparams));
2484 ackmsg->ibm_u.connparams.ibcp_queue_depth = conn->ibc_queue_depth;
2485 ackmsg->ibm_u.connparams.ibcp_max_frags = conn->ibc_max_frags;
2486 ackmsg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2488 kiblnd_pack_msg(ni, ackmsg, version, 0, nid, reqmsg->ibm_srcstamp);
2490 memset(&cp, 0, sizeof(cp));
2491 cp.private_data = ackmsg;
2492 cp.private_data_len = ackmsg->ibm_nob;
2493 cp.responder_resources = 0; /* No atomic ops or RDMA reads */
2494 cp.initiator_depth = 0;
2495 cp.flow_control = 1;
2496 cp.retry_count = *kiblnd_tunables.kib_retry_count;
2497 cp.rnr_retry_count = *kiblnd_tunables.kib_rnr_retry_count;
2499 CDEBUG(D_NET, "Accept %s\n", libcfs_nid2str(nid));
2501 rc = rdma_accept(cmid, &cp);
2503 CERROR("Can't accept %s: %d\n", libcfs_nid2str(nid), rc);
2504 rej.ibr_version = version;
2505 rej.ibr_why = IBLND_REJECT_FATAL;
2507 kiblnd_reject(cmid, &rej);
2508 kiblnd_connreq_done(conn, rc);
2509 kiblnd_conn_decref(conn);
2517 rej.ibr_cp.ibcp_queue_depth =
2518 kiblnd_msg_queue_size(version, ni);
2519 rej.ibr_cp.ibcp_max_frags = kiblnd_rdma_frags(version, ni);
2523 rej.ibr_version = version;
2524 kiblnd_reject(cmid, &rej);
2526 return -ECONNREFUSED;
2530 kiblnd_check_reconnect(kib_conn_t *conn, int version,
2531 __u64 incarnation, int why, kib_connparams_t *cp)
2533 rwlock_t *glock = &kiblnd_data.kib_global_lock;
2534 kib_peer_ni_t *peer_ni = conn->ibc_peer;
2536 int msg_size = IBLND_MSG_SIZE;
2540 unsigned long flags;
2542 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2543 LASSERT(peer_ni->ibp_connecting > 0); /* 'conn' at least */
2544 LASSERT(!peer_ni->ibp_reconnecting);
2547 msg_size = cp->ibcp_max_msg_size;
2548 frag_num = cp->ibcp_max_frags;
2549 queue_dep = cp->ibcp_queue_depth;
2552 write_lock_irqsave(glock, flags);
2553 /* retry connection if it's still needed and no other connection
2554 * attempts (active or passive) are in progress
2555 * NB: reconnect is still needed even when ibp_tx_queue is
2556 * empty if ibp_version != version because reconnect may be
2557 * initiated by kiblnd_query() */
2558 reconnect = (!list_empty(&peer_ni->ibp_tx_queue) ||
2559 peer_ni->ibp_version != version) &&
2560 peer_ni->ibp_connecting == 1 &&
2561 peer_ni->ibp_accepting == 0;
2572 case IBLND_REJECT_RDMA_FRAGS: {
2573 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2576 reason = "can't negotiate max frags";
2579 tunables = &peer_ni->ibp_ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2580 if (!tunables->lnd_map_on_demand) {
2581 reason = "map_on_demand must be enabled";
2584 if (conn->ibc_max_frags <= frag_num) {
2585 reason = "unsupported max frags";
2589 peer_ni->ibp_max_frags = frag_num;
2590 reason = "rdma fragments";
2593 case IBLND_REJECT_MSG_QUEUE_SIZE:
2595 reason = "can't negotiate queue depth";
2598 if (conn->ibc_queue_depth <= queue_dep) {
2599 reason = "unsupported queue depth";
2603 peer_ni->ibp_queue_depth = queue_dep;
2604 reason = "queue depth";
2607 case IBLND_REJECT_CONN_STALE:
2611 case IBLND_REJECT_CONN_RACE:
2612 reason = "conn race";
2615 case IBLND_REJECT_CONN_UNCOMPAT:
2616 reason = "version negotiation";
2619 case IBLND_REJECT_INVALID_SRV_ID:
2620 reason = "invalid service id";
2624 conn->ibc_reconnect = 1;
2625 peer_ni->ibp_reconnecting = 1;
2626 peer_ni->ibp_version = version;
2627 if (incarnation != 0)
2628 peer_ni->ibp_incarnation = incarnation;
2630 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2632 CNETERR("%s: %s (%s), %x, %x, msg_size: %d, queue_depth: %d/%d, max_frags: %d/%d\n",
2633 libcfs_nid2str(peer_ni->ibp_nid),
2634 reconnect ? "reconnect" : "don't reconnect",
2635 reason, IBLND_MSG_VERSION, version, msg_size,
2636 conn->ibc_queue_depth, queue_dep,
2637 conn->ibc_max_frags, frag_num);
2639 * if conn::ibc_reconnect is TRUE, connd will reconnect to the peer_ni
2640 * while destroying the zombie
2645 kiblnd_rejected (kib_conn_t *conn, int reason, void *priv, int priv_nob)
2647 kib_peer_ni_t *peer_ni = conn->ibc_peer;
2649 LASSERT (!in_interrupt());
2650 LASSERT (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2653 case IB_CM_REJ_STALE_CONN:
2654 kiblnd_check_reconnect(conn, IBLND_MSG_VERSION, 0,
2655 IBLND_REJECT_CONN_STALE, NULL);
2658 case IB_CM_REJ_INVALID_SERVICE_ID:
2659 kiblnd_check_reconnect(conn, IBLND_MSG_VERSION, 0,
2660 IBLND_REJECT_INVALID_SRV_ID, NULL);
2661 CNETERR("%s rejected: no listener at %d\n",
2662 libcfs_nid2str(peer_ni->ibp_nid),
2663 *kiblnd_tunables.kib_service);
2666 case IB_CM_REJ_CONSUMER_DEFINED:
2667 if (priv_nob >= offsetof(kib_rej_t, ibr_padding)) {
2668 kib_rej_t *rej = priv;
2669 kib_connparams_t *cp = NULL;
2671 __u64 incarnation = -1;
2673 /* NB. default incarnation is -1 because:
2674 * a) V1 will ignore dst incarnation in connreq.
2675 * b) V2 will provide incarnation while rejecting me,
2676 * -1 will be overwrote.
2678 * if I try to connect to a V1 peer_ni with V2 protocol,
2679 * it rejected me then upgrade to V2, I have no idea
2680 * about the upgrading and try to reconnect with V1,
2681 * in this case upgraded V2 can find out I'm trying to
2682 * talk to the old guy and reject me(incarnation is -1).
2685 if (rej->ibr_magic == __swab32(IBLND_MSG_MAGIC) ||
2686 rej->ibr_magic == __swab32(LNET_PROTO_MAGIC)) {
2687 __swab32s(&rej->ibr_magic);
2688 __swab16s(&rej->ibr_version);
2692 if (priv_nob >= sizeof(kib_rej_t) &&
2693 rej->ibr_version > IBLND_MSG_VERSION_1) {
2694 /* priv_nob is always 148 in current version
2695 * of OFED, so we still need to check version.
2696 * (define of IB_CM_REJ_PRIVATE_DATA_SIZE) */
2700 __swab64s(&rej->ibr_incarnation);
2701 __swab16s(&cp->ibcp_queue_depth);
2702 __swab16s(&cp->ibcp_max_frags);
2703 __swab32s(&cp->ibcp_max_msg_size);
2706 incarnation = rej->ibr_incarnation;
2709 if (rej->ibr_magic != IBLND_MSG_MAGIC &&
2710 rej->ibr_magic != LNET_PROTO_MAGIC) {
2711 CERROR("%s rejected: consumer defined fatal error\n",
2712 libcfs_nid2str(peer_ni->ibp_nid));
2716 if (rej->ibr_version != IBLND_MSG_VERSION &&
2717 rej->ibr_version != IBLND_MSG_VERSION_1) {
2718 CERROR("%s rejected: o2iblnd version %x error\n",
2719 libcfs_nid2str(peer_ni->ibp_nid),
2724 if (rej->ibr_why == IBLND_REJECT_FATAL &&
2725 rej->ibr_version == IBLND_MSG_VERSION_1) {
2726 CDEBUG(D_NET, "rejected by old version peer_ni %s: %x\n",
2727 libcfs_nid2str(peer_ni->ibp_nid), rej->ibr_version);
2729 if (conn->ibc_version != IBLND_MSG_VERSION_1)
2730 rej->ibr_why = IBLND_REJECT_CONN_UNCOMPAT;
2733 switch (rej->ibr_why) {
2734 case IBLND_REJECT_CONN_RACE:
2735 case IBLND_REJECT_CONN_STALE:
2736 case IBLND_REJECT_CONN_UNCOMPAT:
2737 case IBLND_REJECT_MSG_QUEUE_SIZE:
2738 case IBLND_REJECT_RDMA_FRAGS:
2739 kiblnd_check_reconnect(conn, rej->ibr_version,
2740 incarnation, rej->ibr_why, cp);
2743 case IBLND_REJECT_NO_RESOURCES:
2744 CERROR("%s rejected: o2iblnd no resources\n",
2745 libcfs_nid2str(peer_ni->ibp_nid));
2748 case IBLND_REJECT_FATAL:
2749 CERROR("%s rejected: o2iblnd fatal error\n",
2750 libcfs_nid2str(peer_ni->ibp_nid));
2754 CERROR("%s rejected: o2iblnd reason %d\n",
2755 libcfs_nid2str(peer_ni->ibp_nid),
2763 CNETERR("%s rejected: reason %d, size %d\n",
2764 libcfs_nid2str(peer_ni->ibp_nid), reason, priv_nob);
2768 kiblnd_connreq_done(conn, -ECONNREFUSED);
2772 kiblnd_check_connreply (kib_conn_t *conn, void *priv, int priv_nob)
2774 kib_peer_ni_t *peer_ni = conn->ibc_peer;
2775 struct lnet_ni *ni = peer_ni->ibp_ni;
2776 kib_net_t *net = ni->ni_data;
2777 kib_msg_t *msg = priv;
2778 int ver = conn->ibc_version;
2779 int rc = kiblnd_unpack_msg(msg, priv_nob);
2780 unsigned long flags;
2782 LASSERT (net != NULL);
2785 CERROR("Can't unpack connack from %s: %d\n",
2786 libcfs_nid2str(peer_ni->ibp_nid), rc);
2790 if (msg->ibm_type != IBLND_MSG_CONNACK) {
2791 CERROR("Unexpected message %d from %s\n",
2792 msg->ibm_type, libcfs_nid2str(peer_ni->ibp_nid));
2797 if (ver != msg->ibm_version) {
2798 CERROR("%s replied version %x is different with "
2799 "requested version %x\n",
2800 libcfs_nid2str(peer_ni->ibp_nid), msg->ibm_version, ver);
2805 if (msg->ibm_u.connparams.ibcp_queue_depth >
2806 conn->ibc_queue_depth) {
2807 CERROR("%s has incompatible queue depth %d (<=%d wanted)\n",
2808 libcfs_nid2str(peer_ni->ibp_nid),
2809 msg->ibm_u.connparams.ibcp_queue_depth,
2810 conn->ibc_queue_depth);
2815 if (msg->ibm_u.connparams.ibcp_max_frags >
2816 conn->ibc_max_frags) {
2817 CERROR("%s has incompatible max_frags %d (<=%d wanted)\n",
2818 libcfs_nid2str(peer_ni->ibp_nid),
2819 msg->ibm_u.connparams.ibcp_max_frags,
2820 conn->ibc_max_frags);
2825 if (msg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2826 CERROR("%s max message size %d too big (%d max)\n",
2827 libcfs_nid2str(peer_ni->ibp_nid),
2828 msg->ibm_u.connparams.ibcp_max_msg_size,
2834 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2835 if (msg->ibm_dstnid == ni->ni_nid &&
2836 msg->ibm_dststamp == net->ibn_incarnation)
2840 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2843 CERROR("Bad connection reply from %s, rc = %d, "
2844 "version: %x max_frags: %d\n",
2845 libcfs_nid2str(peer_ni->ibp_nid), rc,
2846 msg->ibm_version, msg->ibm_u.connparams.ibcp_max_frags);
2850 conn->ibc_incarnation = msg->ibm_srcstamp;
2851 conn->ibc_credits = msg->ibm_u.connparams.ibcp_queue_depth;
2852 conn->ibc_reserved_credits = msg->ibm_u.connparams.ibcp_queue_depth;
2853 conn->ibc_queue_depth = msg->ibm_u.connparams.ibcp_queue_depth;
2854 conn->ibc_max_frags = msg->ibm_u.connparams.ibcp_max_frags;
2855 LASSERT(conn->ibc_credits + conn->ibc_reserved_credits +
2856 IBLND_OOB_MSGS(ver) <= IBLND_RX_MSGS(conn));
2858 kiblnd_connreq_done(conn, 0);
2862 /* NB My QP has already established itself, so I handle anything going
2863 * wrong here by setting ibc_comms_error.
2864 * kiblnd_connreq_done(0) moves the conn state to ESTABLISHED, but then
2865 * immediately tears it down. */
2868 conn->ibc_comms_error = rc;
2869 kiblnd_connreq_done(conn, 0);
2873 kiblnd_active_connect (struct rdma_cm_id *cmid)
2875 kib_peer_ni_t *peer_ni = (kib_peer_ni_t *)cmid->context;
2878 struct rdma_conn_param cp;
2881 unsigned long flags;
2884 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2886 incarnation = peer_ni->ibp_incarnation;
2887 version = (peer_ni->ibp_version == 0) ? IBLND_MSG_VERSION :
2888 peer_ni->ibp_version;
2890 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2892 conn = kiblnd_create_conn(peer_ni, cmid, IBLND_CONN_ACTIVE_CONNECT,
2895 kiblnd_peer_connect_failed(peer_ni, 1, -ENOMEM);
2896 kiblnd_peer_decref(peer_ni); /* lose cmid's ref */
2900 /* conn "owns" cmid now, so I return success from here on to ensure the
2901 * CM callback doesn't destroy cmid. conn also takes over cmid's ref
2904 msg = &conn->ibc_connvars->cv_msg;
2906 memset(msg, 0, sizeof(*msg));
2907 kiblnd_init_msg(msg, IBLND_MSG_CONNREQ, sizeof(msg->ibm_u.connparams));
2908 msg->ibm_u.connparams.ibcp_queue_depth = conn->ibc_queue_depth;
2909 msg->ibm_u.connparams.ibcp_max_frags = conn->ibc_max_frags;
2910 msg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2912 kiblnd_pack_msg(peer_ni->ibp_ni, msg, version,
2913 0, peer_ni->ibp_nid, incarnation);
2915 memset(&cp, 0, sizeof(cp));
2916 cp.private_data = msg;
2917 cp.private_data_len = msg->ibm_nob;
2918 cp.responder_resources = 0; /* No atomic ops or RDMA reads */
2919 cp.initiator_depth = 0;
2920 cp.flow_control = 1;
2921 cp.retry_count = *kiblnd_tunables.kib_retry_count;
2922 cp.rnr_retry_count = *kiblnd_tunables.kib_rnr_retry_count;
2924 LASSERT(cmid->context == (void *)conn);
2925 LASSERT(conn->ibc_cmid == cmid);
2927 rc = rdma_connect(cmid, &cp);
2929 CERROR("Can't connect to %s: %d\n",
2930 libcfs_nid2str(peer_ni->ibp_nid), rc);
2931 kiblnd_connreq_done(conn, rc);
2932 kiblnd_conn_decref(conn);
2939 kiblnd_cm_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2941 kib_peer_ni_t *peer_ni;
2945 switch (event->event) {
2947 CERROR("Unexpected event: %d, status: %d\n",
2948 event->event, event->status);
2951 case RDMA_CM_EVENT_CONNECT_REQUEST:
2952 /* destroy cmid on failure */
2953 rc = kiblnd_passive_connect(cmid,
2954 (void *)KIBLND_CONN_PARAM(event),
2955 KIBLND_CONN_PARAM_LEN(event));
2956 CDEBUG(D_NET, "connreq: %d\n", rc);
2959 case RDMA_CM_EVENT_ADDR_ERROR:
2960 peer_ni = (kib_peer_ni_t *)cmid->context;
2961 CNETERR("%s: ADDR ERROR %d\n",
2962 libcfs_nid2str(peer_ni->ibp_nid), event->status);
2963 kiblnd_peer_connect_failed(peer_ni, 1, -EHOSTUNREACH);
2964 kiblnd_peer_decref(peer_ni);
2965 return -EHOSTUNREACH; /* rc != 0 destroys cmid */
2967 case RDMA_CM_EVENT_ADDR_RESOLVED:
2968 peer_ni = (kib_peer_ni_t *)cmid->context;
2970 CDEBUG(D_NET,"%s Addr resolved: %d\n",
2971 libcfs_nid2str(peer_ni->ibp_nid), event->status);
2973 if (event->status != 0) {
2974 CNETERR("Can't resolve address for %s: %d\n",
2975 libcfs_nid2str(peer_ni->ibp_nid), event->status);
2978 rc = rdma_resolve_route(
2979 cmid, *kiblnd_tunables.kib_timeout * 1000);
2982 /* Can't initiate route resolution */
2983 CERROR("Can't resolve route for %s: %d\n",
2984 libcfs_nid2str(peer_ni->ibp_nid), rc);
2986 kiblnd_peer_connect_failed(peer_ni, 1, rc);
2987 kiblnd_peer_decref(peer_ni);
2988 return rc; /* rc != 0 destroys cmid */
2990 case RDMA_CM_EVENT_ROUTE_ERROR:
2991 peer_ni = (kib_peer_ni_t *)cmid->context;
2992 CNETERR("%s: ROUTE ERROR %d\n",
2993 libcfs_nid2str(peer_ni->ibp_nid), event->status);
2994 kiblnd_peer_connect_failed(peer_ni, 1, -EHOSTUNREACH);
2995 kiblnd_peer_decref(peer_ni);
2996 return -EHOSTUNREACH; /* rc != 0 destroys cmid */
2998 case RDMA_CM_EVENT_ROUTE_RESOLVED:
2999 peer_ni = (kib_peer_ni_t *)cmid->context;
3000 CDEBUG(D_NET,"%s Route resolved: %d\n",
3001 libcfs_nid2str(peer_ni->ibp_nid), event->status);
3003 if (event->status == 0)
3004 return kiblnd_active_connect(cmid);
3006 CNETERR("Can't resolve route for %s: %d\n",
3007 libcfs_nid2str(peer_ni->ibp_nid), event->status);
3008 kiblnd_peer_connect_failed(peer_ni, 1, event->status);
3009 kiblnd_peer_decref(peer_ni);
3010 return event->status; /* rc != 0 destroys cmid */
3012 case RDMA_CM_EVENT_UNREACHABLE:
3013 conn = (kib_conn_t *)cmid->context;
3014 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
3015 conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
3016 CNETERR("%s: UNREACHABLE %d\n",
3017 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
3018 kiblnd_connreq_done(conn, -ENETDOWN);
3019 kiblnd_conn_decref(conn);
3022 case RDMA_CM_EVENT_CONNECT_ERROR:
3023 conn = (kib_conn_t *)cmid->context;
3024 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
3025 conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
3026 CNETERR("%s: CONNECT ERROR %d\n",
3027 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
3028 kiblnd_connreq_done(conn, -ENOTCONN);
3029 kiblnd_conn_decref(conn);
3032 case RDMA_CM_EVENT_REJECTED:
3033 conn = (kib_conn_t *)cmid->context;
3034 switch (conn->ibc_state) {
3038 case IBLND_CONN_PASSIVE_WAIT:
3039 CERROR ("%s: REJECTED %d\n",
3040 libcfs_nid2str(conn->ibc_peer->ibp_nid),
3042 kiblnd_connreq_done(conn, -ECONNRESET);
3045 case IBLND_CONN_ACTIVE_CONNECT:
3046 kiblnd_rejected(conn, event->status,
3047 (void *)KIBLND_CONN_PARAM(event),
3048 KIBLND_CONN_PARAM_LEN(event));
3051 kiblnd_conn_decref(conn);
3054 case RDMA_CM_EVENT_ESTABLISHED:
3055 conn = (kib_conn_t *)cmid->context;
3056 switch (conn->ibc_state) {
3060 case IBLND_CONN_PASSIVE_WAIT:
3061 CDEBUG(D_NET, "ESTABLISHED (passive): %s\n",
3062 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3063 kiblnd_connreq_done(conn, 0);
3066 case IBLND_CONN_ACTIVE_CONNECT:
3067 CDEBUG(D_NET, "ESTABLISHED(active): %s\n",
3068 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3069 kiblnd_check_connreply(conn,
3070 (void *)KIBLND_CONN_PARAM(event),
3071 KIBLND_CONN_PARAM_LEN(event));
3074 /* net keeps its ref on conn! */
3077 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
3078 CDEBUG(D_NET, "Ignore TIMEWAIT_EXIT event\n");
3081 case RDMA_CM_EVENT_DISCONNECTED:
3082 conn = (kib_conn_t *)cmid->context;
3083 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
3084 CERROR("%s DISCONNECTED\n",
3085 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3086 kiblnd_connreq_done(conn, -ECONNRESET);
3088 kiblnd_close_conn(conn, 0);
3090 kiblnd_conn_decref(conn);
3091 cmid->context = NULL;
3094 case RDMA_CM_EVENT_DEVICE_REMOVAL:
3095 LCONSOLE_ERROR_MSG(0x131,
3096 "Received notification of device removal\n"
3097 "Please shutdown LNET to allow this to proceed\n");
3098 /* Can't remove network from underneath LNET for now, so I have
3102 case RDMA_CM_EVENT_ADDR_CHANGE:
3103 LCONSOLE_INFO("Physical link changed (eg hca/port)\n");
3109 kiblnd_check_txs_locked(kib_conn_t *conn, struct list_head *txs)
3112 struct list_head *ttmp;
3114 list_for_each(ttmp, txs) {
3115 tx = list_entry(ttmp, kib_tx_t, tx_list);
3117 if (txs != &conn->ibc_active_txs) {
3118 LASSERT(tx->tx_queued);
3120 LASSERT(!tx->tx_queued);
3121 LASSERT(tx->tx_waiting || tx->tx_sending != 0);
3124 if (cfs_time_aftereq(jiffies, tx->tx_deadline)) {
3125 CERROR("Timed out tx: %s, %lu seconds\n",
3126 kiblnd_queue2str(conn, txs),
3127 cfs_duration_sec(jiffies - tx->tx_deadline));
3136 kiblnd_conn_timed_out_locked(kib_conn_t *conn)
3138 return kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue) ||
3139 kiblnd_check_txs_locked(conn, &conn->ibc_tx_noops) ||
3140 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_rsrvd) ||
3141 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_nocred) ||
3142 kiblnd_check_txs_locked(conn, &conn->ibc_active_txs);
3146 kiblnd_check_conns (int idx)
3148 struct list_head closes = LIST_HEAD_INIT(closes);
3149 struct list_head checksends = LIST_HEAD_INIT(checksends);
3150 struct list_head *peers = &kiblnd_data.kib_peers[idx];
3151 struct list_head *ptmp;
3152 kib_peer_ni_t *peer_ni;
3154 struct list_head *ctmp;
3155 unsigned long flags;
3157 /* NB. We expect to have a look at all the peers and not find any
3158 * RDMAs to time out, so we just use a shared lock while we
3160 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3162 list_for_each(ptmp, peers) {
3163 peer_ni = list_entry(ptmp, kib_peer_ni_t, ibp_list);
3165 list_for_each(ctmp, &peer_ni->ibp_conns) {
3169 conn = list_entry(ctmp, kib_conn_t, ibc_list);
3171 LASSERT(conn->ibc_state == IBLND_CONN_ESTABLISHED);
3173 spin_lock(&conn->ibc_lock);
3175 sendnoop = kiblnd_need_noop(conn);
3176 timedout = kiblnd_conn_timed_out_locked(conn);
3177 if (!sendnoop && !timedout) {
3178 spin_unlock(&conn->ibc_lock);
3183 CERROR("Timed out RDMA with %s (%lu): "
3184 "c: %u, oc: %u, rc: %u\n",
3185 libcfs_nid2str(peer_ni->ibp_nid),
3186 cfs_duration_sec(cfs_time_current() -
3187 peer_ni->ibp_last_alive),
3189 conn->ibc_outstanding_credits,
3190 conn->ibc_reserved_credits);
3191 list_add(&conn->ibc_connd_list, &closes);
3193 list_add(&conn->ibc_connd_list, &checksends);
3195 /* +ref for 'closes' or 'checksends' */
3196 kiblnd_conn_addref(conn);
3198 spin_unlock(&conn->ibc_lock);
3202 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3204 /* Handle timeout by closing the whole
3205 * connection. We can only be sure RDMA activity
3206 * has ceased once the QP has been modified. */
3207 while (!list_empty(&closes)) {
3208 conn = list_entry(closes.next,
3209 kib_conn_t, ibc_connd_list);
3210 list_del(&conn->ibc_connd_list);
3211 kiblnd_close_conn(conn, -ETIMEDOUT);
3212 kiblnd_conn_decref(conn);
3215 /* In case we have enough credits to return via a
3216 * NOOP, but there were no non-blocking tx descs
3217 * free to do it last time... */
3218 while (!list_empty(&checksends)) {
3219 conn = list_entry(checksends.next,
3220 kib_conn_t, ibc_connd_list);
3221 list_del(&conn->ibc_connd_list);
3223 spin_lock(&conn->ibc_lock);
3224 kiblnd_check_sends_locked(conn);
3225 spin_unlock(&conn->ibc_lock);
3227 kiblnd_conn_decref(conn);
3232 kiblnd_disconnect_conn (kib_conn_t *conn)
3234 LASSERT (!in_interrupt());
3235 LASSERT (current == kiblnd_data.kib_connd);
3236 LASSERT (conn->ibc_state == IBLND_CONN_CLOSING);
3238 rdma_disconnect(conn->ibc_cmid);
3239 kiblnd_finalise_conn(conn);
3241 kiblnd_peer_notify(conn->ibc_peer);
3245 * High-water for reconnection to the same peer_ni, reconnection attempt should
3246 * be delayed after trying more than KIB_RECONN_HIGH_RACE.
3248 #define KIB_RECONN_HIGH_RACE 10
3250 * Allow connd to take a break and handle other things after consecutive
3251 * reconnection attemps.
3253 #define KIB_RECONN_BREAK 100
3256 kiblnd_connd (void *arg)
3258 spinlock_t *lock= &kiblnd_data.kib_connd_lock;
3260 unsigned long flags;
3266 unsigned long deadline = jiffies;
3268 cfs_block_allsigs();
3270 init_waitqueue_entry(&wait, current);
3271 kiblnd_data.kib_connd = current;
3273 spin_lock_irqsave(lock, flags);
3275 while (!kiblnd_data.kib_shutdown) {
3280 if (!list_empty(&kiblnd_data.kib_connd_zombies)) {
3281 kib_peer_ni_t *peer_ni = NULL;
3283 conn = list_entry(kiblnd_data.kib_connd_zombies.next,
3284 kib_conn_t, ibc_list);
3285 list_del(&conn->ibc_list);
3286 if (conn->ibc_reconnect) {
3287 peer_ni = conn->ibc_peer;
3288 kiblnd_peer_addref(peer_ni);
3291 spin_unlock_irqrestore(lock, flags);
3294 kiblnd_destroy_conn(conn, !peer_ni);
3296 spin_lock_irqsave(lock, flags);
3300 conn->ibc_peer = peer_ni;
3301 if (peer_ni->ibp_reconnected < KIB_RECONN_HIGH_RACE)
3302 list_add_tail(&conn->ibc_list,
3303 &kiblnd_data.kib_reconn_list);
3305 list_add_tail(&conn->ibc_list,
3306 &kiblnd_data.kib_reconn_wait);
3309 if (!list_empty(&kiblnd_data.kib_connd_conns)) {
3310 conn = list_entry(kiblnd_data.kib_connd_conns.next,
3311 kib_conn_t, ibc_list);
3312 list_del(&conn->ibc_list);
3314 spin_unlock_irqrestore(lock, flags);
3317 kiblnd_disconnect_conn(conn);
3318 kiblnd_conn_decref(conn);
3320 spin_lock_irqsave(lock, flags);
3323 while (reconn < KIB_RECONN_BREAK) {
3324 if (kiblnd_data.kib_reconn_sec !=
3325 ktime_get_real_seconds()) {
3326 kiblnd_data.kib_reconn_sec = ktime_get_real_seconds();
3327 list_splice_init(&kiblnd_data.kib_reconn_wait,
3328 &kiblnd_data.kib_reconn_list);
3331 if (list_empty(&kiblnd_data.kib_reconn_list))
3334 conn = list_entry(kiblnd_data.kib_reconn_list.next,
3335 kib_conn_t, ibc_list);
3336 list_del(&conn->ibc_list);
3338 spin_unlock_irqrestore(lock, flags);
3341 reconn += kiblnd_reconnect_peer(conn->ibc_peer);
3342 kiblnd_peer_decref(conn->ibc_peer);
3343 LIBCFS_FREE(conn, sizeof(*conn));
3345 spin_lock_irqsave(lock, flags);
3348 /* careful with the jiffy wrap... */
3349 timeout = (int)(deadline - jiffies);
3353 int chunk = kiblnd_data.kib_peer_hash_size;
3355 spin_unlock_irqrestore(lock, flags);
3358 /* Time to check for RDMA timeouts on a few more
3359 * peers: I do checks every 'p' seconds on a
3360 * proportion of the peer_ni table and I need to check
3361 * every connection 'n' times within a timeout
3362 * interval, to ensure I detect a timeout on any
3363 * connection within (n+1)/n times the timeout
3366 if (*kiblnd_tunables.kib_timeout > n * p)
3367 chunk = (chunk * n * p) /
3368 *kiblnd_tunables.kib_timeout;
3372 for (i = 0; i < chunk; i++) {
3373 kiblnd_check_conns(peer_index);
3374 peer_index = (peer_index + 1) %
3375 kiblnd_data.kib_peer_hash_size;
3378 deadline += msecs_to_jiffies(p * MSEC_PER_SEC);
3379 spin_lock_irqsave(lock, flags);
3385 /* Nothing to do for 'timeout' */
3386 set_current_state(TASK_INTERRUPTIBLE);
3387 add_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3388 spin_unlock_irqrestore(lock, flags);
3390 schedule_timeout(timeout);
3392 set_current_state(TASK_RUNNING);
3393 remove_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3394 spin_lock_irqsave(lock, flags);
3397 spin_unlock_irqrestore(lock, flags);
3399 kiblnd_thread_fini();
3404 kiblnd_qp_event(struct ib_event *event, void *arg)
3406 kib_conn_t *conn = arg;
3408 switch (event->event) {
3409 case IB_EVENT_COMM_EST:
3410 CDEBUG(D_NET, "%s established\n",
3411 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3412 /* We received a packet but connection isn't established
3413 * probably handshake packet was lost, so free to
3414 * force make connection established */
3415 rdma_notify(conn->ibc_cmid, IB_EVENT_COMM_EST);
3419 CERROR("%s: Async QP event type %d\n",
3420 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3426 kiblnd_complete (struct ib_wc *wc)
3428 switch (kiblnd_wreqid2type(wc->wr_id)) {
3433 if (wc->status != IB_WC_SUCCESS &&
3434 wc->status != IB_WC_WR_FLUSH_ERR)
3435 CNETERR("FastReg failed: %d\n", wc->status);
3438 case IBLND_WID_RDMA:
3439 /* We only get RDMA completion notification if it fails. All
3440 * subsequent work items, including the final SEND will fail
3441 * too. However we can't print out any more info about the
3442 * failing RDMA because 'tx' might be back on the idle list or
3443 * even reused already if we didn't manage to post all our work
3445 CNETERR("RDMA (tx: %p) failed: %d\n",
3446 kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3450 kiblnd_tx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3454 kiblnd_rx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status,
3461 kiblnd_cq_completion(struct ib_cq *cq, void *arg)
3463 /* NB I'm not allowed to schedule this conn once its refcount has
3464 * reached 0. Since fundamentally I'm racing with scheduler threads
3465 * consuming my CQ I could be called after all completions have
3466 * occurred. But in this case, ibc_nrx == 0 && ibc_nsends_posted == 0
3467 * and this CQ is about to be destroyed so I NOOP. */
3468 kib_conn_t *conn = (kib_conn_t *)arg;
3469 struct kib_sched_info *sched = conn->ibc_sched;
3470 unsigned long flags;
3472 LASSERT(cq == conn->ibc_cq);
3474 spin_lock_irqsave(&sched->ibs_lock, flags);
3476 conn->ibc_ready = 1;
3478 if (!conn->ibc_scheduled &&
3479 (conn->ibc_nrx > 0 ||
3480 conn->ibc_nsends_posted > 0)) {
3481 kiblnd_conn_addref(conn); /* +1 ref for sched_conns */
3482 conn->ibc_scheduled = 1;
3483 list_add_tail(&conn->ibc_sched_list, &sched->ibs_conns);
3485 if (waitqueue_active(&sched->ibs_waitq))
3486 wake_up(&sched->ibs_waitq);
3489 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3493 kiblnd_cq_event(struct ib_event *event, void *arg)
3495 kib_conn_t *conn = arg;
3497 CERROR("%s: async CQ event type %d\n",
3498 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3502 kiblnd_scheduler(void *arg)
3504 long id = (long)arg;
3505 struct kib_sched_info *sched;
3508 unsigned long flags;
3514 cfs_block_allsigs();
3516 init_waitqueue_entry(&wait, current);
3518 sched = kiblnd_data.kib_scheds[KIB_THREAD_CPT(id)];
3520 rc = cfs_cpt_bind(lnet_cpt_table(), sched->ibs_cpt);
3522 CWARN("Unable to bind on CPU partition %d, please verify "
3523 "whether all CPUs are healthy and reload modules if "
3524 "necessary, otherwise your system might under risk of "
3525 "low performance\n", sched->ibs_cpt);
3528 spin_lock_irqsave(&sched->ibs_lock, flags);
3530 while (!kiblnd_data.kib_shutdown) {
3531 if (busy_loops++ >= IBLND_RESCHED) {
3532 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3537 spin_lock_irqsave(&sched->ibs_lock, flags);
3542 if (!list_empty(&sched->ibs_conns)) {
3543 conn = list_entry(sched->ibs_conns.next,
3544 kib_conn_t, ibc_sched_list);
3545 /* take over kib_sched_conns' ref on conn... */
3546 LASSERT(conn->ibc_scheduled);
3547 list_del(&conn->ibc_sched_list);
3548 conn->ibc_ready = 0;
3550 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3552 wc.wr_id = IBLND_WID_INVAL;
3554 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3556 rc = ib_req_notify_cq(conn->ibc_cq,
3559 CWARN("%s: ib_req_notify_cq failed: %d, "
3560 "closing connection\n",
3561 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
3562 kiblnd_close_conn(conn, -EIO);
3563 kiblnd_conn_decref(conn);
3564 spin_lock_irqsave(&sched->ibs_lock,
3569 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3572 if (unlikely(rc > 0 && wc.wr_id == IBLND_WID_INVAL)) {
3574 "ib_poll_cq (rc: %d) returned invalid "
3575 "wr_id, opcode %d, status: %d, "
3576 "vendor_err: %d, conn: %s status: %d\n"
3577 "please upgrade firmware and OFED or "
3578 "contact vendor.\n", rc,
3579 wc.opcode, wc.status, wc.vendor_err,
3580 libcfs_nid2str(conn->ibc_peer->ibp_nid),
3586 CWARN("%s: ib_poll_cq failed: %d, "
3587 "closing connection\n",
3588 libcfs_nid2str(conn->ibc_peer->ibp_nid),
3590 kiblnd_close_conn(conn, -EIO);
3591 kiblnd_conn_decref(conn);
3592 spin_lock_irqsave(&sched->ibs_lock, flags);
3596 spin_lock_irqsave(&sched->ibs_lock, flags);
3598 if (rc != 0 || conn->ibc_ready) {
3599 /* There may be another completion waiting; get
3600 * another scheduler to check while I handle
3602 /* +1 ref for sched_conns */
3603 kiblnd_conn_addref(conn);
3604 list_add_tail(&conn->ibc_sched_list,
3606 if (waitqueue_active(&sched->ibs_waitq))
3607 wake_up(&sched->ibs_waitq);
3609 conn->ibc_scheduled = 0;
3613 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3614 kiblnd_complete(&wc);
3616 spin_lock_irqsave(&sched->ibs_lock, flags);
3619 kiblnd_conn_decref(conn); /* ...drop my ref from above */
3626 set_current_state(TASK_INTERRUPTIBLE);
3627 add_wait_queue_exclusive(&sched->ibs_waitq, &wait);
3628 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3633 remove_wait_queue(&sched->ibs_waitq, &wait);
3634 set_current_state(TASK_RUNNING);
3635 spin_lock_irqsave(&sched->ibs_lock, flags);
3638 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3640 kiblnd_thread_fini();
3645 kiblnd_failover_thread(void *arg)
3647 rwlock_t *glock = &kiblnd_data.kib_global_lock;
3650 unsigned long flags;
3653 LASSERT(*kiblnd_tunables.kib_dev_failover != 0);
3655 cfs_block_allsigs();
3657 init_waitqueue_entry(&wait, current);
3658 write_lock_irqsave(glock, flags);
3660 while (!kiblnd_data.kib_shutdown) {
3661 int do_failover = 0;
3664 list_for_each_entry(dev, &kiblnd_data.kib_failed_devs,
3666 if (cfs_time_before(cfs_time_current(),
3667 dev->ibd_next_failover))
3674 list_del_init(&dev->ibd_fail_list);
3675 dev->ibd_failover = 1;
3676 write_unlock_irqrestore(glock, flags);
3678 rc = kiblnd_dev_failover(dev);
3680 write_lock_irqsave(glock, flags);
3682 LASSERT (dev->ibd_failover);
3683 dev->ibd_failover = 0;
3684 if (rc >= 0) { /* Device is OK or failover succeed */
3685 dev->ibd_next_failover = cfs_time_shift(3);
3689 /* failed to failover, retry later */
3690 dev->ibd_next_failover =
3691 cfs_time_shift(min(dev->ibd_failed_failover, 10));
3692 if (kiblnd_dev_can_failover(dev)) {
3693 list_add_tail(&dev->ibd_fail_list,
3694 &kiblnd_data.kib_failed_devs);
3700 /* long sleep if no more pending failover */
3701 long_sleep = list_empty(&kiblnd_data.kib_failed_devs);
3703 set_current_state(TASK_INTERRUPTIBLE);
3704 add_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3705 write_unlock_irqrestore(glock, flags);
3707 rc = schedule_timeout(long_sleep ? cfs_time_seconds(10) :
3708 cfs_time_seconds(1));
3709 set_current_state(TASK_RUNNING);
3710 remove_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3711 write_lock_irqsave(glock, flags);
3713 if (!long_sleep || rc != 0)
3716 /* have a long sleep, routine check all active devices,
3717 * we need checking like this because if there is not active
3718 * connection on the dev and no SEND from local, we may listen
3719 * on wrong HCA for ever while there is a bonding failover */
3720 list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3721 if (kiblnd_dev_can_failover(dev)) {
3722 list_add_tail(&dev->ibd_fail_list,
3723 &kiblnd_data.kib_failed_devs);
3728 write_unlock_irqrestore(glock, flags);
3730 kiblnd_thread_fini();