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)
804 kib_msg_t *msg = tx->tx_msg;
805 kib_peer_t *peer = conn->ibc_peer;
806 int ver = conn->ibc_version;
809 struct ib_send_wr *bad_wrq;
811 LASSERT (tx->tx_queued);
812 /* We rely on this for QP sizing */
813 LASSERT (tx->tx_nwrq > 0);
814 LASSERT (tx->tx_nwrq <= 1 + IBLND_RDMA_FRAGS(ver));
816 LASSERT (credit == 0 || credit == 1);
817 LASSERT (conn->ibc_outstanding_credits >= 0);
818 LASSERT (conn->ibc_outstanding_credits <= IBLND_MSG_QUEUE_SIZE(ver));
819 LASSERT (conn->ibc_credits >= 0);
820 LASSERT (conn->ibc_credits <= IBLND_MSG_QUEUE_SIZE(ver));
822 if (conn->ibc_nsends_posted == IBLND_CONCURRENT_SENDS(ver)) {
823 /* tx completions outstanding... */
824 CDEBUG(D_NET, "%s: posted enough\n",
825 libcfs_nid2str(peer->ibp_nid));
829 if (credit != 0 && conn->ibc_credits == 0) { /* no credits */
830 CDEBUG(D_NET, "%s: no credits\n",
831 libcfs_nid2str(peer->ibp_nid));
835 if (credit != 0 && !IBLND_OOB_CAPABLE(ver) &&
836 conn->ibc_credits == 1 && /* last credit reserved */
837 msg->ibm_type != IBLND_MSG_NOOP) { /* for NOOP */
838 CDEBUG(D_NET, "%s: not using last credit\n",
839 libcfs_nid2str(peer->ibp_nid));
843 /* NB don't drop ibc_lock before bumping tx_sending */
844 list_del(&tx->tx_list);
847 if (msg->ibm_type == IBLND_MSG_NOOP &&
848 (!kiblnd_need_noop(conn) || /* redundant NOOP */
849 (IBLND_OOB_CAPABLE(ver) && /* posted enough NOOP */
850 conn->ibc_noops_posted == IBLND_OOB_MSGS(ver)))) {
851 /* OK to drop when posted enough NOOPs, since
852 * kiblnd_check_sends will queue NOOP again when
853 * posted NOOPs complete */
854 spin_unlock(&conn->ibc_lock);
855 kiblnd_tx_done(peer->ibp_ni, tx);
856 spin_lock(&conn->ibc_lock);
857 CDEBUG(D_NET, "%s(%d): redundant or enough NOOP\n",
858 libcfs_nid2str(peer->ibp_nid),
859 conn->ibc_noops_posted);
863 kiblnd_pack_msg(peer->ibp_ni, msg, ver, conn->ibc_outstanding_credits,
864 peer->ibp_nid, conn->ibc_incarnation);
866 conn->ibc_credits -= credit;
867 conn->ibc_outstanding_credits = 0;
868 conn->ibc_nsends_posted++;
869 if (msg->ibm_type == IBLND_MSG_NOOP)
870 conn->ibc_noops_posted++;
872 /* CAVEAT EMPTOR! This tx could be the PUT_DONE of an RDMA
873 * PUT. If so, it was first queued here as a PUT_REQ, sent and
874 * stashed on ibc_active_txs, matched by an incoming PUT_ACK,
875 * and then re-queued here. It's (just) possible that
876 * tx_sending is non-zero if we've not done the tx_complete()
877 * from the first send; hence the ++ rather than = below. */
879 list_add(&tx->tx_list, &conn->ibc_active_txs);
881 /* I'm still holding ibc_lock! */
882 if (conn->ibc_state != IBLND_CONN_ESTABLISHED) {
884 } else if (tx->tx_pool->tpo_pool.po_failed ||
885 conn->ibc_hdev != tx->tx_pool->tpo_hdev) {
886 /* close_conn will launch failover */
889 rc = ib_post_send(conn->ibc_cmid->qp,
890 tx->tx_wrq, &bad_wrq);
893 conn->ibc_last_send = jiffies;
898 /* NB credits are transferred in the actual
899 * message, which can only be the last work item */
900 conn->ibc_credits += credit;
901 conn->ibc_outstanding_credits += msg->ibm_credits;
902 conn->ibc_nsends_posted--;
903 if (msg->ibm_type == IBLND_MSG_NOOP)
904 conn->ibc_noops_posted--;
910 done = (tx->tx_sending == 0);
912 list_del(&tx->tx_list);
914 spin_unlock(&conn->ibc_lock);
916 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
917 CERROR("Error %d posting transmit to %s\n",
918 rc, libcfs_nid2str(peer->ibp_nid));
920 CDEBUG(D_NET, "Error %d posting transmit to %s\n",
921 rc, libcfs_nid2str(peer->ibp_nid));
923 kiblnd_close_conn(conn, rc);
926 kiblnd_tx_done(peer->ibp_ni, tx);
928 spin_lock(&conn->ibc_lock);
934 kiblnd_check_sends (kib_conn_t *conn)
936 int ver = conn->ibc_version;
937 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
940 /* Don't send anything until after the connection is established */
941 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
942 CDEBUG(D_NET, "%s too soon\n",
943 libcfs_nid2str(conn->ibc_peer->ibp_nid));
947 spin_lock(&conn->ibc_lock);
949 LASSERT (conn->ibc_nsends_posted <= IBLND_CONCURRENT_SENDS(ver));
950 LASSERT (!IBLND_OOB_CAPABLE(ver) ||
951 conn->ibc_noops_posted <= IBLND_OOB_MSGS(ver));
952 LASSERT (conn->ibc_reserved_credits >= 0);
954 while (conn->ibc_reserved_credits > 0 &&
955 !list_empty(&conn->ibc_tx_queue_rsrvd)) {
956 tx = list_entry(conn->ibc_tx_queue_rsrvd.next,
958 list_del(&tx->tx_list);
959 list_add_tail(&tx->tx_list, &conn->ibc_tx_queue);
960 conn->ibc_reserved_credits--;
963 if (kiblnd_need_noop(conn)) {
964 spin_unlock(&conn->ibc_lock);
966 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
968 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_NOOP, 0);
970 spin_lock(&conn->ibc_lock);
972 kiblnd_queue_tx_locked(tx, conn);
975 kiblnd_conn_addref(conn); /* 1 ref for me.... (see b21911) */
980 if (!list_empty(&conn->ibc_tx_queue_nocred)) {
982 tx = list_entry(conn->ibc_tx_queue_nocred.next,
984 } else if (!list_empty(&conn->ibc_tx_noops)) {
985 LASSERT (!IBLND_OOB_CAPABLE(ver));
987 tx = list_entry(conn->ibc_tx_noops.next,
989 } else if (!list_empty(&conn->ibc_tx_queue)) {
991 tx = list_entry(conn->ibc_tx_queue.next,
996 if (kiblnd_post_tx_locked(conn, tx, credit) != 0)
1000 spin_unlock(&conn->ibc_lock);
1002 kiblnd_conn_decref(conn); /* ...until here */
1006 kiblnd_tx_complete (kib_tx_t *tx, int status)
1008 int failed = (status != IB_WC_SUCCESS);
1009 kib_conn_t *conn = tx->tx_conn;
1012 LASSERT (tx->tx_sending > 0);
1015 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
1016 CNETERR("Tx -> %s cookie "LPX64
1017 " sending %d waiting %d: failed %d\n",
1018 libcfs_nid2str(conn->ibc_peer->ibp_nid),
1019 tx->tx_cookie, tx->tx_sending, tx->tx_waiting,
1022 kiblnd_close_conn(conn, -EIO);
1024 kiblnd_peer_alive(conn->ibc_peer);
1027 spin_lock(&conn->ibc_lock);
1029 /* I could be racing with rdma completion. Whoever makes 'tx' idle
1030 * gets to free it, which also drops its ref on 'conn'. */
1033 conn->ibc_nsends_posted--;
1034 if (tx->tx_msg->ibm_type == IBLND_MSG_NOOP)
1035 conn->ibc_noops_posted--;
1038 tx->tx_waiting = 0; /* don't wait for peer */
1039 tx->tx_status = -EIO;
1042 idle = (tx->tx_sending == 0) && /* This is the final callback */
1043 !tx->tx_waiting && /* Not waiting for peer */
1044 !tx->tx_queued; /* Not re-queued (PUT_DONE) */
1046 list_del(&tx->tx_list);
1048 kiblnd_conn_addref(conn); /* 1 ref for me.... */
1050 spin_unlock(&conn->ibc_lock);
1053 kiblnd_tx_done(conn->ibc_peer->ibp_ni, tx);
1055 kiblnd_check_sends(conn);
1057 kiblnd_conn_decref(conn); /* ...until here */
1061 kiblnd_init_tx_msg (lnet_ni_t *ni, kib_tx_t *tx, int type, int body_nob)
1063 kib_hca_dev_t *hdev = tx->tx_pool->tpo_hdev;
1064 struct ib_sge *sge = &tx->tx_sge[tx->tx_nwrq];
1065 struct ib_send_wr *wrq = &tx->tx_wrq[tx->tx_nwrq];
1066 int nob = offsetof (kib_msg_t, ibm_u) + body_nob;
1069 LASSERT (tx->tx_nwrq >= 0);
1070 LASSERT (tx->tx_nwrq < IBLND_MAX_RDMA_FRAGS + 1);
1071 LASSERT (nob <= IBLND_MSG_SIZE);
1073 kiblnd_init_msg(tx->tx_msg, type, body_nob);
1075 mr = kiblnd_find_dma_mr(hdev, tx->tx_msgaddr, nob);
1076 LASSERT (mr != NULL);
1078 sge->lkey = mr->lkey;
1079 sge->addr = tx->tx_msgaddr;
1082 memset(wrq, 0, sizeof(*wrq));
1085 wrq->wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_TX);
1088 wrq->opcode = IB_WR_SEND;
1089 wrq->send_flags = IB_SEND_SIGNALED;
1095 kiblnd_init_rdma(kib_conn_t *conn, kib_tx_t *tx, int type,
1096 int resid, kib_rdma_desc_t *dstrd, __u64 dstcookie)
1098 kib_msg_t *ibmsg = tx->tx_msg;
1099 kib_rdma_desc_t *srcrd = tx->tx_rd;
1100 struct ib_sge *sge = &tx->tx_sge[0];
1101 struct ib_send_wr *wrq = &tx->tx_wrq[0];
1107 LASSERT (!in_interrupt());
1108 LASSERT (tx->tx_nwrq == 0);
1109 LASSERT (type == IBLND_MSG_GET_DONE ||
1110 type == IBLND_MSG_PUT_DONE);
1112 srcidx = dstidx = 0;
1115 if (srcidx >= srcrd->rd_nfrags) {
1116 CERROR("Src buffer exhausted: %d frags\n", srcidx);
1121 if (dstidx == dstrd->rd_nfrags) {
1122 CERROR("Dst buffer exhausted: %d frags\n", dstidx);
1127 if (tx->tx_nwrq == IBLND_RDMA_FRAGS(conn->ibc_version)) {
1128 CERROR("RDMA too fragmented for %s (%d): "
1129 "%d/%d src %d/%d dst frags\n",
1130 libcfs_nid2str(conn->ibc_peer->ibp_nid),
1131 IBLND_RDMA_FRAGS(conn->ibc_version),
1132 srcidx, srcrd->rd_nfrags,
1133 dstidx, dstrd->rd_nfrags);
1138 wrknob = MIN(MIN(kiblnd_rd_frag_size(srcrd, srcidx),
1139 kiblnd_rd_frag_size(dstrd, dstidx)), resid);
1141 sge = &tx->tx_sge[tx->tx_nwrq];
1142 sge->addr = kiblnd_rd_frag_addr(srcrd, srcidx);
1143 sge->lkey = kiblnd_rd_frag_key(srcrd, srcidx);
1144 sge->length = wrknob;
1146 wrq = &tx->tx_wrq[tx->tx_nwrq];
1148 wrq->next = wrq + 1;
1149 wrq->wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_RDMA);
1152 wrq->opcode = IB_WR_RDMA_WRITE;
1153 wrq->send_flags = 0;
1155 wrq->wr.rdma.remote_addr = kiblnd_rd_frag_addr(dstrd, dstidx);
1156 wrq->wr.rdma.rkey = kiblnd_rd_frag_key(dstrd, dstidx);
1158 srcidx = kiblnd_rd_consume_frag(srcrd, srcidx, wrknob);
1159 dstidx = kiblnd_rd_consume_frag(dstrd, dstidx, wrknob);
1168 if (rc < 0) /* no RDMA if completing with failure */
1171 ibmsg->ibm_u.completion.ibcm_status = rc;
1172 ibmsg->ibm_u.completion.ibcm_cookie = dstcookie;
1173 kiblnd_init_tx_msg(conn->ibc_peer->ibp_ni, tx,
1174 type, sizeof (kib_completion_msg_t));
1180 kiblnd_queue_tx_locked(kib_tx_t *tx, kib_conn_t *conn)
1182 struct list_head *q;
1184 LASSERT(tx->tx_nwrq > 0); /* work items set up */
1185 LASSERT(!tx->tx_queued); /* not queued for sending already */
1186 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1189 tx->tx_deadline = jiffies + (*kiblnd_tunables.kib_timeout * HZ);
1191 if (tx->tx_conn == NULL) {
1192 kiblnd_conn_addref(conn);
1194 LASSERT (tx->tx_msg->ibm_type != IBLND_MSG_PUT_DONE);
1196 /* PUT_DONE first attached to conn as a PUT_REQ */
1197 LASSERT (tx->tx_conn == conn);
1198 LASSERT (tx->tx_msg->ibm_type == IBLND_MSG_PUT_DONE);
1201 switch (tx->tx_msg->ibm_type) {
1205 case IBLND_MSG_PUT_REQ:
1206 case IBLND_MSG_GET_REQ:
1207 q = &conn->ibc_tx_queue_rsrvd;
1210 case IBLND_MSG_PUT_NAK:
1211 case IBLND_MSG_PUT_ACK:
1212 case IBLND_MSG_PUT_DONE:
1213 case IBLND_MSG_GET_DONE:
1214 q = &conn->ibc_tx_queue_nocred;
1217 case IBLND_MSG_NOOP:
1218 if (IBLND_OOB_CAPABLE(conn->ibc_version))
1219 q = &conn->ibc_tx_queue_nocred;
1221 q = &conn->ibc_tx_noops;
1224 case IBLND_MSG_IMMEDIATE:
1225 q = &conn->ibc_tx_queue;
1229 list_add_tail(&tx->tx_list, q);
1233 kiblnd_queue_tx (kib_tx_t *tx, kib_conn_t *conn)
1235 spin_lock(&conn->ibc_lock);
1236 kiblnd_queue_tx_locked(tx, conn);
1237 spin_unlock(&conn->ibc_lock);
1239 kiblnd_check_sends(conn);
1242 static int kiblnd_resolve_addr(struct rdma_cm_id *cmid,
1243 struct sockaddr_in *srcaddr,
1244 struct sockaddr_in *dstaddr,
1247 unsigned short port;
1250 #ifdef HAVE_OFED_RDMA_SET_REUSEADDR
1251 /* allow the port to be reused */
1252 rc = rdma_set_reuseaddr(cmid, 1);
1254 CERROR("Unable to set reuse on cmid: %d\n", rc);
1259 /* look for a free privileged port */
1260 for (port = PROT_SOCK-1; port > 0; port--) {
1261 srcaddr->sin_port = htons(port);
1262 rc = rdma_resolve_addr(cmid,
1263 (struct sockaddr *)srcaddr,
1264 (struct sockaddr *)dstaddr,
1267 CDEBUG(D_NET, "bound to port %hu\n", port);
1269 } else if (rc == -EADDRINUSE || rc == -EADDRNOTAVAIL) {
1270 CDEBUG(D_NET, "bind to port %hu failed: %d\n",
1277 CERROR("Failed to bind to a free privileged port\n");
1278 #ifndef HAVE_OFED_RDMA_SET_REUSEADDR
1279 CERROR("You may need IB verbs that supports rdma_set_reuseaddr()\n");
1285 kiblnd_connect_peer (kib_peer_t *peer)
1287 struct rdma_cm_id *cmid;
1289 kib_net_t *net = peer->ibp_ni->ni_data;
1290 struct sockaddr_in srcaddr;
1291 struct sockaddr_in dstaddr;
1294 LASSERT (net != NULL);
1295 LASSERT (peer->ibp_connecting > 0);
1297 cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, peer, RDMA_PS_TCP,
1301 CERROR("Can't create CMID for %s: %ld\n",
1302 libcfs_nid2str(peer->ibp_nid), PTR_ERR(cmid));
1308 memset(&srcaddr, 0, sizeof(srcaddr));
1309 srcaddr.sin_family = AF_INET;
1310 srcaddr.sin_addr.s_addr = htonl(dev->ibd_ifip);
1312 memset(&dstaddr, 0, sizeof(dstaddr));
1313 dstaddr.sin_family = AF_INET;
1314 dstaddr.sin_port = htons(*kiblnd_tunables.kib_service);
1315 dstaddr.sin_addr.s_addr = htonl(LNET_NIDADDR(peer->ibp_nid));
1317 kiblnd_peer_addref(peer); /* cmid's ref */
1319 if (*kiblnd_tunables.kib_use_priv_port) {
1320 rc = kiblnd_resolve_addr(cmid, &srcaddr, &dstaddr,
1321 *kiblnd_tunables.kib_timeout * 1000);
1323 rc = rdma_resolve_addr(cmid,
1324 (struct sockaddr *)&srcaddr,
1325 (struct sockaddr *)&dstaddr,
1326 *kiblnd_tunables.kib_timeout * 1000);
1329 /* Can't initiate address resolution: */
1330 CERROR("Can't resolve addr for %s: %d\n",
1331 libcfs_nid2str(peer->ibp_nid), rc);
1335 LASSERT (cmid->device != NULL);
1336 CDEBUG(D_NET, "%s: connection bound to %s:%u.%u.%u.%u:%s\n",
1337 libcfs_nid2str(peer->ibp_nid), dev->ibd_ifname,
1338 HIPQUAD(dev->ibd_ifip), cmid->device->name);
1343 kiblnd_peer_decref(peer); /* cmid's ref */
1344 rdma_destroy_id(cmid);
1346 kiblnd_peer_connect_failed(peer, 1, rc);
1350 kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid)
1355 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
1356 unsigned long flags;
1359 /* If I get here, I've committed to send, so I complete the tx with
1360 * failure on any problems */
1362 LASSERT (tx == NULL || tx->tx_conn == NULL); /* only set when assigned a conn */
1363 LASSERT (tx == NULL || tx->tx_nwrq > 0); /* work items have been set up */
1365 /* First time, just use a read lock since I expect to find my peer
1367 read_lock_irqsave(g_lock, flags);
1369 peer = kiblnd_find_peer_locked(nid);
1370 if (peer != NULL && !list_empty(&peer->ibp_conns)) {
1371 /* Found a peer with an established connection */
1372 conn = kiblnd_get_conn_locked(peer);
1373 kiblnd_conn_addref(conn); /* 1 ref for me... */
1375 read_unlock_irqrestore(g_lock, flags);
1378 kiblnd_queue_tx(tx, conn);
1379 kiblnd_conn_decref(conn); /* ...to here */
1383 read_unlock(g_lock);
1384 /* Re-try with a write lock */
1387 peer = kiblnd_find_peer_locked(nid);
1389 if (list_empty(&peer->ibp_conns)) {
1390 /* found a peer, but it's still connecting... */
1391 LASSERT (peer->ibp_connecting != 0 ||
1392 peer->ibp_accepting != 0);
1394 list_add_tail(&tx->tx_list,
1395 &peer->ibp_tx_queue);
1396 write_unlock_irqrestore(g_lock, flags);
1398 conn = kiblnd_get_conn_locked(peer);
1399 kiblnd_conn_addref(conn); /* 1 ref for me... */
1401 write_unlock_irqrestore(g_lock, flags);
1404 kiblnd_queue_tx(tx, conn);
1405 kiblnd_conn_decref(conn); /* ...to here */
1410 write_unlock_irqrestore(g_lock, flags);
1412 /* Allocate a peer ready to add to the peer table and retry */
1413 rc = kiblnd_create_peer(ni, &peer, nid);
1415 CERROR("Can't create peer %s\n", libcfs_nid2str(nid));
1417 tx->tx_status = -EHOSTUNREACH;
1419 kiblnd_tx_done(ni, tx);
1424 write_lock_irqsave(g_lock, flags);
1426 peer2 = kiblnd_find_peer_locked(nid);
1427 if (peer2 != NULL) {
1428 if (list_empty(&peer2->ibp_conns)) {
1429 /* found a peer, but it's still connecting... */
1430 LASSERT (peer2->ibp_connecting != 0 ||
1431 peer2->ibp_accepting != 0);
1433 list_add_tail(&tx->tx_list,
1434 &peer2->ibp_tx_queue);
1435 write_unlock_irqrestore(g_lock, flags);
1437 conn = kiblnd_get_conn_locked(peer2);
1438 kiblnd_conn_addref(conn); /* 1 ref for me... */
1440 write_unlock_irqrestore(g_lock, flags);
1443 kiblnd_queue_tx(tx, conn);
1444 kiblnd_conn_decref(conn); /* ...to here */
1447 kiblnd_peer_decref(peer);
1451 /* Brand new peer */
1452 LASSERT (peer->ibp_connecting == 0);
1453 peer->ibp_connecting = 1;
1455 /* always called with a ref on ni, which prevents ni being shutdown */
1456 LASSERT (((kib_net_t *)ni->ni_data)->ibn_shutdown == 0);
1459 list_add_tail(&tx->tx_list, &peer->ibp_tx_queue);
1461 kiblnd_peer_addref(peer);
1462 list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid));
1464 write_unlock_irqrestore(g_lock, flags);
1466 kiblnd_connect_peer(peer);
1467 kiblnd_peer_decref(peer);
1471 kiblnd_send (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg)
1473 lnet_hdr_t *hdr = &lntmsg->msg_hdr;
1474 int type = lntmsg->msg_type;
1475 lnet_process_id_t target = lntmsg->msg_target;
1476 int target_is_router = lntmsg->msg_target_is_router;
1477 int routing = lntmsg->msg_routing;
1478 unsigned int payload_niov = lntmsg->msg_niov;
1479 struct iovec *payload_iov = lntmsg->msg_iov;
1480 lnet_kiov_t *payload_kiov = lntmsg->msg_kiov;
1481 unsigned int payload_offset = lntmsg->msg_offset;
1482 unsigned int payload_nob = lntmsg->msg_len;
1488 /* NB 'private' is different depending on what we're sending.... */
1490 CDEBUG(D_NET, "sending %d bytes in %d frags to %s\n",
1491 payload_nob, payload_niov, libcfs_id2str(target));
1493 LASSERT (payload_nob == 0 || payload_niov > 0);
1494 LASSERT (payload_niov <= LNET_MAX_IOV);
1496 /* Thread context */
1497 LASSERT (!in_interrupt());
1498 /* payload is either all vaddrs or all pages */
1499 LASSERT (!(payload_kiov != NULL && payload_iov != NULL));
1507 LASSERT (payload_nob == 0);
1511 if (routing || target_is_router)
1512 break; /* send IMMEDIATE */
1514 /* is the REPLY message too small for RDMA? */
1515 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[lntmsg->msg_md->md_length]);
1516 if (nob <= IBLND_MSG_SIZE)
1517 break; /* send IMMEDIATE */
1519 tx = kiblnd_get_idle_tx(ni, target.nid);
1521 CERROR("Can't allocate txd for GET to %s\n",
1522 libcfs_nid2str(target.nid));
1528 if ((lntmsg->msg_md->md_options & LNET_MD_KIOV) == 0)
1529 rc = kiblnd_setup_rd_iov(ni, tx,
1530 &ibmsg->ibm_u.get.ibgm_rd,
1531 lntmsg->msg_md->md_niov,
1532 lntmsg->msg_md->md_iov.iov,
1533 0, lntmsg->msg_md->md_length);
1535 rc = kiblnd_setup_rd_kiov(ni, tx,
1536 &ibmsg->ibm_u.get.ibgm_rd,
1537 lntmsg->msg_md->md_niov,
1538 lntmsg->msg_md->md_iov.kiov,
1539 0, lntmsg->msg_md->md_length);
1541 CERROR("Can't setup GET sink for %s: %d\n",
1542 libcfs_nid2str(target.nid), rc);
1543 kiblnd_tx_done(ni, tx);
1547 nob = offsetof(kib_get_msg_t, ibgm_rd.rd_frags[tx->tx_nfrags]);
1548 ibmsg->ibm_u.get.ibgm_cookie = tx->tx_cookie;
1549 ibmsg->ibm_u.get.ibgm_hdr = *hdr;
1551 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_GET_REQ, nob);
1553 tx->tx_lntmsg[1] = lnet_create_reply_msg(ni, lntmsg);
1554 if (tx->tx_lntmsg[1] == NULL) {
1555 CERROR("Can't create reply for GET -> %s\n",
1556 libcfs_nid2str(target.nid));
1557 kiblnd_tx_done(ni, tx);
1561 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg[0,1] on completion */
1562 tx->tx_waiting = 1; /* waiting for GET_DONE */
1563 kiblnd_launch_tx(ni, tx, target.nid);
1566 case LNET_MSG_REPLY:
1568 /* Is the payload small enough not to need RDMA? */
1569 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob]);
1570 if (nob <= IBLND_MSG_SIZE)
1571 break; /* send IMMEDIATE */
1573 tx = kiblnd_get_idle_tx(ni, target.nid);
1575 CERROR("Can't allocate %s txd for %s\n",
1576 type == LNET_MSG_PUT ? "PUT" : "REPLY",
1577 libcfs_nid2str(target.nid));
1581 if (payload_kiov == NULL)
1582 rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1583 payload_niov, payload_iov,
1584 payload_offset, payload_nob);
1586 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1587 payload_niov, payload_kiov,
1588 payload_offset, payload_nob);
1590 CERROR("Can't setup PUT src for %s: %d\n",
1591 libcfs_nid2str(target.nid), rc);
1592 kiblnd_tx_done(ni, tx);
1597 ibmsg->ibm_u.putreq.ibprm_hdr = *hdr;
1598 ibmsg->ibm_u.putreq.ibprm_cookie = tx->tx_cookie;
1599 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_REQ, sizeof(kib_putreq_msg_t));
1601 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1602 tx->tx_waiting = 1; /* waiting for PUT_{ACK,NAK} */
1603 kiblnd_launch_tx(ni, tx, target.nid);
1607 /* send IMMEDIATE */
1609 LASSERT (offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob])
1612 tx = kiblnd_get_idle_tx(ni, target.nid);
1614 CERROR ("Can't send %d to %s: tx descs exhausted\n",
1615 type, libcfs_nid2str(target.nid));
1620 ibmsg->ibm_u.immediate.ibim_hdr = *hdr;
1622 if (payload_kiov != NULL)
1623 lnet_copy_kiov2flat(IBLND_MSG_SIZE, ibmsg,
1624 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1625 payload_niov, payload_kiov,
1626 payload_offset, payload_nob);
1628 lnet_copy_iov2flat(IBLND_MSG_SIZE, ibmsg,
1629 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1630 payload_niov, payload_iov,
1631 payload_offset, payload_nob);
1633 nob = offsetof(kib_immediate_msg_t, ibim_payload[payload_nob]);
1634 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_IMMEDIATE, nob);
1636 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1637 kiblnd_launch_tx(ni, tx, target.nid);
1642 kiblnd_reply (lnet_ni_t *ni, kib_rx_t *rx, lnet_msg_t *lntmsg)
1644 lnet_process_id_t target = lntmsg->msg_target;
1645 unsigned int niov = lntmsg->msg_niov;
1646 struct iovec *iov = lntmsg->msg_iov;
1647 lnet_kiov_t *kiov = lntmsg->msg_kiov;
1648 unsigned int offset = lntmsg->msg_offset;
1649 unsigned int nob = lntmsg->msg_len;
1653 tx = kiblnd_get_idle_tx(ni, rx->rx_conn->ibc_peer->ibp_nid);
1655 CERROR("Can't get tx for REPLY to %s\n",
1656 libcfs_nid2str(target.nid));
1662 else if (kiov == NULL)
1663 rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1664 niov, iov, offset, nob);
1666 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1667 niov, kiov, offset, nob);
1670 CERROR("Can't setup GET src for %s: %d\n",
1671 libcfs_nid2str(target.nid), rc);
1675 rc = kiblnd_init_rdma(rx->rx_conn, tx,
1676 IBLND_MSG_GET_DONE, nob,
1677 &rx->rx_msg->ibm_u.get.ibgm_rd,
1678 rx->rx_msg->ibm_u.get.ibgm_cookie);
1680 CERROR("Can't setup rdma for GET from %s: %d\n",
1681 libcfs_nid2str(target.nid), rc);
1686 /* No RDMA: local completion may happen now! */
1687 lnet_finalize(ni, lntmsg, 0);
1689 /* RDMA: lnet_finalize(lntmsg) when it
1691 tx->tx_lntmsg[0] = lntmsg;
1694 kiblnd_queue_tx(tx, rx->rx_conn);
1698 kiblnd_tx_done(ni, tx);
1700 lnet_finalize(ni, lntmsg, -EIO);
1704 kiblnd_recv (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed,
1705 unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov,
1706 unsigned int offset, unsigned int mlen, unsigned int rlen)
1708 kib_rx_t *rx = private;
1709 kib_msg_t *rxmsg = rx->rx_msg;
1710 kib_conn_t *conn = rx->rx_conn;
1714 int post_credit = IBLND_POSTRX_PEER_CREDIT;
1717 LASSERT (mlen <= rlen);
1718 LASSERT (!in_interrupt());
1719 /* Either all pages or all vaddrs */
1720 LASSERT (!(kiov != NULL && iov != NULL));
1722 switch (rxmsg->ibm_type) {
1726 case IBLND_MSG_IMMEDIATE:
1727 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[rlen]);
1728 if (nob > rx->rx_nob) {
1729 CERROR ("Immediate message from %s too big: %d(%d)\n",
1730 libcfs_nid2str(rxmsg->ibm_u.immediate.ibim_hdr.src_nid),
1737 lnet_copy_flat2kiov(niov, kiov, offset,
1738 IBLND_MSG_SIZE, rxmsg,
1739 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1742 lnet_copy_flat2iov(niov, iov, offset,
1743 IBLND_MSG_SIZE, rxmsg,
1744 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1746 lnet_finalize (ni, lntmsg, 0);
1749 case IBLND_MSG_PUT_REQ:
1751 lnet_finalize(ni, lntmsg, 0);
1752 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK, 0,
1753 rxmsg->ibm_u.putreq.ibprm_cookie);
1757 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
1759 CERROR("Can't allocate tx for %s\n",
1760 libcfs_nid2str(conn->ibc_peer->ibp_nid));
1761 /* Not replying will break the connection */
1768 rc = kiblnd_setup_rd_iov(ni, tx,
1769 &txmsg->ibm_u.putack.ibpam_rd,
1770 niov, iov, offset, mlen);
1772 rc = kiblnd_setup_rd_kiov(ni, tx,
1773 &txmsg->ibm_u.putack.ibpam_rd,
1774 niov, kiov, offset, mlen);
1776 CERROR("Can't setup PUT sink for %s: %d\n",
1777 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
1778 kiblnd_tx_done(ni, tx);
1779 /* tell peer it's over */
1780 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK, rc,
1781 rxmsg->ibm_u.putreq.ibprm_cookie);
1785 nob = offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[tx->tx_nfrags]);
1786 txmsg->ibm_u.putack.ibpam_src_cookie = rxmsg->ibm_u.putreq.ibprm_cookie;
1787 txmsg->ibm_u.putack.ibpam_dst_cookie = tx->tx_cookie;
1789 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_ACK, nob);
1791 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1792 tx->tx_waiting = 1; /* waiting for PUT_DONE */
1793 kiblnd_queue_tx(tx, conn);
1795 /* reposted buffer reserved for PUT_DONE */
1796 post_credit = IBLND_POSTRX_NO_CREDIT;
1799 case IBLND_MSG_GET_REQ:
1800 if (lntmsg != NULL) {
1801 /* Optimized GET; RDMA lntmsg's payload */
1802 kiblnd_reply(ni, rx, lntmsg);
1804 /* GET didn't match anything */
1805 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_GET_DONE,
1807 rxmsg->ibm_u.get.ibgm_cookie);
1812 kiblnd_post_rx(rx, post_credit);
1817 kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name)
1819 struct task_struct *task = kthread_run(fn, arg, name);
1822 return PTR_ERR(task);
1824 atomic_inc(&kiblnd_data.kib_nthreads);
1829 kiblnd_thread_fini (void)
1831 atomic_dec (&kiblnd_data.kib_nthreads);
1835 kiblnd_peer_alive (kib_peer_t *peer)
1837 /* This is racy, but everyone's only writing cfs_time_current() */
1838 peer->ibp_last_alive = cfs_time_current();
1843 kiblnd_peer_notify (kib_peer_t *peer)
1846 cfs_time_t last_alive = 0;
1847 unsigned long flags;
1849 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1851 if (list_empty(&peer->ibp_conns) &&
1852 peer->ibp_accepting == 0 &&
1853 peer->ibp_connecting == 0 &&
1854 peer->ibp_error != 0) {
1855 error = peer->ibp_error;
1856 peer->ibp_error = 0;
1858 last_alive = peer->ibp_last_alive;
1861 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1864 lnet_notify(peer->ibp_ni,
1865 peer->ibp_nid, 0, last_alive);
1869 kiblnd_close_conn_locked (kib_conn_t *conn, int error)
1871 /* This just does the immediate housekeeping. 'error' is zero for a
1872 * normal shutdown which can happen only after the connection has been
1873 * established. If the connection is established, schedule the
1874 * connection to be finished off by the connd. Otherwise the connd is
1875 * already dealing with it (either to set it up or tear it down).
1876 * Caller holds kib_global_lock exclusively in irq context */
1877 kib_peer_t *peer = conn->ibc_peer;
1879 unsigned long flags;
1881 LASSERT (error != 0 || conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1883 if (error != 0 && conn->ibc_comms_error == 0)
1884 conn->ibc_comms_error = error;
1886 if (conn->ibc_state != IBLND_CONN_ESTABLISHED)
1887 return; /* already being handled */
1890 list_empty(&conn->ibc_tx_noops) &&
1891 list_empty(&conn->ibc_tx_queue) &&
1892 list_empty(&conn->ibc_tx_queue_rsrvd) &&
1893 list_empty(&conn->ibc_tx_queue_nocred) &&
1894 list_empty(&conn->ibc_active_txs)) {
1895 CDEBUG(D_NET, "closing conn to %s\n",
1896 libcfs_nid2str(peer->ibp_nid));
1898 CNETERR("Closing conn to %s: error %d%s%s%s%s%s\n",
1899 libcfs_nid2str(peer->ibp_nid), error,
1900 list_empty(&conn->ibc_tx_queue) ? "" : "(sending)",
1901 list_empty(&conn->ibc_tx_noops) ? "" : "(sending_noops)",
1902 list_empty(&conn->ibc_tx_queue_rsrvd) ?
1903 "" : "(sending_rsrvd)",
1904 list_empty(&conn->ibc_tx_queue_nocred) ?
1905 "" : "(sending_nocred)",
1906 list_empty(&conn->ibc_active_txs) ? "" : "(waiting)");
1909 dev = ((kib_net_t *)peer->ibp_ni->ni_data)->ibn_dev;
1910 list_del(&conn->ibc_list);
1911 /* connd (see below) takes over ibc_list's ref */
1913 if (list_empty(&peer->ibp_conns) && /* no more conns */
1914 kiblnd_peer_active(peer)) { /* still in peer table */
1915 kiblnd_unlink_peer_locked(peer);
1917 /* set/clear error on last conn */
1918 peer->ibp_error = conn->ibc_comms_error;
1921 kiblnd_set_conn_state(conn, IBLND_CONN_CLOSING);
1924 kiblnd_dev_can_failover(dev)) {
1925 list_add_tail(&dev->ibd_fail_list,
1926 &kiblnd_data.kib_failed_devs);
1927 wake_up(&kiblnd_data.kib_failover_waitq);
1930 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
1932 list_add_tail(&conn->ibc_list, &kiblnd_data.kib_connd_conns);
1933 wake_up(&kiblnd_data.kib_connd_waitq);
1935 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
1939 kiblnd_close_conn(kib_conn_t *conn, int error)
1941 unsigned long flags;
1943 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1945 kiblnd_close_conn_locked(conn, error);
1947 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1951 kiblnd_handle_early_rxs(kib_conn_t *conn)
1953 unsigned long flags;
1956 LASSERT(!in_interrupt());
1957 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1959 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1960 while (!list_empty(&conn->ibc_early_rxs)) {
1961 rx = list_entry(conn->ibc_early_rxs.next,
1963 list_del(&rx->rx_list);
1964 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1966 kiblnd_handle_rx(rx);
1968 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1970 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1974 kiblnd_abort_txs(kib_conn_t *conn, struct list_head *txs)
1976 struct list_head zombies = LIST_HEAD_INIT(zombies);
1977 struct list_head *tmp;
1978 struct list_head *nxt;
1981 spin_lock(&conn->ibc_lock);
1983 list_for_each_safe(tmp, nxt, txs) {
1984 tx = list_entry(tmp, kib_tx_t, tx_list);
1986 if (txs == &conn->ibc_active_txs) {
1987 LASSERT(!tx->tx_queued);
1988 LASSERT(tx->tx_waiting ||
1989 tx->tx_sending != 0);
1991 LASSERT(tx->tx_queued);
1994 tx->tx_status = -ECONNABORTED;
1997 if (tx->tx_sending == 0) {
1999 list_del(&tx->tx_list);
2000 list_add(&tx->tx_list, &zombies);
2004 spin_unlock(&conn->ibc_lock);
2006 kiblnd_txlist_done(conn->ibc_peer->ibp_ni, &zombies, -ECONNABORTED);
2010 kiblnd_finalise_conn (kib_conn_t *conn)
2012 LASSERT (!in_interrupt());
2013 LASSERT (conn->ibc_state > IBLND_CONN_INIT);
2015 kiblnd_set_conn_state(conn, IBLND_CONN_DISCONNECTED);
2017 /* abort_receives moves QP state to IB_QPS_ERR. This is only required
2018 * for connections that didn't get as far as being connected, because
2019 * rdma_disconnect() does this for free. */
2020 kiblnd_abort_receives(conn);
2022 /* Complete all tx descs not waiting for sends to complete.
2023 * NB we should be safe from RDMA now that the QP has changed state */
2025 kiblnd_abort_txs(conn, &conn->ibc_tx_noops);
2026 kiblnd_abort_txs(conn, &conn->ibc_tx_queue);
2027 kiblnd_abort_txs(conn, &conn->ibc_tx_queue_rsrvd);
2028 kiblnd_abort_txs(conn, &conn->ibc_tx_queue_nocred);
2029 kiblnd_abort_txs(conn, &conn->ibc_active_txs);
2031 kiblnd_handle_early_rxs(conn);
2035 kiblnd_peer_connect_failed(kib_peer_t *peer, int active, int error)
2037 struct list_head zombies = LIST_HEAD_INIT(zombies);
2038 unsigned long flags;
2040 LASSERT (error != 0);
2041 LASSERT (!in_interrupt());
2043 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2046 LASSERT (peer->ibp_connecting > 0);
2047 peer->ibp_connecting--;
2049 LASSERT (peer->ibp_accepting > 0);
2050 peer->ibp_accepting--;
2053 if (peer->ibp_connecting != 0 ||
2054 peer->ibp_accepting != 0) {
2055 /* another connection attempt under way... */
2056 write_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2061 if (list_empty(&peer->ibp_conns)) {
2062 /* Take peer's blocked transmits to complete with error */
2063 list_add(&zombies, &peer->ibp_tx_queue);
2064 list_del_init(&peer->ibp_tx_queue);
2066 if (kiblnd_peer_active(peer))
2067 kiblnd_unlink_peer_locked(peer);
2069 peer->ibp_error = error;
2071 /* Can't have blocked transmits if there are connections */
2072 LASSERT(list_empty(&peer->ibp_tx_queue));
2075 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2077 kiblnd_peer_notify(peer);
2079 if (list_empty(&zombies))
2082 CNETERR("Deleting messages for %s: connection failed\n",
2083 libcfs_nid2str(peer->ibp_nid));
2085 kiblnd_txlist_done(peer->ibp_ni, &zombies, -EHOSTUNREACH);
2089 kiblnd_connreq_done(kib_conn_t *conn, int status)
2091 kib_peer_t *peer = conn->ibc_peer;
2093 struct list_head txs;
2094 unsigned long flags;
2097 active = (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2099 CDEBUG(D_NET,"%s: active(%d), version(%x), status(%d)\n",
2100 libcfs_nid2str(peer->ibp_nid), active,
2101 conn->ibc_version, status);
2103 LASSERT (!in_interrupt());
2104 LASSERT ((conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT &&
2105 peer->ibp_connecting > 0) ||
2106 (conn->ibc_state == IBLND_CONN_PASSIVE_WAIT &&
2107 peer->ibp_accepting > 0));
2109 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
2110 conn->ibc_connvars = NULL;
2113 /* failed to establish connection */
2114 kiblnd_peer_connect_failed(peer, active, status);
2115 kiblnd_finalise_conn(conn);
2119 /* connection established */
2120 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2122 conn->ibc_last_send = jiffies;
2123 kiblnd_set_conn_state(conn, IBLND_CONN_ESTABLISHED);
2124 kiblnd_peer_alive(peer);
2126 /* Add conn to peer's list and nuke any dangling conns from a different
2127 * peer instance... */
2128 kiblnd_conn_addref(conn); /* +1 ref for ibc_list */
2129 list_add(&conn->ibc_list, &peer->ibp_conns);
2131 peer->ibp_connecting--;
2133 peer->ibp_accepting--;
2135 if (peer->ibp_version == 0) {
2136 peer->ibp_version = conn->ibc_version;
2137 peer->ibp_incarnation = conn->ibc_incarnation;
2140 if (peer->ibp_version != conn->ibc_version ||
2141 peer->ibp_incarnation != conn->ibc_incarnation) {
2142 kiblnd_close_stale_conns_locked(peer, conn->ibc_version,
2143 conn->ibc_incarnation);
2144 peer->ibp_version = conn->ibc_version;
2145 peer->ibp_incarnation = conn->ibc_incarnation;
2148 /* grab pending txs while I have the lock */
2149 list_add(&txs, &peer->ibp_tx_queue);
2150 list_del_init(&peer->ibp_tx_queue);
2152 if (!kiblnd_peer_active(peer) || /* peer has been deleted */
2153 conn->ibc_comms_error != 0) { /* error has happened already */
2154 lnet_ni_t *ni = peer->ibp_ni;
2156 /* start to shut down connection */
2157 kiblnd_close_conn_locked(conn, -ECONNABORTED);
2158 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2160 kiblnd_txlist_done(ni, &txs, -ECONNABORTED);
2165 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2167 /* Schedule blocked txs */
2168 spin_lock(&conn->ibc_lock);
2169 while (!list_empty(&txs)) {
2170 tx = list_entry(txs.next, kib_tx_t, tx_list);
2171 list_del(&tx->tx_list);
2173 kiblnd_queue_tx_locked(tx, conn);
2175 spin_unlock(&conn->ibc_lock);
2177 kiblnd_check_sends(conn);
2179 /* schedule blocked rxs */
2180 kiblnd_handle_early_rxs(conn);
2184 kiblnd_reject(struct rdma_cm_id *cmid, kib_rej_t *rej)
2188 rc = rdma_reject(cmid, rej, sizeof(*rej));
2191 CWARN("Error %d sending reject\n", rc);
2195 kiblnd_passive_connect (struct rdma_cm_id *cmid, void *priv, int priv_nob)
2197 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2198 kib_msg_t *reqmsg = priv;
2204 lnet_ni_t *ni = NULL;
2205 kib_net_t *net = NULL;
2207 struct rdma_conn_param cp;
2209 int version = IBLND_MSG_VERSION;
2210 unsigned long flags;
2212 struct sockaddr_in *peer_addr;
2213 LASSERT (!in_interrupt());
2215 /* cmid inherits 'context' from the corresponding listener id */
2216 ibdev = (kib_dev_t *)cmid->context;
2217 LASSERT (ibdev != NULL);
2219 memset(&rej, 0, sizeof(rej));
2220 rej.ibr_magic = IBLND_MSG_MAGIC;
2221 rej.ibr_why = IBLND_REJECT_FATAL;
2222 rej.ibr_cp.ibcp_max_msg_size = IBLND_MSG_SIZE;
2224 peer_addr = (struct sockaddr_in *)&(cmid->route.addr.dst_addr);
2225 if (*kiblnd_tunables.kib_require_priv_port &&
2226 ntohs(peer_addr->sin_port) >= PROT_SOCK) {
2227 __u32 ip = ntohl(peer_addr->sin_addr.s_addr);
2228 CERROR("Peer's port (%u.%u.%u.%u:%hu) is not privileged\n",
2229 HIPQUAD(ip), ntohs(peer_addr->sin_port));
2233 if (priv_nob < offsetof(kib_msg_t, ibm_type)) {
2234 CERROR("Short connection request\n");
2238 /* Future protocol version compatibility support! If the
2239 * o2iblnd-specific protocol changes, or when LNET unifies
2240 * protocols over all LNDs, the initial connection will
2241 * negotiate a protocol version. I trap this here to avoid
2242 * console errors; the reject tells the peer which protocol I
2244 if (reqmsg->ibm_magic == LNET_PROTO_MAGIC ||
2245 reqmsg->ibm_magic == __swab32(LNET_PROTO_MAGIC))
2247 if (reqmsg->ibm_magic == IBLND_MSG_MAGIC &&
2248 reqmsg->ibm_version != IBLND_MSG_VERSION &&
2249 reqmsg->ibm_version != IBLND_MSG_VERSION_1)
2251 if (reqmsg->ibm_magic == __swab32(IBLND_MSG_MAGIC) &&
2252 reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION) &&
2253 reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION_1))
2256 rc = kiblnd_unpack_msg(reqmsg, priv_nob);
2258 CERROR("Can't parse connection request: %d\n", rc);
2262 nid = reqmsg->ibm_srcnid;
2263 ni = lnet_net2ni(LNET_NIDNET(reqmsg->ibm_dstnid));
2266 net = (kib_net_t *)ni->ni_data;
2267 rej.ibr_incarnation = net->ibn_incarnation;
2270 if (ni == NULL || /* no matching net */
2271 ni->ni_nid != reqmsg->ibm_dstnid || /* right NET, wrong NID! */
2272 net->ibn_dev != ibdev) { /* wrong device */
2273 CERROR("Can't accept %s on %s (%s:%d:%u.%u.%u.%u): "
2274 "bad dst nid %s\n", libcfs_nid2str(nid),
2275 ni == NULL ? "NA" : libcfs_nid2str(ni->ni_nid),
2276 ibdev->ibd_ifname, ibdev->ibd_nnets,
2277 HIPQUAD(ibdev->ibd_ifip),
2278 libcfs_nid2str(reqmsg->ibm_dstnid));
2283 /* check time stamp as soon as possible */
2284 if (reqmsg->ibm_dststamp != 0 &&
2285 reqmsg->ibm_dststamp != net->ibn_incarnation) {
2286 CWARN("Stale connection request\n");
2287 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2291 /* I can accept peer's version */
2292 version = reqmsg->ibm_version;
2294 if (reqmsg->ibm_type != IBLND_MSG_CONNREQ) {
2295 CERROR("Unexpected connreq msg type: %x from %s\n",
2296 reqmsg->ibm_type, libcfs_nid2str(nid));
2300 if (reqmsg->ibm_u.connparams.ibcp_queue_depth !=
2301 IBLND_MSG_QUEUE_SIZE(version)) {
2302 CERROR("Can't accept %s: incompatible queue depth %d (%d wanted)\n",
2303 libcfs_nid2str(nid), reqmsg->ibm_u.connparams.ibcp_queue_depth,
2304 IBLND_MSG_QUEUE_SIZE(version));
2306 if (version == IBLND_MSG_VERSION)
2307 rej.ibr_why = IBLND_REJECT_MSG_QUEUE_SIZE;
2312 if (reqmsg->ibm_u.connparams.ibcp_max_frags !=
2313 IBLND_RDMA_FRAGS(version)) {
2314 CERROR("Can't accept %s(version %x): "
2315 "incompatible max_frags %d (%d wanted)\n",
2316 libcfs_nid2str(nid), version,
2317 reqmsg->ibm_u.connparams.ibcp_max_frags,
2318 IBLND_RDMA_FRAGS(version));
2320 if (version == IBLND_MSG_VERSION)
2321 rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2327 if (reqmsg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2328 CERROR("Can't accept %s: message size %d too big (%d max)\n",
2329 libcfs_nid2str(nid),
2330 reqmsg->ibm_u.connparams.ibcp_max_msg_size,
2335 /* assume 'nid' is a new peer; create */
2336 rc = kiblnd_create_peer(ni, &peer, nid);
2338 CERROR("Can't create peer for %s\n", libcfs_nid2str(nid));
2339 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2343 write_lock_irqsave(g_lock, flags);
2345 peer2 = kiblnd_find_peer_locked(nid);
2346 if (peer2 != NULL) {
2347 if (peer2->ibp_version == 0) {
2348 peer2->ibp_version = version;
2349 peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2352 /* not the guy I've talked with */
2353 if (peer2->ibp_incarnation != reqmsg->ibm_srcstamp ||
2354 peer2->ibp_version != version) {
2355 kiblnd_close_peer_conns_locked(peer2, -ESTALE);
2356 write_unlock_irqrestore(g_lock, flags);
2358 CWARN("Conn stale %s [old ver: %x, new ver: %x]\n",
2359 libcfs_nid2str(nid), peer2->ibp_version, version);
2361 kiblnd_peer_decref(peer);
2362 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2366 /* tie-break connection race in favour of the higher NID */
2367 if (peer2->ibp_connecting != 0 &&
2369 write_unlock_irqrestore(g_lock, flags);
2371 CWARN("Conn race %s\n", libcfs_nid2str(peer2->ibp_nid));
2373 kiblnd_peer_decref(peer);
2374 rej.ibr_why = IBLND_REJECT_CONN_RACE;
2378 peer2->ibp_accepting++;
2379 kiblnd_peer_addref(peer2);
2381 write_unlock_irqrestore(g_lock, flags);
2382 kiblnd_peer_decref(peer);
2385 /* Brand new peer */
2386 LASSERT (peer->ibp_accepting == 0);
2387 LASSERT (peer->ibp_version == 0 &&
2388 peer->ibp_incarnation == 0);
2390 peer->ibp_accepting = 1;
2391 peer->ibp_version = version;
2392 peer->ibp_incarnation = reqmsg->ibm_srcstamp;
2394 /* I have a ref on ni that prevents it being shutdown */
2395 LASSERT (net->ibn_shutdown == 0);
2397 kiblnd_peer_addref(peer);
2398 list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid));
2400 write_unlock_irqrestore(g_lock, flags);
2403 conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_PASSIVE_WAIT, version);
2405 kiblnd_peer_connect_failed(peer, 0, -ENOMEM);
2406 kiblnd_peer_decref(peer);
2407 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2411 /* conn now "owns" cmid, so I return success from here on to ensure the
2412 * CM callback doesn't destroy cmid. */
2414 conn->ibc_incarnation = reqmsg->ibm_srcstamp;
2415 conn->ibc_credits = IBLND_MSG_QUEUE_SIZE(version);
2416 conn->ibc_reserved_credits = IBLND_MSG_QUEUE_SIZE(version);
2417 LASSERT (conn->ibc_credits + conn->ibc_reserved_credits + IBLND_OOB_MSGS(version)
2418 <= IBLND_RX_MSGS(version));
2420 ackmsg = &conn->ibc_connvars->cv_msg;
2421 memset(ackmsg, 0, sizeof(*ackmsg));
2423 kiblnd_init_msg(ackmsg, IBLND_MSG_CONNACK,
2424 sizeof(ackmsg->ibm_u.connparams));
2425 ackmsg->ibm_u.connparams.ibcp_queue_depth = IBLND_MSG_QUEUE_SIZE(version);
2426 ackmsg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2427 ackmsg->ibm_u.connparams.ibcp_max_frags = IBLND_RDMA_FRAGS(version);
2429 kiblnd_pack_msg(ni, ackmsg, version, 0, nid, reqmsg->ibm_srcstamp);
2431 memset(&cp, 0, sizeof(cp));
2432 cp.private_data = ackmsg;
2433 cp.private_data_len = ackmsg->ibm_nob;
2434 cp.responder_resources = 0; /* No atomic ops or RDMA reads */
2435 cp.initiator_depth = 0;
2436 cp.flow_control = 1;
2437 cp.retry_count = *kiblnd_tunables.kib_retry_count;
2438 cp.rnr_retry_count = *kiblnd_tunables.kib_rnr_retry_count;
2440 CDEBUG(D_NET, "Accept %s\n", libcfs_nid2str(nid));
2442 rc = rdma_accept(cmid, &cp);
2444 CERROR("Can't accept %s: %d\n", libcfs_nid2str(nid), rc);
2445 rej.ibr_version = version;
2446 rej.ibr_why = IBLND_REJECT_FATAL;
2448 kiblnd_reject(cmid, &rej);
2449 kiblnd_connreq_done(conn, rc);
2450 kiblnd_conn_decref(conn);
2460 rej.ibr_version = version;
2461 rej.ibr_cp.ibcp_queue_depth = IBLND_MSG_QUEUE_SIZE(version);
2462 rej.ibr_cp.ibcp_max_frags = IBLND_RDMA_FRAGS(version);
2463 kiblnd_reject(cmid, &rej);
2465 return -ECONNREFUSED;
2469 kiblnd_reconnect (kib_conn_t *conn, int version,
2470 __u64 incarnation, int why, kib_connparams_t *cp)
2472 kib_peer_t *peer = conn->ibc_peer;
2475 unsigned long flags;
2477 LASSERT (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2478 LASSERT (peer->ibp_connecting > 0); /* 'conn' at least */
2480 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2482 /* retry connection if it's still needed and no other connection
2483 * attempts (active or passive) are in progress
2484 * NB: reconnect is still needed even when ibp_tx_queue is
2485 * empty if ibp_version != version because reconnect may be
2486 * initiated by kiblnd_query() */
2487 if ((!list_empty(&peer->ibp_tx_queue) ||
2488 peer->ibp_version != version) &&
2489 peer->ibp_connecting == 1 &&
2490 peer->ibp_accepting == 0) {
2492 peer->ibp_connecting++;
2494 peer->ibp_version = version;
2495 peer->ibp_incarnation = incarnation;
2498 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2508 case IBLND_REJECT_CONN_STALE:
2512 case IBLND_REJECT_CONN_RACE:
2513 reason = "conn race";
2516 case IBLND_REJECT_CONN_UNCOMPAT:
2517 reason = "version negotiation";
2521 CNETERR("%s: retrying (%s), %x, %x, "
2522 "queue_dep: %d, max_frag: %d, msg_size: %d\n",
2523 libcfs_nid2str(peer->ibp_nid),
2524 reason, IBLND_MSG_VERSION, version,
2525 cp != NULL? cp->ibcp_queue_depth :IBLND_MSG_QUEUE_SIZE(version),
2526 cp != NULL? cp->ibcp_max_frags : IBLND_RDMA_FRAGS(version),
2527 cp != NULL? cp->ibcp_max_msg_size: IBLND_MSG_SIZE);
2529 kiblnd_connect_peer(peer);
2533 kiblnd_rejected (kib_conn_t *conn, int reason, void *priv, int priv_nob)
2535 kib_peer_t *peer = conn->ibc_peer;
2537 LASSERT (!in_interrupt());
2538 LASSERT (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2541 case IB_CM_REJ_STALE_CONN:
2542 kiblnd_reconnect(conn, IBLND_MSG_VERSION, 0,
2543 IBLND_REJECT_CONN_STALE, NULL);
2546 case IB_CM_REJ_INVALID_SERVICE_ID:
2547 CNETERR("%s rejected: no listener at %d\n",
2548 libcfs_nid2str(peer->ibp_nid),
2549 *kiblnd_tunables.kib_service);
2552 case IB_CM_REJ_CONSUMER_DEFINED:
2553 if (priv_nob >= offsetof(kib_rej_t, ibr_padding)) {
2554 kib_rej_t *rej = priv;
2555 kib_connparams_t *cp = NULL;
2557 __u64 incarnation = -1;
2559 /* NB. default incarnation is -1 because:
2560 * a) V1 will ignore dst incarnation in connreq.
2561 * b) V2 will provide incarnation while rejecting me,
2562 * -1 will be overwrote.
2564 * if I try to connect to a V1 peer with V2 protocol,
2565 * it rejected me then upgrade to V2, I have no idea
2566 * about the upgrading and try to reconnect with V1,
2567 * in this case upgraded V2 can find out I'm trying to
2568 * talk to the old guy and reject me(incarnation is -1).
2571 if (rej->ibr_magic == __swab32(IBLND_MSG_MAGIC) ||
2572 rej->ibr_magic == __swab32(LNET_PROTO_MAGIC)) {
2573 __swab32s(&rej->ibr_magic);
2574 __swab16s(&rej->ibr_version);
2578 if (priv_nob >= sizeof(kib_rej_t) &&
2579 rej->ibr_version > IBLND_MSG_VERSION_1) {
2580 /* priv_nob is always 148 in current version
2581 * of OFED, so we still need to check version.
2582 * (define of IB_CM_REJ_PRIVATE_DATA_SIZE) */
2586 __swab64s(&rej->ibr_incarnation);
2587 __swab16s(&cp->ibcp_queue_depth);
2588 __swab16s(&cp->ibcp_max_frags);
2589 __swab32s(&cp->ibcp_max_msg_size);
2592 incarnation = rej->ibr_incarnation;
2595 if (rej->ibr_magic != IBLND_MSG_MAGIC &&
2596 rej->ibr_magic != LNET_PROTO_MAGIC) {
2597 CERROR("%s rejected: consumer defined fatal error\n",
2598 libcfs_nid2str(peer->ibp_nid));
2602 if (rej->ibr_version != IBLND_MSG_VERSION &&
2603 rej->ibr_version != IBLND_MSG_VERSION_1) {
2604 CERROR("%s rejected: o2iblnd version %x error\n",
2605 libcfs_nid2str(peer->ibp_nid),
2610 if (rej->ibr_why == IBLND_REJECT_FATAL &&
2611 rej->ibr_version == IBLND_MSG_VERSION_1) {
2612 CDEBUG(D_NET, "rejected by old version peer %s: %x\n",
2613 libcfs_nid2str(peer->ibp_nid), rej->ibr_version);
2615 if (conn->ibc_version != IBLND_MSG_VERSION_1)
2616 rej->ibr_why = IBLND_REJECT_CONN_UNCOMPAT;
2619 switch (rej->ibr_why) {
2620 case IBLND_REJECT_CONN_RACE:
2621 case IBLND_REJECT_CONN_STALE:
2622 case IBLND_REJECT_CONN_UNCOMPAT:
2623 kiblnd_reconnect(conn, rej->ibr_version,
2624 incarnation, rej->ibr_why, cp);
2627 case IBLND_REJECT_MSG_QUEUE_SIZE:
2628 CERROR("%s rejected: incompatible message queue depth %d, %d\n",
2629 libcfs_nid2str(peer->ibp_nid),
2630 cp != NULL ? cp->ibcp_queue_depth :
2631 IBLND_MSG_QUEUE_SIZE(rej->ibr_version),
2632 IBLND_MSG_QUEUE_SIZE(conn->ibc_version));
2635 case IBLND_REJECT_RDMA_FRAGS:
2636 CERROR("%s rejected: incompatible # of RDMA fragments %d, %d\n",
2637 libcfs_nid2str(peer->ibp_nid),
2638 cp != NULL ? cp->ibcp_max_frags :
2639 IBLND_RDMA_FRAGS(rej->ibr_version),
2640 IBLND_RDMA_FRAGS(conn->ibc_version));
2643 case IBLND_REJECT_NO_RESOURCES:
2644 CERROR("%s rejected: o2iblnd no resources\n",
2645 libcfs_nid2str(peer->ibp_nid));
2648 case IBLND_REJECT_FATAL:
2649 CERROR("%s rejected: o2iblnd fatal error\n",
2650 libcfs_nid2str(peer->ibp_nid));
2654 CERROR("%s rejected: o2iblnd reason %d\n",
2655 libcfs_nid2str(peer->ibp_nid),
2663 CNETERR("%s rejected: reason %d, size %d\n",
2664 libcfs_nid2str(peer->ibp_nid), reason, priv_nob);
2668 kiblnd_connreq_done(conn, -ECONNREFUSED);
2672 kiblnd_check_connreply (kib_conn_t *conn, void *priv, int priv_nob)
2674 kib_peer_t *peer = conn->ibc_peer;
2675 lnet_ni_t *ni = peer->ibp_ni;
2676 kib_net_t *net = ni->ni_data;
2677 kib_msg_t *msg = priv;
2678 int ver = conn->ibc_version;
2679 int rc = kiblnd_unpack_msg(msg, priv_nob);
2680 unsigned long flags;
2682 LASSERT (net != NULL);
2685 CERROR("Can't unpack connack from %s: %d\n",
2686 libcfs_nid2str(peer->ibp_nid), rc);
2690 if (msg->ibm_type != IBLND_MSG_CONNACK) {
2691 CERROR("Unexpected message %d from %s\n",
2692 msg->ibm_type, libcfs_nid2str(peer->ibp_nid));
2697 if (ver != msg->ibm_version) {
2698 CERROR("%s replied version %x is different with "
2699 "requested version %x\n",
2700 libcfs_nid2str(peer->ibp_nid), msg->ibm_version, ver);
2705 if (msg->ibm_u.connparams.ibcp_queue_depth !=
2706 IBLND_MSG_QUEUE_SIZE(ver)) {
2707 CERROR("%s has incompatible queue depth %d(%d wanted)\n",
2708 libcfs_nid2str(peer->ibp_nid),
2709 msg->ibm_u.connparams.ibcp_queue_depth,
2710 IBLND_MSG_QUEUE_SIZE(ver));
2715 if (msg->ibm_u.connparams.ibcp_max_frags !=
2716 IBLND_RDMA_FRAGS(ver)) {
2717 CERROR("%s has incompatible max_frags %d (%d wanted)\n",
2718 libcfs_nid2str(peer->ibp_nid),
2719 msg->ibm_u.connparams.ibcp_max_frags,
2720 IBLND_RDMA_FRAGS(ver));
2725 if (msg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2726 CERROR("%s max message size %d too big (%d max)\n",
2727 libcfs_nid2str(peer->ibp_nid),
2728 msg->ibm_u.connparams.ibcp_max_msg_size,
2734 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2735 if (msg->ibm_dstnid == ni->ni_nid &&
2736 msg->ibm_dststamp == net->ibn_incarnation)
2740 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2743 CERROR("Bad connection reply from %s, rc = %d, "
2744 "version: %x max_frags: %d\n",
2745 libcfs_nid2str(peer->ibp_nid), rc,
2746 msg->ibm_version, msg->ibm_u.connparams.ibcp_max_frags);
2750 conn->ibc_incarnation = msg->ibm_srcstamp;
2752 conn->ibc_reserved_credits = IBLND_MSG_QUEUE_SIZE(ver);
2753 LASSERT (conn->ibc_credits + conn->ibc_reserved_credits + IBLND_OOB_MSGS(ver)
2754 <= IBLND_RX_MSGS(ver));
2756 kiblnd_connreq_done(conn, 0);
2760 /* NB My QP has already established itself, so I handle anything going
2761 * wrong here by setting ibc_comms_error.
2762 * kiblnd_connreq_done(0) moves the conn state to ESTABLISHED, but then
2763 * immediately tears it down. */
2766 conn->ibc_comms_error = rc;
2767 kiblnd_connreq_done(conn, 0);
2771 kiblnd_active_connect (struct rdma_cm_id *cmid)
2773 kib_peer_t *peer = (kib_peer_t *)cmid->context;
2776 struct rdma_conn_param cp;
2779 unsigned long flags;
2782 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2784 incarnation = peer->ibp_incarnation;
2785 version = (peer->ibp_version == 0) ? IBLND_MSG_VERSION :
2788 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2790 conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_ACTIVE_CONNECT, version);
2792 kiblnd_peer_connect_failed(peer, 1, -ENOMEM);
2793 kiblnd_peer_decref(peer); /* lose cmid's ref */
2797 /* conn "owns" cmid now, so I return success from here on to ensure the
2798 * CM callback doesn't destroy cmid. conn also takes over cmid's ref
2801 msg = &conn->ibc_connvars->cv_msg;
2803 memset(msg, 0, sizeof(*msg));
2804 kiblnd_init_msg(msg, IBLND_MSG_CONNREQ, sizeof(msg->ibm_u.connparams));
2805 msg->ibm_u.connparams.ibcp_queue_depth = IBLND_MSG_QUEUE_SIZE(version);
2806 msg->ibm_u.connparams.ibcp_max_frags = IBLND_RDMA_FRAGS(version);
2807 msg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2809 kiblnd_pack_msg(peer->ibp_ni, msg, version,
2810 0, peer->ibp_nid, incarnation);
2812 memset(&cp, 0, sizeof(cp));
2813 cp.private_data = msg;
2814 cp.private_data_len = msg->ibm_nob;
2815 cp.responder_resources = 0; /* No atomic ops or RDMA reads */
2816 cp.initiator_depth = 0;
2817 cp.flow_control = 1;
2818 cp.retry_count = *kiblnd_tunables.kib_retry_count;
2819 cp.rnr_retry_count = *kiblnd_tunables.kib_rnr_retry_count;
2821 LASSERT(cmid->context == (void *)conn);
2822 LASSERT(conn->ibc_cmid == cmid);
2824 rc = rdma_connect(cmid, &cp);
2826 CERROR("Can't connect to %s: %d\n",
2827 libcfs_nid2str(peer->ibp_nid), rc);
2828 kiblnd_connreq_done(conn, rc);
2829 kiblnd_conn_decref(conn);
2836 kiblnd_cm_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2842 switch (event->event) {
2844 CERROR("Unexpected event: %d, status: %d\n",
2845 event->event, event->status);
2848 case RDMA_CM_EVENT_CONNECT_REQUEST:
2849 /* destroy cmid on failure */
2850 rc = kiblnd_passive_connect(cmid,
2851 (void *)KIBLND_CONN_PARAM(event),
2852 KIBLND_CONN_PARAM_LEN(event));
2853 CDEBUG(D_NET, "connreq: %d\n", rc);
2856 case RDMA_CM_EVENT_ADDR_ERROR:
2857 peer = (kib_peer_t *)cmid->context;
2858 CNETERR("%s: ADDR ERROR %d\n",
2859 libcfs_nid2str(peer->ibp_nid), event->status);
2860 kiblnd_peer_connect_failed(peer, 1, -EHOSTUNREACH);
2861 kiblnd_peer_decref(peer);
2862 return -EHOSTUNREACH; /* rc != 0 destroys cmid */
2864 case RDMA_CM_EVENT_ADDR_RESOLVED:
2865 peer = (kib_peer_t *)cmid->context;
2867 CDEBUG(D_NET,"%s Addr resolved: %d\n",
2868 libcfs_nid2str(peer->ibp_nid), event->status);
2870 if (event->status != 0) {
2871 CNETERR("Can't resolve address for %s: %d\n",
2872 libcfs_nid2str(peer->ibp_nid), event->status);
2875 rc = rdma_resolve_route(
2876 cmid, *kiblnd_tunables.kib_timeout * 1000);
2879 /* Can't initiate route resolution */
2880 CERROR("Can't resolve route for %s: %d\n",
2881 libcfs_nid2str(peer->ibp_nid), rc);
2883 kiblnd_peer_connect_failed(peer, 1, rc);
2884 kiblnd_peer_decref(peer);
2885 return rc; /* rc != 0 destroys cmid */
2887 case RDMA_CM_EVENT_ROUTE_ERROR:
2888 peer = (kib_peer_t *)cmid->context;
2889 CNETERR("%s: ROUTE ERROR %d\n",
2890 libcfs_nid2str(peer->ibp_nid), event->status);
2891 kiblnd_peer_connect_failed(peer, 1, -EHOSTUNREACH);
2892 kiblnd_peer_decref(peer);
2893 return -EHOSTUNREACH; /* rc != 0 destroys cmid */
2895 case RDMA_CM_EVENT_ROUTE_RESOLVED:
2896 peer = (kib_peer_t *)cmid->context;
2897 CDEBUG(D_NET,"%s Route resolved: %d\n",
2898 libcfs_nid2str(peer->ibp_nid), event->status);
2900 if (event->status == 0)
2901 return kiblnd_active_connect(cmid);
2903 CNETERR("Can't resolve route for %s: %d\n",
2904 libcfs_nid2str(peer->ibp_nid), event->status);
2905 kiblnd_peer_connect_failed(peer, 1, event->status);
2906 kiblnd_peer_decref(peer);
2907 return event->status; /* rc != 0 destroys cmid */
2909 case RDMA_CM_EVENT_UNREACHABLE:
2910 conn = (kib_conn_t *)cmid->context;
2911 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
2912 conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
2913 CNETERR("%s: UNREACHABLE %d\n",
2914 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
2915 kiblnd_connreq_done(conn, -ENETDOWN);
2916 kiblnd_conn_decref(conn);
2919 case RDMA_CM_EVENT_CONNECT_ERROR:
2920 conn = (kib_conn_t *)cmid->context;
2921 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
2922 conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
2923 CNETERR("%s: CONNECT ERROR %d\n",
2924 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
2925 kiblnd_connreq_done(conn, -ENOTCONN);
2926 kiblnd_conn_decref(conn);
2929 case RDMA_CM_EVENT_REJECTED:
2930 conn = (kib_conn_t *)cmid->context;
2931 switch (conn->ibc_state) {
2935 case IBLND_CONN_PASSIVE_WAIT:
2936 CERROR ("%s: REJECTED %d\n",
2937 libcfs_nid2str(conn->ibc_peer->ibp_nid),
2939 kiblnd_connreq_done(conn, -ECONNRESET);
2942 case IBLND_CONN_ACTIVE_CONNECT:
2943 kiblnd_rejected(conn, event->status,
2944 (void *)KIBLND_CONN_PARAM(event),
2945 KIBLND_CONN_PARAM_LEN(event));
2948 kiblnd_conn_decref(conn);
2951 case RDMA_CM_EVENT_ESTABLISHED:
2952 conn = (kib_conn_t *)cmid->context;
2953 switch (conn->ibc_state) {
2957 case IBLND_CONN_PASSIVE_WAIT:
2958 CDEBUG(D_NET, "ESTABLISHED (passive): %s\n",
2959 libcfs_nid2str(conn->ibc_peer->ibp_nid));
2960 kiblnd_connreq_done(conn, 0);
2963 case IBLND_CONN_ACTIVE_CONNECT:
2964 CDEBUG(D_NET, "ESTABLISHED(active): %s\n",
2965 libcfs_nid2str(conn->ibc_peer->ibp_nid));
2966 kiblnd_check_connreply(conn,
2967 (void *)KIBLND_CONN_PARAM(event),
2968 KIBLND_CONN_PARAM_LEN(event));
2971 /* net keeps its ref on conn! */
2974 #ifdef HAVE_OFED_RDMA_CMEV_TIMEWAIT_EXIT
2975 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
2976 CDEBUG(D_NET, "Ignore TIMEWAIT_EXIT event\n");
2979 case RDMA_CM_EVENT_DISCONNECTED:
2980 conn = (kib_conn_t *)cmid->context;
2981 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
2982 CERROR("%s DISCONNECTED\n",
2983 libcfs_nid2str(conn->ibc_peer->ibp_nid));
2984 kiblnd_connreq_done(conn, -ECONNRESET);
2986 kiblnd_close_conn(conn, 0);
2988 kiblnd_conn_decref(conn);
2989 cmid->context = NULL;
2992 case RDMA_CM_EVENT_DEVICE_REMOVAL:
2993 LCONSOLE_ERROR_MSG(0x131,
2994 "Received notification of device removal\n"
2995 "Please shutdown LNET to allow this to proceed\n");
2996 /* Can't remove network from underneath LNET for now, so I have
3000 #ifdef HAVE_OFED_RDMA_CMEV_ADDRCHANGE
3001 case RDMA_CM_EVENT_ADDR_CHANGE:
3002 LCONSOLE_INFO("Physical link changed (eg hca/port)\n");
3009 kiblnd_check_txs_locked(kib_conn_t *conn, struct list_head *txs)
3012 struct list_head *ttmp;
3014 list_for_each(ttmp, txs) {
3015 tx = list_entry(ttmp, kib_tx_t, tx_list);
3017 if (txs != &conn->ibc_active_txs) {
3018 LASSERT(tx->tx_queued);
3020 LASSERT(!tx->tx_queued);
3021 LASSERT(tx->tx_waiting || tx->tx_sending != 0);
3024 if (cfs_time_aftereq(jiffies, tx->tx_deadline)) {
3025 CERROR("Timed out tx: %s, %lu seconds\n",
3026 kiblnd_queue2str(conn, txs),
3027 cfs_duration_sec(jiffies - tx->tx_deadline));
3036 kiblnd_conn_timed_out_locked(kib_conn_t *conn)
3038 return kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue) ||
3039 kiblnd_check_txs_locked(conn, &conn->ibc_tx_noops) ||
3040 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_rsrvd) ||
3041 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_nocred) ||
3042 kiblnd_check_txs_locked(conn, &conn->ibc_active_txs);
3046 kiblnd_check_conns (int idx)
3048 struct list_head closes = LIST_HEAD_INIT(closes);
3049 struct list_head checksends = LIST_HEAD_INIT(checksends);
3050 struct list_head *peers = &kiblnd_data.kib_peers[idx];
3051 struct list_head *ptmp;
3054 struct list_head *ctmp;
3055 unsigned long flags;
3057 /* NB. We expect to have a look at all the peers and not find any
3058 * RDMAs to time out, so we just use a shared lock while we
3060 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3062 list_for_each(ptmp, peers) {
3063 peer = list_entry(ptmp, kib_peer_t, ibp_list);
3065 list_for_each(ctmp, &peer->ibp_conns) {
3069 conn = list_entry(ctmp, kib_conn_t, ibc_list);
3071 LASSERT(conn->ibc_state == IBLND_CONN_ESTABLISHED);
3073 spin_lock(&conn->ibc_lock);
3075 sendnoop = kiblnd_need_noop(conn);
3076 timedout = kiblnd_conn_timed_out_locked(conn);
3077 if (!sendnoop && !timedout) {
3078 spin_unlock(&conn->ibc_lock);
3083 CERROR("Timed out RDMA with %s (%lu): "
3084 "c: %u, oc: %u, rc: %u\n",
3085 libcfs_nid2str(peer->ibp_nid),
3086 cfs_duration_sec(cfs_time_current() -
3087 peer->ibp_last_alive),
3089 conn->ibc_outstanding_credits,
3090 conn->ibc_reserved_credits);
3091 list_add(&conn->ibc_connd_list, &closes);
3093 list_add(&conn->ibc_connd_list, &checksends);
3095 /* +ref for 'closes' or 'checksends' */
3096 kiblnd_conn_addref(conn);
3098 spin_unlock(&conn->ibc_lock);
3102 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3104 /* Handle timeout by closing the whole
3105 * connection. We can only be sure RDMA activity
3106 * has ceased once the QP has been modified. */
3107 while (!list_empty(&closes)) {
3108 conn = list_entry(closes.next,
3109 kib_conn_t, ibc_connd_list);
3110 list_del(&conn->ibc_connd_list);
3111 kiblnd_close_conn(conn, -ETIMEDOUT);
3112 kiblnd_conn_decref(conn);
3115 /* In case we have enough credits to return via a
3116 * NOOP, but there were no non-blocking tx descs
3117 * free to do it last time... */
3118 while (!list_empty(&checksends)) {
3119 conn = list_entry(checksends.next,
3120 kib_conn_t, ibc_connd_list);
3121 list_del(&conn->ibc_connd_list);
3122 kiblnd_check_sends(conn);
3123 kiblnd_conn_decref(conn);
3128 kiblnd_disconnect_conn (kib_conn_t *conn)
3130 LASSERT (!in_interrupt());
3131 LASSERT (current == kiblnd_data.kib_connd);
3132 LASSERT (conn->ibc_state == IBLND_CONN_CLOSING);
3134 rdma_disconnect(conn->ibc_cmid);
3135 kiblnd_finalise_conn(conn);
3137 kiblnd_peer_notify(conn->ibc_peer);
3141 kiblnd_connd (void *arg)
3144 unsigned long flags;
3150 unsigned long deadline = jiffies;
3152 cfs_block_allsigs ();
3154 init_waitqueue_entry_current (&wait);
3155 kiblnd_data.kib_connd = current;
3157 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3159 while (!kiblnd_data.kib_shutdown) {
3163 if (!list_empty(&kiblnd_data.kib_connd_zombies)) {
3164 conn = list_entry(kiblnd_data. \
3165 kib_connd_zombies.next,
3166 kib_conn_t, ibc_list);
3167 list_del(&conn->ibc_list);
3169 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock,
3173 kiblnd_destroy_conn(conn);
3175 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3178 if (!list_empty(&kiblnd_data.kib_connd_conns)) {
3179 conn = list_entry(kiblnd_data.kib_connd_conns.next,
3180 kib_conn_t, ibc_list);
3181 list_del(&conn->ibc_list);
3183 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock,
3187 kiblnd_disconnect_conn(conn);
3188 kiblnd_conn_decref(conn);
3190 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3193 /* careful with the jiffy wrap... */
3194 timeout = (int)(deadline - jiffies);
3198 int chunk = kiblnd_data.kib_peer_hash_size;
3200 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
3203 /* Time to check for RDMA timeouts on a few more
3204 * peers: I do checks every 'p' seconds on a
3205 * proportion of the peer table and I need to check
3206 * every connection 'n' times within a timeout
3207 * interval, to ensure I detect a timeout on any
3208 * connection within (n+1)/n times the timeout
3211 if (*kiblnd_tunables.kib_timeout > n * p)
3212 chunk = (chunk * n * p) /
3213 *kiblnd_tunables.kib_timeout;
3217 for (i = 0; i < chunk; i++) {
3218 kiblnd_check_conns(peer_index);
3219 peer_index = (peer_index + 1) %
3220 kiblnd_data.kib_peer_hash_size;
3224 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3230 /* Nothing to do for 'timeout' */
3231 set_current_state(TASK_INTERRUPTIBLE);
3232 add_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3233 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
3235 waitq_timedwait(&wait, TASK_INTERRUPTIBLE, timeout);
3237 set_current_state(TASK_RUNNING);
3238 remove_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3239 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3242 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
3244 kiblnd_thread_fini();
3249 kiblnd_qp_event(struct ib_event *event, void *arg)
3251 kib_conn_t *conn = arg;
3253 switch (event->event) {
3254 case IB_EVENT_COMM_EST:
3255 CDEBUG(D_NET, "%s established\n",
3256 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3260 CERROR("%s: Async QP event type %d\n",
3261 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3267 kiblnd_complete (struct ib_wc *wc)
3269 switch (kiblnd_wreqid2type(wc->wr_id)) {
3273 case IBLND_WID_RDMA:
3274 /* We only get RDMA completion notification if it fails. All
3275 * subsequent work items, including the final SEND will fail
3276 * too. However we can't print out any more info about the
3277 * failing RDMA because 'tx' might be back on the idle list or
3278 * even reused already if we didn't manage to post all our work
3280 CNETERR("RDMA (tx: %p) failed: %d\n",
3281 kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3285 kiblnd_tx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3289 kiblnd_rx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status,
3296 kiblnd_cq_completion(struct ib_cq *cq, void *arg)
3298 /* NB I'm not allowed to schedule this conn once its refcount has
3299 * reached 0. Since fundamentally I'm racing with scheduler threads
3300 * consuming my CQ I could be called after all completions have
3301 * occurred. But in this case, ibc_nrx == 0 && ibc_nsends_posted == 0
3302 * and this CQ is about to be destroyed so I NOOP. */
3303 kib_conn_t *conn = (kib_conn_t *)arg;
3304 struct kib_sched_info *sched = conn->ibc_sched;
3305 unsigned long flags;
3307 LASSERT(cq == conn->ibc_cq);
3309 spin_lock_irqsave(&sched->ibs_lock, flags);
3311 conn->ibc_ready = 1;
3313 if (!conn->ibc_scheduled &&
3314 (conn->ibc_nrx > 0 ||
3315 conn->ibc_nsends_posted > 0)) {
3316 kiblnd_conn_addref(conn); /* +1 ref for sched_conns */
3317 conn->ibc_scheduled = 1;
3318 list_add_tail(&conn->ibc_sched_list, &sched->ibs_conns);
3320 if (waitqueue_active(&sched->ibs_waitq))
3321 wake_up(&sched->ibs_waitq);
3324 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3328 kiblnd_cq_event(struct ib_event *event, void *arg)
3330 kib_conn_t *conn = arg;
3332 CERROR("%s: async CQ event type %d\n",
3333 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3337 kiblnd_scheduler(void *arg)
3339 long id = (long)arg;
3340 struct kib_sched_info *sched;
3343 unsigned long flags;
3349 cfs_block_allsigs();
3351 init_waitqueue_entry_current(&wait);
3353 sched = kiblnd_data.kib_scheds[KIB_THREAD_CPT(id)];
3355 rc = cfs_cpt_bind(lnet_cpt_table(), sched->ibs_cpt);
3357 CWARN("Failed to bind on CPT %d, please verify whether "
3358 "all CPUs are healthy and reload modules if necessary, "
3359 "otherwise your system might under risk of low "
3360 "performance\n", sched->ibs_cpt);
3363 spin_lock_irqsave(&sched->ibs_lock, flags);
3365 while (!kiblnd_data.kib_shutdown) {
3366 if (busy_loops++ >= IBLND_RESCHED) {
3367 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3372 spin_lock_irqsave(&sched->ibs_lock, flags);
3377 if (!list_empty(&sched->ibs_conns)) {
3378 conn = list_entry(sched->ibs_conns.next,
3379 kib_conn_t, ibc_sched_list);
3380 /* take over kib_sched_conns' ref on conn... */
3381 LASSERT(conn->ibc_scheduled);
3382 list_del(&conn->ibc_sched_list);
3383 conn->ibc_ready = 0;
3385 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3387 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3389 rc = ib_req_notify_cq(conn->ibc_cq,
3392 CWARN("%s: ib_req_notify_cq failed: %d, "
3393 "closing connection\n",
3394 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
3395 kiblnd_close_conn(conn, -EIO);
3396 kiblnd_conn_decref(conn);
3397 spin_lock_irqsave(&sched->ibs_lock,
3402 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3406 CWARN("%s: ib_poll_cq failed: %d, "
3407 "closing connection\n",
3408 libcfs_nid2str(conn->ibc_peer->ibp_nid),
3410 kiblnd_close_conn(conn, -EIO);
3411 kiblnd_conn_decref(conn);
3412 spin_lock_irqsave(&sched->ibs_lock, flags);
3416 spin_lock_irqsave(&sched->ibs_lock, flags);
3418 if (rc != 0 || conn->ibc_ready) {
3419 /* There may be another completion waiting; get
3420 * another scheduler to check while I handle
3422 /* +1 ref for sched_conns */
3423 kiblnd_conn_addref(conn);
3424 list_add_tail(&conn->ibc_sched_list,
3426 if (waitqueue_active(&sched->ibs_waitq))
3427 wake_up(&sched->ibs_waitq);
3429 conn->ibc_scheduled = 0;
3433 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3434 kiblnd_complete(&wc);
3436 spin_lock_irqsave(&sched->ibs_lock, flags);
3439 kiblnd_conn_decref(conn); /* ...drop my ref from above */
3446 set_current_state(TASK_INTERRUPTIBLE);
3447 add_wait_queue_exclusive(&sched->ibs_waitq, &wait);
3448 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3450 waitq_wait(&wait, TASK_INTERRUPTIBLE);
3453 remove_wait_queue(&sched->ibs_waitq, &wait);
3454 set_current_state(TASK_RUNNING);
3455 spin_lock_irqsave(&sched->ibs_lock, flags);
3458 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3460 kiblnd_thread_fini();
3465 kiblnd_failover_thread(void *arg)
3467 rwlock_t *glock = &kiblnd_data.kib_global_lock;
3470 unsigned long flags;
3473 LASSERT (*kiblnd_tunables.kib_dev_failover != 0);
3475 cfs_block_allsigs ();
3477 init_waitqueue_entry_current(&wait);
3478 write_lock_irqsave(glock, flags);
3480 while (!kiblnd_data.kib_shutdown) {
3481 int do_failover = 0;
3484 list_for_each_entry(dev, &kiblnd_data.kib_failed_devs,
3486 if (cfs_time_before(cfs_time_current(),
3487 dev->ibd_next_failover))
3494 list_del_init(&dev->ibd_fail_list);
3495 dev->ibd_failover = 1;
3496 write_unlock_irqrestore(glock, flags);
3498 rc = kiblnd_dev_failover(dev);
3500 write_lock_irqsave(glock, flags);
3502 LASSERT (dev->ibd_failover);
3503 dev->ibd_failover = 0;
3504 if (rc >= 0) { /* Device is OK or failover succeed */
3505 dev->ibd_next_failover = cfs_time_shift(3);
3509 /* failed to failover, retry later */
3510 dev->ibd_next_failover =
3511 cfs_time_shift(min(dev->ibd_failed_failover, 10));
3512 if (kiblnd_dev_can_failover(dev)) {
3513 list_add_tail(&dev->ibd_fail_list,
3514 &kiblnd_data.kib_failed_devs);
3520 /* long sleep if no more pending failover */
3521 long_sleep = list_empty(&kiblnd_data.kib_failed_devs);
3523 set_current_state(TASK_INTERRUPTIBLE);
3524 add_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3525 write_unlock_irqrestore(glock, flags);
3527 rc = schedule_timeout(long_sleep ? cfs_time_seconds(10) :
3528 cfs_time_seconds(1));
3529 set_current_state(TASK_RUNNING);
3530 remove_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3531 write_lock_irqsave(glock, flags);
3533 if (!long_sleep || rc != 0)
3536 /* have a long sleep, routine check all active devices,
3537 * we need checking like this because if there is not active
3538 * connection on the dev and no SEND from local, we may listen
3539 * on wrong HCA for ever while there is a bonding failover */
3540 list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3541 if (kiblnd_dev_can_failover(dev)) {
3542 list_add_tail(&dev->ibd_fail_list,
3543 &kiblnd_data.kib_failed_devs);
3548 write_unlock_irqrestore(glock, flags);
3550 kiblnd_thread_fini();