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=73b4c13422bab353f393219612648de672d0a690;hb=35a1096a4741b24dbcd674bda5daf7b8a13cfe28;hpb=9b02beed217855f295eb51555fb9b4d2110af233 diff --git a/lnet/klnds/o2iblnd/o2iblnd_cb.c b/lnet/klnds/o2iblnd/o2iblnd_cb.c index 73b4c13..eeef534 100644 --- a/lnet/klnds/o2iblnd/o2iblnd_cb.c +++ b/lnet/klnds/o2iblnd/o2iblnd_cb.c @@ -1,6 +1,4 @@ -/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*- - * vim:expandtab:shiftwidth=8:tabstop=8: - * +/* * GPL HEADER START * * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. @@ -26,8 +24,10 @@ * GPL HEADER END */ /* - * Copyright 2008 Sun Microsystems, Inc. All rights reserved + * 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/ @@ -43,55 +43,55 @@ 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; + 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); + 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); + 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; + /* 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); + if (tx->tx_conn != NULL) { + LASSERT (ni == tx->tx_conn->ibc_peer->ibp_ni); - kiblnd_conn_decref(tx->tx_conn); - tx->tx_conn = NULL; - } + kiblnd_conn_decref(tx->tx_conn); + tx->tx_conn = NULL; + } - tx->tx_nwrq = 0; - tx->tx_status = 0; + tx->tx_nwrq = 0; + tx->tx_status = 0; - kiblnd_pool_free_node(&tx->tx_pool->tpo_pool, &tx->tx_list); + 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); + while (!cfs_list_empty (txlist)) { + tx = cfs_list_entry (txlist->next, kib_tx_t, tx_list); - list_del(&tx->tx_list); + cfs_list_del(&tx->tx_list); /* complete now */ tx->tx_waiting = 0; tx->tx_status = status; @@ -100,13 +100,15 @@ 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 = (kib_net_t *)ni->ni_data; - struct list_head *node; - kib_tx_t *tx; + kib_net_t *net = (kib_net_t *)ni->ni_data; + cfs_list_t *node; + kib_tx_t *tx; + kib_tx_poolset_t *tps; - node = kiblnd_pool_alloc_node(&net->ibn_tx_ps.tps_poolset); + tps = net->ibn_tx_ps[lnet_cpt_of_nid(target)]; + node = kiblnd_pool_alloc_node(&tps->tps_poolset); if (node == NULL) return NULL; tx = container_of(node, kib_tx_t, tx_list); @@ -126,36 +128,37 @@ kiblnd_get_idle_tx (lnet_ni_t *ni) } 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; - kiblnd_conn_decref(conn); + 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); } 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 = NULL; - struct ib_mr *mr; - int rc; + 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); + 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(net, rx->rx_msgaddr, IBLND_MSG_SIZE); - LASSERT (mr != NULL); + 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; @@ -195,24 +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); @@ -234,15 +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)); @@ -262,19 +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", @@ -304,20 +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(conn->ibc_version)) { - 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(conn->ibc_version)); + rc2, credits, + IBLND_MSG_QUEUE_SIZE(conn->ibc_version)); kiblnd_close_conn(conn, -EPROTO); kiblnd_post_rx(rx, IBLND_POSTRX_NO_CREDIT); @@ -326,7 +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); } @@ -339,9 +349,14 @@ kiblnd_handle_rx (kib_rx_t *rx) break; case IBLND_MSG_NOOP: - if (IBLND_OOB_CAPABLE(conn->ibc_version)) + if (IBLND_OOB_CAPABLE(conn->ibc_version)) { post_credit = IBLND_POSTRX_NO_CREDIT; - else + break; + } + + if (credits != 0) /* credit already posted */ + post_credit = IBLND_POSTRX_NO_CREDIT; + else /* a keepalive NOOP */ post_credit = IBLND_POSTRX_PEER_CREDIT; break; @@ -362,7 +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, @@ -372,12 +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", @@ -401,11 +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; @@ -455,8 +471,8 @@ kiblnd_rx_complete (kib_rx_t *rx, int status, int nob) 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; } @@ -486,17 +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) { - 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) { - list_add_tail(&rx->rx_list, &conn->ibc_early_rxs); - write_unlock_irqrestore(g_lock, flags); - return; - } - write_unlock_irqrestore(g_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; @@ -535,22 +551,32 @@ kiblnd_kvaddr_to_page (unsigned long vaddr) static int kiblnd_fmr_map_tx(kib_net_t *net, kib_tx_t *tx, kib_rdma_desc_t *rd, int nob) { - kib_dev_t *ibdev = net->ibn_dev; - __u64 *pages = tx->tx_pages; - int npages; - int size; - int rc; - int i; + 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 += ibdev->ibd_page_size) { + size += hdev->ibh_page_size) { pages[npages ++] = (rd->rd_frags[i].rf_addr & - ibdev->ibd_page_mask) + size; + hdev->ibh_page_mask) + size; } } - rc = kiblnd_fmr_pool_map(&net->ibn_fmr_ps, pages, npages, 0, &tx->tx_u.fmr); + 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; @@ -560,7 +586,7 @@ kiblnd_fmr_map_tx(kib_net_t *net, kib_tx_t *tx, kib_rdma_desc_t *rd, int nob) * 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 &= ~ibdev->ibd_page_mask; + rd->rd_frags[0].rf_addr &= ~hdev->ibh_page_mask; rd->rd_frags[0].rf_nob = nob; rd->rd_nfrags = 1; @@ -570,12 +596,23 @@ kiblnd_fmr_map_tx(kib_net_t *net, kib_tx_t *tx, kib_rdma_desc_t *rd, int nob) static int kiblnd_pmr_map_tx(kib_net_t *net, kib_tx_t *tx, kib_rdma_desc_t *rd, int nob) { - __u64 iova; - int rc; + 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; - iova = rd->rd_frags[0].rf_addr & ~net->ibn_dev->ibd_page_mask; + cpt = tx->tx_pool->tpo_pool.po_owner->ps_cpt; - rc = kiblnd_pmr_pool_map(&net->ibn_pmr_ps, rd, &iova, &tx->tx_u.pmr); + 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; @@ -595,20 +632,21 @@ kiblnd_pmr_map_tx(kib_net_t *net, kib_tx_t *tx, kib_rdma_desc_t *rd, int nob) void kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx) { - kib_net_t *net = ni->ni_data; + kib_net_t *net = ni->ni_data; - LASSERT (net != NULL); + LASSERT(net != NULL); - if (net->ibn_with_fmr && 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_with_pmr && tx->tx_u.pmr != NULL) { - kiblnd_pmr_pool_unmap(tx->tx_u.pmr); - tx->tx_u.pmr = 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(net->ibn_dev->ibd_cmid->device, + kiblnd_dma_unmap_sg(tx->tx_pool->tpo_hdev->ibh_ibdev, tx->tx_frags, tx->tx_nfrags, tx->tx_dmadir); tx->tx_nfrags = 0; } @@ -618,6 +656,7 @@ int 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; @@ -629,31 +668,31 @@ kiblnd_map_tx(lnet_ni_t *ni, kib_tx_t *tx, tx->tx_nfrags = nfrags; rd->rd_nfrags = - kiblnd_dma_map_sg(net->ibn_dev->ibd_cmid->device, + 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( - net->ibn_dev->ibd_cmid->device, &tx->tx_frags[i]); + hdev->ibh_ibdev, &tx->tx_frags[i]); rd->rd_frags[i].rf_addr = kiblnd_sg_dma_address( - net->ibn_dev->ibd_cmid->device, &tx->tx_frags[i]); + 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(net, rd); + 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_with_fmr) - return kiblnd_fmr_map_tx(net, tx, rd, nob); - else if (net->ibn_with_pmr) - return kiblnd_pmr_map_tx(net, tx, rd, nob); + 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; + return -EINVAL; } @@ -737,7 +776,6 @@ 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_set_page(sg, kiov->kiov_page, fragnob, kiov->kiov_offset + offset); sg++; @@ -786,27 +824,27 @@ kiblnd_post_tx_locked (kib_conn_t *conn, kib_tx_t *tx, int credit) } if (credit != 0 && !IBLND_OOB_CAPABLE(ver) && - conn->ibc_credits == 1 && /* last credit reserved for */ - conn->ibc_outstanding_credits == 0) { /* giving back credits */ + 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 */ - list_del(&tx->tx_list); + cfs_list_del(&tx->tx_list); tx->tx_queued = 0; if (msg->ibm_type == IBLND_MSG_NOOP && - (!kiblnd_send_noop(conn) || /* redundant 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); + 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); @@ -829,14 +867,20 @@ kiblnd_post_tx_locked (kib_conn_t *conn, kib_tx_t *tx, int credit) * 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); + cfs_list_add(&tx->tx_list, &conn->ibc_active_txs); /* I'm still holding ibc_lock! */ - if (conn->ibc_state != IBLND_CONN_ESTABLISHED) + if (conn->ibc_state != IBLND_CONN_ESTABLISHED) { rc = -ECONNABORTED; - else + } 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); + } + conn->ibc_last_send = jiffies; if (rc == 0) @@ -856,9 +900,9 @@ kiblnd_post_tx_locked (kib_conn_t *conn, kib_tx_t *tx, int credit) done = (tx->tx_sending == 0); if (done) - list_del(&tx->tx_list); + cfs_list_del(&tx->tx_list); - spin_unlock(&conn->ibc_lock); + spin_unlock(&conn->ibc_lock); if (conn->ibc_state == IBLND_CONN_ESTABLISHED) CERROR("Error %d posting transmit to %s\n", @@ -871,7 +915,10 @@ kiblnd_post_tx_locked (kib_conn_t *conn, kib_tx_t *tx, int credit) if (done) kiblnd_tx_done(peer->ibp_ni, tx); - return -EIO; + + spin_lock(&conn->ibc_lock); + + return -EIO; } void @@ -888,7 +935,7 @@ kiblnd_check_sends (kib_conn_t *conn) return; } - spin_lock(&conn->ibc_lock); + spin_lock(&conn->ibc_lock); LASSERT (conn->ibc_nsends_posted <= IBLND_CONCURRENT_SENDS(ver)); LASSERT (!IBLND_OOB_CAPABLE(ver) || @@ -896,37 +943,44 @@ kiblnd_check_sends (kib_conn_t *conn) 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 (kiblnd_send_noop(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); + 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); + 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 (;;) { int credit; - if (!list_empty(&conn->ibc_tx_queue_nocred)) { + if (!cfs_list_empty(&conn->ibc_tx_queue_nocred)) { credit = 0; - tx = list_entry(conn->ibc_tx_queue_nocred.next, - kib_tx_t, tx_list); - } else if (!list_empty(&conn->ibc_tx_queue)) { + 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 = list_entry(conn->ibc_tx_queue.next, + 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; @@ -934,7 +988,9 @@ kiblnd_check_sends (kib_conn_t *conn) break; } - spin_unlock(&conn->ibc_lock); + spin_unlock(&conn->ibc_lock); + + kiblnd_conn_decref(conn); /* ...until here */ } void @@ -948,18 +1004,18 @@ 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'. */ @@ -978,11 +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); @@ -995,20 +1051,19 @@ 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); - mr = kiblnd_find_dma_mr(net, tx->tx_msgaddr, nob); + mr = kiblnd_find_dma_mr(hdev, tx->tx_msgaddr, nob); LASSERT (mr != NULL); sge->lkey = mr->lkey; @@ -1029,23 +1084,23 @@ kiblnd_init_tx_msg (lnet_ni_t *ni, kib_tx_t *tx, int type, int body_nob) int kiblnd_init_rdma (kib_conn_t *conn, kib_tx_t *tx, int type, - int resid, kib_rdma_desc_t *dstrd, __u64 dstcookie) + 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 = 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; + 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) { @@ -1115,14 +1170,14 @@ kiblnd_init_rdma (kib_conn_t *conn, 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); @@ -1154,7 +1209,7 @@ kiblnd_queue_tx_locked (kib_tx_t *tx, kib_conn_t *conn) if (IBLND_OOB_CAPABLE(conn->ibc_version)) q = &conn->ibc_tx_queue_nocred; else - q = &conn->ibc_tx_queue; + q = &conn->ibc_tx_noops; break; case IBLND_MSG_IMMEDIATE: @@ -1162,17 +1217,59 @@ kiblnd_queue_tx_locked (kib_tx_t *tx, kib_conn_t *conn) 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 @@ -1188,7 +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)); @@ -1208,22 +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) { - 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; + 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: @@ -1236,7 +1343,7 @@ 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; @@ -1248,15 +1355,15 @@ kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid) /* 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); if (tx != NULL) kiblnd_queue_tx(tx, conn); @@ -1264,24 +1371,25 @@ kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid) 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); if (tx != NULL) - 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... */ + 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); + write_unlock_irqrestore(g_lock, flags); if (tx != NULL) kiblnd_queue_tx(tx, conn); @@ -1290,7 +1398,7 @@ kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid) 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); @@ -1304,22 +1412,23 @@ kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid) 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); if (tx != NULL) - 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... */ + 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); + write_unlock_irqrestore(g_lock, flags); if (tx != NULL) kiblnd_queue_tx(tx, conn); @@ -1338,12 +1447,12 @@ kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid) LASSERT (((kib_net_t *)ni->ni_data)->ibn_shutdown == 0); if (tx != NULL) - list_add_tail(&tx->tx_list, &peer->ibp_tx_queue); + 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); @@ -1372,18 +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); @@ -1398,12 +1507,12 @@ 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't 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; @@ -1452,7 +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", @@ -1491,7 +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)); @@ -1532,7 +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)); @@ -1592,18 +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]); @@ -1636,7 +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)); @@ -1696,53 +1805,51 @@ 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); + read_lock_irqsave(&kiblnd_data.kib_global_lock, flags); - if (list_empty(&peer->ibp_conns) && + 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, @@ -1758,8 +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); @@ -1770,25 +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); @@ -1798,61 +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); @@ -1867,71 +1985,72 @@ kiblnd_abort_txs(kib_conn_t *conn, struct list_head *txs) if (tx->tx_sending == 0) { tx->tx_queued = 0; - list_del (&tx->tx_list); - list_add (&tx->tx_list, &zombies); + 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); @@ -1939,17 +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); @@ -1960,21 +2079,21 @@ kiblnd_connreq_done(kib_conn_t *conn, int status) { kib_peer_t *peer = conn->ibc_peer; kib_tx_t *tx; - struct list_head txs; + cfs_list_t txs; unsigned long flags; int active; active = (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT); - CDEBUG(D_NET,"%s: active(%d), version(%x), status(%d)\n", - libcfs_nid2str(peer->ibp_nid), active, - conn->ibc_version, 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; @@ -1987,7 +2106,7 @@ kiblnd_connreq_done(kib_conn_t *conn, int status) } /* 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); @@ -1996,7 +2115,7 @@ 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 @@ -2016,8 +2135,8 @@ kiblnd_connreq_done(kib_conn_t *conn, int status) } /* grab pending txs while I have the lock */ - list_add(&txs, &peer->ibp_tx_queue); - list_del_init(&peer->ibp_tx_queue); + 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 */ @@ -2025,29 +2144,29 @@ kiblnd_connreq_done(kib_conn_t *conn, int status) /* start to shut down connection */ kiblnd_close_conn_locked(conn, -ECONNABORTED); - write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); + write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); - kiblnd_txlist_done(ni, &txs, -ECONNABORTED); + kiblnd_txlist_done(ni, &txs, -ECONNABORTED); - return; - } + 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 @@ -2064,7 +2183,7 @@ kiblnd_reject(struct rdma_cm_id *cmid, kib_rej_t *rej) int kiblnd_passive_connect (struct rdma_cm_id *cmid, void *priv, int priv_nob) { - rwlock_t *g_lock = &kiblnd_data.kib_global_lock; + rwlock_t *g_lock = &kiblnd_data.kib_global_lock; kib_msg_t *reqmsg = priv; kib_msg_t *ackmsg; kib_dev_t *ibdev; @@ -2076,21 +2195,30 @@ kiblnd_passive_connect (struct rdma_cm_id *cmid, void *priv, int priv_nob) lnet_nid_t nid; struct rdma_conn_param cp; kib_rej_t rej; - int version = IBLND_MSG_VERSION; - unsigned long flags; - int rc; - - LASSERT (!in_interrupt()); + 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); + /* 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; @@ -2201,7 +2329,7 @@ kiblnd_passive_connect (struct rdma_cm_id *cmid, void *priv, int priv_nob) goto failed; } - write_lock_irqsave(g_lock, flags); + write_lock_irqsave(g_lock, flags); peer2 = kiblnd_find_peer_locked(nid); if (peer2 != NULL) { @@ -2214,7 +2342,7 @@ kiblnd_passive_connect (struct rdma_cm_id *cmid, void *priv, int priv_nob) if (peer2->ibp_incarnation != reqmsg->ibm_srcstamp || peer2->ibp_version != version) { kiblnd_close_peer_conns_locked(peer2, -ESTALE); - write_unlock_irqrestore(g_lock, flags); + write_unlock_irqrestore(g_lock, flags); CWARN("Conn stale %s [old ver: %x, new ver: %x]\n", libcfs_nid2str(nid), peer2->ibp_version, version); @@ -2227,7 +2355,7 @@ kiblnd_passive_connect (struct rdma_cm_id *cmid, void *priv, int priv_nob) /* 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)); @@ -2239,7 +2367,7 @@ kiblnd_passive_connect (struct rdma_cm_id *cmid, void *priv, int priv_nob) 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 { @@ -2256,9 +2384,9 @@ kiblnd_passive_connect (struct rdma_cm_id *cmid, void *priv, int priv_nob) 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, version); @@ -2338,11 +2466,15 @@ kiblnd_reconnect (kib_conn_t *conn, int version, 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; @@ -2352,7 +2484,7 @@ kiblnd_reconnect (kib_conn_t *conn, int version, peer->ibp_incarnation = incarnation; } - write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); + write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); if (!retry) return; @@ -2375,13 +2507,13 @@ kiblnd_reconnect (kib_conn_t *conn, int version, break; } - CDEBUG(D_NETERROR, "%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); + 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); } @@ -2389,21 +2521,21 @@ kiblnd_reconnect (kib_conn_t *conn, int version, 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, IBLND_MSG_VERSION, 0, + IBLND_REJECT_CONN_STALE, NULL); + break; case IB_CM_REJ_INVALID_SERVICE_ID: - CDEBUG(D_NETERROR, "%s rejected: no listener at %d\n", - libcfs_nid2str(peer->ibp_nid), - *kiblnd_tunables.kib_service); + CNETERR("%s rejected: no listener at %d\n", + libcfs_nid2str(peer->ibp_nid), + *kiblnd_tunables.kib_service); break; case IB_CM_REJ_CONSUMER_DEFINED: @@ -2513,8 +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; } @@ -2584,13 +2716,13 @@ 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("Bad connection reply from %s, rc = %d, " @@ -2629,15 +2761,16 @@ kiblnd_active_connect (struct rdma_cm_id *cmid) struct rdma_conn_param cp; int version; __u64 incarnation; - long flags; + unsigned long flags; int rc; - read_lock_irqsave(&kiblnd_data.kib_global_lock, flags); + read_lock_irqsave(&kiblnd_data.kib_global_lock, flags); - incarnation = peer->ibp_incarnation; - version = (peer->ibp_version == 0) ? IBLND_MSG_VERSION : peer->ibp_version; + incarnation = peer->ibp_incarnation; + version = (peer->ibp_version == 0) ? IBLND_MSG_VERSION : + peer->ibp_version; - read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); + read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_ACTIVE_CONNECT, version); if (conn == NULL) { @@ -2707,7 +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); @@ -2720,8 +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( @@ -2738,8 +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 */ @@ -2752,7 +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); @@ -2762,7 +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); @@ -2772,8 +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; @@ -2857,17 +2990,14 @@ kiblnd_cm_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event) } } -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; - - spin_lock(&conn->ibc_lock); + cfs_list_t *ttmp; - 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); @@ -2876,147 +3006,174 @@ kiblnd_check_txs (kib_conn_t *conn, struct list_head *txs) LASSERT (tx->tx_waiting || tx->tx_sending != 0); } - if (time_after_eq (jiffies, tx->tx_deadline)) { - timed_out = 1; + 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)); - break; + 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); - list_for_each (ptmp, peers) { - peer = list_entry (ptmp, kib_peer_t, ibp_list); + cfs_list_for_each (ptmp, peers) { + peer = cfs_list_entry (ptmp, kib_peer_t, ibp_list); - list_for_each (ctmp, &peer->ibp_conns) { - conn = list_entry (ctmp, kib_conn_t, ibc_list); + cfs_list_for_each (ctmp, &peer->ibp_conns) { + int timedout; + int sendnoop; + + 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); - - CERROR("Timed out RDMA with %s (%lu)\n", - libcfs_nid2str(peer->ibp_nid), - cfs_duration_sec(cfs_time_current() - - peer->ibp_last_alive)); - - kiblnd_close_conn(conn, -ETIMEDOUT); - kiblnd_conn_decref(conn); /* ...until here */ - - /* start again now I've dropped the lock */ - goto again; - } + 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); + + spin_unlock(&conn->ibc_lock); + } + } + + read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); + + /* 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... */ @@ -3026,7 +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 @@ -3043,35 +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 @@ -3084,7 +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); @@ -3106,8 +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: @@ -3122,33 +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); + /* 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; - spin_lock_irqsave(&kiblnd_data.kib_sched_lock, flags); + LASSERT(cq == conn->ibc_cq); - conn->ibc_ready = 1; + spin_lock_irqsave(&sched->ibs_lock, flags); - 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); - } + 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 @@ -3163,47 +3322,53 @@ 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; + 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; - snprintf(name, sizeof(name), "kiblnd_sd_%02ld", id); - cfs_daemonize(name); - cfs_block_allsigs(); + cfs_block_allsigs(); - init_waitqueue_entry(&wait, current); + init_waitqueue_entry_current(&wait); - spin_lock_irqsave(&kiblnd_data.kib_sched_lock, flags); + sched = kiblnd_data.kib_scheds[KIB_THREAD_CPT(id)]; - while (!kiblnd_data.kib_shutdown) { - if (busy_loops++ >= IBLND_RESCHED) { - spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock, - flags); + 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); + } - our_cond_resched(); - busy_loops = 0; + 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); + + cond_resched(); + busy_loops = 0; + + spin_lock_irqsave(&sched->ibs_lock, flags); + } - did_something = 0; + did_something = 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; + 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; - 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) { @@ -3215,69 +3380,159 @@ kiblnd_scheduler(void *arg) libcfs_nid2str(conn->ibc_peer->ibp_nid), rc); kiblnd_close_conn(conn, -EIO); kiblnd_conn_decref(conn); - spin_lock_irqsave(&kiblnd_data.kib_sched_lock, flags); - continue; - } - - rc = ib_poll_cq(conn->ibc_cq, 1, &wc); + 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); } - 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(&kiblnd_data.kib_sched_lock, flags); + kiblnd_conn_decref(conn); /* ...drop my ref from above */ + did_something = 1; + } + + if (did_something) + continue; + + set_current_state(TASK_INTERRUPTIBLE); + add_wait_queue_exclusive(&sched->ibs_waitq, &wait); + spin_unlock_irqrestore(&sched->ibs_lock, flags); + + waitq_wait(&wait, TASK_INTERRUPTIBLE); + busy_loops = 0; + + remove_wait_queue(&sched->ibs_waitq, &wait); + set_current_state(TASK_RUNNING); + spin_lock_irqsave(&sched->ibs_lock, flags); + } + + 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; + } - spin_lock_irqsave(&kiblnd_data.kib_sched_lock, - flags); - - 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); + 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); - kiblnd_complete(&wc); + write_lock_irqsave(glock, flags); - spin_lock_irqsave(&kiblnd_data.kib_sched_lock, - 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; } - kiblnd_conn_decref(conn); /* ...drop my ref from above */ - did_something = 1; - } + /* 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); + } - if (did_something) continue; + } - set_current_state(TASK_INTERRUPTIBLE); - add_wait_queue_exclusive(&kiblnd_data.kib_sched_waitq, &wait); - spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock, flags); + /* long sleep if no more pending failover */ + long_sleep = cfs_list_empty(&kiblnd_data.kib_failed_devs); - schedule(); - busy_loops = 0; + set_current_state(TASK_INTERRUPTIBLE); + add_wait_queue(&kiblnd_data.kib_failover_waitq, &wait); + write_unlock_irqrestore(glock, flags); - remove_wait_queue(&kiblnd_data.kib_sched_waitq, &wait); - set_current_state(TASK_RUNNING); - spin_lock_irqsave(&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); + + if (!long_sleep || rc != 0) + continue; + + /* 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; }