X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lnet%2Fklnds%2Fo2iblnd%2Fo2iblnd_cb.c;h=eeef534b89d6971eac7d69d0bb27e0d67503fb32;hp=e4504ae858a8e4cb7747c5bcf99f2983072c305b;hb=35a1096a4741b24dbcd674bda5daf7b8a13cfe28;hpb=85ead6a9ef708f7338dcfed2b63eb3a2ba246d96 diff --git a/lnet/klnds/o2iblnd/o2iblnd_cb.c b/lnet/klnds/o2iblnd/o2iblnd_cb.c index e4504ae..eeef534 100644 --- a/lnet/klnds/o2iblnd/o2iblnd_cb.c +++ b/lnet/klnds/o2iblnd/o2iblnd_cb.c @@ -1,139 +1,97 @@ -/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*- - * vim:expandtab:shiftwidth=8:tabstop=8: +/* + * GPL HEADER START * - * Copyright (C) 2006 Cluster File Systems, Inc. - * Author: Eric Barton + * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. * - * This file is part of Lustre, http://www.lustre.org. + * This program is free software; you can redistribute it and/or modify + * it under the terms of the GNU General Public License version 2 only, + * as published by the Free Software Foundation. * - * Lustre is free software; you can redistribute it and/or - * modify it under the terms of version 2 of the GNU General Public - * License as published by the Free Software Foundation. + * This program is distributed in the hope that it will be useful, but + * WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + * General Public License version 2 for more details (a copy is included + * in the LICENSE file that accompanied this code). * - * Lustre is distributed in the hope that it will be useful, - * but WITHOUT ANY WARRANTY; without even the implied warranty of - * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the - * GNU General Public License for more details. + * You should have received a copy of the GNU General Public License + * version 2 along with this program; If not, see + * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf * - * You should have received a copy of the GNU General Public License - * along with Lustre; if not, write to the Free Software - * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. + * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, + * CA 95054 USA or visit www.sun.com if you need additional information or + * have any questions. * + * GPL HEADER END + */ +/* + * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved. + * Use is subject to license terms. + * + * Copyright (c) 2012, Intel Corporation. + */ +/* + * This file is part of Lustre, http://www.lustre.org/ + * Lustre is a trademark of Sun Microsystems, Inc. + * + * lnet/klnds/o2iblnd/o2iblnd_cb.c + * + * Author: Eric Barton */ #include "o2iblnd.h" -char * -kiblnd_msgtype2str(int type) -{ - switch (type) { - case IBLND_MSG_CONNREQ: - return "CONNREQ"; - - case IBLND_MSG_CONNACK: - return "CONNACK"; - - case IBLND_MSG_NOOP: - return "NOOP"; - - case IBLND_MSG_IMMEDIATE: - return "IMMEDIATE"; - - case IBLND_MSG_PUT_REQ: - return "PUT_REQ"; - - case IBLND_MSG_PUT_NAK: - return "PUT_NAK"; - - case IBLND_MSG_PUT_ACK: - return "PUT_ACK"; - - case IBLND_MSG_PUT_DONE: - return "PUT_DONE"; - - case IBLND_MSG_GET_REQ: - return "GET_REQ"; - - case IBLND_MSG_GET_DONE: - return "GET_DONE"; - - default: - return "???"; - } -} - void kiblnd_tx_done (lnet_ni_t *ni, kib_tx_t *tx) { - lnet_msg_t *lntmsg[2]; - kib_net_t *net = ni->ni_data; - int rc; - int i; - - LASSERT (net != NULL); - LASSERT (!in_interrupt()); - LASSERT (!tx->tx_queued); /* mustn't be queued for sending */ - LASSERT (tx->tx_sending == 0); /* mustn't be awaiting sent callback */ - LASSERT (!tx->tx_waiting); /* mustn't be awaiting peer response */ - -#if IBLND_MAP_ON_DEMAND - if (tx->tx_fmr != NULL) { - rc = ib_fmr_pool_unmap(tx->tx_fmr); - LASSERT (rc == 0); - - if (tx->tx_status != 0) { - rc = ib_flush_fmr_pool(net->ibn_fmrpool); - LASSERT (rc == 0); - } - - tx->tx_fmr = NULL; - } -#else - if (tx->tx_nfrags != 0) { - kiblnd_dma_unmap_sg(net->ibn_dev->ibd_cmid->device, - tx->tx_frags, tx->tx_nfrags, tx->tx_dmadir); - tx->tx_nfrags = 0; - } -#endif - /* tx may have up to 2 lnet msgs to finalise */ - lntmsg[0] = tx->tx_lntmsg[0]; tx->tx_lntmsg[0] = NULL; - lntmsg[1] = tx->tx_lntmsg[1]; tx->tx_lntmsg[1] = NULL; - rc = tx->tx_status; - - if (tx->tx_conn != NULL) { - LASSERT (ni == tx->tx_conn->ibc_peer->ibp_ni); - - kiblnd_conn_decref(tx->tx_conn); - tx->tx_conn = NULL; - } - - tx->tx_nwrq = 0; - tx->tx_status = 0; - - spin_lock(&net->ibn_tx_lock); + lnet_msg_t *lntmsg[2]; + kib_net_t *net = ni->ni_data; + int rc; + int i; + + LASSERT (net != NULL); + LASSERT (!in_interrupt()); + LASSERT (!tx->tx_queued); /* mustn't be queued for sending */ + LASSERT (tx->tx_sending == 0); /* mustn't be awaiting sent callback */ + LASSERT (!tx->tx_waiting); /* mustn't be awaiting peer response */ + LASSERT (tx->tx_pool != NULL); + + kiblnd_unmap_tx(ni, tx); + + /* tx may have up to 2 lnet msgs to finalise */ + lntmsg[0] = tx->tx_lntmsg[0]; tx->tx_lntmsg[0] = NULL; + lntmsg[1] = tx->tx_lntmsg[1]; tx->tx_lntmsg[1] = NULL; + rc = tx->tx_status; + + if (tx->tx_conn != NULL) { + LASSERT (ni == tx->tx_conn->ibc_peer->ibp_ni); + + kiblnd_conn_decref(tx->tx_conn); + tx->tx_conn = NULL; + } - list_add(&tx->tx_list, &net->ibn_idle_txs); + tx->tx_nwrq = 0; + tx->tx_status = 0; - spin_unlock(&net->ibn_tx_lock); + kiblnd_pool_free_node(&tx->tx_pool->tpo_pool, &tx->tx_list); - /* delay finalize until my descs have been freed */ - for (i = 0; i < 2; i++) { - if (lntmsg[i] == NULL) - continue; + /* delay finalize until my descs have been freed */ + for (i = 0; i < 2; i++) { + if (lntmsg[i] == NULL) + continue; - lnet_finalize(ni, lntmsg[i], rc); - } + lnet_finalize(ni, lntmsg[i], rc); + } } void -kiblnd_txlist_done (lnet_ni_t *ni, struct list_head *txlist, int status) +kiblnd_txlist_done (lnet_ni_t *ni, cfs_list_t *txlist, int status) { kib_tx_t *tx; - - while (!list_empty (txlist)) { - tx = list_entry (txlist->next, kib_tx_t, tx_list); - list_del (&tx->tx_list); + while (!cfs_list_empty (txlist)) { + tx = cfs_list_entry (txlist->next, kib_tx_t, tx_list); + + cfs_list_del(&tx->tx_list); /* complete now */ tx->tx_waiting = 0; tx->tx_status = status; @@ -142,29 +100,18 @@ kiblnd_txlist_done (lnet_ni_t *ni, struct list_head *txlist, int status) } kib_tx_t * -kiblnd_get_idle_tx (lnet_ni_t *ni) +kiblnd_get_idle_tx(lnet_ni_t *ni, lnet_nid_t target) { - kib_net_t *net = ni->ni_data; - kib_tx_t *tx; - - LASSERT (net != NULL); - - spin_lock(&net->ibn_tx_lock); - - if (list_empty(&net->ibn_idle_txs)) { - spin_unlock(&net->ibn_tx_lock); + kib_net_t *net = (kib_net_t *)ni->ni_data; + cfs_list_t *node; + kib_tx_t *tx; + kib_tx_poolset_t *tps; + + tps = net->ibn_tx_ps[lnet_cpt_of_nid(target)]; + node = kiblnd_pool_alloc_node(&tps->tps_poolset); + if (node == NULL) return NULL; - } - - tx = list_entry(net->ibn_idle_txs.next, kib_tx_t, tx_list); - list_del(&tx->tx_list); - - /* Allocate a new completion cookie. It might not be needed, - * but we've got a lock right now and we're unlikely to - * wrap... */ - tx->tx_cookie = kiblnd_data.kib_next_tx_cookie++; - - spin_unlock(&net->ibn_tx_lock); + tx = container_of(node, kib_tx_t, tx_list); LASSERT (tx->tx_nwrq == 0); LASSERT (!tx->tx_queued); @@ -174,46 +121,48 @@ kiblnd_get_idle_tx (lnet_ni_t *ni) LASSERT (tx->tx_conn == NULL); LASSERT (tx->tx_lntmsg[0] == NULL); LASSERT (tx->tx_lntmsg[1] == NULL); -#if IBLND_MAP_ON_DEMAND - LASSERT (tx->tx_fmr == NULL); -#else + LASSERT (tx->tx_u.pmr == NULL); LASSERT (tx->tx_nfrags == 0); -#endif return tx; } void -kiblnd_drop_rx (kib_rx_t *rx) +kiblnd_drop_rx(kib_rx_t *rx) { - kib_conn_t *conn = rx->rx_conn; - unsigned long flags; - - spin_lock_irqsave(&kiblnd_data.kib_sched_lock, flags); - LASSERT (conn->ibc_nrx > 0); - conn->ibc_nrx--; - spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock, flags); + kib_conn_t *conn = rx->rx_conn; + struct kib_sched_info *sched = conn->ibc_sched; + unsigned long flags; + + spin_lock_irqsave(&sched->ibs_lock, flags); + LASSERT(conn->ibc_nrx > 0); + conn->ibc_nrx--; + spin_unlock_irqrestore(&sched->ibs_lock, flags); - kiblnd_conn_decref(conn); + kiblnd_conn_decref(conn); } int kiblnd_post_rx (kib_rx_t *rx, int credit) { - kib_conn_t *conn = rx->rx_conn; - kib_net_t *net = conn->ibc_peer->ibp_ni->ni_data; - struct ib_recv_wr *bad_wrq; - int rc; - - LASSERT (net != NULL); - LASSERT (!in_interrupt()); - LASSERT (credit == IBLND_POSTRX_NO_CREDIT || - credit == IBLND_POSTRX_PEER_CREDIT || - credit == IBLND_POSTRX_RSRVD_CREDIT); - + kib_conn_t *conn = rx->rx_conn; + kib_net_t *net = conn->ibc_peer->ibp_ni->ni_data; + struct ib_recv_wr *bad_wrq = NULL; + struct ib_mr *mr; + int rc; + + LASSERT (net != NULL); + LASSERT (!in_interrupt()); + LASSERT (credit == IBLND_POSTRX_NO_CREDIT || + credit == IBLND_POSTRX_PEER_CREDIT || + credit == IBLND_POSTRX_RSRVD_CREDIT); + + mr = kiblnd_find_dma_mr(conn->ibc_hdev, rx->rx_msgaddr, IBLND_MSG_SIZE); + LASSERT (mr != NULL); + + rx->rx_sge.lkey = mr->lkey; + rx->rx_sge.addr = rx->rx_msgaddr; rx->rx_sge.length = IBLND_MSG_SIZE; - rx->rx_sge.lkey = net->ibn_dev->ibd_mr->lkey; - rx->rx_sge.addr = rx->rx_msgaddr; rx->rx_wrq.next = NULL; rx->rx_wrq.sg_list = &rx->rx_sge; @@ -231,13 +180,16 @@ kiblnd_post_rx (kib_rx_t *rx, int credit) rx->rx_nob = -1; /* flag posted */ rc = ib_post_recv(conn->ibc_cmid->qp, &rx->rx_wrq, &bad_wrq); + if (rc != 0) { + CERROR("Can't post rx for %s: %d, bad_wrq: %p\n", + libcfs_nid2str(conn->ibc_peer->ibp_nid), rc, bad_wrq); + rx->rx_nob = 0; + } if (conn->ibc_state < IBLND_CONN_ESTABLISHED) /* Initial post */ return rc; if (rc != 0) { - CERROR("Can't post rx for %s: %d\n", - libcfs_nid2str(conn->ibc_peer->ibp_nid), rc); kiblnd_close_conn(conn, rc); kiblnd_drop_rx(rx); /* No more posts for this rx */ return rc; @@ -246,24 +198,24 @@ kiblnd_post_rx (kib_rx_t *rx, int credit) if (credit == IBLND_POSTRX_NO_CREDIT) return 0; - spin_lock(&conn->ibc_lock); - if (credit == IBLND_POSTRX_PEER_CREDIT) - conn->ibc_outstanding_credits++; - else - conn->ibc_reserved_credits++; - spin_unlock(&conn->ibc_lock); + spin_lock(&conn->ibc_lock); + if (credit == IBLND_POSTRX_PEER_CREDIT) + conn->ibc_outstanding_credits++; + else + conn->ibc_reserved_credits++; + spin_unlock(&conn->ibc_lock); - kiblnd_check_sends(conn); - return 0; + kiblnd_check_sends(conn); + return 0; } kib_tx_t * kiblnd_find_waiting_tx_locked(kib_conn_t *conn, int txtype, __u64 cookie) { - struct list_head *tmp; + cfs_list_t *tmp; - list_for_each(tmp, &conn->ibc_active_txs) { - kib_tx_t *tx = list_entry(tmp, kib_tx_t, tx_list); + cfs_list_for_each(tmp, &conn->ibc_active_txs) { + kib_tx_t *tx = cfs_list_entry(tmp, kib_tx_t, tx_list); LASSERT (!tx->tx_queued); LASSERT (tx->tx_sending != 0 || tx->tx_waiting); @@ -285,15 +237,15 @@ kiblnd_find_waiting_tx_locked(kib_conn_t *conn, int txtype, __u64 cookie) void kiblnd_handle_completion(kib_conn_t *conn, int txtype, int status, __u64 cookie) { - kib_tx_t *tx; - lnet_ni_t *ni = conn->ibc_peer->ibp_ni; - int idle; + kib_tx_t *tx; + lnet_ni_t *ni = conn->ibc_peer->ibp_ni; + int idle; - spin_lock(&conn->ibc_lock); + spin_lock(&conn->ibc_lock); - tx = kiblnd_find_waiting_tx_locked(conn, txtype, cookie); - if (tx == NULL) { - spin_unlock(&conn->ibc_lock); + tx = kiblnd_find_waiting_tx_locked(conn, txtype, cookie); + if (tx == NULL) { + spin_unlock(&conn->ibc_lock); CWARN("Unmatched completion type %x cookie "LPX64" from %s\n", txtype, cookie, libcfs_nid2str(conn->ibc_peer->ibp_nid)); @@ -313,19 +265,19 @@ kiblnd_handle_completion(kib_conn_t *conn, int txtype, int status, __u64 cookie) idle = !tx->tx_queued && (tx->tx_sending == 0); if (idle) - list_del(&tx->tx_list); + cfs_list_del(&tx->tx_list); - spin_unlock(&conn->ibc_lock); + spin_unlock(&conn->ibc_lock); - if (idle) - kiblnd_tx_done(ni, tx); + if (idle) + kiblnd_tx_done(ni, tx); } void -kiblnd_send_completion (kib_conn_t *conn, int type, int status, __u64 cookie) +kiblnd_send_completion(kib_conn_t *conn, int type, int status, __u64 cookie) { - lnet_ni_t *ni = conn->ibc_peer->ibp_ni; - kib_tx_t *tx = kiblnd_get_idle_tx(ni); + lnet_ni_t *ni = conn->ibc_peer->ibp_ni; + kib_tx_t *tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid); if (tx == NULL) { CERROR("Can't get tx for completion %x for %s\n", @@ -355,19 +307,22 @@ kiblnd_handle_rx (kib_rx_t *rx) LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED); CDEBUG (D_NET, "Received %x[%d] from %s\n", - msg->ibm_type, credits, libcfs_nid2str(conn->ibc_peer->ibp_nid)); + msg->ibm_type, credits, + libcfs_nid2str(conn->ibc_peer->ibp_nid)); if (credits != 0) { /* Have I received credits that will let me send? */ - spin_lock(&conn->ibc_lock); + spin_lock(&conn->ibc_lock); - if (conn->ibc_credits + credits > IBLND_MSG_QUEUE_SIZE) { - rc2 = conn->ibc_credits; - spin_unlock(&conn->ibc_lock); + if (conn->ibc_credits + credits > + IBLND_MSG_QUEUE_SIZE(conn->ibc_version)) { + rc2 = conn->ibc_credits; + spin_unlock(&conn->ibc_lock); CERROR("Bad credits from %s: %d + %d > %d\n", libcfs_nid2str(conn->ibc_peer->ibp_nid), - rc2, credits, IBLND_MSG_QUEUE_SIZE); + rc2, credits, + IBLND_MSG_QUEUE_SIZE(conn->ibc_version)); kiblnd_close_conn(conn, -EPROTO); kiblnd_post_rx(rx, IBLND_POSTRX_NO_CREDIT); @@ -376,7 +331,12 @@ kiblnd_handle_rx (kib_rx_t *rx) conn->ibc_credits += credits; - spin_unlock(&conn->ibc_lock); + /* This ensures the credit taken by NOOP can be returned */ + if (msg->ibm_type == IBLND_MSG_NOOP && + !IBLND_OOB_CAPABLE(conn->ibc_version)) /* v1 only */ + conn->ibc_outstanding_credits++; + + spin_unlock(&conn->ibc_lock); kiblnd_check_sends(conn); } @@ -389,7 +349,15 @@ kiblnd_handle_rx (kib_rx_t *rx) break; case IBLND_MSG_NOOP: - post_credit = IBLND_POSTRX_PEER_CREDIT; + if (IBLND_OOB_CAPABLE(conn->ibc_version)) { + post_credit = IBLND_POSTRX_NO_CREDIT; + break; + } + + if (credits != 0) /* credit already posted */ + post_credit = IBLND_POSTRX_NO_CREDIT; + else /* a keepalive NOOP */ + post_credit = IBLND_POSTRX_PEER_CREDIT; break; case IBLND_MSG_IMMEDIATE: @@ -409,7 +377,8 @@ kiblnd_handle_rx (kib_rx_t *rx) break; case IBLND_MSG_PUT_NAK: - CWARN ("PUT_NACK from %s\n", libcfs_nid2str(conn->ibc_peer->ibp_nid)); + CWARN ("PUT_NACK from %s\n", + libcfs_nid2str(conn->ibc_peer->ibp_nid)); post_credit = IBLND_POSTRX_RSRVD_CREDIT; kiblnd_handle_completion(conn, IBLND_MSG_PUT_REQ, msg->ibm_u.completion.ibcm_status, @@ -419,12 +388,12 @@ kiblnd_handle_rx (kib_rx_t *rx) case IBLND_MSG_PUT_ACK: post_credit = IBLND_POSTRX_RSRVD_CREDIT; - spin_lock(&conn->ibc_lock); - tx = kiblnd_find_waiting_tx_locked(conn, IBLND_MSG_PUT_REQ, - msg->ibm_u.putack.ibpam_src_cookie); - if (tx != NULL) - list_del(&tx->tx_list); - spin_unlock(&conn->ibc_lock); + spin_lock(&conn->ibc_lock); + tx = kiblnd_find_waiting_tx_locked(conn, IBLND_MSG_PUT_REQ, + msg->ibm_u.putack.ibpam_src_cookie); + if (tx != NULL) + cfs_list_del(&tx->tx_list); + spin_unlock(&conn->ibc_lock); if (tx == NULL) { CERROR("Unmatched PUT_ACK from %s\n", @@ -440,7 +409,7 @@ kiblnd_handle_rx (kib_rx_t *rx) tx->tx_nwrq = 0; /* overwrite PUT_REQ */ - rc2 = kiblnd_init_rdma(ni, tx, IBLND_MSG_PUT_DONE, + rc2 = kiblnd_init_rdma(conn, tx, IBLND_MSG_PUT_DONE, kiblnd_rd_size(&msg->ibm_u.putack.ibpam_rd), &msg->ibm_u.putack.ibpam_rd, msg->ibm_u.putack.ibpam_dst_cookie); @@ -448,11 +417,11 @@ kiblnd_handle_rx (kib_rx_t *rx) CERROR("Can't setup rdma for PUT to %s: %d\n", libcfs_nid2str(conn->ibc_peer->ibp_nid), rc2); - spin_lock(&conn->ibc_lock); - tx->tx_waiting = 0; /* clear waiting and queue atomically */ - kiblnd_queue_tx_locked(tx, conn); - spin_unlock(&conn->ibc_lock); - break; + spin_lock(&conn->ibc_lock); + tx->tx_waiting = 0; /* clear waiting and queue atomically */ + kiblnd_queue_tx_locked(tx, conn); + spin_unlock(&conn->ibc_lock); + break; case IBLND_MSG_PUT_DONE: post_credit = IBLND_POSTRX_PEER_CREDIT; @@ -491,20 +460,19 @@ kiblnd_rx_complete (kib_rx_t *rx, int status, int nob) kib_conn_t *conn = rx->rx_conn; lnet_ni_t *ni = conn->ibc_peer->ibp_ni; kib_net_t *net = ni->ni_data; - unsigned long flags; int rc; int err = -EIO; LASSERT (net != NULL); LASSERT (rx->rx_nob < 0); /* was posted */ rx->rx_nob = 0; /* isn't now */ - + if (conn->ibc_state > IBLND_CONN_ESTABLISHED) goto ignore; if (status != IB_WC_SUCCESS) { - CDEBUG(D_NETERROR, "Rx from %s failed: %d\n", - libcfs_nid2str(conn->ibc_peer->ibp_nid), status); + CNETERR("Rx from %s failed: %d\n", + libcfs_nid2str(conn->ibc_peer->ibp_nid), status); goto failed; } @@ -534,16 +502,17 @@ kiblnd_rx_complete (kib_rx_t *rx, int status, int nob) /* racing with connection establishment/teardown! */ if (conn->ibc_state < IBLND_CONN_ESTABLISHED) { - write_lock_irqsave(&kiblnd_data.kib_global_lock, flags); - /* must check holding global lock to eliminate race */ - if (conn->ibc_state < IBLND_CONN_ESTABLISHED) { - list_add_tail(&rx->rx_list, &conn->ibc_early_rxs); - write_unlock_irqrestore(&kiblnd_data.kib_global_lock, - flags); - return; - } - write_unlock_irqrestore(&kiblnd_data.kib_global_lock, - flags); + rwlock_t *g_lock = &kiblnd_data.kib_global_lock; + unsigned long flags; + + write_lock_irqsave(g_lock, flags); + /* must check holding global lock to eliminate race */ + if (conn->ibc_state < IBLND_CONN_ESTABLISHED) { + cfs_list_add_tail(&rx->rx_list, &conn->ibc_early_rxs); + write_unlock_irqrestore(g_lock, flags); + return; + } + write_unlock_irqrestore(g_lock, flags); } kiblnd_handle_rx(rx); return; @@ -579,19 +548,164 @@ kiblnd_kvaddr_to_page (unsigned long vaddr) return page; } -#if !IBLND_MAP_ON_DEMAND +static int +kiblnd_fmr_map_tx(kib_net_t *net, kib_tx_t *tx, kib_rdma_desc_t *rd, int nob) +{ + kib_hca_dev_t *hdev; + __u64 *pages = tx->tx_pages; + kib_fmr_poolset_t *fps; + int npages; + int size; + int cpt; + int rc; + int i; + + LASSERT(tx->tx_pool != NULL); + LASSERT(tx->tx_pool->tpo_pool.po_owner != NULL); + + hdev = tx->tx_pool->tpo_hdev; + + for (i = 0, npages = 0; i < rd->rd_nfrags; i++) { + for (size = 0; size < rd->rd_frags[i].rf_nob; + size += hdev->ibh_page_size) { + pages[npages ++] = (rd->rd_frags[i].rf_addr & + hdev->ibh_page_mask) + size; + } + } + + cpt = tx->tx_pool->tpo_pool.po_owner->ps_cpt; + + fps = net->ibn_fmr_ps[cpt]; + rc = kiblnd_fmr_pool_map(fps, pages, npages, 0, &tx->tx_u.fmr); + if (rc != 0) { + CERROR ("Can't map %d pages: %d\n", npages, rc); + return rc; + } + + /* If rd is not tx_rd, it's going to get sent to a peer, who will need + * the rkey */ + rd->rd_key = (rd != tx->tx_rd) ? tx->tx_u.fmr.fmr_pfmr->fmr->rkey : + tx->tx_u.fmr.fmr_pfmr->fmr->lkey; + rd->rd_frags[0].rf_addr &= ~hdev->ibh_page_mask; + rd->rd_frags[0].rf_nob = nob; + rd->rd_nfrags = 1; + + return 0; +} + +static int +kiblnd_pmr_map_tx(kib_net_t *net, kib_tx_t *tx, kib_rdma_desc_t *rd, int nob) +{ + kib_hca_dev_t *hdev; + kib_pmr_poolset_t *pps; + __u64 iova; + int cpt; + int rc; + + LASSERT(tx->tx_pool != NULL); + LASSERT(tx->tx_pool->tpo_pool.po_owner != NULL); + + hdev = tx->tx_pool->tpo_hdev; + + iova = rd->rd_frags[0].rf_addr & ~hdev->ibh_page_mask; + + cpt = tx->tx_pool->tpo_pool.po_owner->ps_cpt; + + pps = net->ibn_pmr_ps[cpt]; + rc = kiblnd_pmr_pool_map(pps, hdev, rd, &iova, &tx->tx_u.pmr); + if (rc != 0) { + CERROR("Failed to create MR by phybuf: %d\n", rc); + return rc; + } + + /* If rd is not tx_rd, it's going to get sent to a peer, who will need + * the rkey */ + rd->rd_key = (rd != tx->tx_rd) ? tx->tx_u.pmr->pmr_mr->rkey : + tx->tx_u.pmr->pmr_mr->lkey; + rd->rd_nfrags = 1; + rd->rd_frags[0].rf_addr = iova; + rd->rd_frags[0].rf_nob = nob; + + return 0; +} + +void +kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx) +{ + kib_net_t *net = ni->ni_data; + + LASSERT(net != NULL); + + if (net->ibn_fmr_ps != NULL && tx->tx_u.fmr.fmr_pfmr != NULL) { + kiblnd_fmr_pool_unmap(&tx->tx_u.fmr, tx->tx_status); + tx->tx_u.fmr.fmr_pfmr = NULL; + + } else if (net->ibn_pmr_ps != NULL && tx->tx_u.pmr != NULL) { + kiblnd_pmr_pool_unmap(tx->tx_u.pmr); + tx->tx_u.pmr = NULL; + } + + if (tx->tx_nfrags != 0) { + kiblnd_dma_unmap_sg(tx->tx_pool->tpo_hdev->ibh_ibdev, + tx->tx_frags, tx->tx_nfrags, tx->tx_dmadir); + tx->tx_nfrags = 0; + } +} + int -kiblnd_setup_rd_iov(lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd, +kiblnd_map_tx(lnet_ni_t *ni, kib_tx_t *tx, + kib_rdma_desc_t *rd, int nfrags) +{ + kib_hca_dev_t *hdev = tx->tx_pool->tpo_hdev; + kib_net_t *net = ni->ni_data; + struct ib_mr *mr = NULL; + __u32 nob; + int i; + + /* If rd is not tx_rd, it's going to get sent to a peer and I'm the + * RDMA sink */ + tx->tx_dmadir = (rd != tx->tx_rd) ? DMA_FROM_DEVICE : DMA_TO_DEVICE; + tx->tx_nfrags = nfrags; + + rd->rd_nfrags = + kiblnd_dma_map_sg(hdev->ibh_ibdev, + tx->tx_frags, tx->tx_nfrags, tx->tx_dmadir); + + for (i = 0, nob = 0; i < rd->rd_nfrags; i++) { + rd->rd_frags[i].rf_nob = kiblnd_sg_dma_len( + hdev->ibh_ibdev, &tx->tx_frags[i]); + rd->rd_frags[i].rf_addr = kiblnd_sg_dma_address( + hdev->ibh_ibdev, &tx->tx_frags[i]); + nob += rd->rd_frags[i].rf_nob; + } + + /* looking for pre-mapping MR */ + mr = kiblnd_find_rd_dma_mr(hdev, rd); + if (mr != NULL) { + /* found pre-mapping MR */ + rd->rd_key = (rd != tx->tx_rd) ? mr->rkey : mr->lkey; + return 0; + } + + if (net->ibn_fmr_ps != NULL) + return kiblnd_fmr_map_tx(net, tx, rd, nob); + else if (net->ibn_pmr_ps != NULL) + return kiblnd_pmr_map_tx(net, tx, rd, nob); + + return -EINVAL; +} + + +int +kiblnd_setup_rd_iov(lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd, unsigned int niov, struct iovec *iov, int offset, int nob) - { + kib_net_t *net = ni->ni_data; + struct page *page; struct scatterlist *sg; - int i; - int fragnob; unsigned long vaddr; - struct page *page; + int fragnob; int page_offset; - kib_net_t *net = ni->ni_data; LASSERT (nob > 0); LASSERT (niov > 0); @@ -619,9 +733,7 @@ kiblnd_setup_rd_iov(lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd, fragnob = min((int)(iov->iov_len - offset), nob); fragnob = min(fragnob, (int)PAGE_SIZE - page_offset); - sg->page = page; - sg->offset = page_offset; - sg->length = fragnob; + sg_set_page(sg, page, fragnob, page_offset); sg++; if (offset + fragnob < iov->iov_len) { @@ -633,36 +745,17 @@ kiblnd_setup_rd_iov(lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd, } nob -= fragnob; } while (nob > 0); - - /* If rd is not tx_rd, it's going to get sent to a peer and I'm the - * RDMA sink */ - tx->tx_nfrags = sg - tx->tx_frags; - tx->tx_dmadir = (rd != tx->tx_rd) ? DMA_FROM_DEVICE : DMA_TO_DEVICE; - - rd->rd_nfrags = kiblnd_dma_map_sg(net->ibn_dev->ibd_cmid->device, - tx->tx_frags, tx->tx_nfrags, - tx->tx_dmadir); - rd->rd_key = (rd != tx->tx_rd) ? - net->ibn_dev->ibd_mr->rkey : net->ibn_dev->ibd_mr->lkey; - for (i = 0; i < rd->rd_nfrags; i++) { - rd->rd_frags[i].rf_nob = kiblnd_sg_dma_len( - net->ibn_dev->ibd_cmid->device, &tx->tx_frags[i]); - rd->rd_frags[i].rf_addr = kiblnd_sg_dma_address( - net->ibn_dev->ibd_cmid->device, &tx->tx_frags[i]); - } - - return 0; + return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags); } int -kiblnd_setup_rd_kiov (lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd, +kiblnd_setup_rd_kiov (lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd, int nkiov, lnet_kiov_t *kiov, int offset, int nob) { + kib_net_t *net = ni->ni_data; struct scatterlist *sg; - int i; int fragnob; - kib_net_t *net = ni->ni_data; CDEBUG(D_NET, "niov %d offset %d nob %d\n", nkiov, offset, nob); @@ -683,187 +776,157 @@ kiblnd_setup_rd_kiov (lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd, fragnob = min((int)(kiov->kiov_len - offset), nob); - memset(sg, 0, sizeof(*sg)); - sg->page = kiov->kiov_page; - sg->offset = kiov->kiov_offset + offset; - sg->length = fragnob; + sg_set_page(sg, kiov->kiov_page, fragnob, + kiov->kiov_offset + offset); sg++; - + offset = 0; kiov++; nkiov--; nob -= fragnob; } while (nob > 0); - /* If rd is not tx_rd, it's going to get sent to a peer and I'm the - * RDMA sink */ - tx->tx_nfrags = sg - tx->tx_frags; - tx->tx_dmadir = (rd != tx->tx_rd) ? DMA_FROM_DEVICE : DMA_TO_DEVICE; - - rd->rd_nfrags = kiblnd_dma_map_sg(net->ibn_dev->ibd_cmid->device, - tx->tx_frags, tx->tx_nfrags, tx->tx_dmadir); - rd->rd_key = (rd != tx->tx_rd) ? - net->ibn_dev->ibd_mr->rkey : net->ibn_dev->ibd_mr->lkey; - - for (i = 0; i < tx->tx_nfrags; i++) { - rd->rd_frags[i].rf_nob = kiblnd_sg_dma_len( - net->ibn_dev->ibd_cmid->device, &tx->tx_frags[i]); - rd->rd_frags[i].rf_addr = kiblnd_sg_dma_address( - net->ibn_dev->ibd_cmid->device, &tx->tx_frags[i]); -#if 0 - CDEBUG(D_WARNING,"frag[%d]: "LPX64" for %d\n", - i, rd->rd_frags[i].rf_addr, rd->rd_frags[i].rf_nob); -#endif - } - - return 0; + return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags); } -#else + int -kiblnd_map_tx (lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd, - int npages, unsigned long page_offset, int nob) +kiblnd_post_tx_locked (kib_conn_t *conn, kib_tx_t *tx, int credit) { - struct ib_pool_fmr *fmr; - kib_net_t *net = ni->ni_data; + kib_msg_t *msg = tx->tx_msg; + kib_peer_t *peer = conn->ibc_peer; + int ver = conn->ibc_version; + int rc; + int done; + struct ib_send_wr *bad_wrq; - LASSERT (net != NULL); - LASSERT (tx->tx_fmr == NULL); - LASSERT (page_offset < PAGE_SIZE); - LASSERT (npages >= (1 + ((page_offset + nob - 1)>>PAGE_SHIFT))); - LASSERT (npages <= LNET_MAX_IOV); + LASSERT (tx->tx_queued); + /* We rely on this for QP sizing */ + LASSERT (tx->tx_nwrq > 0); + LASSERT (tx->tx_nwrq <= 1 + IBLND_RDMA_FRAGS(ver)); - rd->rd_addr = 0; + LASSERT (credit == 0 || credit == 1); + LASSERT (conn->ibc_outstanding_credits >= 0); + LASSERT (conn->ibc_outstanding_credits <= IBLND_MSG_QUEUE_SIZE(ver)); + LASSERT (conn->ibc_credits >= 0); + LASSERT (conn->ibc_credits <= IBLND_MSG_QUEUE_SIZE(ver)); - fmr = ib_fmr_pool_map_phys(net->ibn_fmrpool, tx->tx_pages, - npages, rd->rd_addr); - if (IS_ERR(fmr)) { - CERROR ("Can't map %d pages: %ld\n", npages, PTR_ERR(fmr)); - return PTR_ERR(fmr); + if (conn->ibc_nsends_posted == IBLND_CONCURRENT_SENDS(ver)) { + /* tx completions outstanding... */ + CDEBUG(D_NET, "%s: posted enough\n", + libcfs_nid2str(peer->ibp_nid)); + return -EAGAIN; } - /* If rd is not tx_rd, it's going to get sent to a peer, who will need - * the rkey */ - - rd->rd_key = (rd != tx->tx_rd) ? fmr->fmr->rkey : fmr->fmr->lkey; - rd->rd_nob = nob; - - tx->tx_fmr = fmr; - return 0; -} - -int -kiblnd_setup_rd_iov (lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd, - unsigned int niov, struct iovec *iov, int offset, int nob) - -{ - int resid; - int fragnob; - struct page *page; - int npages; - unsigned long page_offset; - unsigned long vaddr; - - LASSERT (nob > 0); - LASSERT (niov > 0); - - while (offset >= iov->iov_len) { - offset -= iov->iov_len; - niov--; - iov++; - LASSERT (niov > 0); + if (credit != 0 && conn->ibc_credits == 0) { /* no credits */ + CDEBUG(D_NET, "%s: no credits\n", + libcfs_nid2str(peer->ibp_nid)); + return -EAGAIN; } - if (nob > iov->iov_len - offset) { - CERROR ("Can't map multiple vaddr fragments\n"); - return (-EMSGSIZE); + if (credit != 0 && !IBLND_OOB_CAPABLE(ver) && + conn->ibc_credits == 1 && /* last credit reserved */ + msg->ibm_type != IBLND_MSG_NOOP) { /* for NOOP */ + CDEBUG(D_NET, "%s: not using last credit\n", + libcfs_nid2str(peer->ibp_nid)); + return -EAGAIN; + } + + /* NB don't drop ibc_lock before bumping tx_sending */ + cfs_list_del(&tx->tx_list); + tx->tx_queued = 0; + + if (msg->ibm_type == IBLND_MSG_NOOP && + (!kiblnd_need_noop(conn) || /* redundant NOOP */ + (IBLND_OOB_CAPABLE(ver) && /* posted enough NOOP */ + conn->ibc_noops_posted == IBLND_OOB_MSGS(ver)))) { + /* OK to drop when posted enough NOOPs, since + * kiblnd_check_sends will queue NOOP again when + * posted NOOPs complete */ + spin_unlock(&conn->ibc_lock); + kiblnd_tx_done(peer->ibp_ni, tx); + spin_lock(&conn->ibc_lock); + CDEBUG(D_NET, "%s(%d): redundant or enough NOOP\n", + libcfs_nid2str(peer->ibp_nid), + conn->ibc_noops_posted); + return 0; } - vaddr = ((unsigned long)iov->iov_base) + offset; - - page_offset = vaddr & (PAGE_SIZE - 1); - resid = nob; - npages = 0; - - do { - LASSERT (npages < LNET_MAX_IOV); - - page = kiblnd_kvaddr_to_page(vaddr); - if (page == NULL) { - CERROR("Can't find page for %lu\n", vaddr); - return -EFAULT; - } - - tx->tx_pages[npages++] = lnet_page2phys(page); - - fragnob = PAGE_SIZE - (vaddr & (PAGE_SIZE - 1)); - vaddr += fragnob; - resid -= fragnob; - - } while (resid > 0); - - return kiblnd_map_tx(ni, tx, rd, npages, page_offset, nob); -} + kiblnd_pack_msg(peer->ibp_ni, msg, ver, conn->ibc_outstanding_credits, + peer->ibp_nid, conn->ibc_incarnation); + + conn->ibc_credits -= credit; + conn->ibc_outstanding_credits = 0; + conn->ibc_nsends_posted++; + if (msg->ibm_type == IBLND_MSG_NOOP) + conn->ibc_noops_posted++; + + /* CAVEAT EMPTOR! This tx could be the PUT_DONE of an RDMA + * PUT. If so, it was first queued here as a PUT_REQ, sent and + * stashed on ibc_active_txs, matched by an incoming PUT_ACK, + * and then re-queued here. It's (just) possible that + * tx_sending is non-zero if we've not done the tx_complete() + * from the first send; hence the ++ rather than = below. */ + tx->tx_sending++; + cfs_list_add(&tx->tx_list, &conn->ibc_active_txs); + + /* I'm still holding ibc_lock! */ + if (conn->ibc_state != IBLND_CONN_ESTABLISHED) { + rc = -ECONNABORTED; + } else if (tx->tx_pool->tpo_pool.po_failed || + conn->ibc_hdev != tx->tx_pool->tpo_hdev) { + /* close_conn will launch failover */ + rc = -ENETDOWN; + } else { + rc = ib_post_send(conn->ibc_cmid->qp, + tx->tx_wrq, &bad_wrq); + } -int -kiblnd_setup_rd_kiov (lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd, - int nkiov, lnet_kiov_t *kiov, int offset, int nob) -{ - int resid; - int npages; - unsigned long page_offset; + conn->ibc_last_send = jiffies; - CDEBUG(D_NET, "niov %d offset %d nob %d\n", nkiov, offset, nob); + if (rc == 0) + return 0; - LASSERT (nob > 0); - LASSERT (nkiov > 0); - LASSERT (nkiov <= LNET_MAX_IOV); + /* NB credits are transferred in the actual + * message, which can only be the last work item */ + conn->ibc_credits += credit; + conn->ibc_outstanding_credits += msg->ibm_credits; + conn->ibc_nsends_posted--; + if (msg->ibm_type == IBLND_MSG_NOOP) + conn->ibc_noops_posted--; - while (offset >= kiov->kiov_len) { - offset -= kiov->kiov_len; - nkiov--; - kiov++; - LASSERT (nkiov > 0); - } + tx->tx_status = rc; + tx->tx_waiting = 0; + tx->tx_sending--; - page_offset = kiov->kiov_offset + offset; + done = (tx->tx_sending == 0); + if (done) + cfs_list_del(&tx->tx_list); - resid = offset + nob; - npages = 0; + spin_unlock(&conn->ibc_lock); - do { - LASSERT (npages < LNET_MAX_IOV); - LASSERT (nkiov > 0); + if (conn->ibc_state == IBLND_CONN_ESTABLISHED) + CERROR("Error %d posting transmit to %s\n", + rc, libcfs_nid2str(peer->ibp_nid)); + else + CDEBUG(D_NET, "Error %d posting transmit to %s\n", + rc, libcfs_nid2str(peer->ibp_nid)); - if ((npages > 0 && kiov->kiov_offset != 0) || - (resid > kiov->kiov_len && - (kiov->kiov_offset + kiov->kiov_len) != PAGE_SIZE)) { - /* Can't have gaps */ - CERROR ("Can't make payload contiguous in I/O VM:" - "page %d, offset %d, len %d \n", - npages, kiov->kiov_offset, kiov->kiov_len); + kiblnd_close_conn(conn, rc); - return -EINVAL; - } + if (done) + kiblnd_tx_done(peer->ibp_ni, tx); - tx->tx_pages[npages++] = lnet_page2phys(kiov->kiov_page); - resid -= kiov->kiov_len; - kiov++; - nkiov--; - } while (resid > 0); + spin_lock(&conn->ibc_lock); - return kiblnd_map_tx(ni, tx, rd, npages, page_offset, nob); + return -EIO; } -#endif void kiblnd_check_sends (kib_conn_t *conn) { - kib_tx_t *tx; - lnet_ni_t *ni = conn->ibc_peer->ibp_ni; - int rc; - int consume_cred = 0; - struct ib_send_wr *bad_wrq; - int done; + int ver = conn->ibc_version; + lnet_ni_t *ni = conn->ibc_peer->ibp_ni; + kib_tx_t *tx; /* Don't send anything until after the connection is established */ if (conn->ibc_state < IBLND_CONN_ESTABLISHED) { @@ -872,190 +935,62 @@ kiblnd_check_sends (kib_conn_t *conn) return; } - spin_lock(&conn->ibc_lock); + spin_lock(&conn->ibc_lock); - LASSERT (conn->ibc_nsends_posted <= - *kiblnd_tunables.kib_concurrent_sends); + LASSERT (conn->ibc_nsends_posted <= IBLND_CONCURRENT_SENDS(ver)); + LASSERT (!IBLND_OOB_CAPABLE(ver) || + conn->ibc_noops_posted <= IBLND_OOB_MSGS(ver)); LASSERT (conn->ibc_reserved_credits >= 0); while (conn->ibc_reserved_credits > 0 && - !list_empty(&conn->ibc_tx_queue_rsrvd)) { - tx = list_entry(conn->ibc_tx_queue_rsrvd.next, - kib_tx_t, tx_list); - list_del(&tx->tx_list); - list_add_tail(&tx->tx_list, &conn->ibc_tx_queue); + !cfs_list_empty(&conn->ibc_tx_queue_rsrvd)) { + tx = cfs_list_entry(conn->ibc_tx_queue_rsrvd.next, + kib_tx_t, tx_list); + cfs_list_del(&tx->tx_list); + cfs_list_add_tail(&tx->tx_list, &conn->ibc_tx_queue); conn->ibc_reserved_credits--; } - if (list_empty(&conn->ibc_tx_queue) && - list_empty(&conn->ibc_tx_queue_nocred) && - (conn->ibc_outstanding_credits >= IBLND_CREDIT_HIGHWATER || - kiblnd_send_keepalive(conn))) { - spin_unlock(&conn->ibc_lock); + if (kiblnd_need_noop(conn)) { + spin_unlock(&conn->ibc_lock); - tx = kiblnd_get_idle_tx(ni); - if (tx != NULL) - kiblnd_init_tx_msg(ni, tx, IBLND_MSG_NOOP, 0); - - spin_lock(&conn->ibc_lock); + tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid); + if (tx != NULL) + kiblnd_init_tx_msg(ni, tx, IBLND_MSG_NOOP, 0); + spin_lock(&conn->ibc_lock); if (tx != NULL) kiblnd_queue_tx_locked(tx, conn); } + kiblnd_conn_addref(conn); /* 1 ref for me.... (see b21911) */ + for (;;) { - if (!list_empty (&conn->ibc_tx_queue_nocred)) { - tx = list_entry (conn->ibc_tx_queue_nocred.next, - kib_tx_t, tx_list); - consume_cred = 0; - } else if (!list_empty (&conn->ibc_tx_queue)) { - tx = list_entry (conn->ibc_tx_queue.next, - kib_tx_t, tx_list); - consume_cred = 1; - } else { - /* nothing to send right now */ - break; - } - - LASSERT (tx->tx_queued); - /* We rely on this for QP sizing */ - LASSERT (tx->tx_nwrq > 0 && - tx->tx_nwrq <= 1 + IBLND_MAX_RDMA_FRAGS); - - LASSERT (conn->ibc_outstanding_credits >= 0); - LASSERT (conn->ibc_outstanding_credits <= IBLND_MSG_QUEUE_SIZE); - LASSERT (conn->ibc_credits >= 0); - LASSERT (conn->ibc_credits <= IBLND_MSG_QUEUE_SIZE); - - if (conn->ibc_nsends_posted == - *kiblnd_tunables.kib_concurrent_sends) { - /* tx completions outstanding... */ - CDEBUG(D_NET, "%s: posted enough\n", - libcfs_nid2str(conn->ibc_peer->ibp_nid)); + int credit; + + if (!cfs_list_empty(&conn->ibc_tx_queue_nocred)) { + credit = 0; + tx = cfs_list_entry(conn->ibc_tx_queue_nocred.next, + kib_tx_t, tx_list); + } else if (!cfs_list_empty(&conn->ibc_tx_noops)) { + LASSERT (!IBLND_OOB_CAPABLE(ver)); + credit = 1; + tx = cfs_list_entry(conn->ibc_tx_noops.next, + kib_tx_t, tx_list); + } else if (!cfs_list_empty(&conn->ibc_tx_queue)) { + credit = 1; + tx = cfs_list_entry(conn->ibc_tx_queue.next, + kib_tx_t, tx_list); + } else break; - } - - if (consume_cred) { - if (conn->ibc_credits == 0) { /* no credits */ - CDEBUG(D_NET, "%s: no credits\n", - libcfs_nid2str(conn->ibc_peer->ibp_nid)); - break; - } - - if (conn->ibc_credits == 1 && /* last credit reserved for */ - conn->ibc_outstanding_credits == 0) { /* giving back credits */ - CDEBUG(D_NET, "%s: not using last credit\n", - libcfs_nid2str(conn->ibc_peer->ibp_nid)); - break; - } - } - - list_del (&tx->tx_list); - tx->tx_queued = 0; - - /* NB don't drop ibc_lock before bumping tx_sending */ - - if (tx->tx_msg->ibm_type == IBLND_MSG_NOOP && - (!list_empty(&conn->ibc_tx_queue) || - !list_empty(&conn->ibc_tx_queue_nocred) || - (conn->ibc_outstanding_credits < IBLND_CREDIT_HIGHWATER && - !kiblnd_send_keepalive(conn)))) { - /* redundant NOOP */ - spin_unlock(&conn->ibc_lock); - kiblnd_tx_done(ni, tx); - spin_lock(&conn->ibc_lock); - CDEBUG(D_NET, "%s: redundant noop\n", - libcfs_nid2str(conn->ibc_peer->ibp_nid)); - continue; - } - - kiblnd_pack_msg(ni, tx->tx_msg, conn->ibc_outstanding_credits, - conn->ibc_peer->ibp_nid, conn->ibc_incarnation); - - conn->ibc_outstanding_credits = 0; - conn->ibc_nsends_posted++; - if (consume_cred) - conn->ibc_credits--; - - /* CAVEAT EMPTOR! This tx could be the PUT_DONE of an RDMA - * PUT. If so, it was first queued here as a PUT_REQ, sent and - * stashed on ibc_active_txs, matched by an incoming PUT_ACK, - * and then re-queued here. It's (just) possible that - * tx_sending is non-zero if we've not done the tx_complete() from - * the first send; hence the ++ rather than = below. */ - tx->tx_sending++; - - list_add (&tx->tx_list, &conn->ibc_active_txs); -#if 0 - { - int i; - - for (i = 0; i < tx->tx_nwrq - 1; i++) { - LASSERT (tx->tx_wrq[i].opcode == IB_WR_RDMA_WRITE); - LASSERT (tx->tx_wrq[i].next == &tx->tx_wrq[i+1]); - LASSERT (tx->tx_wrq[i].sg_list == &tx->tx_sge[i]); - - CDEBUG(D_WARNING, "WORK[%d]: RDMA "LPX64 - " for %d k %x -> "LPX64" k %x\n", i, - tx->tx_wrq[i].sg_list->addr, - tx->tx_wrq[i].sg_list->length, - tx->tx_wrq[i].sg_list->lkey, - tx->tx_wrq[i].wr.rdma.remote_addr, - tx->tx_wrq[i].wr.rdma.rkey); - } - - LASSERT (tx->tx_wrq[i].opcode == IB_WR_SEND); - LASSERT (tx->tx_wrq[i].next == NULL); - LASSERT (tx->tx_wrq[i].sg_list == &tx->tx_sge[i]); - - CDEBUG(D_WARNING, "WORK[%d]: SEND "LPX64" for %d k %x\n", i, - tx->tx_wrq[i].sg_list->addr, - tx->tx_wrq[i].sg_list->length, - tx->tx_wrq[i].sg_list->lkey); - } -#endif - /* I'm still holding ibc_lock! */ - if (conn->ibc_state != IBLND_CONN_ESTABLISHED) - rc = -ECONNABORTED; - else - rc = ib_post_send(conn->ibc_cmid->qp, tx->tx_wrq, &bad_wrq); - - conn->ibc_last_send = jiffies; - - if (rc != 0) { - /* NB credits are transferred in the actual - * message, which can only be the last work item */ - conn->ibc_outstanding_credits += tx->tx_msg->ibm_credits; - if (consume_cred) - conn->ibc_credits++; - conn->ibc_nsends_posted--; - - tx->tx_status = rc; - tx->tx_waiting = 0; - tx->tx_sending--; - - done = (tx->tx_sending == 0); - if (done) - list_del (&tx->tx_list); - - spin_unlock(&conn->ibc_lock); - if (conn->ibc_state == IBLND_CONN_ESTABLISHED) - CERROR("Error %d posting transmit to %s\n", - rc, libcfs_nid2str(conn->ibc_peer->ibp_nid)); - else - CDEBUG(D_NET, "Error %d posting transmit to %s\n", - rc, libcfs_nid2str(conn->ibc_peer->ibp_nid)); - - kiblnd_close_conn(conn, rc); - - if (done) - kiblnd_tx_done(ni, tx); - return; - } + if (kiblnd_post_tx_locked(conn, tx, credit) != 0) + break; } - spin_unlock(&conn->ibc_lock); + spin_unlock(&conn->ibc_lock); + + kiblnd_conn_decref(conn); /* ...until here */ } void @@ -1069,24 +1004,26 @@ kiblnd_tx_complete (kib_tx_t *tx, int status) if (failed) { if (conn->ibc_state == IBLND_CONN_ESTABLISHED) - CDEBUG(D_NETERROR, "Tx -> %s cookie "LPX64 - "sending %d waiting %d: failed %d\n", - libcfs_nid2str(conn->ibc_peer->ibp_nid), - tx->tx_cookie, tx->tx_sending, tx->tx_waiting, - status); + CNETERR("Tx -> %s cookie "LPX64 + " sending %d waiting %d: failed %d\n", + libcfs_nid2str(conn->ibc_peer->ibp_nid), + tx->tx_cookie, tx->tx_sending, tx->tx_waiting, + status); kiblnd_close_conn(conn, -EIO); } else { kiblnd_peer_alive(conn->ibc_peer); } - spin_lock(&conn->ibc_lock); + spin_lock(&conn->ibc_lock); /* I could be racing with rdma completion. Whoever makes 'tx' idle * gets to free it, which also drops its ref on 'conn'. */ tx->tx_sending--; conn->ibc_nsends_posted--; + if (tx->tx_msg->ibm_type == IBLND_MSG_NOOP) + conn->ibc_noops_posted--; if (failed) { tx->tx_waiting = 0; /* don't wait for peer */ @@ -1097,11 +1034,11 @@ kiblnd_tx_complete (kib_tx_t *tx, int status) !tx->tx_waiting && /* Not waiting for peer */ !tx->tx_queued; /* Not re-queued (PUT_DONE) */ if (idle) - list_del(&tx->tx_list); + cfs_list_del(&tx->tx_list); kiblnd_conn_addref(conn); /* 1 ref for me.... */ - spin_unlock(&conn->ibc_lock); + spin_unlock(&conn->ibc_lock); if (idle) kiblnd_tx_done(conn->ibc_peer->ibp_ni, tx); @@ -1114,20 +1051,23 @@ kiblnd_tx_complete (kib_tx_t *tx, int status) void kiblnd_init_tx_msg (lnet_ni_t *ni, kib_tx_t *tx, int type, int body_nob) { - kib_net_t *net = ni->ni_data; + kib_hca_dev_t *hdev = tx->tx_pool->tpo_hdev; struct ib_sge *sge = &tx->tx_sge[tx->tx_nwrq]; struct ib_send_wr *wrq = &tx->tx_wrq[tx->tx_nwrq]; int nob = offsetof (kib_msg_t, ibm_u) + body_nob; + struct ib_mr *mr; - LASSERT (net != NULL); LASSERT (tx->tx_nwrq >= 0); LASSERT (tx->tx_nwrq < IBLND_MAX_RDMA_FRAGS + 1); LASSERT (nob <= IBLND_MSG_SIZE); kiblnd_init_msg(tx->tx_msg, type, body_nob); - sge->addr = tx->tx_msgaddr; - sge->lkey = net->ibn_dev->ibd_mr->lkey; + mr = kiblnd_find_dma_mr(hdev, tx->tx_msgaddr, nob); + LASSERT (mr != NULL); + + sge->lkey = mr->lkey; + sge->addr = tx->tx_msgaddr; sge->length = nob; memset(wrq, 0, sizeof(*wrq)); @@ -1143,55 +1083,24 @@ kiblnd_init_tx_msg (lnet_ni_t *ni, kib_tx_t *tx, int type, int body_nob) } int -kiblnd_init_rdma (lnet_ni_t *ni, kib_tx_t *tx, int type, - int nob, kib_rdma_desc_t *dstrd, __u64 dstcookie) +kiblnd_init_rdma (kib_conn_t *conn, kib_tx_t *tx, int type, + int resid, kib_rdma_desc_t *dstrd, __u64 dstcookie) { - kib_msg_t *ibmsg = tx->tx_msg; - kib_rdma_desc_t *srcrd = tx->tx_rd; - struct ib_sge *sge = &tx->tx_sge[0]; - struct ib_send_wr *wrq = &tx->tx_wrq[0]; - int rc = nob; - -#if IBLND_MAP_ON_DEMAND - LASSERT (!in_interrupt()); - LASSERT (tx->tx_nwrq == 0); - LASSERT (type == IBLND_MSG_GET_DONE || - type == IBLND_MSG_PUT_DONE); - - sge->addr = srcrd->rd_addr; - sge->lkey = srcrd->rd_key; - sge->length = nob; - - wrq = &tx->tx_wrq[0]; - - wrq->next = &tx->tx_wrq[1]; - wrq->wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_RDMA); - wrq->sg_list = sge; - wrq->num_sge = 1; - wrq->opcode = IB_WR_RDMA_WRITE; - wrq->send_flags = 0; - - wrq->wr.rdma.remote_addr = dstrd->rd_addr; - wrq->wr.rdma.rkey = dstrd->rd_key; - - tx->tx_nwrq = 1; -#else - /* CAVEAT EMPTOR: this 'consumes' the frags in 'dstrd' */ - int resid = nob; - kib_rdma_frag_t *srcfrag; - int srcidx; - kib_rdma_frag_t *dstfrag; - int dstidx; - int wrknob; - - LASSERT (!in_interrupt()); - LASSERT (tx->tx_nwrq == 0); - LASSERT (type == IBLND_MSG_GET_DONE || - type == IBLND_MSG_PUT_DONE); - - srcidx = dstidx = 0; - srcfrag = &srcrd->rd_frags[0]; - dstfrag = &dstrd->rd_frags[0]; + kib_msg_t *ibmsg = tx->tx_msg; + kib_rdma_desc_t *srcrd = tx->tx_rd; + struct ib_sge *sge = &tx->tx_sge[0]; + struct ib_send_wr *wrq = &tx->tx_wrq[0]; + int rc = resid; + int srcidx; + int dstidx; + int wrknob; + + LASSERT (!in_interrupt()); + LASSERT (tx->tx_nwrq == 0); + LASSERT (type == IBLND_MSG_GET_DONE || + type == IBLND_MSG_PUT_DONE); + + srcidx = dstidx = 0; while (resid > 0) { if (srcidx >= srcrd->rd_nfrags) { @@ -1199,27 +1108,31 @@ kiblnd_init_rdma (lnet_ni_t *ni, kib_tx_t *tx, int type, rc = -EPROTO; break; } - + if (dstidx == dstrd->rd_nfrags) { CERROR("Dst buffer exhausted: %d frags\n", dstidx); rc = -EPROTO; break; } - if (tx->tx_nwrq == IBLND_MAX_RDMA_FRAGS) { - CERROR("RDMA too fragmented: %d/%d src %d/%d dst frags\n", + if (tx->tx_nwrq == IBLND_RDMA_FRAGS(conn->ibc_version)) { + CERROR("RDMA too fragmented for %s (%d): " + "%d/%d src %d/%d dst frags\n", + libcfs_nid2str(conn->ibc_peer->ibp_nid), + IBLND_RDMA_FRAGS(conn->ibc_version), srcidx, srcrd->rd_nfrags, dstidx, dstrd->rd_nfrags); rc = -EMSGSIZE; break; } - wrknob = MIN(MIN(srcfrag->rf_nob, dstfrag->rf_nob), resid); + wrknob = MIN(MIN(kiblnd_rd_frag_size(srcrd, srcidx), + kiblnd_rd_frag_size(dstrd, dstidx)), resid); sge = &tx->tx_sge[tx->tx_nwrq]; - sge->addr = srcfrag->rf_addr; + sge->addr = kiblnd_rd_frag_addr(srcrd, srcidx); + sge->lkey = kiblnd_rd_frag_key(srcrd, srcidx); sge->length = wrknob; - sge->lkey = srcrd->rd_key; wrq = &tx->tx_wrq[tx->tx_nwrq]; @@ -1230,38 +1143,26 @@ kiblnd_init_rdma (lnet_ni_t *ni, kib_tx_t *tx, int type, wrq->opcode = IB_WR_RDMA_WRITE; wrq->send_flags = 0; - wrq->wr.rdma.remote_addr = dstfrag->rf_addr; - wrq->wr.rdma.rkey = dstrd->rd_key; + wrq->wr.rdma.remote_addr = kiblnd_rd_frag_addr(dstrd, dstidx); + wrq->wr.rdma.rkey = kiblnd_rd_frag_key(dstrd, dstidx); - wrq++; - sge++; + srcidx = kiblnd_rd_consume_frag(srcrd, srcidx, wrknob); + dstidx = kiblnd_rd_consume_frag(dstrd, dstidx, wrknob); resid -= wrknob; - if (wrknob < srcfrag->rf_nob) { - srcfrag->rf_nob -= wrknob; - srcfrag->rf_addr += wrknob; - } else { - srcfrag++; - srcidx++; - } - - if (wrknob < dstfrag->rf_nob) { - dstfrag->rf_nob -= wrknob; - dstfrag->rf_addr += wrknob; - } else { - dstfrag++; - dstidx++; - } - + tx->tx_nwrq++; + wrq++; + sge++; } if (rc < 0) /* no RDMA if completing with failure */ tx->tx_nwrq = 0; -#endif + ibmsg->ibm_u.completion.ibcm_status = rc; ibmsg->ibm_u.completion.ibcm_cookie = dstcookie; - kiblnd_init_tx_msg(ni, tx, type, sizeof (kib_completion_msg_t)); + kiblnd_init_tx_msg(conn->ibc_peer->ibp_ni, tx, + type, sizeof (kib_completion_msg_t)); return rc; } @@ -1269,13 +1170,14 @@ kiblnd_init_rdma (lnet_ni_t *ni, kib_tx_t *tx, int type, void kiblnd_queue_tx_locked (kib_tx_t *tx, kib_conn_t *conn) { - struct list_head *q; + cfs_list_t *q; LASSERT (tx->tx_nwrq > 0); /* work items set up */ LASSERT (!tx->tx_queued); /* not queued for sending already */ + LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED); - tx->tx_queued = 1; - tx->tx_deadline = jiffies + (*kiblnd_tunables.kib_timeout * HZ); + tx->tx_queued = 1; + tx->tx_deadline = jiffies + (*kiblnd_tunables.kib_timeout * HZ); if (tx->tx_conn == NULL) { kiblnd_conn_addref(conn); @@ -1304,28 +1206,77 @@ kiblnd_queue_tx_locked (kib_tx_t *tx, kib_conn_t *conn) break; case IBLND_MSG_NOOP: + if (IBLND_OOB_CAPABLE(conn->ibc_version)) + q = &conn->ibc_tx_queue_nocred; + else + q = &conn->ibc_tx_noops; + break; + case IBLND_MSG_IMMEDIATE: q = &conn->ibc_tx_queue; break; } - list_add_tail(&tx->tx_list, q); + cfs_list_add_tail(&tx->tx_list, q); } void kiblnd_queue_tx (kib_tx_t *tx, kib_conn_t *conn) { - spin_lock(&conn->ibc_lock); - kiblnd_queue_tx_locked(tx, conn); - spin_unlock(&conn->ibc_lock); + spin_lock(&conn->ibc_lock); + kiblnd_queue_tx_locked(tx, conn); + spin_unlock(&conn->ibc_lock); - kiblnd_check_sends(conn); + kiblnd_check_sends(conn); +} + +static int kiblnd_resolve_addr(struct rdma_cm_id *cmid, + struct sockaddr_in *srcaddr, + struct sockaddr_in *dstaddr, + int timeout_ms) +{ + unsigned short port; + int rc; + +#ifdef HAVE_OFED_RDMA_SET_REUSEADDR + /* allow the port to be reused */ + rc = rdma_set_reuseaddr(cmid, 1); + if (rc != 0) { + CERROR("Unable to set reuse on cmid: %d\n", rc); + return rc; + } +#endif + + /* look for a free privileged port */ + for (port = PROT_SOCK-1; port > 0; port--) { + srcaddr->sin_port = htons(port); + rc = rdma_resolve_addr(cmid, + (struct sockaddr *)srcaddr, + (struct sockaddr *)dstaddr, + timeout_ms); + if (rc == 0) { + CDEBUG(D_NET, "bound to port %hu\n", port); + return 0; + } else if (rc == -EADDRINUSE || rc == -EADDRNOTAVAIL) { + CDEBUG(D_NET, "bind to port %hu failed: %d\n", + port, rc); + } else { + return rc; + } + } + + CERROR("Failed to bind to a free privileged port\n"); +#ifndef HAVE_OFED_RDMA_SET_REUSEADDR + CERROR("You may need IB verbs that supports rdma_set_reuseaddr()\n"); +#endif + return rc; } void kiblnd_connect_peer (kib_peer_t *peer) { struct rdma_cm_id *cmid; + kib_dev_t *dev; kib_net_t *net = peer->ibp_ni->ni_data; struct sockaddr_in srcaddr; struct sockaddr_in dstaddr; @@ -1334,7 +1285,9 @@ kiblnd_connect_peer (kib_peer_t *peer) LASSERT (net != NULL); LASSERT (peer->ibp_connecting > 0); - cmid = rdma_create_id(kiblnd_cm_callback, peer, RDMA_PS_TCP); + cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, peer, RDMA_PS_TCP, + IB_QPT_RC); + if (IS_ERR(cmid)) { CERROR("Can't create CMID for %s: %ld\n", libcfs_nid2str(peer->ibp_nid), PTR_ERR(cmid)); @@ -1342,9 +1295,10 @@ kiblnd_connect_peer (kib_peer_t *peer) goto failed; } + dev = net->ibn_dev; memset(&srcaddr, 0, sizeof(srcaddr)); srcaddr.sin_family = AF_INET; - srcaddr.sin_addr.s_addr = htonl(net->ibn_dev->ibd_ifip); + srcaddr.sin_addr.s_addr = htonl(dev->ibd_ifip); memset(&dstaddr, 0, sizeof(dstaddr)); dstaddr.sin_family = AF_INET; @@ -1353,17 +1307,30 @@ kiblnd_connect_peer (kib_peer_t *peer) kiblnd_peer_addref(peer); /* cmid's ref */ - rc = rdma_resolve_addr(cmid, - (struct sockaddr *)&srcaddr, - (struct sockaddr *)&dstaddr, - *kiblnd_tunables.kib_timeout * 1000); - if (rc == 0) - return; + if (*kiblnd_tunables.kib_use_priv_port) { + rc = kiblnd_resolve_addr(cmid, &srcaddr, &dstaddr, + *kiblnd_tunables.kib_timeout * 1000); + } else { + rc = rdma_resolve_addr(cmid, + (struct sockaddr *)&srcaddr, + (struct sockaddr *)&dstaddr, + *kiblnd_tunables.kib_timeout * 1000); + } + if (rc != 0) { + /* Can't initiate address resolution: */ + CERROR("Can't resolve addr for %s: %d\n", + libcfs_nid2str(peer->ibp_nid), rc); + goto failed2; + } - /* Can't initiate address resolution: */ - CERROR("Can't resolve addr for %s: %d\n", - libcfs_nid2str(peer->ibp_nid), rc); + LASSERT (cmid->device != NULL); + CDEBUG(D_NET, "%s: connection bound to %s:%u.%u.%u.%u:%s\n", + libcfs_nid2str(peer->ibp_nid), dev->ibd_ifname, + HIPQUAD(dev->ibd_ifip), cmid->device->name); + return; + + failed2: kiblnd_peer_decref(peer); /* cmid's ref */ rdma_destroy_id(cmid); failed: @@ -1376,86 +1343,95 @@ kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid) kib_peer_t *peer; kib_peer_t *peer2; kib_conn_t *conn; - rwlock_t *g_lock = &kiblnd_data.kib_global_lock; + rwlock_t *g_lock = &kiblnd_data.kib_global_lock; unsigned long flags; int rc; /* If I get here, I've committed to send, so I complete the tx with * failure on any problems */ - LASSERT (tx->tx_conn == NULL); /* only set when assigned a conn */ - LASSERT (tx->tx_nwrq > 0); /* work items have been set up */ + LASSERT (tx == NULL || tx->tx_conn == NULL); /* only set when assigned a conn */ + LASSERT (tx == NULL || tx->tx_nwrq > 0); /* work items have been set up */ /* First time, just use a read lock since I expect to find my peer * connected */ - read_lock_irqsave(g_lock, flags); + read_lock_irqsave(g_lock, flags); peer = kiblnd_find_peer_locked(nid); - if (peer != NULL && !list_empty(&peer->ibp_conns)) { + if (peer != NULL && !cfs_list_empty(&peer->ibp_conns)) { /* Found a peer with an established connection */ conn = kiblnd_get_conn_locked(peer); kiblnd_conn_addref(conn); /* 1 ref for me... */ - read_unlock_irqrestore(g_lock, flags); + read_unlock_irqrestore(g_lock, flags); - kiblnd_queue_tx(tx, conn); + if (tx != NULL) + kiblnd_queue_tx(tx, conn); kiblnd_conn_decref(conn); /* ...to here */ return; } - read_unlock(g_lock); - /* Re-try with a write lock */ - write_lock(g_lock); + read_unlock(g_lock); + /* Re-try with a write lock */ + write_lock(g_lock); peer = kiblnd_find_peer_locked(nid); if (peer != NULL) { - if (list_empty(&peer->ibp_conns)) { + if (cfs_list_empty(&peer->ibp_conns)) { /* found a peer, but it's still connecting... */ LASSERT (peer->ibp_connecting != 0 || peer->ibp_accepting != 0); - list_add_tail (&tx->tx_list, &peer->ibp_tx_queue); - write_unlock_irqrestore(g_lock, flags); - } else { - conn = kiblnd_get_conn_locked(peer); - kiblnd_conn_addref(conn); /* 1 ref for me... */ - - write_unlock_irqrestore(g_lock, flags); - - kiblnd_queue_tx(tx, conn); + if (tx != NULL) + cfs_list_add_tail(&tx->tx_list, + &peer->ibp_tx_queue); + write_unlock_irqrestore(g_lock, flags); + } else { + conn = kiblnd_get_conn_locked(peer); + kiblnd_conn_addref(conn); /* 1 ref for me... */ + + write_unlock_irqrestore(g_lock, flags); + + if (tx != NULL) + kiblnd_queue_tx(tx, conn); kiblnd_conn_decref(conn); /* ...to here */ } return; } - write_unlock_irqrestore(g_lock, flags); + write_unlock_irqrestore(g_lock, flags); /* Allocate a peer ready to add to the peer table and retry */ rc = kiblnd_create_peer(ni, &peer, nid); if (rc != 0) { CERROR("Can't create peer %s\n", libcfs_nid2str(nid)); - tx->tx_status = -EHOSTUNREACH; - tx->tx_waiting = 0; - kiblnd_tx_done(ni, tx); + if (tx != NULL) { + tx->tx_status = -EHOSTUNREACH; + tx->tx_waiting = 0; + kiblnd_tx_done(ni, tx); + } return; } - write_lock_irqsave(g_lock, flags); + write_lock_irqsave(g_lock, flags); peer2 = kiblnd_find_peer_locked(nid); if (peer2 != NULL) { - if (list_empty(&peer2->ibp_conns)) { + if (cfs_list_empty(&peer2->ibp_conns)) { /* found a peer, but it's still connecting... */ LASSERT (peer2->ibp_connecting != 0 || peer2->ibp_accepting != 0); - list_add_tail (&tx->tx_list, &peer2->ibp_tx_queue); - write_unlock_irqrestore(g_lock, flags); - } else { - conn = kiblnd_get_conn_locked(peer2); - kiblnd_conn_addref(conn); /* 1 ref for me... */ - - write_unlock_irqrestore(g_lock, flags); - - kiblnd_queue_tx(tx, conn); + if (tx != NULL) + cfs_list_add_tail(&tx->tx_list, + &peer2->ibp_tx_queue); + write_unlock_irqrestore(g_lock, flags); + } else { + conn = kiblnd_get_conn_locked(peer2); + kiblnd_conn_addref(conn); /* 1 ref for me... */ + + write_unlock_irqrestore(g_lock, flags); + + if (tx != NULL) + kiblnd_queue_tx(tx, conn); kiblnd_conn_decref(conn); /* ...to here */ } @@ -1470,12 +1446,13 @@ kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid) /* always called with a ref on ni, which prevents ni being shutdown */ LASSERT (((kib_net_t *)ni->ni_data)->ibn_shutdown == 0); - list_add_tail(&tx->tx_list, &peer->ibp_tx_queue); + if (tx != NULL) + cfs_list_add_tail(&tx->tx_list, &peer->ibp_tx_queue); kiblnd_peer_addref(peer); - list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid)); + cfs_list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid)); - write_unlock_irqrestore(g_lock, flags); + write_unlock_irqrestore(g_lock, flags); kiblnd_connect_peer(peer); kiblnd_peer_decref(peer); @@ -1504,18 +1481,18 @@ kiblnd_send (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg) CDEBUG(D_NET, "sending %d bytes in %d frags to %s\n", payload_nob, payload_niov, libcfs_id2str(target)); - LASSERT (payload_nob == 0 || payload_niov > 0); - LASSERT (payload_niov <= LNET_MAX_IOV); + LASSERT (payload_nob == 0 || payload_niov > 0); + LASSERT (payload_niov <= LNET_MAX_IOV); - /* Thread context */ - LASSERT (!in_interrupt()); - /* payload is either all vaddrs or all pages */ - LASSERT (!(payload_kiov != NULL && payload_iov != NULL)); + /* Thread context */ + LASSERT (!in_interrupt()); + /* payload is either all vaddrs or all pages */ + LASSERT (!(payload_kiov != NULL && payload_iov != NULL)); - switch (type) { - default: - LBUG(); - return (-EIO); + switch (type) { + default: + LBUG(); + return (-EIO); case LNET_MSG_ACK: LASSERT (payload_nob == 0); @@ -1530,16 +1507,14 @@ kiblnd_send (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg) if (nob <= IBLND_MSG_SIZE) break; /* send IMMEDIATE */ - tx = kiblnd_get_idle_tx(ni); - if (tx == NULL) { - CERROR("Can allocate txd for GET to %s: \n", - libcfs_nid2str(target.nid)); - return -ENOMEM; - } - + tx = kiblnd_get_idle_tx(ni, target.nid); + if (tx == NULL) { + CERROR("Can't allocate txd for GET to %s\n", + libcfs_nid2str(target.nid)); + return -ENOMEM; + } + ibmsg = tx->tx_msg; - ibmsg->ibm_u.get.ibgm_hdr = *hdr; - ibmsg->ibm_u.get.ibgm_cookie = tx->tx_cookie; if ((lntmsg->msg_md->md_options & LNET_MD_KIOV) == 0) rc = kiblnd_setup_rd_iov(ni, tx, @@ -1559,11 +1534,11 @@ kiblnd_send (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg) kiblnd_tx_done(ni, tx); return -EIO; } -#if IBLND_MAP_ON_DEMAND - nob = sizeof(kib_get_msg_t); -#else + nob = offsetof(kib_get_msg_t, ibgm_rd.rd_frags[tx->tx_nfrags]); -#endif + ibmsg->ibm_u.get.ibgm_cookie = tx->tx_cookie; + ibmsg->ibm_u.get.ibgm_hdr = *hdr; + kiblnd_init_tx_msg(ni, tx, IBLND_MSG_GET_REQ, nob); tx->tx_lntmsg[1] = lnet_create_reply_msg(ni, lntmsg); @@ -1586,7 +1561,7 @@ kiblnd_send (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg) if (nob <= IBLND_MSG_SIZE) break; /* send IMMEDIATE */ - tx = kiblnd_get_idle_tx(ni); + tx = kiblnd_get_idle_tx(ni, target.nid); if (tx == NULL) { CERROR("Can't allocate %s txd for %s\n", type == LNET_MSG_PUT ? "PUT" : "REPLY", @@ -1625,7 +1600,7 @@ kiblnd_send (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg) LASSERT (offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob]) <= IBLND_MSG_SIZE); - tx = kiblnd_get_idle_tx(ni); + tx = kiblnd_get_idle_tx(ni, target.nid); if (tx == NULL) { CERROR ("Can't send %d to %s: tx descs exhausted\n", type, libcfs_nid2str(target.nid)); @@ -1666,7 +1641,7 @@ kiblnd_reply (lnet_ni_t *ni, kib_rx_t *rx, lnet_msg_t *lntmsg) kib_tx_t *tx; int rc; - tx = kiblnd_get_idle_tx(ni); + tx = kiblnd_get_idle_tx(ni, rx->rx_conn->ibc_peer->ibp_nid); if (tx == NULL) { CERROR("Can't get tx for REPLY to %s\n", libcfs_nid2str(target.nid)); @@ -1688,7 +1663,8 @@ kiblnd_reply (lnet_ni_t *ni, kib_rx_t *rx, lnet_msg_t *lntmsg) goto failed_1; } - rc = kiblnd_init_rdma(ni, tx, IBLND_MSG_GET_DONE, nob, + rc = kiblnd_init_rdma(rx->rx_conn, tx, + IBLND_MSG_GET_DONE, nob, &rx->rx_msg->ibm_u.get.ibgm_rd, rx->rx_msg->ibm_u.get.ibgm_cookie); if (rc < 0) { @@ -1725,18 +1701,18 @@ kiblnd_recv (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed, kib_conn_t *conn = rx->rx_conn; kib_tx_t *tx; kib_msg_t *txmsg; - int nob; - int post_credit = IBLND_POSTRX_PEER_CREDIT; - int rc = 0; + int nob; + int post_credit = IBLND_POSTRX_PEER_CREDIT; + int rc = 0; - LASSERT (mlen <= rlen); - LASSERT (!in_interrupt()); - /* Either all pages or all vaddrs */ - LASSERT (!(kiov != NULL && iov != NULL)); + LASSERT (mlen <= rlen); + LASSERT (!in_interrupt()); + /* Either all pages or all vaddrs */ + LASSERT (!(kiov != NULL && iov != NULL)); - switch (rxmsg->ibm_type) { - default: - LBUG(); + switch (rxmsg->ibm_type) { + default: + LBUG(); case IBLND_MSG_IMMEDIATE: nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[rlen]); @@ -1769,7 +1745,7 @@ kiblnd_recv (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed, break; } - tx = kiblnd_get_idle_tx(ni); + tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid); if (tx == NULL) { CERROR("Can't allocate tx for %s\n", libcfs_nid2str(conn->ibc_peer->ibp_nid)); @@ -1784,7 +1760,7 @@ kiblnd_recv (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed, &txmsg->ibm_u.putack.ibpam_rd, niov, iov, offset, mlen); else - rc = kiblnd_setup_rd_kiov(ni, tx, + rc = kiblnd_setup_rd_kiov(ni, tx, &txmsg->ibm_u.putack.ibpam_rd, niov, kiov, offset, mlen); if (rc != 0) { @@ -1797,13 +1773,10 @@ kiblnd_recv (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed, break; } + nob = offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[tx->tx_nfrags]); txmsg->ibm_u.putack.ibpam_src_cookie = rxmsg->ibm_u.putreq.ibprm_cookie; txmsg->ibm_u.putack.ibpam_dst_cookie = tx->tx_cookie; -#if IBLND_MAP_ON_DEMAND - nob = sizeof(kib_putack_msg_t); -#else - nob = offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[tx->tx_nfrags]); -#endif + kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_ACK, nob); tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */ @@ -1832,54 +1805,52 @@ kiblnd_recv (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed, } int -kiblnd_thread_start (int (*fn)(void *arg), void *arg) +kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name) { - long pid = kernel_thread (fn, arg, 0); + struct task_struct *task = kthread_run(fn, arg, name); - if (pid < 0) - return ((int)pid); + if (IS_ERR(task)) + return PTR_ERR(task); - atomic_inc (&kiblnd_data.kib_nthreads); - return (0); + atomic_inc(&kiblnd_data.kib_nthreads); + return 0; } void kiblnd_thread_fini (void) { - atomic_dec (&kiblnd_data.kib_nthreads); + atomic_dec (&kiblnd_data.kib_nthreads); } void kiblnd_peer_alive (kib_peer_t *peer) { - /* This is racy, but everyone's only writing cfs_time_current() */ - peer->ibp_last_alive = cfs_time_current(); - mb(); + /* This is racy, but everyone's only writing cfs_time_current() */ + peer->ibp_last_alive = cfs_time_current(); + smp_mb(); } void kiblnd_peer_notify (kib_peer_t *peer) { - time_t last_alive = 0; int error = 0; + cfs_time_t last_alive = 0; unsigned long flags; - - read_lock_irqsave(&kiblnd_data.kib_global_lock, flags); - if (list_empty(&peer->ibp_conns) && + read_lock_irqsave(&kiblnd_data.kib_global_lock, flags); + + if (cfs_list_empty(&peer->ibp_conns) && peer->ibp_accepting == 0 && peer->ibp_connecting == 0 && peer->ibp_error != 0) { error = peer->ibp_error; peer->ibp_error = 0; - - last_alive = cfs_time_current_sec() - - cfs_duration_sec(cfs_time_current() - - peer->ibp_last_alive); + + last_alive = peer->ibp_last_alive; } - - read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); - + + read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); + if (error != 0) lnet_notify(peer->ibp_ni, peer->ibp_nid, 0, last_alive); @@ -1894,8 +1865,9 @@ kiblnd_close_conn_locked (kib_conn_t *conn, int error) * connection to be finished off by the connd. Otherwise the connd is * already dealing with it (either to set it up or tear it down). * Caller holds kib_global_lock exclusively in irq context */ - unsigned long flags; kib_peer_t *peer = conn->ibc_peer; + kib_dev_t *dev; + unsigned long flags; LASSERT (error != 0 || conn->ibc_state >= IBLND_CONN_ESTABLISHED); @@ -1906,25 +1878,28 @@ kiblnd_close_conn_locked (kib_conn_t *conn, int error) return; /* already being handled */ if (error == 0 && - list_empty(&conn->ibc_tx_queue) && - list_empty(&conn->ibc_tx_queue_rsrvd) && - list_empty(&conn->ibc_tx_queue_nocred) && - list_empty(&conn->ibc_active_txs)) { + cfs_list_empty(&conn->ibc_tx_noops) && + cfs_list_empty(&conn->ibc_tx_queue) && + cfs_list_empty(&conn->ibc_tx_queue_rsrvd) && + cfs_list_empty(&conn->ibc_tx_queue_nocred) && + cfs_list_empty(&conn->ibc_active_txs)) { CDEBUG(D_NET, "closing conn to %s\n", libcfs_nid2str(peer->ibp_nid)); } else { - CDEBUG(D_NETERROR, "Closing conn to %s: error %d%s%s%s%s\n", + CNETERR("Closing conn to %s: error %d%s%s%s%s%s\n", libcfs_nid2str(peer->ibp_nid), error, - list_empty(&conn->ibc_tx_queue) ? "" : "(sending)", - list_empty(&conn->ibc_tx_queue_rsrvd) ? "" : "(sending_rsrvd)", - list_empty(&conn->ibc_tx_queue_nocred) ? "" : "(sending_nocred)", - list_empty(&conn->ibc_active_txs) ? "" : "(waiting)"); + cfs_list_empty(&conn->ibc_tx_queue) ? "" : "(sending)", + cfs_list_empty(&conn->ibc_tx_noops) ? "" : "(sending_noops)", + cfs_list_empty(&conn->ibc_tx_queue_rsrvd) ? "" : "(sending_rsrvd)", + cfs_list_empty(&conn->ibc_tx_queue_nocred) ? "" : "(sending_nocred)", + cfs_list_empty(&conn->ibc_active_txs) ? "" : "(waiting)"); } - list_del (&conn->ibc_list); + dev = ((kib_net_t *)peer->ibp_ni->ni_data)->ibn_dev; + cfs_list_del(&conn->ibc_list); /* connd (see below) takes over ibc_list's ref */ - if (list_empty (&peer->ibp_conns) && /* no more conns */ + if (cfs_list_empty (&peer->ibp_conns) && /* no more conns */ kiblnd_peer_active(peer)) { /* still in peer table */ kiblnd_unlink_peer_locked(peer); @@ -1934,61 +1909,68 @@ kiblnd_close_conn_locked (kib_conn_t *conn, int error) kiblnd_set_conn_state(conn, IBLND_CONN_CLOSING); - spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); + if (error != 0 && + kiblnd_dev_can_failover(dev)) { + cfs_list_add_tail(&dev->ibd_fail_list, + &kiblnd_data.kib_failed_devs); + wake_up(&kiblnd_data.kib_failover_waitq); + } + + spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); - list_add_tail (&conn->ibc_list, &kiblnd_data.kib_connd_conns); - wake_up (&kiblnd_data.kib_connd_waitq); + cfs_list_add_tail(&conn->ibc_list, &kiblnd_data.kib_connd_conns); + wake_up(&kiblnd_data.kib_connd_waitq); - spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags); + spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags); } void -kiblnd_close_conn (kib_conn_t *conn, int error) +kiblnd_close_conn(kib_conn_t *conn, int error) { - unsigned long flags; + unsigned long flags; - write_lock_irqsave(&kiblnd_data.kib_global_lock, flags); + write_lock_irqsave(&kiblnd_data.kib_global_lock, flags); - kiblnd_close_conn_locked(conn, error); + kiblnd_close_conn_locked(conn, error); - write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); + write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); } void kiblnd_handle_early_rxs(kib_conn_t *conn) { - unsigned long flags; - kib_rx_t *rx; + unsigned long flags; + kib_rx_t *rx; - LASSERT (!in_interrupt()); - LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED); + LASSERT(!in_interrupt()); + LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED); - write_lock_irqsave(&kiblnd_data.kib_global_lock, flags); - while (!list_empty(&conn->ibc_early_rxs)) { - rx = list_entry(conn->ibc_early_rxs.next, - kib_rx_t, rx_list); - list_del(&rx->rx_list); - write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); + write_lock_irqsave(&kiblnd_data.kib_global_lock, flags); + while (!cfs_list_empty(&conn->ibc_early_rxs)) { + rx = cfs_list_entry(conn->ibc_early_rxs.next, + kib_rx_t, rx_list); + cfs_list_del(&rx->rx_list); + write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); - kiblnd_handle_rx(rx); + kiblnd_handle_rx(rx); - write_lock_irqsave(&kiblnd_data.kib_global_lock, flags); - } - write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); + write_lock_irqsave(&kiblnd_data.kib_global_lock, flags); + } + write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); } void -kiblnd_abort_txs(kib_conn_t *conn, struct list_head *txs) +kiblnd_abort_txs(kib_conn_t *conn, cfs_list_t *txs) { - LIST_HEAD (zombies); - struct list_head *tmp; - struct list_head *nxt; + CFS_LIST_HEAD (zombies); + cfs_list_t *tmp; + cfs_list_t *nxt; kib_tx_t *tx; - spin_lock(&conn->ibc_lock); + spin_lock(&conn->ibc_lock); - list_for_each_safe (tmp, nxt, txs) { - tx = list_entry (tmp, kib_tx_t, tx_list); + cfs_list_for_each_safe (tmp, nxt, txs) { + tx = cfs_list_entry (tmp, kib_tx_t, tx_list); if (txs == &conn->ibc_active_txs) { LASSERT (!tx->tx_queued); @@ -1997,77 +1979,78 @@ kiblnd_abort_txs(kib_conn_t *conn, struct list_head *txs) } else { LASSERT (tx->tx_queued); } - + tx->tx_status = -ECONNABORTED; - tx->tx_queued = 0; tx->tx_waiting = 0; - + if (tx->tx_sending == 0) { - list_del (&tx->tx_list); - list_add (&tx->tx_list, &zombies); + tx->tx_queued = 0; + cfs_list_del (&tx->tx_list); + cfs_list_add (&tx->tx_list, &zombies); } } - spin_unlock(&conn->ibc_lock); + spin_unlock(&conn->ibc_lock); - kiblnd_txlist_done(conn->ibc_peer->ibp_ni, - &zombies, -ECONNABORTED); + kiblnd_txlist_done(conn->ibc_peer->ibp_ni, &zombies, -ECONNABORTED); } void kiblnd_finalise_conn (kib_conn_t *conn) { - LASSERT (!in_interrupt()); - LASSERT (conn->ibc_state > IBLND_CONN_INIT); + LASSERT (!in_interrupt()); + LASSERT (conn->ibc_state > IBLND_CONN_INIT); - kiblnd_set_conn_state(conn, IBLND_CONN_DISCONNECTED); + kiblnd_set_conn_state(conn, IBLND_CONN_DISCONNECTED); - /* abort_receives moves QP state to IB_QPS_ERR. This is only required - * for connections that didn't get as far as being connected, because - * rdma_disconnect() does this for free. */ - kiblnd_abort_receives(conn); + /* abort_receives moves QP state to IB_QPS_ERR. This is only required + * for connections that didn't get as far as being connected, because + * rdma_disconnect() does this for free. */ + kiblnd_abort_receives(conn); - /* Complete all tx descs not waiting for sends to complete. - * NB we should be safe from RDMA now that the QP has changed state */ + /* Complete all tx descs not waiting for sends to complete. + * NB we should be safe from RDMA now that the QP has changed state */ - kiblnd_abort_txs(conn, &conn->ibc_tx_queue); - kiblnd_abort_txs(conn, &conn->ibc_tx_queue_rsrvd); - kiblnd_abort_txs(conn, &conn->ibc_tx_queue_nocred); - kiblnd_abort_txs(conn, &conn->ibc_active_txs); + kiblnd_abort_txs(conn, &conn->ibc_tx_noops); + kiblnd_abort_txs(conn, &conn->ibc_tx_queue); + kiblnd_abort_txs(conn, &conn->ibc_tx_queue_rsrvd); + kiblnd_abort_txs(conn, &conn->ibc_tx_queue_nocred); + kiblnd_abort_txs(conn, &conn->ibc_active_txs); - kiblnd_handle_early_rxs(conn); + kiblnd_handle_early_rxs(conn); } void kiblnd_peer_connect_failed (kib_peer_t *peer, int active, int error) { - LIST_HEAD (zombies); - unsigned long flags; + CFS_LIST_HEAD (zombies); + unsigned long flags; - LASSERT (error != 0); - LASSERT (!in_interrupt()); + LASSERT (error != 0); + LASSERT (!in_interrupt()); - write_lock_irqsave(&kiblnd_data.kib_global_lock, flags); + write_lock_irqsave(&kiblnd_data.kib_global_lock, flags); - if (active) { - LASSERT (peer->ibp_connecting > 0); - peer->ibp_connecting--; - } else { - LASSERT (peer->ibp_accepting > 0); - peer->ibp_accepting--; - } + if (active) { + LASSERT (peer->ibp_connecting > 0); + peer->ibp_connecting--; + } else { + LASSERT (peer->ibp_accepting > 0); + peer->ibp_accepting--; + } - if (peer->ibp_connecting != 0 || + if (peer->ibp_connecting != 0 || peer->ibp_accepting != 0) { /* another connection attempt under way... */ - write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); + write_unlock_irqrestore(&kiblnd_data.kib_global_lock, + flags); return; } - if (list_empty(&peer->ibp_conns)) { + if (cfs_list_empty(&peer->ibp_conns)) { /* Take peer's blocked transmits to complete with error */ - list_add(&zombies, &peer->ibp_tx_queue); - list_del_init(&peer->ibp_tx_queue); + cfs_list_add(&zombies, &peer->ibp_tx_queue); + cfs_list_del_init(&peer->ibp_tx_queue); if (kiblnd_peer_active(peer)) kiblnd_unlink_peer_locked(peer); @@ -2075,17 +2058,17 @@ kiblnd_peer_connect_failed (kib_peer_t *peer, int active, int error) peer->ibp_error = error; } else { /* Can't have blocked transmits if there are connections */ - LASSERT (list_empty(&peer->ibp_tx_queue)); + LASSERT (cfs_list_empty(&peer->ibp_tx_queue)); } - write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); + write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); kiblnd_peer_notify(peer); - if (list_empty (&zombies)) + if (cfs_list_empty (&zombies)) return; - CDEBUG (D_NETERROR, "Deleting messages for %s: connection failed\n", + CNETERR("Deleting messages for %s: connection failed\n", libcfs_nid2str(peer->ibp_nid)); kiblnd_txlist_done(peer->ibp_ni, &zombies, -EHOSTUNREACH); @@ -2094,36 +2077,36 @@ kiblnd_peer_connect_failed (kib_peer_t *peer, int active, int error) void kiblnd_connreq_done(kib_conn_t *conn, int status) { - struct list_head txs; - kib_peer_t *peer = conn->ibc_peer; - int active; - unsigned long flags; kib_tx_t *tx; + cfs_list_t txs; + unsigned long flags; + int active; active = (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT); - CDEBUG(D_NET,"%s: %d, %d\n", libcfs_nid2str(peer->ibp_nid), - active, status); + CDEBUG(D_NET,"%s: active(%d), version(%x), status(%d)\n", + libcfs_nid2str(peer->ibp_nid), active, + conn->ibc_version, status); - LASSERT (!in_interrupt()); - LASSERT ((conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT && - peer->ibp_connecting > 0) || - (conn->ibc_state == IBLND_CONN_PASSIVE_WAIT && - peer->ibp_accepting > 0)); + LASSERT (!in_interrupt()); + LASSERT ((conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT && + peer->ibp_connecting > 0) || + (conn->ibc_state == IBLND_CONN_PASSIVE_WAIT && + peer->ibp_accepting > 0)); LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars)); conn->ibc_connvars = NULL; if (status != 0) { /* failed to establish connection */ - kiblnd_peer_connect_failed(conn->ibc_peer, active, status); + kiblnd_peer_connect_failed(peer, active, status); kiblnd_finalise_conn(conn); return; } /* connection established */ - write_lock_irqsave(&kiblnd_data.kib_global_lock, flags); + write_lock_irqsave(&kiblnd_data.kib_global_lock, flags); conn->ibc_last_send = jiffies; kiblnd_set_conn_state(conn, IBLND_CONN_ESTABLISHED); @@ -2132,55 +2115,66 @@ kiblnd_connreq_done(kib_conn_t *conn, int status) /* Add conn to peer's list and nuke any dangling conns from a different * peer instance... */ kiblnd_conn_addref(conn); /* +1 ref for ibc_list */ - list_add(&conn->ibc_list, &peer->ibp_conns); + cfs_list_add(&conn->ibc_list, &peer->ibp_conns); if (active) peer->ibp_connecting--; else peer->ibp_accepting--; - kiblnd_close_stale_conns_locked(conn->ibc_peer, - conn->ibc_incarnation); + if (peer->ibp_version == 0) { + peer->ibp_version = conn->ibc_version; + peer->ibp_incarnation = conn->ibc_incarnation; + } + + if (peer->ibp_version != conn->ibc_version || + peer->ibp_incarnation != conn->ibc_incarnation) { + kiblnd_close_stale_conns_locked(peer, conn->ibc_version, + conn->ibc_incarnation); + peer->ibp_version = conn->ibc_version; + peer->ibp_incarnation = conn->ibc_incarnation; + } + + /* grab pending txs while I have the lock */ + cfs_list_add(&txs, &peer->ibp_tx_queue); + cfs_list_del_init(&peer->ibp_tx_queue); if (!kiblnd_peer_active(peer) || /* peer has been deleted */ conn->ibc_comms_error != 0) { /* error has happened already */ + lnet_ni_t *ni = peer->ibp_ni; /* start to shut down connection */ kiblnd_close_conn_locked(conn, -ECONNABORTED); - write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); - return; - } + write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); - /* grab pending txs while I have the lock */ - list_add(&txs, &peer->ibp_tx_queue); - list_del_init(&peer->ibp_tx_queue); + kiblnd_txlist_done(ni, &txs, -ECONNABORTED); + + return; + } - write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); + write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); - /* Schedule blocked txs */ - spin_lock (&conn->ibc_lock); - while (!list_empty (&txs)) { - tx = list_entry (txs.next, kib_tx_t, tx_list); - list_del (&tx->tx_list); + /* Schedule blocked txs */ + spin_lock(&conn->ibc_lock); + while (!cfs_list_empty(&txs)) { + tx = cfs_list_entry(txs.next, kib_tx_t, tx_list); + cfs_list_del(&tx->tx_list); - kiblnd_queue_tx_locked(tx, conn); - } - spin_unlock (&conn->ibc_lock); + kiblnd_queue_tx_locked(tx, conn); + } + spin_unlock(&conn->ibc_lock); - kiblnd_check_sends(conn); + kiblnd_check_sends(conn); - /* schedule blocked rxs */ - kiblnd_handle_early_rxs(conn); + /* schedule blocked rxs */ + kiblnd_handle_early_rxs(conn); } void -kiblnd_reject(struct rdma_cm_id *cmid, int why) +kiblnd_reject(struct rdma_cm_id *cmid, kib_rej_t *rej) { int rc; - kib_rej_t rej = {.ibr_magic = IBLND_MSG_MAGIC, - .ibr_version = IBLND_MSG_VERSION, - .ibr_why = why}; - rc = rdma_reject(cmid, &rej, sizeof(rej)); + rc = rdma_reject(cmid, rej, sizeof(*rej)); if (rc != 0) CWARN("Error %d sending reject\n", rc); @@ -2189,45 +2183,64 @@ kiblnd_reject(struct rdma_cm_id *cmid, int why) int kiblnd_passive_connect (struct rdma_cm_id *cmid, void *priv, int priv_nob) { - kib_msg_t *ackmsg; + rwlock_t *g_lock = &kiblnd_data.kib_global_lock; kib_msg_t *reqmsg = priv; - rwlock_t *g_lock = &kiblnd_data.kib_global_lock; - struct rdma_conn_param cp; - unsigned long flags; - lnet_ni_t *ni = NULL; + kib_msg_t *ackmsg; kib_dev_t *ibdev; kib_peer_t *peer; kib_peer_t *peer2; kib_conn_t *conn; + lnet_ni_t *ni = NULL; + kib_net_t *net = NULL; lnet_nid_t nid; - int rc; - int rej = IBLND_REJECT_FATAL; - - LASSERT (!in_interrupt()); - - /* cmid inherits 'context' from the corresponding listener id */ - ibdev = (kib_dev_t *)cmid->context; - LASSERT (ibdev != NULL); + struct rdma_conn_param cp; + kib_rej_t rej; + int version = IBLND_MSG_VERSION; + unsigned long flags; + int rc; + struct sockaddr_in *peer_addr; + LASSERT (!in_interrupt()); + + /* cmid inherits 'context' from the corresponding listener id */ + ibdev = (kib_dev_t *)cmid->context; + LASSERT (ibdev != NULL); + + memset(&rej, 0, sizeof(rej)); + rej.ibr_magic = IBLND_MSG_MAGIC; + rej.ibr_why = IBLND_REJECT_FATAL; + rej.ibr_cp.ibcp_max_msg_size = IBLND_MSG_SIZE; + + peer_addr = (struct sockaddr_in *)&(cmid->route.addr.dst_addr); + if (*kiblnd_tunables.kib_require_priv_port && + ntohs(peer_addr->sin_port) >= PROT_SOCK) { + __u32 ip = ntohl(peer_addr->sin_addr.s_addr); + CERROR("Peer's port (%u.%u.%u.%u:%hu) is not privileged\n", + HIPQUAD(ip), ntohs(peer_addr->sin_port)); + goto failed; + } if (priv_nob < offsetof(kib_msg_t, ibm_type)) { CERROR("Short connection request\n"); goto failed; } + /* Future protocol version compatibility support! If the + * o2iblnd-specific protocol changes, or when LNET unifies + * protocols over all LNDs, the initial connection will + * negotiate a protocol version. I trap this here to avoid + * console errors; the reject tells the peer which protocol I + * speak. */ if (reqmsg->ibm_magic == LNET_PROTO_MAGIC || - reqmsg->ibm_magic == __swab32(LNET_PROTO_MAGIC) || - (reqmsg->ibm_magic == IBLND_MSG_MAGIC && - reqmsg->ibm_version != IBLND_MSG_VERSION) || - (reqmsg->ibm_magic == __swab32(IBLND_MSG_MAGIC) && - reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION))) { - /* Future protocol version compatibility support! If the - * o2iblnd-specific protocol changes, or when LNET unifies - * protocols over all LNDs, the initial connection will - * negotiate a protocol version. I trap this here to avoid - * console errors; the reject tells the peer which protocol I - * speak. */ + reqmsg->ibm_magic == __swab32(LNET_PROTO_MAGIC)) + goto failed; + if (reqmsg->ibm_magic == IBLND_MSG_MAGIC && + reqmsg->ibm_version != IBLND_MSG_VERSION && + reqmsg->ibm_version != IBLND_MSG_VERSION_1) + goto failed; + if (reqmsg->ibm_magic == __swab32(IBLND_MSG_MAGIC) && + reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION) && + reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION_1)) goto failed; - } rc = kiblnd_unpack_msg(reqmsg, priv_nob); if (rc != 0) { @@ -2236,6 +2249,36 @@ kiblnd_passive_connect (struct rdma_cm_id *cmid, void *priv, int priv_nob) } nid = reqmsg->ibm_srcnid; + ni = lnet_net2ni(LNET_NIDNET(reqmsg->ibm_dstnid)); + + if (ni != NULL) { + net = (kib_net_t *)ni->ni_data; + rej.ibr_incarnation = net->ibn_incarnation; + } + + if (ni == NULL || /* no matching net */ + ni->ni_nid != reqmsg->ibm_dstnid || /* right NET, wrong NID! */ + net->ibn_dev != ibdev) { /* wrong device */ + CERROR("Can't accept %s on %s (%s:%d:%u.%u.%u.%u): " + "bad dst nid %s\n", libcfs_nid2str(nid), + ni == NULL ? "NA" : libcfs_nid2str(ni->ni_nid), + ibdev->ibd_ifname, ibdev->ibd_nnets, + HIPQUAD(ibdev->ibd_ifip), + libcfs_nid2str(reqmsg->ibm_dstnid)); + + goto failed; + } + + /* check time stamp as soon as possible */ + if (reqmsg->ibm_dststamp != 0 && + reqmsg->ibm_dststamp != net->ibn_incarnation) { + CWARN("Stale connection request\n"); + rej.ibr_why = IBLND_REJECT_CONN_STALE; + goto failed; + } + + /* I can accept peer's version */ + version = reqmsg->ibm_version; if (reqmsg->ibm_type != IBLND_MSG_CONNREQ) { CERROR("Unexpected connreq msg type: %x from %s\n", @@ -2243,20 +2286,31 @@ kiblnd_passive_connect (struct rdma_cm_id *cmid, void *priv, int priv_nob) goto failed; } - if (reqmsg->ibm_u.connparams.ibcp_queue_depth != IBLND_MSG_QUEUE_SIZE) { + if (reqmsg->ibm_u.connparams.ibcp_queue_depth != + IBLND_MSG_QUEUE_SIZE(version)) { CERROR("Can't accept %s: incompatible queue depth %d (%d wanted)\n", - libcfs_nid2str(nid), - reqmsg->ibm_u.connparams.ibcp_queue_depth, - IBLND_MSG_QUEUE_SIZE); + libcfs_nid2str(nid), reqmsg->ibm_u.connparams.ibcp_queue_depth, + IBLND_MSG_QUEUE_SIZE(version)); + + if (version == IBLND_MSG_VERSION) + rej.ibr_why = IBLND_REJECT_MSG_QUEUE_SIZE; + goto failed; } - if (reqmsg->ibm_u.connparams.ibcp_max_frags != IBLND_MAX_RDMA_FRAGS) { - CERROR("Can't accept %s: incompatible max_frags %d (%d wanted)\n", - libcfs_nid2str(nid), + if (reqmsg->ibm_u.connparams.ibcp_max_frags != + IBLND_RDMA_FRAGS(version)) { + CERROR("Can't accept %s(version %x): " + "incompatible max_frags %d (%d wanted)\n", + libcfs_nid2str(nid), version, reqmsg->ibm_u.connparams.ibcp_max_frags, - IBLND_MAX_RDMA_FRAGS); + IBLND_RDMA_FRAGS(version)); + + if (version == IBLND_MSG_VERSION) + rej.ibr_why = IBLND_REJECT_RDMA_FRAGS; + goto failed; + } if (reqmsg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) { @@ -2267,88 +2321,101 @@ kiblnd_passive_connect (struct rdma_cm_id *cmid, void *priv, int priv_nob) goto failed; } - ni = lnet_net2ni(LNET_NIDNET(reqmsg->ibm_dstnid)); - if (ni == NULL || /* no matching net */ - ni->ni_nid != reqmsg->ibm_dstnid || /* right NET, wrong NID! */ - ((kib_net_t*)ni->ni_data)->ibn_dev != ibdev) { /* wrong device */ - CERROR("Can't accept %s: bad dst nid %s\n", - libcfs_nid2str(nid), - libcfs_nid2str(reqmsg->ibm_dstnid)); - - goto failed; - } - /* assume 'nid' is a new peer; create */ rc = kiblnd_create_peer(ni, &peer, nid); if (rc != 0) { CERROR("Can't create peer for %s\n", libcfs_nid2str(nid)); - rej = IBLND_REJECT_NO_RESOURCES; + rej.ibr_why = IBLND_REJECT_NO_RESOURCES; goto failed; } - write_lock_irqsave(g_lock, flags); + write_lock_irqsave(g_lock, flags); peer2 = kiblnd_find_peer_locked(nid); if (peer2 != NULL) { - /* tie-break connection race in favour of the higher NID */ + if (peer2->ibp_version == 0) { + peer2->ibp_version = version; + peer2->ibp_incarnation = reqmsg->ibm_srcstamp; + } + + /* not the guy I've talked with */ + if (peer2->ibp_incarnation != reqmsg->ibm_srcstamp || + peer2->ibp_version != version) { + kiblnd_close_peer_conns_locked(peer2, -ESTALE); + write_unlock_irqrestore(g_lock, flags); + + CWARN("Conn stale %s [old ver: %x, new ver: %x]\n", + libcfs_nid2str(nid), peer2->ibp_version, version); + + kiblnd_peer_decref(peer); + rej.ibr_why = IBLND_REJECT_CONN_STALE; + goto failed; + } + + /* tie-break connection race in favour of the higher NID */ if (peer2->ibp_connecting != 0 && nid < ni->ni_nid) { - write_unlock_irqrestore(g_lock, flags); + write_unlock_irqrestore(g_lock, flags); - CWARN("Conn race %s\n", - libcfs_nid2str(peer2->ibp_nid)); + CWARN("Conn race %s\n", libcfs_nid2str(peer2->ibp_nid)); kiblnd_peer_decref(peer); - rej = IBLND_REJECT_CONN_RACE; + rej.ibr_why = IBLND_REJECT_CONN_RACE; goto failed; } peer2->ibp_accepting++; kiblnd_peer_addref(peer2); - write_unlock_irqrestore(g_lock, flags); + write_unlock_irqrestore(g_lock, flags); kiblnd_peer_decref(peer); peer = peer2; } else { /* Brand new peer */ LASSERT (peer->ibp_accepting == 0); - peer->ibp_accepting = 1; + LASSERT (peer->ibp_version == 0 && + peer->ibp_incarnation == 0); + + peer->ibp_accepting = 1; + peer->ibp_version = version; + peer->ibp_incarnation = reqmsg->ibm_srcstamp; /* I have a ref on ni that prevents it being shutdown */ - LASSERT (((kib_net_t *)ni->ni_data)->ibn_shutdown == 0); + LASSERT (net->ibn_shutdown == 0); kiblnd_peer_addref(peer); - list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid)); + cfs_list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid)); - write_unlock_irqrestore(g_lock, flags); + write_unlock_irqrestore(g_lock, flags); } - conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_PASSIVE_WAIT); + conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_PASSIVE_WAIT, version); if (conn == NULL) { kiblnd_peer_connect_failed(peer, 0, -ENOMEM); kiblnd_peer_decref(peer); - rej = IBLND_REJECT_NO_RESOURCES; + rej.ibr_why = IBLND_REJECT_NO_RESOURCES; goto failed; } /* conn now "owns" cmid, so I return success from here on to ensure the * CM callback doesn't destroy cmid. */ - conn->ibc_incarnation = reqmsg->ibm_srcstamp; - conn->ibc_credits = IBLND_MSG_QUEUE_SIZE; - conn->ibc_reserved_credits = IBLND_MSG_QUEUE_SIZE; - LASSERT (conn->ibc_credits + conn->ibc_reserved_credits - <= IBLND_RX_MSGS); + conn->ibc_incarnation = reqmsg->ibm_srcstamp; + conn->ibc_credits = IBLND_MSG_QUEUE_SIZE(version); + conn->ibc_reserved_credits = IBLND_MSG_QUEUE_SIZE(version); + LASSERT (conn->ibc_credits + conn->ibc_reserved_credits + IBLND_OOB_MSGS(version) + <= IBLND_RX_MSGS(version)); ackmsg = &conn->ibc_connvars->cv_msg; memset(ackmsg, 0, sizeof(*ackmsg)); kiblnd_init_msg(ackmsg, IBLND_MSG_CONNACK, sizeof(ackmsg->ibm_u.connparams)); - ackmsg->ibm_u.connparams.ibcp_queue_depth = IBLND_MSG_QUEUE_SIZE; - ackmsg->ibm_u.connparams.ibcp_max_frags = IBLND_MAX_RDMA_FRAGS; + ackmsg->ibm_u.connparams.ibcp_queue_depth = IBLND_MSG_QUEUE_SIZE(version); ackmsg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE; - kiblnd_pack_msg(ni, ackmsg, 0, nid, reqmsg->ibm_srcstamp); + ackmsg->ibm_u.connparams.ibcp_max_frags = IBLND_RDMA_FRAGS(version); + + kiblnd_pack_msg(ni, ackmsg, version, 0, nid, reqmsg->ibm_srcstamp); memset(&cp, 0, sizeof(cp)); cp.private_data = ackmsg; @@ -2364,7 +2431,10 @@ kiblnd_passive_connect (struct rdma_cm_id *cmid, void *priv, int priv_nob) rc = rdma_accept(cmid, &cp); if (rc != 0) { CERROR("Can't accept %s: %d\n", libcfs_nid2str(nid), rc); - kiblnd_reject(cmid, IBLND_REJECT_FATAL); + rej.ibr_version = version; + rej.ibr_why = IBLND_REJECT_FATAL; + + kiblnd_reject(cmid, &rej); kiblnd_connreq_done(conn, rc); kiblnd_conn_decref(conn); } @@ -2376,61 +2446,139 @@ kiblnd_passive_connect (struct rdma_cm_id *cmid, void *priv, int priv_nob) if (ni != NULL) lnet_ni_decref(ni); - kiblnd_reject(cmid, rej); + rej.ibr_version = version; + rej.ibr_cp.ibcp_queue_depth = IBLND_MSG_QUEUE_SIZE(version); + rej.ibr_cp.ibcp_max_frags = IBLND_RDMA_FRAGS(version); + kiblnd_reject(cmid, &rej); + return -ECONNREFUSED; } void -kiblnd_reconnect (kib_conn_t *conn, char *why) +kiblnd_reconnect (kib_conn_t *conn, int version, + __u64 incarnation, int why, kib_connparams_t *cp) { kib_peer_t *peer = conn->ibc_peer; + char *reason; int retry = 0; unsigned long flags; - + LASSERT (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT); LASSERT (peer->ibp_connecting > 0); /* 'conn' at least */ - write_lock_irqsave(&kiblnd_data.kib_global_lock, flags); + write_lock_irqsave(&kiblnd_data.kib_global_lock, flags); /* retry connection if it's still needed and no other connection - * attempts (active or passive) are in progress */ - if (!list_empty(&peer->ibp_tx_queue) && + * attempts (active or passive) are in progress + * NB: reconnect is still needed even when ibp_tx_queue is + * empty if ibp_version != version because reconnect may be + * initiated by kiblnd_query() */ + if ((!cfs_list_empty(&peer->ibp_tx_queue) || + peer->ibp_version != version) && peer->ibp_connecting == 1 && peer->ibp_accepting == 0) { retry = 1; peer->ibp_connecting++; + + peer->ibp_version = version; + peer->ibp_incarnation = incarnation; } - - write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); - if (retry) { - CDEBUG(D_NETERROR, "%s: retrying (%s)\n", - libcfs_nid2str(peer->ibp_nid), why); - kiblnd_connect_peer(peer); + write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); + + if (!retry) + return; + + switch (why) { + default: + reason = "Unknown"; + break; + + case IBLND_REJECT_CONN_STALE: + reason = "stale"; + break; + + case IBLND_REJECT_CONN_RACE: + reason = "conn race"; + break; + + case IBLND_REJECT_CONN_UNCOMPAT: + reason = "version negotiation"; + break; } + + CNETERR("%s: retrying (%s), %x, %x, " + "queue_dep: %d, max_frag: %d, msg_size: %d\n", + libcfs_nid2str(peer->ibp_nid), + reason, IBLND_MSG_VERSION, version, + cp != NULL? cp->ibcp_queue_depth :IBLND_MSG_QUEUE_SIZE(version), + cp != NULL? cp->ibcp_max_frags : IBLND_RDMA_FRAGS(version), + cp != NULL? cp->ibcp_max_msg_size: IBLND_MSG_SIZE); + + kiblnd_connect_peer(peer); } void kiblnd_rejected (kib_conn_t *conn, int reason, void *priv, int priv_nob) { - kib_peer_t *peer = conn->ibc_peer; + kib_peer_t *peer = conn->ibc_peer; - LASSERT (!in_interrupt()); - LASSERT (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT); + LASSERT (!in_interrupt()); + LASSERT (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT); + + switch (reason) { + case IB_CM_REJ_STALE_CONN: + kiblnd_reconnect(conn, IBLND_MSG_VERSION, 0, + IBLND_REJECT_CONN_STALE, NULL); + break; - switch (reason) { - case IB_CM_REJ_STALE_CONN: - kiblnd_reconnect(conn, "stale"); + case IB_CM_REJ_INVALID_SERVICE_ID: + CNETERR("%s rejected: no listener at %d\n", + libcfs_nid2str(peer->ibp_nid), + *kiblnd_tunables.kib_service); break; case IB_CM_REJ_CONSUMER_DEFINED: - if (priv_nob >= sizeof(kib_rej_t)) { - kib_rej_t *rej = priv; + if (priv_nob >= offsetof(kib_rej_t, ibr_padding)) { + kib_rej_t *rej = priv; + kib_connparams_t *cp = NULL; + int flip = 0; + __u64 incarnation = -1; + + /* NB. default incarnation is -1 because: + * a) V1 will ignore dst incarnation in connreq. + * b) V2 will provide incarnation while rejecting me, + * -1 will be overwrote. + * + * if I try to connect to a V1 peer with V2 protocol, + * it rejected me then upgrade to V2, I have no idea + * about the upgrading and try to reconnect with V1, + * in this case upgraded V2 can find out I'm trying to + * talk to the old guy and reject me(incarnation is -1). + */ if (rej->ibr_magic == __swab32(IBLND_MSG_MAGIC) || rej->ibr_magic == __swab32(LNET_PROTO_MAGIC)) { __swab32s(&rej->ibr_magic); __swab16s(&rej->ibr_version); + flip = 1; + } + + if (priv_nob >= sizeof(kib_rej_t) && + rej->ibr_version > IBLND_MSG_VERSION_1) { + /* priv_nob is always 148 in current version + * of OFED, so we still need to check version. + * (define of IB_CM_REJ_PRIVATE_DATA_SIZE) */ + cp = &rej->ibr_cp; + + if (flip) { + __swab64s(&rej->ibr_incarnation); + __swab16s(&cp->ibcp_queue_depth); + __swab16s(&cp->ibcp_max_frags); + __swab32s(&cp->ibcp_max_msg_size); + } + + incarnation = rej->ibr_incarnation; } if (rej->ibr_magic != IBLND_MSG_MAGIC && @@ -2439,27 +2587,54 @@ kiblnd_rejected (kib_conn_t *conn, int reason, void *priv, int priv_nob) libcfs_nid2str(peer->ibp_nid)); break; } - - if (rej->ibr_version != IBLND_MSG_VERSION) { - CERROR("%s rejected: o2iblnd version %d error\n", + + if (rej->ibr_version != IBLND_MSG_VERSION && + rej->ibr_version != IBLND_MSG_VERSION_1) { + CERROR("%s rejected: o2iblnd version %x error\n", libcfs_nid2str(peer->ibp_nid), rej->ibr_version); break; } - + + if (rej->ibr_why == IBLND_REJECT_FATAL && + rej->ibr_version == IBLND_MSG_VERSION_1) { + CDEBUG(D_NET, "rejected by old version peer %s: %x\n", + libcfs_nid2str(peer->ibp_nid), rej->ibr_version); + + if (conn->ibc_version != IBLND_MSG_VERSION_1) + rej->ibr_why = IBLND_REJECT_CONN_UNCOMPAT; + } + switch (rej->ibr_why) { case IBLND_REJECT_CONN_RACE: - kiblnd_reconnect(conn, "conn race"); + case IBLND_REJECT_CONN_STALE: + case IBLND_REJECT_CONN_UNCOMPAT: + kiblnd_reconnect(conn, rej->ibr_version, + incarnation, rej->ibr_why, cp); + break; + + case IBLND_REJECT_MSG_QUEUE_SIZE: + CERROR("%s rejected: incompatible message queue depth %d, %d\n", + libcfs_nid2str(peer->ibp_nid), cp->ibcp_queue_depth, + IBLND_MSG_QUEUE_SIZE(conn->ibc_version)); + break; + + case IBLND_REJECT_RDMA_FRAGS: + CERROR("%s rejected: incompatible # of RDMA fragments %d, %d\n", + libcfs_nid2str(peer->ibp_nid), cp->ibcp_max_frags, + IBLND_RDMA_FRAGS(conn->ibc_version)); break; - + case IBLND_REJECT_NO_RESOURCES: CERROR("%s rejected: o2iblnd no resources\n", libcfs_nid2str(peer->ibp_nid)); break; + case IBLND_REJECT_FATAL: CERROR("%s rejected: o2iblnd fatal error\n", libcfs_nid2str(peer->ibp_nid)); break; + default: CERROR("%s rejected: o2iblnd reason %d\n", libcfs_nid2str(peer->ibp_nid), @@ -2470,8 +2645,8 @@ kiblnd_rejected (kib_conn_t *conn, int reason, void *priv, int priv_nob) } /* fall through */ default: - CDEBUG(D_NETERROR, "%s rejected: reason %d, size %d\n", - libcfs_nid2str(peer->ibp_nid), reason, priv_nob); + CNETERR("%s rejected: reason %d, size %d\n", + libcfs_nid2str(peer->ibp_nid), reason, priv_nob); break; } @@ -2482,10 +2657,11 @@ void kiblnd_check_connreply (kib_conn_t *conn, void *priv, int priv_nob) { kib_peer_t *peer = conn->ibc_peer; - lnet_ni_t *ni = peer->ibp_ni; - kib_net_t *net = ni->ni_data; - kib_msg_t *msg = priv; - int rc = kiblnd_unpack_msg(msg, priv_nob); + lnet_ni_t *ni = peer->ibp_ni; + kib_net_t *net = ni->ni_data; + kib_msg_t *msg = priv; + int ver = conn->ibc_version; + int rc = kiblnd_unpack_msg(msg, priv_nob); unsigned long flags; LASSERT (net != NULL); @@ -2503,20 +2679,30 @@ kiblnd_check_connreply (kib_conn_t *conn, void *priv, int priv_nob) goto failed; } - if (msg->ibm_u.connparams.ibcp_queue_depth != IBLND_MSG_QUEUE_SIZE) { + if (ver != msg->ibm_version) { + CERROR("%s replied version %x is different with " + "requested version %x\n", + libcfs_nid2str(peer->ibp_nid), msg->ibm_version, ver); + rc = -EPROTO; + goto failed; + } + + if (msg->ibm_u.connparams.ibcp_queue_depth != + IBLND_MSG_QUEUE_SIZE(ver)) { CERROR("%s has incompatible queue depth %d(%d wanted)\n", libcfs_nid2str(peer->ibp_nid), msg->ibm_u.connparams.ibcp_queue_depth, - IBLND_MSG_QUEUE_SIZE); + IBLND_MSG_QUEUE_SIZE(ver)); rc = -EPROTO; goto failed; } - if (msg->ibm_u.connparams.ibcp_max_frags != IBLND_MAX_RDMA_FRAGS) { + if (msg->ibm_u.connparams.ibcp_max_frags != + IBLND_RDMA_FRAGS(ver)) { CERROR("%s has incompatible max_frags %d (%d wanted)\n", libcfs_nid2str(peer->ibp_nid), msg->ibm_u.connparams.ibcp_max_frags, - IBLND_MAX_RDMA_FRAGS); + IBLND_RDMA_FRAGS(ver)); rc = -EPROTO; goto failed; } @@ -2530,25 +2716,27 @@ kiblnd_check_connreply (kib_conn_t *conn, void *priv, int priv_nob) goto failed; } - read_lock_irqsave(&kiblnd_data.kib_global_lock, flags); - if (msg->ibm_dstnid == ni->ni_nid && - msg->ibm_dststamp == net->ibn_incarnation) - rc = 0; - else - rc = -ESTALE; - read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); + read_lock_irqsave(&kiblnd_data.kib_global_lock, flags); + if (msg->ibm_dstnid == ni->ni_nid && + msg->ibm_dststamp == net->ibn_incarnation) + rc = 0; + else + rc = -ESTALE; + read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); if (rc != 0) { - CERROR("Stale connection reply from %s\n", - libcfs_nid2str(peer->ibp_nid)); + CERROR("Bad connection reply from %s, rc = %d, " + "version: %x max_frags: %d\n", + libcfs_nid2str(peer->ibp_nid), rc, + msg->ibm_version, msg->ibm_u.connparams.ibcp_max_frags); goto failed; } - conn->ibc_incarnation = msg->ibm_srcstamp; - conn->ibc_credits = IBLND_MSG_QUEUE_SIZE; - conn->ibc_reserved_credits = IBLND_MSG_QUEUE_SIZE; - LASSERT (conn->ibc_credits + conn->ibc_reserved_credits - <= IBLND_RX_MSGS); + conn->ibc_incarnation = msg->ibm_srcstamp; + conn->ibc_credits = + conn->ibc_reserved_credits = IBLND_MSG_QUEUE_SIZE(ver); + LASSERT (conn->ibc_credits + conn->ibc_reserved_credits + IBLND_OOB_MSGS(ver) + <= IBLND_RX_MSGS(ver)); kiblnd_connreq_done(conn, 0); return; @@ -2571,9 +2759,20 @@ kiblnd_active_connect (struct rdma_cm_id *cmid) kib_conn_t *conn; kib_msg_t *msg; struct rdma_conn_param cp; + int version; + __u64 incarnation; + unsigned long flags; int rc; - conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_ACTIVE_CONNECT); + read_lock_irqsave(&kiblnd_data.kib_global_lock, flags); + + incarnation = peer->ibp_incarnation; + version = (peer->ibp_version == 0) ? IBLND_MSG_VERSION : + peer->ibp_version; + + read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); + + conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_ACTIVE_CONNECT, version); if (conn == NULL) { kiblnd_peer_connect_failed(peer, 1, -ENOMEM); kiblnd_peer_decref(peer); /* lose cmid's ref */ @@ -2588,11 +2787,13 @@ kiblnd_active_connect (struct rdma_cm_id *cmid) memset(msg, 0, sizeof(*msg)); kiblnd_init_msg(msg, IBLND_MSG_CONNREQ, sizeof(msg->ibm_u.connparams)); - msg->ibm_u.connparams.ibcp_queue_depth = IBLND_MSG_QUEUE_SIZE; - msg->ibm_u.connparams.ibcp_max_frags = IBLND_MAX_RDMA_FRAGS; + msg->ibm_u.connparams.ibcp_queue_depth = IBLND_MSG_QUEUE_SIZE(version); + msg->ibm_u.connparams.ibcp_max_frags = IBLND_RDMA_FRAGS(version); msg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE; - kiblnd_pack_msg(peer->ibp_ni, msg, 0, peer->ibp_nid, 0); - + + kiblnd_pack_msg(peer->ibp_ni, msg, version, + 0, peer->ibp_nid, incarnation); + memset(&cp, 0, sizeof(cp)); cp.private_data = msg; cp.private_data_len = msg->ibm_nob; @@ -2604,7 +2805,7 @@ kiblnd_active_connect (struct rdma_cm_id *cmid) LASSERT(cmid->context == (void *)conn); LASSERT(conn->ibc_cmid == cmid); - + rc = rdma_connect(cmid, &cp); if (rc != 0) { CERROR("Can't connect to %s: %d\n", @@ -2625,6 +2826,8 @@ kiblnd_cm_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event) switch (event->event) { default: + CERROR("Unexpected event: %d, status: %d\n", + event->event, event->status); LBUG(); case RDMA_CM_EVENT_CONNECT_REQUEST: @@ -2637,7 +2840,7 @@ kiblnd_cm_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event) case RDMA_CM_EVENT_ADDR_ERROR: peer = (kib_peer_t *)cmid->context; - CDEBUG(D_NETERROR, "%s: ADDR ERROR %d\n", + CNETERR("%s: ADDR ERROR %d\n", libcfs_nid2str(peer->ibp_nid), event->status); kiblnd_peer_connect_failed(peer, 1, -EHOSTUNREACH); kiblnd_peer_decref(peer); @@ -2650,8 +2853,8 @@ kiblnd_cm_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event) libcfs_nid2str(peer->ibp_nid), event->status); if (event->status != 0) { - CDEBUG(D_NETERROR, "Can't resolve address for %s: %d\n", - libcfs_nid2str(peer->ibp_nid), event->status); + CNETERR("Can't resolve address for %s: %d\n", + libcfs_nid2str(peer->ibp_nid), event->status); rc = event->status; } else { rc = rdma_resolve_route( @@ -2668,8 +2871,8 @@ kiblnd_cm_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event) case RDMA_CM_EVENT_ROUTE_ERROR: peer = (kib_peer_t *)cmid->context; - CDEBUG(D_NETERROR, "%s: ROUTE ERROR %d\n", - libcfs_nid2str(peer->ibp_nid), event->status); + CNETERR("%s: ROUTE ERROR %d\n", + libcfs_nid2str(peer->ibp_nid), event->status); kiblnd_peer_connect_failed(peer, 1, -EHOSTUNREACH); kiblnd_peer_decref(peer); return -EHOSTUNREACH; /* rc != 0 destroys cmid */ @@ -2682,7 +2885,7 @@ kiblnd_cm_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event) if (event->status == 0) return kiblnd_active_connect(cmid); - CDEBUG(D_NETERROR, "Can't resolve route for %s: %d\n", + CNETERR("Can't resolve route for %s: %d\n", libcfs_nid2str(peer->ibp_nid), event->status); kiblnd_peer_connect_failed(peer, 1, event->status); kiblnd_peer_decref(peer); @@ -2692,7 +2895,7 @@ kiblnd_cm_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event) conn = (kib_conn_t *)cmid->context; LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT || conn->ibc_state == IBLND_CONN_PASSIVE_WAIT); - CDEBUG(D_NETERROR, "%s: UNREACHABLE %d\n", + CNETERR("%s: UNREACHABLE %d\n", libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status); kiblnd_connreq_done(conn, -ENETDOWN); kiblnd_conn_decref(conn); @@ -2702,8 +2905,8 @@ kiblnd_cm_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event) conn = (kib_conn_t *)cmid->context; LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT || conn->ibc_state == IBLND_CONN_PASSIVE_WAIT); - CDEBUG(D_NETERROR, "%s: CONNECT ERROR %d\n", - libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status); + CNETERR("%s: CONNECT ERROR %d\n", + libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status); kiblnd_connreq_done(conn, -ENOTCONN); kiblnd_conn_decref(conn); return 0; @@ -2753,6 +2956,11 @@ kiblnd_cm_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event) /* net keeps its ref on conn! */ return 0; +#ifdef HAVE_OFED_RDMA_CMEV_TIMEWAIT_EXIT + case RDMA_CM_EVENT_TIMEWAIT_EXIT: + CDEBUG(D_NET, "Ignore TIMEWAIT_EXIT event\n"); + return 0; +#endif case RDMA_CM_EVENT_DISCONNECTED: conn = (kib_conn_t *)cmid->context; if (conn->ibc_state < IBLND_CONN_ESTABLISHED) { @@ -2763,173 +2971,209 @@ kiblnd_cm_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event) kiblnd_close_conn(conn, 0); } kiblnd_conn_decref(conn); + cmid->context = NULL; return 0; - case RDMA_CM_EVENT_DEVICE_REMOVAL: - LCONSOLE_ERROR_MSG(0x131, + case RDMA_CM_EVENT_DEVICE_REMOVAL: + LCONSOLE_ERROR_MSG(0x131, "Received notification of device removal\n" "Please shutdown LNET to allow this to proceed\n"); /* Can't remove network from underneath LNET for now, so I have * to ignore this */ - return 0; - } + return 0; + +#ifdef HAVE_OFED_RDMA_CMEV_ADDRCHANGE + case RDMA_CM_EVENT_ADDR_CHANGE: + LCONSOLE_INFO("Physical link changed (eg hca/port)\n"); + return 0; +#endif + } } -int -kiblnd_check_txs (kib_conn_t *conn, struct list_head *txs) +static int +kiblnd_check_txs_locked(kib_conn_t *conn, cfs_list_t *txs) { kib_tx_t *tx; - struct list_head *ttmp; - int timed_out = 0; + cfs_list_t *ttmp; - spin_lock(&conn->ibc_lock); - - list_for_each (ttmp, txs) { - tx = list_entry (ttmp, kib_tx_t, tx_list); + cfs_list_for_each (ttmp, txs) { + tx = cfs_list_entry (ttmp, kib_tx_t, tx_list); if (txs != &conn->ibc_active_txs) { LASSERT (tx->tx_queued); } else { LASSERT (!tx->tx_queued); LASSERT (tx->tx_waiting || tx->tx_sending != 0); - } + } - if (time_after_eq (jiffies, tx->tx_deadline)) { - timed_out = 1; - break; + if (cfs_time_aftereq (jiffies, tx->tx_deadline)) { + CERROR("Timed out tx: %s, %lu seconds\n", + kiblnd_queue2str(conn, txs), + cfs_duration_sec(jiffies - tx->tx_deadline)); + return 1; } } - spin_unlock(&conn->ibc_lock); - return timed_out; + return 0; } -int -kiblnd_conn_timed_out (kib_conn_t *conn) +static int +kiblnd_conn_timed_out_locked(kib_conn_t *conn) { - return kiblnd_check_txs(conn, &conn->ibc_tx_queue) || - kiblnd_check_txs(conn, &conn->ibc_tx_queue_rsrvd) || - kiblnd_check_txs(conn, &conn->ibc_tx_queue_nocred) || - kiblnd_check_txs(conn, &conn->ibc_active_txs); + return kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue) || + kiblnd_check_txs_locked(conn, &conn->ibc_tx_noops) || + kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_rsrvd) || + kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_nocred) || + kiblnd_check_txs_locked(conn, &conn->ibc_active_txs); } void kiblnd_check_conns (int idx) { - struct list_head *peers = &kiblnd_data.kib_peers[idx]; - struct list_head *ptmp; - kib_peer_t *peer; - kib_conn_t *conn; - struct list_head *ctmp; - unsigned long flags; + CFS_LIST_HEAD (closes); + CFS_LIST_HEAD (checksends); + cfs_list_t *peers = &kiblnd_data.kib_peers[idx]; + cfs_list_t *ptmp; + kib_peer_t *peer; + kib_conn_t *conn; + cfs_list_t *ctmp; + unsigned long flags; - again: /* NB. We expect to have a look at all the peers and not find any - * rdmas to time out, so we just use a shared lock while we + * RDMAs to time out, so we just use a shared lock while we * take a look... */ - read_lock_irqsave(&kiblnd_data.kib_global_lock, flags); + read_lock_irqsave(&kiblnd_data.kib_global_lock, flags); + + cfs_list_for_each (ptmp, peers) { + peer = cfs_list_entry (ptmp, kib_peer_t, ibp_list); - list_for_each (ptmp, peers) { - peer = list_entry (ptmp, kib_peer_t, ibp_list); + cfs_list_for_each (ctmp, &peer->ibp_conns) { + int timedout; + int sendnoop; - list_for_each (ctmp, &peer->ibp_conns) { - conn = list_entry (ctmp, kib_conn_t, ibc_list); + conn = cfs_list_entry(ctmp, kib_conn_t, ibc_list); LASSERT (conn->ibc_state == IBLND_CONN_ESTABLISHED); - /* In case we have enough credits to return via a - * NOOP, but there were no non-blocking tx descs - * free to do it last time... */ - kiblnd_check_sends(conn); + spin_lock(&conn->ibc_lock); - if (!kiblnd_conn_timed_out(conn)) + sendnoop = kiblnd_need_noop(conn); + timedout = kiblnd_conn_timed_out_locked(conn); + if (!sendnoop && !timedout) { + spin_unlock(&conn->ibc_lock); continue; + } - /* Handle timeout by closing the whole connection. We - * can only be sure RDMA activity has ceased once the - * QP has been modified. */ - - kiblnd_conn_addref(conn); /* 1 ref for me... */ - - read_unlock_irqrestore(&kiblnd_data.kib_global_lock, - flags); + if (timedout) { + CERROR("Timed out RDMA with %s (%lu): " + "c: %u, oc: %u, rc: %u\n", + libcfs_nid2str(peer->ibp_nid), + cfs_duration_sec(cfs_time_current() - + peer->ibp_last_alive), + conn->ibc_credits, + conn->ibc_outstanding_credits, + conn->ibc_reserved_credits); + cfs_list_add(&conn->ibc_connd_list, &closes); + } else { + cfs_list_add(&conn->ibc_connd_list, + &checksends); + } + /* +ref for 'closes' or 'checksends' */ + kiblnd_conn_addref(conn); - CERROR("Timed out RDMA with %s\n", - libcfs_nid2str(peer->ibp_nid)); + spin_unlock(&conn->ibc_lock); + } + } - kiblnd_close_conn(conn, -ETIMEDOUT); - kiblnd_conn_decref(conn); /* ...until here */ + read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); - /* start again now I've dropped the lock */ - goto again; - } + /* Handle timeout by closing the whole + * connection. We can only be sure RDMA activity + * has ceased once the QP has been modified. */ + while (!cfs_list_empty(&closes)) { + conn = cfs_list_entry(closes.next, + kib_conn_t, ibc_connd_list); + cfs_list_del(&conn->ibc_connd_list); + kiblnd_close_conn(conn, -ETIMEDOUT); + kiblnd_conn_decref(conn); } - read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); + /* In case we have enough credits to return via a + * NOOP, but there were no non-blocking tx descs + * free to do it last time... */ + while (!cfs_list_empty(&checksends)) { + conn = cfs_list_entry(checksends.next, + kib_conn_t, ibc_connd_list); + cfs_list_del(&conn->ibc_connd_list); + kiblnd_check_sends(conn); + kiblnd_conn_decref(conn); + } } void kiblnd_disconnect_conn (kib_conn_t *conn) { - LASSERT (!in_interrupt()); - LASSERT (current == kiblnd_data.kib_connd); - LASSERT (conn->ibc_state == IBLND_CONN_CLOSING); + LASSERT (!in_interrupt()); + LASSERT (current == kiblnd_data.kib_connd); + LASSERT (conn->ibc_state == IBLND_CONN_CLOSING); - rdma_disconnect(conn->ibc_cmid); - kiblnd_finalise_conn(conn); + rdma_disconnect(conn->ibc_cmid); + kiblnd_finalise_conn(conn); - kiblnd_peer_notify(conn->ibc_peer); + kiblnd_peer_notify(conn->ibc_peer); } int kiblnd_connd (void *arg) { - wait_queue_t wait; - unsigned long flags; - kib_conn_t *conn; - int timeout; - int i; - int dropped_lock; - int peer_index = 0; - unsigned long deadline = jiffies; + wait_queue_t wait; + unsigned long flags; + kib_conn_t *conn; + int timeout; + int i; + int dropped_lock; + int peer_index = 0; + unsigned long deadline = jiffies; - cfs_daemonize ("kiblnd_connd"); - cfs_block_allsigs (); + cfs_block_allsigs (); - init_waitqueue_entry (&wait, current); - kiblnd_data.kib_connd = current; + init_waitqueue_entry_current (&wait); + kiblnd_data.kib_connd = current; - spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); + spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); while (!kiblnd_data.kib_shutdown) { dropped_lock = 0; - if (!list_empty (&kiblnd_data.kib_connd_zombies)) { - conn = list_entry (kiblnd_data.kib_connd_zombies.next, - kib_conn_t, ibc_list); - list_del (&conn->ibc_list); + if (!cfs_list_empty (&kiblnd_data.kib_connd_zombies)) { + conn = cfs_list_entry(kiblnd_data. \ + kib_connd_zombies.next, + kib_conn_t, ibc_list); + cfs_list_del(&conn->ibc_list); - spin_unlock_irqrestore (&kiblnd_data.kib_connd_lock, flags); - dropped_lock = 1; + spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, + flags); + dropped_lock = 1; - kiblnd_destroy_conn(conn); + kiblnd_destroy_conn(conn); - spin_lock_irqsave (&kiblnd_data.kib_connd_lock, flags); - } + spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); + } - if (!list_empty (&kiblnd_data.kib_connd_conns)) { - conn = list_entry (kiblnd_data.kib_connd_conns.next, - kib_conn_t, ibc_list); - list_del (&conn->ibc_list); + if (!cfs_list_empty(&kiblnd_data.kib_connd_conns)) { + conn = cfs_list_entry(kiblnd_data.kib_connd_conns.next, + kib_conn_t, ibc_list); + cfs_list_del(&conn->ibc_list); - spin_unlock_irqrestore (&kiblnd_data.kib_connd_lock, flags); - dropped_lock = 1; + spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, + flags); + dropped_lock = 1; - kiblnd_disconnect_conn(conn); - kiblnd_conn_decref(conn); + kiblnd_disconnect_conn(conn); + kiblnd_conn_decref(conn); - spin_lock_irqsave (&kiblnd_data.kib_connd_lock, flags); + spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); } /* careful with the jiffy wrap... */ @@ -2939,7 +3183,7 @@ kiblnd_connd (void *arg) const int p = 1; int chunk = kiblnd_data.kib_peer_hash_size; - spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags); + spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags); dropped_lock = 1; /* Time to check for RDMA timeouts on a few more @@ -2956,35 +3200,35 @@ kiblnd_connd (void *arg) if (chunk == 0) chunk = 1; - for (i = 0; i < chunk; i++) { - kiblnd_check_conns(peer_index); - peer_index = (peer_index + 1) % - kiblnd_data.kib_peer_hash_size; - } + for (i = 0; i < chunk; i++) { + kiblnd_check_conns(peer_index); + peer_index = (peer_index + 1) % + kiblnd_data.kib_peer_hash_size; + } - deadline += p * HZ; - spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); - } + deadline += p * HZ; + spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); + } - if (dropped_lock) - continue; + if (dropped_lock) + continue; - /* Nothing to do for 'timeout' */ - set_current_state (TASK_INTERRUPTIBLE); - add_wait_queue (&kiblnd_data.kib_connd_waitq, &wait); - spin_unlock_irqrestore (&kiblnd_data.kib_connd_lock, flags); + /* Nothing to do for 'timeout' */ + set_current_state(TASK_INTERRUPTIBLE); + add_wait_queue(&kiblnd_data.kib_connd_waitq, &wait); + spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags); - schedule_timeout (timeout); + waitq_timedwait(&wait, TASK_INTERRUPTIBLE, timeout); - set_current_state (TASK_RUNNING); - remove_wait_queue (&kiblnd_data.kib_connd_waitq, &wait); - spin_lock_irqsave (&kiblnd_data.kib_connd_lock, flags); - } + set_current_state(TASK_RUNNING); + remove_wait_queue(&kiblnd_data.kib_connd_waitq, &wait); + spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); + } - spin_unlock_irqrestore (&kiblnd_data.kib_connd_lock, flags); + spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags); - kiblnd_thread_fini(); - return (0); + kiblnd_thread_fini(); + return 0; } void @@ -2997,7 +3241,7 @@ kiblnd_qp_event(struct ib_event *event, void *arg) CDEBUG(D_NET, "%s established\n", libcfs_nid2str(conn->ibc_peer->ibp_nid)); return; - + default: CERROR("%s: Async QP event type %d\n", libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event); @@ -3019,8 +3263,8 @@ kiblnd_complete (struct ib_wc *wc) * failing RDMA because 'tx' might be back on the idle list or * even reused already if we didn't manage to post all our work * items */ - CDEBUG(D_NETERROR, "RDMA (tx: %p) failed: %d\n", - kiblnd_wreqid2ptr(wc->wr_id), wc->status); + CNETERR("RDMA (tx: %p) failed: %d\n", + kiblnd_wreqid2ptr(wc->wr_id), wc->status); return; case IBLND_WID_TX: @@ -3035,33 +3279,35 @@ kiblnd_complete (struct ib_wc *wc) } void -kiblnd_cq_completion (struct ib_cq *cq, void *arg) +kiblnd_cq_completion(struct ib_cq *cq, void *arg) { - /* NB I'm not allowed to schedule this conn once its refcount has - * reached 0. Since fundamentally I'm racing with scheduler threads - * consuming my CQ I could be called after all completions have - * occurred. But in this case, ibc_nrx == 0 && ibc_nsends_posted == 0 - * and this CQ is about to be destroyed so I NOOP. */ - kib_conn_t *conn = (kib_conn_t *)arg; - unsigned long flags; - - LASSERT (cq == conn->ibc_cq); - - spin_lock_irqsave(&kiblnd_data.kib_sched_lock, flags); - - conn->ibc_ready = 1; - - if (!conn->ibc_scheduled && - (conn->ibc_nrx > 0 || - conn->ibc_nsends_posted > 0)) { - kiblnd_conn_addref(conn); /* +1 ref for sched_conns */ - conn->ibc_scheduled = 1; - list_add_tail(&conn->ibc_sched_list, - &kiblnd_data.kib_sched_conns); - wake_up(&kiblnd_data.kib_sched_waitq); - } + /* NB I'm not allowed to schedule this conn once its refcount has + * reached 0. Since fundamentally I'm racing with scheduler threads + * consuming my CQ I could be called after all completions have + * occurred. But in this case, ibc_nrx == 0 && ibc_nsends_posted == 0 + * and this CQ is about to be destroyed so I NOOP. */ + kib_conn_t *conn = (kib_conn_t *)arg; + struct kib_sched_info *sched = conn->ibc_sched; + unsigned long flags; + + LASSERT(cq == conn->ibc_cq); + + spin_lock_irqsave(&sched->ibs_lock, flags); + + conn->ibc_ready = 1; + + if (!conn->ibc_scheduled && + (conn->ibc_nrx > 0 || + conn->ibc_nsends_posted > 0)) { + kiblnd_conn_addref(conn); /* +1 ref for sched_conns */ + conn->ibc_scheduled = 1; + cfs_list_add_tail(&conn->ibc_sched_list, &sched->ibs_conns); + + if (waitqueue_active(&sched->ibs_waitq)) + wake_up(&sched->ibs_waitq); + } - spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock, flags); + spin_unlock_irqrestore(&sched->ibs_lock, flags); } void @@ -3076,105 +3322,217 @@ kiblnd_cq_event(struct ib_event *event, void *arg) int kiblnd_scheduler(void *arg) { - long id = (long)arg; - wait_queue_t wait; - char name[16]; - unsigned long flags; - kib_conn_t *conn; - struct ib_wc wc; - int rc; - int did_something; - int busy_loops = 0; - - snprintf(name, sizeof(name), "kiblnd_sd_%02ld", id); - cfs_daemonize(name); - cfs_block_allsigs(); + long id = (long)arg; + struct kib_sched_info *sched; + kib_conn_t *conn; + wait_queue_t wait; + unsigned long flags; + struct ib_wc wc; + int did_something; + int busy_loops = 0; + int rc; + + cfs_block_allsigs(); + + init_waitqueue_entry_current(&wait); + + sched = kiblnd_data.kib_scheds[KIB_THREAD_CPT(id)]; + + rc = cfs_cpt_bind(lnet_cpt_table(), sched->ibs_cpt); + if (rc != 0) { + CWARN("Failed to bind on CPT %d, please verify whether " + "all CPUs are healthy and reload modules if necessary, " + "otherwise your system might under risk of low " + "performance\n", sched->ibs_cpt); + } - init_waitqueue_entry(&wait, current); + spin_lock_irqsave(&sched->ibs_lock, flags); - spin_lock_irqsave(&kiblnd_data.kib_sched_lock, flags); + while (!kiblnd_data.kib_shutdown) { + if (busy_loops++ >= IBLND_RESCHED) { + spin_unlock_irqrestore(&sched->ibs_lock, flags); - while (!kiblnd_data.kib_shutdown) { - if (busy_loops++ >= IBLND_RESCHED) { - spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock, - flags); + cond_resched(); + busy_loops = 0; - our_cond_resched(); - busy_loops = 0; + spin_lock_irqsave(&sched->ibs_lock, flags); + } - spin_lock_irqsave(&kiblnd_data.kib_sched_lock, flags); - } + did_something = 0; - did_something = 0; + if (!cfs_list_empty(&sched->ibs_conns)) { + conn = cfs_list_entry(sched->ibs_conns.next, + kib_conn_t, ibc_sched_list); + /* take over kib_sched_conns' ref on conn... */ + LASSERT(conn->ibc_scheduled); + cfs_list_del(&conn->ibc_sched_list); + conn->ibc_ready = 0; - if (!list_empty(&kiblnd_data.kib_sched_conns)) { - conn = list_entry(kiblnd_data.kib_sched_conns.next, - kib_conn_t, ibc_sched_list); - /* take over kib_sched_conns' ref on conn... */ - LASSERT(conn->ibc_scheduled); - list_del(&conn->ibc_sched_list); - conn->ibc_ready = 0; - - spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock, - flags); + spin_unlock_irqrestore(&sched->ibs_lock, flags); rc = ib_poll_cq(conn->ibc_cq, 1, &wc); if (rc == 0) { rc = ib_req_notify_cq(conn->ibc_cq, IB_CQ_NEXT_COMP); - LASSERT (rc >= 0); - - rc = ib_poll_cq(conn->ibc_cq, 1, &wc); + if (rc < 0) { + CWARN("%s: ib_req_notify_cq failed: %d, " + "closing connection\n", + libcfs_nid2str(conn->ibc_peer->ibp_nid), rc); + kiblnd_close_conn(conn, -EIO); + kiblnd_conn_decref(conn); + spin_lock_irqsave(&sched->ibs_lock, + flags); + continue; + } + + rc = ib_poll_cq(conn->ibc_cq, 1, &wc); + } + + if (rc < 0) { + CWARN("%s: ib_poll_cq failed: %d, " + "closing connection\n", + libcfs_nid2str(conn->ibc_peer->ibp_nid), + rc); + kiblnd_close_conn(conn, -EIO); + kiblnd_conn_decref(conn); + spin_lock_irqsave(&sched->ibs_lock, flags); + continue; + } + + spin_lock_irqsave(&sched->ibs_lock, flags); + + if (rc != 0 || conn->ibc_ready) { + /* There may be another completion waiting; get + * another scheduler to check while I handle + * this one... */ + /* +1 ref for sched_conns */ + kiblnd_conn_addref(conn); + cfs_list_add_tail(&conn->ibc_sched_list, + &sched->ibs_conns); + if (waitqueue_active(&sched->ibs_waitq)) + wake_up(&sched->ibs_waitq); + } else { + conn->ibc_scheduled = 0; + } + + if (rc != 0) { + spin_unlock_irqrestore(&sched->ibs_lock, flags); + kiblnd_complete(&wc); + + spin_lock_irqsave(&sched->ibs_lock, flags); } - LASSERT (rc >= 0); + kiblnd_conn_decref(conn); /* ...drop my ref from above */ + did_something = 1; + } - spin_lock_irqsave(&kiblnd_data.kib_sched_lock, - flags); + if (did_something) + continue; - if (rc != 0 || conn->ibc_ready) { - /* There may be another completion waiting; get - * another scheduler to check while I handle - * this one... */ - kiblnd_conn_addref(conn); /* +1 ref for sched_conns */ - list_add_tail(&conn->ibc_sched_list, - &kiblnd_data.kib_sched_conns); - wake_up(&kiblnd_data.kib_sched_waitq); - } else { - conn->ibc_scheduled = 0; - } - - if (rc != 0) { - spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock, - flags); + set_current_state(TASK_INTERRUPTIBLE); + add_wait_queue_exclusive(&sched->ibs_waitq, &wait); + spin_unlock_irqrestore(&sched->ibs_lock, flags); - kiblnd_complete(&wc); + waitq_wait(&wait, TASK_INTERRUPTIBLE); + busy_loops = 0; - spin_lock_irqsave(&kiblnd_data.kib_sched_lock, - flags); - } + remove_wait_queue(&sched->ibs_waitq, &wait); + set_current_state(TASK_RUNNING); + spin_lock_irqsave(&sched->ibs_lock, flags); + } - kiblnd_conn_decref(conn); /* ...drop my ref from above */ - did_something = 1; + spin_unlock_irqrestore(&sched->ibs_lock, flags); + + kiblnd_thread_fini(); + return 0; +} + +int +kiblnd_failover_thread(void *arg) +{ + rwlock_t *glock = &kiblnd_data.kib_global_lock; + kib_dev_t *dev; + wait_queue_t wait; + unsigned long flags; + int rc; + + LASSERT (*kiblnd_tunables.kib_dev_failover != 0); + + cfs_block_allsigs (); + + init_waitqueue_entry_current(&wait); + write_lock_irqsave(glock, flags); + + while (!kiblnd_data.kib_shutdown) { + int do_failover = 0; + int long_sleep; + + cfs_list_for_each_entry(dev, &kiblnd_data.kib_failed_devs, + ibd_fail_list) { + if (cfs_time_before(cfs_time_current(), + dev->ibd_next_failover)) + continue; + do_failover = 1; + break; } - if (did_something) + if (do_failover) { + cfs_list_del_init(&dev->ibd_fail_list); + dev->ibd_failover = 1; + write_unlock_irqrestore(glock, flags); + + rc = kiblnd_dev_failover(dev); + + write_lock_irqsave(glock, flags); + + LASSERT (dev->ibd_failover); + dev->ibd_failover = 0; + if (rc >= 0) { /* Device is OK or failover succeed */ + dev->ibd_next_failover = cfs_time_shift(3); + continue; + } + + /* failed to failover, retry later */ + dev->ibd_next_failover = + cfs_time_shift(min(dev->ibd_failed_failover, 10)); + if (kiblnd_dev_can_failover(dev)) { + cfs_list_add_tail(&dev->ibd_fail_list, + &kiblnd_data.kib_failed_devs); + } + continue; + } + + /* long sleep if no more pending failover */ + long_sleep = cfs_list_empty(&kiblnd_data.kib_failed_devs); + + set_current_state(TASK_INTERRUPTIBLE); + add_wait_queue(&kiblnd_data.kib_failover_waitq, &wait); + write_unlock_irqrestore(glock, flags); - set_current_state(TASK_INTERRUPTIBLE); - add_wait_queue_exclusive(&kiblnd_data.kib_sched_waitq, &wait); - spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock, flags); + rc = schedule_timeout(long_sleep ? cfs_time_seconds(10) : + cfs_time_seconds(1)); + set_current_state(TASK_RUNNING); + remove_wait_queue(&kiblnd_data.kib_failover_waitq, &wait); + write_lock_irqsave(glock, flags); - schedule(); - busy_loops = 0; + if (!long_sleep || rc != 0) + continue; - remove_wait_queue(&kiblnd_data.kib_sched_waitq, &wait); - set_current_state(TASK_RUNNING); - spin_lock_irqsave(&kiblnd_data.kib_sched_lock, flags); + /* have a long sleep, routine check all active devices, + * we need checking like this because if there is not active + * connection on the dev and no SEND from local, we may listen + * on wrong HCA for ever while there is a bonding failover */ + cfs_list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) { + if (kiblnd_dev_can_failover(dev)) { + cfs_list_add_tail(&dev->ibd_fail_list, + &kiblnd_data.kib_failed_devs); + } + } } - spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock, flags); + write_unlock_irqrestore(glock, flags); kiblnd_thread_fini(); - return (0); + return 0; }