4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2012, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lnet/klnds/o2iblnd/o2iblnd_cb.c
38 * Author: Eric Barton <eric@bartonsoftware.com>
43 static void kiblnd_peer_alive(kib_peer_t *peer);
44 static void kiblnd_peer_connect_failed(kib_peer_t *peer, int active, int error);
45 static void kiblnd_check_sends(kib_conn_t *conn);
46 static void kiblnd_init_tx_msg(lnet_ni_t *ni, kib_tx_t *tx,
47 int type, int body_nob);
48 static int kiblnd_init_rdma(kib_conn_t *conn, kib_tx_t *tx, int type,
49 int resid, kib_rdma_desc_t *dstrd, __u64 dstcookie);
50 static void kiblnd_queue_tx_locked(kib_tx_t *tx, kib_conn_t *conn);
51 static void kiblnd_queue_tx(kib_tx_t *tx, kib_conn_t *conn);
54 kiblnd_tx_done (lnet_ni_t *ni, kib_tx_t *tx)
56 lnet_msg_t *lntmsg[2];
57 kib_net_t *net = ni->ni_data;
61 LASSERT (net != NULL);
62 LASSERT (!in_interrupt());
63 LASSERT (!tx->tx_queued); /* mustn't be queued for sending */
64 LASSERT (tx->tx_sending == 0); /* mustn't be awaiting sent callback */
65 LASSERT (!tx->tx_waiting); /* mustn't be awaiting peer response */
66 LASSERT (tx->tx_pool != NULL);
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;
75 if (tx->tx_conn != 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++) {
89 if (lntmsg[i] == NULL)
92 lnet_finalize(ni, lntmsg[i], rc);
97 kiblnd_txlist_done(lnet_ni_t *ni, struct list_head *txlist, int status)
101 while (!list_empty(txlist)) {
102 tx = list_entry(txlist->next, kib_tx_t, tx_list);
104 list_del(&tx->tx_list);
107 tx->tx_status = status;
108 kiblnd_tx_done(ni, tx);
113 kiblnd_get_idle_tx(lnet_ni_t *ni, lnet_nid_t target)
115 kib_net_t *net = (kib_net_t *)ni->ni_data;
116 struct list_head *node;
118 kib_tx_poolset_t *tps;
120 tps = net->ibn_tx_ps[lnet_cpt_of_nid(target)];
121 node = kiblnd_pool_alloc_node(&tps->tps_poolset);
124 tx = container_of(node, kib_tx_t, tx_list);
126 LASSERT (tx->tx_nwrq == 0);
127 LASSERT (!tx->tx_queued);
128 LASSERT (tx->tx_sending == 0);
129 LASSERT (!tx->tx_waiting);
130 LASSERT (tx->tx_status == 0);
131 LASSERT (tx->tx_conn == NULL);
132 LASSERT (tx->tx_lntmsg[0] == NULL);
133 LASSERT (tx->tx_lntmsg[1] == NULL);
134 LASSERT (tx->tx_u.pmr == NULL);
135 LASSERT (tx->tx_nfrags == 0);
141 kiblnd_drop_rx(kib_rx_t *rx)
143 kib_conn_t *conn = rx->rx_conn;
144 struct kib_sched_info *sched = conn->ibc_sched;
147 spin_lock_irqsave(&sched->ibs_lock, flags);
148 LASSERT(conn->ibc_nrx > 0);
150 spin_unlock_irqrestore(&sched->ibs_lock, flags);
152 kiblnd_conn_decref(conn);
156 kiblnd_post_rx (kib_rx_t *rx, int credit)
158 kib_conn_t *conn = rx->rx_conn;
159 kib_net_t *net = conn->ibc_peer->ibp_ni->ni_data;
160 struct ib_recv_wr *bad_wrq = NULL;
164 LASSERT (net != NULL);
165 LASSERT (!in_interrupt());
166 LASSERT (credit == IBLND_POSTRX_NO_CREDIT ||
167 credit == IBLND_POSTRX_PEER_CREDIT ||
168 credit == IBLND_POSTRX_RSRVD_CREDIT);
170 mr = kiblnd_find_dma_mr(conn->ibc_hdev, rx->rx_msgaddr, IBLND_MSG_SIZE);
171 LASSERT (mr != NULL);
173 rx->rx_sge.lkey = mr->lkey;
174 rx->rx_sge.addr = rx->rx_msgaddr;
175 rx->rx_sge.length = IBLND_MSG_SIZE;
177 rx->rx_wrq.next = NULL;
178 rx->rx_wrq.sg_list = &rx->rx_sge;
179 rx->rx_wrq.num_sge = 1;
180 rx->rx_wrq.wr_id = kiblnd_ptr2wreqid(rx, IBLND_WID_RX);
182 LASSERT (conn->ibc_state >= IBLND_CONN_INIT);
183 LASSERT (rx->rx_nob >= 0); /* not posted */
185 if (conn->ibc_state > IBLND_CONN_ESTABLISHED) {
186 kiblnd_drop_rx(rx); /* No more posts for this rx */
190 rx->rx_nob = -1; /* flag posted */
192 rc = ib_post_recv(conn->ibc_cmid->qp, &rx->rx_wrq, &bad_wrq);
194 CERROR("Can't post rx for %s: %d, bad_wrq: %p\n",
195 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc, bad_wrq);
199 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) /* Initial post */
203 kiblnd_close_conn(conn, rc);
204 kiblnd_drop_rx(rx); /* No more posts for this rx */
208 if (credit == IBLND_POSTRX_NO_CREDIT)
211 spin_lock(&conn->ibc_lock);
212 if (credit == IBLND_POSTRX_PEER_CREDIT)
213 conn->ibc_outstanding_credits++;
215 conn->ibc_reserved_credits++;
216 spin_unlock(&conn->ibc_lock);
218 kiblnd_check_sends(conn);
223 kiblnd_find_waiting_tx_locked(kib_conn_t *conn, int txtype, __u64 cookie)
225 struct list_head *tmp;
227 list_for_each(tmp, &conn->ibc_active_txs) {
228 kib_tx_t *tx = list_entry(tmp, kib_tx_t, tx_list);
230 LASSERT(!tx->tx_queued);
231 LASSERT(tx->tx_sending != 0 || tx->tx_waiting);
233 if (tx->tx_cookie != cookie)
236 if (tx->tx_waiting &&
237 tx->tx_msg->ibm_type == txtype)
240 CWARN("Bad completion: %swaiting, type %x (wanted %x)\n",
241 tx->tx_waiting ? "" : "NOT ",
242 tx->tx_msg->ibm_type, txtype);
248 kiblnd_handle_completion(kib_conn_t *conn, int txtype, int status, __u64 cookie)
251 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
254 spin_lock(&conn->ibc_lock);
256 tx = kiblnd_find_waiting_tx_locked(conn, txtype, cookie);
258 spin_unlock(&conn->ibc_lock);
260 CWARN("Unmatched completion type %x cookie "LPX64" from %s\n",
261 txtype, cookie, libcfs_nid2str(conn->ibc_peer->ibp_nid));
262 kiblnd_close_conn(conn, -EPROTO);
266 if (tx->tx_status == 0) { /* success so far */
267 if (status < 0) { /* failed? */
268 tx->tx_status = status;
269 } else if (txtype == IBLND_MSG_GET_REQ) {
270 lnet_set_reply_msg_len(ni, tx->tx_lntmsg[1], status);
276 idle = !tx->tx_queued && (tx->tx_sending == 0);
278 list_del(&tx->tx_list);
280 spin_unlock(&conn->ibc_lock);
283 kiblnd_tx_done(ni, tx);
287 kiblnd_send_completion(kib_conn_t *conn, int type, int status, __u64 cookie)
289 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
290 kib_tx_t *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(kib_completion_msg_t));
302 kiblnd_queue_tx(tx, conn);
306 kiblnd_handle_rx (kib_rx_t *rx)
308 kib_msg_t *msg = rx->rx_msg;
309 kib_conn_t *conn = rx->rx_conn;
310 lnet_ni_t *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 IBLND_MSG_QUEUE_SIZE(conn->ibc_version)) {
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),
335 IBLND_MSG_QUEUE_SIZE(conn->ibc_version));
337 kiblnd_close_conn(conn, -EPROTO);
338 kiblnd_post_rx(rx, IBLND_POSTRX_NO_CREDIT);
342 conn->ibc_credits += credits;
344 /* This ensures the credit taken by NOOP can be returned */
345 if (msg->ibm_type == IBLND_MSG_NOOP &&
346 !IBLND_OOB_CAPABLE(conn->ibc_version)) /* v1 only */
347 conn->ibc_outstanding_credits++;
349 spin_unlock(&conn->ibc_lock);
350 kiblnd_check_sends(conn);
353 switch (msg->ibm_type) {
355 CERROR("Bad IBLND message type %x from %s\n",
356 msg->ibm_type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
357 post_credit = IBLND_POSTRX_NO_CREDIT;
362 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
363 post_credit = IBLND_POSTRX_NO_CREDIT;
367 if (credits != 0) /* credit already posted */
368 post_credit = IBLND_POSTRX_NO_CREDIT;
369 else /* a keepalive NOOP */
370 post_credit = IBLND_POSTRX_PEER_CREDIT;
373 case IBLND_MSG_IMMEDIATE:
374 post_credit = IBLND_POSTRX_DONT_POST;
375 rc = lnet_parse(ni, &msg->ibm_u.immediate.ibim_hdr,
376 msg->ibm_srcnid, rx, 0);
377 if (rc < 0) /* repost on error */
378 post_credit = IBLND_POSTRX_PEER_CREDIT;
381 case IBLND_MSG_PUT_REQ:
382 post_credit = IBLND_POSTRX_DONT_POST;
383 rc = lnet_parse(ni, &msg->ibm_u.putreq.ibprm_hdr,
384 msg->ibm_srcnid, rx, 1);
385 if (rc < 0) /* repost on error */
386 post_credit = IBLND_POSTRX_PEER_CREDIT;
389 case IBLND_MSG_PUT_NAK:
390 CWARN ("PUT_NACK from %s\n",
391 libcfs_nid2str(conn->ibc_peer->ibp_nid));
392 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
393 kiblnd_handle_completion(conn, IBLND_MSG_PUT_REQ,
394 msg->ibm_u.completion.ibcm_status,
395 msg->ibm_u.completion.ibcm_cookie);
398 case IBLND_MSG_PUT_ACK:
399 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
401 spin_lock(&conn->ibc_lock);
402 tx = kiblnd_find_waiting_tx_locked(conn, IBLND_MSG_PUT_REQ,
403 msg->ibm_u.putack.ibpam_src_cookie);
405 list_del(&tx->tx_list);
406 spin_unlock(&conn->ibc_lock);
409 CERROR("Unmatched PUT_ACK from %s\n",
410 libcfs_nid2str(conn->ibc_peer->ibp_nid));
415 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 (kib_rx_t *rx, int status, int nob)
469 kib_msg_t *msg = rx->rx_msg;
470 kib_conn_t *conn = rx->rx_conn;
471 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
472 kib_net_t *net = ni->ni_data;
476 LASSERT (net != NULL);
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);
544 LASSERT (page != NULL);
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);
556 LASSERT (page != NULL);
561 kiblnd_fmr_map_tx(kib_net_t *net, kib_tx_t *tx, kib_rdma_desc_t *rd, int nob)
564 __u64 *pages = tx->tx_pages;
565 kib_fmr_poolset_t *fps;
572 LASSERT(tx->tx_pool != NULL);
573 LASSERT(tx->tx_pool->tpo_pool.po_owner != NULL);
575 hdev = tx->tx_pool->tpo_hdev;
577 for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
578 for (size = 0; size < rd->rd_frags[i].rf_nob;
579 size += hdev->ibh_page_size) {
580 pages[npages ++] = (rd->rd_frags[i].rf_addr &
581 hdev->ibh_page_mask) + size;
585 cpt = tx->tx_pool->tpo_pool.po_owner->ps_cpt;
587 fps = net->ibn_fmr_ps[cpt];
588 rc = kiblnd_fmr_pool_map(fps, pages, npages, 0, &tx->tx_u.fmr);
590 CERROR ("Can't map %d pages: %d\n", npages, rc);
594 /* If rd is not tx_rd, it's going to get sent to a peer, who will need
596 rd->rd_key = (rd != tx->tx_rd) ? tx->tx_u.fmr.fmr_pfmr->fmr->rkey :
597 tx->tx_u.fmr.fmr_pfmr->fmr->lkey;
598 rd->rd_frags[0].rf_addr &= ~hdev->ibh_page_mask;
599 rd->rd_frags[0].rf_nob = nob;
606 kiblnd_pmr_map_tx(kib_net_t *net, kib_tx_t *tx, kib_rdma_desc_t *rd, int nob)
609 kib_pmr_poolset_t *pps;
614 LASSERT(tx->tx_pool != NULL);
615 LASSERT(tx->tx_pool->tpo_pool.po_owner != NULL);
617 hdev = tx->tx_pool->tpo_hdev;
619 iova = rd->rd_frags[0].rf_addr & ~hdev->ibh_page_mask;
621 cpt = tx->tx_pool->tpo_pool.po_owner->ps_cpt;
623 pps = net->ibn_pmr_ps[cpt];
624 rc = kiblnd_pmr_pool_map(pps, hdev, rd, &iova, &tx->tx_u.pmr);
626 CERROR("Failed to create MR by phybuf: %d\n", rc);
630 /* If rd is not tx_rd, it's going to get sent to a peer, who will need
632 rd->rd_key = (rd != tx->tx_rd) ? tx->tx_u.pmr->pmr_mr->rkey :
633 tx->tx_u.pmr->pmr_mr->lkey;
635 rd->rd_frags[0].rf_addr = iova;
636 rd->rd_frags[0].rf_nob = nob;
642 kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx)
644 kib_net_t *net = ni->ni_data;
646 LASSERT(net != NULL);
648 if (net->ibn_fmr_ps != NULL && tx->tx_u.fmr.fmr_pfmr != NULL) {
649 kiblnd_fmr_pool_unmap(&tx->tx_u.fmr, tx->tx_status);
650 tx->tx_u.fmr.fmr_pfmr = NULL;
652 } else if (net->ibn_pmr_ps != NULL && tx->tx_u.pmr != NULL) {
653 kiblnd_pmr_pool_unmap(tx->tx_u.pmr);
657 if (tx->tx_nfrags != 0) {
658 kiblnd_dma_unmap_sg(tx->tx_pool->tpo_hdev->ibh_ibdev,
659 tx->tx_frags, tx->tx_nfrags, tx->tx_dmadir);
665 kiblnd_map_tx(lnet_ni_t *ni, kib_tx_t *tx,
666 kib_rdma_desc_t *rd, int nfrags)
668 kib_hca_dev_t *hdev = tx->tx_pool->tpo_hdev;
669 kib_net_t *net = ni->ni_data;
670 struct ib_mr *mr = NULL;
674 /* If rd is not tx_rd, it's going to get sent to a peer and I'm the
676 tx->tx_dmadir = (rd != tx->tx_rd) ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
677 tx->tx_nfrags = nfrags;
680 kiblnd_dma_map_sg(hdev->ibh_ibdev,
681 tx->tx_frags, tx->tx_nfrags, tx->tx_dmadir);
683 for (i = 0, nob = 0; i < rd->rd_nfrags; i++) {
684 rd->rd_frags[i].rf_nob = kiblnd_sg_dma_len(
685 hdev->ibh_ibdev, &tx->tx_frags[i]);
686 rd->rd_frags[i].rf_addr = kiblnd_sg_dma_address(
687 hdev->ibh_ibdev, &tx->tx_frags[i]);
688 nob += rd->rd_frags[i].rf_nob;
691 /* looking for pre-mapping MR */
692 mr = kiblnd_find_rd_dma_mr(hdev, rd);
694 /* found pre-mapping MR */
695 rd->rd_key = (rd != tx->tx_rd) ? mr->rkey : mr->lkey;
699 if (net->ibn_fmr_ps != NULL)
700 return kiblnd_fmr_map_tx(net, tx, rd, nob);
701 else if (net->ibn_pmr_ps != NULL)
702 return kiblnd_pmr_map_tx(net, tx, rd, nob);
709 kiblnd_setup_rd_iov(lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd,
710 unsigned int niov, struct iovec *iov, int offset, int nob)
712 kib_net_t *net = ni->ni_data;
714 struct scatterlist *sg;
721 LASSERT (net != NULL);
723 while (offset >= iov->iov_len) {
724 offset -= iov->iov_len;
734 vaddr = ((unsigned long)iov->iov_base) + offset;
735 page_offset = vaddr & (PAGE_SIZE - 1);
736 page = kiblnd_kvaddr_to_page(vaddr);
738 CERROR ("Can't find page\n");
742 fragnob = min((int)(iov->iov_len - offset), nob);
743 fragnob = min(fragnob, (int)PAGE_SIZE - page_offset);
745 sg_set_page(sg, page, fragnob, page_offset);
748 if (offset + fragnob < iov->iov_len) {
758 return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags);
762 kiblnd_setup_rd_kiov (lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd,
763 int nkiov, lnet_kiov_t *kiov, int offset, int nob)
765 kib_net_t *net = ni->ni_data;
766 struct scatterlist *sg;
769 CDEBUG(D_NET, "niov %d offset %d nob %d\n", nkiov, offset, nob);
773 LASSERT (net != NULL);
775 while (offset >= kiov->kiov_len) {
776 offset -= kiov->kiov_len;
786 fragnob = min((int)(kiov->kiov_len - offset), nob);
788 sg_set_page(sg, kiov->kiov_page, fragnob,
789 kiov->kiov_offset + offset);
798 return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags);
802 kiblnd_post_tx_locked (kib_conn_t *conn, kib_tx_t *tx, int credit)
803 __must_hold(&conn->ibc_lock)
805 kib_msg_t *msg = tx->tx_msg;
806 kib_peer_t *peer = conn->ibc_peer;
807 int ver = conn->ibc_version;
810 struct ib_send_wr *bad_wrq;
812 LASSERT (tx->tx_queued);
813 /* We rely on this for QP sizing */
814 LASSERT (tx->tx_nwrq > 0);
815 LASSERT (tx->tx_nwrq <= 1 + IBLND_RDMA_FRAGS(ver));
817 LASSERT (credit == 0 || credit == 1);
818 LASSERT (conn->ibc_outstanding_credits >= 0);
819 LASSERT (conn->ibc_outstanding_credits <= IBLND_MSG_QUEUE_SIZE(ver));
820 LASSERT (conn->ibc_credits >= 0);
821 LASSERT (conn->ibc_credits <= IBLND_MSG_QUEUE_SIZE(ver));
823 if (conn->ibc_nsends_posted == IBLND_CONCURRENT_SENDS(ver)) {
824 /* tx completions outstanding... */
825 CDEBUG(D_NET, "%s: posted enough\n",
826 libcfs_nid2str(peer->ibp_nid));
830 if (credit != 0 && conn->ibc_credits == 0) { /* no credits */
831 CDEBUG(D_NET, "%s: no credits\n",
832 libcfs_nid2str(peer->ibp_nid));
836 if (credit != 0 && !IBLND_OOB_CAPABLE(ver) &&
837 conn->ibc_credits == 1 && /* last credit reserved */
838 msg->ibm_type != IBLND_MSG_NOOP) { /* for NOOP */
839 CDEBUG(D_NET, "%s: not using last credit\n",
840 libcfs_nid2str(peer->ibp_nid));
844 /* NB don't drop ibc_lock before bumping tx_sending */
845 list_del(&tx->tx_list);
848 if (msg->ibm_type == IBLND_MSG_NOOP &&
849 (!kiblnd_need_noop(conn) || /* redundant NOOP */
850 (IBLND_OOB_CAPABLE(ver) && /* posted enough NOOP */
851 conn->ibc_noops_posted == IBLND_OOB_MSGS(ver)))) {
852 /* OK to drop when posted enough NOOPs, since
853 * kiblnd_check_sends will queue NOOP again when
854 * posted NOOPs complete */
855 spin_unlock(&conn->ibc_lock);
856 kiblnd_tx_done(peer->ibp_ni, tx);
857 spin_lock(&conn->ibc_lock);
858 CDEBUG(D_NET, "%s(%d): redundant or enough NOOP\n",
859 libcfs_nid2str(peer->ibp_nid),
860 conn->ibc_noops_posted);
864 kiblnd_pack_msg(peer->ibp_ni, msg, ver, conn->ibc_outstanding_credits,
865 peer->ibp_nid, conn->ibc_incarnation);
867 conn->ibc_credits -= credit;
868 conn->ibc_outstanding_credits = 0;
869 conn->ibc_nsends_posted++;
870 if (msg->ibm_type == IBLND_MSG_NOOP)
871 conn->ibc_noops_posted++;
873 /* CAVEAT EMPTOR! This tx could be the PUT_DONE of an RDMA
874 * PUT. If so, it was first queued here as a PUT_REQ, sent and
875 * stashed on ibc_active_txs, matched by an incoming PUT_ACK,
876 * and then re-queued here. It's (just) possible that
877 * tx_sending is non-zero if we've not done the tx_complete()
878 * from the first send; hence the ++ rather than = below. */
880 list_add(&tx->tx_list, &conn->ibc_active_txs);
882 /* I'm still holding ibc_lock! */
883 if (conn->ibc_state != IBLND_CONN_ESTABLISHED) {
885 } else if (tx->tx_pool->tpo_pool.po_failed ||
886 conn->ibc_hdev != tx->tx_pool->tpo_hdev) {
887 /* close_conn will launch failover */
890 rc = ib_post_send(conn->ibc_cmid->qp,
891 tx->tx_wrq, &bad_wrq);
894 conn->ibc_last_send = jiffies;
899 /* NB credits are transferred in the actual
900 * message, which can only be the last work item */
901 conn->ibc_credits += credit;
902 conn->ibc_outstanding_credits += msg->ibm_credits;
903 conn->ibc_nsends_posted--;
904 if (msg->ibm_type == IBLND_MSG_NOOP)
905 conn->ibc_noops_posted--;
911 done = (tx->tx_sending == 0);
913 list_del(&tx->tx_list);
915 spin_unlock(&conn->ibc_lock);
917 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
918 CERROR("Error %d posting transmit to %s\n",
919 rc, libcfs_nid2str(peer->ibp_nid));
921 CDEBUG(D_NET, "Error %d posting transmit to %s\n",
922 rc, libcfs_nid2str(peer->ibp_nid));
924 kiblnd_close_conn(conn, rc);
927 kiblnd_tx_done(peer->ibp_ni, tx);
929 spin_lock(&conn->ibc_lock);
935 kiblnd_check_sends (kib_conn_t *conn)
937 int ver = conn->ibc_version;
938 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
941 /* Don't send anything until after the connection is established */
942 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
943 CDEBUG(D_NET, "%s too soon\n",
944 libcfs_nid2str(conn->ibc_peer->ibp_nid));
948 spin_lock(&conn->ibc_lock);
950 LASSERT (conn->ibc_nsends_posted <= IBLND_CONCURRENT_SENDS(ver));
951 LASSERT (!IBLND_OOB_CAPABLE(ver) ||
952 conn->ibc_noops_posted <= IBLND_OOB_MSGS(ver));
953 LASSERT (conn->ibc_reserved_credits >= 0);
955 while (conn->ibc_reserved_credits > 0 &&
956 !list_empty(&conn->ibc_tx_queue_rsrvd)) {
957 tx = list_entry(conn->ibc_tx_queue_rsrvd.next,
959 list_del(&tx->tx_list);
960 list_add_tail(&tx->tx_list, &conn->ibc_tx_queue);
961 conn->ibc_reserved_credits--;
964 if (kiblnd_need_noop(conn)) {
965 spin_unlock(&conn->ibc_lock);
967 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
969 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_NOOP, 0);
971 spin_lock(&conn->ibc_lock);
973 kiblnd_queue_tx_locked(tx, conn);
976 kiblnd_conn_addref(conn); /* 1 ref for me.... (see b21911) */
981 if (!list_empty(&conn->ibc_tx_queue_nocred)) {
983 tx = list_entry(conn->ibc_tx_queue_nocred.next,
985 } else if (!list_empty(&conn->ibc_tx_noops)) {
986 LASSERT (!IBLND_OOB_CAPABLE(ver));
988 tx = list_entry(conn->ibc_tx_noops.next,
990 } else if (!list_empty(&conn->ibc_tx_queue)) {
992 tx = list_entry(conn->ibc_tx_queue.next,
997 if (kiblnd_post_tx_locked(conn, tx, credit) != 0)
1001 spin_unlock(&conn->ibc_lock);
1003 kiblnd_conn_decref(conn); /* ...until here */
1007 kiblnd_tx_complete (kib_tx_t *tx, int status)
1009 int failed = (status != IB_WC_SUCCESS);
1010 kib_conn_t *conn = tx->tx_conn;
1013 LASSERT (tx->tx_sending > 0);
1016 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
1017 CNETERR("Tx -> %s cookie "LPX64
1018 " sending %d waiting %d: failed %d\n",
1019 libcfs_nid2str(conn->ibc_peer->ibp_nid),
1020 tx->tx_cookie, tx->tx_sending, tx->tx_waiting,
1023 kiblnd_close_conn(conn, -EIO);
1025 kiblnd_peer_alive(conn->ibc_peer);
1028 spin_lock(&conn->ibc_lock);
1030 /* I could be racing with rdma completion. Whoever makes 'tx' idle
1031 * gets to free it, which also drops its ref on 'conn'. */
1034 conn->ibc_nsends_posted--;
1035 if (tx->tx_msg->ibm_type == IBLND_MSG_NOOP)
1036 conn->ibc_noops_posted--;
1039 tx->tx_waiting = 0; /* don't wait for peer */
1040 tx->tx_status = -EIO;
1043 idle = (tx->tx_sending == 0) && /* This is the final callback */
1044 !tx->tx_waiting && /* Not waiting for peer */
1045 !tx->tx_queued; /* Not re-queued (PUT_DONE) */
1047 list_del(&tx->tx_list);
1049 kiblnd_conn_addref(conn); /* 1 ref for me.... */
1051 spin_unlock(&conn->ibc_lock);
1054 kiblnd_tx_done(conn->ibc_peer->ibp_ni, tx);
1056 kiblnd_check_sends(conn);
1058 kiblnd_conn_decref(conn); /* ...until here */
1062 kiblnd_init_tx_msg (lnet_ni_t *ni, kib_tx_t *tx, int type, int body_nob)
1064 kib_hca_dev_t *hdev = tx->tx_pool->tpo_hdev;
1065 struct ib_sge *sge = &tx->tx_sge[tx->tx_nwrq];
1066 struct ib_send_wr *wrq = &tx->tx_wrq[tx->tx_nwrq];
1067 int nob = offsetof (kib_msg_t, ibm_u) + body_nob;
1070 LASSERT (tx->tx_nwrq >= 0);
1071 LASSERT (tx->tx_nwrq < IBLND_MAX_RDMA_FRAGS + 1);
1072 LASSERT (nob <= IBLND_MSG_SIZE);
1074 kiblnd_init_msg(tx->tx_msg, type, body_nob);
1076 mr = kiblnd_find_dma_mr(hdev, tx->tx_msgaddr, nob);
1077 LASSERT (mr != NULL);
1079 sge->lkey = mr->lkey;
1080 sge->addr = tx->tx_msgaddr;
1083 memset(wrq, 0, sizeof(*wrq));
1086 wrq->wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_TX);
1089 wrq->opcode = IB_WR_SEND;
1090 wrq->send_flags = IB_SEND_SIGNALED;
1096 kiblnd_init_rdma(kib_conn_t *conn, kib_tx_t *tx, int type,
1097 int resid, kib_rdma_desc_t *dstrd, __u64 dstcookie)
1099 kib_msg_t *ibmsg = tx->tx_msg;
1100 kib_rdma_desc_t *srcrd = tx->tx_rd;
1101 struct ib_sge *sge = &tx->tx_sge[0];
1102 struct ib_send_wr *wrq = &tx->tx_wrq[0];
1108 LASSERT (!in_interrupt());
1109 LASSERT (tx->tx_nwrq == 0);
1110 LASSERT (type == IBLND_MSG_GET_DONE ||
1111 type == IBLND_MSG_PUT_DONE);
1113 srcidx = dstidx = 0;
1116 if (srcidx >= srcrd->rd_nfrags) {
1117 CERROR("Src buffer exhausted: %d frags\n", srcidx);
1122 if (dstidx == dstrd->rd_nfrags) {
1123 CERROR("Dst buffer exhausted: %d frags\n", dstidx);
1128 if (tx->tx_nwrq == IBLND_RDMA_FRAGS(conn->ibc_version)) {
1129 CERROR("RDMA too fragmented for %s (%d): "
1130 "%d/%d src %d/%d dst frags\n",
1131 libcfs_nid2str(conn->ibc_peer->ibp_nid),
1132 IBLND_RDMA_FRAGS(conn->ibc_version),
1133 srcidx, srcrd->rd_nfrags,
1134 dstidx, dstrd->rd_nfrags);
1139 wrknob = MIN(MIN(kiblnd_rd_frag_size(srcrd, srcidx),
1140 kiblnd_rd_frag_size(dstrd, dstidx)), resid);
1142 sge = &tx->tx_sge[tx->tx_nwrq];
1143 sge->addr = kiblnd_rd_frag_addr(srcrd, srcidx);
1144 sge->lkey = kiblnd_rd_frag_key(srcrd, srcidx);
1145 sge->length = wrknob;
1147 wrq = &tx->tx_wrq[tx->tx_nwrq];
1149 wrq->next = wrq + 1;
1150 wrq->wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_RDMA);
1153 wrq->opcode = IB_WR_RDMA_WRITE;
1154 wrq->send_flags = 0;
1156 wrq->wr.rdma.remote_addr = kiblnd_rd_frag_addr(dstrd, dstidx);
1157 wrq->wr.rdma.rkey = kiblnd_rd_frag_key(dstrd, dstidx);
1159 srcidx = kiblnd_rd_consume_frag(srcrd, srcidx, wrknob);
1160 dstidx = kiblnd_rd_consume_frag(dstrd, dstidx, wrknob);
1169 if (rc < 0) /* no RDMA if completing with failure */
1172 ibmsg->ibm_u.completion.ibcm_status = rc;
1173 ibmsg->ibm_u.completion.ibcm_cookie = dstcookie;
1174 kiblnd_init_tx_msg(conn->ibc_peer->ibp_ni, tx,
1175 type, sizeof (kib_completion_msg_t));
1181 kiblnd_queue_tx_locked(kib_tx_t *tx, kib_conn_t *conn)
1183 struct list_head *q;
1185 LASSERT(tx->tx_nwrq > 0); /* work items set up */
1186 LASSERT(!tx->tx_queued); /* not queued for sending already */
1187 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1190 tx->tx_deadline = jiffies +
1191 msecs_to_jiffies(*kiblnd_tunables.kib_timeout *
1194 if (tx->tx_conn == NULL) {
1195 kiblnd_conn_addref(conn);
1197 LASSERT (tx->tx_msg->ibm_type != IBLND_MSG_PUT_DONE);
1199 /* PUT_DONE first attached to conn as a PUT_REQ */
1200 LASSERT (tx->tx_conn == conn);
1201 LASSERT (tx->tx_msg->ibm_type == IBLND_MSG_PUT_DONE);
1204 switch (tx->tx_msg->ibm_type) {
1208 case IBLND_MSG_PUT_REQ:
1209 case IBLND_MSG_GET_REQ:
1210 q = &conn->ibc_tx_queue_rsrvd;
1213 case IBLND_MSG_PUT_NAK:
1214 case IBLND_MSG_PUT_ACK:
1215 case IBLND_MSG_PUT_DONE:
1216 case IBLND_MSG_GET_DONE:
1217 q = &conn->ibc_tx_queue_nocred;
1220 case IBLND_MSG_NOOP:
1221 if (IBLND_OOB_CAPABLE(conn->ibc_version))
1222 q = &conn->ibc_tx_queue_nocred;
1224 q = &conn->ibc_tx_noops;
1227 case IBLND_MSG_IMMEDIATE:
1228 q = &conn->ibc_tx_queue;
1232 list_add_tail(&tx->tx_list, q);
1236 kiblnd_queue_tx (kib_tx_t *tx, kib_conn_t *conn)
1238 spin_lock(&conn->ibc_lock);
1239 kiblnd_queue_tx_locked(tx, conn);
1240 spin_unlock(&conn->ibc_lock);
1242 kiblnd_check_sends(conn);
1245 static int kiblnd_resolve_addr(struct rdma_cm_id *cmid,
1246 struct sockaddr_in *srcaddr,
1247 struct sockaddr_in *dstaddr,
1250 unsigned short port;
1253 #ifdef HAVE_OFED_RDMA_SET_REUSEADDR
1254 /* allow the port to be reused */
1255 rc = rdma_set_reuseaddr(cmid, 1);
1257 CERROR("Unable to set reuse on cmid: %d\n", rc);
1262 /* look for a free privileged port */
1263 for (port = PROT_SOCK-1; port > 0; port--) {
1264 srcaddr->sin_port = htons(port);
1265 rc = rdma_resolve_addr(cmid,
1266 (struct sockaddr *)srcaddr,
1267 (struct sockaddr *)dstaddr,
1270 CDEBUG(D_NET, "bound to port %hu\n", port);
1272 } else if (rc == -EADDRINUSE || rc == -EADDRNOTAVAIL) {
1273 CDEBUG(D_NET, "bind to port %hu failed: %d\n",
1280 CERROR("Failed to bind to a free privileged port\n");
1281 #ifndef HAVE_OFED_RDMA_SET_REUSEADDR
1282 CERROR("You may need IB verbs that supports rdma_set_reuseaddr()\n");
1288 kiblnd_connect_peer (kib_peer_t *peer)
1290 struct rdma_cm_id *cmid;
1292 kib_net_t *net = peer->ibp_ni->ni_data;
1293 struct sockaddr_in srcaddr;
1294 struct sockaddr_in dstaddr;
1297 LASSERT (net != NULL);
1298 LASSERT (peer->ibp_connecting > 0);
1300 cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, peer, RDMA_PS_TCP,
1304 CERROR("Can't create CMID for %s: %ld\n",
1305 libcfs_nid2str(peer->ibp_nid), PTR_ERR(cmid));
1311 memset(&srcaddr, 0, sizeof(srcaddr));
1312 srcaddr.sin_family = AF_INET;
1313 srcaddr.sin_addr.s_addr = htonl(dev->ibd_ifip);
1315 memset(&dstaddr, 0, sizeof(dstaddr));
1316 dstaddr.sin_family = AF_INET;
1317 dstaddr.sin_port = htons(*kiblnd_tunables.kib_service);
1318 dstaddr.sin_addr.s_addr = htonl(LNET_NIDADDR(peer->ibp_nid));
1320 kiblnd_peer_addref(peer); /* cmid's ref */
1322 if (*kiblnd_tunables.kib_use_priv_port) {
1323 rc = kiblnd_resolve_addr(cmid, &srcaddr, &dstaddr,
1324 *kiblnd_tunables.kib_timeout * 1000);
1326 rc = rdma_resolve_addr(cmid,
1327 (struct sockaddr *)&srcaddr,
1328 (struct sockaddr *)&dstaddr,
1329 *kiblnd_tunables.kib_timeout * 1000);
1332 /* Can't initiate address resolution: */
1333 CERROR("Can't resolve addr for %s: %d\n",
1334 libcfs_nid2str(peer->ibp_nid), rc);
1338 LASSERT (cmid->device != NULL);
1339 CDEBUG(D_NET, "%s: connection bound to %s:%u.%u.%u.%u:%s\n",
1340 libcfs_nid2str(peer->ibp_nid), dev->ibd_ifname,
1341 HIPQUAD(dev->ibd_ifip), cmid->device->name);
1346 kiblnd_peer_decref(peer); /* cmid's ref */
1347 rdma_destroy_id(cmid);
1349 kiblnd_peer_connect_failed(peer, 1, rc);
1353 kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid)
1358 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
1359 unsigned long flags;
1362 /* If I get here, I've committed to send, so I complete the tx with
1363 * failure on any problems */
1365 LASSERT (tx == NULL || tx->tx_conn == NULL); /* only set when assigned a conn */
1366 LASSERT (tx == NULL || tx->tx_nwrq > 0); /* work items have been set up */
1368 /* First time, just use a read lock since I expect to find my peer
1370 read_lock_irqsave(g_lock, flags);
1372 peer = kiblnd_find_peer_locked(nid);
1373 if (peer != NULL && !list_empty(&peer->ibp_conns)) {
1374 /* Found a peer with an established connection */
1375 conn = kiblnd_get_conn_locked(peer);
1376 kiblnd_conn_addref(conn); /* 1 ref for me... */
1378 read_unlock_irqrestore(g_lock, flags);
1381 kiblnd_queue_tx(tx, conn);
1382 kiblnd_conn_decref(conn); /* ...to here */
1386 read_unlock(g_lock);
1387 /* Re-try with a write lock */
1390 peer = kiblnd_find_peer_locked(nid);
1392 if (list_empty(&peer->ibp_conns)) {
1393 /* found a peer, but it's still connecting... */
1394 LASSERT (peer->ibp_connecting != 0 ||
1395 peer->ibp_accepting != 0);
1397 list_add_tail(&tx->tx_list,
1398 &peer->ibp_tx_queue);
1399 write_unlock_irqrestore(g_lock, flags);
1401 conn = kiblnd_get_conn_locked(peer);
1402 kiblnd_conn_addref(conn); /* 1 ref for me... */
1404 write_unlock_irqrestore(g_lock, flags);
1407 kiblnd_queue_tx(tx, conn);
1408 kiblnd_conn_decref(conn); /* ...to here */
1413 write_unlock_irqrestore(g_lock, flags);
1415 /* Allocate a peer ready to add to the peer table and retry */
1416 rc = kiblnd_create_peer(ni, &peer, nid);
1418 CERROR("Can't create peer %s\n", libcfs_nid2str(nid));
1420 tx->tx_status = -EHOSTUNREACH;
1422 kiblnd_tx_done(ni, tx);
1427 write_lock_irqsave(g_lock, flags);
1429 peer2 = kiblnd_find_peer_locked(nid);
1430 if (peer2 != NULL) {
1431 if (list_empty(&peer2->ibp_conns)) {
1432 /* found a peer, but it's still connecting... */
1433 LASSERT (peer2->ibp_connecting != 0 ||
1434 peer2->ibp_accepting != 0);
1436 list_add_tail(&tx->tx_list,
1437 &peer2->ibp_tx_queue);
1438 write_unlock_irqrestore(g_lock, flags);
1440 conn = kiblnd_get_conn_locked(peer2);
1441 kiblnd_conn_addref(conn); /* 1 ref for me... */
1443 write_unlock_irqrestore(g_lock, flags);
1446 kiblnd_queue_tx(tx, conn);
1447 kiblnd_conn_decref(conn); /* ...to here */
1450 kiblnd_peer_decref(peer);
1454 /* Brand new peer */
1455 LASSERT (peer->ibp_connecting == 0);
1456 peer->ibp_connecting = 1;
1458 /* always called with a ref on ni, which prevents ni being shutdown */
1459 LASSERT (((kib_net_t *)ni->ni_data)->ibn_shutdown == 0);
1462 list_add_tail(&tx->tx_list, &peer->ibp_tx_queue);
1464 kiblnd_peer_addref(peer);
1465 list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid));
1467 write_unlock_irqrestore(g_lock, flags);
1469 kiblnd_connect_peer(peer);
1470 kiblnd_peer_decref(peer);
1474 kiblnd_send (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg)
1476 lnet_hdr_t *hdr = &lntmsg->msg_hdr;
1477 int type = lntmsg->msg_type;
1478 lnet_process_id_t target = lntmsg->msg_target;
1479 int target_is_router = lntmsg->msg_target_is_router;
1480 int routing = lntmsg->msg_routing;
1481 unsigned int payload_niov = lntmsg->msg_niov;
1482 struct iovec *payload_iov = lntmsg->msg_iov;
1483 lnet_kiov_t *payload_kiov = lntmsg->msg_kiov;
1484 unsigned int payload_offset = lntmsg->msg_offset;
1485 unsigned int payload_nob = lntmsg->msg_len;
1491 /* NB 'private' is different depending on what we're sending.... */
1493 CDEBUG(D_NET, "sending %d bytes in %d frags to %s\n",
1494 payload_nob, payload_niov, libcfs_id2str(target));
1496 LASSERT (payload_nob == 0 || payload_niov > 0);
1497 LASSERT (payload_niov <= LNET_MAX_IOV);
1499 /* Thread context */
1500 LASSERT (!in_interrupt());
1501 /* payload is either all vaddrs or all pages */
1502 LASSERT (!(payload_kiov != NULL && payload_iov != NULL));
1510 LASSERT (payload_nob == 0);
1514 if (routing || target_is_router)
1515 break; /* send IMMEDIATE */
1517 /* is the REPLY message too small for RDMA? */
1518 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[lntmsg->msg_md->md_length]);
1519 if (nob <= IBLND_MSG_SIZE)
1520 break; /* send IMMEDIATE */
1522 tx = kiblnd_get_idle_tx(ni, target.nid);
1524 CERROR("Can't allocate txd for GET to %s\n",
1525 libcfs_nid2str(target.nid));
1531 if ((lntmsg->msg_md->md_options & LNET_MD_KIOV) == 0)
1532 rc = kiblnd_setup_rd_iov(ni, tx,
1533 &ibmsg->ibm_u.get.ibgm_rd,
1534 lntmsg->msg_md->md_niov,
1535 lntmsg->msg_md->md_iov.iov,
1536 0, lntmsg->msg_md->md_length);
1538 rc = kiblnd_setup_rd_kiov(ni, tx,
1539 &ibmsg->ibm_u.get.ibgm_rd,
1540 lntmsg->msg_md->md_niov,
1541 lntmsg->msg_md->md_iov.kiov,
1542 0, lntmsg->msg_md->md_length);
1544 CERROR("Can't setup GET sink for %s: %d\n",
1545 libcfs_nid2str(target.nid), rc);
1546 kiblnd_tx_done(ni, tx);
1550 nob = offsetof(kib_get_msg_t, ibgm_rd.rd_frags[tx->tx_nfrags]);
1551 ibmsg->ibm_u.get.ibgm_cookie = tx->tx_cookie;
1552 ibmsg->ibm_u.get.ibgm_hdr = *hdr;
1554 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_GET_REQ, nob);
1556 tx->tx_lntmsg[1] = lnet_create_reply_msg(ni, lntmsg);
1557 if (tx->tx_lntmsg[1] == NULL) {
1558 CERROR("Can't create reply for GET -> %s\n",
1559 libcfs_nid2str(target.nid));
1560 kiblnd_tx_done(ni, tx);
1564 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg[0,1] on completion */
1565 tx->tx_waiting = 1; /* waiting for GET_DONE */
1566 kiblnd_launch_tx(ni, tx, target.nid);
1569 case LNET_MSG_REPLY:
1571 /* Is the payload small enough not to need RDMA? */
1572 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob]);
1573 if (nob <= IBLND_MSG_SIZE)
1574 break; /* send IMMEDIATE */
1576 tx = kiblnd_get_idle_tx(ni, target.nid);
1578 CERROR("Can't allocate %s txd for %s\n",
1579 type == LNET_MSG_PUT ? "PUT" : "REPLY",
1580 libcfs_nid2str(target.nid));
1584 if (payload_kiov == NULL)
1585 rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1586 payload_niov, payload_iov,
1587 payload_offset, payload_nob);
1589 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1590 payload_niov, payload_kiov,
1591 payload_offset, payload_nob);
1593 CERROR("Can't setup PUT src for %s: %d\n",
1594 libcfs_nid2str(target.nid), rc);
1595 kiblnd_tx_done(ni, tx);
1600 ibmsg->ibm_u.putreq.ibprm_hdr = *hdr;
1601 ibmsg->ibm_u.putreq.ibprm_cookie = tx->tx_cookie;
1602 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_REQ, sizeof(kib_putreq_msg_t));
1604 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1605 tx->tx_waiting = 1; /* waiting for PUT_{ACK,NAK} */
1606 kiblnd_launch_tx(ni, tx, target.nid);
1610 /* send IMMEDIATE */
1612 LASSERT (offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob])
1615 tx = kiblnd_get_idle_tx(ni, target.nid);
1617 CERROR ("Can't send %d to %s: tx descs exhausted\n",
1618 type, libcfs_nid2str(target.nid));
1623 ibmsg->ibm_u.immediate.ibim_hdr = *hdr;
1625 if (payload_kiov != NULL)
1626 lnet_copy_kiov2flat(IBLND_MSG_SIZE, ibmsg,
1627 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1628 payload_niov, payload_kiov,
1629 payload_offset, payload_nob);
1631 lnet_copy_iov2flat(IBLND_MSG_SIZE, ibmsg,
1632 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1633 payload_niov, payload_iov,
1634 payload_offset, payload_nob);
1636 nob = offsetof(kib_immediate_msg_t, ibim_payload[payload_nob]);
1637 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_IMMEDIATE, nob);
1639 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1640 kiblnd_launch_tx(ni, tx, target.nid);
1645 kiblnd_reply (lnet_ni_t *ni, kib_rx_t *rx, lnet_msg_t *lntmsg)
1647 lnet_process_id_t target = lntmsg->msg_target;
1648 unsigned int niov = lntmsg->msg_niov;
1649 struct iovec *iov = lntmsg->msg_iov;
1650 lnet_kiov_t *kiov = lntmsg->msg_kiov;
1651 unsigned int offset = lntmsg->msg_offset;
1652 unsigned int nob = lntmsg->msg_len;
1656 tx = kiblnd_get_idle_tx(ni, rx->rx_conn->ibc_peer->ibp_nid);
1658 CERROR("Can't get tx for REPLY to %s\n",
1659 libcfs_nid2str(target.nid));
1665 else if (kiov == NULL)
1666 rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1667 niov, iov, offset, nob);
1669 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1670 niov, kiov, offset, nob);
1673 CERROR("Can't setup GET src for %s: %d\n",
1674 libcfs_nid2str(target.nid), rc);
1678 rc = kiblnd_init_rdma(rx->rx_conn, tx,
1679 IBLND_MSG_GET_DONE, nob,
1680 &rx->rx_msg->ibm_u.get.ibgm_rd,
1681 rx->rx_msg->ibm_u.get.ibgm_cookie);
1683 CERROR("Can't setup rdma for GET from %s: %d\n",
1684 libcfs_nid2str(target.nid), rc);
1689 /* No RDMA: local completion may happen now! */
1690 lnet_finalize(ni, lntmsg, 0);
1692 /* RDMA: lnet_finalize(lntmsg) when it
1694 tx->tx_lntmsg[0] = lntmsg;
1697 kiblnd_queue_tx(tx, rx->rx_conn);
1701 kiblnd_tx_done(ni, tx);
1703 lnet_finalize(ni, lntmsg, -EIO);
1707 kiblnd_recv (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed,
1708 unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov,
1709 unsigned int offset, unsigned int mlen, unsigned int rlen)
1711 kib_rx_t *rx = private;
1712 kib_msg_t *rxmsg = rx->rx_msg;
1713 kib_conn_t *conn = rx->rx_conn;
1717 int post_credit = IBLND_POSTRX_PEER_CREDIT;
1720 LASSERT (mlen <= rlen);
1721 LASSERT (!in_interrupt());
1722 /* Either all pages or all vaddrs */
1723 LASSERT (!(kiov != NULL && iov != NULL));
1725 switch (rxmsg->ibm_type) {
1729 case IBLND_MSG_IMMEDIATE:
1730 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[rlen]);
1731 if (nob > rx->rx_nob) {
1732 CERROR ("Immediate message from %s too big: %d(%d)\n",
1733 libcfs_nid2str(rxmsg->ibm_u.immediate.ibim_hdr.src_nid),
1740 lnet_copy_flat2kiov(niov, kiov, offset,
1741 IBLND_MSG_SIZE, rxmsg,
1742 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1745 lnet_copy_flat2iov(niov, iov, offset,
1746 IBLND_MSG_SIZE, rxmsg,
1747 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1749 lnet_finalize (ni, lntmsg, 0);
1752 case IBLND_MSG_PUT_REQ:
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 */
1771 rc = kiblnd_setup_rd_iov(ni, tx,
1772 &txmsg->ibm_u.putack.ibpam_rd,
1773 niov, iov, offset, mlen);
1775 rc = kiblnd_setup_rd_kiov(ni, tx,
1776 &txmsg->ibm_u.putack.ibpam_rd,
1777 niov, kiov, offset, mlen);
1779 CERROR("Can't setup PUT sink for %s: %d\n",
1780 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
1781 kiblnd_tx_done(ni, tx);
1782 /* tell peer it's over */
1783 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK, rc,
1784 rxmsg->ibm_u.putreq.ibprm_cookie);
1788 nob = offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[tx->tx_nfrags]);
1789 txmsg->ibm_u.putack.ibpam_src_cookie = rxmsg->ibm_u.putreq.ibprm_cookie;
1790 txmsg->ibm_u.putack.ibpam_dst_cookie = tx->tx_cookie;
1792 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_ACK, nob);
1794 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1795 tx->tx_waiting = 1; /* waiting for PUT_DONE */
1796 kiblnd_queue_tx(tx, conn);
1798 /* reposted buffer reserved for PUT_DONE */
1799 post_credit = IBLND_POSTRX_NO_CREDIT;
1802 case IBLND_MSG_GET_REQ:
1803 if (lntmsg != NULL) {
1804 /* Optimized GET; RDMA lntmsg's payload */
1805 kiblnd_reply(ni, rx, lntmsg);
1807 /* GET didn't match anything */
1808 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_GET_DONE,
1810 rxmsg->ibm_u.get.ibgm_cookie);
1815 kiblnd_post_rx(rx, post_credit);
1820 kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name)
1822 struct task_struct *task = kthread_run(fn, arg, name);
1825 return PTR_ERR(task);
1827 atomic_inc(&kiblnd_data.kib_nthreads);
1832 kiblnd_thread_fini (void)
1834 atomic_dec (&kiblnd_data.kib_nthreads);
1838 kiblnd_peer_alive (kib_peer_t *peer)
1840 /* This is racy, but everyone's only writing cfs_time_current() */
1841 peer->ibp_last_alive = cfs_time_current();
1846 kiblnd_peer_notify (kib_peer_t *peer)
1849 cfs_time_t last_alive = 0;
1850 unsigned long flags;
1852 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1854 if (list_empty(&peer->ibp_conns) &&
1855 peer->ibp_accepting == 0 &&
1856 peer->ibp_connecting == 0 &&
1857 peer->ibp_error != 0) {
1858 error = peer->ibp_error;
1859 peer->ibp_error = 0;
1861 last_alive = peer->ibp_last_alive;
1864 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1867 lnet_notify(peer->ibp_ni,
1868 peer->ibp_nid, 0, last_alive);
1872 kiblnd_close_conn_locked (kib_conn_t *conn, int error)
1874 /* This just does the immediate housekeeping. 'error' is zero for a
1875 * normal shutdown which can happen only after the connection has been
1876 * established. If the connection is established, schedule the
1877 * connection to be finished off by the connd. Otherwise the connd is
1878 * already dealing with it (either to set it up or tear it down).
1879 * Caller holds kib_global_lock exclusively in irq context */
1880 kib_peer_t *peer = conn->ibc_peer;
1882 unsigned long flags;
1884 LASSERT (error != 0 || conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1886 if (error != 0 && conn->ibc_comms_error == 0)
1887 conn->ibc_comms_error = error;
1889 if (conn->ibc_state != IBLND_CONN_ESTABLISHED)
1890 return; /* already being handled */
1893 list_empty(&conn->ibc_tx_noops) &&
1894 list_empty(&conn->ibc_tx_queue) &&
1895 list_empty(&conn->ibc_tx_queue_rsrvd) &&
1896 list_empty(&conn->ibc_tx_queue_nocred) &&
1897 list_empty(&conn->ibc_active_txs)) {
1898 CDEBUG(D_NET, "closing conn to %s\n",
1899 libcfs_nid2str(peer->ibp_nid));
1901 CNETERR("Closing conn to %s: error %d%s%s%s%s%s\n",
1902 libcfs_nid2str(peer->ibp_nid), error,
1903 list_empty(&conn->ibc_tx_queue) ? "" : "(sending)",
1904 list_empty(&conn->ibc_tx_noops) ? "" : "(sending_noops)",
1905 list_empty(&conn->ibc_tx_queue_rsrvd) ?
1906 "" : "(sending_rsrvd)",
1907 list_empty(&conn->ibc_tx_queue_nocred) ?
1908 "" : "(sending_nocred)",
1909 list_empty(&conn->ibc_active_txs) ? "" : "(waiting)");
1912 dev = ((kib_net_t *)peer->ibp_ni->ni_data)->ibn_dev;
1913 list_del(&conn->ibc_list);
1914 /* connd (see below) takes over ibc_list's ref */
1916 if (list_empty(&peer->ibp_conns) && /* no more conns */
1917 kiblnd_peer_active(peer)) { /* still in peer table */
1918 kiblnd_unlink_peer_locked(peer);
1920 /* set/clear error on last conn */
1921 peer->ibp_error = conn->ibc_comms_error;
1924 kiblnd_set_conn_state(conn, IBLND_CONN_CLOSING);
1927 kiblnd_dev_can_failover(dev)) {
1928 list_add_tail(&dev->ibd_fail_list,
1929 &kiblnd_data.kib_failed_devs);
1930 wake_up(&kiblnd_data.kib_failover_waitq);
1933 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
1935 list_add_tail(&conn->ibc_list, &kiblnd_data.kib_connd_conns);
1936 wake_up(&kiblnd_data.kib_connd_waitq);
1938 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
1942 kiblnd_close_conn(kib_conn_t *conn, int error)
1944 unsigned long flags;
1946 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1948 kiblnd_close_conn_locked(conn, error);
1950 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1954 kiblnd_handle_early_rxs(kib_conn_t *conn)
1956 unsigned long flags;
1959 LASSERT(!in_interrupt());
1960 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1962 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1963 while (!list_empty(&conn->ibc_early_rxs)) {
1964 rx = list_entry(conn->ibc_early_rxs.next,
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(kib_conn_t *conn, struct list_head *txs)
1979 struct list_head zombies = LIST_HEAD_INIT(zombies);
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, kib_tx_t, tx_list);
1989 if (txs == &conn->ibc_active_txs) {
1990 LASSERT(!tx->tx_queued);
1991 LASSERT(tx->tx_waiting ||
1992 tx->tx_sending != 0);
1994 LASSERT(tx->tx_queued);
1997 tx->tx_status = -ECONNABORTED;
2000 if (tx->tx_sending == 0) {
2002 list_del(&tx->tx_list);
2003 list_add(&tx->tx_list, &zombies);
2007 spin_unlock(&conn->ibc_lock);
2009 kiblnd_txlist_done(conn->ibc_peer->ibp_ni, &zombies, -ECONNABORTED);
2013 kiblnd_finalise_conn (kib_conn_t *conn)
2015 LASSERT (!in_interrupt());
2016 LASSERT (conn->ibc_state > IBLND_CONN_INIT);
2018 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. */
2023 kiblnd_abort_receives(conn);
2025 /* Complete all tx descs not waiting for sends to complete.
2026 * NB we should be safe from RDMA now that the QP has changed state */
2028 kiblnd_abort_txs(conn, &conn->ibc_tx_noops);
2029 kiblnd_abort_txs(conn, &conn->ibc_tx_queue);
2030 kiblnd_abort_txs(conn, &conn->ibc_tx_queue_rsrvd);
2031 kiblnd_abort_txs(conn, &conn->ibc_tx_queue_nocred);
2032 kiblnd_abort_txs(conn, &conn->ibc_active_txs);
2034 kiblnd_handle_early_rxs(conn);
2038 kiblnd_peer_connect_failed(kib_peer_t *peer, int active, int error)
2040 struct list_head zombies = LIST_HEAD_INIT(zombies);
2041 unsigned long flags;
2043 LASSERT (error != 0);
2044 LASSERT (!in_interrupt());
2046 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2049 LASSERT (peer->ibp_connecting > 0);
2050 peer->ibp_connecting--;
2052 LASSERT (peer->ibp_accepting > 0);
2053 peer->ibp_accepting--;
2056 if (peer->ibp_connecting != 0 ||
2057 peer->ibp_accepting != 0) {
2058 /* another connection attempt under way... */
2059 write_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2064 if (list_empty(&peer->ibp_conns)) {
2065 /* Take peer's blocked transmits to complete with error */
2066 list_add(&zombies, &peer->ibp_tx_queue);
2067 list_del_init(&peer->ibp_tx_queue);
2069 if (kiblnd_peer_active(peer))
2070 kiblnd_unlink_peer_locked(peer);
2072 peer->ibp_error = error;
2074 /* Can't have blocked transmits if there are connections */
2075 LASSERT(list_empty(&peer->ibp_tx_queue));
2078 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2080 kiblnd_peer_notify(peer);
2082 if (list_empty(&zombies))
2085 CNETERR("Deleting messages for %s: connection failed\n",
2086 libcfs_nid2str(peer->ibp_nid));
2088 kiblnd_txlist_done(peer->ibp_ni, &zombies, -EHOSTUNREACH);
2092 kiblnd_connreq_done(kib_conn_t *conn, int status)
2094 kib_peer_t *peer = conn->ibc_peer;
2096 struct list_head txs;
2097 unsigned long flags;
2100 active = (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2102 CDEBUG(D_NET,"%s: active(%d), version(%x), status(%d)\n",
2103 libcfs_nid2str(peer->ibp_nid), active,
2104 conn->ibc_version, status);
2106 LASSERT (!in_interrupt());
2107 LASSERT ((conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT &&
2108 peer->ibp_connecting > 0) ||
2109 (conn->ibc_state == IBLND_CONN_PASSIVE_WAIT &&
2110 peer->ibp_accepting > 0));
2112 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
2113 conn->ibc_connvars = NULL;
2116 /* failed to establish connection */
2117 kiblnd_peer_connect_failed(peer, active, status);
2118 kiblnd_finalise_conn(conn);
2122 /* connection established */
2123 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2125 conn->ibc_last_send = jiffies;
2126 kiblnd_set_conn_state(conn, IBLND_CONN_ESTABLISHED);
2127 kiblnd_peer_alive(peer);
2129 /* Add conn to peer's list and nuke any dangling conns from a different
2130 * peer instance... */
2131 kiblnd_conn_addref(conn); /* +1 ref for ibc_list */
2132 list_add(&conn->ibc_list, &peer->ibp_conns);
2134 peer->ibp_connecting--;
2136 peer->ibp_accepting--;
2138 if (peer->ibp_version == 0) {
2139 peer->ibp_version = conn->ibc_version;
2140 peer->ibp_incarnation = conn->ibc_incarnation;
2143 if (peer->ibp_version != conn->ibc_version ||
2144 peer->ibp_incarnation != conn->ibc_incarnation) {
2145 kiblnd_close_stale_conns_locked(peer, conn->ibc_version,
2146 conn->ibc_incarnation);
2147 peer->ibp_version = conn->ibc_version;
2148 peer->ibp_incarnation = conn->ibc_incarnation;
2151 /* grab pending txs while I have the lock */
2152 list_add(&txs, &peer->ibp_tx_queue);
2153 list_del_init(&peer->ibp_tx_queue);
2155 if (!kiblnd_peer_active(peer) || /* peer has been deleted */
2156 conn->ibc_comms_error != 0) { /* error has happened already */
2157 lnet_ni_t *ni = peer->ibp_ni;
2159 /* start to shut down connection */
2160 kiblnd_close_conn_locked(conn, -ECONNABORTED);
2161 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2163 kiblnd_txlist_done(ni, &txs, -ECONNABORTED);
2168 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2170 /* Schedule blocked txs */
2171 spin_lock(&conn->ibc_lock);
2172 while (!list_empty(&txs)) {
2173 tx = list_entry(txs.next, kib_tx_t, tx_list);
2174 list_del(&tx->tx_list);
2176 kiblnd_queue_tx_locked(tx, conn);
2178 spin_unlock(&conn->ibc_lock);
2180 kiblnd_check_sends(conn);
2182 /* schedule blocked rxs */
2183 kiblnd_handle_early_rxs(conn);
2187 kiblnd_reject(struct rdma_cm_id *cmid, kib_rej_t *rej)
2191 rc = rdma_reject(cmid, rej, sizeof(*rej));
2194 CWARN("Error %d sending reject\n", rc);
2198 kiblnd_passive_connect (struct rdma_cm_id *cmid, void *priv, int priv_nob)
2200 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2201 kib_msg_t *reqmsg = priv;
2207 lnet_ni_t *ni = NULL;
2208 kib_net_t *net = NULL;
2210 struct rdma_conn_param cp;
2212 int version = IBLND_MSG_VERSION;
2213 unsigned long flags;
2215 struct sockaddr_in *peer_addr;
2216 LASSERT (!in_interrupt());
2218 /* cmid inherits 'context' from the corresponding listener id */
2219 ibdev = (kib_dev_t *)cmid->context;
2220 LASSERT (ibdev != NULL);
2222 memset(&rej, 0, sizeof(rej));
2223 rej.ibr_magic = IBLND_MSG_MAGIC;
2224 rej.ibr_why = IBLND_REJECT_FATAL;
2225 rej.ibr_cp.ibcp_max_msg_size = IBLND_MSG_SIZE;
2227 peer_addr = (struct sockaddr_in *)&(cmid->route.addr.dst_addr);
2228 if (*kiblnd_tunables.kib_require_priv_port &&
2229 ntohs(peer_addr->sin_port) >= PROT_SOCK) {
2230 __u32 ip = ntohl(peer_addr->sin_addr.s_addr);
2231 CERROR("Peer's port (%u.%u.%u.%u:%hu) is not privileged\n",
2232 HIPQUAD(ip), ntohs(peer_addr->sin_port));
2236 if (priv_nob < offsetof(kib_msg_t, ibm_type)) {
2237 CERROR("Short connection request\n");
2241 /* Future protocol version compatibility support! If the
2242 * o2iblnd-specific protocol changes, or when LNET unifies
2243 * protocols over all LNDs, the initial connection will
2244 * negotiate a protocol version. I trap this here to avoid
2245 * console errors; the reject tells the peer which protocol I
2247 if (reqmsg->ibm_magic == LNET_PROTO_MAGIC ||
2248 reqmsg->ibm_magic == __swab32(LNET_PROTO_MAGIC))
2250 if (reqmsg->ibm_magic == IBLND_MSG_MAGIC &&
2251 reqmsg->ibm_version != IBLND_MSG_VERSION &&
2252 reqmsg->ibm_version != IBLND_MSG_VERSION_1)
2254 if (reqmsg->ibm_magic == __swab32(IBLND_MSG_MAGIC) &&
2255 reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION) &&
2256 reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION_1))
2259 rc = kiblnd_unpack_msg(reqmsg, priv_nob);
2261 CERROR("Can't parse connection request: %d\n", rc);
2265 nid = reqmsg->ibm_srcnid;
2266 ni = lnet_net2ni(LNET_NIDNET(reqmsg->ibm_dstnid));
2269 net = (kib_net_t *)ni->ni_data;
2270 rej.ibr_incarnation = net->ibn_incarnation;
2273 if (ni == NULL || /* no matching net */
2274 ni->ni_nid != reqmsg->ibm_dstnid || /* right NET, wrong NID! */
2275 net->ibn_dev != ibdev) { /* wrong device */
2276 CERROR("Can't accept %s on %s (%s:%d:%u.%u.%u.%u): "
2277 "bad dst nid %s\n", libcfs_nid2str(nid),
2278 ni == NULL ? "NA" : libcfs_nid2str(ni->ni_nid),
2279 ibdev->ibd_ifname, ibdev->ibd_nnets,
2280 HIPQUAD(ibdev->ibd_ifip),
2281 libcfs_nid2str(reqmsg->ibm_dstnid));
2286 /* check time stamp as soon as possible */
2287 if (reqmsg->ibm_dststamp != 0 &&
2288 reqmsg->ibm_dststamp != net->ibn_incarnation) {
2289 CWARN("Stale connection request\n");
2290 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2294 /* I can accept peer's version */
2295 version = reqmsg->ibm_version;
2297 if (reqmsg->ibm_type != IBLND_MSG_CONNREQ) {
2298 CERROR("Unexpected connreq msg type: %x from %s\n",
2299 reqmsg->ibm_type, libcfs_nid2str(nid));
2303 if (reqmsg->ibm_u.connparams.ibcp_queue_depth !=
2304 IBLND_MSG_QUEUE_SIZE(version)) {
2305 CERROR("Can't accept %s: incompatible queue depth %d (%d wanted)\n",
2306 libcfs_nid2str(nid), reqmsg->ibm_u.connparams.ibcp_queue_depth,
2307 IBLND_MSG_QUEUE_SIZE(version));
2309 if (version == IBLND_MSG_VERSION)
2310 rej.ibr_why = IBLND_REJECT_MSG_QUEUE_SIZE;
2315 if (reqmsg->ibm_u.connparams.ibcp_max_frags !=
2316 IBLND_RDMA_FRAGS(version)) {
2317 CERROR("Can't accept %s(version %x): "
2318 "incompatible max_frags %d (%d wanted)\n",
2319 libcfs_nid2str(nid), version,
2320 reqmsg->ibm_u.connparams.ibcp_max_frags,
2321 IBLND_RDMA_FRAGS(version));
2323 if (version == IBLND_MSG_VERSION)
2324 rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2330 if (reqmsg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2331 CERROR("Can't accept %s: message size %d too big (%d max)\n",
2332 libcfs_nid2str(nid),
2333 reqmsg->ibm_u.connparams.ibcp_max_msg_size,
2338 /* assume 'nid' is a new peer; create */
2339 rc = kiblnd_create_peer(ni, &peer, nid);
2341 CERROR("Can't create peer for %s\n", libcfs_nid2str(nid));
2342 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2346 write_lock_irqsave(g_lock, flags);
2348 peer2 = kiblnd_find_peer_locked(nid);
2349 if (peer2 != NULL) {
2350 if (peer2->ibp_version == 0) {
2351 peer2->ibp_version = version;
2352 peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2355 /* not the guy I've talked with */
2356 if (peer2->ibp_incarnation != reqmsg->ibm_srcstamp ||
2357 peer2->ibp_version != version) {
2358 kiblnd_close_peer_conns_locked(peer2, -ESTALE);
2359 write_unlock_irqrestore(g_lock, flags);
2361 CWARN("Conn stale %s [old ver: %x, new ver: %x]\n",
2362 libcfs_nid2str(nid), peer2->ibp_version, version);
2364 kiblnd_peer_decref(peer);
2365 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2369 /* tie-break connection race in favour of the higher NID */
2370 if (peer2->ibp_connecting != 0 &&
2372 write_unlock_irqrestore(g_lock, flags);
2374 CWARN("Conn race %s\n", libcfs_nid2str(peer2->ibp_nid));
2376 kiblnd_peer_decref(peer);
2377 rej.ibr_why = IBLND_REJECT_CONN_RACE;
2381 peer2->ibp_accepting++;
2382 kiblnd_peer_addref(peer2);
2384 write_unlock_irqrestore(g_lock, flags);
2385 kiblnd_peer_decref(peer);
2388 /* Brand new peer */
2389 LASSERT (peer->ibp_accepting == 0);
2390 LASSERT (peer->ibp_version == 0 &&
2391 peer->ibp_incarnation == 0);
2393 peer->ibp_accepting = 1;
2394 peer->ibp_version = version;
2395 peer->ibp_incarnation = reqmsg->ibm_srcstamp;
2397 /* I have a ref on ni that prevents it being shutdown */
2398 LASSERT (net->ibn_shutdown == 0);
2400 kiblnd_peer_addref(peer);
2401 list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid));
2403 write_unlock_irqrestore(g_lock, flags);
2406 conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_PASSIVE_WAIT, version);
2408 kiblnd_peer_connect_failed(peer, 0, -ENOMEM);
2409 kiblnd_peer_decref(peer);
2410 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2414 /* conn now "owns" cmid, so I return success from here on to ensure the
2415 * CM callback doesn't destroy cmid. */
2417 conn->ibc_incarnation = reqmsg->ibm_srcstamp;
2418 conn->ibc_credits = IBLND_MSG_QUEUE_SIZE(version);
2419 conn->ibc_reserved_credits = IBLND_MSG_QUEUE_SIZE(version);
2420 LASSERT (conn->ibc_credits + conn->ibc_reserved_credits + IBLND_OOB_MSGS(version)
2421 <= IBLND_RX_MSGS(version));
2423 ackmsg = &conn->ibc_connvars->cv_msg;
2424 memset(ackmsg, 0, sizeof(*ackmsg));
2426 kiblnd_init_msg(ackmsg, IBLND_MSG_CONNACK,
2427 sizeof(ackmsg->ibm_u.connparams));
2428 ackmsg->ibm_u.connparams.ibcp_queue_depth = IBLND_MSG_QUEUE_SIZE(version);
2429 ackmsg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2430 ackmsg->ibm_u.connparams.ibcp_max_frags = IBLND_RDMA_FRAGS(version);
2432 kiblnd_pack_msg(ni, ackmsg, version, 0, nid, reqmsg->ibm_srcstamp);
2434 memset(&cp, 0, sizeof(cp));
2435 cp.private_data = ackmsg;
2436 cp.private_data_len = ackmsg->ibm_nob;
2437 cp.responder_resources = 0; /* No atomic ops or RDMA reads */
2438 cp.initiator_depth = 0;
2439 cp.flow_control = 1;
2440 cp.retry_count = *kiblnd_tunables.kib_retry_count;
2441 cp.rnr_retry_count = *kiblnd_tunables.kib_rnr_retry_count;
2443 CDEBUG(D_NET, "Accept %s\n", libcfs_nid2str(nid));
2445 rc = rdma_accept(cmid, &cp);
2447 CERROR("Can't accept %s: %d\n", libcfs_nid2str(nid), rc);
2448 rej.ibr_version = version;
2449 rej.ibr_why = IBLND_REJECT_FATAL;
2451 kiblnd_reject(cmid, &rej);
2452 kiblnd_connreq_done(conn, rc);
2453 kiblnd_conn_decref(conn);
2463 rej.ibr_version = version;
2464 rej.ibr_cp.ibcp_queue_depth = IBLND_MSG_QUEUE_SIZE(version);
2465 rej.ibr_cp.ibcp_max_frags = IBLND_RDMA_FRAGS(version);
2466 kiblnd_reject(cmid, &rej);
2468 return -ECONNREFUSED;
2472 kiblnd_reconnect (kib_conn_t *conn, int version,
2473 __u64 incarnation, int why, kib_connparams_t *cp)
2475 kib_peer_t *peer = conn->ibc_peer;
2478 unsigned long flags;
2480 LASSERT (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2481 LASSERT (peer->ibp_connecting > 0); /* 'conn' at least */
2483 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2485 /* retry connection if it's still needed and no other connection
2486 * attempts (active or passive) are in progress
2487 * NB: reconnect is still needed even when ibp_tx_queue is
2488 * empty if ibp_version != version because reconnect may be
2489 * initiated by kiblnd_query() */
2490 if ((!list_empty(&peer->ibp_tx_queue) ||
2491 peer->ibp_version != version) &&
2492 peer->ibp_connecting == 1 &&
2493 peer->ibp_accepting == 0) {
2495 peer->ibp_connecting++;
2497 peer->ibp_version = version;
2498 peer->ibp_incarnation = incarnation;
2501 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2511 case IBLND_REJECT_CONN_STALE:
2515 case IBLND_REJECT_CONN_RACE:
2516 reason = "conn race";
2519 case IBLND_REJECT_CONN_UNCOMPAT:
2520 reason = "version negotiation";
2524 CNETERR("%s: retrying (%s), %x, %x, "
2525 "queue_dep: %d, max_frag: %d, msg_size: %d\n",
2526 libcfs_nid2str(peer->ibp_nid),
2527 reason, IBLND_MSG_VERSION, version,
2528 cp != NULL? cp->ibcp_queue_depth :IBLND_MSG_QUEUE_SIZE(version),
2529 cp != NULL? cp->ibcp_max_frags : IBLND_RDMA_FRAGS(version),
2530 cp != NULL? cp->ibcp_max_msg_size: IBLND_MSG_SIZE);
2532 kiblnd_connect_peer(peer);
2536 kiblnd_rejected (kib_conn_t *conn, int reason, void *priv, int priv_nob)
2538 kib_peer_t *peer = conn->ibc_peer;
2540 LASSERT (!in_interrupt());
2541 LASSERT (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2544 case IB_CM_REJ_STALE_CONN:
2545 kiblnd_reconnect(conn, IBLND_MSG_VERSION, 0,
2546 IBLND_REJECT_CONN_STALE, NULL);
2549 case IB_CM_REJ_INVALID_SERVICE_ID:
2550 CNETERR("%s rejected: no listener at %d\n",
2551 libcfs_nid2str(peer->ibp_nid),
2552 *kiblnd_tunables.kib_service);
2555 case IB_CM_REJ_CONSUMER_DEFINED:
2556 if (priv_nob >= offsetof(kib_rej_t, ibr_padding)) {
2557 kib_rej_t *rej = priv;
2558 kib_connparams_t *cp = NULL;
2560 __u64 incarnation = -1;
2562 /* NB. default incarnation is -1 because:
2563 * a) V1 will ignore dst incarnation in connreq.
2564 * b) V2 will provide incarnation while rejecting me,
2565 * -1 will be overwrote.
2567 * if I try to connect to a V1 peer with V2 protocol,
2568 * it rejected me then upgrade to V2, I have no idea
2569 * about the upgrading and try to reconnect with V1,
2570 * in this case upgraded V2 can find out I'm trying to
2571 * talk to the old guy and reject me(incarnation is -1).
2574 if (rej->ibr_magic == __swab32(IBLND_MSG_MAGIC) ||
2575 rej->ibr_magic == __swab32(LNET_PROTO_MAGIC)) {
2576 __swab32s(&rej->ibr_magic);
2577 __swab16s(&rej->ibr_version);
2581 if (priv_nob >= sizeof(kib_rej_t) &&
2582 rej->ibr_version > IBLND_MSG_VERSION_1) {
2583 /* priv_nob is always 148 in current version
2584 * of OFED, so we still need to check version.
2585 * (define of IB_CM_REJ_PRIVATE_DATA_SIZE) */
2589 __swab64s(&rej->ibr_incarnation);
2590 __swab16s(&cp->ibcp_queue_depth);
2591 __swab16s(&cp->ibcp_max_frags);
2592 __swab32s(&cp->ibcp_max_msg_size);
2595 incarnation = rej->ibr_incarnation;
2598 if (rej->ibr_magic != IBLND_MSG_MAGIC &&
2599 rej->ibr_magic != LNET_PROTO_MAGIC) {
2600 CERROR("%s rejected: consumer defined fatal error\n",
2601 libcfs_nid2str(peer->ibp_nid));
2605 if (rej->ibr_version != IBLND_MSG_VERSION &&
2606 rej->ibr_version != IBLND_MSG_VERSION_1) {
2607 CERROR("%s rejected: o2iblnd version %x error\n",
2608 libcfs_nid2str(peer->ibp_nid),
2613 if (rej->ibr_why == IBLND_REJECT_FATAL &&
2614 rej->ibr_version == IBLND_MSG_VERSION_1) {
2615 CDEBUG(D_NET, "rejected by old version peer %s: %x\n",
2616 libcfs_nid2str(peer->ibp_nid), rej->ibr_version);
2618 if (conn->ibc_version != IBLND_MSG_VERSION_1)
2619 rej->ibr_why = IBLND_REJECT_CONN_UNCOMPAT;
2622 switch (rej->ibr_why) {
2623 case IBLND_REJECT_CONN_RACE:
2624 case IBLND_REJECT_CONN_STALE:
2625 case IBLND_REJECT_CONN_UNCOMPAT:
2626 kiblnd_reconnect(conn, rej->ibr_version,
2627 incarnation, rej->ibr_why, cp);
2630 case IBLND_REJECT_MSG_QUEUE_SIZE:
2631 CERROR("%s rejected: incompatible message queue depth %d, %d\n",
2632 libcfs_nid2str(peer->ibp_nid),
2633 cp != NULL ? cp->ibcp_queue_depth :
2634 IBLND_MSG_QUEUE_SIZE(rej->ibr_version),
2635 IBLND_MSG_QUEUE_SIZE(conn->ibc_version));
2638 case IBLND_REJECT_RDMA_FRAGS:
2639 CERROR("%s rejected: incompatible # of RDMA fragments %d, %d\n",
2640 libcfs_nid2str(peer->ibp_nid),
2641 cp != NULL ? cp->ibcp_max_frags :
2642 IBLND_RDMA_FRAGS(rej->ibr_version),
2643 IBLND_RDMA_FRAGS(conn->ibc_version));
2646 case IBLND_REJECT_NO_RESOURCES:
2647 CERROR("%s rejected: o2iblnd no resources\n",
2648 libcfs_nid2str(peer->ibp_nid));
2651 case IBLND_REJECT_FATAL:
2652 CERROR("%s rejected: o2iblnd fatal error\n",
2653 libcfs_nid2str(peer->ibp_nid));
2657 CERROR("%s rejected: o2iblnd reason %d\n",
2658 libcfs_nid2str(peer->ibp_nid),
2666 CNETERR("%s rejected: reason %d, size %d\n",
2667 libcfs_nid2str(peer->ibp_nid), reason, priv_nob);
2671 kiblnd_connreq_done(conn, -ECONNREFUSED);
2675 kiblnd_check_connreply (kib_conn_t *conn, void *priv, int priv_nob)
2677 kib_peer_t *peer = conn->ibc_peer;
2678 lnet_ni_t *ni = peer->ibp_ni;
2679 kib_net_t *net = ni->ni_data;
2680 kib_msg_t *msg = priv;
2681 int ver = conn->ibc_version;
2682 int rc = kiblnd_unpack_msg(msg, priv_nob);
2683 unsigned long flags;
2685 LASSERT (net != NULL);
2688 CERROR("Can't unpack connack from %s: %d\n",
2689 libcfs_nid2str(peer->ibp_nid), rc);
2693 if (msg->ibm_type != IBLND_MSG_CONNACK) {
2694 CERROR("Unexpected message %d from %s\n",
2695 msg->ibm_type, libcfs_nid2str(peer->ibp_nid));
2700 if (ver != msg->ibm_version) {
2701 CERROR("%s replied version %x is different with "
2702 "requested version %x\n",
2703 libcfs_nid2str(peer->ibp_nid), msg->ibm_version, ver);
2708 if (msg->ibm_u.connparams.ibcp_queue_depth !=
2709 IBLND_MSG_QUEUE_SIZE(ver)) {
2710 CERROR("%s has incompatible queue depth %d(%d wanted)\n",
2711 libcfs_nid2str(peer->ibp_nid),
2712 msg->ibm_u.connparams.ibcp_queue_depth,
2713 IBLND_MSG_QUEUE_SIZE(ver));
2718 if (msg->ibm_u.connparams.ibcp_max_frags !=
2719 IBLND_RDMA_FRAGS(ver)) {
2720 CERROR("%s has incompatible max_frags %d (%d wanted)\n",
2721 libcfs_nid2str(peer->ibp_nid),
2722 msg->ibm_u.connparams.ibcp_max_frags,
2723 IBLND_RDMA_FRAGS(ver));
2728 if (msg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2729 CERROR("%s max message size %d too big (%d max)\n",
2730 libcfs_nid2str(peer->ibp_nid),
2731 msg->ibm_u.connparams.ibcp_max_msg_size,
2737 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2738 if (msg->ibm_dstnid == ni->ni_nid &&
2739 msg->ibm_dststamp == net->ibn_incarnation)
2743 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2746 CERROR("Bad connection reply from %s, rc = %d, "
2747 "version: %x max_frags: %d\n",
2748 libcfs_nid2str(peer->ibp_nid), rc,
2749 msg->ibm_version, msg->ibm_u.connparams.ibcp_max_frags);
2753 conn->ibc_incarnation = msg->ibm_srcstamp;
2755 conn->ibc_reserved_credits = IBLND_MSG_QUEUE_SIZE(ver);
2756 LASSERT (conn->ibc_credits + conn->ibc_reserved_credits + IBLND_OOB_MSGS(ver)
2757 <= IBLND_RX_MSGS(ver));
2759 kiblnd_connreq_done(conn, 0);
2763 /* NB My QP has already established itself, so I handle anything going
2764 * wrong here by setting ibc_comms_error.
2765 * kiblnd_connreq_done(0) moves the conn state to ESTABLISHED, but then
2766 * immediately tears it down. */
2769 conn->ibc_comms_error = rc;
2770 kiblnd_connreq_done(conn, 0);
2774 kiblnd_active_connect (struct rdma_cm_id *cmid)
2776 kib_peer_t *peer = (kib_peer_t *)cmid->context;
2779 struct rdma_conn_param cp;
2782 unsigned long flags;
2785 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2787 incarnation = peer->ibp_incarnation;
2788 version = (peer->ibp_version == 0) ? IBLND_MSG_VERSION :
2791 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2793 conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_ACTIVE_CONNECT, version);
2795 kiblnd_peer_connect_failed(peer, 1, -ENOMEM);
2796 kiblnd_peer_decref(peer); /* lose cmid's ref */
2800 /* conn "owns" cmid now, so I return success from here on to ensure the
2801 * CM callback doesn't destroy cmid. conn also takes over cmid's ref
2804 msg = &conn->ibc_connvars->cv_msg;
2806 memset(msg, 0, sizeof(*msg));
2807 kiblnd_init_msg(msg, IBLND_MSG_CONNREQ, sizeof(msg->ibm_u.connparams));
2808 msg->ibm_u.connparams.ibcp_queue_depth = IBLND_MSG_QUEUE_SIZE(version);
2809 msg->ibm_u.connparams.ibcp_max_frags = IBLND_RDMA_FRAGS(version);
2810 msg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2812 kiblnd_pack_msg(peer->ibp_ni, msg, version,
2813 0, peer->ibp_nid, incarnation);
2815 memset(&cp, 0, sizeof(cp));
2816 cp.private_data = msg;
2817 cp.private_data_len = msg->ibm_nob;
2818 cp.responder_resources = 0; /* No atomic ops or RDMA reads */
2819 cp.initiator_depth = 0;
2820 cp.flow_control = 1;
2821 cp.retry_count = *kiblnd_tunables.kib_retry_count;
2822 cp.rnr_retry_count = *kiblnd_tunables.kib_rnr_retry_count;
2824 LASSERT(cmid->context == (void *)conn);
2825 LASSERT(conn->ibc_cmid == cmid);
2827 rc = rdma_connect(cmid, &cp);
2829 CERROR("Can't connect to %s: %d\n",
2830 libcfs_nid2str(peer->ibp_nid), rc);
2831 kiblnd_connreq_done(conn, rc);
2832 kiblnd_conn_decref(conn);
2839 kiblnd_cm_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2845 switch (event->event) {
2847 CERROR("Unexpected event: %d, status: %d\n",
2848 event->event, event->status);
2851 case RDMA_CM_EVENT_CONNECT_REQUEST:
2852 /* destroy cmid on failure */
2853 rc = kiblnd_passive_connect(cmid,
2854 (void *)KIBLND_CONN_PARAM(event),
2855 KIBLND_CONN_PARAM_LEN(event));
2856 CDEBUG(D_NET, "connreq: %d\n", rc);
2859 case RDMA_CM_EVENT_ADDR_ERROR:
2860 peer = (kib_peer_t *)cmid->context;
2861 CNETERR("%s: ADDR ERROR %d\n",
2862 libcfs_nid2str(peer->ibp_nid), event->status);
2863 kiblnd_peer_connect_failed(peer, 1, -EHOSTUNREACH);
2864 kiblnd_peer_decref(peer);
2865 return -EHOSTUNREACH; /* rc != 0 destroys cmid */
2867 case RDMA_CM_EVENT_ADDR_RESOLVED:
2868 peer = (kib_peer_t *)cmid->context;
2870 CDEBUG(D_NET,"%s Addr resolved: %d\n",
2871 libcfs_nid2str(peer->ibp_nid), event->status);
2873 if (event->status != 0) {
2874 CNETERR("Can't resolve address for %s: %d\n",
2875 libcfs_nid2str(peer->ibp_nid), event->status);
2878 rc = rdma_resolve_route(
2879 cmid, *kiblnd_tunables.kib_timeout * 1000);
2882 /* Can't initiate route resolution */
2883 CERROR("Can't resolve route for %s: %d\n",
2884 libcfs_nid2str(peer->ibp_nid), rc);
2886 kiblnd_peer_connect_failed(peer, 1, rc);
2887 kiblnd_peer_decref(peer);
2888 return rc; /* rc != 0 destroys cmid */
2890 case RDMA_CM_EVENT_ROUTE_ERROR:
2891 peer = (kib_peer_t *)cmid->context;
2892 CNETERR("%s: ROUTE ERROR %d\n",
2893 libcfs_nid2str(peer->ibp_nid), event->status);
2894 kiblnd_peer_connect_failed(peer, 1, -EHOSTUNREACH);
2895 kiblnd_peer_decref(peer);
2896 return -EHOSTUNREACH; /* rc != 0 destroys cmid */
2898 case RDMA_CM_EVENT_ROUTE_RESOLVED:
2899 peer = (kib_peer_t *)cmid->context;
2900 CDEBUG(D_NET,"%s Route resolved: %d\n",
2901 libcfs_nid2str(peer->ibp_nid), event->status);
2903 if (event->status == 0)
2904 return kiblnd_active_connect(cmid);
2906 CNETERR("Can't resolve route for %s: %d\n",
2907 libcfs_nid2str(peer->ibp_nid), event->status);
2908 kiblnd_peer_connect_failed(peer, 1, event->status);
2909 kiblnd_peer_decref(peer);
2910 return event->status; /* rc != 0 destroys cmid */
2912 case RDMA_CM_EVENT_UNREACHABLE:
2913 conn = (kib_conn_t *)cmid->context;
2914 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
2915 conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
2916 CNETERR("%s: UNREACHABLE %d\n",
2917 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
2918 kiblnd_connreq_done(conn, -ENETDOWN);
2919 kiblnd_conn_decref(conn);
2922 case RDMA_CM_EVENT_CONNECT_ERROR:
2923 conn = (kib_conn_t *)cmid->context;
2924 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
2925 conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
2926 CNETERR("%s: CONNECT ERROR %d\n",
2927 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
2928 kiblnd_connreq_done(conn, -ENOTCONN);
2929 kiblnd_conn_decref(conn);
2932 case RDMA_CM_EVENT_REJECTED:
2933 conn = (kib_conn_t *)cmid->context;
2934 switch (conn->ibc_state) {
2938 case IBLND_CONN_PASSIVE_WAIT:
2939 CERROR ("%s: REJECTED %d\n",
2940 libcfs_nid2str(conn->ibc_peer->ibp_nid),
2942 kiblnd_connreq_done(conn, -ECONNRESET);
2945 case IBLND_CONN_ACTIVE_CONNECT:
2946 kiblnd_rejected(conn, event->status,
2947 (void *)KIBLND_CONN_PARAM(event),
2948 KIBLND_CONN_PARAM_LEN(event));
2951 kiblnd_conn_decref(conn);
2954 case RDMA_CM_EVENT_ESTABLISHED:
2955 conn = (kib_conn_t *)cmid->context;
2956 switch (conn->ibc_state) {
2960 case IBLND_CONN_PASSIVE_WAIT:
2961 CDEBUG(D_NET, "ESTABLISHED (passive): %s\n",
2962 libcfs_nid2str(conn->ibc_peer->ibp_nid));
2963 kiblnd_connreq_done(conn, 0);
2966 case IBLND_CONN_ACTIVE_CONNECT:
2967 CDEBUG(D_NET, "ESTABLISHED(active): %s\n",
2968 libcfs_nid2str(conn->ibc_peer->ibp_nid));
2969 kiblnd_check_connreply(conn,
2970 (void *)KIBLND_CONN_PARAM(event),
2971 KIBLND_CONN_PARAM_LEN(event));
2974 /* net keeps its ref on conn! */
2977 #ifdef HAVE_OFED_RDMA_CMEV_TIMEWAIT_EXIT
2978 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
2979 CDEBUG(D_NET, "Ignore TIMEWAIT_EXIT event\n");
2982 case RDMA_CM_EVENT_DISCONNECTED:
2983 conn = (kib_conn_t *)cmid->context;
2984 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
2985 CERROR("%s DISCONNECTED\n",
2986 libcfs_nid2str(conn->ibc_peer->ibp_nid));
2987 kiblnd_connreq_done(conn, -ECONNRESET);
2989 kiblnd_close_conn(conn, 0);
2991 kiblnd_conn_decref(conn);
2992 cmid->context = NULL;
2995 case RDMA_CM_EVENT_DEVICE_REMOVAL:
2996 LCONSOLE_ERROR_MSG(0x131,
2997 "Received notification of device removal\n"
2998 "Please shutdown LNET to allow this to proceed\n");
2999 /* Can't remove network from underneath LNET for now, so I have
3003 #ifdef HAVE_OFED_RDMA_CMEV_ADDRCHANGE
3004 case RDMA_CM_EVENT_ADDR_CHANGE:
3005 LCONSOLE_INFO("Physical link changed (eg hca/port)\n");
3012 kiblnd_check_txs_locked(kib_conn_t *conn, struct list_head *txs)
3015 struct list_head *ttmp;
3017 list_for_each(ttmp, txs) {
3018 tx = list_entry(ttmp, kib_tx_t, tx_list);
3020 if (txs != &conn->ibc_active_txs) {
3021 LASSERT(tx->tx_queued);
3023 LASSERT(!tx->tx_queued);
3024 LASSERT(tx->tx_waiting || tx->tx_sending != 0);
3027 if (cfs_time_aftereq(jiffies, tx->tx_deadline)) {
3028 CERROR("Timed out tx: %s, %lu seconds\n",
3029 kiblnd_queue2str(conn, txs),
3030 cfs_duration_sec(jiffies - tx->tx_deadline));
3039 kiblnd_conn_timed_out_locked(kib_conn_t *conn)
3041 return kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue) ||
3042 kiblnd_check_txs_locked(conn, &conn->ibc_tx_noops) ||
3043 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_rsrvd) ||
3044 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_nocred) ||
3045 kiblnd_check_txs_locked(conn, &conn->ibc_active_txs);
3049 kiblnd_check_conns (int idx)
3051 struct list_head closes = LIST_HEAD_INIT(closes);
3052 struct list_head checksends = LIST_HEAD_INIT(checksends);
3053 struct list_head *peers = &kiblnd_data.kib_peers[idx];
3054 struct list_head *ptmp;
3057 struct list_head *ctmp;
3058 unsigned long flags;
3060 /* NB. We expect to have a look at all the peers and not find any
3061 * RDMAs to time out, so we just use a shared lock while we
3063 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3065 list_for_each(ptmp, peers) {
3066 peer = list_entry(ptmp, kib_peer_t, ibp_list);
3068 list_for_each(ctmp, &peer->ibp_conns) {
3072 conn = list_entry(ctmp, kib_conn_t, ibc_list);
3074 LASSERT(conn->ibc_state == IBLND_CONN_ESTABLISHED);
3076 spin_lock(&conn->ibc_lock);
3078 sendnoop = kiblnd_need_noop(conn);
3079 timedout = kiblnd_conn_timed_out_locked(conn);
3080 if (!sendnoop && !timedout) {
3081 spin_unlock(&conn->ibc_lock);
3086 CERROR("Timed out RDMA with %s (%lu): "
3087 "c: %u, oc: %u, rc: %u\n",
3088 libcfs_nid2str(peer->ibp_nid),
3089 cfs_duration_sec(cfs_time_current() -
3090 peer->ibp_last_alive),
3092 conn->ibc_outstanding_credits,
3093 conn->ibc_reserved_credits);
3094 list_add(&conn->ibc_connd_list, &closes);
3096 list_add(&conn->ibc_connd_list, &checksends);
3098 /* +ref for 'closes' or 'checksends' */
3099 kiblnd_conn_addref(conn);
3101 spin_unlock(&conn->ibc_lock);
3105 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3107 /* Handle timeout by closing the whole
3108 * connection. We can only be sure RDMA activity
3109 * has ceased once the QP has been modified. */
3110 while (!list_empty(&closes)) {
3111 conn = list_entry(closes.next,
3112 kib_conn_t, ibc_connd_list);
3113 list_del(&conn->ibc_connd_list);
3114 kiblnd_close_conn(conn, -ETIMEDOUT);
3115 kiblnd_conn_decref(conn);
3118 /* In case we have enough credits to return via a
3119 * NOOP, but there were no non-blocking tx descs
3120 * free to do it last time... */
3121 while (!list_empty(&checksends)) {
3122 conn = list_entry(checksends.next,
3123 kib_conn_t, ibc_connd_list);
3124 list_del(&conn->ibc_connd_list);
3125 kiblnd_check_sends(conn);
3126 kiblnd_conn_decref(conn);
3131 kiblnd_disconnect_conn (kib_conn_t *conn)
3133 LASSERT (!in_interrupt());
3134 LASSERT (current == kiblnd_data.kib_connd);
3135 LASSERT (conn->ibc_state == IBLND_CONN_CLOSING);
3137 rdma_disconnect(conn->ibc_cmid);
3138 kiblnd_finalise_conn(conn);
3140 kiblnd_peer_notify(conn->ibc_peer);
3144 kiblnd_connd (void *arg)
3147 unsigned long flags;
3153 unsigned long deadline = jiffies;
3155 cfs_block_allsigs ();
3157 init_waitqueue_entry_current (&wait);
3158 kiblnd_data.kib_connd = current;
3160 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3162 while (!kiblnd_data.kib_shutdown) {
3166 if (!list_empty(&kiblnd_data.kib_connd_zombies)) {
3167 conn = list_entry(kiblnd_data. \
3168 kib_connd_zombies.next,
3169 kib_conn_t, ibc_list);
3170 list_del(&conn->ibc_list);
3172 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock,
3176 kiblnd_destroy_conn(conn);
3178 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3181 if (!list_empty(&kiblnd_data.kib_connd_conns)) {
3182 conn = list_entry(kiblnd_data.kib_connd_conns.next,
3183 kib_conn_t, ibc_list);
3184 list_del(&conn->ibc_list);
3186 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock,
3190 kiblnd_disconnect_conn(conn);
3191 kiblnd_conn_decref(conn);
3193 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3196 /* careful with the jiffy wrap... */
3197 timeout = (int)(deadline - jiffies);
3201 int chunk = kiblnd_data.kib_peer_hash_size;
3203 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
3206 /* Time to check for RDMA timeouts on a few more
3207 * peers: I do checks every 'p' seconds on a
3208 * proportion of the peer table and I need to check
3209 * every connection 'n' times within a timeout
3210 * interval, to ensure I detect a timeout on any
3211 * connection within (n+1)/n times the timeout
3214 if (*kiblnd_tunables.kib_timeout > n * p)
3215 chunk = (chunk * n * p) /
3216 *kiblnd_tunables.kib_timeout;
3220 for (i = 0; i < chunk; i++) {
3221 kiblnd_check_conns(peer_index);
3222 peer_index = (peer_index + 1) %
3223 kiblnd_data.kib_peer_hash_size;
3226 deadline += msecs_to_jiffies(p * MSEC_PER_SEC);
3227 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3233 /* Nothing to do for 'timeout' */
3234 set_current_state(TASK_INTERRUPTIBLE);
3235 add_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3236 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
3238 waitq_timedwait(&wait, TASK_INTERRUPTIBLE, timeout);
3240 set_current_state(TASK_RUNNING);
3241 remove_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3242 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3245 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
3247 kiblnd_thread_fini();
3252 kiblnd_qp_event(struct ib_event *event, void *arg)
3254 kib_conn_t *conn = arg;
3256 switch (event->event) {
3257 case IB_EVENT_COMM_EST:
3258 CDEBUG(D_NET, "%s established\n",
3259 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3263 CERROR("%s: Async QP event type %d\n",
3264 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3270 kiblnd_complete (struct ib_wc *wc)
3272 switch (kiblnd_wreqid2type(wc->wr_id)) {
3276 case IBLND_WID_RDMA:
3277 /* We only get RDMA completion notification if it fails. All
3278 * subsequent work items, including the final SEND will fail
3279 * too. However we can't print out any more info about the
3280 * failing RDMA because 'tx' might be back on the idle list or
3281 * even reused already if we didn't manage to post all our work
3283 CNETERR("RDMA (tx: %p) failed: %d\n",
3284 kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3288 kiblnd_tx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3292 kiblnd_rx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status,
3299 kiblnd_cq_completion(struct ib_cq *cq, void *arg)
3301 /* NB I'm not allowed to schedule this conn once its refcount has
3302 * reached 0. Since fundamentally I'm racing with scheduler threads
3303 * consuming my CQ I could be called after all completions have
3304 * occurred. But in this case, ibc_nrx == 0 && ibc_nsends_posted == 0
3305 * and this CQ is about to be destroyed so I NOOP. */
3306 kib_conn_t *conn = (kib_conn_t *)arg;
3307 struct kib_sched_info *sched = conn->ibc_sched;
3308 unsigned long flags;
3310 LASSERT(cq == conn->ibc_cq);
3312 spin_lock_irqsave(&sched->ibs_lock, flags);
3314 conn->ibc_ready = 1;
3316 if (!conn->ibc_scheduled &&
3317 (conn->ibc_nrx > 0 ||
3318 conn->ibc_nsends_posted > 0)) {
3319 kiblnd_conn_addref(conn); /* +1 ref for sched_conns */
3320 conn->ibc_scheduled = 1;
3321 list_add_tail(&conn->ibc_sched_list, &sched->ibs_conns);
3323 if (waitqueue_active(&sched->ibs_waitq))
3324 wake_up(&sched->ibs_waitq);
3327 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3331 kiblnd_cq_event(struct ib_event *event, void *arg)
3333 kib_conn_t *conn = arg;
3335 CERROR("%s: async CQ event type %d\n",
3336 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3340 kiblnd_scheduler(void *arg)
3342 long id = (long)arg;
3343 struct kib_sched_info *sched;
3346 unsigned long flags;
3352 cfs_block_allsigs();
3354 init_waitqueue_entry_current(&wait);
3356 sched = kiblnd_data.kib_scheds[KIB_THREAD_CPT(id)];
3358 rc = cfs_cpt_bind(lnet_cpt_table(), sched->ibs_cpt);
3360 CWARN("Failed to bind on CPT %d, please verify whether "
3361 "all CPUs are healthy and reload modules if necessary, "
3362 "otherwise your system might under risk of low "
3363 "performance\n", sched->ibs_cpt);
3366 spin_lock_irqsave(&sched->ibs_lock, flags);
3368 while (!kiblnd_data.kib_shutdown) {
3369 if (busy_loops++ >= IBLND_RESCHED) {
3370 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3375 spin_lock_irqsave(&sched->ibs_lock, flags);
3380 if (!list_empty(&sched->ibs_conns)) {
3381 conn = list_entry(sched->ibs_conns.next,
3382 kib_conn_t, ibc_sched_list);
3383 /* take over kib_sched_conns' ref on conn... */
3384 LASSERT(conn->ibc_scheduled);
3385 list_del(&conn->ibc_sched_list);
3386 conn->ibc_ready = 0;
3388 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3390 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3392 rc = ib_req_notify_cq(conn->ibc_cq,
3395 CWARN("%s: ib_req_notify_cq failed: %d, "
3396 "closing connection\n",
3397 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
3398 kiblnd_close_conn(conn, -EIO);
3399 kiblnd_conn_decref(conn);
3400 spin_lock_irqsave(&sched->ibs_lock,
3405 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3409 CWARN("%s: ib_poll_cq failed: %d, "
3410 "closing connection\n",
3411 libcfs_nid2str(conn->ibc_peer->ibp_nid),
3413 kiblnd_close_conn(conn, -EIO);
3414 kiblnd_conn_decref(conn);
3415 spin_lock_irqsave(&sched->ibs_lock, flags);
3419 spin_lock_irqsave(&sched->ibs_lock, flags);
3421 if (rc != 0 || conn->ibc_ready) {
3422 /* There may be another completion waiting; get
3423 * another scheduler to check while I handle
3425 /* +1 ref for sched_conns */
3426 kiblnd_conn_addref(conn);
3427 list_add_tail(&conn->ibc_sched_list,
3429 if (waitqueue_active(&sched->ibs_waitq))
3430 wake_up(&sched->ibs_waitq);
3432 conn->ibc_scheduled = 0;
3436 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3437 kiblnd_complete(&wc);
3439 spin_lock_irqsave(&sched->ibs_lock, flags);
3442 kiblnd_conn_decref(conn); /* ...drop my ref from above */
3449 set_current_state(TASK_INTERRUPTIBLE);
3450 add_wait_queue_exclusive(&sched->ibs_waitq, &wait);
3451 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3453 waitq_wait(&wait, TASK_INTERRUPTIBLE);
3456 remove_wait_queue(&sched->ibs_waitq, &wait);
3457 set_current_state(TASK_RUNNING);
3458 spin_lock_irqsave(&sched->ibs_lock, flags);
3461 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3463 kiblnd_thread_fini();
3468 kiblnd_failover_thread(void *arg)
3470 rwlock_t *glock = &kiblnd_data.kib_global_lock;
3473 unsigned long flags;
3476 LASSERT (*kiblnd_tunables.kib_dev_failover != 0);
3478 cfs_block_allsigs ();
3480 init_waitqueue_entry_current(&wait);
3481 write_lock_irqsave(glock, flags);
3483 while (!kiblnd_data.kib_shutdown) {
3484 int do_failover = 0;
3487 list_for_each_entry(dev, &kiblnd_data.kib_failed_devs,
3489 if (cfs_time_before(cfs_time_current(),
3490 dev->ibd_next_failover))
3497 list_del_init(&dev->ibd_fail_list);
3498 dev->ibd_failover = 1;
3499 write_unlock_irqrestore(glock, flags);
3501 rc = kiblnd_dev_failover(dev);
3503 write_lock_irqsave(glock, flags);
3505 LASSERT (dev->ibd_failover);
3506 dev->ibd_failover = 0;
3507 if (rc >= 0) { /* Device is OK or failover succeed */
3508 dev->ibd_next_failover = cfs_time_shift(3);
3512 /* failed to failover, retry later */
3513 dev->ibd_next_failover =
3514 cfs_time_shift(min(dev->ibd_failed_failover, 10));
3515 if (kiblnd_dev_can_failover(dev)) {
3516 list_add_tail(&dev->ibd_fail_list,
3517 &kiblnd_data.kib_failed_devs);
3523 /* long sleep if no more pending failover */
3524 long_sleep = list_empty(&kiblnd_data.kib_failed_devs);
3526 set_current_state(TASK_INTERRUPTIBLE);
3527 add_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3528 write_unlock_irqrestore(glock, flags);
3530 rc = schedule_timeout(long_sleep ? cfs_time_seconds(10) :
3531 cfs_time_seconds(1));
3532 set_current_state(TASK_RUNNING);
3533 remove_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3534 write_lock_irqsave(glock, flags);
3536 if (!long_sleep || rc != 0)
3539 /* have a long sleep, routine check all active devices,
3540 * we need checking like this because if there is not active
3541 * connection on the dev and no SEND from local, we may listen
3542 * on wrong HCA for ever while there is a bonding failover */
3543 list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3544 if (kiblnd_dev_can_failover(dev)) {
3545 list_add_tail(&dev->ibd_fail_list,
3546 &kiblnd_data.kib_failed_devs);
3551 write_unlock_irqrestore(glock, flags);
3553 kiblnd_thread_fini();