X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lnet%2Fklnds%2Fo2iblnd%2Fo2iblnd_cb.c;h=41b9cdb6ac19bc7dfe35a8b3a03adb1059d2f52e;hp=de402e73203053d2483f5e3fe6f466506b3bea56;hb=fa7276592a0cc12d651a5b2bd39dd7ca45bc25bf;hpb=8553494bca7381262843f900d14c58f5c242f3ea diff --git a/lnet/klnds/o2iblnd/o2iblnd_cb.c b/lnet/klnds/o2iblnd/o2iblnd_cb.c index de402e7..41b9cdb 100644 --- a/lnet/klnds/o2iblnd/o2iblnd_cb.c +++ b/lnet/klnds/o2iblnd/o2iblnd_cb.c @@ -15,11 +15,7 @@ * * 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 - * - * 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. + * http://www.gnu.org/licenses/gpl-2.0.html * * GPL HEADER END */ @@ -27,7 +23,7 @@ * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. * - * Copyright (c) 2012, Intel Corporation. + * Copyright (c) 2012, 2016, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -40,17 +36,20 @@ #include "o2iblnd.h" +#define MAX_CONN_RACES_BEFORE_ABORT 20 + static void kiblnd_peer_alive(kib_peer_t *peer); static void kiblnd_peer_connect_failed(kib_peer_t *peer, int active, int error); -static void kiblnd_check_sends(kib_conn_t *conn); static void kiblnd_init_tx_msg(lnet_ni_t *ni, kib_tx_t *tx, int type, int body_nob); static int kiblnd_init_rdma(kib_conn_t *conn, kib_tx_t *tx, int type, int resid, kib_rdma_desc_t *dstrd, __u64 dstcookie); static void kiblnd_queue_tx_locked(kib_tx_t *tx, kib_conn_t *conn); static void kiblnd_queue_tx(kib_tx_t *tx, kib_conn_t *conn); +static void kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx); +static void kiblnd_check_sends_locked(kib_conn_t *conn); -static void +void kiblnd_tx_done (lnet_ni_t *ni, kib_tx_t *tx) { lnet_msg_t *lntmsg[2]; @@ -117,7 +116,7 @@ kiblnd_get_idle_tx(lnet_ni_t *ni, lnet_nid_t target) kib_tx_t *tx; kib_tx_poolset_t *tps; - tps = net->ibn_tx_ps[lnet_cpt_of_nid(target)]; + tps = net->ibn_tx_ps[lnet_cpt_of_nid(target, ni)]; node = kiblnd_pool_alloc_node(&tps->tps_poolset); if (node == NULL) return NULL; @@ -131,7 +130,6 @@ kiblnd_get_idle_tx(lnet_ni_t *ni, lnet_nid_t target) LASSERT (tx->tx_conn == NULL); LASSERT (tx->tx_lntmsg[0] == NULL); LASSERT (tx->tx_lntmsg[1] == NULL); - LASSERT (tx->tx_u.pmr == NULL); LASSERT (tx->tx_nfrags == 0); return tx; @@ -158,7 +156,7 @@ 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; + struct ib_mr *mr = conn->ibc_hdev->ibh_mrs; int rc; LASSERT (net != NULL); @@ -166,9 +164,7 @@ kiblnd_post_rx (kib_rx_t *rx, int credit) 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); + LASSERT(mr != NULL); rx->rx_sge.lkey = mr->lkey; rx->rx_sge.addr = rx->rx_msgaddr; @@ -189,34 +185,40 @@ 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; - } + /* NB: need an extra reference after ib_post_recv because we don't + * own this rx (and rx::rx_conn) anymore, LU-5678. + */ + kiblnd_conn_addref(conn); + rc = ib_post_recv(conn->ibc_cmid->qp, &rx->rx_wrq, &bad_wrq); + if (unlikely(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 (conn->ibc_state < IBLND_CONN_ESTABLISHED) /* Initial post */ + goto out; - if (rc != 0) { - kiblnd_close_conn(conn, rc); - kiblnd_drop_rx(rx); /* No more posts for this rx */ - return rc; - } + if (unlikely(rc != 0)) { + kiblnd_close_conn(conn, rc); + kiblnd_drop_rx(rx); /* No more posts for this rx */ + goto out; + } - if (credit == IBLND_POSTRX_NO_CREDIT) - return 0; + if (credit == IBLND_POSTRX_NO_CREDIT) + goto out; spin_lock(&conn->ibc_lock); if (credit == IBLND_POSTRX_PEER_CREDIT) conn->ibc_outstanding_credits++; else conn->ibc_reserved_credits++; + kiblnd_check_sends_locked(conn); spin_unlock(&conn->ibc_lock); - kiblnd_check_sends(conn); - return 0; +out: + kiblnd_conn_decref(conn); + return rc; } static kib_tx_t * @@ -257,7 +259,7 @@ kiblnd_handle_completion(kib_conn_t *conn, int txtype, int status, __u64 cookie) if (tx == NULL) { spin_unlock(&conn->ibc_lock); - CWARN("Unmatched completion type %x cookie "LPX64" from %s\n", + CWARN("Unmatched completion type %x cookie %#llx from %s\n", txtype, cookie, libcfs_nid2str(conn->ibc_peer->ibp_nid)); kiblnd_close_conn(conn, -EPROTO); return; @@ -325,19 +327,19 @@ kiblnd_handle_rx (kib_rx_t *rx) spin_lock(&conn->ibc_lock); if (conn->ibc_credits + credits > - IBLND_MSG_QUEUE_SIZE(conn->ibc_version)) { + conn->ibc_queue_depth) { 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)); + CERROR("Bad credits from %s: %d + %d > %d\n", + libcfs_nid2str(conn->ibc_peer->ibp_nid), + rc2, credits, + conn->ibc_queue_depth); - kiblnd_close_conn(conn, -EPROTO); - kiblnd_post_rx(rx, IBLND_POSTRX_NO_CREDIT); - return; - } + kiblnd_close_conn(conn, -EPROTO); + kiblnd_post_rx(rx, IBLND_POSTRX_NO_CREDIT); + return; + } conn->ibc_credits += credits; @@ -346,8 +348,8 @@ kiblnd_handle_rx (kib_rx_t *rx) !IBLND_OOB_CAPABLE(conn->ibc_version)) /* v1 only */ conn->ibc_outstanding_credits++; + kiblnd_check_sends_locked(conn); spin_unlock(&conn->ibc_lock); - kiblnd_check_sends(conn); } switch (msg->ibm_type) { @@ -558,56 +560,10 @@ 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) +kiblnd_fmr_map_tx(kib_net_t *net, kib_tx_t *tx, kib_rdma_desc_t *rd, __u32 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; @@ -615,44 +571,34 @@ kiblnd_pmr_map_tx(kib_net_t *net, kib_tx_t *tx, kib_rdma_desc_t *rd, int nob) 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; - } + fps = net->ibn_fmr_ps[cpt]; + rc = kiblnd_fmr_pool_map(fps, tx, rd, nob, 0, &tx->fmr); + if (rc != 0) { + CERROR("Can't map %u pages: %d\n", nob, 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; + /* If rd is not tx_rd, it's going to get sent to a peer, who will need + * the rkey */ + rd->rd_key = tx->fmr.fmr_key; + rd->rd_frags[0].rf_addr &= ~hdev->ibh_page_mask; + rd->rd_frags[0].rf_nob = nob; + rd->rd_nfrags = 1; - return 0; + return 0; } -void +static 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 (net->ibn_fmr_ps != NULL) + kiblnd_fmr_pool_unmap(&tx->fmr, tx->tx_status); if (tx->tx_nfrags != 0) { kiblnd_dma_unmap_sg(tx->tx_pool->tpo_hdev->ibh_ibdev, @@ -661,24 +607,22 @@ kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx) } } -int -kiblnd_map_tx(lnet_ni_t *ni, kib_tx_t *tx, - kib_rdma_desc_t *rd, int nfrags) +static 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; - int i; + kib_net_t *net = ni->ni_data; + kib_hca_dev_t *hdev = net->ibn_dev->ibd_hdev; + 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; + tx->tx_nfrags = nfrags; - rd->rd_nfrags = - kiblnd_dma_map_sg(hdev->ibh_ibdev, - tx->tx_frags, tx->tx_nfrags, tx->tx_dmadir); + 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( @@ -688,18 +632,17 @@ kiblnd_map_tx(lnet_ni_t *ni, kib_tx_t *tx, 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; - } + mr = kiblnd_find_rd_dma_mr(ni, rd, + (tx->tx_conn != NULL) ? + tx->tx_conn->ibc_max_frags : -1); + 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; } @@ -707,7 +650,7 @@ kiblnd_map_tx(lnet_ni_t *ni, kib_tx_t *tx, static 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) + unsigned int niov, struct kvec *iov, int offset, int nob) { kib_net_t *net = ni->ni_data; struct page *page; @@ -743,7 +686,11 @@ kiblnd_setup_rd_iov(lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd, fragnob = min(fragnob, (int)PAGE_SIZE - page_offset); sg_set_page(sg, page, fragnob, page_offset); - sg++; + sg = sg_next(sg); + if (!sg) { + CERROR("lacking enough sg entries to map tx\n"); + return -EFAULT; + } if (offset + fragnob < iov->iov_len) { offset += fragnob; @@ -785,9 +732,13 @@ 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); - sg_set_page(sg, kiov->kiov_page, fragnob, - kiov->kiov_offset + offset); - sg++; + sg_set_page(sg, kiov->kiov_page, fragnob, + kiov->kiov_offset + offset); + sg = sg_next(sg); + if (!sg) { + CERROR("lacking enough sg entries to map tx\n"); + return -EFAULT; + } offset = 0; kiov++; @@ -804,23 +755,24 @@ __must_hold(&conn->ibc_lock) { kib_msg_t *msg = tx->tx_msg; kib_peer_t *peer = conn->ibc_peer; + struct lnet_ni *ni = peer->ibp_ni; int ver = conn->ibc_version; int rc; int done; - struct ib_send_wr *bad_wrq; - 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)); + LASSERT(tx->tx_queued); + /* We rely on this for QP sizing */ + LASSERT(tx->tx_nwrq > 0); + LASSERT(tx->tx_nwrq <= 1 + conn->ibc_max_frags); - 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)); + LASSERT(credit == 0 || credit == 1); + LASSERT(conn->ibc_outstanding_credits >= 0); + LASSERT(conn->ibc_outstanding_credits <= conn->ibc_queue_depth); + LASSERT(conn->ibc_credits >= 0); + LASSERT(conn->ibc_credits <= conn->ibc_queue_depth); - if (conn->ibc_nsends_posted == IBLND_CONCURRENT_SENDS(ver)) { + if (conn->ibc_nsends_posted == + kiblnd_concurrent_sends(ver, ni)) { /* tx completions outstanding... */ CDEBUG(D_NET, "%s: posted enough\n", libcfs_nid2str(peer->ibp_nid)); @@ -849,9 +801,9 @@ __must_hold(&conn->ibc_lock) (!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 */ + /* OK to drop when posted enough NOOPs, since + * kiblnd_check_sends_locked 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); @@ -887,9 +839,28 @@ __must_hold(&conn->ibc_lock) /* close_conn will launch failover */ rc = -ENETDOWN; } else { - rc = ib_post_send(conn->ibc_cmid->qp, - tx->tx_wrq, &bad_wrq); - } + struct kib_fast_reg_descriptor *frd = tx->fmr.fmr_frd; + struct ib_send_wr *bad = &tx->tx_wrq[tx->tx_nwrq - 1].wr; + struct ib_send_wr *wr = &tx->tx_wrq[0].wr; + + if (frd != NULL) { + if (!frd->frd_valid) { + wr = &frd->frd_inv_wr.wr; + wr->next = &frd->frd_fastreg_wr.wr; + } else { + wr = &frd->frd_fastreg_wr.wr; + } + frd->frd_fastreg_wr.wr.next = &tx->tx_wrq[0].wr; + } + + LASSERTF(bad->wr_id == kiblnd_ptr2wreqid(tx, IBLND_WID_TX), + "bad wr_id %#llx, opc %d, flags %d, peer: %s\n", + bad->wr_id, bad->opcode, bad->send_flags, + libcfs_nid2str(conn->ibc_peer->ibp_nid)); + + bad = NULL; + rc = ib_post_send(conn->ibc_cmid->qp, wr, &bad); + } conn->ibc_last_send = jiffies; @@ -932,7 +903,7 @@ __must_hold(&conn->ibc_lock) } static void -kiblnd_check_sends (kib_conn_t *conn) +kiblnd_check_sends_locked(kib_conn_t *conn) { int ver = conn->ibc_version; lnet_ni_t *ni = conn->ibc_peer->ibp_ni; @@ -945,9 +916,8 @@ kiblnd_check_sends (kib_conn_t *conn) return; } - spin_lock(&conn->ibc_lock); - - LASSERT (conn->ibc_nsends_posted <= IBLND_CONCURRENT_SENDS(ver)); + LASSERT(conn->ibc_nsends_posted <= + kiblnd_concurrent_sends(ver, ni)); LASSERT (!IBLND_OOB_CAPABLE(ver) || conn->ibc_noops_posted <= IBLND_OOB_MSGS(ver)); LASSERT (conn->ibc_reserved_credits >= 0); @@ -973,8 +943,6 @@ kiblnd_check_sends (kib_conn_t *conn) kiblnd_queue_tx_locked(tx, conn); } - kiblnd_conn_addref(conn); /* 1 ref for me.... (see b21911) */ - for (;;) { int credit; @@ -997,10 +965,6 @@ kiblnd_check_sends (kib_conn_t *conn) if (kiblnd_post_tx_locked(conn, tx, credit) != 0) break; } - - spin_unlock(&conn->ibc_lock); - - kiblnd_conn_decref(conn); /* ...until here */ } static void @@ -1014,7 +978,7 @@ kiblnd_tx_complete (kib_tx_t *tx, int status) if (failed) { if (conn->ibc_state == IBLND_CONN_ESTABLISHED) - CNETERR("Tx -> %s cookie "LPX64 + CNETERR("Tx -> %s cookie %#llx" " sending %d waiting %d: failed %d\n", libcfs_nid2str(conn->ibc_peer->ibp_nid), tx->tx_cookie, tx->tx_sending, tx->tx_waiting, @@ -1046,50 +1010,44 @@ kiblnd_tx_complete (kib_tx_t *tx, int status) if (idle) list_del(&tx->tx_list); - kiblnd_conn_addref(conn); /* 1 ref for me.... */ - + kiblnd_check_sends_locked(conn); spin_unlock(&conn->ibc_lock); if (idle) kiblnd_tx_done(conn->ibc_peer->ibp_ni, tx); - - kiblnd_check_sends(conn); - - kiblnd_conn_decref(conn); /* ...until here */ } static void kiblnd_init_tx_msg (lnet_ni_t *ni, kib_tx_t *tx, int type, int body_nob) { - 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; + kib_hca_dev_t *hdev = tx->tx_pool->tpo_hdev; + struct ib_sge *sge = &tx->tx_sge[tx->tx_nwrq]; + struct ib_rdma_wr *wrq; + int nob = offsetof(kib_msg_t, ibm_u) + body_nob; + struct ib_mr *mr = hdev->ibh_mrs; - LASSERT (tx->tx_nwrq >= 0); - LASSERT (tx->tx_nwrq < IBLND_MAX_RDMA_FRAGS + 1); - LASSERT (nob <= IBLND_MSG_SIZE); + LASSERT(tx->tx_nwrq >= 0); + LASSERT(tx->tx_nwrq < IBLND_MAX_RDMA_FRAGS + 1); + LASSERT(nob <= IBLND_MSG_SIZE); + LASSERT(mr != NULL); kiblnd_init_msg(tx->tx_msg, type, body_nob); - 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)); + wrq = &tx->tx_wrq[tx->tx_nwrq]; + memset(wrq, 0, sizeof(*wrq)); - wrq->next = NULL; - wrq->wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_TX); - wrq->sg_list = sge; - wrq->num_sge = 1; - wrq->opcode = IB_WR_SEND; - wrq->send_flags = IB_SEND_SIGNALED; + wrq->wr.next = NULL; + wrq->wr.wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_TX); + wrq->wr.sg_list = sge; + wrq->wr.num_sge = 1; + wrq->wr.opcode = IB_WR_SEND; + wrq->wr.send_flags = IB_SEND_SIGNALED; - tx->tx_nwrq++; + tx->tx_nwrq++; } static int @@ -1099,7 +1057,7 @@ kiblnd_init_rdma(kib_conn_t *conn, kib_tx_t *tx, int type, 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]; + struct ib_rdma_wr *wrq; int rc = resid; int srcidx; int dstidx; @@ -1125,16 +1083,16 @@ kiblnd_init_rdma(kib_conn_t *conn, kib_tx_t *tx, int type, break; } - 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; - } + if (tx->tx_nwrq >= conn->ibc_max_frags) { + CERROR("RDMA has too many fragments for peer %s (%d), " + "src idx/frags: %d/%d dst idx/frags: %d/%d\n", + libcfs_nid2str(conn->ibc_peer->ibp_nid), + conn->ibc_max_frags, + srcidx, srcrd->rd_nfrags, + dstidx, dstrd->rd_nfrags); + rc = -EMSGSIZE; + break; + } wrknob = MIN(MIN(kiblnd_rd_frag_size(srcrd, srcidx), kiblnd_rd_frag_size(dstrd, dstidx)), resid); @@ -1146,15 +1104,20 @@ kiblnd_init_rdma(kib_conn_t *conn, kib_tx_t *tx, int type, wrq = &tx->tx_wrq[tx->tx_nwrq]; - wrq->next = 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 = kiblnd_rd_frag_addr(dstrd, dstidx); - wrq->wr.rdma.rkey = kiblnd_rd_frag_key(dstrd, dstidx); + wrq->wr.next = &(wrq + 1)->wr; + wrq->wr.wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_RDMA); + wrq->wr.sg_list = sge; + wrq->wr.num_sge = 1; + wrq->wr.opcode = IB_WR_RDMA_WRITE; + wrq->wr.send_flags = 0; + +#ifdef HAVE_IB_RDMA_WR + wrq->remote_addr = kiblnd_rd_frag_addr(dstrd, dstidx); + wrq->rkey = kiblnd_rd_frag_key(dstrd, dstidx); +#else + wrq->wr.wr.rdma.remote_addr = kiblnd_rd_frag_addr(dstrd, dstidx); + wrq->wr.wr.rdma.rkey = kiblnd_rd_frag_key(dstrd, dstidx); +#endif srcidx = kiblnd_rd_consume_frag(srcrd, srcidx, wrknob); dstidx = kiblnd_rd_consume_frag(dstrd, dstidx, wrknob); @@ -1187,7 +1150,9 @@ kiblnd_queue_tx_locked(kib_tx_t *tx, kib_conn_t *conn) LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED); tx->tx_queued = 1; - tx->tx_deadline = jiffies + (*kiblnd_tunables.kib_timeout * HZ); + tx->tx_deadline = jiffies + + msecs_to_jiffies(*kiblnd_tunables.kib_timeout * + MSEC_PER_SEC); if (tx->tx_conn == NULL) { kiblnd_conn_addref(conn); @@ -1235,9 +1200,8 @@ kiblnd_queue_tx (kib_tx_t *tx, kib_conn_t *conn) { spin_lock(&conn->ibc_lock); kiblnd_queue_tx_locked(tx, conn); + kiblnd_check_sends_locked(conn); spin_unlock(&conn->ibc_lock); - - kiblnd_check_sends(conn); } static int kiblnd_resolve_addr(struct rdma_cm_id *cmid, @@ -1248,14 +1212,12 @@ static int kiblnd_resolve_addr(struct rdma_cm_id *cmid, 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--) { @@ -1276,9 +1238,6 @@ static int kiblnd_resolve_addr(struct rdma_cm_id *cmid, } 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; } @@ -1294,6 +1253,7 @@ kiblnd_connect_peer (kib_peer_t *peer) LASSERT (net != NULL); LASSERT (peer->ibp_connecting > 0); + LASSERT(!peer->ibp_reconnecting); cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, peer, RDMA_PS_TCP, IB_QPT_RC); @@ -1334,17 +1294,69 @@ kiblnd_connect_peer (kib_peer_t *peer) } LASSERT (cmid->device != NULL); - CDEBUG(D_NET, "%s: connection bound to %s:%u.%u.%u.%u:%s\n", + CDEBUG(D_NET, "%s: connection bound to %s:%pI4h:%s\n", libcfs_nid2str(peer->ibp_nid), dev->ibd_ifname, - HIPQUAD(dev->ibd_ifip), cmid->device->name); + &dev->ibd_ifip, cmid->device->name); - return; + return; failed2: - kiblnd_peer_decref(peer); /* cmid's ref */ - rdma_destroy_id(cmid); + kiblnd_peer_connect_failed(peer, 1, rc); + kiblnd_peer_decref(peer); /* cmid's ref */ + rdma_destroy_id(cmid); + return; failed: - kiblnd_peer_connect_failed(peer, 1, rc); + kiblnd_peer_connect_failed(peer, 1, rc); +} + +bool +kiblnd_reconnect_peer(kib_peer_t *peer) +{ + rwlock_t *glock = &kiblnd_data.kib_global_lock; + char *reason = NULL; + struct list_head txs; + unsigned long flags; + + INIT_LIST_HEAD(&txs); + + write_lock_irqsave(glock, flags); + if (peer->ibp_reconnecting == 0) { + if (peer->ibp_accepting) + reason = "accepting"; + else if (peer->ibp_connecting) + reason = "connecting"; + else if (!list_empty(&peer->ibp_conns)) + reason = "connected"; + else /* connected then closed */ + reason = "closed"; + + goto no_reconnect; + } + + LASSERT(!peer->ibp_accepting && !peer->ibp_connecting && + list_empty(&peer->ibp_conns)); + peer->ibp_reconnecting = 0; + + if (!kiblnd_peer_active(peer)) { + list_splice_init(&peer->ibp_tx_queue, &txs); + reason = "unlinked"; + goto no_reconnect; + } + + peer->ibp_connecting++; + peer->ibp_reconnected++; + write_unlock_irqrestore(glock, flags); + + kiblnd_connect_peer(peer); + return true; + + no_reconnect: + write_unlock_irqrestore(glock, flags); + + CWARN("Abort reconnection of %s: %s\n", + libcfs_nid2str(peer->ibp_nid), reason); + kiblnd_txlist_done(peer->ibp_ni, &txs, -ECONNABORTED); + return false; } void @@ -1367,7 +1379,7 @@ kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid) * connected */ read_lock_irqsave(g_lock, flags); - peer = kiblnd_find_peer_locked(nid); + peer = kiblnd_find_peer_locked(ni, nid); if (peer != NULL && !list_empty(&peer->ibp_conns)) { /* Found a peer with an established connection */ conn = kiblnd_get_conn_locked(peer); @@ -1385,12 +1397,11 @@ kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid) /* Re-try with a write lock */ write_lock(g_lock); - peer = kiblnd_find_peer_locked(nid); + peer = kiblnd_find_peer_locked(ni, nid); if (peer != NULL) { if (list_empty(&peer->ibp_conns)) { /* found a peer, but it's still connecting... */ - LASSERT (peer->ibp_connecting != 0 || - peer->ibp_accepting != 0); + LASSERT(kiblnd_peer_connecting(peer)); if (tx != NULL) list_add_tail(&tx->tx_list, &peer->ibp_tx_queue); @@ -1410,26 +1421,25 @@ kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid) 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)); - if (tx != NULL) { - tx->tx_status = -EHOSTUNREACH; - tx->tx_waiting = 0; - kiblnd_tx_done(ni, tx); - } - return; - } + /* 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)); + if (tx != NULL) { + tx->tx_status = -EHOSTUNREACH; + tx->tx_waiting = 0; + kiblnd_tx_done(ni, tx); + } + return; + } write_lock_irqsave(g_lock, flags); - peer2 = kiblnd_find_peer_locked(nid); + peer2 = kiblnd_find_peer_locked(ni, nid); if (peer2 != NULL) { if (list_empty(&peer2->ibp_conns)) { /* found a peer, but it's still connecting... */ - LASSERT (peer2->ibp_connecting != 0 || - peer2->ibp_accepting != 0); + LASSERT(kiblnd_peer_connecting(peer2)); if (tx != NULL) list_add_tail(&tx->tx_list, &peer2->ibp_tx_queue); @@ -1472,19 +1482,20 @@ int kiblnd_send (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg) { lnet_hdr_t *hdr = &lntmsg->msg_hdr; - int type = lntmsg->msg_type; - lnet_process_id_t target = lntmsg->msg_target; - int target_is_router = lntmsg->msg_target_is_router; - int routing = lntmsg->msg_routing; - unsigned int payload_niov = lntmsg->msg_niov; - struct iovec *payload_iov = lntmsg->msg_iov; - lnet_kiov_t *payload_kiov = lntmsg->msg_kiov; - unsigned int payload_offset = lntmsg->msg_offset; - unsigned int payload_nob = lntmsg->msg_len; - kib_msg_t *ibmsg; - kib_tx_t *tx; - int nob; - int rc; + int type = lntmsg->msg_type; + lnet_process_id_t target = lntmsg->msg_target; + int target_is_router = lntmsg->msg_target_is_router; + int routing = lntmsg->msg_routing; + unsigned int payload_niov = lntmsg->msg_niov; + struct kvec *payload_iov = lntmsg->msg_iov; + lnet_kiov_t *payload_kiov = lntmsg->msg_kiov; + unsigned int payload_offset = lntmsg->msg_offset; + unsigned int payload_nob = lntmsg->msg_len; + kib_msg_t *ibmsg; + kib_rdma_desc_t *rd; + kib_tx_t *tx; + int nob; + int rc; /* NB 'private' is different depending on what we're sending.... */ @@ -1524,30 +1535,28 @@ kiblnd_send (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg) return -ENOMEM; } - ibmsg = tx->tx_msg; - - if ((lntmsg->msg_md->md_options & LNET_MD_KIOV) == 0) - rc = kiblnd_setup_rd_iov(ni, tx, - &ibmsg->ibm_u.get.ibgm_rd, - lntmsg->msg_md->md_niov, - lntmsg->msg_md->md_iov.iov, - 0, lntmsg->msg_md->md_length); - else - rc = kiblnd_setup_rd_kiov(ni, tx, - &ibmsg->ibm_u.get.ibgm_rd, - lntmsg->msg_md->md_niov, - lntmsg->msg_md->md_iov.kiov, - 0, lntmsg->msg_md->md_length); - if (rc != 0) { - CERROR("Can't setup GET sink for %s: %d\n", - libcfs_nid2str(target.nid), rc); - kiblnd_tx_done(ni, tx); - return -EIO; - } + ibmsg = tx->tx_msg; + rd = &ibmsg->ibm_u.get.ibgm_rd; + if ((lntmsg->msg_md->md_options & LNET_MD_KIOV) == 0) + rc = kiblnd_setup_rd_iov(ni, tx, rd, + lntmsg->msg_md->md_niov, + lntmsg->msg_md->md_iov.iov, + 0, lntmsg->msg_md->md_length); + else + rc = kiblnd_setup_rd_kiov(ni, tx, rd, + lntmsg->msg_md->md_niov, + lntmsg->msg_md->md_iov.kiov, + 0, lntmsg->msg_md->md_length); + if (rc != 0) { + CERROR("Can't setup GET sink for %s: %d\n", + libcfs_nid2str(target.nid), rc); + kiblnd_tx_done(ni, tx); + return -EIO; + } - nob = offsetof(kib_get_msg_t, ibgm_rd.rd_frags[tx->tx_nfrags]); - ibmsg->ibm_u.get.ibgm_cookie = tx->tx_cookie; - ibmsg->ibm_u.get.ibgm_hdr = *hdr; + nob = offsetof(kib_get_msg_t, ibgm_rd.rd_frags[rd->rd_nfrags]); + 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); @@ -1644,7 +1653,7 @@ kiblnd_reply (lnet_ni_t *ni, kib_rx_t *rx, lnet_msg_t *lntmsg) { lnet_process_id_t target = lntmsg->msg_target; unsigned int niov = lntmsg->msg_niov; - struct iovec *iov = lntmsg->msg_iov; + struct kvec *iov = lntmsg->msg_iov; lnet_kiov_t *kiov = lntmsg->msg_kiov; unsigned int offset = lntmsg->msg_offset; unsigned int nob = lntmsg->msg_len; @@ -1702,15 +1711,14 @@ kiblnd_reply (lnet_ni_t *ni, kib_rx_t *rx, lnet_msg_t *lntmsg) } int -kiblnd_recv (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed, - unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov, - unsigned int offset, unsigned int mlen, unsigned int rlen) +kiblnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed, + unsigned int niov, struct kvec *iov, lnet_kiov_t *kiov, + unsigned int offset, unsigned int mlen, unsigned int rlen) { kib_rx_t *rx = private; kib_msg_t *rxmsg = rx->rx_msg; 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; @@ -1747,7 +1755,10 @@ kiblnd_recv (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed, lnet_finalize (ni, lntmsg, 0); break; - case IBLND_MSG_PUT_REQ: + case IBLND_MSG_PUT_REQ: { + kib_msg_t *txmsg; + kib_rdma_desc_t *rd; + if (mlen == 0) { lnet_finalize(ni, lntmsg, 0); kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK, 0, @@ -1764,28 +1775,27 @@ kiblnd_recv (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed, break; } - txmsg = tx->tx_msg; - if (kiov == NULL) - rc = kiblnd_setup_rd_iov(ni, tx, - &txmsg->ibm_u.putack.ibpam_rd, - niov, iov, offset, mlen); - else - rc = kiblnd_setup_rd_kiov(ni, tx, - &txmsg->ibm_u.putack.ibpam_rd, - niov, kiov, offset, mlen); - if (rc != 0) { - CERROR("Can't setup PUT sink for %s: %d\n", - libcfs_nid2str(conn->ibc_peer->ibp_nid), rc); - kiblnd_tx_done(ni, tx); - /* tell peer it's over */ - kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK, rc, - rxmsg->ibm_u.putreq.ibprm_cookie); - break; - } + txmsg = tx->tx_msg; + rd = &txmsg->ibm_u.putack.ibpam_rd; + if (kiov == NULL) + rc = kiblnd_setup_rd_iov(ni, tx, rd, + niov, iov, offset, mlen); + else + rc = kiblnd_setup_rd_kiov(ni, tx, rd, + niov, kiov, offset, mlen); + if (rc != 0) { + CERROR("Can't setup PUT sink for %s: %d\n", + libcfs_nid2str(conn->ibc_peer->ibp_nid), rc); + kiblnd_tx_done(ni, tx); + /* tell peer it's over */ + kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK, rc, + rxmsg->ibm_u.putreq.ibprm_cookie); + 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; + nob = offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[rd->rd_nfrags]); + txmsg->ibm_u.putack.ibpam_src_cookie = rxmsg->ibm_u.putreq.ibprm_cookie; + txmsg->ibm_u.putack.ibpam_dst_cookie = tx->tx_cookie; kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_ACK, nob); @@ -1796,6 +1806,7 @@ kiblnd_recv (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed, /* reposted buffer reserved for PUT_DONE */ post_credit = IBLND_POSTRX_NO_CREDIT; break; + } case IBLND_MSG_GET_REQ: if (lntmsg != NULL) { @@ -1849,10 +1860,7 @@ kiblnd_peer_notify (kib_peer_t *peer) read_lock_irqsave(&kiblnd_data.kib_global_lock, flags); - if (list_empty(&peer->ibp_conns) && - peer->ibp_accepting == 0 && - peer->ibp_connecting == 0 && - peer->ibp_error != 0) { + if (kiblnd_peer_idle(peer) && peer->ibp_error != 0) { error = peer->ibp_error; peer->ibp_error = 0; @@ -2051,14 +2059,14 @@ kiblnd_peer_connect_failed(kib_peer_t *peer, int active, int error) peer->ibp_accepting--; } - if (peer->ibp_connecting != 0 || - peer->ibp_accepting != 0) { + if (kiblnd_peer_connecting(peer)) { /* another connection attempt under way... */ write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); return; } + peer->ibp_reconnected = 0; if (list_empty(&peer->ibp_conns)) { /* Take peer's blocked transmits to complete with error */ list_add(&zombies, &peer->ibp_tx_queue); @@ -2128,6 +2136,7 @@ kiblnd_connreq_done(kib_conn_t *conn, int status) * peer instance... */ kiblnd_conn_addref(conn); /* +1 ref for ibc_list */ list_add(&conn->ibc_list, &peer->ibp_conns); + peer->ibp_reconnected = 0; if (active) peer->ibp_connecting--; else @@ -2163,6 +2172,12 @@ kiblnd_connreq_done(kib_conn_t *conn, int status) return; } + /* +1 ref for myself, this connection is visible to other threads + * now, refcount of peer:ibp_conns can be released by connection + * close from either a different thread, or the calling of + * kiblnd_check_sends_locked() below. See bz21911 for details. + */ + kiblnd_conn_addref(conn); write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); /* Schedule blocked txs */ @@ -2173,12 +2188,12 @@ kiblnd_connreq_done(kib_conn_t *conn, int status) kiblnd_queue_tx_locked(tx, conn); } + kiblnd_check_sends_locked(conn); spin_unlock(&conn->ibc_lock); - kiblnd_check_sends(conn); - /* schedule blocked rxs */ kiblnd_handle_early_rxs(conn); + kiblnd_conn_decref(conn); } static void @@ -2193,7 +2208,7 @@ kiblnd_reject(struct rdma_cm_id *cmid, kib_rej_t *rej) } static int -kiblnd_passive_connect (struct rdma_cm_id *cmid, void *priv, int priv_nob) +kiblnd_passive_connect(struct rdma_cm_id *cmid, void *priv, int priv_nob) { rwlock_t *g_lock = &kiblnd_data.kib_global_lock; kib_msg_t *reqmsg = priv; @@ -2225,105 +2240,120 @@ kiblnd_passive_connect (struct rdma_cm_id *cmid, void *priv, int priv_nob) 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)) - 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; + __u32 ip = ntohl(peer_addr->sin_addr.s_addr); + CERROR("Peer's port (%pI4h:%hu) is not privileged\n", + &ip, ntohs(peer_addr->sin_port)); + goto failed; + } - rc = kiblnd_unpack_msg(reqmsg, priv_nob); - if (rc != 0) { - CERROR("Can't parse connection request: %d\n", rc); - goto failed; - } + if (priv_nob < offsetof(kib_msg_t, ibm_type)) { + CERROR("Short connection request\n"); + goto failed; + } - nid = reqmsg->ibm_srcnid; - ni = lnet_net2ni(LNET_NIDNET(reqmsg->ibm_dstnid)); + /* 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)) + 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) { + CERROR("Can't parse connection request: %d\n", rc); + goto failed; + } - if (ni != NULL) { - net = (kib_net_t *)ni->ni_data; - rej.ibr_incarnation = net->ibn_incarnation; - } + nid = reqmsg->ibm_srcnid; + ni = lnet_nid2ni_addref(reqmsg->ibm_dstnid); - 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)); + if (ni != NULL) { + net = (kib_net_t *)ni->ni_data; + rej.ibr_incarnation = net->ibn_incarnation; + } - goto failed; - } + 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 conn from %s on %s (%s:%d:%pI4h): " + "bad dst nid %s\n", libcfs_nid2str(nid), + ni == NULL ? "NA" : libcfs_nid2str(ni->ni_nid), + ibdev->ibd_ifname, ibdev->ibd_nnets, + &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", - reqmsg->ibm_type, libcfs_nid2str(nid)); - goto failed; - } - - 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(version)); + 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; + } - if (version == IBLND_MSG_VERSION) - rej.ibr_why = IBLND_REJECT_MSG_QUEUE_SIZE; + /* I can accept peer's version */ + version = reqmsg->ibm_version; - goto failed; - } + if (reqmsg->ibm_type != IBLND_MSG_CONNREQ) { + CERROR("Unexpected connreq msg type: %x from %s\n", + reqmsg->ibm_type, libcfs_nid2str(nid)); + goto failed; + } - 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_RDMA_FRAGS(version)); + if (reqmsg->ibm_u.connparams.ibcp_queue_depth > + kiblnd_msg_queue_size(version, ni)) { + CERROR("Can't accept conn from %s, queue depth too large: " + " %d (<=%d wanted)\n", + libcfs_nid2str(nid), + reqmsg->ibm_u.connparams.ibcp_queue_depth, + kiblnd_msg_queue_size(version, ni)); - if (version == IBLND_MSG_VERSION) - rej.ibr_why = IBLND_REJECT_RDMA_FRAGS; + if (version == IBLND_MSG_VERSION) + rej.ibr_why = IBLND_REJECT_MSG_QUEUE_SIZE; - goto failed; + goto failed; + } - } + if (reqmsg->ibm_u.connparams.ibcp_max_frags > + kiblnd_rdma_frags(version, ni)) { + CWARN("Can't accept conn from %s (version %x): " + "max_frags %d too large (%d wanted)\n", + libcfs_nid2str(nid), version, + reqmsg->ibm_u.connparams.ibcp_max_frags, + kiblnd_rdma_frags(version, ni)); + + if (version >= IBLND_MSG_VERSION) + rej.ibr_why = IBLND_REJECT_RDMA_FRAGS; + + goto failed; + } else if (reqmsg->ibm_u.connparams.ibcp_max_frags < + kiblnd_rdma_frags(version, ni) && + net->ibn_fmr_ps == NULL) { + CWARN("Can't accept conn from %s (version %x): " + "max_frags %d incompatible without FMR pool " + "(%d wanted)\n", + libcfs_nid2str(nid), version, + reqmsg->ibm_u.connparams.ibcp_max_frags, + kiblnd_rdma_frags(version, ni)); + + 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) { CERROR("Can't accept %s: message size %d too big (%d max)\n", @@ -2333,17 +2363,21 @@ kiblnd_passive_connect (struct rdma_cm_id *cmid, void *priv, int priv_nob) 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.ibr_why = IBLND_REJECT_NO_RESOURCES; - 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.ibr_why = IBLND_REJECT_NO_RESOURCES; + goto failed; + } + + /* We have validated the peer's parameters so use those */ + peer->ibp_max_frags = reqmsg->ibm_u.connparams.ibcp_max_frags; + peer->ibp_queue_depth = reqmsg->ibm_u.connparams.ibcp_queue_depth; write_lock_irqsave(g_lock, flags); - peer2 = kiblnd_find_peer_locked(nid); + peer2 = kiblnd_find_peer_locked(ni, nid); if (peer2 != NULL) { if (peer2->ibp_version == 0) { peer2->ibp_version = version; @@ -2353,31 +2387,61 @@ kiblnd_passive_connect (struct rdma_cm_id *cmid, void *priv, int priv_nob) /* 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); + kiblnd_close_peer_conns_locked(peer2, -ESTALE); + + if (kiblnd_peer_active(peer2)) { + peer2->ibp_incarnation = reqmsg->ibm_srcstamp; + peer2->ibp_version = version; + } write_unlock_irqrestore(g_lock, flags); - CWARN("Conn stale %s [old ver: %x, new ver: %x]\n", - libcfs_nid2str(nid), peer2->ibp_version, version); + CWARN("Conn stale %s version %x/%x incarnation %llu/%llu\n", + libcfs_nid2str(nid), peer2->ibp_version, version, + peer2->ibp_incarnation, reqmsg->ibm_srcstamp); - kiblnd_peer_decref(peer); - rej.ibr_why = IBLND_REJECT_CONN_STALE; - goto failed; - } + 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) { + /* Tie-break connection race in favour of the higher NID. + * If we keep running into a race condition multiple times, + * we have to assume that the connection attempt with the + * higher NID is stuck in a connecting state and will never + * recover. As such, we pass through this if-block and let + * the lower NID connection win so we can move forward. + */ + if (peer2->ibp_connecting != 0 && + nid < ni->ni_nid && peer2->ibp_races < + MAX_CONN_RACES_BEFORE_ABORT) { + peer2->ibp_races++; write_unlock_irqrestore(g_lock, flags); - CWARN("Conn race %s\n", libcfs_nid2str(peer2->ibp_nid)); - - kiblnd_peer_decref(peer); - rej.ibr_why = IBLND_REJECT_CONN_RACE; - goto failed; - } + CDEBUG(D_NET, "Conn race %s\n", + libcfs_nid2str(peer2->ibp_nid)); - peer2->ibp_accepting++; - kiblnd_peer_addref(peer2); + kiblnd_peer_decref(peer); + rej.ibr_why = IBLND_REJECT_CONN_RACE; + goto failed; + } + if (peer2->ibp_races >= MAX_CONN_RACES_BEFORE_ABORT) + CNETERR("Conn race %s: unresolved after %d attempts, letting lower NID win\n", + libcfs_nid2str(peer2->ibp_nid), + MAX_CONN_RACES_BEFORE_ABORT); + /* + * passive connection is allowed even this peer is waiting for + * reconnection. + */ + peer2->ibp_reconnecting = 0; + peer2->ibp_races = 0; + peer2->ibp_accepting++; + kiblnd_peer_addref(peer2); + + /* Race with kiblnd_launch_tx (active connect) to create peer + * so copy validated parameters since we now know what the + * peer's limits are */ + peer2->ibp_max_frags = peer->ibp_max_frags; + peer2->ibp_queue_depth = peer->ibp_queue_depth; write_unlock_irqrestore(g_lock, flags); kiblnd_peer_decref(peer); @@ -2401,7 +2465,7 @@ kiblnd_passive_connect (struct rdma_cm_id *cmid, void *priv, int priv_nob) write_unlock_irqrestore(g_lock, flags); } - conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_PASSIVE_WAIT, version); + 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); @@ -2411,21 +2475,20 @@ kiblnd_passive_connect (struct rdma_cm_id *cmid, void *priv, int priv_nob) /* 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(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)); + conn->ibc_incarnation = reqmsg->ibm_srcstamp; + conn->ibc_credits = conn->ibc_queue_depth; + conn->ibc_reserved_credits = conn->ibc_queue_depth; + LASSERT(conn->ibc_credits + conn->ibc_reserved_credits + + IBLND_OOB_MSGS(version) <= IBLND_RX_MSGS(conn)); 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(version); - ackmsg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE; - ackmsg->ibm_u.connparams.ibcp_max_frags = IBLND_RDMA_FRAGS(version); + ackmsg->ibm_u.connparams.ibcp_queue_depth = conn->ibc_queue_depth; + ackmsg->ibm_u.connparams.ibcp_max_frags = conn->ibc_max_frags; + ackmsg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE; kiblnd_pack_msg(ni, ackmsg, version, 0, nid, reqmsg->ibm_srcstamp); @@ -2455,57 +2518,97 @@ kiblnd_passive_connect (struct rdma_cm_id *cmid, void *priv, int priv_nob) return 0; failed: - if (ni != NULL) - lnet_ni_decref(ni); + if (ni != NULL) { + rej.ibr_cp.ibcp_queue_depth = + kiblnd_msg_queue_size(version, ni); + rej.ibr_cp.ibcp_max_frags = kiblnd_rdma_frags(version, ni); + lnet_ni_decref(ni); + } - 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); + rej.ibr_version = version; + kiblnd_reject(cmid, &rej); - return -ECONNREFUSED; + return -ECONNREFUSED; } static void -kiblnd_reconnect (kib_conn_t *conn, int version, - __u64 incarnation, int why, kib_connparams_t *cp) +kiblnd_check_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; + rwlock_t *glock = &kiblnd_data.kib_global_lock; + kib_peer_t *peer = conn->ibc_peer; + char *reason; + int msg_size = IBLND_MSG_SIZE; + int frag_num = -1; + int queue_dep = -1; + bool reconnect; + unsigned long flags; - LASSERT (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT); - LASSERT (peer->ibp_connecting > 0); /* 'conn' at least */ + LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT); + LASSERT(peer->ibp_connecting > 0); /* 'conn' at least */ + LASSERT(!peer->ibp_reconnecting); - write_lock_irqsave(&kiblnd_data.kib_global_lock, flags); + if (cp) { + msg_size = cp->ibcp_max_msg_size; + frag_num = cp->ibcp_max_frags; + queue_dep = cp->ibcp_queue_depth; + } + write_lock_irqsave(glock, flags); /* retry connection if it's still needed and no other connection * 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 ((!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) - return; + reconnect = (!list_empty(&peer->ibp_tx_queue) || + peer->ibp_version != version) && + peer->ibp_connecting == 1 && + peer->ibp_accepting == 0; + if (!reconnect) { + reason = "no need"; + goto out; + } switch (why) { default: reason = "Unknown"; break; + case IBLND_REJECT_RDMA_FRAGS: { + struct lnet_ioctl_config_o2iblnd_tunables *tunables; + + if (!cp) { + reason = "can't negotiate max frags"; + goto out; + } + tunables = &peer->ibp_ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib; + if (!tunables->lnd_map_on_demand) { + reason = "map_on_demand must be enabled"; + goto out; + } + if (conn->ibc_max_frags <= frag_num) { + reason = "unsupported max frags"; + goto out; + } + + peer->ibp_max_frags = frag_num; + reason = "rdma fragments"; + break; + } + case IBLND_REJECT_MSG_QUEUE_SIZE: + if (!cp) { + reason = "can't negotiate queue depth"; + goto out; + } + if (conn->ibc_queue_depth <= queue_dep) { + reason = "unsupported queue depth"; + goto out; + } + + peer->ibp_queue_depth = queue_dep; + reason = "queue depth"; + break; + case IBLND_REJECT_CONN_STALE: reason = "stale"; break; @@ -2519,15 +2622,24 @@ kiblnd_reconnect (kib_conn_t *conn, int version, 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); + conn->ibc_reconnect = 1; + peer->ibp_reconnecting = 1; + peer->ibp_version = version; + if (incarnation != 0) + peer->ibp_incarnation = incarnation; + out: + write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); - kiblnd_connect_peer(peer); + CNETERR("%s: %s (%s), %x, %x, msg_size: %d, queue_depth: %d/%d, max_frags: %d/%d\n", + libcfs_nid2str(peer->ibp_nid), + reconnect ? "reconnect" : "don't reconnect", + reason, IBLND_MSG_VERSION, version, msg_size, + conn->ibc_queue_depth, queue_dep, + conn->ibc_max_frags, frag_num); + /* + * if conn::ibc_reconnect is TRUE, connd will reconnect to the peer + * while destroying the zombie + */ } static void @@ -2540,8 +2652,8 @@ kiblnd_rejected (kib_conn_t *conn, int reason, void *priv, int priv_nob) switch (reason) { case IB_CM_REJ_STALE_CONN: - kiblnd_reconnect(conn, IBLND_MSG_VERSION, 0, - IBLND_REJECT_CONN_STALE, NULL); + kiblnd_check_reconnect(conn, IBLND_MSG_VERSION, 0, + IBLND_REJECT_CONN_STALE, NULL); break; case IB_CM_REJ_INVALID_SERVICE_ID: @@ -2621,24 +2733,10 @@ kiblnd_rejected (kib_conn_t *conn, int reason, void *priv, int priv_nob) case IBLND_REJECT_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 != NULL ? cp->ibcp_queue_depth : - IBLND_MSG_QUEUE_SIZE(rej->ibr_version), - 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 != NULL ? cp->ibcp_max_frags : - IBLND_RDMA_FRAGS(rej->ibr_version), - IBLND_RDMA_FRAGS(conn->ibc_version)); + case IBLND_REJECT_MSG_QUEUE_SIZE: + case IBLND_REJECT_RDMA_FRAGS: + kiblnd_check_reconnect(conn, rej->ibr_version, + incarnation, rej->ibr_why, cp); break; case IBLND_REJECT_NO_RESOURCES: @@ -2703,25 +2801,25 @@ 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(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(ver)); - rc = -EPROTO; - goto failed; - } + if (msg->ibm_u.connparams.ibcp_queue_depth > + conn->ibc_queue_depth) { + CERROR("%s has incompatible queue depth %d (<=%d wanted)\n", + libcfs_nid2str(peer->ibp_nid), + msg->ibm_u.connparams.ibcp_queue_depth, + conn->ibc_queue_depth); + rc = -EPROTO; + goto failed; + } - 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_RDMA_FRAGS(ver)); - rc = -EPROTO; - goto failed; - } + if (msg->ibm_u.connparams.ibcp_max_frags > + conn->ibc_max_frags) { + CERROR("%s has incompatible max_frags %d (<=%d wanted)\n", + libcfs_nid2str(peer->ibp_nid), + msg->ibm_u.connparams.ibcp_max_frags, + conn->ibc_max_frags); + rc = -EPROTO; + goto failed; + } if (msg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) { CERROR("%s max message size %d too big (%d max)\n", @@ -2748,11 +2846,13 @@ kiblnd_check_connreply (kib_conn_t *conn, void *priv, int priv_nob) goto failed; } - 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)); + conn->ibc_incarnation = msg->ibm_srcstamp; + conn->ibc_credits = msg->ibm_u.connparams.ibcp_queue_depth; + conn->ibc_reserved_credits = msg->ibm_u.connparams.ibcp_queue_depth; + conn->ibc_queue_depth = msg->ibm_u.connparams.ibcp_queue_depth; + conn->ibc_max_frags = msg->ibm_u.connparams.ibcp_max_frags; + LASSERT(conn->ibc_credits + conn->ibc_reserved_credits + + IBLND_OOB_MSGS(ver) <= IBLND_RX_MSGS(conn)); kiblnd_connreq_done(conn, 0); return; @@ -2788,7 +2888,8 @@ kiblnd_active_connect (struct rdma_cm_id *cmid) read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); - conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_ACTIVE_CONNECT, version); + 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 */ @@ -2801,11 +2902,11 @@ kiblnd_active_connect (struct rdma_cm_id *cmid) msg = &conn->ibc_connvars->cv_msg; - 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(version); - msg->ibm_u.connparams.ibcp_max_frags = IBLND_RDMA_FRAGS(version); - msg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE; + memset(msg, 0, sizeof(*msg)); + kiblnd_init_msg(msg, IBLND_MSG_CONNREQ, sizeof(msg->ibm_u.connparams)); + msg->ibm_u.connparams.ibcp_queue_depth = conn->ibc_queue_depth; + msg->ibm_u.connparams.ibcp_max_frags = conn->ibc_max_frags; + msg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE; kiblnd_pack_msg(peer->ibp_ni, msg, version, 0, peer->ibp_nid, incarnation); @@ -2972,11 +3073,10 @@ 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) { @@ -2998,11 +3098,9 @@ kiblnd_cm_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event) * to ignore this */ 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 } } @@ -3055,9 +3153,9 @@ kiblnd_check_conns (int idx) struct list_head *ctmp; unsigned long flags; - /* 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 - * take a look... */ + /* 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 + * take a look... */ read_lock_irqsave(&kiblnd_data.kib_global_lock, flags); list_for_each(ptmp, peers) { @@ -3120,7 +3218,11 @@ kiblnd_check_conns (int idx) conn = list_entry(checksends.next, kib_conn_t, ibc_connd_list); list_del(&conn->ibc_connd_list); - kiblnd_check_sends(conn); + + spin_lock(&conn->ibc_lock); + kiblnd_check_sends_locked(conn); + spin_unlock(&conn->ibc_lock); + kiblnd_conn_decref(conn); } } @@ -3138,10 +3240,22 @@ kiblnd_disconnect_conn (kib_conn_t *conn) kiblnd_peer_notify(conn->ibc_peer); } +/* + * High-water for reconnection to the same peer, reconnection attempt should + * be delayed after trying more than KIB_RECONN_HIGH_RACE. + */ +#define KIB_RECONN_HIGH_RACE 10 +/* + * Allow connd to take a break and handle other things after consecutive + * reconnection attemps. + */ +#define KIB_RECONN_BREAK 100 + int kiblnd_connd (void *arg) { - wait_queue_t wait; + spinlock_t *lock= &kiblnd_data.kib_connd_lock; + wait_queue_t wait; unsigned long flags; kib_conn_t *conn; int timeout; @@ -3150,30 +3264,45 @@ kiblnd_connd (void *arg) int peer_index = 0; unsigned long deadline = jiffies; - cfs_block_allsigs (); + cfs_block_allsigs(); - init_waitqueue_entry_current (&wait); + init_waitqueue_entry(&wait, current); kiblnd_data.kib_connd = current; - spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); + spin_lock_irqsave(lock, flags); - while (!kiblnd_data.kib_shutdown) { + while (!kiblnd_data.kib_shutdown) { + int reconn = 0; 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); + kib_peer_t *peer = NULL; + + conn = list_entry(kiblnd_data.kib_connd_zombies.next, + kib_conn_t, ibc_list); list_del(&conn->ibc_list); + if (conn->ibc_reconnect) { + peer = conn->ibc_peer; + kiblnd_peer_addref(peer); + } - spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, - flags); + spin_unlock_irqrestore(lock, flags); dropped_lock = 1; - kiblnd_destroy_conn(conn); + kiblnd_destroy_conn(conn, !peer); + + spin_lock_irqsave(lock, flags); + if (!peer) + continue; - spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); + conn->ibc_peer = peer; + if (peer->ibp_reconnected < KIB_RECONN_HIGH_RACE) + list_add_tail(&conn->ibc_list, + &kiblnd_data.kib_reconn_list); + else + list_add_tail(&conn->ibc_list, + &kiblnd_data.kib_reconn_wait); } if (!list_empty(&kiblnd_data.kib_connd_conns)) { @@ -3181,16 +3310,40 @@ kiblnd_connd (void *arg) kib_conn_t, ibc_list); list_del(&conn->ibc_list); - spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, - flags); + spin_unlock_irqrestore(lock, flags); dropped_lock = 1; kiblnd_disconnect_conn(conn); kiblnd_conn_decref(conn); - spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); + spin_lock_irqsave(lock, flags); } + while (reconn < KIB_RECONN_BREAK) { + if (kiblnd_data.kib_reconn_sec != + ktime_get_real_seconds()) { + kiblnd_data.kib_reconn_sec = ktime_get_real_seconds(); + list_splice_init(&kiblnd_data.kib_reconn_wait, + &kiblnd_data.kib_reconn_list); + } + + if (list_empty(&kiblnd_data.kib_reconn_list)) + break; + + conn = list_entry(kiblnd_data.kib_reconn_list.next, + kib_conn_t, ibc_list); + list_del(&conn->ibc_list); + + spin_unlock_irqrestore(lock, flags); + dropped_lock = 1; + + reconn += kiblnd_reconnect_peer(conn->ibc_peer); + kiblnd_peer_decref(conn->ibc_peer); + LIBCFS_FREE(conn, sizeof(*conn)); + + spin_lock_irqsave(lock, flags); + } + /* careful with the jiffy wrap... */ timeout = (int)(deadline - jiffies); if (timeout <= 0) { @@ -3198,7 +3351,7 @@ 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(lock, flags); dropped_lock = 1; /* Time to check for RDMA timeouts on a few more @@ -3221,8 +3374,8 @@ kiblnd_connd (void *arg) kiblnd_data.kib_peer_hash_size; } - deadline += p * HZ; - spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); + deadline += msecs_to_jiffies(p * MSEC_PER_SEC); + spin_lock_irqsave(lock, flags); } if (dropped_lock) @@ -3231,16 +3384,16 @@ kiblnd_connd (void *arg) /* 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); + spin_unlock_irqrestore(lock, flags); - waitq_timedwait(&wait, TASK_INTERRUPTIBLE, timeout); + schedule_timeout(timeout); set_current_state(TASK_RUNNING); remove_wait_queue(&kiblnd_data.kib_connd_waitq, &wait); - spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); + spin_lock_irqsave(lock, flags); } - spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags); + spin_unlock_irqrestore(lock, flags); kiblnd_thread_fini(); return 0; @@ -3255,6 +3408,10 @@ kiblnd_qp_event(struct ib_event *event, void *arg) case IB_EVENT_COMM_EST: CDEBUG(D_NET, "%s established\n", libcfs_nid2str(conn->ibc_peer->ibp_nid)); + /* We received a packet but connection isn't established + * probably handshake packet was lost, so free to + * force make connection established */ + rdma_notify(conn->ibc_cmid, IB_EVENT_COMM_EST); return; default: @@ -3267,9 +3424,15 @@ kiblnd_qp_event(struct ib_event *event, void *arg) static void kiblnd_complete (struct ib_wc *wc) { - switch (kiblnd_wreqid2type(wc->wr_id)) { - default: - LBUG(); + switch (kiblnd_wreqid2type(wc->wr_id)) { + default: + LBUG(); + + case IBLND_WID_MR: + if (wc->status != IB_WC_SUCCESS && + wc->status != IB_WC_WR_FLUSH_ERR) + CNETERR("FastReg failed: %d\n", wc->status); + return; case IBLND_WID_RDMA: /* We only get RDMA completion notification if it fails. All @@ -3349,16 +3512,16 @@ kiblnd_scheduler(void *arg) cfs_block_allsigs(); - init_waitqueue_entry_current(&wait); + init_waitqueue_entry(&wait, current); 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); + CWARN("Unable to bind on CPU partition %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); } spin_lock_irqsave(&sched->ibs_lock, flags); @@ -3385,6 +3548,8 @@ kiblnd_scheduler(void *arg) spin_unlock_irqrestore(&sched->ibs_lock, flags); + wc.wr_id = IBLND_WID_INVAL; + rc = ib_poll_cq(conn->ibc_cq, 1, &wc); if (rc == 0) { rc = ib_req_notify_cq(conn->ibc_cq, @@ -3403,6 +3568,19 @@ kiblnd_scheduler(void *arg) rc = ib_poll_cq(conn->ibc_cq, 1, &wc); } + if (unlikely(rc > 0 && wc.wr_id == IBLND_WID_INVAL)) { + LCONSOLE_ERROR( + "ib_poll_cq (rc: %d) returned invalid " + "wr_id, opcode %d, status: %d, " + "vendor_err: %d, conn: %s status: %d\n" + "please upgrade firmware and OFED or " + "contact vendor.\n", rc, + wc.opcode, wc.status, wc.vendor_err, + libcfs_nid2str(conn->ibc_peer->ibp_nid), + conn->ibc_state); + rc = -EINVAL; + } + if (rc < 0) { CWARN("%s: ib_poll_cq failed: %d, " "closing connection\n", @@ -3448,7 +3626,7 @@ kiblnd_scheduler(void *arg) add_wait_queue_exclusive(&sched->ibs_waitq, &wait); spin_unlock_irqrestore(&sched->ibs_lock, flags); - waitq_wait(&wait, TASK_INTERRUPTIBLE); + schedule(); busy_loops = 0; remove_wait_queue(&sched->ibs_waitq, &wait); @@ -3466,16 +3644,16 @@ 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; + kib_dev_t *dev; + wait_queue_t wait; + unsigned long flags; + int rc; - LASSERT (*kiblnd_tunables.kib_dev_failover != 0); + LASSERT(*kiblnd_tunables.kib_dev_failover != 0); - cfs_block_allsigs (); + cfs_block_allsigs(); - init_waitqueue_entry_current(&wait); + init_waitqueue_entry(&wait, current); write_lock_irqsave(glock, flags); while (!kiblnd_data.kib_shutdown) {