1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (C) 2006 Cluster File Systems, Inc.
5 * Author: Eric Barton <eric@bartonsoftware.com>
7 * This file is part of Lustre, http://www.lustre.org.
9 * Lustre is free software; you can redistribute it and/or
10 * modify it under the terms of version 2 of the GNU General Public
11 * License as published by the Free Software Foundation.
13 * Lustre is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 * GNU General Public License for more details.
18 * You should have received a copy of the GNU General Public License
19 * along with Lustre; if not, write to the Free Software
20 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
27 kiblnd_msgtype2str(int type)
30 case IBLND_MSG_CONNREQ:
33 case IBLND_MSG_CONNACK:
39 case IBLND_MSG_IMMEDIATE:
42 case IBLND_MSG_PUT_REQ:
45 case IBLND_MSG_PUT_NAK:
48 case IBLND_MSG_PUT_ACK:
51 case IBLND_MSG_PUT_DONE:
54 case IBLND_MSG_GET_REQ:
57 case IBLND_MSG_GET_DONE:
66 kiblnd_tx_done (lnet_ni_t *ni, kib_tx_t *tx)
68 lnet_msg_t *lntmsg[2];
69 kib_net_t *net = ni->ni_data;
73 LASSERT (net != NULL);
74 LASSERT (!in_interrupt());
75 LASSERT (!tx->tx_queued); /* mustn't be queued for sending */
76 LASSERT (tx->tx_sending == 0); /* mustn't be awaiting sent callback */
77 LASSERT (!tx->tx_waiting); /* mustn't be awaiting peer response */
79 #if IBLND_MAP_ON_DEMAND
80 if (tx->tx_fmr != NULL) {
81 rc = ib_fmr_pool_unmap(tx->tx_fmr);
84 if (tx->tx_status != 0) {
85 rc = ib_flush_fmr_pool(net->ibn_fmrpool);
92 if (tx->tx_nfrags != 0) {
93 kiblnd_dma_unmap_sg(net->ibn_dev->ibd_cmid->device,
94 tx->tx_frags, tx->tx_nfrags, tx->tx_dmadir);
98 /* tx may have up to 2 lnet msgs to finalise */
99 lntmsg[0] = tx->tx_lntmsg[0]; tx->tx_lntmsg[0] = NULL;
100 lntmsg[1] = tx->tx_lntmsg[1]; tx->tx_lntmsg[1] = NULL;
103 if (tx->tx_conn != NULL) {
104 LASSERT (ni == tx->tx_conn->ibc_peer->ibp_ni);
106 kiblnd_conn_decref(tx->tx_conn);
113 spin_lock(&net->ibn_tx_lock);
115 list_add(&tx->tx_list, &net->ibn_idle_txs);
117 spin_unlock(&net->ibn_tx_lock);
119 /* delay finalize until my descs have been freed */
120 for (i = 0; i < 2; i++) {
121 if (lntmsg[i] == NULL)
124 lnet_finalize(ni, lntmsg[i], rc);
129 kiblnd_txlist_done (lnet_ni_t *ni, struct list_head *txlist, int status)
133 while (!list_empty (txlist)) {
134 tx = list_entry (txlist->next, kib_tx_t, tx_list);
136 list_del (&tx->tx_list);
139 tx->tx_status = status;
140 kiblnd_tx_done(ni, tx);
145 kiblnd_get_idle_tx (lnet_ni_t *ni)
147 kib_net_t *net = ni->ni_data;
150 LASSERT (net != NULL);
152 spin_lock(&net->ibn_tx_lock);
154 if (list_empty(&net->ibn_idle_txs)) {
155 spin_unlock(&net->ibn_tx_lock);
159 tx = list_entry(net->ibn_idle_txs.next, kib_tx_t, tx_list);
160 list_del(&tx->tx_list);
162 /* Allocate a new completion cookie. It might not be needed,
163 * but we've got a lock right now and we're unlikely to
165 tx->tx_cookie = kiblnd_data.kib_next_tx_cookie++;
167 spin_unlock(&net->ibn_tx_lock);
169 LASSERT (tx->tx_nwrq == 0);
170 LASSERT (!tx->tx_queued);
171 LASSERT (tx->tx_sending == 0);
172 LASSERT (!tx->tx_waiting);
173 LASSERT (tx->tx_status == 0);
174 LASSERT (tx->tx_conn == NULL);
175 LASSERT (tx->tx_lntmsg[0] == NULL);
176 LASSERT (tx->tx_lntmsg[1] == NULL);
177 #if IBLND_MAP_ON_DEMAND
178 LASSERT (tx->tx_fmr == NULL);
180 LASSERT (tx->tx_nfrags == 0);
187 kiblnd_drop_rx (kib_rx_t *rx)
189 kib_conn_t *conn = rx->rx_conn;
192 spin_lock_irqsave(&kiblnd_data.kib_sched_lock, flags);
193 LASSERT (conn->ibc_nrx > 0);
195 spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock, flags);
197 kiblnd_conn_decref(conn);
201 kiblnd_post_rx (kib_rx_t *rx, int credit)
203 kib_conn_t *conn = rx->rx_conn;
204 kib_net_t *net = conn->ibc_peer->ibp_ni->ni_data;
205 struct ib_recv_wr *bad_wrq;
208 LASSERT (net != NULL);
209 LASSERT (!in_interrupt());
210 LASSERT (credit == IBLND_POSTRX_NO_CREDIT ||
211 credit == IBLND_POSTRX_PEER_CREDIT ||
212 credit == IBLND_POSTRX_RSRVD_CREDIT);
214 rx->rx_sge.length = IBLND_MSG_SIZE;
215 rx->rx_sge.lkey = net->ibn_dev->ibd_mr->lkey;
216 rx->rx_sge.addr = rx->rx_msgaddr;
218 rx->rx_wrq.next = NULL;
219 rx->rx_wrq.sg_list = &rx->rx_sge;
220 rx->rx_wrq.num_sge = 1;
221 rx->rx_wrq.wr_id = kiblnd_ptr2wreqid(rx, IBLND_WID_RX);
223 LASSERT (conn->ibc_state >= IBLND_CONN_INIT);
224 LASSERT (rx->rx_nob >= 0); /* not posted */
226 if (conn->ibc_state > IBLND_CONN_ESTABLISHED) {
227 kiblnd_drop_rx(rx); /* No more posts for this rx */
231 rx->rx_nob = -1; /* flag posted */
233 rc = ib_post_recv(conn->ibc_cmid->qp, &rx->rx_wrq, &bad_wrq);
235 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) /* Initial post */
239 CERROR("Can't post rx for %s: %d\n",
240 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
241 kiblnd_close_conn(conn, rc);
242 kiblnd_drop_rx(rx); /* No more posts for this rx */
246 if (credit == IBLND_POSTRX_NO_CREDIT)
249 spin_lock(&conn->ibc_lock);
250 if (credit == IBLND_POSTRX_PEER_CREDIT)
251 conn->ibc_outstanding_credits++;
253 conn->ibc_reserved_credits++;
254 spin_unlock(&conn->ibc_lock);
256 kiblnd_check_sends(conn);
261 kiblnd_find_waiting_tx_locked(kib_conn_t *conn, int txtype, __u64 cookie)
263 struct list_head *tmp;
265 list_for_each(tmp, &conn->ibc_active_txs) {
266 kib_tx_t *tx = list_entry(tmp, kib_tx_t, tx_list);
268 LASSERT (!tx->tx_queued);
269 LASSERT (tx->tx_sending != 0 || tx->tx_waiting);
271 if (tx->tx_cookie != cookie)
274 if (tx->tx_waiting &&
275 tx->tx_msg->ibm_type == txtype)
278 CWARN("Bad completion: %swaiting, type %x (wanted %x)\n",
279 tx->tx_waiting ? "" : "NOT ",
280 tx->tx_msg->ibm_type, txtype);
286 kiblnd_handle_completion(kib_conn_t *conn, int txtype, int status, __u64 cookie)
289 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
292 spin_lock(&conn->ibc_lock);
294 tx = kiblnd_find_waiting_tx_locked(conn, txtype, cookie);
296 spin_unlock(&conn->ibc_lock);
298 CWARN("Unmatched completion type %x cookie "LPX64" from %s\n",
299 txtype, cookie, libcfs_nid2str(conn->ibc_peer->ibp_nid));
300 kiblnd_close_conn(conn, -EPROTO);
304 if (tx->tx_status == 0) { /* success so far */
305 if (status < 0) { /* failed? */
306 tx->tx_status = status;
307 } else if (txtype == IBLND_MSG_GET_REQ) {
308 lnet_set_reply_msg_len(ni, tx->tx_lntmsg[1], status);
314 idle = !tx->tx_queued && (tx->tx_sending == 0);
316 list_del(&tx->tx_list);
318 spin_unlock(&conn->ibc_lock);
321 kiblnd_tx_done(ni, tx);
325 kiblnd_send_completion (kib_conn_t *conn, int type, int status, __u64 cookie)
327 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
328 kib_tx_t *tx = kiblnd_get_idle_tx(ni);
331 CERROR("Can't get tx for completion %x for %s\n",
332 type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
336 tx->tx_msg->ibm_u.completion.ibcm_status = status;
337 tx->tx_msg->ibm_u.completion.ibcm_cookie = cookie;
338 kiblnd_init_tx_msg(ni, tx, type, sizeof(kib_completion_msg_t));
340 kiblnd_queue_tx(tx, conn);
344 kiblnd_handle_rx (kib_rx_t *rx)
346 kib_msg_t *msg = rx->rx_msg;
347 kib_conn_t *conn = rx->rx_conn;
348 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
349 int credits = msg->ibm_credits;
355 LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
357 CDEBUG (D_NET, "Received %x[%d] from %s\n",
358 msg->ibm_type, credits, libcfs_nid2str(conn->ibc_peer->ibp_nid));
361 /* Have I received credits that will let me send? */
362 spin_lock(&conn->ibc_lock);
364 if (conn->ibc_credits + credits > IBLND_MSG_QUEUE_SIZE) {
365 rc2 = conn->ibc_credits;
366 spin_unlock(&conn->ibc_lock);
368 CERROR("Bad credits from %s: %d + %d > %d\n",
369 libcfs_nid2str(conn->ibc_peer->ibp_nid),
370 rc2, credits, IBLND_MSG_QUEUE_SIZE);
372 kiblnd_close_conn(conn, -EPROTO);
373 kiblnd_post_rx(rx, IBLND_POSTRX_NO_CREDIT);
377 conn->ibc_credits += credits;
379 /* This ensures the credit taken by NOOP can be returned */
380 if (msg->ibm_type == IBLND_MSG_NOOP)
381 conn->ibc_outstanding_credits++;
383 spin_unlock(&conn->ibc_lock);
384 kiblnd_check_sends(conn);
387 switch (msg->ibm_type) {
389 CERROR("Bad IBLND message type %x from %s\n",
390 msg->ibm_type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
391 post_credit = IBLND_POSTRX_NO_CREDIT;
396 if (credits != 0) /* credit already posted */
397 post_credit = IBLND_POSTRX_NO_CREDIT;
398 else /* a keepalive NOOP */
399 post_credit = IBLND_POSTRX_PEER_CREDIT;
402 case IBLND_MSG_IMMEDIATE:
403 post_credit = IBLND_POSTRX_DONT_POST;
404 rc = lnet_parse(ni, &msg->ibm_u.immediate.ibim_hdr,
405 msg->ibm_srcnid, rx, 0);
406 if (rc < 0) /* repost on error */
407 post_credit = IBLND_POSTRX_PEER_CREDIT;
410 case IBLND_MSG_PUT_REQ:
411 post_credit = IBLND_POSTRX_DONT_POST;
412 rc = lnet_parse(ni, &msg->ibm_u.putreq.ibprm_hdr,
413 msg->ibm_srcnid, rx, 1);
414 if (rc < 0) /* repost on error */
415 post_credit = IBLND_POSTRX_PEER_CREDIT;
418 case IBLND_MSG_PUT_NAK:
419 CWARN ("PUT_NACK from %s\n", libcfs_nid2str(conn->ibc_peer->ibp_nid));
420 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
421 kiblnd_handle_completion(conn, IBLND_MSG_PUT_REQ,
422 msg->ibm_u.completion.ibcm_status,
423 msg->ibm_u.completion.ibcm_cookie);
426 case IBLND_MSG_PUT_ACK:
427 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
429 spin_lock(&conn->ibc_lock);
430 tx = kiblnd_find_waiting_tx_locked(conn, IBLND_MSG_PUT_REQ,
431 msg->ibm_u.putack.ibpam_src_cookie);
433 list_del(&tx->tx_list);
434 spin_unlock(&conn->ibc_lock);
437 CERROR("Unmatched PUT_ACK from %s\n",
438 libcfs_nid2str(conn->ibc_peer->ibp_nid));
443 LASSERT (tx->tx_waiting);
444 /* CAVEAT EMPTOR: I could be racing with tx_complete, but...
445 * (a) I can overwrite tx_msg since my peer has received it!
446 * (b) tx_waiting set tells tx_complete() it's not done. */
448 tx->tx_nwrq = 0; /* overwrite PUT_REQ */
450 rc2 = kiblnd_init_rdma(ni, tx, IBLND_MSG_PUT_DONE,
451 kiblnd_rd_size(&msg->ibm_u.putack.ibpam_rd),
452 &msg->ibm_u.putack.ibpam_rd,
453 msg->ibm_u.putack.ibpam_dst_cookie);
455 CERROR("Can't setup rdma for PUT to %s: %d\n",
456 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc2);
458 spin_lock(&conn->ibc_lock);
459 tx->tx_waiting = 0; /* clear waiting and queue atomically */
460 kiblnd_queue_tx_locked(tx, conn);
461 spin_unlock(&conn->ibc_lock);
464 case IBLND_MSG_PUT_DONE:
465 post_credit = IBLND_POSTRX_PEER_CREDIT;
466 kiblnd_handle_completion(conn, IBLND_MSG_PUT_ACK,
467 msg->ibm_u.completion.ibcm_status,
468 msg->ibm_u.completion.ibcm_cookie);
471 case IBLND_MSG_GET_REQ:
472 post_credit = IBLND_POSTRX_DONT_POST;
473 rc = lnet_parse(ni, &msg->ibm_u.get.ibgm_hdr,
474 msg->ibm_srcnid, rx, 1);
475 if (rc < 0) /* repost on error */
476 post_credit = IBLND_POSTRX_PEER_CREDIT;
479 case IBLND_MSG_GET_DONE:
480 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
481 kiblnd_handle_completion(conn, IBLND_MSG_GET_REQ,
482 msg->ibm_u.completion.ibcm_status,
483 msg->ibm_u.completion.ibcm_cookie);
487 if (rc < 0) /* protocol error */
488 kiblnd_close_conn(conn, rc);
490 if (post_credit != IBLND_POSTRX_DONT_POST)
491 kiblnd_post_rx(rx, post_credit);
495 kiblnd_rx_complete (kib_rx_t *rx, int status, int nob)
497 kib_msg_t *msg = rx->rx_msg;
498 kib_conn_t *conn = rx->rx_conn;
499 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
500 kib_net_t *net = ni->ni_data;
505 LASSERT (net != NULL);
506 LASSERT (rx->rx_nob < 0); /* was posted */
507 rx->rx_nob = 0; /* isn't now */
509 if (conn->ibc_state > IBLND_CONN_ESTABLISHED)
512 if (status != IB_WC_SUCCESS) {
513 CDEBUG(D_NETERROR, "Rx from %s failed: %d\n",
514 libcfs_nid2str(conn->ibc_peer->ibp_nid), status);
521 rc = kiblnd_unpack_msg(msg, rx->rx_nob);
523 CERROR ("Error %d unpacking rx from %s\n",
524 rc, libcfs_nid2str(conn->ibc_peer->ibp_nid));
528 if (msg->ibm_srcnid != conn->ibc_peer->ibp_nid ||
529 msg->ibm_dstnid != ni->ni_nid ||
530 msg->ibm_srcstamp != conn->ibc_incarnation ||
531 msg->ibm_dststamp != net->ibn_incarnation) {
532 CERROR ("Stale rx from %s\n",
533 libcfs_nid2str(conn->ibc_peer->ibp_nid));
538 /* set time last known alive */
539 kiblnd_peer_alive(conn->ibc_peer);
541 /* racing with connection establishment/teardown! */
543 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
544 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
545 /* must check holding global lock to eliminate race */
546 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
547 list_add_tail(&rx->rx_list, &conn->ibc_early_rxs);
548 write_unlock_irqrestore(&kiblnd_data.kib_global_lock,
552 write_unlock_irqrestore(&kiblnd_data.kib_global_lock,
555 kiblnd_handle_rx(rx);
559 CDEBUG(D_NET, "rx %p conn %p\n", rx, conn);
560 kiblnd_close_conn(conn, err);
562 kiblnd_drop_rx(rx); /* Don't re-post rx. */
566 kiblnd_kvaddr_to_page (unsigned long vaddr)
570 if (vaddr >= VMALLOC_START &&
571 vaddr < VMALLOC_END) {
572 page = vmalloc_to_page ((void *)vaddr);
573 LASSERT (page != NULL);
576 #ifdef CONFIG_HIGHMEM
577 if (vaddr >= PKMAP_BASE &&
578 vaddr < (PKMAP_BASE + LAST_PKMAP * PAGE_SIZE)) {
579 /* No highmem pages only used for bulk (kiov) I/O */
580 CERROR("find page for address in highmem\n");
584 page = virt_to_page (vaddr);
585 LASSERT (page != NULL);
589 #if !IBLND_MAP_ON_DEMAND
591 kiblnd_setup_rd_iov(lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd,
592 unsigned int niov, struct iovec *iov, int offset, int nob)
595 struct scatterlist *sg;
601 kib_net_t *net = ni->ni_data;
605 LASSERT (net != NULL);
607 while (offset >= iov->iov_len) {
608 offset -= iov->iov_len;
618 vaddr = ((unsigned long)iov->iov_base) + offset;
619 page_offset = vaddr & (PAGE_SIZE - 1);
620 page = kiblnd_kvaddr_to_page(vaddr);
622 CERROR ("Can't find page\n");
626 fragnob = min((int)(iov->iov_len - offset), nob);
627 fragnob = min(fragnob, (int)PAGE_SIZE - page_offset);
630 sg->offset = page_offset;
631 sg->length = fragnob;
634 if (offset + fragnob < iov->iov_len) {
644 /* If rd is not tx_rd, it's going to get sent to a peer and I'm the
646 tx->tx_nfrags = sg - tx->tx_frags;
647 tx->tx_dmadir = (rd != tx->tx_rd) ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
649 rd->rd_nfrags = kiblnd_dma_map_sg(net->ibn_dev->ibd_cmid->device,
650 tx->tx_frags, tx->tx_nfrags,
652 rd->rd_key = (rd != tx->tx_rd) ?
653 net->ibn_dev->ibd_mr->rkey : net->ibn_dev->ibd_mr->lkey;
655 for (i = 0; i < rd->rd_nfrags; i++) {
656 rd->rd_frags[i].rf_nob = kiblnd_sg_dma_len(
657 net->ibn_dev->ibd_cmid->device, &tx->tx_frags[i]);
658 rd->rd_frags[i].rf_addr = kiblnd_sg_dma_address(
659 net->ibn_dev->ibd_cmid->device, &tx->tx_frags[i]);
666 kiblnd_setup_rd_kiov (lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd,
667 int nkiov, lnet_kiov_t *kiov, int offset, int nob)
669 struct scatterlist *sg;
672 kib_net_t *net = ni->ni_data;
674 CDEBUG(D_NET, "niov %d offset %d nob %d\n", nkiov, offset, nob);
678 LASSERT (net != NULL);
680 while (offset >= kiov->kiov_len) {
681 offset -= kiov->kiov_len;
691 fragnob = min((int)(kiov->kiov_len - offset), nob);
693 memset(sg, 0, sizeof(*sg));
694 sg->page = kiov->kiov_page;
695 sg->offset = kiov->kiov_offset + offset;
696 sg->length = fragnob;
705 /* If rd is not tx_rd, it's going to get sent to a peer and I'm the
707 tx->tx_nfrags = sg - tx->tx_frags;
708 tx->tx_dmadir = (rd != tx->tx_rd) ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
710 rd->rd_nfrags = kiblnd_dma_map_sg(net->ibn_dev->ibd_cmid->device,
711 tx->tx_frags, tx->tx_nfrags, tx->tx_dmadir);
712 rd->rd_key = (rd != tx->tx_rd) ?
713 net->ibn_dev->ibd_mr->rkey : net->ibn_dev->ibd_mr->lkey;
715 for (i = 0; i < tx->tx_nfrags; i++) {
716 rd->rd_frags[i].rf_nob = kiblnd_sg_dma_len(
717 net->ibn_dev->ibd_cmid->device, &tx->tx_frags[i]);
718 rd->rd_frags[i].rf_addr = kiblnd_sg_dma_address(
719 net->ibn_dev->ibd_cmid->device, &tx->tx_frags[i]);
721 CDEBUG(D_WARNING,"frag[%d]: "LPX64" for %d\n",
722 i, rd->rd_frags[i].rf_addr, rd->rd_frags[i].rf_nob);
730 kiblnd_map_tx (lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd,
731 int npages, unsigned long page_offset, int nob)
733 struct ib_pool_fmr *fmr;
734 kib_net_t *net = ni->ni_data;
736 LASSERT (net != NULL);
737 LASSERT (tx->tx_fmr == NULL);
738 LASSERT (page_offset < PAGE_SIZE);
739 LASSERT (npages >= (1 + ((page_offset + nob - 1)>>PAGE_SHIFT)));
740 LASSERT (npages <= LNET_MAX_IOV);
744 fmr = ib_fmr_pool_map_phys(net->ibn_fmrpool, tx->tx_pages,
745 npages, rd->rd_addr);
747 CERROR ("Can't map %d pages: %ld\n", npages, PTR_ERR(fmr));
751 /* If rd is not tx_rd, it's going to get sent to a peer, who will need
754 rd->rd_key = (rd != tx->tx_rd) ? fmr->fmr->rkey : fmr->fmr->lkey;
762 kiblnd_setup_rd_iov (lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd,
763 unsigned int niov, struct iovec *iov, int offset, int nob)
770 unsigned long page_offset;
776 while (offset >= iov->iov_len) {
777 offset -= iov->iov_len;
783 if (nob > iov->iov_len - offset) {
784 CERROR ("Can't map multiple vaddr fragments\n");
788 vaddr = ((unsigned long)iov->iov_base) + offset;
790 page_offset = vaddr & (PAGE_SIZE - 1);
795 LASSERT (npages < LNET_MAX_IOV);
797 page = kiblnd_kvaddr_to_page(vaddr);
799 CERROR("Can't find page for %lu\n", vaddr);
803 tx->tx_pages[npages++] = lnet_page2phys(page);
805 fragnob = PAGE_SIZE - (vaddr & (PAGE_SIZE - 1));
811 return kiblnd_map_tx(ni, tx, rd, npages, page_offset, nob);
815 kiblnd_setup_rd_kiov (lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd,
816 int nkiov, lnet_kiov_t *kiov, int offset, int nob)
820 unsigned long page_offset;
822 CDEBUG(D_NET, "niov %d offset %d nob %d\n", nkiov, offset, nob);
826 LASSERT (nkiov <= LNET_MAX_IOV);
828 while (offset >= kiov->kiov_len) {
829 offset -= kiov->kiov_len;
835 page_offset = kiov->kiov_offset + offset;
837 resid = offset + nob;
841 LASSERT (npages < LNET_MAX_IOV);
844 if ((npages > 0 && kiov->kiov_offset != 0) ||
845 (resid > kiov->kiov_len &&
846 (kiov->kiov_offset + kiov->kiov_len) != PAGE_SIZE)) {
847 /* Can't have gaps */
848 CERROR ("Can't make payload contiguous in I/O VM:"
849 "page %d, offset %d, len %d \n",
850 npages, kiov->kiov_offset, kiov->kiov_len);
855 tx->tx_pages[npages++] = lnet_page2phys(kiov->kiov_page);
856 resid -= kiov->kiov_len;
861 return kiblnd_map_tx(ni, tx, rd, npages, page_offset, nob);
866 kiblnd_check_sends (kib_conn_t *conn)
869 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
871 int consume_cred = 0;
872 struct ib_send_wr *bad_wrq;
875 /* Don't send anything until after the connection is established */
876 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
877 CDEBUG(D_NET, "%s too soon\n",
878 libcfs_nid2str(conn->ibc_peer->ibp_nid));
882 spin_lock(&conn->ibc_lock);
884 LASSERT (conn->ibc_nsends_posted <=
885 *kiblnd_tunables.kib_concurrent_sends);
886 LASSERT (conn->ibc_reserved_credits >= 0);
888 while (conn->ibc_reserved_credits > 0 &&
889 !list_empty(&conn->ibc_tx_queue_rsrvd)) {
890 tx = list_entry(conn->ibc_tx_queue_rsrvd.next,
892 list_del(&tx->tx_list);
893 list_add_tail(&tx->tx_list, &conn->ibc_tx_queue);
894 conn->ibc_reserved_credits--;
897 if (kiblnd_send_noop(conn)) {
898 spin_unlock(&conn->ibc_lock);
900 tx = kiblnd_get_idle_tx(ni);
902 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_NOOP, 0);
904 spin_lock(&conn->ibc_lock);
907 kiblnd_queue_tx_locked(tx, conn);
911 if (!list_empty(&conn->ibc_tx_queue_nocred)) {
912 tx = list_entry(conn->ibc_tx_queue_nocred.next,
915 } else if (!list_empty(&conn->ibc_tx_noops)) {
916 tx = list_entry(conn->ibc_tx_noops.next,
919 } else if (!list_empty(&conn->ibc_tx_queue)) {
920 tx = list_entry(conn->ibc_tx_queue.next,
924 /* nothing to send right now */
928 LASSERT (tx->tx_queued);
929 /* We rely on this for QP sizing */
930 LASSERT (tx->tx_nwrq > 0 &&
931 tx->tx_nwrq <= 1 + IBLND_MAX_RDMA_FRAGS);
933 LASSERT (conn->ibc_outstanding_credits >= 0);
934 LASSERT (conn->ibc_outstanding_credits <= IBLND_MSG_QUEUE_SIZE);
935 LASSERT (conn->ibc_credits >= 0);
936 LASSERT (conn->ibc_credits <= IBLND_MSG_QUEUE_SIZE);
938 if (conn->ibc_nsends_posted ==
939 *kiblnd_tunables.kib_concurrent_sends) {
940 /* tx completions outstanding... */
941 CDEBUG(D_NET, "%s: posted enough\n",
942 libcfs_nid2str(conn->ibc_peer->ibp_nid));
947 if (conn->ibc_credits == 0) { /* no credits */
948 CDEBUG(D_NET, "%s: no credits\n",
949 libcfs_nid2str(conn->ibc_peer->ibp_nid));
950 break; /* NB ibc_tx_queue_nocred checked */
953 /* Last credit reserved for NOOP */
954 if (conn->ibc_credits == 1 &&
955 tx->tx_msg->ibm_type != IBLND_MSG_NOOP) {
956 CDEBUG(D_NET, "%s: not using last credit\n",
957 libcfs_nid2str(conn->ibc_peer->ibp_nid));
958 break; /* NB ibc_tx_noops checked */
962 list_del(&tx->tx_list);
965 /* NB don't drop ibc_lock before bumping tx_sending */
967 if (tx->tx_msg->ibm_type == IBLND_MSG_NOOP &&
968 !kiblnd_send_noop(conn)) {
970 spin_unlock(&conn->ibc_lock);
971 kiblnd_tx_done(ni, tx);
972 spin_lock(&conn->ibc_lock);
973 CDEBUG(D_NET, "%s: redundant noop\n",
974 libcfs_nid2str(conn->ibc_peer->ibp_nid));
978 kiblnd_pack_msg(ni, tx->tx_msg, conn->ibc_outstanding_credits,
979 conn->ibc_peer->ibp_nid, conn->ibc_incarnation);
981 conn->ibc_outstanding_credits = 0;
982 conn->ibc_nsends_posted++;
986 /* CAVEAT EMPTOR! This tx could be the PUT_DONE of an RDMA
987 * PUT. If so, it was first queued here as a PUT_REQ, sent and
988 * stashed on ibc_active_txs, matched by an incoming PUT_ACK,
989 * and then re-queued here. It's (just) possible that
990 * tx_sending is non-zero if we've not done the tx_complete() from
991 * the first send; hence the ++ rather than = below. */
994 list_add (&tx->tx_list, &conn->ibc_active_txs);
999 for (i = 0; i < tx->tx_nwrq - 1; i++) {
1000 LASSERT (tx->tx_wrq[i].opcode == IB_WR_RDMA_WRITE);
1001 LASSERT (tx->tx_wrq[i].next == &tx->tx_wrq[i+1]);
1002 LASSERT (tx->tx_wrq[i].sg_list == &tx->tx_sge[i]);
1004 CDEBUG(D_WARNING, "WORK[%d]: RDMA "LPX64
1005 " for %d k %x -> "LPX64" k %x\n", i,
1006 tx->tx_wrq[i].sg_list->addr,
1007 tx->tx_wrq[i].sg_list->length,
1008 tx->tx_wrq[i].sg_list->lkey,
1009 tx->tx_wrq[i].wr.rdma.remote_addr,
1010 tx->tx_wrq[i].wr.rdma.rkey);
1013 LASSERT (tx->tx_wrq[i].opcode == IB_WR_SEND);
1014 LASSERT (tx->tx_wrq[i].next == NULL);
1015 LASSERT (tx->tx_wrq[i].sg_list == &tx->tx_sge[i]);
1017 CDEBUG(D_WARNING, "WORK[%d]: SEND "LPX64" for %d k %x\n", i,
1018 tx->tx_wrq[i].sg_list->addr,
1019 tx->tx_wrq[i].sg_list->length,
1020 tx->tx_wrq[i].sg_list->lkey);
1023 /* I'm still holding ibc_lock! */
1024 if (conn->ibc_state != IBLND_CONN_ESTABLISHED)
1027 rc = ib_post_send(conn->ibc_cmid->qp, tx->tx_wrq, &bad_wrq);
1029 conn->ibc_last_send = jiffies;
1032 /* NB credits are transferred in the actual
1033 * message, which can only be the last work item */
1034 conn->ibc_outstanding_credits += tx->tx_msg->ibm_credits;
1036 conn->ibc_credits++;
1037 conn->ibc_nsends_posted--;
1043 done = (tx->tx_sending == 0);
1045 list_del (&tx->tx_list);
1047 spin_unlock(&conn->ibc_lock);
1049 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
1050 CERROR("Error %d posting transmit to %s\n",
1051 rc, libcfs_nid2str(conn->ibc_peer->ibp_nid));
1053 CDEBUG(D_NET, "Error %d posting transmit to %s\n",
1054 rc, libcfs_nid2str(conn->ibc_peer->ibp_nid));
1056 kiblnd_close_conn(conn, rc);
1059 kiblnd_tx_done(ni, tx);
1064 spin_unlock(&conn->ibc_lock);
1068 kiblnd_tx_complete (kib_tx_t *tx, int status)
1070 int failed = (status != IB_WC_SUCCESS);
1071 kib_conn_t *conn = tx->tx_conn;
1074 LASSERT (tx->tx_sending > 0);
1077 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
1078 CDEBUG(D_NETERROR, "Tx -> %s cookie "LPX64
1079 "sending %d waiting %d: failed %d\n",
1080 libcfs_nid2str(conn->ibc_peer->ibp_nid),
1081 tx->tx_cookie, tx->tx_sending, tx->tx_waiting,
1084 kiblnd_close_conn(conn, -EIO);
1086 kiblnd_peer_alive(conn->ibc_peer);
1089 spin_lock(&conn->ibc_lock);
1091 /* I could be racing with rdma completion. Whoever makes 'tx' idle
1092 * gets to free it, which also drops its ref on 'conn'. */
1095 conn->ibc_nsends_posted--;
1098 tx->tx_waiting = 0; /* don't wait for peer */
1099 tx->tx_status = -EIO;
1102 idle = (tx->tx_sending == 0) && /* This is the final callback */
1103 !tx->tx_waiting && /* Not waiting for peer */
1104 !tx->tx_queued; /* Not re-queued (PUT_DONE) */
1106 list_del(&tx->tx_list);
1108 kiblnd_conn_addref(conn); /* 1 ref for me.... */
1110 spin_unlock(&conn->ibc_lock);
1113 kiblnd_tx_done(conn->ibc_peer->ibp_ni, tx);
1115 kiblnd_check_sends(conn);
1117 kiblnd_conn_decref(conn); /* ...until here */
1121 kiblnd_init_tx_msg (lnet_ni_t *ni, kib_tx_t *tx, int type, int body_nob)
1123 kib_net_t *net = ni->ni_data;
1124 struct ib_sge *sge = &tx->tx_sge[tx->tx_nwrq];
1125 struct ib_send_wr *wrq = &tx->tx_wrq[tx->tx_nwrq];
1126 int nob = offsetof (kib_msg_t, ibm_u) + body_nob;
1128 LASSERT (net != NULL);
1129 LASSERT (tx->tx_nwrq >= 0);
1130 LASSERT (tx->tx_nwrq < IBLND_MAX_RDMA_FRAGS + 1);
1131 LASSERT (nob <= IBLND_MSG_SIZE);
1133 kiblnd_init_msg(tx->tx_msg, type, body_nob);
1135 sge->addr = tx->tx_msgaddr;
1136 sge->lkey = net->ibn_dev->ibd_mr->lkey;
1139 memset(wrq, 0, sizeof(*wrq));
1142 wrq->wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_TX);
1145 wrq->opcode = IB_WR_SEND;
1146 wrq->send_flags = IB_SEND_SIGNALED;
1152 kiblnd_init_rdma (lnet_ni_t *ni, kib_tx_t *tx, int type,
1153 int nob, kib_rdma_desc_t *dstrd, __u64 dstcookie)
1155 kib_msg_t *ibmsg = tx->tx_msg;
1156 kib_rdma_desc_t *srcrd = tx->tx_rd;
1157 struct ib_sge *sge = &tx->tx_sge[0];
1158 struct ib_send_wr *wrq = &tx->tx_wrq[0];
1161 #if IBLND_MAP_ON_DEMAND
1162 LASSERT (!in_interrupt());
1163 LASSERT (tx->tx_nwrq == 0);
1164 LASSERT (type == IBLND_MSG_GET_DONE ||
1165 type == IBLND_MSG_PUT_DONE);
1167 sge->addr = srcrd->rd_addr;
1168 sge->lkey = srcrd->rd_key;
1171 wrq = &tx->tx_wrq[0];
1173 wrq->next = &tx->tx_wrq[1];
1174 wrq->wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_RDMA);
1177 wrq->opcode = IB_WR_RDMA_WRITE;
1178 wrq->send_flags = 0;
1180 wrq->wr.rdma.remote_addr = dstrd->rd_addr;
1181 wrq->wr.rdma.rkey = dstrd->rd_key;
1185 /* CAVEAT EMPTOR: this 'consumes' the frags in 'dstrd' */
1187 kib_rdma_frag_t *srcfrag;
1189 kib_rdma_frag_t *dstfrag;
1193 LASSERT (!in_interrupt());
1194 LASSERT (tx->tx_nwrq == 0);
1195 LASSERT (type == IBLND_MSG_GET_DONE ||
1196 type == IBLND_MSG_PUT_DONE);
1198 srcidx = dstidx = 0;
1199 srcfrag = &srcrd->rd_frags[0];
1200 dstfrag = &dstrd->rd_frags[0];
1203 if (srcidx >= srcrd->rd_nfrags) {
1204 CERROR("Src buffer exhausted: %d frags\n", srcidx);
1209 if (dstidx == dstrd->rd_nfrags) {
1210 CERROR("Dst buffer exhausted: %d frags\n", dstidx);
1215 if (tx->tx_nwrq == IBLND_MAX_RDMA_FRAGS) {
1216 CERROR("RDMA too fragmented: %d/%d src %d/%d dst frags\n",
1217 srcidx, srcrd->rd_nfrags,
1218 dstidx, dstrd->rd_nfrags);
1223 wrknob = MIN(MIN(srcfrag->rf_nob, dstfrag->rf_nob), resid);
1225 sge = &tx->tx_sge[tx->tx_nwrq];
1226 sge->addr = srcfrag->rf_addr;
1227 sge->length = wrknob;
1228 sge->lkey = srcrd->rd_key;
1230 wrq = &tx->tx_wrq[tx->tx_nwrq];
1232 wrq->next = wrq + 1;
1233 wrq->wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_RDMA);
1236 wrq->opcode = IB_WR_RDMA_WRITE;
1237 wrq->send_flags = 0;
1239 wrq->wr.rdma.remote_addr = dstfrag->rf_addr;
1240 wrq->wr.rdma.rkey = dstrd->rd_key;
1246 if (wrknob < srcfrag->rf_nob) {
1247 srcfrag->rf_nob -= wrknob;
1248 srcfrag->rf_addr += wrknob;
1254 if (wrknob < dstfrag->rf_nob) {
1255 dstfrag->rf_nob -= wrknob;
1256 dstfrag->rf_addr += wrknob;
1265 if (rc < 0) /* no RDMA if completing with failure */
1268 ibmsg->ibm_u.completion.ibcm_status = rc;
1269 ibmsg->ibm_u.completion.ibcm_cookie = dstcookie;
1270 kiblnd_init_tx_msg(ni, tx, type, sizeof (kib_completion_msg_t));
1276 kiblnd_queue_tx_locked (kib_tx_t *tx, kib_conn_t *conn)
1278 struct list_head *q;
1280 LASSERT (tx->tx_nwrq > 0); /* work items set up */
1281 LASSERT (!tx->tx_queued); /* not queued for sending already */
1284 tx->tx_deadline = jiffies + (*kiblnd_tunables.kib_timeout * HZ);
1286 if (tx->tx_conn == NULL) {
1287 kiblnd_conn_addref(conn);
1289 LASSERT (tx->tx_msg->ibm_type != IBLND_MSG_PUT_DONE);
1291 /* PUT_DONE first attached to conn as a PUT_REQ */
1292 LASSERT (tx->tx_conn == conn);
1293 LASSERT (tx->tx_msg->ibm_type == IBLND_MSG_PUT_DONE);
1296 switch (tx->tx_msg->ibm_type) {
1300 case IBLND_MSG_PUT_REQ:
1301 case IBLND_MSG_GET_REQ:
1302 q = &conn->ibc_tx_queue_rsrvd;
1305 case IBLND_MSG_PUT_NAK:
1306 case IBLND_MSG_PUT_ACK:
1307 case IBLND_MSG_PUT_DONE:
1308 case IBLND_MSG_GET_DONE:
1309 q = &conn->ibc_tx_queue_nocred;
1312 case IBLND_MSG_NOOP:
1313 q = &conn->ibc_tx_noops;
1316 case IBLND_MSG_IMMEDIATE:
1317 q = &conn->ibc_tx_queue;
1321 list_add_tail(&tx->tx_list, q);
1325 kiblnd_queue_tx (kib_tx_t *tx, kib_conn_t *conn)
1327 spin_lock(&conn->ibc_lock);
1328 kiblnd_queue_tx_locked(tx, conn);
1329 spin_unlock(&conn->ibc_lock);
1331 kiblnd_check_sends(conn);
1335 kiblnd_connect_peer (kib_peer_t *peer)
1337 struct rdma_cm_id *cmid;
1338 kib_net_t *net = peer->ibp_ni->ni_data;
1339 struct sockaddr_in srcaddr;
1340 struct sockaddr_in dstaddr;
1343 LASSERT (net != NULL);
1344 LASSERT (peer->ibp_connecting > 0);
1346 cmid = rdma_create_id(kiblnd_cm_callback, peer, RDMA_PS_TCP);
1348 CERROR("Can't create CMID for %s: %ld\n",
1349 libcfs_nid2str(peer->ibp_nid), PTR_ERR(cmid));
1354 memset(&srcaddr, 0, sizeof(srcaddr));
1355 srcaddr.sin_family = AF_INET;
1356 srcaddr.sin_addr.s_addr = htonl(net->ibn_dev->ibd_ifip);
1358 memset(&dstaddr, 0, sizeof(dstaddr));
1359 dstaddr.sin_family = AF_INET;
1360 dstaddr.sin_port = htons(*kiblnd_tunables.kib_service);
1361 dstaddr.sin_addr.s_addr = htonl(LNET_NIDADDR(peer->ibp_nid));
1363 kiblnd_peer_addref(peer); /* cmid's ref */
1365 rc = rdma_resolve_addr(cmid,
1366 (struct sockaddr *)&srcaddr,
1367 (struct sockaddr *)&dstaddr,
1368 *kiblnd_tunables.kib_timeout * 1000);
1372 /* Can't initiate address resolution: */
1373 CERROR("Can't resolve addr for %s: %d\n",
1374 libcfs_nid2str(peer->ibp_nid), rc);
1376 kiblnd_peer_decref(peer); /* cmid's ref */
1377 rdma_destroy_id(cmid);
1379 kiblnd_peer_connect_failed(peer, 1, rc);
1383 kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid)
1388 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
1389 unsigned long flags;
1392 /* If I get here, I've committed to send, so I complete the tx with
1393 * failure on any problems */
1395 LASSERT (tx->tx_conn == NULL); /* only set when assigned a conn */
1396 LASSERT (tx->tx_nwrq > 0); /* work items have been set up */
1398 /* First time, just use a read lock since I expect to find my peer
1400 read_lock_irqsave(g_lock, flags);
1402 peer = kiblnd_find_peer_locked(nid);
1403 if (peer != NULL && !list_empty(&peer->ibp_conns)) {
1404 /* Found a peer with an established connection */
1405 conn = kiblnd_get_conn_locked(peer);
1406 kiblnd_conn_addref(conn); /* 1 ref for me... */
1408 read_unlock_irqrestore(g_lock, flags);
1410 kiblnd_queue_tx(tx, conn);
1411 kiblnd_conn_decref(conn); /* ...to here */
1415 read_unlock(g_lock);
1416 /* Re-try with a write lock */
1419 peer = kiblnd_find_peer_locked(nid);
1421 if (list_empty(&peer->ibp_conns)) {
1422 /* found a peer, but it's still connecting... */
1423 LASSERT (peer->ibp_connecting != 0 ||
1424 peer->ibp_accepting != 0);
1425 list_add_tail (&tx->tx_list, &peer->ibp_tx_queue);
1426 write_unlock_irqrestore(g_lock, flags);
1428 conn = kiblnd_get_conn_locked(peer);
1429 kiblnd_conn_addref(conn); /* 1 ref for me... */
1431 write_unlock_irqrestore(g_lock, flags);
1433 kiblnd_queue_tx(tx, conn);
1434 kiblnd_conn_decref(conn); /* ...to here */
1439 write_unlock_irqrestore(g_lock, flags);
1441 /* Allocate a peer ready to add to the peer table and retry */
1442 rc = kiblnd_create_peer(ni, &peer, nid);
1444 CERROR("Can't create peer %s\n", libcfs_nid2str(nid));
1445 tx->tx_status = -EHOSTUNREACH;
1447 kiblnd_tx_done(ni, tx);
1451 write_lock_irqsave(g_lock, flags);
1453 peer2 = kiblnd_find_peer_locked(nid);
1454 if (peer2 != NULL) {
1455 if (list_empty(&peer2->ibp_conns)) {
1456 /* found a peer, but it's still connecting... */
1457 LASSERT (peer2->ibp_connecting != 0 ||
1458 peer2->ibp_accepting != 0);
1459 list_add_tail (&tx->tx_list, &peer2->ibp_tx_queue);
1460 write_unlock_irqrestore(g_lock, flags);
1462 conn = kiblnd_get_conn_locked(peer2);
1463 kiblnd_conn_addref(conn); /* 1 ref for me... */
1465 write_unlock_irqrestore(g_lock, flags);
1467 kiblnd_queue_tx(tx, conn);
1468 kiblnd_conn_decref(conn); /* ...to here */
1471 kiblnd_peer_decref(peer);
1475 /* Brand new peer */
1476 LASSERT (peer->ibp_connecting == 0);
1477 peer->ibp_connecting = 1;
1479 /* always called with a ref on ni, which prevents ni being shutdown */
1480 LASSERT (((kib_net_t *)ni->ni_data)->ibn_shutdown == 0);
1482 list_add_tail(&tx->tx_list, &peer->ibp_tx_queue);
1484 kiblnd_peer_addref(peer);
1485 list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid));
1487 write_unlock_irqrestore(g_lock, flags);
1489 kiblnd_connect_peer(peer);
1490 kiblnd_peer_decref(peer);
1494 kiblnd_send (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg)
1496 lnet_hdr_t *hdr = &lntmsg->msg_hdr;
1497 int type = lntmsg->msg_type;
1498 lnet_process_id_t target = lntmsg->msg_target;
1499 int target_is_router = lntmsg->msg_target_is_router;
1500 int routing = lntmsg->msg_routing;
1501 unsigned int payload_niov = lntmsg->msg_niov;
1502 struct iovec *payload_iov = lntmsg->msg_iov;
1503 lnet_kiov_t *payload_kiov = lntmsg->msg_kiov;
1504 unsigned int payload_offset = lntmsg->msg_offset;
1505 unsigned int payload_nob = lntmsg->msg_len;
1511 /* NB 'private' is different depending on what we're sending.... */
1513 CDEBUG(D_NET, "sending %d bytes in %d frags to %s\n",
1514 payload_nob, payload_niov, libcfs_id2str(target));
1516 LASSERT (payload_nob == 0 || payload_niov > 0);
1517 LASSERT (payload_niov <= LNET_MAX_IOV);
1519 /* Thread context */
1520 LASSERT (!in_interrupt());
1521 /* payload is either all vaddrs or all pages */
1522 LASSERT (!(payload_kiov != NULL && payload_iov != NULL));
1530 LASSERT (payload_nob == 0);
1534 if (routing || target_is_router)
1535 break; /* send IMMEDIATE */
1537 /* is the REPLY message too small for RDMA? */
1538 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[lntmsg->msg_md->md_length]);
1539 if (nob <= IBLND_MSG_SIZE)
1540 break; /* send IMMEDIATE */
1542 tx = kiblnd_get_idle_tx(ni);
1544 CERROR("Can allocate txd for GET to %s: \n",
1545 libcfs_nid2str(target.nid));
1550 ibmsg->ibm_u.get.ibgm_hdr = *hdr;
1551 ibmsg->ibm_u.get.ibgm_cookie = tx->tx_cookie;
1553 if ((lntmsg->msg_md->md_options & LNET_MD_KIOV) == 0)
1554 rc = kiblnd_setup_rd_iov(ni, tx,
1555 &ibmsg->ibm_u.get.ibgm_rd,
1556 lntmsg->msg_md->md_niov,
1557 lntmsg->msg_md->md_iov.iov,
1558 0, lntmsg->msg_md->md_length);
1560 rc = kiblnd_setup_rd_kiov(ni, tx,
1561 &ibmsg->ibm_u.get.ibgm_rd,
1562 lntmsg->msg_md->md_niov,
1563 lntmsg->msg_md->md_iov.kiov,
1564 0, lntmsg->msg_md->md_length);
1566 CERROR("Can't setup GET sink for %s: %d\n",
1567 libcfs_nid2str(target.nid), rc);
1568 kiblnd_tx_done(ni, tx);
1571 #if IBLND_MAP_ON_DEMAND
1572 nob = sizeof(kib_get_msg_t);
1574 nob = offsetof(kib_get_msg_t, ibgm_rd.rd_frags[tx->tx_nfrags]);
1576 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_GET_REQ, nob);
1578 tx->tx_lntmsg[1] = lnet_create_reply_msg(ni, lntmsg);
1579 if (tx->tx_lntmsg[1] == NULL) {
1580 CERROR("Can't create reply for GET -> %s\n",
1581 libcfs_nid2str(target.nid));
1582 kiblnd_tx_done(ni, tx);
1586 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg[0,1] on completion */
1587 tx->tx_waiting = 1; /* waiting for GET_DONE */
1588 kiblnd_launch_tx(ni, tx, target.nid);
1591 case LNET_MSG_REPLY:
1593 /* Is the payload small enough not to need RDMA? */
1594 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob]);
1595 if (nob <= IBLND_MSG_SIZE)
1596 break; /* send IMMEDIATE */
1598 tx = kiblnd_get_idle_tx(ni);
1600 CERROR("Can't allocate %s txd for %s\n",
1601 type == LNET_MSG_PUT ? "PUT" : "REPLY",
1602 libcfs_nid2str(target.nid));
1606 if (payload_kiov == NULL)
1607 rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1608 payload_niov, payload_iov,
1609 payload_offset, payload_nob);
1611 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1612 payload_niov, payload_kiov,
1613 payload_offset, payload_nob);
1615 CERROR("Can't setup PUT src for %s: %d\n",
1616 libcfs_nid2str(target.nid), rc);
1617 kiblnd_tx_done(ni, tx);
1622 ibmsg->ibm_u.putreq.ibprm_hdr = *hdr;
1623 ibmsg->ibm_u.putreq.ibprm_cookie = tx->tx_cookie;
1624 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_REQ, sizeof(kib_putreq_msg_t));
1626 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1627 tx->tx_waiting = 1; /* waiting for PUT_{ACK,NAK} */
1628 kiblnd_launch_tx(ni, tx, target.nid);
1632 /* send IMMEDIATE */
1634 LASSERT (offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob])
1637 tx = kiblnd_get_idle_tx(ni);
1639 CERROR ("Can't send %d to %s: tx descs exhausted\n",
1640 type, libcfs_nid2str(target.nid));
1645 ibmsg->ibm_u.immediate.ibim_hdr = *hdr;
1647 if (payload_kiov != NULL)
1648 lnet_copy_kiov2flat(IBLND_MSG_SIZE, ibmsg,
1649 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1650 payload_niov, payload_kiov,
1651 payload_offset, payload_nob);
1653 lnet_copy_iov2flat(IBLND_MSG_SIZE, ibmsg,
1654 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1655 payload_niov, payload_iov,
1656 payload_offset, payload_nob);
1658 nob = offsetof(kib_immediate_msg_t, ibim_payload[payload_nob]);
1659 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_IMMEDIATE, nob);
1661 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1662 kiblnd_launch_tx(ni, tx, target.nid);
1667 kiblnd_reply (lnet_ni_t *ni, kib_rx_t *rx, lnet_msg_t *lntmsg)
1669 lnet_process_id_t target = lntmsg->msg_target;
1670 unsigned int niov = lntmsg->msg_niov;
1671 struct iovec *iov = lntmsg->msg_iov;
1672 lnet_kiov_t *kiov = lntmsg->msg_kiov;
1673 unsigned int offset = lntmsg->msg_offset;
1674 unsigned int nob = lntmsg->msg_len;
1678 tx = kiblnd_get_idle_tx(ni);
1680 CERROR("Can't get tx for REPLY to %s\n",
1681 libcfs_nid2str(target.nid));
1687 else if (kiov == NULL)
1688 rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1689 niov, iov, offset, nob);
1691 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1692 niov, kiov, offset, nob);
1695 CERROR("Can't setup GET src for %s: %d\n",
1696 libcfs_nid2str(target.nid), rc);
1700 rc = kiblnd_init_rdma(ni, tx, IBLND_MSG_GET_DONE, nob,
1701 &rx->rx_msg->ibm_u.get.ibgm_rd,
1702 rx->rx_msg->ibm_u.get.ibgm_cookie);
1704 CERROR("Can't setup rdma for GET from %s: %d\n",
1705 libcfs_nid2str(target.nid), rc);
1710 /* No RDMA: local completion may happen now! */
1711 lnet_finalize(ni, lntmsg, 0);
1713 /* RDMA: lnet_finalize(lntmsg) when it
1715 tx->tx_lntmsg[0] = lntmsg;
1718 kiblnd_queue_tx(tx, rx->rx_conn);
1722 kiblnd_tx_done(ni, tx);
1724 lnet_finalize(ni, lntmsg, -EIO);
1728 kiblnd_recv (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed,
1729 unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov,
1730 unsigned int offset, unsigned int mlen, unsigned int rlen)
1732 kib_rx_t *rx = private;
1733 kib_msg_t *rxmsg = rx->rx_msg;
1734 kib_conn_t *conn = rx->rx_conn;
1738 int post_credit = IBLND_POSTRX_PEER_CREDIT;
1741 LASSERT (mlen <= rlen);
1742 LASSERT (!in_interrupt());
1743 /* Either all pages or all vaddrs */
1744 LASSERT (!(kiov != NULL && iov != NULL));
1746 switch (rxmsg->ibm_type) {
1750 case IBLND_MSG_IMMEDIATE:
1751 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[rlen]);
1752 if (nob > rx->rx_nob) {
1753 CERROR ("Immediate message from %s too big: %d(%d)\n",
1754 libcfs_nid2str(rxmsg->ibm_u.immediate.ibim_hdr.src_nid),
1761 lnet_copy_flat2kiov(niov, kiov, offset,
1762 IBLND_MSG_SIZE, rxmsg,
1763 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1766 lnet_copy_flat2iov(niov, iov, offset,
1767 IBLND_MSG_SIZE, rxmsg,
1768 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1770 lnet_finalize (ni, lntmsg, 0);
1773 case IBLND_MSG_PUT_REQ:
1775 lnet_finalize(ni, lntmsg, 0);
1776 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK, 0,
1777 rxmsg->ibm_u.putreq.ibprm_cookie);
1781 tx = kiblnd_get_idle_tx(ni);
1783 CERROR("Can't allocate tx for %s\n",
1784 libcfs_nid2str(conn->ibc_peer->ibp_nid));
1785 /* Not replying will break the connection */
1792 rc = kiblnd_setup_rd_iov(ni, tx,
1793 &txmsg->ibm_u.putack.ibpam_rd,
1794 niov, iov, offset, mlen);
1796 rc = kiblnd_setup_rd_kiov(ni, tx,
1797 &txmsg->ibm_u.putack.ibpam_rd,
1798 niov, kiov, offset, mlen);
1800 CERROR("Can't setup PUT sink for %s: %d\n",
1801 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
1802 kiblnd_tx_done(ni, tx);
1803 /* tell peer it's over */
1804 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK, rc,
1805 rxmsg->ibm_u.putreq.ibprm_cookie);
1809 txmsg->ibm_u.putack.ibpam_src_cookie = rxmsg->ibm_u.putreq.ibprm_cookie;
1810 txmsg->ibm_u.putack.ibpam_dst_cookie = tx->tx_cookie;
1811 #if IBLND_MAP_ON_DEMAND
1812 nob = sizeof(kib_putack_msg_t);
1814 nob = offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[tx->tx_nfrags]);
1816 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_ACK, nob);
1818 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1819 tx->tx_waiting = 1; /* waiting for PUT_DONE */
1820 kiblnd_queue_tx(tx, conn);
1822 /* reposted buffer reserved for PUT_DONE */
1823 post_credit = IBLND_POSTRX_NO_CREDIT;
1826 case IBLND_MSG_GET_REQ:
1827 if (lntmsg != NULL) {
1828 /* Optimized GET; RDMA lntmsg's payload */
1829 kiblnd_reply(ni, rx, lntmsg);
1831 /* GET didn't match anything */
1832 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_GET_DONE,
1834 rxmsg->ibm_u.get.ibgm_cookie);
1839 kiblnd_post_rx(rx, post_credit);
1844 kiblnd_thread_start (int (*fn)(void *arg), void *arg)
1846 long pid = kernel_thread (fn, arg, 0);
1851 atomic_inc (&kiblnd_data.kib_nthreads);
1856 kiblnd_thread_fini (void)
1858 atomic_dec (&kiblnd_data.kib_nthreads);
1862 kiblnd_peer_alive (kib_peer_t *peer)
1864 /* This is racy, but everyone's only writing cfs_time_current() */
1865 peer->ibp_last_alive = cfs_time_current();
1870 kiblnd_peer_notify (kib_peer_t *peer)
1872 time_t last_alive = 0;
1874 unsigned long flags;
1876 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1878 if (list_empty(&peer->ibp_conns) &&
1879 peer->ibp_accepting == 0 &&
1880 peer->ibp_connecting == 0 &&
1881 peer->ibp_error != 0) {
1882 error = peer->ibp_error;
1883 peer->ibp_error = 0;
1885 last_alive = cfs_time_current_sec() -
1886 cfs_duration_sec(cfs_time_current() -
1887 peer->ibp_last_alive);
1890 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1893 lnet_notify(peer->ibp_ni,
1894 peer->ibp_nid, 0, last_alive);
1898 kiblnd_close_conn_locked (kib_conn_t *conn, int error)
1900 /* This just does the immediate housekeeping. 'error' is zero for a
1901 * normal shutdown which can happen only after the connection has been
1902 * established. If the connection is established, schedule the
1903 * connection to be finished off by the connd. Otherwise the connd is
1904 * already dealing with it (either to set it up or tear it down).
1905 * Caller holds kib_global_lock exclusively in irq context */
1906 unsigned long flags;
1907 kib_peer_t *peer = conn->ibc_peer;
1909 LASSERT (error != 0 || conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1911 if (error != 0 && conn->ibc_comms_error == 0)
1912 conn->ibc_comms_error = error;
1914 if (conn->ibc_state != IBLND_CONN_ESTABLISHED)
1915 return; /* already being handled */
1918 list_empty(&conn->ibc_tx_noops) &&
1919 list_empty(&conn->ibc_tx_queue) &&
1920 list_empty(&conn->ibc_tx_queue_rsrvd) &&
1921 list_empty(&conn->ibc_tx_queue_nocred) &&
1922 list_empty(&conn->ibc_active_txs)) {
1923 CDEBUG(D_NET, "closing conn to %s\n",
1924 libcfs_nid2str(peer->ibp_nid));
1926 CDEBUG(D_NETERROR, "Closing conn to %s: error %d%s%s%s%s%s\n",
1927 libcfs_nid2str(peer->ibp_nid), error,
1928 list_empty(&conn->ibc_tx_queue) ? "" : "(sending)",
1929 list_empty(&conn->ibc_tx_noops) ? "" : "(sending_noops)",
1930 list_empty(&conn->ibc_tx_queue_rsrvd) ? "" : "(sending_rsrvd)",
1931 list_empty(&conn->ibc_tx_queue_nocred) ? "" : "(sending_nocred)",
1932 list_empty(&conn->ibc_active_txs) ? "" : "(waiting)");
1935 list_del (&conn->ibc_list);
1936 /* connd (see below) takes over ibc_list's ref */
1938 if (list_empty (&peer->ibp_conns) && /* no more conns */
1939 kiblnd_peer_active(peer)) { /* still in peer table */
1940 kiblnd_unlink_peer_locked(peer);
1942 /* set/clear error on last conn */
1943 peer->ibp_error = conn->ibc_comms_error;
1946 kiblnd_set_conn_state(conn, IBLND_CONN_CLOSING);
1948 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
1950 list_add_tail (&conn->ibc_list, &kiblnd_data.kib_connd_conns);
1951 wake_up (&kiblnd_data.kib_connd_waitq);
1953 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
1957 kiblnd_close_conn (kib_conn_t *conn, int error)
1959 unsigned long flags;
1961 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1963 kiblnd_close_conn_locked(conn, error);
1965 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1969 kiblnd_handle_early_rxs(kib_conn_t *conn)
1971 unsigned long flags;
1974 LASSERT (!in_interrupt());
1975 LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1977 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1978 while (!list_empty(&conn->ibc_early_rxs)) {
1979 rx = list_entry(conn->ibc_early_rxs.next,
1981 list_del(&rx->rx_list);
1982 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1984 kiblnd_handle_rx(rx);
1986 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1988 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1992 kiblnd_abort_txs(kib_conn_t *conn, struct list_head *txs)
1994 LIST_HEAD (zombies);
1995 struct list_head *tmp;
1996 struct list_head *nxt;
1999 spin_lock(&conn->ibc_lock);
2001 list_for_each_safe (tmp, nxt, txs) {
2002 tx = list_entry (tmp, kib_tx_t, tx_list);
2004 if (txs == &conn->ibc_active_txs) {
2005 LASSERT (!tx->tx_queued);
2006 LASSERT (tx->tx_waiting ||
2007 tx->tx_sending != 0);
2009 LASSERT (tx->tx_queued);
2012 tx->tx_status = -ECONNABORTED;
2016 if (tx->tx_sending == 0) {
2017 list_del (&tx->tx_list);
2018 list_add (&tx->tx_list, &zombies);
2022 spin_unlock(&conn->ibc_lock);
2024 kiblnd_txlist_done(conn->ibc_peer->ibp_ni,
2025 &zombies, -ECONNABORTED);
2029 kiblnd_finalise_conn (kib_conn_t *conn)
2031 LASSERT (!in_interrupt());
2032 LASSERT (conn->ibc_state > IBLND_CONN_INIT);
2034 kiblnd_set_conn_state(conn, IBLND_CONN_DISCONNECTED);
2036 /* abort_receives moves QP state to IB_QPS_ERR. This is only required
2037 * for connections that didn't get as far as being connected, because
2038 * rdma_disconnect() does this for free. */
2039 kiblnd_abort_receives(conn);
2041 /* Complete all tx descs not waiting for sends to complete.
2042 * NB we should be safe from RDMA now that the QP has changed state */
2044 kiblnd_abort_txs(conn, &conn->ibc_tx_noops);
2045 kiblnd_abort_txs(conn, &conn->ibc_tx_queue);
2046 kiblnd_abort_txs(conn, &conn->ibc_tx_queue_rsrvd);
2047 kiblnd_abort_txs(conn, &conn->ibc_tx_queue_nocred);
2048 kiblnd_abort_txs(conn, &conn->ibc_active_txs);
2050 kiblnd_handle_early_rxs(conn);
2054 kiblnd_peer_connect_failed (kib_peer_t *peer, int active, int error)
2056 LIST_HEAD (zombies);
2057 unsigned long flags;
2059 LASSERT (error != 0);
2060 LASSERT (!in_interrupt());
2062 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2065 LASSERT (peer->ibp_connecting > 0);
2066 peer->ibp_connecting--;
2068 LASSERT (peer->ibp_accepting > 0);
2069 peer->ibp_accepting--;
2072 if (peer->ibp_connecting != 0 ||
2073 peer->ibp_accepting != 0) {
2074 /* another connection attempt under way... */
2075 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2079 if (list_empty(&peer->ibp_conns)) {
2080 /* Take peer's blocked transmits to complete with error */
2081 list_add(&zombies, &peer->ibp_tx_queue);
2082 list_del_init(&peer->ibp_tx_queue);
2084 if (kiblnd_peer_active(peer))
2085 kiblnd_unlink_peer_locked(peer);
2087 peer->ibp_error = error;
2089 /* Can't have blocked transmits if there are connections */
2090 LASSERT (list_empty(&peer->ibp_tx_queue));
2093 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2095 kiblnd_peer_notify(peer);
2097 if (list_empty (&zombies))
2100 CDEBUG (D_NETERROR, "Deleting messages for %s: connection failed\n",
2101 libcfs_nid2str(peer->ibp_nid));
2103 kiblnd_txlist_done(peer->ibp_ni, &zombies, -EHOSTUNREACH);
2107 kiblnd_connreq_done(kib_conn_t *conn, int status)
2109 struct list_head txs;
2111 kib_peer_t *peer = conn->ibc_peer;
2113 unsigned long flags;
2116 active = (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2118 CDEBUG(D_NET,"%s: %d, %d\n", libcfs_nid2str(peer->ibp_nid),
2121 LASSERT (!in_interrupt());
2122 LASSERT ((conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT &&
2123 peer->ibp_connecting > 0) ||
2124 (conn->ibc_state == IBLND_CONN_PASSIVE_WAIT &&
2125 peer->ibp_accepting > 0));
2127 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
2128 conn->ibc_connvars = NULL;
2131 /* failed to establish connection */
2132 kiblnd_peer_connect_failed(peer, active, status);
2133 kiblnd_finalise_conn(conn);
2137 /* connection established */
2138 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2140 conn->ibc_last_send = jiffies;
2141 kiblnd_set_conn_state(conn, IBLND_CONN_ESTABLISHED);
2142 kiblnd_peer_alive(peer);
2144 /* Add conn to peer's list and nuke any dangling conns from a different
2145 * peer instance... */
2146 kiblnd_conn_addref(conn); /* +1 ref for ibc_list */
2147 list_add(&conn->ibc_list, &peer->ibp_conns);
2149 peer->ibp_connecting--;
2151 peer->ibp_accepting--;
2153 kiblnd_close_stale_conns_locked(peer, conn->ibc_incarnation);
2155 /* grab pending txs while I have the lock */
2156 list_add(&txs, &peer->ibp_tx_queue);
2157 list_del_init(&peer->ibp_tx_queue);
2159 if (!kiblnd_peer_active(peer) || /* peer has been deleted */
2160 conn->ibc_comms_error != 0) { /* error has happened already */
2161 lnet_ni_t *ni = peer->ibp_ni;
2163 /* start to shut down connection */
2164 kiblnd_close_conn_locked(conn, -ECONNABORTED);
2165 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2167 kiblnd_txlist_done(ni, &txs, -ECONNABORTED);
2172 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2174 /* Schedule blocked txs */
2175 spin_lock (&conn->ibc_lock);
2176 while (!list_empty (&txs)) {
2177 tx = list_entry (txs.next, kib_tx_t, tx_list);
2178 list_del (&tx->tx_list);
2180 kiblnd_queue_tx_locked(tx, conn);
2182 spin_unlock (&conn->ibc_lock);
2184 kiblnd_check_sends(conn);
2186 /* schedule blocked rxs */
2187 kiblnd_handle_early_rxs(conn);
2191 kiblnd_reject(struct rdma_cm_id *cmid, int why)
2194 kib_rej_t rej = {.ibr_magic = IBLND_MSG_MAGIC,
2195 .ibr_version = IBLND_MSG_VERSION,
2198 rc = rdma_reject(cmid, &rej, sizeof(rej));
2201 CWARN("Error %d sending reject\n", rc);
2205 kiblnd_passive_connect (struct rdma_cm_id *cmid, void *priv, int priv_nob)
2208 kib_msg_t *reqmsg = priv;
2209 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2210 struct rdma_conn_param cp;
2211 unsigned long flags;
2212 lnet_ni_t *ni = NULL;
2219 int rej = IBLND_REJECT_FATAL;
2221 LASSERT (!in_interrupt());
2223 /* cmid inherits 'context' from the corresponding listener id */
2224 ibdev = (kib_dev_t *)cmid->context;
2225 LASSERT (ibdev != NULL);
2227 if (priv_nob < offsetof(kib_msg_t, ibm_type)) {
2228 CERROR("Short connection request\n");
2232 if (reqmsg->ibm_magic == LNET_PROTO_MAGIC ||
2233 reqmsg->ibm_magic == __swab32(LNET_PROTO_MAGIC) ||
2234 (reqmsg->ibm_magic == IBLND_MSG_MAGIC &&
2235 reqmsg->ibm_version != IBLND_MSG_VERSION) ||
2236 (reqmsg->ibm_magic == __swab32(IBLND_MSG_MAGIC) &&
2237 reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION))) {
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
2247 rc = kiblnd_unpack_msg(reqmsg, priv_nob);
2249 CERROR("Can't parse connection request: %d\n", rc);
2253 nid = reqmsg->ibm_srcnid;
2255 if (reqmsg->ibm_type != IBLND_MSG_CONNREQ) {
2256 CERROR("Unexpected connreq msg type: %x from %s\n",
2257 reqmsg->ibm_type, libcfs_nid2str(nid));
2261 if (reqmsg->ibm_u.connparams.ibcp_queue_depth != IBLND_MSG_QUEUE_SIZE) {
2262 CERROR("Can't accept %s: incompatible queue depth %d (%d wanted)\n",
2263 libcfs_nid2str(nid),
2264 reqmsg->ibm_u.connparams.ibcp_queue_depth,
2265 IBLND_MSG_QUEUE_SIZE);
2269 if (reqmsg->ibm_u.connparams.ibcp_max_frags != IBLND_MAX_RDMA_FRAGS) {
2270 CERROR("Can't accept %s: incompatible max_frags %d (%d wanted)\n",
2271 libcfs_nid2str(nid),
2272 reqmsg->ibm_u.connparams.ibcp_max_frags,
2273 IBLND_MAX_RDMA_FRAGS);
2277 if (reqmsg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2278 CERROR("Can't accept %s: message size %d too big (%d max)\n",
2279 libcfs_nid2str(nid),
2280 reqmsg->ibm_u.connparams.ibcp_max_msg_size,
2285 ni = lnet_net2ni(LNET_NIDNET(reqmsg->ibm_dstnid));
2286 if (ni == NULL || /* no matching net */
2287 ni->ni_nid != reqmsg->ibm_dstnid || /* right NET, wrong NID! */
2288 ((kib_net_t*)ni->ni_data)->ibn_dev != ibdev) { /* wrong device */
2289 CERROR("Can't accept %s: bad dst nid %s\n",
2290 libcfs_nid2str(nid),
2291 libcfs_nid2str(reqmsg->ibm_dstnid));
2296 /* assume 'nid' is a new peer; create */
2297 rc = kiblnd_create_peer(ni, &peer, nid);
2299 CERROR("Can't create peer for %s\n", libcfs_nid2str(nid));
2300 rej = IBLND_REJECT_NO_RESOURCES;
2304 write_lock_irqsave(g_lock, flags);
2306 peer2 = kiblnd_find_peer_locked(nid);
2307 if (peer2 != NULL) {
2308 /* tie-break connection race in favour of the higher NID */
2309 if (peer2->ibp_connecting != 0 &&
2311 write_unlock_irqrestore(g_lock, flags);
2313 CWARN("Conn race %s\n",
2314 libcfs_nid2str(peer2->ibp_nid));
2316 kiblnd_peer_decref(peer);
2317 rej = IBLND_REJECT_CONN_RACE;
2321 peer2->ibp_accepting++;
2322 kiblnd_peer_addref(peer2);
2324 write_unlock_irqrestore(g_lock, flags);
2325 kiblnd_peer_decref(peer);
2328 /* Brand new peer */
2329 LASSERT (peer->ibp_accepting == 0);
2330 peer->ibp_accepting = 1;
2332 /* I have a ref on ni that prevents it being shutdown */
2333 LASSERT (((kib_net_t *)ni->ni_data)->ibn_shutdown == 0);
2335 kiblnd_peer_addref(peer);
2336 list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid));
2338 write_unlock_irqrestore(g_lock, flags);
2341 conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_PASSIVE_WAIT);
2343 kiblnd_peer_connect_failed(peer, 0, -ENOMEM);
2344 kiblnd_peer_decref(peer);
2345 rej = IBLND_REJECT_NO_RESOURCES;
2349 /* conn now "owns" cmid, so I return success from here on to ensure the
2350 * CM callback doesn't destroy cmid. */
2352 conn->ibc_incarnation = reqmsg->ibm_srcstamp;
2353 conn->ibc_credits = IBLND_MSG_QUEUE_SIZE;
2354 conn->ibc_reserved_credits = IBLND_MSG_QUEUE_SIZE;
2355 LASSERT (conn->ibc_credits + conn->ibc_reserved_credits
2358 ackmsg = &conn->ibc_connvars->cv_msg;
2359 memset(ackmsg, 0, sizeof(*ackmsg));
2361 kiblnd_init_msg(ackmsg, IBLND_MSG_CONNACK,
2362 sizeof(ackmsg->ibm_u.connparams));
2363 ackmsg->ibm_u.connparams.ibcp_queue_depth = IBLND_MSG_QUEUE_SIZE;
2364 ackmsg->ibm_u.connparams.ibcp_max_frags = IBLND_MAX_RDMA_FRAGS;
2365 ackmsg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2366 kiblnd_pack_msg(ni, ackmsg, 0, nid, reqmsg->ibm_srcstamp);
2368 memset(&cp, 0, sizeof(cp));
2369 cp.private_data = ackmsg;
2370 cp.private_data_len = ackmsg->ibm_nob;
2371 cp.responder_resources = 0; /* No atomic ops or RDMA reads */
2372 cp.initiator_depth = 0;
2373 cp.flow_control = 1;
2374 cp.retry_count = *kiblnd_tunables.kib_retry_count;
2375 cp.rnr_retry_count = *kiblnd_tunables.kib_rnr_retry_count;
2377 CDEBUG(D_NET, "Accept %s\n", libcfs_nid2str(nid));
2379 rc = rdma_accept(cmid, &cp);
2381 CERROR("Can't accept %s: %d\n", libcfs_nid2str(nid), rc);
2382 kiblnd_reject(cmid, IBLND_REJECT_FATAL);
2383 kiblnd_connreq_done(conn, rc);
2384 kiblnd_conn_decref(conn);
2394 kiblnd_reject(cmid, rej);
2395 return -ECONNREFUSED;
2399 kiblnd_reconnect (kib_conn_t *conn, char *why)
2401 kib_peer_t *peer = conn->ibc_peer;
2403 unsigned long flags;
2405 LASSERT (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2406 LASSERT (peer->ibp_connecting > 0); /* 'conn' at least */
2408 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2410 /* retry connection if it's still needed and no other connection
2411 * attempts (active or passive) are in progress */
2412 if (!list_empty(&peer->ibp_tx_queue) &&
2413 peer->ibp_connecting == 1 &&
2414 peer->ibp_accepting == 0) {
2416 peer->ibp_connecting++;
2419 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2422 CDEBUG(D_NETERROR, "%s: retrying (%s)\n",
2423 libcfs_nid2str(peer->ibp_nid), why);
2424 kiblnd_connect_peer(peer);
2429 kiblnd_rejected (kib_conn_t *conn, int reason, void *priv, int priv_nob)
2431 kib_peer_t *peer = conn->ibc_peer;
2433 LASSERT (!in_interrupt());
2434 LASSERT (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2437 case IB_CM_REJ_STALE_CONN:
2438 kiblnd_reconnect(conn, "stale");
2441 case IB_CM_REJ_CONSUMER_DEFINED:
2442 if (priv_nob >= sizeof(kib_rej_t)) {
2443 kib_rej_t *rej = priv;
2445 if (rej->ibr_magic == __swab32(IBLND_MSG_MAGIC) ||
2446 rej->ibr_magic == __swab32(LNET_PROTO_MAGIC)) {
2447 __swab32s(&rej->ibr_magic);
2448 __swab16s(&rej->ibr_version);
2451 if (rej->ibr_magic != IBLND_MSG_MAGIC &&
2452 rej->ibr_magic != LNET_PROTO_MAGIC) {
2453 CERROR("%s rejected: consumer defined fatal error\n",
2454 libcfs_nid2str(peer->ibp_nid));
2458 if (rej->ibr_version != IBLND_MSG_VERSION) {
2459 CERROR("%s rejected: o2iblnd version %d error\n",
2460 libcfs_nid2str(peer->ibp_nid),
2465 switch (rej->ibr_why) {
2466 case IBLND_REJECT_CONN_RACE:
2467 kiblnd_reconnect(conn, "conn race");
2470 case IBLND_REJECT_NO_RESOURCES:
2471 CERROR("%s rejected: o2iblnd no resources\n",
2472 libcfs_nid2str(peer->ibp_nid));
2474 case IBLND_REJECT_FATAL:
2475 CERROR("%s rejected: o2iblnd fatal error\n",
2476 libcfs_nid2str(peer->ibp_nid));
2479 CERROR("%s rejected: o2iblnd reason %d\n",
2480 libcfs_nid2str(peer->ibp_nid),
2488 CDEBUG(D_NETERROR, "%s rejected: reason %d, size %d\n",
2489 libcfs_nid2str(peer->ibp_nid), reason, priv_nob);
2493 kiblnd_connreq_done(conn, -ECONNREFUSED);
2497 kiblnd_check_connreply (kib_conn_t *conn, void *priv, int priv_nob)
2499 kib_peer_t *peer = conn->ibc_peer;
2500 lnet_ni_t *ni = peer->ibp_ni;
2501 kib_net_t *net = ni->ni_data;
2502 kib_msg_t *msg = priv;
2503 int rc = kiblnd_unpack_msg(msg, priv_nob);
2504 unsigned long flags;
2506 LASSERT (net != NULL);
2509 CERROR("Can't unpack connack from %s: %d\n",
2510 libcfs_nid2str(peer->ibp_nid), rc);
2514 if (msg->ibm_type != IBLND_MSG_CONNACK) {
2515 CERROR("Unexpected message %d from %s\n",
2516 msg->ibm_type, libcfs_nid2str(peer->ibp_nid));
2521 if (msg->ibm_u.connparams.ibcp_queue_depth != IBLND_MSG_QUEUE_SIZE) {
2522 CERROR("%s has incompatible queue depth %d(%d wanted)\n",
2523 libcfs_nid2str(peer->ibp_nid),
2524 msg->ibm_u.connparams.ibcp_queue_depth,
2525 IBLND_MSG_QUEUE_SIZE);
2530 if (msg->ibm_u.connparams.ibcp_max_frags != IBLND_MAX_RDMA_FRAGS) {
2531 CERROR("%s has incompatible max_frags %d (%d wanted)\n",
2532 libcfs_nid2str(peer->ibp_nid),
2533 msg->ibm_u.connparams.ibcp_max_frags,
2534 IBLND_MAX_RDMA_FRAGS);
2539 if (msg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2540 CERROR("%s max message size %d too big (%d max)\n",
2541 libcfs_nid2str(peer->ibp_nid),
2542 msg->ibm_u.connparams.ibcp_max_msg_size,
2548 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2549 if (msg->ibm_dstnid == ni->ni_nid &&
2550 msg->ibm_dststamp == net->ibn_incarnation)
2554 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2557 CERROR("Stale connection reply from %s\n",
2558 libcfs_nid2str(peer->ibp_nid));
2562 conn->ibc_incarnation = msg->ibm_srcstamp;
2563 conn->ibc_credits = IBLND_MSG_QUEUE_SIZE;
2564 conn->ibc_reserved_credits = IBLND_MSG_QUEUE_SIZE;
2565 LASSERT (conn->ibc_credits + conn->ibc_reserved_credits
2568 kiblnd_connreq_done(conn, 0);
2572 /* NB My QP has already established itself, so I handle anything going
2573 * wrong here by setting ibc_comms_error.
2574 * kiblnd_connreq_done(0) moves the conn state to ESTABLISHED, but then
2575 * immediately tears it down. */
2578 conn->ibc_comms_error = rc;
2579 kiblnd_connreq_done(conn, 0);
2583 kiblnd_active_connect (struct rdma_cm_id *cmid)
2585 kib_peer_t *peer = (kib_peer_t *)cmid->context;
2588 struct rdma_conn_param cp;
2591 conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_ACTIVE_CONNECT);
2593 kiblnd_peer_connect_failed(peer, 1, -ENOMEM);
2594 kiblnd_peer_decref(peer); /* lose cmid's ref */
2598 /* conn "owns" cmid now, so I return success from here on to ensure the
2599 * CM callback doesn't destroy cmid. conn also takes over cmid's ref
2602 msg = &conn->ibc_connvars->cv_msg;
2604 memset(msg, 0, sizeof(*msg));
2605 kiblnd_init_msg(msg, IBLND_MSG_CONNREQ, sizeof(msg->ibm_u.connparams));
2606 msg->ibm_u.connparams.ibcp_queue_depth = IBLND_MSG_QUEUE_SIZE;
2607 msg->ibm_u.connparams.ibcp_max_frags = IBLND_MAX_RDMA_FRAGS;
2608 msg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2609 kiblnd_pack_msg(peer->ibp_ni, msg, 0, peer->ibp_nid, 0);
2611 memset(&cp, 0, sizeof(cp));
2612 cp.private_data = msg;
2613 cp.private_data_len = msg->ibm_nob;
2614 cp.responder_resources = 0; /* No atomic ops or RDMA reads */
2615 cp.initiator_depth = 0;
2616 cp.flow_control = 1;
2617 cp.retry_count = *kiblnd_tunables.kib_retry_count;
2618 cp.rnr_retry_count = *kiblnd_tunables.kib_rnr_retry_count;
2620 LASSERT(cmid->context == (void *)conn);
2621 LASSERT(conn->ibc_cmid == cmid);
2623 rc = rdma_connect(cmid, &cp);
2625 CERROR("Can't connect to %s: %d\n",
2626 libcfs_nid2str(peer->ibp_nid), rc);
2627 kiblnd_connreq_done(conn, rc);
2628 kiblnd_conn_decref(conn);
2635 kiblnd_cm_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2641 switch (event->event) {
2645 case RDMA_CM_EVENT_CONNECT_REQUEST:
2646 /* destroy cmid on failure */
2647 rc = kiblnd_passive_connect(cmid,
2648 (void *)KIBLND_CONN_PARAM(event),
2649 KIBLND_CONN_PARAM_LEN(event));
2650 CDEBUG(D_NET, "connreq: %d\n", rc);
2653 case RDMA_CM_EVENT_ADDR_ERROR:
2654 peer = (kib_peer_t *)cmid->context;
2655 CDEBUG(D_NETERROR, "%s: ADDR ERROR %d\n",
2656 libcfs_nid2str(peer->ibp_nid), event->status);
2657 kiblnd_peer_connect_failed(peer, 1, -EHOSTUNREACH);
2658 kiblnd_peer_decref(peer);
2659 return -EHOSTUNREACH; /* rc != 0 destroys cmid */
2661 case RDMA_CM_EVENT_ADDR_RESOLVED:
2662 peer = (kib_peer_t *)cmid->context;
2664 CDEBUG(D_NET,"%s Addr resolved: %d\n",
2665 libcfs_nid2str(peer->ibp_nid), event->status);
2667 if (event->status != 0) {
2668 CDEBUG(D_NETERROR, "Can't resolve address for %s: %d\n",
2669 libcfs_nid2str(peer->ibp_nid), event->status);
2672 rc = rdma_resolve_route(
2673 cmid, *kiblnd_tunables.kib_timeout * 1000);
2676 /* Can't initiate route resolution */
2677 CERROR("Can't resolve route for %s: %d\n",
2678 libcfs_nid2str(peer->ibp_nid), rc);
2680 kiblnd_peer_connect_failed(peer, 1, rc);
2681 kiblnd_peer_decref(peer);
2682 return rc; /* rc != 0 destroys cmid */
2684 case RDMA_CM_EVENT_ROUTE_ERROR:
2685 peer = (kib_peer_t *)cmid->context;
2686 CDEBUG(D_NETERROR, "%s: ROUTE ERROR %d\n",
2687 libcfs_nid2str(peer->ibp_nid), event->status);
2688 kiblnd_peer_connect_failed(peer, 1, -EHOSTUNREACH);
2689 kiblnd_peer_decref(peer);
2690 return -EHOSTUNREACH; /* rc != 0 destroys cmid */
2692 case RDMA_CM_EVENT_ROUTE_RESOLVED:
2693 peer = (kib_peer_t *)cmid->context;
2694 CDEBUG(D_NET,"%s Route resolved: %d\n",
2695 libcfs_nid2str(peer->ibp_nid), event->status);
2697 if (event->status == 0)
2698 return kiblnd_active_connect(cmid);
2700 CDEBUG(D_NETERROR, "Can't resolve route for %s: %d\n",
2701 libcfs_nid2str(peer->ibp_nid), event->status);
2702 kiblnd_peer_connect_failed(peer, 1, event->status);
2703 kiblnd_peer_decref(peer);
2704 return event->status; /* rc != 0 destroys cmid */
2706 case RDMA_CM_EVENT_UNREACHABLE:
2707 conn = (kib_conn_t *)cmid->context;
2708 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
2709 conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
2710 CDEBUG(D_NETERROR, "%s: UNREACHABLE %d\n",
2711 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
2712 kiblnd_connreq_done(conn, -ENETDOWN);
2713 kiblnd_conn_decref(conn);
2716 case RDMA_CM_EVENT_CONNECT_ERROR:
2717 conn = (kib_conn_t *)cmid->context;
2718 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
2719 conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
2720 CDEBUG(D_NETERROR, "%s: CONNECT ERROR %d\n",
2721 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
2722 kiblnd_connreq_done(conn, -ENOTCONN);
2723 kiblnd_conn_decref(conn);
2726 case RDMA_CM_EVENT_REJECTED:
2727 conn = (kib_conn_t *)cmid->context;
2728 switch (conn->ibc_state) {
2732 case IBLND_CONN_PASSIVE_WAIT:
2733 CERROR ("%s: REJECTED %d\n",
2734 libcfs_nid2str(conn->ibc_peer->ibp_nid),
2736 kiblnd_connreq_done(conn, -ECONNRESET);
2739 case IBLND_CONN_ACTIVE_CONNECT:
2740 kiblnd_rejected(conn, event->status,
2741 (void *)KIBLND_CONN_PARAM(event),
2742 KIBLND_CONN_PARAM_LEN(event));
2745 kiblnd_conn_decref(conn);
2748 case RDMA_CM_EVENT_ESTABLISHED:
2749 conn = (kib_conn_t *)cmid->context;
2750 switch (conn->ibc_state) {
2754 case IBLND_CONN_PASSIVE_WAIT:
2755 CDEBUG(D_NET, "ESTABLISHED (passive): %s\n",
2756 libcfs_nid2str(conn->ibc_peer->ibp_nid));
2757 kiblnd_connreq_done(conn, 0);
2760 case IBLND_CONN_ACTIVE_CONNECT:
2761 CDEBUG(D_NET, "ESTABLISHED(active): %s\n",
2762 libcfs_nid2str(conn->ibc_peer->ibp_nid));
2763 kiblnd_check_connreply(conn,
2764 (void *)KIBLND_CONN_PARAM(event),
2765 KIBLND_CONN_PARAM_LEN(event));
2768 /* net keeps its ref on conn! */
2771 case RDMA_CM_EVENT_DISCONNECTED:
2772 conn = (kib_conn_t *)cmid->context;
2773 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
2774 CERROR("%s DISCONNECTED\n",
2775 libcfs_nid2str(conn->ibc_peer->ibp_nid));
2776 kiblnd_connreq_done(conn, -ECONNRESET);
2778 kiblnd_close_conn(conn, 0);
2780 kiblnd_conn_decref(conn);
2783 case RDMA_CM_EVENT_DEVICE_REMOVAL:
2784 LCONSOLE_ERROR_MSG(0x131,
2785 "Received notification of device removal\n"
2786 "Please shutdown LNET to allow this to proceed\n");
2787 /* Can't remove network from underneath LNET for now, so I have
2794 kiblnd_check_txs (kib_conn_t *conn, struct list_head *txs)
2797 struct list_head *ttmp;
2800 spin_lock(&conn->ibc_lock);
2802 list_for_each (ttmp, txs) {
2803 tx = list_entry (ttmp, kib_tx_t, tx_list);
2805 if (txs != &conn->ibc_active_txs) {
2806 LASSERT (tx->tx_queued);
2808 LASSERT (!tx->tx_queued);
2809 LASSERT (tx->tx_waiting || tx->tx_sending != 0);
2812 if (time_after_eq (jiffies, tx->tx_deadline)) {
2818 spin_unlock(&conn->ibc_lock);
2823 kiblnd_conn_timed_out (kib_conn_t *conn)
2825 return kiblnd_check_txs(conn, &conn->ibc_tx_queue) ||
2826 kiblnd_check_txs(conn, &conn->ibc_tx_noops) ||
2827 kiblnd_check_txs(conn, &conn->ibc_tx_queue_rsrvd) ||
2828 kiblnd_check_txs(conn, &conn->ibc_tx_queue_nocred) ||
2829 kiblnd_check_txs(conn, &conn->ibc_active_txs);
2833 kiblnd_check_conns (int idx)
2835 struct list_head *peers = &kiblnd_data.kib_peers[idx];
2836 struct list_head *ptmp;
2839 struct list_head *ctmp;
2840 unsigned long flags;
2843 /* NB. We expect to have a look at all the peers and not find any
2844 * rdmas to time out, so we just use a shared lock while we
2846 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2848 list_for_each (ptmp, peers) {
2849 peer = list_entry (ptmp, kib_peer_t, ibp_list);
2851 list_for_each (ctmp, &peer->ibp_conns) {
2852 conn = list_entry (ctmp, kib_conn_t, ibc_list);
2854 LASSERT (conn->ibc_state == IBLND_CONN_ESTABLISHED);
2856 /* In case we have enough credits to return via a
2857 * NOOP, but there were no non-blocking tx descs
2858 * free to do it last time... */
2859 kiblnd_check_sends(conn);
2861 if (!kiblnd_conn_timed_out(conn))
2864 /* Handle timeout by closing the whole connection. We
2865 * can only be sure RDMA activity has ceased once the
2866 * QP has been modified. */
2868 kiblnd_conn_addref(conn); /* 1 ref for me... */
2870 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2873 CERROR("Timed out RDMA with %s\n",
2874 libcfs_nid2str(peer->ibp_nid));
2876 kiblnd_close_conn(conn, -ETIMEDOUT);
2877 kiblnd_conn_decref(conn); /* ...until here */
2879 /* start again now I've dropped the lock */
2884 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2888 kiblnd_disconnect_conn (kib_conn_t *conn)
2890 LASSERT (!in_interrupt());
2891 LASSERT (current == kiblnd_data.kib_connd);
2892 LASSERT (conn->ibc_state == IBLND_CONN_CLOSING);
2894 rdma_disconnect(conn->ibc_cmid);
2895 kiblnd_finalise_conn(conn);
2897 kiblnd_peer_notify(conn->ibc_peer);
2901 kiblnd_connd (void *arg)
2904 unsigned long flags;
2910 unsigned long deadline = jiffies;
2912 cfs_daemonize ("kiblnd_connd");
2913 cfs_block_allsigs ();
2915 init_waitqueue_entry (&wait, current);
2916 kiblnd_data.kib_connd = current;
2918 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
2920 while (!kiblnd_data.kib_shutdown) {
2924 if (!list_empty (&kiblnd_data.kib_connd_zombies)) {
2925 conn = list_entry (kiblnd_data.kib_connd_zombies.next,
2926 kib_conn_t, ibc_list);
2927 list_del (&conn->ibc_list);
2929 spin_unlock_irqrestore (&kiblnd_data.kib_connd_lock, flags);
2932 kiblnd_destroy_conn(conn);
2934 spin_lock_irqsave (&kiblnd_data.kib_connd_lock, flags);
2937 if (!list_empty (&kiblnd_data.kib_connd_conns)) {
2938 conn = list_entry (kiblnd_data.kib_connd_conns.next,
2939 kib_conn_t, ibc_list);
2940 list_del (&conn->ibc_list);
2942 spin_unlock_irqrestore (&kiblnd_data.kib_connd_lock, flags);
2945 kiblnd_disconnect_conn(conn);
2946 kiblnd_conn_decref(conn);
2948 spin_lock_irqsave (&kiblnd_data.kib_connd_lock, flags);
2951 /* careful with the jiffy wrap... */
2952 timeout = (int)(deadline - jiffies);
2956 int chunk = kiblnd_data.kib_peer_hash_size;
2958 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
2961 /* Time to check for RDMA timeouts on a few more
2962 * peers: I do checks every 'p' seconds on a
2963 * proportion of the peer table and I need to check
2964 * every connection 'n' times within a timeout
2965 * interval, to ensure I detect a timeout on any
2966 * connection within (n+1)/n times the timeout
2969 if (*kiblnd_tunables.kib_timeout > n * p)
2970 chunk = (chunk * n * p) /
2971 *kiblnd_tunables.kib_timeout;
2975 for (i = 0; i < chunk; i++) {
2976 kiblnd_check_conns(peer_index);
2977 peer_index = (peer_index + 1) %
2978 kiblnd_data.kib_peer_hash_size;
2982 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
2988 /* Nothing to do for 'timeout' */
2989 set_current_state (TASK_INTERRUPTIBLE);
2990 add_wait_queue (&kiblnd_data.kib_connd_waitq, &wait);
2991 spin_unlock_irqrestore (&kiblnd_data.kib_connd_lock, flags);
2993 schedule_timeout (timeout);
2995 set_current_state (TASK_RUNNING);
2996 remove_wait_queue (&kiblnd_data.kib_connd_waitq, &wait);
2997 spin_lock_irqsave (&kiblnd_data.kib_connd_lock, flags);
3000 spin_unlock_irqrestore (&kiblnd_data.kib_connd_lock, flags);
3002 kiblnd_thread_fini();
3007 kiblnd_qp_event(struct ib_event *event, void *arg)
3009 kib_conn_t *conn = arg;
3011 switch (event->event) {
3012 case IB_EVENT_COMM_EST:
3013 CDEBUG(D_NET, "%s established\n",
3014 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3018 CERROR("%s: Async QP event type %d\n",
3019 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3025 kiblnd_complete (struct ib_wc *wc)
3027 switch (kiblnd_wreqid2type(wc->wr_id)) {
3031 case IBLND_WID_RDMA:
3032 /* We only get RDMA completion notification if it fails. All
3033 * subsequent work items, including the final SEND will fail
3034 * too. However we can't print out any more info about the
3035 * failing RDMA because 'tx' might be back on the idle list or
3036 * even reused already if we didn't manage to post all our work
3038 CDEBUG(D_NETERROR, "RDMA (tx: %p) failed: %d\n",
3039 kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3043 kiblnd_tx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3047 kiblnd_rx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status,
3054 kiblnd_cq_completion (struct ib_cq *cq, void *arg)
3056 /* NB I'm not allowed to schedule this conn once its refcount has
3057 * reached 0. Since fundamentally I'm racing with scheduler threads
3058 * consuming my CQ I could be called after all completions have
3059 * occurred. But in this case, ibc_nrx == 0 && ibc_nsends_posted == 0
3060 * and this CQ is about to be destroyed so I NOOP. */
3061 kib_conn_t *conn = (kib_conn_t *)arg;
3062 unsigned long flags;
3064 LASSERT (cq == conn->ibc_cq);
3066 spin_lock_irqsave(&kiblnd_data.kib_sched_lock, flags);
3068 conn->ibc_ready = 1;
3070 if (!conn->ibc_scheduled &&
3071 (conn->ibc_nrx > 0 ||
3072 conn->ibc_nsends_posted > 0)) {
3073 kiblnd_conn_addref(conn); /* +1 ref for sched_conns */
3074 conn->ibc_scheduled = 1;
3075 list_add_tail(&conn->ibc_sched_list,
3076 &kiblnd_data.kib_sched_conns);
3077 wake_up(&kiblnd_data.kib_sched_waitq);
3080 spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock, flags);
3084 kiblnd_cq_event(struct ib_event *event, void *arg)
3086 kib_conn_t *conn = arg;
3088 CERROR("%s: async CQ event type %d\n",
3089 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3093 kiblnd_scheduler(void *arg)
3095 long id = (long)arg;
3098 unsigned long flags;
3105 snprintf(name, sizeof(name), "kiblnd_sd_%02ld", id);
3106 cfs_daemonize(name);
3107 cfs_block_allsigs();
3109 init_waitqueue_entry(&wait, current);
3111 spin_lock_irqsave(&kiblnd_data.kib_sched_lock, flags);
3113 while (!kiblnd_data.kib_shutdown) {
3114 if (busy_loops++ >= IBLND_RESCHED) {
3115 spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock,
3121 spin_lock_irqsave(&kiblnd_data.kib_sched_lock, flags);
3126 if (!list_empty(&kiblnd_data.kib_sched_conns)) {
3127 conn = list_entry(kiblnd_data.kib_sched_conns.next,
3128 kib_conn_t, ibc_sched_list);
3129 /* take over kib_sched_conns' ref on conn... */
3130 LASSERT(conn->ibc_scheduled);
3131 list_del(&conn->ibc_sched_list);
3132 conn->ibc_ready = 0;
3134 spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock,
3137 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3139 rc = ib_req_notify_cq(conn->ibc_cq,
3143 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3148 spin_lock_irqsave(&kiblnd_data.kib_sched_lock,
3151 if (rc != 0 || conn->ibc_ready) {
3152 /* There may be another completion waiting; get
3153 * another scheduler to check while I handle
3155 kiblnd_conn_addref(conn); /* +1 ref for sched_conns */
3156 list_add_tail(&conn->ibc_sched_list,
3157 &kiblnd_data.kib_sched_conns);
3158 wake_up(&kiblnd_data.kib_sched_waitq);
3160 conn->ibc_scheduled = 0;
3164 spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock,
3167 kiblnd_complete(&wc);
3169 spin_lock_irqsave(&kiblnd_data.kib_sched_lock,
3173 kiblnd_conn_decref(conn); /* ...drop my ref from above */
3180 set_current_state(TASK_INTERRUPTIBLE);
3181 add_wait_queue_exclusive(&kiblnd_data.kib_sched_waitq, &wait);
3182 spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock, flags);
3187 remove_wait_queue(&kiblnd_data.kib_sched_waitq, &wait);
3188 set_current_state(TASK_RUNNING);
3189 spin_lock_irqsave(&kiblnd_data.kib_sched_lock, flags);
3192 spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock, flags);
3194 kiblnd_thread_fini();