4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2015, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lnet/klnds/o2iblnd/o2iblnd_cb.c
34 * Author: Eric Barton <eric@bartonsoftware.com>
39 #define MAX_CONN_RACES_BEFORE_ABORT 20
41 static void kiblnd_peer_alive(struct kib_peer *peer);
42 static void kiblnd_peer_connect_failed(struct kib_peer *peer, int active, int error);
43 static void kiblnd_init_tx_msg(struct lnet_ni *ni, struct kib_tx *tx,
44 int type, int body_nob);
45 static int kiblnd_init_rdma(struct kib_conn *conn, struct kib_tx *tx, int type,
46 int resid, struct kib_rdma_desc *dstrd,
48 static void kiblnd_queue_tx_locked(struct kib_tx *tx, struct kib_conn *conn);
49 static void kiblnd_queue_tx(struct kib_tx *tx, struct kib_conn *conn);
50 static void kiblnd_unmap_tx(struct lnet_ni *ni, struct kib_tx *tx);
51 static void kiblnd_check_sends_locked(struct kib_conn *conn);
54 kiblnd_tx_done(struct lnet_ni *ni, struct kib_tx *tx)
56 struct lnet_msg *lntmsg[2];
57 struct kib_net *net = ni->ni_data;
62 LASSERT(!in_interrupt());
63 LASSERT(!tx->tx_queued); /* mustn't be queued for sending */
64 LASSERT(!tx->tx_sending); /* mustn't be awaiting sent callback */
65 LASSERT(!tx->tx_waiting); /* mustn't be awaiting peer response */
68 kiblnd_unmap_tx(ni, tx);
70 /* tx may have up to 2 lnet msgs to finalise */
71 lntmsg[0] = tx->tx_lntmsg[0]; tx->tx_lntmsg[0] = NULL;
72 lntmsg[1] = tx->tx_lntmsg[1]; tx->tx_lntmsg[1] = NULL;
76 LASSERT(ni == tx->tx_conn->ibc_peer->ibp_ni);
78 kiblnd_conn_decref(tx->tx_conn);
85 kiblnd_pool_free_node(&tx->tx_pool->tpo_pool, &tx->tx_list);
87 /* delay finalize until my descs have been freed */
88 for (i = 0; i < 2; i++) {
92 lnet_finalize(ni, lntmsg[i], rc);
97 kiblnd_txlist_done(struct lnet_ni *ni, struct list_head *txlist, int status)
101 while (!list_empty(txlist)) {
102 tx = list_entry(txlist->next, struct kib_tx, tx_list);
104 list_del(&tx->tx_list);
107 tx->tx_status = status;
108 kiblnd_tx_done(ni, tx);
112 static struct kib_tx *
113 kiblnd_get_idle_tx(struct lnet_ni *ni, lnet_nid_t target)
115 struct kib_net *net = (struct kib_net *)ni->ni_data;
116 struct list_head *node;
118 struct kib_tx_poolset *tps;
120 tps = net->ibn_tx_ps[lnet_cpt_of_nid(target)];
121 node = kiblnd_pool_alloc_node(&tps->tps_poolset);
124 tx = list_entry(node, struct kib_tx, tx_list);
126 LASSERT(!tx->tx_nwrq);
127 LASSERT(!tx->tx_queued);
128 LASSERT(!tx->tx_sending);
129 LASSERT(!tx->tx_waiting);
130 LASSERT(!tx->tx_status);
131 LASSERT(!tx->tx_conn);
132 LASSERT(!tx->tx_lntmsg[0]);
133 LASSERT(!tx->tx_lntmsg[1]);
134 LASSERT(!tx->tx_nfrags);
140 kiblnd_drop_rx(struct kib_rx *rx)
142 struct kib_conn *conn = rx->rx_conn;
143 struct kib_sched_info *sched = conn->ibc_sched;
146 spin_lock_irqsave(&sched->ibs_lock, flags);
147 LASSERT(conn->ibc_nrx > 0);
149 spin_unlock_irqrestore(&sched->ibs_lock, flags);
151 kiblnd_conn_decref(conn);
155 kiblnd_post_rx(struct kib_rx *rx, int credit)
157 struct kib_conn *conn = rx->rx_conn;
158 struct kib_net *net = conn->ibc_peer->ibp_ni->ni_data;
159 struct ib_recv_wr *bad_wrq = NULL;
163 LASSERT(!in_interrupt());
164 LASSERT(credit == IBLND_POSTRX_NO_CREDIT ||
165 credit == IBLND_POSTRX_PEER_CREDIT ||
166 credit == IBLND_POSTRX_RSRVD_CREDIT);
168 rx->rx_sge.lkey = conn->ibc_hdev->ibh_pd->local_dma_lkey;
169 rx->rx_sge.addr = rx->rx_msgaddr;
170 rx->rx_sge.length = IBLND_MSG_SIZE;
172 rx->rx_wrq.next = NULL;
173 rx->rx_wrq.sg_list = &rx->rx_sge;
174 rx->rx_wrq.num_sge = 1;
175 rx->rx_wrq.wr_id = kiblnd_ptr2wreqid(rx, IBLND_WID_RX);
177 LASSERT(conn->ibc_state >= IBLND_CONN_INIT);
178 LASSERT(rx->rx_nob >= 0); /* not posted */
180 if (conn->ibc_state > IBLND_CONN_ESTABLISHED) {
181 kiblnd_drop_rx(rx); /* No more posts for this rx */
185 rx->rx_nob = -1; /* flag posted */
187 /* NB: need an extra reference after ib_post_recv because we don't
188 * own this rx (and rx::rx_conn) anymore, LU-5678.
190 kiblnd_conn_addref(conn);
191 rc = ib_post_recv(conn->ibc_cmid->qp, &rx->rx_wrq, &bad_wrq);
193 CERROR("Can't post rx for %s: %d, bad_wrq: %p\n",
194 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc, bad_wrq);
198 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) /* Initial post */
202 kiblnd_close_conn(conn, rc);
203 kiblnd_drop_rx(rx); /* No more posts for this rx */
207 if (credit == IBLND_POSTRX_NO_CREDIT)
210 spin_lock(&conn->ibc_lock);
211 if (credit == IBLND_POSTRX_PEER_CREDIT)
212 conn->ibc_outstanding_credits++;
214 conn->ibc_reserved_credits++;
215 kiblnd_check_sends_locked(conn);
216 spin_unlock(&conn->ibc_lock);
219 kiblnd_conn_decref(conn);
223 static struct kib_tx *
224 kiblnd_find_waiting_tx_locked(struct kib_conn *conn, int txtype, __u64 cookie)
226 struct list_head *tmp;
228 list_for_each(tmp, &conn->ibc_active_txs) {
229 struct kib_tx *tx = list_entry(tmp, struct kib_tx, tx_list);
231 LASSERT(!tx->tx_queued);
232 LASSERT(tx->tx_sending || tx->tx_waiting);
234 if (tx->tx_cookie != cookie)
237 if (tx->tx_waiting &&
238 tx->tx_msg->ibm_type == txtype)
241 CWARN("Bad completion: %swaiting, type %x (wanted %x)\n",
242 tx->tx_waiting ? "" : "NOT ",
243 tx->tx_msg->ibm_type, txtype);
249 kiblnd_handle_completion(struct kib_conn *conn, int txtype, int status, __u64 cookie)
252 struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
255 spin_lock(&conn->ibc_lock);
257 tx = kiblnd_find_waiting_tx_locked(conn, txtype, cookie);
259 spin_unlock(&conn->ibc_lock);
261 CWARN("Unmatched completion type %x cookie %#llx from %s\n",
262 txtype, cookie, libcfs_nid2str(conn->ibc_peer->ibp_nid));
263 kiblnd_close_conn(conn, -EPROTO);
267 if (!tx->tx_status) { /* success so far */
268 if (status < 0) /* failed? */
269 tx->tx_status = status;
270 else if (txtype == IBLND_MSG_GET_REQ)
271 lnet_set_reply_msg_len(ni, tx->tx_lntmsg[1], status);
276 idle = !tx->tx_queued && !tx->tx_sending;
278 list_del(&tx->tx_list);
280 spin_unlock(&conn->ibc_lock);
283 kiblnd_tx_done(ni, tx);
287 kiblnd_send_completion(struct kib_conn *conn, int type, int status, __u64 cookie)
289 struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
290 struct kib_tx *tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
293 CERROR("Can't get tx for completion %x for %s\n",
294 type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
298 tx->tx_msg->ibm_u.completion.ibcm_status = status;
299 tx->tx_msg->ibm_u.completion.ibcm_cookie = cookie;
300 kiblnd_init_tx_msg(ni, tx, type, sizeof(struct kib_completion_msg));
302 kiblnd_queue_tx(tx, conn);
306 kiblnd_handle_rx(struct kib_rx *rx)
308 struct kib_msg *msg = rx->rx_msg;
309 struct kib_conn *conn = rx->rx_conn;
310 struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
311 int credits = msg->ibm_credits;
317 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
319 CDEBUG(D_NET, "Received %x[%d] from %s\n",
320 msg->ibm_type, credits,
321 libcfs_nid2str(conn->ibc_peer->ibp_nid));
324 /* Have I received credits that will let me send? */
325 spin_lock(&conn->ibc_lock);
327 if (conn->ibc_credits + credits >
328 conn->ibc_queue_depth) {
329 rc2 = conn->ibc_credits;
330 spin_unlock(&conn->ibc_lock);
332 CERROR("Bad credits from %s: %d + %d > %d\n",
333 libcfs_nid2str(conn->ibc_peer->ibp_nid),
334 rc2, credits, 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) /* 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);
416 * CAVEAT EMPTOR: I could be racing with tx_complete, but...
417 * (a) I can overwrite tx_msg since my peer has received it!
418 * (b) tx_waiting set tells tx_complete() it's not done.
420 tx->tx_nwrq = 0; /* overwrite PUT_REQ */
422 rc2 = kiblnd_init_rdma(conn, tx, IBLND_MSG_PUT_DONE,
423 kiblnd_rd_size(&msg->ibm_u.putack.ibpam_rd),
424 &msg->ibm_u.putack.ibpam_rd,
425 msg->ibm_u.putack.ibpam_dst_cookie);
427 CERROR("Can't setup rdma for PUT to %s: %d\n",
428 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc2);
430 spin_lock(&conn->ibc_lock);
431 tx->tx_waiting = 0; /* clear waiting and queue atomically */
432 kiblnd_queue_tx_locked(tx, conn);
433 spin_unlock(&conn->ibc_lock);
436 case IBLND_MSG_PUT_DONE:
437 post_credit = IBLND_POSTRX_PEER_CREDIT;
438 kiblnd_handle_completion(conn, IBLND_MSG_PUT_ACK,
439 msg->ibm_u.completion.ibcm_status,
440 msg->ibm_u.completion.ibcm_cookie);
443 case IBLND_MSG_GET_REQ:
444 post_credit = IBLND_POSTRX_DONT_POST;
445 rc = lnet_parse(ni, &msg->ibm_u.get.ibgm_hdr,
446 msg->ibm_srcnid, rx, 1);
447 if (rc < 0) /* repost on error */
448 post_credit = IBLND_POSTRX_PEER_CREDIT;
451 case IBLND_MSG_GET_DONE:
452 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
453 kiblnd_handle_completion(conn, IBLND_MSG_GET_REQ,
454 msg->ibm_u.completion.ibcm_status,
455 msg->ibm_u.completion.ibcm_cookie);
459 if (rc < 0) /* protocol error */
460 kiblnd_close_conn(conn, rc);
462 if (post_credit != IBLND_POSTRX_DONT_POST)
463 kiblnd_post_rx(rx, post_credit);
467 kiblnd_rx_complete(struct kib_rx *rx, int status, int nob)
469 struct kib_msg *msg = rx->rx_msg;
470 struct kib_conn *conn = rx->rx_conn;
471 struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
472 struct kib_net *net = ni->ni_data;
477 LASSERT(rx->rx_nob < 0); /* was posted */
478 rx->rx_nob = 0; /* isn't now */
480 if (conn->ibc_state > IBLND_CONN_ESTABLISHED)
483 if (status != IB_WC_SUCCESS) {
484 CNETERR("Rx from %s failed: %d\n",
485 libcfs_nid2str(conn->ibc_peer->ibp_nid), status);
492 rc = kiblnd_unpack_msg(msg, rx->rx_nob);
494 CERROR("Error %d unpacking rx from %s\n",
495 rc, libcfs_nid2str(conn->ibc_peer->ibp_nid));
499 if (msg->ibm_srcnid != conn->ibc_peer->ibp_nid ||
500 msg->ibm_dstnid != ni->ni_nid ||
501 msg->ibm_srcstamp != conn->ibc_incarnation ||
502 msg->ibm_dststamp != net->ibn_incarnation) {
503 CERROR("Stale rx from %s\n",
504 libcfs_nid2str(conn->ibc_peer->ibp_nid));
509 /* set time last known alive */
510 kiblnd_peer_alive(conn->ibc_peer);
512 /* racing with connection establishment/teardown! */
514 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
515 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
518 write_lock_irqsave(g_lock, flags);
519 /* must check holding global lock to eliminate race */
520 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
521 list_add_tail(&rx->rx_list, &conn->ibc_early_rxs);
522 write_unlock_irqrestore(g_lock, flags);
525 write_unlock_irqrestore(g_lock, flags);
527 kiblnd_handle_rx(rx);
531 CDEBUG(D_NET, "rx %p conn %p\n", rx, conn);
532 kiblnd_close_conn(conn, err);
534 kiblnd_drop_rx(rx); /* Don't re-post rx. */
538 kiblnd_kvaddr_to_page(unsigned long vaddr)
542 if (is_vmalloc_addr((void *)vaddr)) {
543 page = vmalloc_to_page((void *)vaddr);
547 #ifdef CONFIG_HIGHMEM
548 if (vaddr >= PKMAP_BASE &&
549 vaddr < (PKMAP_BASE + LAST_PKMAP * PAGE_SIZE)) {
550 /* No highmem pages only used for bulk (kiov) I/O */
551 CERROR("find page for address in highmem\n");
555 page = virt_to_page(vaddr);
561 kiblnd_fmr_map_tx(struct kib_net *net, struct kib_tx *tx, struct kib_rdma_desc *rd, __u32 nob)
563 struct kib_hca_dev *hdev;
564 struct kib_fmr_poolset *fps;
568 LASSERT(tx->tx_pool);
569 LASSERT(tx->tx_pool->tpo_pool.po_owner);
571 hdev = tx->tx_pool->tpo_hdev;
572 cpt = tx->tx_pool->tpo_pool.po_owner->ps_cpt;
574 fps = net->ibn_fmr_ps[cpt];
575 rc = kiblnd_fmr_pool_map(fps, tx, rd, nob, 0, &tx->fmr);
577 CERROR("Can't map %u bytes: %d\n", nob, rc);
582 * If rd is not tx_rd, it's going to get sent to a peer, who will need
585 rd->rd_key = tx->fmr.fmr_key;
586 rd->rd_frags[0].rf_addr &= ~hdev->ibh_page_mask;
587 rd->rd_frags[0].rf_nob = nob;
593 static void kiblnd_unmap_tx(struct lnet_ni *ni, struct kib_tx *tx)
595 struct kib_net *net = ni->ni_data;
600 kiblnd_fmr_pool_unmap(&tx->fmr, tx->tx_status);
603 kiblnd_dma_unmap_sg(tx->tx_pool->tpo_hdev->ibh_ibdev,
604 tx->tx_frags, tx->tx_nfrags, tx->tx_dmadir);
609 static int kiblnd_map_tx(struct lnet_ni *ni, struct kib_tx *tx,
610 struct kib_rdma_desc *rd, int nfrags)
612 struct kib_net *net = ni->ni_data;
613 struct kib_hca_dev *hdev = net->ibn_dev->ibd_hdev;
618 * If rd is not tx_rd, it's going to get sent to a peer and I'm the
621 tx->tx_dmadir = (rd != tx->tx_rd) ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
622 tx->tx_nfrags = nfrags;
624 rd->rd_nfrags = kiblnd_dma_map_sg(hdev->ibh_ibdev, tx->tx_frags,
625 tx->tx_nfrags, tx->tx_dmadir);
627 for (i = 0, nob = 0; i < rd->rd_nfrags; i++) {
628 rd->rd_frags[i].rf_nob = kiblnd_sg_dma_len(
629 hdev->ibh_ibdev, &tx->tx_frags[i]);
630 rd->rd_frags[i].rf_addr = kiblnd_sg_dma_address(
631 hdev->ibh_ibdev, &tx->tx_frags[i]);
632 nob += rd->rd_frags[i].rf_nob;
636 return kiblnd_fmr_map_tx(net, tx, rd, nob);
642 kiblnd_setup_rd_iov(struct lnet_ni *ni, struct kib_tx *tx,
643 struct kib_rdma_desc *rd, unsigned int niov,
644 const struct kvec *iov, int offset, int nob)
646 struct kib_net *net = ni->ni_data;
648 struct scatterlist *sg;
657 while (offset >= iov->iov_len) {
658 offset -= iov->iov_len;
668 vaddr = ((unsigned long)iov->iov_base) + offset;
669 page_offset = vaddr & (PAGE_SIZE - 1);
670 page = kiblnd_kvaddr_to_page(vaddr);
672 CERROR("Can't find page\n");
676 fragnob = min((int)(iov->iov_len - offset), nob);
677 fragnob = min(fragnob, (int)PAGE_SIZE - page_offset);
679 sg_set_page(sg, page, fragnob, page_offset);
682 CERROR("lacking enough sg entries to map tx\n");
686 if (offset + fragnob < iov->iov_len) {
696 return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags);
700 kiblnd_setup_rd_kiov(struct lnet_ni *ni, struct kib_tx *tx,
701 struct kib_rdma_desc *rd, int nkiov,
702 const struct bio_vec *kiov, int offset, int nob)
704 struct kib_net *net = ni->ni_data;
705 struct scatterlist *sg;
708 CDEBUG(D_NET, "niov %d offset %d nob %d\n", nkiov, offset, nob);
714 while (offset >= kiov->bv_len) {
715 offset -= kiov->bv_len;
725 fragnob = min((int)(kiov->bv_len - offset), nob);
727 sg_set_page(sg, kiov->bv_page, fragnob,
728 kiov->bv_offset + offset);
731 CERROR("lacking enough sg entries to map tx\n");
741 return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags);
745 kiblnd_post_tx_locked(struct kib_conn *conn, struct kib_tx *tx, int credit)
746 __must_hold(&conn->ibc_lock)
748 struct kib_msg *msg = tx->tx_msg;
749 struct kib_peer *peer = conn->ibc_peer;
750 struct lnet_ni *ni = peer->ibp_ni;
751 int ver = conn->ibc_version;
755 LASSERT(tx->tx_queued);
756 /* We rely on this for QP sizing */
757 LASSERT(tx->tx_nwrq > 0);
759 LASSERT(!credit || credit == 1);
760 LASSERT(conn->ibc_outstanding_credits >= 0);
761 LASSERT(conn->ibc_outstanding_credits <= conn->ibc_queue_depth);
762 LASSERT(conn->ibc_credits >= 0);
763 LASSERT(conn->ibc_credits <= conn->ibc_queue_depth);
765 if (conn->ibc_nsends_posted == kiblnd_concurrent_sends(ver, ni)) {
766 /* tx completions outstanding... */
767 CDEBUG(D_NET, "%s: posted enough\n",
768 libcfs_nid2str(peer->ibp_nid));
772 if (credit && !conn->ibc_credits) { /* no credits */
773 CDEBUG(D_NET, "%s: no credits\n",
774 libcfs_nid2str(peer->ibp_nid));
778 if (credit && !IBLND_OOB_CAPABLE(ver) &&
779 conn->ibc_credits == 1 && /* last credit reserved */
780 msg->ibm_type != IBLND_MSG_NOOP) { /* for NOOP */
781 CDEBUG(D_NET, "%s: not using last credit\n",
782 libcfs_nid2str(peer->ibp_nid));
786 /* NB don't drop ibc_lock before bumping tx_sending */
787 list_del(&tx->tx_list);
790 if (msg->ibm_type == IBLND_MSG_NOOP &&
791 (!kiblnd_need_noop(conn) || /* redundant NOOP */
792 (IBLND_OOB_CAPABLE(ver) && /* posted enough NOOP */
793 conn->ibc_noops_posted == IBLND_OOB_MSGS(ver)))) {
795 * OK to drop when posted enough NOOPs, since
796 * kiblnd_check_sends_locked will queue NOOP again when
797 * posted NOOPs complete
799 spin_unlock(&conn->ibc_lock);
800 kiblnd_tx_done(peer->ibp_ni, tx);
801 spin_lock(&conn->ibc_lock);
802 CDEBUG(D_NET, "%s(%d): redundant or enough NOOP\n",
803 libcfs_nid2str(peer->ibp_nid),
804 conn->ibc_noops_posted);
808 kiblnd_pack_msg(peer->ibp_ni, msg, ver, conn->ibc_outstanding_credits,
809 peer->ibp_nid, conn->ibc_incarnation);
811 conn->ibc_credits -= credit;
812 conn->ibc_outstanding_credits = 0;
813 conn->ibc_nsends_posted++;
814 if (msg->ibm_type == IBLND_MSG_NOOP)
815 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.
826 list_add(&tx->tx_list, &conn->ibc_active_txs);
828 /* I'm still holding ibc_lock! */
829 if (conn->ibc_state != IBLND_CONN_ESTABLISHED) {
831 } else if (tx->tx_pool->tpo_pool.po_failed ||
832 conn->ibc_hdev != tx->tx_pool->tpo_hdev) {
833 /* close_conn will launch failover */
836 struct kib_fast_reg_descriptor *frd = tx->fmr.fmr_frd;
837 struct ib_send_wr *bad = &tx->tx_wrq[tx->tx_nwrq - 1].wr;
838 struct ib_send_wr *wrq = &tx->tx_wrq[0].wr;
841 if (!frd->frd_valid) {
842 wrq = &frd->frd_inv_wr;
843 wrq->next = &frd->frd_fastreg_wr.wr;
845 wrq = &frd->frd_fastreg_wr.wr;
847 frd->frd_fastreg_wr.wr.next = &tx->tx_wrq[0].wr;
850 LASSERTF(bad->wr_id == kiblnd_ptr2wreqid(tx, IBLND_WID_TX),
851 "bad wr_id %llx, opc %d, flags %d, peer: %s\n",
852 bad->wr_id, bad->opcode, bad->send_flags,
853 libcfs_nid2str(conn->ibc_peer->ibp_nid));
855 rc = ib_post_send(conn->ibc_cmid->qp, wrq, &bad);
858 conn->ibc_last_send = jiffies;
864 * NB credits are transferred in the actual
865 * message, which can only be the last work item
867 conn->ibc_credits += credit;
868 conn->ibc_outstanding_credits += msg->ibm_credits;
869 conn->ibc_nsends_posted--;
870 if (msg->ibm_type == IBLND_MSG_NOOP)
871 conn->ibc_noops_posted--;
877 done = !tx->tx_sending;
879 list_del(&tx->tx_list);
881 spin_unlock(&conn->ibc_lock);
883 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
884 CERROR("Error %d posting transmit to %s\n",
885 rc, libcfs_nid2str(peer->ibp_nid));
887 CDEBUG(D_NET, "Error %d posting transmit to %s\n",
888 rc, libcfs_nid2str(peer->ibp_nid));
890 kiblnd_close_conn(conn, rc);
893 kiblnd_tx_done(peer->ibp_ni, tx);
895 spin_lock(&conn->ibc_lock);
901 kiblnd_check_sends_locked(struct kib_conn *conn)
903 int ver = conn->ibc_version;
904 struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
907 /* Don't send anything until after the connection is established */
908 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
909 CDEBUG(D_NET, "%s too soon\n",
910 libcfs_nid2str(conn->ibc_peer->ibp_nid));
914 LASSERT(conn->ibc_nsends_posted <= kiblnd_concurrent_sends(ver, ni));
915 LASSERT(!IBLND_OOB_CAPABLE(ver) ||
916 conn->ibc_noops_posted <= IBLND_OOB_MSGS(ver));
917 LASSERT(conn->ibc_reserved_credits >= 0);
919 while (conn->ibc_reserved_credits > 0 &&
920 !list_empty(&conn->ibc_tx_queue_rsrvd)) {
921 tx = list_entry(conn->ibc_tx_queue_rsrvd.next,
922 struct kib_tx, tx_list);
923 list_del(&tx->tx_list);
924 list_add_tail(&tx->tx_list, &conn->ibc_tx_queue);
925 conn->ibc_reserved_credits--;
928 if (kiblnd_need_noop(conn)) {
929 spin_unlock(&conn->ibc_lock);
931 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
933 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_NOOP, 0);
935 spin_lock(&conn->ibc_lock);
937 kiblnd_queue_tx_locked(tx, conn);
943 if (!list_empty(&conn->ibc_tx_queue_nocred)) {
945 tx = list_entry(conn->ibc_tx_queue_nocred.next,
946 struct kib_tx, tx_list);
947 } else if (!list_empty(&conn->ibc_tx_noops)) {
948 LASSERT(!IBLND_OOB_CAPABLE(ver));
950 tx = list_entry(conn->ibc_tx_noops.next,
951 struct kib_tx, tx_list);
952 } else if (!list_empty(&conn->ibc_tx_queue)) {
954 tx = list_entry(conn->ibc_tx_queue.next,
955 struct kib_tx, tx_list);
960 if (kiblnd_post_tx_locked(conn, tx, credit))
966 kiblnd_tx_complete(struct kib_tx *tx, int status)
968 int failed = (status != IB_WC_SUCCESS);
969 struct kib_conn *conn = tx->tx_conn;
972 LASSERT(tx->tx_sending > 0);
975 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
976 CNETERR("Tx -> %s cookie %#llx sending %d waiting %d: failed %d\n",
977 libcfs_nid2str(conn->ibc_peer->ibp_nid),
978 tx->tx_cookie, tx->tx_sending, tx->tx_waiting,
981 kiblnd_close_conn(conn, -EIO);
983 kiblnd_peer_alive(conn->ibc_peer);
986 spin_lock(&conn->ibc_lock);
989 * I could be racing with rdma completion. Whoever makes 'tx' idle
990 * gets to free it, which also drops its ref on 'conn'.
993 conn->ibc_nsends_posted--;
994 if (tx->tx_msg->ibm_type == IBLND_MSG_NOOP)
995 conn->ibc_noops_posted--;
998 tx->tx_waiting = 0; /* don't wait for peer */
999 tx->tx_status = -EIO;
1002 idle = !tx->tx_sending && /* This is the final callback */
1003 !tx->tx_waiting && /* Not waiting for peer */
1004 !tx->tx_queued; /* Not re-queued (PUT_DONE) */
1006 list_del(&tx->tx_list);
1008 kiblnd_check_sends_locked(conn);
1009 spin_unlock(&conn->ibc_lock);
1012 kiblnd_tx_done(conn->ibc_peer->ibp_ni, tx);
1016 kiblnd_init_tx_msg(struct lnet_ni *ni, struct kib_tx *tx, int type,
1019 struct kib_hca_dev *hdev = tx->tx_pool->tpo_hdev;
1020 struct ib_sge *sge = &tx->tx_sge[tx->tx_nwrq];
1021 struct ib_rdma_wr *wrq = &tx->tx_wrq[tx->tx_nwrq];
1022 int nob = offsetof(struct kib_msg, ibm_u) + body_nob;
1024 LASSERT(tx->tx_nwrq >= 0);
1025 LASSERT(tx->tx_nwrq < IBLND_MAX_RDMA_FRAGS + 1);
1026 LASSERT(nob <= IBLND_MSG_SIZE);
1028 kiblnd_init_msg(tx->tx_msg, type, body_nob);
1030 sge->lkey = hdev->ibh_pd->local_dma_lkey;
1031 sge->addr = tx->tx_msgaddr;
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(struct kib_conn *conn, struct kib_tx *tx, int type,
1048 int resid, struct kib_rdma_desc *dstrd, __u64 dstcookie)
1050 struct kib_msg *ibmsg = tx->tx_msg;
1051 struct kib_rdma_desc *srcrd = tx->tx_rd;
1052 struct ib_sge *sge = &tx->tx_sge[0];
1053 struct ib_rdma_wr *wrq, *next;
1059 LASSERT(!in_interrupt());
1060 LASSERT(!tx->tx_nwrq);
1061 LASSERT(type == IBLND_MSG_GET_DONE ||
1062 type == IBLND_MSG_PUT_DONE);
1064 if (kiblnd_rd_size(srcrd) > conn->ibc_max_frags << PAGE_SHIFT) {
1065 CERROR("RDMA is too large for peer %s (%d), src size: %d dst size: %d\n",
1066 libcfs_nid2str(conn->ibc_peer->ibp_nid),
1067 conn->ibc_max_frags << PAGE_SHIFT,
1068 kiblnd_rd_size(srcrd), kiblnd_rd_size(dstrd));
1074 if (srcidx >= srcrd->rd_nfrags) {
1075 CERROR("Src buffer exhausted: %d frags\n", srcidx);
1080 if (dstidx == dstrd->rd_nfrags) {
1081 CERROR("Dst buffer exhausted: %d frags\n", dstidx);
1086 if (tx->tx_nwrq >= IBLND_MAX_RDMA_FRAGS) {
1087 CERROR("RDMA has too many fragments for peer %s (%d), src idx/frags: %d/%d dst idx/frags: %d/%d\n",
1088 libcfs_nid2str(conn->ibc_peer->ibp_nid),
1089 IBLND_MAX_RDMA_FRAGS,
1090 srcidx, srcrd->rd_nfrags,
1091 dstidx, dstrd->rd_nfrags);
1096 wrknob = min3(kiblnd_rd_frag_size(srcrd, srcidx),
1097 kiblnd_rd_frag_size(dstrd, dstidx),
1100 sge = &tx->tx_sge[tx->tx_nwrq];
1101 sge->addr = kiblnd_rd_frag_addr(srcrd, srcidx);
1102 sge->lkey = kiblnd_rd_frag_key(srcrd, srcidx);
1103 sge->length = wrknob;
1105 wrq = &tx->tx_wrq[tx->tx_nwrq];
1108 wrq->wr.next = &next->wr;
1109 wrq->wr.wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_RDMA);
1110 wrq->wr.sg_list = sge;
1111 wrq->wr.num_sge = 1;
1112 wrq->wr.opcode = IB_WR_RDMA_WRITE;
1113 wrq->wr.send_flags = 0;
1115 wrq->remote_addr = kiblnd_rd_frag_addr(dstrd, dstidx);
1116 wrq->rkey = kiblnd_rd_frag_key(dstrd, dstidx);
1118 srcidx = kiblnd_rd_consume_frag(srcrd, srcidx, wrknob);
1119 dstidx = kiblnd_rd_consume_frag(dstrd, dstidx, wrknob);
1128 if (rc < 0) /* no RDMA if completing with failure */
1131 ibmsg->ibm_u.completion.ibcm_status = rc;
1132 ibmsg->ibm_u.completion.ibcm_cookie = dstcookie;
1133 kiblnd_init_tx_msg(conn->ibc_peer->ibp_ni, tx,
1134 type, sizeof(struct kib_completion_msg));
1140 kiblnd_queue_tx_locked(struct kib_tx *tx, struct kib_conn *conn)
1142 struct list_head *q;
1144 LASSERT(tx->tx_nwrq > 0); /* work items set up */
1145 LASSERT(!tx->tx_queued); /* not queued for sending already */
1146 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1149 tx->tx_deadline = jiffies +
1150 msecs_to_jiffies(*kiblnd_tunables.kib_timeout *
1154 kiblnd_conn_addref(conn);
1156 LASSERT(tx->tx_msg->ibm_type != IBLND_MSG_PUT_DONE);
1158 /* PUT_DONE first attached to conn as a PUT_REQ */
1159 LASSERT(tx->tx_conn == conn);
1160 LASSERT(tx->tx_msg->ibm_type == IBLND_MSG_PUT_DONE);
1163 switch (tx->tx_msg->ibm_type) {
1167 case IBLND_MSG_PUT_REQ:
1168 case IBLND_MSG_GET_REQ:
1169 q = &conn->ibc_tx_queue_rsrvd;
1172 case IBLND_MSG_PUT_NAK:
1173 case IBLND_MSG_PUT_ACK:
1174 case IBLND_MSG_PUT_DONE:
1175 case IBLND_MSG_GET_DONE:
1176 q = &conn->ibc_tx_queue_nocred;
1179 case IBLND_MSG_NOOP:
1180 if (IBLND_OOB_CAPABLE(conn->ibc_version))
1181 q = &conn->ibc_tx_queue_nocred;
1183 q = &conn->ibc_tx_noops;
1186 case IBLND_MSG_IMMEDIATE:
1187 q = &conn->ibc_tx_queue;
1191 list_add_tail(&tx->tx_list, q);
1195 kiblnd_queue_tx(struct kib_tx *tx, struct kib_conn *conn)
1197 spin_lock(&conn->ibc_lock);
1198 kiblnd_queue_tx_locked(tx, conn);
1199 kiblnd_check_sends_locked(conn);
1200 spin_unlock(&conn->ibc_lock);
1203 static int kiblnd_resolve_addr(struct rdma_cm_id *cmid,
1204 struct sockaddr_in *srcaddr,
1205 struct sockaddr_in *dstaddr,
1208 unsigned short port;
1211 /* allow the port to be reused */
1212 rc = rdma_set_reuseaddr(cmid, 1);
1214 CERROR("Unable to set reuse on cmid: %d\n", rc);
1218 /* look for a free privileged port */
1219 for (port = PROT_SOCK - 1; port > 0; port--) {
1220 srcaddr->sin_port = htons(port);
1221 rc = rdma_resolve_addr(cmid,
1222 (struct sockaddr *)srcaddr,
1223 (struct sockaddr *)dstaddr,
1226 CDEBUG(D_NET, "bound to port %hu\n", port);
1228 } else if (rc == -EADDRINUSE || rc == -EADDRNOTAVAIL) {
1229 CDEBUG(D_NET, "bind to port %hu failed: %d\n",
1236 CERROR("Failed to bind to a free privileged port\n");
1241 kiblnd_connect_peer(struct kib_peer *peer)
1243 struct rdma_cm_id *cmid;
1244 struct kib_dev *dev;
1245 struct kib_net *net = peer->ibp_ni->ni_data;
1246 struct sockaddr_in srcaddr;
1247 struct sockaddr_in dstaddr;
1251 LASSERT(peer->ibp_connecting > 0);
1252 LASSERT(!peer->ibp_reconnecting);
1254 cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, peer, RDMA_PS_TCP,
1258 CERROR("Can't create CMID for %s: %ld\n",
1259 libcfs_nid2str(peer->ibp_nid), PTR_ERR(cmid));
1265 memset(&srcaddr, 0, sizeof(srcaddr));
1266 srcaddr.sin_family = AF_INET;
1267 srcaddr.sin_addr.s_addr = htonl(dev->ibd_ifip);
1269 memset(&dstaddr, 0, sizeof(dstaddr));
1270 dstaddr.sin_family = AF_INET;
1271 dstaddr.sin_port = htons(*kiblnd_tunables.kib_service);
1272 dstaddr.sin_addr.s_addr = htonl(LNET_NIDADDR(peer->ibp_nid));
1274 kiblnd_peer_addref(peer); /* cmid's ref */
1276 if (*kiblnd_tunables.kib_use_priv_port) {
1277 rc = kiblnd_resolve_addr(cmid, &srcaddr, &dstaddr,
1278 *kiblnd_tunables.kib_timeout * 1000);
1280 rc = rdma_resolve_addr(cmid,
1281 (struct sockaddr *)&srcaddr,
1282 (struct sockaddr *)&dstaddr,
1283 *kiblnd_tunables.kib_timeout * 1000);
1286 /* Can't initiate address resolution: */
1287 CERROR("Can't resolve addr for %s: %d\n",
1288 libcfs_nid2str(peer->ibp_nid), rc);
1292 LASSERT(cmid->device);
1293 CDEBUG(D_NET, "%s: connection bound to %s:%pI4h:%s\n",
1294 libcfs_nid2str(peer->ibp_nid), dev->ibd_ifname,
1295 &dev->ibd_ifip, cmid->device->name);
1300 kiblnd_peer_connect_failed(peer, 1, rc);
1301 kiblnd_peer_decref(peer); /* cmid's ref */
1302 rdma_destroy_id(cmid);
1305 kiblnd_peer_connect_failed(peer, 1, rc);
1309 kiblnd_reconnect_peer(struct kib_peer *peer)
1311 rwlock_t *glock = &kiblnd_data.kib_global_lock;
1312 char *reason = NULL;
1313 struct list_head txs;
1314 unsigned long flags;
1316 INIT_LIST_HEAD(&txs);
1318 write_lock_irqsave(glock, flags);
1319 if (!peer->ibp_reconnecting) {
1320 if (peer->ibp_accepting)
1321 reason = "accepting";
1322 else if (peer->ibp_connecting)
1323 reason = "connecting";
1324 else if (!list_empty(&peer->ibp_conns))
1325 reason = "connected";
1326 else /* connected then closed */
1332 LASSERT(!peer->ibp_accepting && !peer->ibp_connecting &&
1333 list_empty(&peer->ibp_conns));
1334 peer->ibp_reconnecting = 0;
1336 if (!kiblnd_peer_active(peer)) {
1337 list_splice_init(&peer->ibp_tx_queue, &txs);
1338 reason = "unlinked";
1342 peer->ibp_connecting++;
1343 peer->ibp_reconnected++;
1344 write_unlock_irqrestore(glock, flags);
1346 kiblnd_connect_peer(peer);
1350 write_unlock_irqrestore(glock, flags);
1352 CWARN("Abort reconnection of %s: %s\n",
1353 libcfs_nid2str(peer->ibp_nid), reason);
1354 kiblnd_txlist_done(peer->ibp_ni, &txs, -ECONNABORTED);
1359 kiblnd_launch_tx(struct lnet_ni *ni, struct kib_tx *tx, lnet_nid_t nid)
1361 struct kib_peer *peer;
1362 struct kib_peer *peer2;
1363 struct kib_conn *conn;
1364 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
1365 unsigned long flags;
1369 * If I get here, I've committed to send, so I complete the tx with
1370 * failure on any problems
1372 LASSERT(!tx || !tx->tx_conn); /* only set when assigned a conn */
1373 LASSERT(!tx || tx->tx_nwrq > 0); /* work items have been set up */
1376 * First time, just use a read lock since I expect to find my peer
1379 read_lock_irqsave(g_lock, flags);
1381 peer = kiblnd_find_peer_locked(nid);
1382 if (peer && !list_empty(&peer->ibp_conns)) {
1383 /* Found a peer with an established connection */
1384 conn = kiblnd_get_conn_locked(peer);
1385 kiblnd_conn_addref(conn); /* 1 ref for me... */
1387 read_unlock_irqrestore(g_lock, flags);
1390 kiblnd_queue_tx(tx, conn);
1391 kiblnd_conn_decref(conn); /* ...to here */
1395 read_unlock(g_lock);
1396 /* Re-try with a write lock */
1399 peer = kiblnd_find_peer_locked(nid);
1401 if (list_empty(&peer->ibp_conns)) {
1402 /* found a peer, but it's still connecting... */
1403 LASSERT(kiblnd_peer_connecting(peer));
1405 list_add_tail(&tx->tx_list,
1406 &peer->ibp_tx_queue);
1407 write_unlock_irqrestore(g_lock, flags);
1409 conn = kiblnd_get_conn_locked(peer);
1410 kiblnd_conn_addref(conn); /* 1 ref for me... */
1412 write_unlock_irqrestore(g_lock, flags);
1415 kiblnd_queue_tx(tx, conn);
1416 kiblnd_conn_decref(conn); /* ...to here */
1421 write_unlock_irqrestore(g_lock, flags);
1423 /* Allocate a peer ready to add to the peer table and retry */
1424 rc = kiblnd_create_peer(ni, &peer, nid);
1426 CERROR("Can't create peer %s\n", libcfs_nid2str(nid));
1428 tx->tx_status = -EHOSTUNREACH;
1430 kiblnd_tx_done(ni, tx);
1435 write_lock_irqsave(g_lock, flags);
1437 peer2 = kiblnd_find_peer_locked(nid);
1439 if (list_empty(&peer2->ibp_conns)) {
1440 /* found a peer, but it's still connecting... */
1441 LASSERT(kiblnd_peer_connecting(peer2));
1443 list_add_tail(&tx->tx_list,
1444 &peer2->ibp_tx_queue);
1445 write_unlock_irqrestore(g_lock, flags);
1447 conn = kiblnd_get_conn_locked(peer2);
1448 kiblnd_conn_addref(conn); /* 1 ref for me... */
1450 write_unlock_irqrestore(g_lock, flags);
1453 kiblnd_queue_tx(tx, conn);
1454 kiblnd_conn_decref(conn); /* ...to here */
1457 kiblnd_peer_decref(peer);
1461 /* Brand new peer */
1462 LASSERT(!peer->ibp_connecting);
1463 peer->ibp_connecting = 1;
1465 /* always called with a ref on ni, which prevents ni being shutdown */
1466 LASSERT(!((struct kib_net *)ni->ni_data)->ibn_shutdown);
1469 list_add_tail(&tx->tx_list, &peer->ibp_tx_queue);
1471 kiblnd_peer_addref(peer);
1472 list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid));
1474 write_unlock_irqrestore(g_lock, flags);
1476 kiblnd_connect_peer(peer);
1477 kiblnd_peer_decref(peer);
1481 kiblnd_send(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg)
1483 struct lnet_hdr *hdr = &lntmsg->msg_hdr;
1484 int type = lntmsg->msg_type;
1485 struct lnet_process_id target = lntmsg->msg_target;
1486 int target_is_router = lntmsg->msg_target_is_router;
1487 int routing = lntmsg->msg_routing;
1488 unsigned int payload_niov = lntmsg->msg_niov;
1489 struct kvec *payload_iov = lntmsg->msg_iov;
1490 struct bio_vec *payload_kiov = lntmsg->msg_kiov;
1491 unsigned int payload_offset = lntmsg->msg_offset;
1492 unsigned int payload_nob = lntmsg->msg_len;
1493 struct iov_iter from;
1494 struct kib_msg *ibmsg;
1495 struct kib_rdma_desc *rd;
1500 /* NB 'private' is different depending on what we're sending.... */
1502 CDEBUG(D_NET, "sending %d bytes in %d frags to %s\n",
1503 payload_nob, payload_niov, libcfs_id2str(target));
1505 LASSERT(!payload_nob || payload_niov > 0);
1506 LASSERT(payload_niov <= LNET_MAX_IOV);
1508 /* Thread context */
1509 LASSERT(!in_interrupt());
1510 /* payload is either all vaddrs or all pages */
1511 LASSERT(!(payload_kiov && payload_iov));
1514 iov_iter_bvec(&from, ITER_BVEC | WRITE,
1515 payload_kiov, payload_niov,
1516 payload_nob + payload_offset);
1518 iov_iter_kvec(&from, ITER_KVEC | WRITE,
1519 payload_iov, payload_niov,
1520 payload_nob + payload_offset);
1522 iov_iter_advance(&from, payload_offset);
1530 LASSERT(!payload_nob);
1534 if (routing || target_is_router)
1535 break; /* send IMMEDIATE */
1537 /* is the REPLY message too small for RDMA? */
1538 nob = offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[lntmsg->msg_md->md_length]);
1539 if (nob <= IBLND_MSG_SIZE)
1540 break; /* send IMMEDIATE */
1542 tx = kiblnd_get_idle_tx(ni, target.nid);
1544 CERROR("Can't allocate txd for GET to %s\n",
1545 libcfs_nid2str(target.nid));
1550 rd = &ibmsg->ibm_u.get.ibgm_rd;
1551 if (!(lntmsg->msg_md->md_options & LNET_MD_KIOV))
1552 rc = kiblnd_setup_rd_iov(ni, tx, rd,
1553 lntmsg->msg_md->md_niov,
1554 lntmsg->msg_md->md_iov.iov,
1555 0, lntmsg->msg_md->md_length);
1557 rc = kiblnd_setup_rd_kiov(ni, tx, rd,
1558 lntmsg->msg_md->md_niov,
1559 lntmsg->msg_md->md_iov.kiov,
1560 0, lntmsg->msg_md->md_length);
1562 CERROR("Can't setup GET sink for %s: %d\n",
1563 libcfs_nid2str(target.nid), rc);
1564 kiblnd_tx_done(ni, tx);
1568 nob = offsetof(struct kib_get_msg, ibgm_rd.rd_frags[rd->rd_nfrags]);
1569 ibmsg->ibm_u.get.ibgm_cookie = tx->tx_cookie;
1570 ibmsg->ibm_u.get.ibgm_hdr = *hdr;
1572 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_GET_REQ, nob);
1574 tx->tx_lntmsg[1] = lnet_create_reply_msg(ni, lntmsg);
1575 if (!tx->tx_lntmsg[1]) {
1576 CERROR("Can't create reply for GET -> %s\n",
1577 libcfs_nid2str(target.nid));
1578 kiblnd_tx_done(ni, tx);
1582 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg[0,1] on completion */
1583 tx->tx_waiting = 1; /* waiting for GET_DONE */
1584 kiblnd_launch_tx(ni, tx, target.nid);
1587 case LNET_MSG_REPLY:
1589 /* Is the payload small enough not to need RDMA? */
1590 nob = offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[payload_nob]);
1591 if (nob <= IBLND_MSG_SIZE)
1592 break; /* send IMMEDIATE */
1594 tx = kiblnd_get_idle_tx(ni, target.nid);
1596 CERROR("Can't allocate %s txd for %s\n",
1597 type == LNET_MSG_PUT ? "PUT" : "REPLY",
1598 libcfs_nid2str(target.nid));
1603 rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1604 payload_niov, payload_iov,
1605 payload_offset, payload_nob);
1607 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1608 payload_niov, payload_kiov,
1609 payload_offset, payload_nob);
1611 CERROR("Can't setup PUT src for %s: %d\n",
1612 libcfs_nid2str(target.nid), rc);
1613 kiblnd_tx_done(ni, tx);
1618 ibmsg->ibm_u.putreq.ibprm_hdr = *hdr;
1619 ibmsg->ibm_u.putreq.ibprm_cookie = tx->tx_cookie;
1620 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_REQ, sizeof(struct kib_putreq_msg));
1622 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1623 tx->tx_waiting = 1; /* waiting for PUT_{ACK,NAK} */
1624 kiblnd_launch_tx(ni, tx, target.nid);
1628 /* send IMMEDIATE */
1630 LASSERT(offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[payload_nob])
1633 tx = kiblnd_get_idle_tx(ni, target.nid);
1635 CERROR("Can't send %d to %s: tx descs exhausted\n",
1636 type, libcfs_nid2str(target.nid));
1641 ibmsg->ibm_u.immediate.ibim_hdr = *hdr;
1643 copy_from_iter(&ibmsg->ibm_u.immediate.ibim_payload, IBLND_MSG_SIZE,
1645 nob = offsetof(struct kib_immediate_msg, ibim_payload[payload_nob]);
1646 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_IMMEDIATE, nob);
1648 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1649 kiblnd_launch_tx(ni, tx, target.nid);
1654 kiblnd_reply(struct lnet_ni *ni, struct kib_rx *rx, struct lnet_msg *lntmsg)
1656 struct lnet_process_id target = lntmsg->msg_target;
1657 unsigned int niov = lntmsg->msg_niov;
1658 struct kvec *iov = lntmsg->msg_iov;
1659 struct bio_vec *kiov = lntmsg->msg_kiov;
1660 unsigned int offset = lntmsg->msg_offset;
1661 unsigned int nob = lntmsg->msg_len;
1665 tx = kiblnd_get_idle_tx(ni, rx->rx_conn->ibc_peer->ibp_nid);
1667 CERROR("Can't get tx for REPLY to %s\n",
1668 libcfs_nid2str(target.nid));
1675 rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1676 niov, iov, offset, nob);
1678 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1679 niov, kiov, offset, nob);
1682 CERROR("Can't setup GET src for %s: %d\n",
1683 libcfs_nid2str(target.nid), rc);
1687 rc = kiblnd_init_rdma(rx->rx_conn, tx,
1688 IBLND_MSG_GET_DONE, nob,
1689 &rx->rx_msg->ibm_u.get.ibgm_rd,
1690 rx->rx_msg->ibm_u.get.ibgm_cookie);
1692 CERROR("Can't setup rdma for GET from %s: %d\n",
1693 libcfs_nid2str(target.nid), rc);
1698 /* No RDMA: local completion may happen now! */
1699 lnet_finalize(ni, lntmsg, 0);
1701 /* RDMA: lnet_finalize(lntmsg) when it completes */
1702 tx->tx_lntmsg[0] = lntmsg;
1705 kiblnd_queue_tx(tx, rx->rx_conn);
1709 kiblnd_tx_done(ni, tx);
1711 lnet_finalize(ni, lntmsg, -EIO);
1715 kiblnd_recv(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg,
1716 int delayed, struct iov_iter *to, unsigned int rlen)
1718 struct kib_rx *rx = private;
1719 struct kib_msg *rxmsg = rx->rx_msg;
1720 struct kib_conn *conn = rx->rx_conn;
1723 int post_credit = IBLND_POSTRX_PEER_CREDIT;
1726 LASSERT(iov_iter_count(to) <= rlen);
1727 LASSERT(!in_interrupt());
1728 /* Either all pages or all vaddrs */
1730 switch (rxmsg->ibm_type) {
1734 case IBLND_MSG_IMMEDIATE:
1735 nob = offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[rlen]);
1736 if (nob > rx->rx_nob) {
1737 CERROR("Immediate message from %s too big: %d(%d)\n",
1738 libcfs_nid2str(rxmsg->ibm_u.immediate.ibim_hdr.src_nid),
1744 copy_to_iter(&rxmsg->ibm_u.immediate.ibim_payload,
1745 IBLND_MSG_SIZE, to);
1746 lnet_finalize(ni, lntmsg, 0);
1749 case IBLND_MSG_PUT_REQ: {
1750 struct kib_msg *txmsg;
1751 struct kib_rdma_desc *rd;
1753 if (!iov_iter_count(to)) {
1754 lnet_finalize(ni, lntmsg, 0);
1755 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK, 0,
1756 rxmsg->ibm_u.putreq.ibprm_cookie);
1760 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
1762 CERROR("Can't allocate tx for %s\n",
1763 libcfs_nid2str(conn->ibc_peer->ibp_nid));
1764 /* Not replying will break the connection */
1770 rd = &txmsg->ibm_u.putack.ibpam_rd;
1771 if (!(to->type & ITER_BVEC))
1772 rc = kiblnd_setup_rd_iov(ni, tx, rd,
1773 to->nr_segs, to->kvec,
1775 iov_iter_count(to));
1777 rc = kiblnd_setup_rd_kiov(ni, tx, rd,
1778 to->nr_segs, to->bvec,
1780 iov_iter_count(to));
1782 CERROR("Can't setup PUT sink for %s: %d\n",
1783 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
1784 kiblnd_tx_done(ni, tx);
1785 /* tell peer it's over */
1786 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK, rc,
1787 rxmsg->ibm_u.putreq.ibprm_cookie);
1791 nob = offsetof(struct kib_putack_msg, ibpam_rd.rd_frags[rd->rd_nfrags]);
1792 txmsg->ibm_u.putack.ibpam_src_cookie = rxmsg->ibm_u.putreq.ibprm_cookie;
1793 txmsg->ibm_u.putack.ibpam_dst_cookie = tx->tx_cookie;
1795 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_ACK, nob);
1797 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1798 tx->tx_waiting = 1; /* waiting for PUT_DONE */
1799 kiblnd_queue_tx(tx, conn);
1801 /* reposted buffer reserved for PUT_DONE */
1802 post_credit = IBLND_POSTRX_NO_CREDIT;
1806 case IBLND_MSG_GET_REQ:
1808 /* Optimized GET; RDMA lntmsg's payload */
1809 kiblnd_reply(ni, rx, lntmsg);
1811 /* GET didn't match anything */
1812 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_GET_DONE,
1814 rxmsg->ibm_u.get.ibgm_cookie);
1819 kiblnd_post_rx(rx, post_credit);
1824 kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name)
1826 struct task_struct *task = kthread_run(fn, arg, "%s", name);
1829 return PTR_ERR(task);
1831 atomic_inc(&kiblnd_data.kib_nthreads);
1836 kiblnd_thread_fini(void)
1838 atomic_dec(&kiblnd_data.kib_nthreads);
1842 kiblnd_peer_alive(struct kib_peer *peer)
1844 /* This is racy, but everyone's only writing cfs_time_current() */
1845 peer->ibp_last_alive = cfs_time_current();
1850 kiblnd_peer_notify(struct kib_peer *peer)
1853 unsigned long last_alive = 0;
1854 unsigned long flags;
1856 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1858 if (kiblnd_peer_idle(peer) && peer->ibp_error) {
1859 error = peer->ibp_error;
1860 peer->ibp_error = 0;
1862 last_alive = peer->ibp_last_alive;
1865 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1868 lnet_notify(peer->ibp_ni,
1869 peer->ibp_nid, 0, last_alive);
1873 kiblnd_close_conn_locked(struct kib_conn *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
1883 struct kib_peer *peer = conn->ibc_peer;
1884 struct kib_dev *dev;
1885 unsigned long flags;
1887 LASSERT(error || conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1889 if (error && !conn->ibc_comms_error)
1890 conn->ibc_comms_error = error;
1892 if (conn->ibc_state != IBLND_CONN_ESTABLISHED)
1893 return; /* already being handled */
1896 list_empty(&conn->ibc_tx_noops) &&
1897 list_empty(&conn->ibc_tx_queue) &&
1898 list_empty(&conn->ibc_tx_queue_rsrvd) &&
1899 list_empty(&conn->ibc_tx_queue_nocred) &&
1900 list_empty(&conn->ibc_active_txs)) {
1901 CDEBUG(D_NET, "closing conn to %s\n",
1902 libcfs_nid2str(peer->ibp_nid));
1904 CNETERR("Closing conn to %s: error %d%s%s%s%s%s\n",
1905 libcfs_nid2str(peer->ibp_nid), error,
1906 list_empty(&conn->ibc_tx_queue) ? "" : "(sending)",
1907 list_empty(&conn->ibc_tx_noops) ? "" : "(sending_noops)",
1908 list_empty(&conn->ibc_tx_queue_rsrvd) ? "" : "(sending_rsrvd)",
1909 list_empty(&conn->ibc_tx_queue_nocred) ? "" : "(sending_nocred)",
1910 list_empty(&conn->ibc_active_txs) ? "" : "(waiting)");
1913 dev = ((struct kib_net *)peer->ibp_ni->ni_data)->ibn_dev;
1914 list_del(&conn->ibc_list);
1915 /* connd (see below) takes over ibc_list's ref */
1917 if (list_empty(&peer->ibp_conns) && /* no more conns */
1918 kiblnd_peer_active(peer)) { /* still in peer table */
1919 kiblnd_unlink_peer_locked(peer);
1921 /* set/clear error on last conn */
1922 peer->ibp_error = conn->ibc_comms_error;
1925 kiblnd_set_conn_state(conn, IBLND_CONN_CLOSING);
1928 kiblnd_dev_can_failover(dev)) {
1929 list_add_tail(&dev->ibd_fail_list,
1930 &kiblnd_data.kib_failed_devs);
1931 wake_up(&kiblnd_data.kib_failover_waitq);
1934 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
1936 list_add_tail(&conn->ibc_list, &kiblnd_data.kib_connd_conns);
1937 wake_up(&kiblnd_data.kib_connd_waitq);
1939 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
1943 kiblnd_close_conn(struct kib_conn *conn, int error)
1945 unsigned long flags;
1947 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1949 kiblnd_close_conn_locked(conn, error);
1951 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1955 kiblnd_handle_early_rxs(struct kib_conn *conn)
1957 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 list_for_each_entry_safe(rx, tmp, &conn->ibc_early_rxs, rx_list) {
1966 list_del(&rx->rx_list);
1967 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1969 kiblnd_handle_rx(rx);
1971 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1973 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1977 kiblnd_abort_txs(struct kib_conn *conn, struct list_head *txs)
1980 struct list_head *tmp;
1981 struct list_head *nxt;
1984 spin_lock(&conn->ibc_lock);
1986 list_for_each_safe(tmp, nxt, txs) {
1987 tx = list_entry(tmp, struct kib_tx, tx_list);
1989 if (txs == &conn->ibc_active_txs) {
1990 LASSERT(!tx->tx_queued);
1991 LASSERT(tx->tx_waiting || tx->tx_sending);
1993 LASSERT(tx->tx_queued);
1996 tx->tx_status = -ECONNABORTED;
1999 if (!tx->tx_sending) {
2001 list_del(&tx->tx_list);
2002 list_add(&tx->tx_list, &zombies);
2006 spin_unlock(&conn->ibc_lock);
2008 kiblnd_txlist_done(conn->ibc_peer->ibp_ni, &zombies, -ECONNABORTED);
2012 kiblnd_finalise_conn(struct kib_conn *conn)
2014 LASSERT(!in_interrupt());
2015 LASSERT(conn->ibc_state > IBLND_CONN_INIT);
2017 kiblnd_set_conn_state(conn, IBLND_CONN_DISCONNECTED);
2020 * abort_receives moves QP state to IB_QPS_ERR. This is only required
2021 * for connections that didn't get as far as being connected, because
2022 * rdma_disconnect() does this for free.
2024 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(struct kib_peer *peer, int active, int error)
2043 unsigned long flags;
2046 LASSERT(!in_interrupt());
2048 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2051 LASSERT(peer->ibp_connecting > 0);
2052 peer->ibp_connecting--;
2054 LASSERT(peer->ibp_accepting > 0);
2055 peer->ibp_accepting--;
2058 if (kiblnd_peer_connecting(peer)) {
2059 /* another connection attempt under way... */
2060 write_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2065 peer->ibp_reconnected = 0;
2066 if (list_empty(&peer->ibp_conns)) {
2067 /* Take peer's blocked transmits to complete with error */
2068 list_add(&zombies, &peer->ibp_tx_queue);
2069 list_del_init(&peer->ibp_tx_queue);
2071 if (kiblnd_peer_active(peer))
2072 kiblnd_unlink_peer_locked(peer);
2074 peer->ibp_error = error;
2076 /* Can't have blocked transmits if there are connections */
2077 LASSERT(list_empty(&peer->ibp_tx_queue));
2080 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2082 kiblnd_peer_notify(peer);
2084 if (list_empty(&zombies))
2087 CNETERR("Deleting messages for %s: connection failed\n",
2088 libcfs_nid2str(peer->ibp_nid));
2090 kiblnd_txlist_done(peer->ibp_ni, &zombies, -EHOSTUNREACH);
2094 kiblnd_connreq_done(struct kib_conn *conn, int status)
2096 struct kib_peer *peer = conn->ibc_peer;
2099 struct list_head txs;
2100 unsigned long flags;
2103 active = (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2105 CDEBUG(D_NET, "%s: active(%d), version(%x), status(%d)\n",
2106 libcfs_nid2str(peer->ibp_nid), active,
2107 conn->ibc_version, status);
2109 LASSERT(!in_interrupt());
2110 LASSERT((conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT &&
2111 peer->ibp_connecting > 0) ||
2112 (conn->ibc_state == IBLND_CONN_PASSIVE_WAIT &&
2113 peer->ibp_accepting > 0));
2115 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
2116 conn->ibc_connvars = NULL;
2119 /* failed to establish connection */
2120 kiblnd_peer_connect_failed(peer, active, status);
2121 kiblnd_finalise_conn(conn);
2125 /* connection established */
2126 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2128 conn->ibc_last_send = jiffies;
2129 kiblnd_set_conn_state(conn, IBLND_CONN_ESTABLISHED);
2130 kiblnd_peer_alive(peer);
2133 * Add conn to peer's list and nuke any dangling conns from a different
2136 kiblnd_conn_addref(conn); /* +1 ref for ibc_list */
2137 list_add(&conn->ibc_list, &peer->ibp_conns);
2138 peer->ibp_reconnected = 0;
2140 peer->ibp_connecting--;
2142 peer->ibp_accepting--;
2144 if (!peer->ibp_version) {
2145 peer->ibp_version = conn->ibc_version;
2146 peer->ibp_incarnation = conn->ibc_incarnation;
2149 if (peer->ibp_version != conn->ibc_version ||
2150 peer->ibp_incarnation != conn->ibc_incarnation) {
2151 kiblnd_close_stale_conns_locked(peer, conn->ibc_version,
2152 conn->ibc_incarnation);
2153 peer->ibp_version = conn->ibc_version;
2154 peer->ibp_incarnation = conn->ibc_incarnation;
2157 /* grab pending txs while I have the lock */
2158 list_add(&txs, &peer->ibp_tx_queue);
2159 list_del_init(&peer->ibp_tx_queue);
2161 if (!kiblnd_peer_active(peer) || /* peer has been deleted */
2162 conn->ibc_comms_error) { /* error has happened already */
2163 struct lnet_ni *ni = peer->ibp_ni;
2165 /* start to shut down connection */
2166 kiblnd_close_conn_locked(conn, -ECONNABORTED);
2167 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2169 kiblnd_txlist_done(ni, &txs, -ECONNABORTED);
2175 * +1 ref for myself, this connection is visible to other threads
2176 * now, refcount of peer:ibp_conns can be released by connection
2177 * close from either a different thread, or the calling of
2178 * kiblnd_check_sends_locked() below. See bz21911 for details.
2180 kiblnd_conn_addref(conn);
2181 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2183 /* Schedule blocked txs */
2184 spin_lock(&conn->ibc_lock);
2185 list_for_each_entry_safe(tx, tmp, &txs, tx_list) {
2186 list_del(&tx->tx_list);
2188 kiblnd_queue_tx_locked(tx, conn);
2190 kiblnd_check_sends_locked(conn);
2191 spin_unlock(&conn->ibc_lock);
2193 /* schedule blocked rxs */
2194 kiblnd_handle_early_rxs(conn);
2196 kiblnd_conn_decref(conn);
2200 kiblnd_reject(struct rdma_cm_id *cmid, struct kib_rej *rej)
2204 rc = rdma_reject(cmid, rej, sizeof(*rej));
2207 CWARN("Error %d sending reject\n", rc);
2211 kiblnd_passive_connect(struct rdma_cm_id *cmid, void *priv, int priv_nob)
2213 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2214 struct kib_msg *reqmsg = priv;
2215 struct kib_msg *ackmsg;
2216 struct kib_dev *ibdev;
2217 struct kib_peer *peer;
2218 struct kib_peer *peer2;
2219 struct kib_conn *conn;
2220 struct lnet_ni *ni = NULL;
2221 struct kib_net *net = NULL;
2223 struct rdma_conn_param cp;
2225 int version = IBLND_MSG_VERSION;
2226 unsigned long flags;
2229 struct sockaddr_in *peer_addr;
2231 LASSERT(!in_interrupt());
2233 /* cmid inherits 'context' from the corresponding listener id */
2234 ibdev = (struct kib_dev *)cmid->context;
2237 memset(&rej, 0, sizeof(rej));
2238 rej.ibr_magic = IBLND_MSG_MAGIC;
2239 rej.ibr_why = IBLND_REJECT_FATAL;
2240 rej.ibr_cp.ibcp_max_msg_size = IBLND_MSG_SIZE;
2242 peer_addr = (struct sockaddr_in *)&cmid->route.addr.dst_addr;
2243 if (*kiblnd_tunables.kib_require_priv_port &&
2244 ntohs(peer_addr->sin_port) >= PROT_SOCK) {
2245 __u32 ip = ntohl(peer_addr->sin_addr.s_addr);
2247 CERROR("Peer's port (%pI4h:%hu) is not privileged\n",
2248 &ip, ntohs(peer_addr->sin_port));
2252 if (priv_nob < offsetof(struct kib_msg, ibm_type)) {
2253 CERROR("Short connection request\n");
2258 * Future protocol version compatibility support! If the
2259 * o2iblnd-specific protocol changes, or when LNET unifies
2260 * protocols over all LNDs, the initial connection will
2261 * negotiate a protocol version. I trap this here to avoid
2262 * console errors; the reject tells the peer which protocol I
2265 if (reqmsg->ibm_magic == LNET_PROTO_MAGIC ||
2266 reqmsg->ibm_magic == __swab32(LNET_PROTO_MAGIC))
2268 if (reqmsg->ibm_magic == IBLND_MSG_MAGIC &&
2269 reqmsg->ibm_version != IBLND_MSG_VERSION &&
2270 reqmsg->ibm_version != IBLND_MSG_VERSION_1)
2272 if (reqmsg->ibm_magic == __swab32(IBLND_MSG_MAGIC) &&
2273 reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION) &&
2274 reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION_1))
2277 rc = kiblnd_unpack_msg(reqmsg, priv_nob);
2279 CERROR("Can't parse connection request: %d\n", rc);
2283 nid = reqmsg->ibm_srcnid;
2284 ni = lnet_net2ni(LNET_NIDNET(reqmsg->ibm_dstnid));
2287 net = (struct kib_net *)ni->ni_data;
2288 rej.ibr_incarnation = net->ibn_incarnation;
2291 if (!ni || /* no matching net */
2292 ni->ni_nid != reqmsg->ibm_dstnid || /* right NET, wrong NID! */
2293 net->ibn_dev != ibdev) { /* wrong device */
2294 CERROR("Can't accept conn from %s on %s (%s:%d:%pI4h): bad dst nid %s\n",
2295 libcfs_nid2str(nid),
2296 !ni ? "NA" : libcfs_nid2str(ni->ni_nid),
2297 ibdev->ibd_ifname, ibdev->ibd_nnets,
2299 libcfs_nid2str(reqmsg->ibm_dstnid));
2304 /* check time stamp as soon as possible */
2305 if (reqmsg->ibm_dststamp &&
2306 reqmsg->ibm_dststamp != net->ibn_incarnation) {
2307 CWARN("Stale connection request\n");
2308 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2312 /* I can accept peer's version */
2313 version = reqmsg->ibm_version;
2315 if (reqmsg->ibm_type != IBLND_MSG_CONNREQ) {
2316 CERROR("Unexpected connreq msg type: %x from %s\n",
2317 reqmsg->ibm_type, libcfs_nid2str(nid));
2321 if (reqmsg->ibm_u.connparams.ibcp_queue_depth >
2322 kiblnd_msg_queue_size(version, ni)) {
2323 CERROR("Can't accept conn from %s, queue depth too large: %d (<=%d wanted)\n",
2324 libcfs_nid2str(nid),
2325 reqmsg->ibm_u.connparams.ibcp_queue_depth,
2326 kiblnd_msg_queue_size(version, ni));
2328 if (version == IBLND_MSG_VERSION)
2329 rej.ibr_why = IBLND_REJECT_MSG_QUEUE_SIZE;
2334 max_frags = reqmsg->ibm_u.connparams.ibcp_max_frags >> IBLND_FRAG_SHIFT;
2335 if (max_frags > kiblnd_rdma_frags(version, ni)) {
2336 CWARN("Can't accept conn from %s (version %x): max message size %d is too large (%d wanted)\n",
2337 libcfs_nid2str(nid), version, max_frags,
2338 kiblnd_rdma_frags(version, ni));
2340 if (version >= IBLND_MSG_VERSION)
2341 rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2344 } else if (max_frags < kiblnd_rdma_frags(version, ni) &&
2346 CWARN("Can't accept conn from %s (version %x): max message size %d incompatible without FMR pool (%d wanted)\n",
2347 libcfs_nid2str(nid), version, max_frags,
2348 kiblnd_rdma_frags(version, ni));
2350 if (version == IBLND_MSG_VERSION)
2351 rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2356 if (reqmsg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2357 CERROR("Can't accept %s: message size %d too big (%d max)\n",
2358 libcfs_nid2str(nid),
2359 reqmsg->ibm_u.connparams.ibcp_max_msg_size,
2364 /* assume 'nid' is a new peer; create */
2365 rc = kiblnd_create_peer(ni, &peer, nid);
2367 CERROR("Can't create peer for %s\n", libcfs_nid2str(nid));
2368 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2372 /* We have validated the peer's parameters so use those */
2373 peer->ibp_max_frags = max_frags;
2374 peer->ibp_queue_depth = reqmsg->ibm_u.connparams.ibcp_queue_depth;
2376 write_lock_irqsave(g_lock, flags);
2378 peer2 = kiblnd_find_peer_locked(nid);
2380 if (!peer2->ibp_version) {
2381 peer2->ibp_version = version;
2382 peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2385 /* not the guy I've talked with */
2386 if (peer2->ibp_incarnation != reqmsg->ibm_srcstamp ||
2387 peer2->ibp_version != version) {
2388 kiblnd_close_peer_conns_locked(peer2, -ESTALE);
2390 if (kiblnd_peer_active(peer2)) {
2391 peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2392 peer2->ibp_version = version;
2394 write_unlock_irqrestore(g_lock, flags);
2396 CWARN("Conn stale %s version %x/%x incarnation %llu/%llu\n",
2397 libcfs_nid2str(nid), peer2->ibp_version, version,
2398 peer2->ibp_incarnation, reqmsg->ibm_srcstamp);
2400 kiblnd_peer_decref(peer);
2401 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2406 * Tie-break connection race in favour of the higher NID.
2407 * If we keep running into a race condition multiple times,
2408 * we have to assume that the connection attempt with the
2409 * higher NID is stuck in a connecting state and will never
2410 * recover. As such, we pass through this if-block and let
2411 * the lower NID connection win so we can move forward.
2413 if (peer2->ibp_connecting &&
2414 nid < ni->ni_nid && peer2->ibp_races <
2415 MAX_CONN_RACES_BEFORE_ABORT) {
2417 write_unlock_irqrestore(g_lock, flags);
2419 CDEBUG(D_NET, "Conn race %s\n",
2420 libcfs_nid2str(peer2->ibp_nid));
2422 kiblnd_peer_decref(peer);
2423 rej.ibr_why = IBLND_REJECT_CONN_RACE;
2426 if (peer2->ibp_races >= MAX_CONN_RACES_BEFORE_ABORT)
2427 CNETERR("Conn race %s: unresolved after %d attempts, letting lower NID win\n",
2428 libcfs_nid2str(peer2->ibp_nid),
2429 MAX_CONN_RACES_BEFORE_ABORT);
2431 * passive connection is allowed even this peer is waiting for
2434 peer2->ibp_reconnecting = 0;
2435 peer2->ibp_races = 0;
2436 peer2->ibp_accepting++;
2437 kiblnd_peer_addref(peer2);
2440 * Race with kiblnd_launch_tx (active connect) to create peer
2441 * so copy validated parameters since we now know what the
2444 peer2->ibp_max_frags = peer->ibp_max_frags;
2445 peer2->ibp_queue_depth = peer->ibp_queue_depth;
2447 write_unlock_irqrestore(g_lock, flags);
2448 kiblnd_peer_decref(peer);
2451 /* Brand new peer */
2452 LASSERT(!peer->ibp_accepting);
2453 LASSERT(!peer->ibp_version &&
2454 !peer->ibp_incarnation);
2456 peer->ibp_accepting = 1;
2457 peer->ibp_version = version;
2458 peer->ibp_incarnation = reqmsg->ibm_srcstamp;
2460 /* I have a ref on ni that prevents it being shutdown */
2461 LASSERT(!net->ibn_shutdown);
2463 kiblnd_peer_addref(peer);
2464 list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid));
2466 write_unlock_irqrestore(g_lock, flags);
2469 conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_PASSIVE_WAIT,
2472 kiblnd_peer_connect_failed(peer, 0, -ENOMEM);
2473 kiblnd_peer_decref(peer);
2474 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2479 * conn now "owns" cmid, so I return success from here on to ensure the
2480 * CM callback doesn't destroy cmid.
2482 conn->ibc_incarnation = reqmsg->ibm_srcstamp;
2483 conn->ibc_credits = conn->ibc_queue_depth;
2484 conn->ibc_reserved_credits = conn->ibc_queue_depth;
2485 LASSERT(conn->ibc_credits + conn->ibc_reserved_credits +
2486 IBLND_OOB_MSGS(version) <= IBLND_RX_MSGS(conn));
2488 ackmsg = &conn->ibc_connvars->cv_msg;
2489 memset(ackmsg, 0, sizeof(*ackmsg));
2491 kiblnd_init_msg(ackmsg, IBLND_MSG_CONNACK,
2492 sizeof(ackmsg->ibm_u.connparams));
2493 ackmsg->ibm_u.connparams.ibcp_queue_depth = conn->ibc_queue_depth;
2494 ackmsg->ibm_u.connparams.ibcp_max_frags = conn->ibc_max_frags << IBLND_FRAG_SHIFT;
2495 ackmsg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2497 kiblnd_pack_msg(ni, ackmsg, version, 0, nid, reqmsg->ibm_srcstamp);
2499 memset(&cp, 0, sizeof(cp));
2500 cp.private_data = ackmsg;
2501 cp.private_data_len = ackmsg->ibm_nob;
2502 cp.responder_resources = 0; /* No atomic ops or RDMA reads */
2503 cp.initiator_depth = 0;
2504 cp.flow_control = 1;
2505 cp.retry_count = *kiblnd_tunables.kib_retry_count;
2506 cp.rnr_retry_count = *kiblnd_tunables.kib_rnr_retry_count;
2508 CDEBUG(D_NET, "Accept %s\n", libcfs_nid2str(nid));
2510 rc = rdma_accept(cmid, &cp);
2512 CERROR("Can't accept %s: %d\n", libcfs_nid2str(nid), rc);
2513 rej.ibr_version = version;
2514 rej.ibr_why = IBLND_REJECT_FATAL;
2516 kiblnd_reject(cmid, &rej);
2517 kiblnd_connreq_done(conn, rc);
2518 kiblnd_conn_decref(conn);
2526 rej.ibr_cp.ibcp_queue_depth = kiblnd_msg_queue_size(version, ni);
2527 rej.ibr_cp.ibcp_max_frags = kiblnd_rdma_frags(version, ni);
2531 rej.ibr_version = version;
2532 kiblnd_reject(cmid, &rej);
2534 return -ECONNREFUSED;
2538 kiblnd_check_reconnect(struct kib_conn *conn, int version,
2539 __u64 incarnation, int why, struct kib_connparams *cp)
2541 rwlock_t *glock = &kiblnd_data.kib_global_lock;
2542 struct kib_peer *peer = conn->ibc_peer;
2544 int msg_size = IBLND_MSG_SIZE;
2548 unsigned long flags;
2550 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2551 LASSERT(peer->ibp_connecting > 0); /* 'conn' at least */
2552 LASSERT(!peer->ibp_reconnecting);
2555 msg_size = cp->ibcp_max_msg_size;
2556 frag_num = cp->ibcp_max_frags << IBLND_FRAG_SHIFT;
2557 queue_dep = cp->ibcp_queue_depth;
2560 write_lock_irqsave(glock, flags);
2562 * retry connection if it's still needed and no other connection
2563 * attempts (active or passive) are in progress
2564 * NB: reconnect is still needed even when ibp_tx_queue is
2565 * empty if ibp_version != version because reconnect may be
2566 * initiated by kiblnd_query()
2568 reconnect = (!list_empty(&peer->ibp_tx_queue) ||
2569 peer->ibp_version != version) &&
2570 peer->ibp_connecting == 1 &&
2571 !peer->ibp_accepting;
2582 case IBLND_REJECT_RDMA_FRAGS: {
2583 struct lnet_ioctl_config_lnd_tunables *tunables;
2586 reason = "can't negotiate max frags";
2589 tunables = peer->ibp_ni->ni_lnd_tunables;
2590 if (!tunables->lt_tun_u.lt_o2ib.lnd_map_on_demand) {
2591 reason = "map_on_demand must be enabled";
2594 if (conn->ibc_max_frags <= frag_num) {
2595 reason = "unsupported max frags";
2599 peer->ibp_max_frags = frag_num;
2600 reason = "rdma fragments";
2603 case IBLND_REJECT_MSG_QUEUE_SIZE:
2605 reason = "can't negotiate queue depth";
2608 if (conn->ibc_queue_depth <= queue_dep) {
2609 reason = "unsupported queue depth";
2613 peer->ibp_queue_depth = queue_dep;
2614 reason = "queue depth";
2617 case IBLND_REJECT_CONN_STALE:
2621 case IBLND_REJECT_CONN_RACE:
2622 reason = "conn race";
2625 case IBLND_REJECT_CONN_UNCOMPAT:
2626 reason = "version negotiation";
2630 conn->ibc_reconnect = 1;
2631 peer->ibp_reconnecting = 1;
2632 peer->ibp_version = version;
2634 peer->ibp_incarnation = incarnation;
2636 write_unlock_irqrestore(glock, flags);
2638 CNETERR("%s: %s (%s), %x, %x, msg_size: %d, queue_depth: %d/%d, max_frags: %d/%d\n",
2639 libcfs_nid2str(peer->ibp_nid),
2640 reconnect ? "reconnect" : "don't reconnect",
2641 reason, IBLND_MSG_VERSION, version, msg_size,
2642 conn->ibc_queue_depth, queue_dep,
2643 conn->ibc_max_frags, frag_num);
2645 * if conn::ibc_reconnect is TRUE, connd will reconnect to the peer
2646 * while destroying the zombie
2651 kiblnd_rejected(struct kib_conn *conn, int reason, void *priv, int priv_nob)
2653 struct kib_peer *peer = conn->ibc_peer;
2655 LASSERT(!in_interrupt());
2656 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2659 case IB_CM_REJ_STALE_CONN:
2660 kiblnd_check_reconnect(conn, IBLND_MSG_VERSION, 0,
2661 IBLND_REJECT_CONN_STALE, NULL);
2664 case IB_CM_REJ_INVALID_SERVICE_ID:
2665 CNETERR("%s rejected: no listener at %d\n",
2666 libcfs_nid2str(peer->ibp_nid),
2667 *kiblnd_tunables.kib_service);
2670 case IB_CM_REJ_CONSUMER_DEFINED:
2671 if (priv_nob >= offsetof(struct kib_rej, ibr_padding)) {
2672 struct kib_rej *rej = priv;
2673 struct kib_connparams *cp = NULL;
2675 __u64 incarnation = -1;
2677 /* NB. default incarnation is -1 because:
2678 * a) V1 will ignore dst incarnation in connreq.
2679 * b) V2 will provide incarnation while rejecting me,
2680 * -1 will be overwrote.
2682 * if I try to connect to a V1 peer with V2 protocol,
2683 * it rejected me then upgrade to V2, I have no idea
2684 * about the upgrading and try to reconnect with V1,
2685 * in this case upgraded V2 can find out I'm trying to
2686 * talk to the old guy and reject me(incarnation is -1).
2689 if (rej->ibr_magic == __swab32(IBLND_MSG_MAGIC) ||
2690 rej->ibr_magic == __swab32(LNET_PROTO_MAGIC)) {
2691 __swab32s(&rej->ibr_magic);
2692 __swab16s(&rej->ibr_version);
2696 if (priv_nob >= sizeof(struct kib_rej) &&
2697 rej->ibr_version > IBLND_MSG_VERSION_1) {
2699 * priv_nob is always 148 in current version
2700 * of OFED, so we still need to check version.
2701 * (define of IB_CM_REJ_PRIVATE_DATA_SIZE)
2706 __swab64s(&rej->ibr_incarnation);
2707 __swab16s(&cp->ibcp_queue_depth);
2708 __swab16s(&cp->ibcp_max_frags);
2709 __swab32s(&cp->ibcp_max_msg_size);
2712 incarnation = rej->ibr_incarnation;
2715 if (rej->ibr_magic != IBLND_MSG_MAGIC &&
2716 rej->ibr_magic != LNET_PROTO_MAGIC) {
2717 CERROR("%s rejected: consumer defined fatal error\n",
2718 libcfs_nid2str(peer->ibp_nid));
2722 if (rej->ibr_version != IBLND_MSG_VERSION &&
2723 rej->ibr_version != IBLND_MSG_VERSION_1) {
2724 CERROR("%s rejected: o2iblnd version %x error\n",
2725 libcfs_nid2str(peer->ibp_nid),
2730 if (rej->ibr_why == IBLND_REJECT_FATAL &&
2731 rej->ibr_version == IBLND_MSG_VERSION_1) {
2732 CDEBUG(D_NET, "rejected by old version peer %s: %x\n",
2733 libcfs_nid2str(peer->ibp_nid), rej->ibr_version);
2735 if (conn->ibc_version != IBLND_MSG_VERSION_1)
2736 rej->ibr_why = IBLND_REJECT_CONN_UNCOMPAT;
2739 switch (rej->ibr_why) {
2740 case IBLND_REJECT_CONN_RACE:
2741 case IBLND_REJECT_CONN_STALE:
2742 case IBLND_REJECT_CONN_UNCOMPAT:
2743 case IBLND_REJECT_MSG_QUEUE_SIZE:
2744 case IBLND_REJECT_RDMA_FRAGS:
2745 kiblnd_check_reconnect(conn, rej->ibr_version,
2750 case IBLND_REJECT_NO_RESOURCES:
2751 CERROR("%s rejected: o2iblnd no resources\n",
2752 libcfs_nid2str(peer->ibp_nid));
2755 case IBLND_REJECT_FATAL:
2756 CERROR("%s rejected: o2iblnd fatal error\n",
2757 libcfs_nid2str(peer->ibp_nid));
2761 CERROR("%s rejected: o2iblnd reason %d\n",
2762 libcfs_nid2str(peer->ibp_nid),
2770 CNETERR("%s rejected: reason %d, size %d\n",
2771 libcfs_nid2str(peer->ibp_nid), reason, priv_nob);
2775 kiblnd_connreq_done(conn, -ECONNREFUSED);
2779 kiblnd_check_connreply(struct kib_conn *conn, void *priv, int priv_nob)
2781 struct kib_peer *peer = conn->ibc_peer;
2782 struct lnet_ni *ni = peer->ibp_ni;
2783 struct kib_net *net = ni->ni_data;
2784 struct kib_msg *msg = priv;
2785 int ver = conn->ibc_version;
2786 int rc = kiblnd_unpack_msg(msg, priv_nob);
2787 unsigned long flags;
2792 CERROR("Can't unpack connack from %s: %d\n",
2793 libcfs_nid2str(peer->ibp_nid), rc);
2797 if (msg->ibm_type != IBLND_MSG_CONNACK) {
2798 CERROR("Unexpected message %d from %s\n",
2799 msg->ibm_type, libcfs_nid2str(peer->ibp_nid));
2804 if (ver != msg->ibm_version) {
2805 CERROR("%s replied version %x is different with requested version %x\n",
2806 libcfs_nid2str(peer->ibp_nid), msg->ibm_version, ver);
2811 if (msg->ibm_u.connparams.ibcp_queue_depth >
2812 conn->ibc_queue_depth) {
2813 CERROR("%s has incompatible queue depth %d (<=%d wanted)\n",
2814 libcfs_nid2str(peer->ibp_nid),
2815 msg->ibm_u.connparams.ibcp_queue_depth,
2816 conn->ibc_queue_depth);
2821 if ((msg->ibm_u.connparams.ibcp_max_frags >> IBLND_FRAG_SHIFT) >
2822 conn->ibc_max_frags) {
2823 CERROR("%s has incompatible max_frags %d (<=%d wanted)\n",
2824 libcfs_nid2str(peer->ibp_nid),
2825 msg->ibm_u.connparams.ibcp_max_frags >> IBLND_FRAG_SHIFT,
2826 conn->ibc_max_frags);
2831 if (msg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2832 CERROR("%s max message size %d too big (%d max)\n",
2833 libcfs_nid2str(peer->ibp_nid),
2834 msg->ibm_u.connparams.ibcp_max_msg_size,
2840 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2841 if (msg->ibm_dstnid == ni->ni_nid &&
2842 msg->ibm_dststamp == net->ibn_incarnation)
2846 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2849 CERROR("Bad connection reply from %s, rc = %d, version: %x max_frags: %d\n",
2850 libcfs_nid2str(peer->ibp_nid), rc,
2851 msg->ibm_version, msg->ibm_u.connparams.ibcp_max_frags);
2855 conn->ibc_incarnation = msg->ibm_srcstamp;
2856 conn->ibc_credits = msg->ibm_u.connparams.ibcp_queue_depth;
2857 conn->ibc_reserved_credits = msg->ibm_u.connparams.ibcp_queue_depth;
2858 conn->ibc_queue_depth = msg->ibm_u.connparams.ibcp_queue_depth;
2859 conn->ibc_max_frags = msg->ibm_u.connparams.ibcp_max_frags >> IBLND_FRAG_SHIFT;
2860 LASSERT(conn->ibc_credits + conn->ibc_reserved_credits +
2861 IBLND_OOB_MSGS(ver) <= IBLND_RX_MSGS(conn));
2863 kiblnd_connreq_done(conn, 0);
2868 * NB My QP has already established itself, so I handle anything going
2869 * wrong here by setting ibc_comms_error.
2870 * kiblnd_connreq_done(0) moves the conn state to ESTABLISHED, but then
2871 * immediately tears it down.
2874 conn->ibc_comms_error = rc;
2875 kiblnd_connreq_done(conn, 0);
2879 kiblnd_active_connect(struct rdma_cm_id *cmid)
2881 struct kib_peer *peer = (struct kib_peer *)cmid->context;
2882 struct kib_conn *conn;
2883 struct kib_msg *msg;
2884 struct rdma_conn_param cp;
2887 unsigned long flags;
2890 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2892 incarnation = peer->ibp_incarnation;
2893 version = !peer->ibp_version ? IBLND_MSG_VERSION :
2896 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2898 conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_ACTIVE_CONNECT,
2901 kiblnd_peer_connect_failed(peer, 1, -ENOMEM);
2902 kiblnd_peer_decref(peer); /* lose cmid's ref */
2907 * conn "owns" cmid now, so I return success from here on to ensure the
2908 * CM callback doesn't destroy cmid. conn also takes over cmid's ref
2911 msg = &conn->ibc_connvars->cv_msg;
2913 memset(msg, 0, sizeof(*msg));
2914 kiblnd_init_msg(msg, IBLND_MSG_CONNREQ, sizeof(msg->ibm_u.connparams));
2915 msg->ibm_u.connparams.ibcp_queue_depth = conn->ibc_queue_depth;
2916 msg->ibm_u.connparams.ibcp_max_frags = conn->ibc_max_frags << IBLND_FRAG_SHIFT;
2917 msg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2919 kiblnd_pack_msg(peer->ibp_ni, msg, version,
2920 0, peer->ibp_nid, incarnation);
2922 memset(&cp, 0, sizeof(cp));
2923 cp.private_data = msg;
2924 cp.private_data_len = msg->ibm_nob;
2925 cp.responder_resources = 0; /* No atomic ops or RDMA reads */
2926 cp.initiator_depth = 0;
2927 cp.flow_control = 1;
2928 cp.retry_count = *kiblnd_tunables.kib_retry_count;
2929 cp.rnr_retry_count = *kiblnd_tunables.kib_rnr_retry_count;
2931 LASSERT(cmid->context == (void *)conn);
2932 LASSERT(conn->ibc_cmid == cmid);
2934 rc = rdma_connect(cmid, &cp);
2936 CERROR("Can't connect to %s: %d\n",
2937 libcfs_nid2str(peer->ibp_nid), rc);
2938 kiblnd_connreq_done(conn, rc);
2939 kiblnd_conn_decref(conn);
2946 kiblnd_cm_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2948 struct kib_peer *peer;
2949 struct kib_conn *conn;
2952 switch (event->event) {
2954 CERROR("Unexpected event: %d, status: %d\n",
2955 event->event, event->status);
2958 case RDMA_CM_EVENT_CONNECT_REQUEST:
2959 /* destroy cmid on failure */
2960 rc = kiblnd_passive_connect(cmid,
2961 (void *)KIBLND_CONN_PARAM(event),
2962 KIBLND_CONN_PARAM_LEN(event));
2963 CDEBUG(D_NET, "connreq: %d\n", rc);
2966 case RDMA_CM_EVENT_ADDR_ERROR:
2967 peer = (struct kib_peer *)cmid->context;
2968 CNETERR("%s: ADDR ERROR %d\n",
2969 libcfs_nid2str(peer->ibp_nid), event->status);
2970 kiblnd_peer_connect_failed(peer, 1, -EHOSTUNREACH);
2971 kiblnd_peer_decref(peer);
2972 return -EHOSTUNREACH; /* rc destroys cmid */
2974 case RDMA_CM_EVENT_ADDR_RESOLVED:
2975 peer = (struct kib_peer *)cmid->context;
2977 CDEBUG(D_NET, "%s Addr resolved: %d\n",
2978 libcfs_nid2str(peer->ibp_nid), event->status);
2980 if (event->status) {
2981 CNETERR("Can't resolve address for %s: %d\n",
2982 libcfs_nid2str(peer->ibp_nid), event->status);
2985 rc = rdma_resolve_route(
2986 cmid, *kiblnd_tunables.kib_timeout * 1000);
2989 /* Can't initiate route resolution */
2990 CERROR("Can't resolve route for %s: %d\n",
2991 libcfs_nid2str(peer->ibp_nid), rc);
2993 kiblnd_peer_connect_failed(peer, 1, rc);
2994 kiblnd_peer_decref(peer);
2995 return rc; /* rc destroys cmid */
2997 case RDMA_CM_EVENT_ROUTE_ERROR:
2998 peer = (struct kib_peer *)cmid->context;
2999 CNETERR("%s: ROUTE ERROR %d\n",
3000 libcfs_nid2str(peer->ibp_nid), event->status);
3001 kiblnd_peer_connect_failed(peer, 1, -EHOSTUNREACH);
3002 kiblnd_peer_decref(peer);
3003 return -EHOSTUNREACH; /* rc destroys cmid */
3005 case RDMA_CM_EVENT_ROUTE_RESOLVED:
3006 peer = (struct kib_peer *)cmid->context;
3007 CDEBUG(D_NET, "%s Route resolved: %d\n",
3008 libcfs_nid2str(peer->ibp_nid), event->status);
3011 return kiblnd_active_connect(cmid);
3013 CNETERR("Can't resolve route for %s: %d\n",
3014 libcfs_nid2str(peer->ibp_nid), event->status);
3015 kiblnd_peer_connect_failed(peer, 1, event->status);
3016 kiblnd_peer_decref(peer);
3017 return event->status; /* rc destroys cmid */
3019 case RDMA_CM_EVENT_UNREACHABLE:
3020 conn = (struct kib_conn *)cmid->context;
3021 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
3022 conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
3023 CNETERR("%s: UNREACHABLE %d\n",
3024 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
3025 kiblnd_connreq_done(conn, -ENETDOWN);
3026 kiblnd_conn_decref(conn);
3029 case RDMA_CM_EVENT_CONNECT_ERROR:
3030 conn = (struct kib_conn *)cmid->context;
3031 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
3032 conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
3033 CNETERR("%s: CONNECT ERROR %d\n",
3034 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
3035 kiblnd_connreq_done(conn, -ENOTCONN);
3036 kiblnd_conn_decref(conn);
3039 case RDMA_CM_EVENT_REJECTED:
3040 conn = (struct kib_conn *)cmid->context;
3041 switch (conn->ibc_state) {
3045 case IBLND_CONN_PASSIVE_WAIT:
3046 CERROR("%s: REJECTED %d\n",
3047 libcfs_nid2str(conn->ibc_peer->ibp_nid),
3049 kiblnd_connreq_done(conn, -ECONNRESET);
3052 case IBLND_CONN_ACTIVE_CONNECT:
3053 kiblnd_rejected(conn, event->status,
3054 (void *)KIBLND_CONN_PARAM(event),
3055 KIBLND_CONN_PARAM_LEN(event));
3058 kiblnd_conn_decref(conn);
3061 case RDMA_CM_EVENT_ESTABLISHED:
3062 conn = (struct kib_conn *)cmid->context;
3063 switch (conn->ibc_state) {
3067 case IBLND_CONN_PASSIVE_WAIT:
3068 CDEBUG(D_NET, "ESTABLISHED (passive): %s\n",
3069 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3070 kiblnd_connreq_done(conn, 0);
3073 case IBLND_CONN_ACTIVE_CONNECT:
3074 CDEBUG(D_NET, "ESTABLISHED(active): %s\n",
3075 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3076 kiblnd_check_connreply(conn,
3077 (void *)KIBLND_CONN_PARAM(event),
3078 KIBLND_CONN_PARAM_LEN(event));
3081 /* net keeps its ref on conn! */
3084 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
3085 CDEBUG(D_NET, "Ignore TIMEWAIT_EXIT event\n");
3087 case RDMA_CM_EVENT_DISCONNECTED:
3088 conn = (struct kib_conn *)cmid->context;
3089 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
3090 CERROR("%s DISCONNECTED\n",
3091 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3092 kiblnd_connreq_done(conn, -ECONNRESET);
3094 kiblnd_close_conn(conn, 0);
3096 kiblnd_conn_decref(conn);
3097 cmid->context = NULL;
3100 case RDMA_CM_EVENT_DEVICE_REMOVAL:
3101 LCONSOLE_ERROR_MSG(0x131,
3102 "Received notification of device removal\n"
3103 "Please shutdown LNET to allow this to proceed\n");
3105 * Can't remove network from underneath LNET for now, so I have
3110 case RDMA_CM_EVENT_ADDR_CHANGE:
3111 LCONSOLE_INFO("Physical link changed (eg hca/port)\n");
3117 kiblnd_check_txs_locked(struct kib_conn *conn, struct list_head *txs)
3120 struct list_head *ttmp;
3122 list_for_each(ttmp, txs) {
3123 tx = list_entry(ttmp, struct kib_tx, tx_list);
3125 if (txs != &conn->ibc_active_txs) {
3126 LASSERT(tx->tx_queued);
3128 LASSERT(!tx->tx_queued);
3129 LASSERT(tx->tx_waiting || tx->tx_sending);
3132 if (cfs_time_aftereq(jiffies, tx->tx_deadline)) {
3133 CERROR("Timed out tx: %s, %lu seconds\n",
3134 kiblnd_queue2str(conn, txs),
3135 cfs_duration_sec(jiffies - tx->tx_deadline));
3144 kiblnd_conn_timed_out_locked(struct kib_conn *conn)
3146 return kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue) ||
3147 kiblnd_check_txs_locked(conn, &conn->ibc_tx_noops) ||
3148 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_rsrvd) ||
3149 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_nocred) ||
3150 kiblnd_check_txs_locked(conn, &conn->ibc_active_txs);
3154 kiblnd_check_conns(int idx)
3157 LIST_HEAD(checksends);
3158 struct list_head *peers = &kiblnd_data.kib_peers[idx];
3159 struct list_head *ptmp;
3160 struct kib_peer *peer;
3161 struct kib_conn *conn;
3162 struct kib_conn *temp;
3163 struct kib_conn *tmp;
3164 struct list_head *ctmp;
3165 unsigned long flags;
3168 * NB. We expect to have a look at all the peers and not find any
3169 * RDMAs to time out, so we just use a shared lock while we
3172 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3174 list_for_each(ptmp, peers) {
3175 peer = list_entry(ptmp, struct kib_peer, ibp_list);
3177 list_for_each(ctmp, &peer->ibp_conns) {
3181 conn = list_entry(ctmp, struct kib_conn, ibc_list);
3183 LASSERT(conn->ibc_state == IBLND_CONN_ESTABLISHED);
3185 spin_lock(&conn->ibc_lock);
3187 sendnoop = kiblnd_need_noop(conn);
3188 timedout = kiblnd_conn_timed_out_locked(conn);
3189 if (!sendnoop && !timedout) {
3190 spin_unlock(&conn->ibc_lock);
3195 CERROR("Timed out RDMA with %s (%lu): c: %u, oc: %u, rc: %u\n",
3196 libcfs_nid2str(peer->ibp_nid),
3197 cfs_duration_sec(cfs_time_current() -
3198 peer->ibp_last_alive),
3200 conn->ibc_outstanding_credits,
3201 conn->ibc_reserved_credits);
3202 list_add(&conn->ibc_connd_list, &closes);
3204 list_add(&conn->ibc_connd_list, &checksends);
3206 /* +ref for 'closes' or 'checksends' */
3207 kiblnd_conn_addref(conn);
3209 spin_unlock(&conn->ibc_lock);
3213 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3216 * Handle timeout by closing the whole
3217 * connection. We can only be sure RDMA activity
3218 * has ceased once the QP has been modified.
3220 list_for_each_entry_safe(conn, tmp, &closes, ibc_connd_list) {
3221 list_del(&conn->ibc_connd_list);
3222 kiblnd_close_conn(conn, -ETIMEDOUT);
3223 kiblnd_conn_decref(conn);
3227 * In case we have enough credits to return via a
3228 * NOOP, but there were no non-blocking tx descs
3229 * free to do it last time...
3231 list_for_each_entry_safe(conn, temp, &checksends, ibc_connd_list) {
3232 list_del(&conn->ibc_connd_list);
3234 spin_lock(&conn->ibc_lock);
3235 kiblnd_check_sends_locked(conn);
3236 spin_unlock(&conn->ibc_lock);
3238 kiblnd_conn_decref(conn);
3243 kiblnd_disconnect_conn(struct kib_conn *conn)
3245 LASSERT(!in_interrupt());
3246 LASSERT(current == kiblnd_data.kib_connd);
3247 LASSERT(conn->ibc_state == IBLND_CONN_CLOSING);
3249 rdma_disconnect(conn->ibc_cmid);
3250 kiblnd_finalise_conn(conn);
3252 kiblnd_peer_notify(conn->ibc_peer);
3256 * High-water for reconnection to the same peer, reconnection attempt should
3257 * be delayed after trying more than KIB_RECONN_HIGH_RACE.
3259 #define KIB_RECONN_HIGH_RACE 10
3261 * Allow connd to take a break and handle other things after consecutive
3262 * reconnection attempts.
3264 #define KIB_RECONN_BREAK 100
3267 kiblnd_connd(void *arg)
3269 spinlock_t *lock = &kiblnd_data.kib_connd_lock;
3270 wait_queue_entry_t wait;
3271 unsigned long flags;
3272 struct kib_conn *conn;
3277 unsigned long deadline = jiffies;
3279 cfs_block_allsigs();
3281 init_waitqueue_entry(&wait, current);
3282 kiblnd_data.kib_connd = current;
3284 spin_lock_irqsave(lock, flags);
3286 while (!kiblnd_data.kib_shutdown) {
3291 if (!list_empty(&kiblnd_data.kib_connd_zombies)) {
3292 struct kib_peer *peer = NULL;
3294 conn = list_entry(kiblnd_data.kib_connd_zombies.next,
3295 struct kib_conn, ibc_list);
3296 list_del(&conn->ibc_list);
3297 if (conn->ibc_reconnect) {
3298 peer = conn->ibc_peer;
3299 kiblnd_peer_addref(peer);
3302 spin_unlock_irqrestore(lock, flags);
3305 kiblnd_destroy_conn(conn, !peer);
3307 spin_lock_irqsave(lock, flags);
3311 conn->ibc_peer = peer;
3312 if (peer->ibp_reconnected < KIB_RECONN_HIGH_RACE)
3313 list_add_tail(&conn->ibc_list,
3314 &kiblnd_data.kib_reconn_list);
3316 list_add_tail(&conn->ibc_list,
3317 &kiblnd_data.kib_reconn_wait);
3320 if (!list_empty(&kiblnd_data.kib_connd_conns)) {
3321 conn = list_entry(kiblnd_data.kib_connd_conns.next,
3322 struct kib_conn, ibc_list);
3323 list_del(&conn->ibc_list);
3325 spin_unlock_irqrestore(lock, flags);
3328 kiblnd_disconnect_conn(conn);
3329 kiblnd_conn_decref(conn);
3331 spin_lock_irqsave(lock, flags);
3334 while (reconn < KIB_RECONN_BREAK) {
3335 if (kiblnd_data.kib_reconn_sec !=
3336 ktime_get_real_seconds()) {
3337 kiblnd_data.kib_reconn_sec = ktime_get_real_seconds();
3338 list_splice_init(&kiblnd_data.kib_reconn_wait,
3339 &kiblnd_data.kib_reconn_list);
3342 if (list_empty(&kiblnd_data.kib_reconn_list))
3345 conn = list_entry(kiblnd_data.kib_reconn_list.next,
3346 struct kib_conn, ibc_list);
3347 list_del(&conn->ibc_list);
3349 spin_unlock_irqrestore(lock, flags);
3352 reconn += kiblnd_reconnect_peer(conn->ibc_peer);
3353 kiblnd_peer_decref(conn->ibc_peer);
3354 LIBCFS_FREE(conn, sizeof(*conn));
3356 spin_lock_irqsave(lock, flags);
3359 /* careful with the jiffy wrap... */
3360 timeout = (int)(deadline - jiffies);
3364 int chunk = kiblnd_data.kib_peer_hash_size;
3366 spin_unlock_irqrestore(lock, flags);
3370 * Time to check for RDMA timeouts on a few more
3371 * peers: I do checks every 'p' seconds on a
3372 * proportion of the peer table and I need to check
3373 * every connection 'n' times within a timeout
3374 * interval, to ensure I detect a timeout on any
3375 * connection within (n+1)/n times the timeout
3378 if (*kiblnd_tunables.kib_timeout > n * p)
3379 chunk = (chunk * n * p) /
3380 *kiblnd_tunables.kib_timeout;
3384 for (i = 0; i < chunk; i++) {
3385 kiblnd_check_conns(peer_index);
3386 peer_index = (peer_index + 1) %
3387 kiblnd_data.kib_peer_hash_size;
3390 deadline += msecs_to_jiffies(p * MSEC_PER_SEC);
3391 spin_lock_irqsave(lock, flags);
3397 /* Nothing to do for 'timeout' */
3398 set_current_state(TASK_INTERRUPTIBLE);
3399 add_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3400 spin_unlock_irqrestore(lock, flags);
3402 schedule_timeout(timeout);
3404 remove_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3405 spin_lock_irqsave(lock, flags);
3408 spin_unlock_irqrestore(lock, flags);
3410 kiblnd_thread_fini();
3415 kiblnd_qp_event(struct ib_event *event, void *arg)
3417 struct kib_conn *conn = arg;
3419 switch (event->event) {
3420 case IB_EVENT_COMM_EST:
3421 CDEBUG(D_NET, "%s established\n",
3422 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3424 * We received a packet but connection isn't established
3425 * probably handshake packet was lost, so free to
3426 * force make connection established
3428 rdma_notify(conn->ibc_cmid, IB_EVENT_COMM_EST);
3432 CERROR("%s: Async QP event type %d\n",
3433 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3439 kiblnd_complete(struct ib_wc *wc)
3441 switch (kiblnd_wreqid2type(wc->wr_id)) {
3446 if (wc->status != IB_WC_SUCCESS &&
3447 wc->status != IB_WC_WR_FLUSH_ERR)
3448 CNETERR("FastReg failed: %d\n", wc->status);
3451 case IBLND_WID_RDMA:
3453 * We only get RDMA completion notification if it fails. All
3454 * subsequent work items, including the final SEND will fail
3455 * too. However we can't print out any more info about the
3456 * failing RDMA because 'tx' might be back on the idle list or
3457 * even reused already if we didn't manage to post all our work
3460 CNETERR("RDMA (tx: %p) failed: %d\n",
3461 kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3465 kiblnd_tx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3469 kiblnd_rx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status,
3476 kiblnd_cq_completion(struct ib_cq *cq, void *arg)
3479 * NB I'm not allowed to schedule this conn once its refcount has
3480 * reached 0. Since fundamentally I'm racing with scheduler threads
3481 * consuming my CQ I could be called after all completions have
3482 * occurred. But in this case, !ibc_nrx && !ibc_nsends_posted
3483 * and this CQ is about to be destroyed so I NOOP.
3485 struct kib_conn *conn = arg;
3486 struct kib_sched_info *sched = conn->ibc_sched;
3487 unsigned long flags;
3489 LASSERT(cq == conn->ibc_cq);
3491 spin_lock_irqsave(&sched->ibs_lock, flags);
3493 conn->ibc_ready = 1;
3495 if (!conn->ibc_scheduled &&
3496 (conn->ibc_nrx > 0 ||
3497 conn->ibc_nsends_posted > 0)) {
3498 kiblnd_conn_addref(conn); /* +1 ref for sched_conns */
3499 conn->ibc_scheduled = 1;
3500 list_add_tail(&conn->ibc_sched_list, &sched->ibs_conns);
3502 if (waitqueue_active(&sched->ibs_waitq))
3503 wake_up(&sched->ibs_waitq);
3506 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3510 kiblnd_cq_event(struct ib_event *event, void *arg)
3512 struct kib_conn *conn = arg;
3514 CERROR("%s: async CQ event type %d\n",
3515 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3519 kiblnd_scheduler(void *arg)
3521 long id = (long)arg;
3522 struct kib_sched_info *sched;
3523 struct kib_conn *conn;
3524 wait_queue_entry_t wait;
3525 unsigned long flags;
3531 cfs_block_allsigs();
3533 init_waitqueue_entry(&wait, current);
3535 sched = kiblnd_data.kib_scheds[KIB_THREAD_CPT(id)];
3537 rc = cfs_cpt_bind(lnet_cpt_table(), sched->ibs_cpt);
3539 CWARN("Unable to bind on CPU partition %d, please verify whether all CPUs are healthy and reload modules if necessary, otherwise your system might under risk of low performance\n",
3543 spin_lock_irqsave(&sched->ibs_lock, flags);
3545 while (!kiblnd_data.kib_shutdown) {
3546 if (busy_loops++ >= IBLND_RESCHED) {
3547 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3552 spin_lock_irqsave(&sched->ibs_lock, flags);
3557 if (!list_empty(&sched->ibs_conns)) {
3558 conn = list_entry(sched->ibs_conns.next, struct kib_conn,
3560 /* take over kib_sched_conns' ref on conn... */
3561 LASSERT(conn->ibc_scheduled);
3562 list_del(&conn->ibc_sched_list);
3563 conn->ibc_ready = 0;
3565 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3567 wc.wr_id = IBLND_WID_INVAL;
3569 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3571 rc = ib_req_notify_cq(conn->ibc_cq,
3574 CWARN("%s: ib_req_notify_cq failed: %d, closing connection\n",
3575 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
3576 kiblnd_close_conn(conn, -EIO);
3577 kiblnd_conn_decref(conn);
3578 spin_lock_irqsave(&sched->ibs_lock,
3583 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3586 if (unlikely(rc > 0 && wc.wr_id == IBLND_WID_INVAL)) {
3587 LCONSOLE_ERROR("ib_poll_cq (rc: %d) returned invalid wr_id, opcode %d, status: %d, vendor_err: %d, conn: %s status: %d\nplease upgrade firmware and OFED or contact vendor.\n",
3588 rc, wc.opcode, wc.status,
3590 libcfs_nid2str(conn->ibc_peer->ibp_nid),
3596 CWARN("%s: ib_poll_cq failed: %d, closing connection\n",
3597 libcfs_nid2str(conn->ibc_peer->ibp_nid),
3599 kiblnd_close_conn(conn, -EIO);
3600 kiblnd_conn_decref(conn);
3601 spin_lock_irqsave(&sched->ibs_lock, flags);
3605 spin_lock_irqsave(&sched->ibs_lock, flags);
3607 if (rc || conn->ibc_ready) {
3609 * There may be another completion waiting; get
3610 * another scheduler to check while I handle
3613 /* +1 ref for sched_conns */
3614 kiblnd_conn_addref(conn);
3615 list_add_tail(&conn->ibc_sched_list,
3617 if (waitqueue_active(&sched->ibs_waitq))
3618 wake_up(&sched->ibs_waitq);
3620 conn->ibc_scheduled = 0;
3624 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3625 kiblnd_complete(&wc);
3627 spin_lock_irqsave(&sched->ibs_lock, flags);
3630 kiblnd_conn_decref(conn); /* ...drop my ref from above */
3637 set_current_state(TASK_INTERRUPTIBLE);
3638 add_wait_queue_exclusive(&sched->ibs_waitq, &wait);
3639 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3644 remove_wait_queue(&sched->ibs_waitq, &wait);
3645 spin_lock_irqsave(&sched->ibs_lock, flags);
3648 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3650 kiblnd_thread_fini();
3655 kiblnd_failover_thread(void *arg)
3657 rwlock_t *glock = &kiblnd_data.kib_global_lock;
3658 struct kib_dev *dev;
3659 wait_queue_entry_t wait;
3660 unsigned long flags;
3663 LASSERT(*kiblnd_tunables.kib_dev_failover);
3665 cfs_block_allsigs();
3667 init_waitqueue_entry(&wait, current);
3668 write_lock_irqsave(glock, flags);
3670 while (!kiblnd_data.kib_shutdown) {
3671 int do_failover = 0;
3674 list_for_each_entry(dev, &kiblnd_data.kib_failed_devs,
3676 if (time_before(cfs_time_current(),
3677 dev->ibd_next_failover))
3684 list_del_init(&dev->ibd_fail_list);
3685 dev->ibd_failover = 1;
3686 write_unlock_irqrestore(glock, flags);
3688 rc = kiblnd_dev_failover(dev);
3690 write_lock_irqsave(glock, flags);
3692 LASSERT(dev->ibd_failover);
3693 dev->ibd_failover = 0;
3694 if (rc >= 0) { /* Device is OK or failover succeed */
3695 dev->ibd_next_failover = cfs_time_shift(3);
3699 /* failed to failover, retry later */
3700 dev->ibd_next_failover =
3701 cfs_time_shift(min(dev->ibd_failed_failover, 10));
3702 if (kiblnd_dev_can_failover(dev)) {
3703 list_add_tail(&dev->ibd_fail_list,
3704 &kiblnd_data.kib_failed_devs);
3710 /* long sleep if no more pending failover */
3711 long_sleep = list_empty(&kiblnd_data.kib_failed_devs);
3713 set_current_state(TASK_INTERRUPTIBLE);
3714 add_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3715 write_unlock_irqrestore(glock, flags);
3717 rc = schedule_timeout(long_sleep ? cfs_time_seconds(10) :
3718 cfs_time_seconds(1));
3719 remove_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3720 write_lock_irqsave(glock, flags);
3722 if (!long_sleep || rc)
3726 * have a long sleep, routine check all active devices,
3727 * we need checking like this because if there is not active
3728 * connection on the dev and no SEND from local, we may listen
3729 * on wrong HCA for ever while there is a bonding failover
3731 list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3732 if (kiblnd_dev_can_failover(dev)) {
3733 list_add_tail(&dev->ibd_fail_list,
3734 &kiblnd_data.kib_failed_devs);
3739 write_unlock_irqrestore(glock, flags);
3741 kiblnd_thread_fini();