X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lnet%2Fklnds%2Fo2iblnd%2Fo2iblnd.c;h=ee5a01f9fa5185e2fb072cfe91004bfe8546ae49;hp=5576ee8672bb9188ba65fb5936aac59b0e533aa1;hb=783428b60a98874b4783f8da48c66019d68d84d6;hpb=4c689a573fafcfa1ca7474a275f958e00b1deddc diff --git a/lnet/klnds/o2iblnd/o2iblnd.c b/lnet/klnds/o2iblnd/o2iblnd.c index 5576ee8..ee5a01f 100644 --- a/lnet/klnds/o2iblnd/o2iblnd.c +++ b/lnet/klnds/o2iblnd/o2iblnd.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) 2011, 2015, Intel Corporation. + * Copyright (c) 2011, 2016, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -337,8 +333,8 @@ kiblnd_create_peer(lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid) peer->ibp_nid = nid; peer->ibp_error = 0; peer->ibp_last_alive = 0; - peer->ibp_max_frags = IBLND_CFG_RDMA_FRAGS; - peer->ibp_queue_depth = *kiblnd_tunables.kib_peertxcredits; + peer->ibp_max_frags = kiblnd_cfg_rdma_frags(peer->ibp_ni); + peer->ibp_queue_depth = ni->ni_peertxcredits; atomic_set(&peer->ibp_refcount, 1); /* 1 ref for caller */ INIT_LIST_HEAD(&peer->ibp_list); /* not in the peer table yet */ @@ -367,9 +363,7 @@ kiblnd_destroy_peer (kib_peer_t *peer) LASSERT(net != NULL); LASSERT (atomic_read(&peer->ibp_refcount) == 0); LASSERT(!kiblnd_peer_active(peer)); - LASSERT(peer->ibp_connecting == 0); - LASSERT(peer->ibp_accepting == 0); - LASSERT(list_empty(&peer->ibp_conns)); + LASSERT(kiblnd_peer_idle(peer)); LASSERT(list_empty(&peer->ibp_tx_queue)); LIBCFS_FREE(peer, sizeof(*peer)); @@ -393,10 +387,7 @@ kiblnd_find_peer_locked (lnet_nid_t nid) list_for_each(tmp, peer_list) { peer = list_entry(tmp, kib_peer_t, ibp_list); - - LASSERT(peer->ibp_connecting > 0 || /* creating conns */ - peer->ibp_accepting > 0 || - !list_empty(&peer->ibp_conns)); /* active conn */ + LASSERT(!kiblnd_peer_idle(peer)); if (peer->ibp_nid != nid) continue; @@ -437,9 +428,7 @@ kiblnd_get_peer_info(lnet_ni_t *ni, int index, list_for_each(ptmp, &kiblnd_data.kib_peers[i]) { peer = list_entry(ptmp, kib_peer_t, ibp_list); - LASSERT(peer->ibp_connecting > 0 || - peer->ibp_accepting > 0 || - !list_empty(&peer->ibp_conns)); + LASSERT(!kiblnd_peer_idle(peer)); if (peer->ibp_ni != ni) continue; @@ -506,9 +495,7 @@ kiblnd_del_peer (lnet_ni_t *ni, lnet_nid_t nid) for (i = lo; i <= hi; i++) { list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) { peer = list_entry(ptmp, kib_peer_t, ibp_list); - LASSERT(peer->ibp_connecting > 0 || - peer->ibp_accepting > 0 || - !list_empty(&peer->ibp_conns)); + LASSERT(!kiblnd_peer_idle(peer)); if (peer->ibp_ni != ni) continue; @@ -551,9 +538,7 @@ kiblnd_get_conn_by_idx(lnet_ni_t *ni, int index) list_for_each(ptmp, &kiblnd_data.kib_peers[i]) { peer = list_entry(ptmp, kib_peer_t, ibp_list); - LASSERT(peer->ibp_connecting > 0 || - peer->ibp_accepting > 0 || - !list_empty(&peer->ibp_conns)); + LASSERT(!kiblnd_peer_idle(peer)); if (peer->ibp_ni != ni) continue; @@ -587,7 +572,7 @@ static void kiblnd_debug_tx (kib_tx_t *tx) { CDEBUG(D_CONSOLE, " %p snd %d q %d w %d rc %d dl %lx " - "cookie "LPX64" msg %s%s type %x cred %d\n", + "cookie %#llx msg %s%s type %x cred %d\n", tx, tx->tx_sending, tx->tx_queued, tx->tx_waiting, tx->tx_status, tx->tx_deadline, tx->tx_cookie, tx->tx_lntmsg[0] == NULL ? "-" : "!", @@ -686,6 +671,7 @@ kiblnd_get_completion_vector(kib_conn_t *conn, int cpt) int vectors; int off; int i; + lnet_nid_t ibp_nid; vectors = conn->ibc_cmid->device->num_comp_vectors; if (vectors <= 1) @@ -694,7 +680,8 @@ kiblnd_get_completion_vector(kib_conn_t *conn, int cpt) mask = cfs_cpt_cpumask(lnet_cpt_table(), cpt); /* hash NID to CPU id in this partition... */ - off = conn->ibc_peer->ibp_nid % cpumask_weight(mask); + ibp_nid = conn->ibc_peer->ibp_nid; + off = do_div(ibp_nid, cpumask_weight(mask)); for_each_cpu(i, mask) { if (off-- == 0) return i % vectors; @@ -859,24 +846,34 @@ kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid, conn->ibc_sched = sched; - rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr); - if (rc != 0) { - CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d\n", - rc, init_qp_attr->cap.max_send_wr, - init_qp_attr->cap.max_recv_wr); - goto failed_2; - } + do { + rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr); + if (!rc || init_qp_attr->cap.max_send_wr < 16) + break; - LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr)); + init_qp_attr->cap.max_send_wr -= init_qp_attr->cap.max_send_wr / 4; + } while (rc); + + if (rc) { + CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d\n", + rc, init_qp_attr->cap.max_send_wr, + init_qp_attr->cap.max_recv_wr); + goto failed_2; + } - /* 1 ref for caller and each rxmsg */ + if (init_qp_attr->cap.max_send_wr != IBLND_SEND_WRS(conn)) + CDEBUG(D_NET, "original send wr %d, created with %d\n", + IBLND_SEND_WRS(conn), init_qp_attr->cap.max_send_wr); + + LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr)); + + /* 1 ref for caller and each rxmsg */ atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(conn)); conn->ibc_nrx = IBLND_RX_MSGS(conn); - /* post receives */ + /* post receives */ for (i = 0; i < IBLND_RX_MSGS(conn); i++) { - rc = kiblnd_post_rx(&conn->ibc_rxs[i], - IBLND_POSTRX_NO_CREDIT); + rc = kiblnd_post_rx(&conn->ibc_rxs[i], IBLND_POSTRX_NO_CREDIT); if (rc != 0) { CERROR("Can't post rxmsg: %d\n", rc); @@ -913,7 +910,7 @@ kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid, return conn; failed_2: - kiblnd_destroy_conn(conn); + kiblnd_destroy_conn(conn, true); failed_1: LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr)); failed_0: @@ -921,7 +918,7 @@ kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid, } void -kiblnd_destroy_conn (kib_conn_t *conn) +kiblnd_destroy_conn(kib_conn_t *conn, bool free_conn) { struct rdma_cm_id *cmid = conn->ibc_cmid; kib_peer_t *peer = conn->ibc_peer; @@ -980,25 +977,13 @@ kiblnd_destroy_conn (kib_conn_t *conn) if (conn->ibc_state != IBLND_CONN_INIT) { kib_net_t *net = peer->ibp_ni->ni_data; + kiblnd_peer_decref(peer); rdma_destroy_id(cmid); atomic_dec(&net->ibn_nconns); - if (conn->ibc_conn_race) { - if (peer->ibp_accepting == 0 && - !list_empty(&peer->ibp_tx_queue)) { - kiblnd_connect_peer(peer); - } else { - rwlock_t *glock = &kiblnd_data.kib_global_lock; - unsigned long flags; - - write_lock_irqsave(glock, flags); - peer->ibp_connecting--; - write_unlock_irqrestore(glock, flags); - } - } - kiblnd_peer_decref(peer); } - LIBCFS_FREE(conn, sizeof(*conn)); + if (free_conn) + LIBCFS_FREE(conn, sizeof(*conn)); } int @@ -1041,7 +1026,7 @@ kiblnd_close_stale_conns_locked(kib_peer_t *peer, continue; CDEBUG(D_NET, "Closing stale conn -> %s version: %x, " - "incarnation:"LPX64"(%x, "LPX64")\n", + "incarnation:%#llx(%x, %#llx)\n", libcfs_nid2str(peer->ibp_nid), conn->ibc_version, conn->ibc_incarnation, version, incarnation); @@ -1078,9 +1063,7 @@ kiblnd_close_matching_conns(lnet_ni_t *ni, lnet_nid_t nid) list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) { peer = list_entry(ptmp, kib_peer_t, ibp_list); - LASSERT(peer->ibp_connecting > 0 || - peer->ibp_accepting > 0 || - !list_empty(&peer->ibp_conns)); + LASSERT(!kiblnd_peer_idle(peer)); if (peer->ibp_ni != ni) continue; @@ -1167,12 +1150,8 @@ kiblnd_query(lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when) read_lock_irqsave(glock, flags); peer = kiblnd_find_peer_locked(nid); - if (peer != NULL) { - LASSERT(peer->ibp_connecting > 0 || /* creating conns */ - peer->ibp_accepting > 0 || - !list_empty(&peer->ibp_conns)); /* active conn */ + if (peer != NULL) last_alive = peer->ibp_last_alive; - } read_unlock_irqrestore(glock, flags); @@ -1283,7 +1262,7 @@ kiblnd_map_rx_descs(kib_conn_t *conn) rx->rx_msgaddr)); KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr); - CDEBUG(D_NET, "rx %d: %p "LPX64"("LPX64")\n", + CDEBUG(D_NET, "rx %d: %p %#llx(%#llx)\n", i, rx->rx_msg, rx->rx_msgaddr, (__u64)(page_to_phys(pg) + pg_off)); @@ -1335,6 +1314,7 @@ kiblnd_current_hdev(kib_dev_t *dev) if (i++ % 50 == 0) CDEBUG(D_NET, "%s: Wait for failover\n", dev->ibd_ifname); + set_current_state(TASK_INTERRUPTIBLE); schedule_timeout(cfs_time_seconds(1) / 100); read_lock_irqsave(&kiblnd_data.kib_global_lock, flags); @@ -1402,100 +1382,269 @@ kiblnd_map_tx_pool(kib_tx_pool_t *tpo) } struct ib_mr * -kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev, kib_rdma_desc_t *rd, +kiblnd_find_rd_dma_mr(struct lnet_ni *ni, kib_rdma_desc_t *rd, int negotiated_nfrags) { - __u16 nfrags = (negotiated_nfrags != -1) ? - negotiated_nfrags : *kiblnd_tunables.kib_map_on_demand; + kib_net_t *net = ni->ni_data; + kib_hca_dev_t *hdev = net->ibn_dev->ibd_hdev; + struct lnet_ioctl_config_o2iblnd_tunables *tunables; + int mod; + __u16 nfrags; + + tunables = &ni->ni_lnd_tunables->lt_tun_u.lt_o2ib; + mod = tunables->lnd_map_on_demand; + nfrags = (negotiated_nfrags != -1) ? negotiated_nfrags : mod; LASSERT(hdev->ibh_mrs != NULL); - if (*kiblnd_tunables.kib_map_on_demand > 0 && - nfrags <= rd->rd_nfrags) + if (mod > 0 && nfrags <= rd->rd_nfrags) return NULL; return hdev->ibh_mrs; } static void -kiblnd_destroy_fmr_pool(kib_fmr_pool_t *pool) +kiblnd_destroy_fmr_pool(kib_fmr_pool_t *fpo) { - LASSERT (pool->fpo_map_count == 0); + LASSERT(fpo->fpo_map_count == 0); - if (pool->fpo_fmr_pool != NULL) - ib_destroy_fmr_pool(pool->fpo_fmr_pool); + if (fpo->fpo_is_fmr) { + if (fpo->fmr.fpo_fmr_pool) + ib_destroy_fmr_pool(fpo->fmr.fpo_fmr_pool); + } else { + struct kib_fast_reg_descriptor *frd, *tmp; + int i = 0; + + list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list, + frd_list) { + list_del(&frd->frd_list); +#ifndef HAVE_IB_MAP_MR_SG + ib_free_fast_reg_page_list(frd->frd_frpl); +#endif + ib_dereg_mr(frd->frd_mr); + LIBCFS_FREE(frd, sizeof(*frd)); + i++; + } + if (i < fpo->fast_reg.fpo_pool_size) + CERROR("FastReg pool still has %d regions registered\n", + fpo->fast_reg.fpo_pool_size - i); + } - if (pool->fpo_hdev != NULL) - kiblnd_hdev_decref(pool->fpo_hdev); + if (fpo->fpo_hdev) + kiblnd_hdev_decref(fpo->fpo_hdev); - LIBCFS_FREE(pool, sizeof(kib_fmr_pool_t)); + LIBCFS_FREE(fpo, sizeof(*fpo)); } static void kiblnd_destroy_fmr_pool_list(struct list_head *head) { - kib_fmr_pool_t *pool; + kib_fmr_pool_t *fpo, *tmp; - while (!list_empty(head)) { - pool = list_entry(head->next, kib_fmr_pool_t, fpo_list); - list_del(&pool->fpo_list); - kiblnd_destroy_fmr_pool(pool); + list_for_each_entry_safe(fpo, tmp, head, fpo_list) { + list_del(&fpo->fpo_list); + kiblnd_destroy_fmr_pool(fpo); } } -static int kiblnd_fmr_pool_size(int ncpts) +static int +kiblnd_fmr_pool_size(struct lnet_ioctl_config_o2iblnd_tunables *tunables, + int ncpts) { - int size = *kiblnd_tunables.kib_fmr_pool_size / ncpts; + int size = tunables->lnd_fmr_pool_size / ncpts; return max(IBLND_FMR_POOL, size); } -static int kiblnd_fmr_flush_trigger(int ncpts) +static int +kiblnd_fmr_flush_trigger(struct lnet_ioctl_config_o2iblnd_tunables *tunables, + int ncpts) { - int size = *kiblnd_tunables.kib_fmr_flush_trigger / ncpts; + int size = tunables->lnd_fmr_flush_trigger / ncpts; return max(IBLND_FMR_POOL_FLUSH, size); } -static int -kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo) +static int kiblnd_alloc_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t *fpo) { - /* FMR pool for RDMA */ - kib_dev_t *dev = fps->fps_net->ibn_dev; - kib_fmr_pool_t *fpo; - struct ib_fmr_pool_param param = { - .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE, - .page_shift = PAGE_SHIFT, - .access = (IB_ACCESS_LOCAL_WRITE | - IB_ACCESS_REMOTE_WRITE), + struct ib_fmr_pool_param param = { + .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE, + .page_shift = PAGE_SHIFT, + .access = (IB_ACCESS_LOCAL_WRITE | + IB_ACCESS_REMOTE_WRITE), .pool_size = fps->fps_pool_size, .dirty_watermark = fps->fps_flush_trigger, .flush_function = NULL, .flush_arg = NULL, - .cache = !!*kiblnd_tunables.kib_fmr_cache}; + .cache = !!fps->fps_cache }; + int rc = 0; + + fpo->fmr.fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd, + ¶m); + if (IS_ERR(fpo->fmr.fpo_fmr_pool)) { + rc = PTR_ERR(fpo->fmr.fpo_fmr_pool); + if (rc != -ENOSYS) + CERROR("Failed to create FMR pool: %d\n", rc); + else + CERROR("FMRs are not supported\n"); + } + + return rc; +} + +static int kiblnd_alloc_freg_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t *fpo) +{ + struct kib_fast_reg_descriptor *frd, *tmp; + int i, rc; + + INIT_LIST_HEAD(&fpo->fast_reg.fpo_pool_list); + fpo->fast_reg.fpo_pool_size = 0; + for (i = 0; i < fps->fps_pool_size; i++) { + LIBCFS_CPT_ALLOC(frd, lnet_cpt_table(), fps->fps_cpt, + sizeof(*frd)); + if (!frd) { + CERROR("Failed to allocate a new fast_reg descriptor\n"); + rc = -ENOMEM; + goto out; + } + frd->frd_mr = NULL; + +#ifndef HAVE_IB_MAP_MR_SG + frd->frd_frpl = ib_alloc_fast_reg_page_list(fpo->fpo_hdev->ibh_ibdev, + LNET_MAX_PAYLOAD/PAGE_SIZE); + if (IS_ERR(frd->frd_frpl)) { + rc = PTR_ERR(frd->frd_frpl); + CERROR("Failed to allocate ib_fast_reg_page_list: %d\n", + rc); + frd->frd_frpl = NULL; + goto out_middle; + } +#endif + +#ifdef HAVE_IB_ALLOC_FAST_REG_MR + frd->frd_mr = ib_alloc_fast_reg_mr(fpo->fpo_hdev->ibh_pd, + LNET_MAX_PAYLOAD/PAGE_SIZE); +#else + frd->frd_mr = ib_alloc_mr(fpo->fpo_hdev->ibh_pd, + IB_MR_TYPE_MEM_REG, + LNET_MAX_PAYLOAD/PAGE_SIZE); +#endif + if (IS_ERR(frd->frd_mr)) { + rc = PTR_ERR(frd->frd_mr); + CERROR("Failed to allocate ib_fast_reg_mr: %d\n", rc); + frd->frd_mr = NULL; + goto out_middle; + } + + /* There appears to be a bug in MLX5 code where you must + * invalidate the rkey of a new FastReg pool before first + * using it. Thus, I am marking the FRD invalid here. */ + frd->frd_valid = false; + + list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list); + fpo->fast_reg.fpo_pool_size++; + } + + return 0; + +out_middle: + if (frd->frd_mr) + ib_dereg_mr(frd->frd_mr); +#ifndef HAVE_IB_MAP_MR_SG + if (frd->frd_frpl) + ib_free_fast_reg_page_list(frd->frd_frpl); +#endif + LIBCFS_FREE(frd, sizeof(*frd)); + +out: + list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list, + frd_list) { + list_del(&frd->frd_list); +#ifndef HAVE_IB_MAP_MR_SG + ib_free_fast_reg_page_list(frd->frd_frpl); +#endif + ib_dereg_mr(frd->frd_mr); + LIBCFS_FREE(frd, sizeof(*frd)); + } + + return rc; +} + +static int +kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo) +{ + struct ib_device_attr *dev_attr; + kib_dev_t *dev = fps->fps_net->ibn_dev; + kib_fmr_pool_t *fpo; int rc; - LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo)); - if (fpo == NULL) +#ifndef HAVE_IB_DEVICE_ATTRS + dev_attr = kmalloc(sizeof(*dev_attr), GFP_KERNEL); + if (!dev_attr) return -ENOMEM; +#endif + + LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo)); + if (!fpo) { + rc = -ENOMEM; + goto out_dev_attr; + } fpo->fpo_hdev = kiblnd_current_hdev(dev); - fpo->fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd, ¶m); - if (IS_ERR(fpo->fpo_fmr_pool)) { - rc = PTR_ERR(fpo->fpo_fmr_pool); - CERROR("Failed to create FMR pool: %d\n", rc); +#ifdef HAVE_IB_DEVICE_ATTRS + dev_attr = &fpo->fpo_hdev->ibh_ibdev->attrs; +#else + rc = ib_query_device(fpo->fpo_hdev->ibh_ibdev, dev_attr); + if (rc) { + CERROR("Query device failed for %s: %d\n", + fpo->fpo_hdev->ibh_ibdev->name, rc); + goto out_dev_attr; + } +#endif - kiblnd_hdev_decref(fpo->fpo_hdev); - LIBCFS_FREE(fpo, sizeof(kib_fmr_pool_t)); - return rc; - } + /* Check for FMR or FastReg support */ + fpo->fpo_is_fmr = 0; + if (fpo->fpo_hdev->ibh_ibdev->alloc_fmr && + fpo->fpo_hdev->ibh_ibdev->dealloc_fmr && + fpo->fpo_hdev->ibh_ibdev->map_phys_fmr && + fpo->fpo_hdev->ibh_ibdev->unmap_fmr) { + LCONSOLE_INFO("Using FMR for registration\n"); + fpo->fpo_is_fmr = 1; + } else if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) { + LCONSOLE_INFO("Using FastReg for registration\n"); + } else { + rc = -ENOSYS; + LCONSOLE_ERROR_MSG(rc, "IB device does not support FMRs nor FastRegs, can't register memory\n"); + goto out_dev_attr; + } - fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE); - fpo->fpo_owner = fps; - *pp_fpo = fpo; + if (fpo->fpo_is_fmr) + rc = kiblnd_alloc_fmr_pool(fps, fpo); + else + rc = kiblnd_alloc_freg_pool(fps, fpo); + if (rc) + goto out_fpo; - return 0; +#ifndef HAVE_IB_DEVICE_ATTRS + kfree(dev_attr); +#endif + fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE); + fpo->fpo_owner = fps; + *pp_fpo = fpo; + + return 0; + +out_fpo: + kiblnd_hdev_decref(fpo->fpo_hdev); + LIBCFS_FREE(fpo, sizeof(*fpo)); + +out_dev_attr: +#ifndef HAVE_IB_DEVICE_ATTRS + kfree(dev_attr); +#endif + + return rc; } static void @@ -1530,8 +1679,9 @@ kiblnd_fini_fmr_poolset(kib_fmr_poolset_t *fps) } static int -kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt, kib_net_t *net, - int pool_size, int flush_trigger) +kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt, int ncpts, + kib_net_t *net, + struct lnet_ioctl_config_o2iblnd_tunables *tunables) { kib_fmr_pool_t *fpo; int rc; @@ -1540,8 +1690,11 @@ kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt, kib_net_t *net, fps->fps_net = net; fps->fps_cpt = cpt; - fps->fps_pool_size = pool_size; - fps->fps_flush_trigger = flush_trigger; + + fps->fps_pool_size = kiblnd_fmr_pool_size(tunables, ncpts); + fps->fps_flush_trigger = kiblnd_fmr_flush_trigger(tunables, ncpts); + fps->fps_cache = tunables->lnd_fmr_cache; + spin_lock_init(&fps->fps_lock); INIT_LIST_HEAD(&fps->fps_pool_list); INIT_LIST_HEAD(&fps->fps_failed_pool_list); @@ -1563,26 +1716,65 @@ kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, cfs_time_t now) return cfs_time_aftereq(now, fpo->fpo_deadline); } +static int +kiblnd_map_tx_pages(kib_tx_t *tx, kib_rdma_desc_t *rd) +{ + kib_hca_dev_t *hdev; + __u64 *pages = tx->tx_pages; + int npages; + int size; + int i; + + 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; + } + } + + return npages; +} + void kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status) { struct list_head zombies = LIST_HEAD_INIT(zombies); kib_fmr_pool_t *fpo = fmr->fmr_pool; - kib_fmr_poolset_t *fps = fpo->fpo_owner; + kib_fmr_poolset_t *fps; cfs_time_t now = cfs_time_current(); kib_fmr_pool_t *tmp; int rc; - rc = ib_fmr_pool_unmap(fmr->fmr_pfmr); - LASSERT(rc == 0); + if (!fpo) + return; - if (status != 0) { - rc = ib_flush_fmr_pool(fpo->fpo_fmr_pool); - LASSERT(rc == 0); - } + fps = fpo->fpo_owner; + if (fpo->fpo_is_fmr) { + if (fmr->fmr_pfmr) { + rc = ib_fmr_pool_unmap(fmr->fmr_pfmr); + LASSERT(!rc); + fmr->fmr_pfmr = NULL; + } + + if (status) { + rc = ib_flush_fmr_pool(fpo->fmr.fpo_fmr_pool); + LASSERT(!rc); + } + } else { + struct kib_fast_reg_descriptor *frd = fmr->fmr_frd; + if (frd) { + frd->frd_valid = false; + spin_lock(&fps->fps_lock); + list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list); + spin_unlock(&fps->fps_lock); + fmr->fmr_frd = NULL; + } + } fmr->fmr_pool = NULL; - fmr->fmr_pfmr = NULL; spin_lock(&fps->fps_lock); fpo->fpo_map_count--; /* decref the pool */ @@ -1604,13 +1796,16 @@ kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status) } int -kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages, - __u64 iov, kib_fmr_t *fmr) +kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, kib_tx_t *tx, kib_rdma_desc_t *rd, + __u32 nob, __u64 iov, kib_fmr_t *fmr) { - struct ib_pool_fmr *pfmr; - kib_fmr_pool_t *fpo; - __u64 version; - int rc; + kib_fmr_pool_t *fpo; + __u64 *pages = tx->tx_pages; + __u64 version; + bool is_rx = (rd != tx->tx_rd); + bool tx_pages_mapped = 0; + int npages = 0; + int rc; again: spin_lock(&fps->fps_lock); @@ -1618,21 +1813,138 @@ again: list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) { fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE); fpo->fpo_map_count++; - spin_unlock(&fps->fps_lock); - pfmr = ib_fmr_pool_map_phys(fpo->fpo_fmr_pool, - pages, npages, iov); - if (likely(!IS_ERR(pfmr))) { - fmr->fmr_pool = fpo; - fmr->fmr_pfmr = pfmr; - return 0; - } + if (fpo->fpo_is_fmr) { + struct ib_pool_fmr *pfmr; + + spin_unlock(&fps->fps_lock); + + if (!tx_pages_mapped) { + npages = kiblnd_map_tx_pages(tx, rd); + tx_pages_mapped = 1; + } + + pfmr = ib_fmr_pool_map_phys(fpo->fmr.fpo_fmr_pool, + pages, npages, iov); + if (likely(!IS_ERR(pfmr))) { + fmr->fmr_key = is_rx ? pfmr->fmr->rkey + : pfmr->fmr->lkey; + fmr->fmr_frd = NULL; + fmr->fmr_pfmr = pfmr; + fmr->fmr_pool = fpo; + return 0; + } + rc = PTR_ERR(pfmr); + } else { + if (!list_empty(&fpo->fast_reg.fpo_pool_list)) { + struct kib_fast_reg_descriptor *frd; +#ifdef HAVE_IB_MAP_MR_SG + struct ib_reg_wr *wr; + int n; +#else + struct ib_rdma_wr *wr; + struct ib_fast_reg_page_list *frpl; +#endif + struct ib_mr *mr; + + frd = list_first_entry(&fpo->fast_reg.fpo_pool_list, + struct kib_fast_reg_descriptor, + frd_list); + list_del(&frd->frd_list); + spin_unlock(&fps->fps_lock); + +#ifndef HAVE_IB_MAP_MR_SG + frpl = frd->frd_frpl; +#endif + mr = frd->frd_mr; + + if (!frd->frd_valid) { + struct ib_rdma_wr *inv_wr; + __u32 key = is_rx ? mr->rkey : mr->lkey; + + inv_wr = &frd->frd_inv_wr; + memset(inv_wr, 0, sizeof(*inv_wr)); + + inv_wr->wr.opcode = IB_WR_LOCAL_INV; + inv_wr->wr.wr_id = IBLND_WID_MR; + inv_wr->wr.ex.invalidate_rkey = key; + + /* Bump the key */ + key = ib_inc_rkey(key); + ib_update_fast_reg_key(mr, key); + } + +#ifdef HAVE_IB_MAP_MR_SG +#ifdef HAVE_IB_MAP_MR_SG_5ARGS + n = ib_map_mr_sg(mr, tx->tx_frags, + tx->tx_nfrags, NULL, PAGE_SIZE); +#else + n = ib_map_mr_sg(mr, tx->tx_frags, + tx->tx_nfrags, PAGE_SIZE); +#endif + if (unlikely(n != tx->tx_nfrags)) { + CERROR("Failed to map mr %d/%d " + "elements\n", n, tx->tx_nfrags); + return n < 0 ? n : -EINVAL; + } + + mr->iova = iov; + + wr = &frd->frd_fastreg_wr; + memset(wr, 0, sizeof(*wr)); + + wr->wr.opcode = IB_WR_REG_MR; + wr->wr.wr_id = IBLND_WID_MR; + wr->wr.num_sge = 0; + wr->wr.send_flags = 0; + wr->mr = mr; + wr->key = is_rx ? mr->rkey : mr->lkey; + wr->access = (IB_ACCESS_LOCAL_WRITE | + IB_ACCESS_REMOTE_WRITE); +#else + if (!tx_pages_mapped) { + npages = kiblnd_map_tx_pages(tx, rd); + tx_pages_mapped = 1; + } + + LASSERT(npages <= frpl->max_page_list_len); + memcpy(frpl->page_list, pages, + sizeof(*pages) * npages); + + /* Prepare FastReg WR */ + wr = &frd->frd_fastreg_wr; + memset(wr, 0, sizeof(*wr)); + + wr->wr.opcode = IB_WR_FAST_REG_MR; + wr->wr.wr_id = IBLND_WID_MR; + + wr->wr.wr.fast_reg.iova_start = iov; + wr->wr.wr.fast_reg.page_list = frpl; + wr->wr.wr.fast_reg.page_list_len = npages; + wr->wr.wr.fast_reg.page_shift = PAGE_SHIFT; + wr->wr.wr.fast_reg.length = nob; + wr->wr.wr.fast_reg.rkey = + is_rx ? mr->rkey : mr->lkey; + wr->wr.wr.fast_reg.access_flags = + (IB_ACCESS_LOCAL_WRITE | + IB_ACCESS_REMOTE_WRITE); +#endif + + fmr->fmr_key = is_rx ? mr->rkey : mr->lkey; + fmr->fmr_frd = frd; + fmr->fmr_pfmr = NULL; + fmr->fmr_pool = fpo; + return 0; + } + spin_unlock(&fps->fps_lock); + rc = -EBUSY; + } spin_lock(&fps->fps_lock); fpo->fpo_map_count--; - if (PTR_ERR(pfmr) != -EAGAIN) { + if (rc != -EAGAIN) { spin_unlock(&fps->fps_lock); - return PTR_ERR(pfmr); + return rc; } /* EAGAIN and ... */ @@ -1822,6 +2134,9 @@ kiblnd_pool_alloc_node(kib_poolset_t *ps) struct list_head *node; kib_pool_t *pool; int rc; + unsigned int interval = 1; + cfs_time_t time_before; + unsigned int trips = 0; again: spin_lock(&ps->ps_lock); @@ -1846,10 +2161,17 @@ again: if (ps->ps_increasing) { /* another thread is allocating a new pool */ spin_unlock(&ps->ps_lock); + trips++; CDEBUG(D_NET, "Another thread is allocating new " - "%s pool, waiting for her to complete\n", - ps->ps_name); - schedule(); + "%s pool, waiting %d HZs for her to complete." + "trips = %d\n", + ps->ps_name, interval, trips); + + set_current_state(TASK_INTERRUPTIBLE); + schedule_timeout(interval); + if (interval < cfs_time_seconds(1)) + interval *= 2; + goto again; } @@ -1863,8 +2185,10 @@ again: spin_unlock(&ps->ps_lock); CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name); - + time_before = cfs_time_current(); rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool); + CDEBUG(D_NET, "ps_pool_create took %lu HZ to complete", + cfs_time_current() - time_before); spin_lock(&ps->ps_lock); ps->ps_increasing = 0; @@ -1906,8 +2230,8 @@ kiblnd_destroy_tx_pool(kib_pool_t *pool) sizeof(*tx->tx_pages)); if (tx->tx_frags != NULL) LIBCFS_FREE(tx->tx_frags, - IBLND_MAX_RDMA_FRAGS * - sizeof(*tx->tx_frags)); + (1 + IBLND_MAX_RDMA_FRAGS) * + sizeof(*tx->tx_frags)); if (tx->tx_wrq != NULL) LIBCFS_FREE(tx->tx_wrq, (1 + IBLND_MAX_RDMA_FRAGS) * @@ -1985,11 +2309,12 @@ kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po) } LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt, - IBLND_MAX_RDMA_FRAGS * sizeof(*tx->tx_frags)); + (1 + IBLND_MAX_RDMA_FRAGS) * + sizeof(*tx->tx_frags)); if (tx->tx_frags == NULL) break; - sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS); + sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS + 1); LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt, (1 + IBLND_MAX_RDMA_FRAGS) * @@ -2062,15 +2387,18 @@ kiblnd_net_fini_pools(kib_net_t *net) } static int -kiblnd_net_init_pools(kib_net_t *net, __u32 *cpts, int ncpts) +kiblnd_net_init_pools(kib_net_t *net, lnet_ni_t *ni, __u32 *cpts, int ncpts) { + struct lnet_ioctl_config_o2iblnd_tunables *tunables; unsigned long flags; int cpt; int rc; int i; + tunables = &ni->ni_lnd_tunables->lt_tun_u.lt_o2ib; + read_lock_irqsave(&kiblnd_data.kib_global_lock, flags); - if (*kiblnd_tunables.kib_map_on_demand == 0) { + if (tunables->lnd_map_on_demand == 0) { read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); goto create_tx_pool; @@ -2078,10 +2406,9 @@ kiblnd_net_init_pools(kib_net_t *net, __u32 *cpts, int ncpts) read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); - if (*kiblnd_tunables.kib_fmr_pool_size < - *kiblnd_tunables.kib_ntx / 4) { + if (tunables->lnd_fmr_pool_size < *kiblnd_tunables.kib_ntx / 4) { CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n", - *kiblnd_tunables.kib_fmr_pool_size, + tunables->lnd_fmr_pool_size, *kiblnd_tunables.kib_ntx / 4); rc = -EINVAL; goto failed; @@ -2104,9 +2431,8 @@ kiblnd_net_init_pools(kib_net_t *net, __u32 *cpts, int ncpts) for (i = 0; i < ncpts; i++) { cpt = (cpts == NULL) ? i : cpts[i]; - rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, net, - kiblnd_fmr_pool_size(ncpts), - kiblnd_fmr_flush_trigger(ncpts)); + rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, ncpts, + net, tunables); if (rc != 0) { CERROR("Can't initialize FMR pool for CPT %d: %d\n", cpt, rc); @@ -2151,8 +2477,10 @@ kiblnd_net_init_pools(kib_net_t *net, __u32 *cpts, int ncpts) static int kiblnd_hdev_get_attr(kib_hca_dev_t *hdev) { - struct ib_device_attr *attr; - int rc; +#ifndef HAVE_IB_DEVICE_ATTRS + struct ib_device_attr *attr; + int rc; +#endif /* It's safe to assume a HCA can handle a page size * matching that of the native system */ @@ -2160,6 +2488,9 @@ kiblnd_hdev_get_attr(kib_hca_dev_t *hdev) hdev->ibh_page_size = 1 << PAGE_SHIFT; hdev->ibh_page_mask = ~((__u64)hdev->ibh_page_size - 1); +#ifdef HAVE_IB_DEVICE_ATTRS + hdev->ibh_mr_size = hdev->ibh_ibdev->attrs.max_mr_size; +#else LIBCFS_ALLOC(attr, sizeof(*attr)); if (attr == NULL) { CERROR("Out of memory\n"); @@ -2176,13 +2507,14 @@ kiblnd_hdev_get_attr(kib_hca_dev_t *hdev) CERROR("Failed to query IB device: %d\n", rc); return rc; } +#endif if (hdev->ibh_mr_size == ~0ULL) { hdev->ibh_mr_shift = 64; return 0; } - CERROR("Invalid mr size: "LPX64"\n", hdev->ibh_mr_size); + CERROR("Invalid mr size: %#llx\n", hdev->ibh_mr_size); return -EINVAL; } @@ -2468,7 +2800,6 @@ kiblnd_create_dev(char *ifname) if (dev == NULL) return NULL; - memset(dev, 0, sizeof(*dev)); netdev = dev_get_by_name(&init_net, ifname); if (netdev == NULL) { dev->ibd_can_failover = 0; @@ -2519,6 +2850,8 @@ kiblnd_base_shutdown(void) } LASSERT(list_empty(&kiblnd_data.kib_connd_zombies)); LASSERT(list_empty(&kiblnd_data.kib_connd_conns)); + LASSERT(list_empty(&kiblnd_data.kib_reconn_list)); + LASSERT(list_empty(&kiblnd_data.kib_reconn_wait)); /* flag threads to terminate; wake and wait for them to die */ kiblnd_data.kib_shutdown = 1; @@ -2670,6 +3003,9 @@ kiblnd_base_startup(void) spin_lock_init(&kiblnd_data.kib_connd_lock); INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns); INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies); + INIT_LIST_HEAD(&kiblnd_data.kib_reconn_list); + INIT_LIST_HEAD(&kiblnd_data.kib_reconn_wait); + init_waitqueue_head(&kiblnd_data.kib_connd_waitq); init_waitqueue_head(&kiblnd_data.kib_failover_waitq); @@ -2854,15 +3190,10 @@ kiblnd_startup (lnet_ni_t *ni) if (net == NULL) goto failed; - memset(net, 0, sizeof(*net)); - do_gettimeofday(&tv); net->ibn_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec; - ni->ni_peertimeout = *kiblnd_tunables.kib_peertimeout; - ni->ni_maxtxcredits = *kiblnd_tunables.kib_credits; - ni->ni_peertxcredits = *kiblnd_tunables.kib_peertxcredits; - ni->ni_peerrtrcredits = *kiblnd_tunables.kib_peerrtrcredits; + kiblnd_tunables_setup(ni); if (ni->ni_interfaces[0] != NULL) { /* Use the IPoIB interface specified in 'networks=' */ @@ -2901,7 +3232,7 @@ kiblnd_startup (lnet_ni_t *ni) if (rc != 0) goto failed; - rc = kiblnd_net_init_pools(net, ni->ni_cpts, ni->ni_ncpts); + rc = kiblnd_net_init_pools(net, ni, ni->ni_cpts, ni->ni_ncpts); if (rc != 0) { CERROR("Failed to initialize NI pools: %d\n", rc); goto failed; @@ -2939,7 +3270,6 @@ static lnd_t the_o2iblnd = { static void __exit ko2iblnd_exit(void) { lnet_unregister_lnd(&the_o2iblnd); - kiblnd_tunables_fini(); } static int __init ko2iblnd_init(void)