X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lnet%2Fklnds%2Fo2iblnd%2Fo2iblnd.c;h=86298dd68e56c205a9a86312fa89d458ff353c30;hb=c8fd9c3c38a66ab3dd479c48e0168577a4ff960e;hp=c5e63eadf3612e0406583f2aeb82f6e7ec29dd87;hpb=e3a7c58aebafce40323db54bf6056029e5af4a70;p=fs%2Flustre-release.git diff --git a/lnet/klnds/o2iblnd/o2iblnd.c b/lnet/klnds/o2iblnd/o2iblnd.c index c5e63ea..86298dd 100644 --- a/lnet/klnds/o2iblnd/o2iblnd.c +++ b/lnet/klnds/o2iblnd/o2iblnd.c @@ -27,7 +27,7 @@ * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. * - * Copyright (c) 2011, Whamcloud, Inc. + * Copyright (c) 2011, 2012, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -324,16 +324,17 @@ kiblnd_unpack_msg(kib_msg_t *msg, int nob) } int -kiblnd_create_peer (lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid) +kiblnd_create_peer(lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid) { - kib_peer_t *peer; - kib_net_t *net = ni->ni_data; - unsigned long flags; + kib_peer_t *peer; + kib_net_t *net = ni->ni_data; + int cpt = lnet_cpt_of_nid(nid); + unsigned long flags; - LASSERT (net != NULL); - LASSERT (nid != LNET_NID_ANY); + LASSERT(net != NULL); + LASSERT(nid != LNET_NID_ANY); - LIBCFS_ALLOC(peer, sizeof(*peer)); + LIBCFS_CPT_ALLOC(peer, lnet_cpt_table(), cpt, sizeof(*peer)); if (peer == NULL) { CERROR("Cannot allocate peer\n"); return -ENOMEM; @@ -351,7 +352,7 @@ kiblnd_create_peer (lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid) CFS_INIT_LIST_HEAD(&peer->ibp_conns); CFS_INIT_LIST_HEAD(&peer->ibp_tx_queue); - cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags); + write_lock_irqsave(&kiblnd_data.kib_global_lock, flags); /* always called with a ref on ni, which prevents ni being shutdown */ LASSERT (net->ibn_shutdown == 0); @@ -359,7 +360,7 @@ kiblnd_create_peer (lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid) /* npeers only grows with the global lock held */ cfs_atomic_inc(&net->ibn_npeers); - cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); + write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); *peerp = peer; return 0; @@ -436,7 +437,7 @@ kiblnd_get_peer_info (lnet_ni_t *ni, int index, int i; unsigned long flags; - cfs_read_lock_irqsave(&kiblnd_data.kib_global_lock, flags); + read_lock_irqsave(&kiblnd_data.kib_global_lock, flags); for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) { @@ -456,13 +457,13 @@ kiblnd_get_peer_info (lnet_ni_t *ni, int index, *nidp = peer->ibp_nid; *count = cfs_atomic_read(&peer->ibp_refcount); - cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, - flags); - return 0; - } - } + read_unlock_irqrestore(&kiblnd_data.kib_global_lock, + flags); + return 0; + } + } - cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); + read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); return -ENOENT; } @@ -500,7 +501,7 @@ kiblnd_del_peer (lnet_ni_t *ni, lnet_nid_t nid) unsigned long flags; int rc = -ENOENT; - cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags); + write_lock_irqsave(&kiblnd_data.kib_global_lock, flags); if (nid != LNET_NID_ANY) { lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers; @@ -534,7 +535,7 @@ kiblnd_del_peer (lnet_ni_t *ni, lnet_nid_t nid) } } - cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); + write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); kiblnd_txlist_done(ni, &zombies, -EIO); @@ -551,7 +552,7 @@ kiblnd_get_conn_by_idx (lnet_ni_t *ni, int index) int i; unsigned long flags; - cfs_read_lock_irqsave(&kiblnd_data.kib_global_lock, flags); + read_lock_irqsave(&kiblnd_data.kib_global_lock, flags); for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) { cfs_list_for_each (ptmp, &kiblnd_data.kib_peers[i]) { @@ -571,14 +572,14 @@ kiblnd_get_conn_by_idx (lnet_ni_t *ni, int index) conn = cfs_list_entry(ctmp, kib_conn_t, ibc_list); kiblnd_conn_addref(conn); - cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, - flags); - return conn; - } - } - } - - cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); + read_unlock_irqrestore(&kiblnd_data.kib_global_lock, + flags); + return conn; + } + } + } + + read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); return NULL; } @@ -605,10 +606,10 @@ kiblnd_debug_tx (kib_tx_t *tx) void kiblnd_debug_conn (kib_conn_t *conn) { - cfs_list_t *tmp; - int i; + cfs_list_t *tmp; + int i; - cfs_spin_lock(&conn->ibc_lock); + spin_lock(&conn->ibc_lock); CDEBUG(D_CONSOLE, "conn[%d] %p [version %x] -> %s: \n", cfs_atomic_read(&conn->ibc_refcount), conn, @@ -647,7 +648,7 @@ kiblnd_debug_conn (kib_conn_t *conn) for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++) kiblnd_debug_rx(&conn->ibc_rxs[i]); - cfs_spin_unlock(&conn->ibc_lock); + spin_unlock(&conn->ibc_lock); } int @@ -686,6 +687,33 @@ kiblnd_setup_mtu_locked(struct rdma_cm_id *cmid) cmid->route.path_rec->mtu = mtu; } +#ifdef HAVE_OFED_IB_COMP_VECTOR +static int +kiblnd_get_completion_vector(kib_conn_t *conn, int cpt) +{ + cpumask_t *mask; + int vectors; + int off; + int i; + + vectors = conn->ibc_cmid->device->num_comp_vectors; + if (vectors <= 1) + return 0; + + mask = cfs_cpt_cpumask(lnet_cpt_table(), cpt); + + /* hash NID to CPU id in this partition... */ + off = conn->ibc_peer->ibp_nid % cpus_weight(*mask); + for_each_cpu_mask(i, *mask) { + if (off-- == 0) + return i % vectors; + } + + LBUG(); + return 1; +} +#endif + kib_conn_t * kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid, int state, int version) @@ -697,35 +725,43 @@ kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid, * she must dispose of 'cmid'. (Actually I'd block forever if I tried * to destroy 'cmid' here since I'm called from the CM which still has * its ref on 'cmid'). */ - cfs_rwlock_t *glock = &kiblnd_data.kib_global_lock; + rwlock_t *glock = &kiblnd_data.kib_global_lock; kib_net_t *net = peer->ibp_ni->ni_data; - kib_dev_t *dev = net->ibn_dev; + kib_dev_t *dev; struct ib_qp_init_attr *init_qp_attr; - kib_conn_t *conn; - struct ib_cq *cq; - unsigned long flags; - int rc; - int i; + struct kib_sched_info *sched; + kib_conn_t *conn; + struct ib_cq *cq; + unsigned long flags; + int cpt; + int rc; + int i; - LASSERT (net != NULL); - LASSERT (!cfs_in_interrupt()); + LASSERT(net != NULL); + LASSERT(!cfs_in_interrupt()); - LIBCFS_ALLOC(init_qp_attr, sizeof(*init_qp_attr)); - if (init_qp_attr == NULL) { - CERROR("Can't allocate qp_attr for %s\n", - libcfs_nid2str(peer->ibp_nid)); - goto failed_0; - } + dev = net->ibn_dev; + + cpt = lnet_cpt_of_nid(peer->ibp_nid); + sched = kiblnd_data.kib_scheds[cpt]; - LIBCFS_ALLOC(conn, sizeof(*conn)); + LASSERT(sched->ibs_nthreads > 0); + + LIBCFS_CPT_ALLOC(init_qp_attr, lnet_cpt_table(), cpt, + sizeof(*init_qp_attr)); + if (init_qp_attr == NULL) { + CERROR("Can't allocate qp_attr for %s\n", + libcfs_nid2str(peer->ibp_nid)); + goto failed_0; + } + + LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn)); if (conn == NULL) { CERROR("Can't allocate connection for %s\n", libcfs_nid2str(peer->ibp_nid)); goto failed_1; } - memset(conn, 0, sizeof(*conn)); /* zero flags, NULL pointers etc... */ - conn->ibc_state = IBLND_CONN_INIT; conn->ibc_version = version; conn->ibc_peer = peer; /* I take the caller's ref */ @@ -738,21 +774,21 @@ kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid, CFS_INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd); CFS_INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred); CFS_INIT_LIST_HEAD(&conn->ibc_active_txs); - cfs_spin_lock_init(&conn->ibc_lock); - - LIBCFS_ALLOC(conn->ibc_connvars, sizeof(*conn->ibc_connvars)); - if (conn->ibc_connvars == NULL) { - CERROR("Can't allocate in-progress connection state\n"); - goto failed_2; - } - memset(conn->ibc_connvars, 0, sizeof(*conn->ibc_connvars)); - - cfs_write_lock_irqsave(glock, flags); - if (dev->ibd_failover) { - cfs_write_unlock_irqrestore(glock, flags); - CERROR("%s: failover in progress\n", dev->ibd_ifname); - goto failed_2; - } + spin_lock_init(&conn->ibc_lock); + + LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt, + sizeof(*conn->ibc_connvars)); + if (conn->ibc_connvars == NULL) { + CERROR("Can't allocate in-progress connection state\n"); + goto failed_2; + } + + write_lock_irqsave(glock, flags); + if (dev->ibd_failover) { + write_unlock_irqrestore(glock, flags); + CERROR("%s: failover in progress\n", dev->ibd_ifname); + goto failed_2; + } if (dev->ibd_hdev->ibh_ibdev != cmid->device) { /* wakeup failover thread and teardown connection */ @@ -762,7 +798,7 @@ kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid, cfs_waitq_signal(&kiblnd_data.kib_failover_waitq); } - cfs_write_unlock_irqrestore(glock, flags); + write_unlock_irqrestore(glock, flags); CERROR("cmid HCA(%s), kib_dev(%s) need failover\n", cmid->device->name, dev->ibd_ifname); goto failed_2; @@ -773,26 +809,27 @@ kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid, kiblnd_setup_mtu_locked(cmid); - cfs_write_unlock_irqrestore(glock, flags); + write_unlock_irqrestore(glock, flags); - LIBCFS_ALLOC(conn->ibc_rxs, IBLND_RX_MSGS(version) * sizeof(kib_rx_t)); - if (conn->ibc_rxs == NULL) { - CERROR("Cannot allocate RX buffers\n"); - goto failed_2; - } - memset(conn->ibc_rxs, 0, IBLND_RX_MSGS(version) * sizeof(kib_rx_t)); + LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt, + IBLND_RX_MSGS(version) * sizeof(kib_rx_t)); + if (conn->ibc_rxs == NULL) { + CERROR("Cannot allocate RX buffers\n"); + goto failed_2; + } - rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, - IBLND_RX_MSG_PAGES(version)); - if (rc != 0) - goto failed_2; + rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt, + IBLND_RX_MSG_PAGES(version)); + if (rc != 0) + goto failed_2; - kiblnd_map_rx_descs(conn); + kiblnd_map_rx_descs(conn); #ifdef HAVE_OFED_IB_COMP_VECTOR - cq = ib_create_cq(cmid->device, - kiblnd_cq_completion, kiblnd_cq_event, conn, - IBLND_CQ_ENTRIES(version), 0); + cq = ib_create_cq(cmid->device, + kiblnd_cq_completion, kiblnd_cq_event, conn, + IBLND_CQ_ENTRIES(version), + kiblnd_get_completion_vector(conn, cpt)); #else cq = ib_create_cq(cmid->device, kiblnd_cq_completion, kiblnd_cq_event, conn, @@ -812,7 +849,6 @@ kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid, goto failed_2; } - memset(init_qp_attr, 0, sizeof(*init_qp_attr)); init_qp_attr->event_handler = kiblnd_qp_event; init_qp_attr->qp_context = conn; init_qp_attr->cap.max_send_wr = IBLND_SEND_WRS(version); @@ -824,6 +860,8 @@ kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid, init_qp_attr->send_cq = cq; init_qp_attr->recv_cq = cq; + 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", @@ -850,11 +888,9 @@ kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid, /* correct # of posted buffers * NB locking needed now I'm racing with completion */ - cfs_spin_lock_irqsave(&kiblnd_data.kib_sched_lock, - flags); - conn->ibc_nrx -= IBLND_RX_MSGS(version) - i; - cfs_spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock, - flags); + spin_lock_irqsave(&sched->ibs_lock, flags); + conn->ibc_nrx -= IBLND_RX_MSGS(version) - i; + spin_unlock_irqrestore(&sched->ibs_lock, flags); /* cmid will be destroyed by CM(ofed) after cm_callback * returned, so we can't refer it anymore @@ -1019,7 +1055,7 @@ kiblnd_close_matching_conns (lnet_ni_t *ni, lnet_nid_t nid) unsigned long flags; int count = 0; - cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags); + write_lock_irqsave(&kiblnd_data.kib_global_lock, flags); if (nid != LNET_NID_ANY) lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers; @@ -1046,7 +1082,7 @@ kiblnd_close_matching_conns (lnet_ni_t *ni, lnet_nid_t nid) } } - cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); + write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); /* wildcards always succeed */ if (nid == LNET_NID_ANY) @@ -1112,13 +1148,13 @@ kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg) void kiblnd_query (lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when) { - cfs_time_t last_alive = 0; - cfs_time_t now = cfs_time_current(); - cfs_rwlock_t *glock = &kiblnd_data.kib_global_lock; - kib_peer_t *peer; - unsigned long flags; + cfs_time_t last_alive = 0; + cfs_time_t now = cfs_time_current(); + rwlock_t *glock = &kiblnd_data.kib_global_lock; + kib_peer_t *peer; + unsigned long flags; - cfs_read_lock_irqsave(glock, flags); + read_lock_irqsave(glock, flags); peer = kiblnd_find_peer_locked(nid); if (peer != NULL) { @@ -1128,7 +1164,7 @@ kiblnd_query (lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when) last_alive = peer->ibp_last_alive; } - cfs_read_unlock_irqrestore(glock, flags); + read_unlock_irqrestore(glock, flags); if (last_alive != 0) *when = last_alive; @@ -1145,25 +1181,27 @@ kiblnd_query (lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when) } void -kiblnd_free_pages (kib_pages_t *p) +kiblnd_free_pages(kib_pages_t *p) { - int npages = p->ibp_npages; - int i; + int npages = p->ibp_npages; + int i; - for (i = 0; i < npages; i++) - if (p->ibp_pages[i] != NULL) - __free_page(p->ibp_pages[i]); + for (i = 0; i < npages; i++) { + if (p->ibp_pages[i] != NULL) + cfs_free_page(p->ibp_pages[i]); + } - LIBCFS_FREE (p, offsetof(kib_pages_t, ibp_pages[npages])); + LIBCFS_FREE(p, offsetof(kib_pages_t, ibp_pages[npages])); } int -kiblnd_alloc_pages (kib_pages_t **pp, int npages) +kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages) { - kib_pages_t *p; - int i; + kib_pages_t *p; + int i; - LIBCFS_ALLOC(p, offsetof(kib_pages_t, ibp_pages[npages])); + LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt, + offsetof(kib_pages_t, ibp_pages[npages])); if (p == NULL) { CERROR("Can't allocate descriptor for %d pages\n", npages); return -ENOMEM; @@ -1173,7 +1211,8 @@ kiblnd_alloc_pages (kib_pages_t **pp, int npages) p->ibp_npages = npages; for (i = 0; i < npages; i++) { - p->ibp_pages[i] = alloc_page(GFP_KERNEL); + p->ibp_pages[i] = cfs_page_cpt_alloc(lnet_cpt_table(), cpt, + CFS_ALLOC_IO); if (p->ibp_pages[i] == NULL) { CERROR("Can't allocate page %d of %d\n", i, npages); kiblnd_free_pages(p); @@ -1280,22 +1319,23 @@ kiblnd_current_hdev(kib_dev_t *dev) unsigned long flags; int i = 0; - cfs_read_lock_irqsave(&kiblnd_data.kib_global_lock, flags); - while (dev->ibd_failover) { - cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); - if (i++ % 50 == 0) - CDEBUG(D_NET, "Wait for dev(%s) failover\n", dev->ibd_ifname); - cfs_schedule_timeout(cfs_time_seconds(1) / 100); + read_lock_irqsave(&kiblnd_data.kib_global_lock, flags); + while (dev->ibd_failover) { + read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); + if (i++ % 50 == 0) + CDEBUG(D_NET, "%s: Wait for failover\n", + dev->ibd_ifname); + cfs_schedule_timeout(cfs_time_seconds(1) / 100); - cfs_read_lock_irqsave(&kiblnd_data.kib_global_lock, flags); - } + read_lock_irqsave(&kiblnd_data.kib_global_lock, flags); + } - kiblnd_hdev_addref_locked(dev->ibd_hdev); - hdev = dev->ibd_hdev; + kiblnd_hdev_addref_locked(dev->ibd_hdev); + hdev = dev->ibd_hdev; - cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); + read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); - return hdev; + return hdev; } static void @@ -1304,7 +1344,7 @@ kiblnd_map_tx_pool(kib_tx_pool_t *tpo) kib_pages_t *txpgs = tpo->tpo_tx_pages; kib_pool_t *pool = &tpo->tpo_pool; kib_net_t *net = pool->po_owner->ps_net; - kib_dev_t *dev = net->ibn_dev; + kib_dev_t *dev; struct page *page; kib_tx_t *tx; int page_offset; @@ -1313,6 +1353,8 @@ kiblnd_map_tx_pool(kib_tx_pool_t *tpo) LASSERT (net != NULL); + dev = net->ibn_dev; + /* pre-mapped messages are not bigger than 1 page */ CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE); @@ -1427,6 +1469,20 @@ kiblnd_destroy_fmr_pool_list(cfs_list_t *head) } } +static int kiblnd_fmr_pool_size(int ncpts) +{ + int size = *kiblnd_tunables.kib_fmr_pool_size / ncpts; + + return max(IBLND_FMR_POOL, size); +} + +static int kiblnd_fmr_flush_trigger(int ncpts) +{ + int size = *kiblnd_tunables.kib_fmr_flush_trigger / ncpts; + + return max(IBLND_FMR_POOL_FLUSH, size); +} + int kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo) { @@ -1438,24 +1494,24 @@ kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo) .page_shift = PAGE_SHIFT, .access = (IB_ACCESS_LOCAL_WRITE | IB_ACCESS_REMOTE_WRITE), - .pool_size = *kiblnd_tunables.kib_fmr_pool_size, - .dirty_watermark = *kiblnd_tunables.kib_fmr_flush_trigger, - .flush_function = NULL, - .flush_arg = NULL, - .cache = !!*kiblnd_tunables.kib_fmr_cache}; - int rc; - - LIBCFS_ALLOC(fpo, sizeof(kib_fmr_pool_t)); - if (fpo == NULL) - return -ENOMEM; + .pool_size = fps->fps_pool_size, + .dirty_watermark = fps->fps_flush_trigger, + .flush_function = NULL, + .flush_arg = NULL, + .cache = !!*kiblnd_tunables.kib_fmr_cache}; + int rc; + + LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo)); + if (fpo == NULL) + return -ENOMEM; + + 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); - memset(fpo, 0, sizeof(kib_fmr_pool_t)); - 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)) { - CERROR("Failed to create FMR pool: %ld\n", - PTR_ERR(fpo->fpo_fmr_pool)); - rc = PTR_ERR(fpo->fpo_fmr_pool); kiblnd_hdev_decref(fpo->fpo_hdev); LIBCFS_FREE(fpo, sizeof(kib_fmr_pool_t)); return rc; @@ -1469,9 +1525,12 @@ kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo) } static void -kiblnd_fail_fmr_pool_set(kib_fmr_poolset_t *fps, cfs_list_t *zombies) +kiblnd_fail_fmr_poolset(kib_fmr_poolset_t *fps, cfs_list_t *zombies) { - cfs_spin_lock(&fps->fps_lock); + if (fps->fps_net == NULL) /* intialized? */ + return; + + spin_lock(&fps->fps_lock); while (!cfs_list_empty(&fps->fps_pool_list)) { kib_fmr_pool_t *fpo = cfs_list_entry(fps->fps_pool_list.next, @@ -1484,18 +1543,21 @@ kiblnd_fail_fmr_pool_set(kib_fmr_poolset_t *fps, cfs_list_t *zombies) cfs_list_add(&fpo->fpo_list, &fps->fps_failed_pool_list); } - cfs_spin_unlock(&fps->fps_lock); + spin_unlock(&fps->fps_lock); } static void -kiblnd_fini_fmr_pool_set(kib_fmr_poolset_t *fps) +kiblnd_fini_fmr_poolset(kib_fmr_poolset_t *fps) { - kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list); - kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list); + if (fps->fps_net != NULL) { /* initialized? */ + kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list); + kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list); + } } static int -kiblnd_init_fmr_pool_set(kib_fmr_poolset_t *fps, kib_net_t *net) +kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt, kib_net_t *net, + int pool_size, int flush_trigger) { kib_fmr_pool_t *fpo; int rc; @@ -1503,9 +1565,13 @@ kiblnd_init_fmr_pool_set(kib_fmr_poolset_t *fps, kib_net_t *net) memset(fps, 0, sizeof(kib_fmr_poolset_t)); fps->fps_net = net; - cfs_spin_lock_init(&fps->fps_lock); - CFS_INIT_LIST_HEAD(&fps->fps_pool_list); - CFS_INIT_LIST_HEAD(&fps->fps_failed_pool_list); + fps->fps_cpt = cpt; + fps->fps_pool_size = pool_size; + fps->fps_flush_trigger = flush_trigger; + spin_lock_init(&fps->fps_lock); + CFS_INIT_LIST_HEAD(&fps->fps_pool_list); + CFS_INIT_LIST_HEAD(&fps->fps_failed_pool_list); + rc = kiblnd_create_fmr_pool(fps, &fpo); if (rc == 0) cfs_list_add_tail(&fpo->fpo_list, &fps->fps_pool_list); @@ -1544,7 +1610,7 @@ kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status) fmr->fmr_pool = NULL; fmr->fmr_pfmr = NULL; - cfs_spin_lock(&fps->fps_lock); + spin_lock(&fps->fps_lock); fpo->fpo_map_count --; /* decref the pool */ cfs_list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) { @@ -1557,7 +1623,7 @@ kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status) fps->fps_version ++; } } - cfs_spin_unlock(&fps->fps_lock); + spin_unlock(&fps->fps_lock); if (!cfs_list_empty(&zombies)) kiblnd_destroy_fmr_pool_list(&zombies); @@ -1572,14 +1638,13 @@ kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages, __u64 version; int rc; - LASSERT (fps->fps_net->ibn_with_fmr); again: - cfs_spin_lock(&fps->fps_lock); - version = fps->fps_version; - cfs_list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) { - fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE); - fpo->fpo_map_count ++; - cfs_spin_unlock(&fps->fps_lock); + spin_lock(&fps->fps_lock); + version = fps->fps_version; + cfs_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); @@ -1589,51 +1654,51 @@ kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages, return 0; } - cfs_spin_lock(&fps->fps_lock); - fpo->fpo_map_count --; - if (PTR_ERR(pfmr) != -EAGAIN) { - cfs_spin_unlock(&fps->fps_lock); - return PTR_ERR(pfmr); - } - - /* EAGAIN and ... */ - if (version != fps->fps_version) { - cfs_spin_unlock(&fps->fps_lock); - goto again; - } - } - - if (fps->fps_increasing) { - cfs_spin_unlock(&fps->fps_lock); - CDEBUG(D_NET, "Another thread is allocating new " - "FMR pool, waiting for her to complete\n"); - cfs_schedule(); - goto again; - - } - - if (cfs_time_before(cfs_time_current(), fps->fps_next_retry)) { - /* someone failed recently */ - cfs_spin_unlock(&fps->fps_lock); - return -EAGAIN; - } - - fps->fps_increasing = 1; - cfs_spin_unlock(&fps->fps_lock); - - CDEBUG(D_NET, "Allocate new FMR pool\n"); - rc = kiblnd_create_fmr_pool(fps, &fpo); - cfs_spin_lock(&fps->fps_lock); - fps->fps_increasing = 0; - if (rc == 0) { - fps->fps_version ++; - cfs_list_add_tail(&fpo->fpo_list, &fps->fps_pool_list); - } else { - fps->fps_next_retry = cfs_time_shift(IBLND_POOL_RETRY); - } - cfs_spin_unlock(&fps->fps_lock); - - goto again; + spin_lock(&fps->fps_lock); + fpo->fpo_map_count--; + if (PTR_ERR(pfmr) != -EAGAIN) { + spin_unlock(&fps->fps_lock); + return PTR_ERR(pfmr); + } + + /* EAGAIN and ... */ + if (version != fps->fps_version) { + spin_unlock(&fps->fps_lock); + goto again; + } + } + + if (fps->fps_increasing) { + spin_unlock(&fps->fps_lock); + CDEBUG(D_NET, "Another thread is allocating new " + "FMR pool, waiting for her to complete\n"); + cfs_schedule(); + goto again; + + } + + if (cfs_time_before(cfs_time_current(), fps->fps_next_retry)) { + /* someone failed recently */ + spin_unlock(&fps->fps_lock); + return -EAGAIN; + } + + fps->fps_increasing = 1; + spin_unlock(&fps->fps_lock); + + CDEBUG(D_NET, "Allocate new FMR pool\n"); + rc = kiblnd_create_fmr_pool(fps, &fpo); + spin_lock(&fps->fps_lock); + fps->fps_increasing = 0; + if (rc == 0) { + fps->fps_version++; + cfs_list_add_tail(&fpo->fpo_list, &fps->fps_pool_list); + } else { + fps->fps_next_retry = cfs_time_shift(IBLND_POOL_RETRY); + } + spin_unlock(&fps->fps_lock); + + goto again; } static void @@ -1672,9 +1737,12 @@ kiblnd_destroy_pool_list(cfs_list_t *head) } static void -kiblnd_fail_pool_set(kib_poolset_t *ps, cfs_list_t *zombies) +kiblnd_fail_poolset(kib_poolset_t *ps, cfs_list_t *zombies) { - cfs_spin_lock(&ps->ps_lock); + if (ps->ps_net == NULL) /* intialized? */ + return; + + spin_lock(&ps->ps_lock); while (!cfs_list_empty(&ps->ps_pool_list)) { kib_pool_t *po = cfs_list_entry(ps->ps_pool_list.next, kib_pool_t, po_list); @@ -1685,37 +1753,42 @@ kiblnd_fail_pool_set(kib_poolset_t *ps, cfs_list_t *zombies) else cfs_list_add(&po->po_list, &ps->ps_failed_pool_list); } - cfs_spin_unlock(&ps->ps_lock); + spin_unlock(&ps->ps_lock); } static void -kiblnd_fini_pool_set(kib_poolset_t *ps) +kiblnd_fini_poolset(kib_poolset_t *ps) { - kiblnd_destroy_pool_list(&ps->ps_failed_pool_list); - kiblnd_destroy_pool_list(&ps->ps_pool_list); + if (ps->ps_net != NULL) { /* initialized? */ + kiblnd_destroy_pool_list(&ps->ps_failed_pool_list); + kiblnd_destroy_pool_list(&ps->ps_pool_list); + } } static int -kiblnd_init_pool_set(kib_poolset_t *ps, kib_net_t *net, - char *name, int size, - kib_ps_pool_create_t po_create, - kib_ps_pool_destroy_t po_destroy, - kib_ps_node_init_t nd_init, - kib_ps_node_fini_t nd_fini) +kiblnd_init_poolset(kib_poolset_t *ps, int cpt, + kib_net_t *net, char *name, int size, + kib_ps_pool_create_t po_create, + kib_ps_pool_destroy_t po_destroy, + kib_ps_node_init_t nd_init, + kib_ps_node_fini_t nd_fini) { - kib_pool_t *pool; - int rc; + kib_pool_t *pool; + int rc; - memset(ps, 0, sizeof(kib_poolset_t)); + memset(ps, 0, sizeof(kib_poolset_t)); + ps->ps_cpt = cpt; ps->ps_net = net; ps->ps_pool_create = po_create; ps->ps_pool_destroy = po_destroy; ps->ps_node_init = nd_init; ps->ps_node_fini = nd_fini; ps->ps_pool_size = size; - strncpy(ps->ps_name, name, IBLND_POOL_NAME_LEN); - cfs_spin_lock_init(&ps->ps_lock); + if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name)) + >= sizeof(ps->ps_name)) + return -E2BIG; + spin_lock_init(&ps->ps_lock); CFS_INIT_LIST_HEAD(&ps->ps_pool_list); CFS_INIT_LIST_HEAD(&ps->ps_failed_pool_list); @@ -1746,7 +1819,7 @@ kiblnd_pool_free_node(kib_pool_t *pool, cfs_list_t *node) kib_pool_t *tmp; cfs_time_t now = cfs_time_current(); - cfs_spin_lock(&ps->ps_lock); + spin_lock(&ps->ps_lock); if (ps->ps_node_fini != NULL) ps->ps_node_fini(pool, node); @@ -1763,10 +1836,10 @@ kiblnd_pool_free_node(kib_pool_t *pool, cfs_list_t *node) if (kiblnd_pool_is_idle(pool, now)) cfs_list_move(&pool->po_list, &zombies); } - cfs_spin_unlock(&ps->ps_lock); + spin_unlock(&ps->ps_lock); - if (!cfs_list_empty(&zombies)) - kiblnd_destroy_pool_list(&zombies); + if (!cfs_list_empty(&zombies)) + kiblnd_destroy_pool_list(&zombies); } cfs_list_t * @@ -1777,7 +1850,7 @@ kiblnd_pool_alloc_node(kib_poolset_t *ps) int rc; again: - cfs_spin_lock(&ps->ps_lock); + spin_lock(&ps->ps_lock); cfs_list_for_each_entry(pool, &ps->ps_pool_list, po_list) { if (cfs_list_empty(&pool->po_free_list)) continue; @@ -1791,14 +1864,14 @@ kiblnd_pool_alloc_node(kib_poolset_t *ps) /* still hold the lock */ ps->ps_node_init(pool, node); } - cfs_spin_unlock(&ps->ps_lock); - return node; - } - - /* no available tx pool and ... */ - if (ps->ps_increasing) { - /* another thread is allocating a new pool */ - cfs_spin_unlock(&ps->ps_lock); + spin_unlock(&ps->ps_lock); + return node; + } + + /* no available tx pool and ... */ + if (ps->ps_increasing) { + /* another thread is allocating a new pool */ + spin_unlock(&ps->ps_lock); CDEBUG(D_NET, "Another thread is allocating new " "%s pool, waiting for her to complete\n", ps->ps_name); @@ -1806,20 +1879,20 @@ kiblnd_pool_alloc_node(kib_poolset_t *ps) goto again; } - if (cfs_time_before(cfs_time_current(), ps->ps_next_retry)) { - /* someone failed recently */ - cfs_spin_unlock(&ps->ps_lock); - return NULL; - } + if (cfs_time_before(cfs_time_current(), ps->ps_next_retry)) { + /* someone failed recently */ + spin_unlock(&ps->ps_lock); + return NULL; + } - ps->ps_increasing = 1; - cfs_spin_unlock(&ps->ps_lock); + ps->ps_increasing = 1; + spin_unlock(&ps->ps_lock); - CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name); + CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name); - rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool); + rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool); - cfs_spin_lock(&ps->ps_lock); + spin_lock(&ps->ps_lock); ps->ps_increasing = 0; if (rc == 0) { cfs_list_add_tail(&pool->po_list, &ps->ps_pool_list); @@ -1828,9 +1901,9 @@ kiblnd_pool_alloc_node(kib_poolset_t *ps) CERROR("Can't allocate new %s pool because out of memory\n", ps->ps_name); } - cfs_spin_unlock(&ps->ps_lock); + spin_unlock(&ps->ps_lock); - goto again; + goto again; } void @@ -1922,15 +1995,23 @@ kiblnd_destroy_pmr_pool(kib_pool_t *pool) LIBCFS_FREE(ppo, sizeof(kib_pmr_pool_t)); } +static inline int kiblnd_pmr_pool_size(int ncpts) +{ + int size = *kiblnd_tunables.kib_pmr_pool_size / ncpts; + + return max(IBLND_PMR_POOL, size); +} + static int kiblnd_create_pmr_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po) { - kib_pmr_pool_t *ppo; - kib_pool_t *pool; - kib_phys_mr_t *pmr; - int i; + struct kib_pmr_pool *ppo; + struct kib_pool *pool; + kib_phys_mr_t *pmr; + int i; - LIBCFS_ALLOC(ppo, sizeof(kib_pmr_pool_t)); + LIBCFS_CPT_ALLOC(ppo, lnet_cpt_table(), + ps->ps_cpt, sizeof(kib_pmr_pool_t)); if (ppo == NULL) { CERROR("Failed to allocate PMR pool\n"); return -ENOMEM; @@ -1940,15 +2021,14 @@ kiblnd_create_pmr_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po) kiblnd_init_pool(ps, pool, size); for (i = 0; i < size; i++) { - LIBCFS_ALLOC(pmr, sizeof(kib_phys_mr_t)); - if (pmr == NULL) - break; - - memset(pmr, 0, sizeof(kib_phys_mr_t)); - pmr->pmr_pool = ppo; - LIBCFS_ALLOC(pmr->pmr_ipb, - IBLND_MAX_RDMA_FRAGS * - sizeof(struct ib_phys_buf)); + LIBCFS_CPT_ALLOC(pmr, lnet_cpt_table(), + ps->ps_cpt, sizeof(kib_phys_mr_t)); + if (pmr == NULL) + break; + + pmr->pmr_pool = ppo; + LIBCFS_CPT_ALLOC(pmr->pmr_ipb, lnet_cpt_table(), ps->ps_cpt, + IBLND_MAX_RDMA_FRAGS * sizeof(*pmr->pmr_ipb)); if (pmr->pmr_ipb == NULL) break; @@ -2014,6 +2094,13 @@ out: LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t)); } +static int kiblnd_tx_pool_size(int ncpts) +{ + int ntx = *kiblnd_tunables.kib_ntx / ncpts; + + return max(IBLND_TX_POOL, ntx); +} + static int kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po) { @@ -2022,7 +2109,7 @@ kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po) kib_pool_t *pool; kib_tx_pool_t *tpo; - LIBCFS_ALLOC(tpo, sizeof(kib_tx_pool_t)); + LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo)); if (tpo == NULL) { CERROR("Failed to allocate TX pool\n"); return -ENOMEM; @@ -2034,13 +2121,14 @@ kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po) tpo->tpo_tx_pages = NULL; npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE; - if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, npg) != 0) { - CERROR("Can't allocate tx pages: %d\n", npg); - LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t)); - return -ENOMEM; - } - - LIBCFS_ALLOC (tpo->tpo_tx_descs, size * sizeof(kib_tx_t)); + if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) { + CERROR("Can't allocate tx pages: %d\n", npg); + LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t)); + return -ENOMEM; + } + + LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt, + size * sizeof(kib_tx_t)); if (tpo->tpo_tx_descs == NULL) { CERROR("Can't allocate %d tx descriptors\n", size); ps->ps_pool_destroy(pool); @@ -2053,36 +2141,38 @@ kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po) kib_tx_t *tx = &tpo->tpo_tx_descs[i]; tx->tx_pool = tpo; - if (ps->ps_net->ibn_with_fmr){ - LIBCFS_ALLOC(tx->tx_pages, LNET_MAX_IOV * - sizeof(*tx->tx_pages)); - if (tx->tx_pages == NULL) - break; - } - - LIBCFS_ALLOC(tx->tx_frags, - IBLND_MAX_RDMA_FRAGS * - sizeof(*tx->tx_frags)); - if (tx->tx_frags == NULL) - break; - - LIBCFS_ALLOC(tx->tx_wrq, - (1 + IBLND_MAX_RDMA_FRAGS) * - sizeof(*tx->tx_wrq)); - if (tx->tx_wrq == NULL) - break; - - LIBCFS_ALLOC(tx->tx_sge, - (1 + IBLND_MAX_RDMA_FRAGS) * - sizeof(*tx->tx_sge)); - if (tx->tx_sge == NULL) - break; - - LIBCFS_ALLOC(tx->tx_rd, - offsetof(kib_rdma_desc_t, - rd_frags[IBLND_MAX_RDMA_FRAGS])); - if (tx->tx_rd == NULL) - break; + if (ps->ps_net->ibn_fmr_ps != NULL) { + LIBCFS_CPT_ALLOC(tx->tx_pages, + lnet_cpt_table(), ps->ps_cpt, + LNET_MAX_IOV * sizeof(*tx->tx_pages)); + if (tx->tx_pages == NULL) + break; + } + + LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt, + IBLND_MAX_RDMA_FRAGS * sizeof(*tx->tx_frags)); + if (tx->tx_frags == NULL) + break; + + sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS); + + LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt, + (1 + IBLND_MAX_RDMA_FRAGS) * + sizeof(*tx->tx_wrq)); + if (tx->tx_wrq == NULL) + break; + + LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt, + (1 + IBLND_MAX_RDMA_FRAGS) * + sizeof(*tx->tx_sge)); + if (tx->tx_sge == NULL) + break; + + LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt, + offsetof(kib_rdma_desc_t, + rd_frags[IBLND_MAX_RDMA_FRAGS])); + if (tx->tx_rd == NULL) + break; } if (i == size) { @@ -2106,76 +2196,174 @@ kiblnd_tx_init(kib_pool_t *pool, cfs_list_t *node) } void -kiblnd_ni_fini_pools(kib_net_t *net) +kiblnd_net_fini_pools(kib_net_t *net) { - kiblnd_fini_pool_set(&net->ibn_tx_ps.tps_poolset); - if (net->ibn_with_fmr) - kiblnd_fini_fmr_pool_set(&net->ibn_fmr_ps); - else if (net->ibn_with_pmr) - kiblnd_fini_pool_set(&net->ibn_pmr_ps.pps_poolset); -} + int i; -int -kiblnd_net_init_pools(kib_net_t *net) -{ - kib_fmr_poolset_t *fps = &net->ibn_fmr_ps; - kib_pmr_poolset_t *pps = &net->ibn_pmr_ps; - kib_tx_poolset_t *tps = &net->ibn_tx_ps; - unsigned long flags; - int rc; + cfs_cpt_for_each(i, lnet_cpt_table()) { + kib_tx_poolset_t *tps; + kib_fmr_poolset_t *fps; + kib_pmr_poolset_t *pps; - if (*kiblnd_tunables.kib_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, - *kiblnd_tunables.kib_ntx / 4); - return -EINVAL; - } + if (net->ibn_tx_ps != NULL) { + tps = net->ibn_tx_ps[i]; + kiblnd_fini_poolset(&tps->tps_poolset); + } - if (*kiblnd_tunables.kib_pmr_pool_size < - *kiblnd_tunables.kib_ntx / 4) { - CERROR("Can't set pmr pool size (%d) < ntx / 4(%d)\n", - *kiblnd_tunables.kib_pmr_pool_size, - *kiblnd_tunables.kib_ntx / 4); - return -EINVAL; - } + if (net->ibn_fmr_ps != NULL) { + fps = net->ibn_fmr_ps[i]; + kiblnd_fini_fmr_poolset(fps); + } - cfs_read_lock_irqsave(&kiblnd_data.kib_global_lock, flags); - if (*kiblnd_tunables.kib_map_on_demand > 0 || - net->ibn_dev->ibd_hdev->ibh_nmrs > 1) { - /* premapping can fail if ibd_nmr > 1, so we always create - * FMR/PMR pool and map-on-demand if premapping failed */ - cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); - rc = kiblnd_init_fmr_pool_set(fps, net); - if (rc == 0) { - net->ibn_with_fmr = 1; - } else if (rc == -ENOSYS) { - rc = kiblnd_init_pool_set(&pps->pps_poolset, net, "PMR", - *kiblnd_tunables.kib_pmr_pool_size, - kiblnd_create_pmr_pool, - kiblnd_destroy_pmr_pool, - NULL, NULL); - if (rc == 0) - net->ibn_with_pmr = 1; - } - if (rc != 0) - return rc; - } else { - cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); - } + if (net->ibn_pmr_ps != NULL) { + pps = net->ibn_pmr_ps[i]; + kiblnd_fini_poolset(&pps->pps_poolset); + } + } - rc = kiblnd_init_pool_set(&tps->tps_poolset, net, "TX", IBLND_TX_MSGS(), - kiblnd_create_tx_pool, kiblnd_destroy_tx_pool, - kiblnd_tx_init, NULL); - if (rc == 0) - return 0; + if (net->ibn_tx_ps != NULL) { + cfs_percpt_free(net->ibn_tx_ps); + net->ibn_tx_ps = NULL; + } - if (net->ibn_with_fmr) - kiblnd_fini_fmr_pool_set(fps); - else if (net->ibn_with_pmr) - kiblnd_fini_pool_set(&pps->pps_poolset); + if (net->ibn_fmr_ps != NULL) { + cfs_percpt_free(net->ibn_fmr_ps); + net->ibn_fmr_ps = NULL; + } - return rc; + if (net->ibn_pmr_ps != NULL) { + cfs_percpt_free(net->ibn_pmr_ps); + net->ibn_pmr_ps = NULL; + } +} + +int +kiblnd_net_init_pools(kib_net_t *net, __u32 *cpts, int ncpts) +{ + unsigned long flags; + int cpt; + int rc; + int i; + + read_lock_irqsave(&kiblnd_data.kib_global_lock, flags); + if (*kiblnd_tunables.kib_map_on_demand == 0 && + net->ibn_dev->ibd_hdev->ibh_nmrs == 1) { + read_unlock_irqrestore(&kiblnd_data.kib_global_lock, + flags); + goto create_tx_pool; + } + + read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); + + if (*kiblnd_tunables.kib_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, + *kiblnd_tunables.kib_ntx / 4); + rc = -EINVAL; + goto failed; + } + + /* TX pool must be created later than FMR/PMR, see LU-2268 + * for details */ + LASSERT(net->ibn_tx_ps == NULL); + + /* premapping can fail if ibd_nmr > 1, so we always create + * FMR/PMR pool and map-on-demand if premapping failed */ + + net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(), + sizeof(kib_fmr_poolset_t)); + if (net->ibn_fmr_ps == NULL) { + CERROR("Failed to allocate FMR pool array\n"); + rc = -ENOMEM; + goto failed; + } + + 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)); + if (rc == -ENOSYS && i == 0) /* no FMR */ + break; /* create PMR pool */ + + if (rc != 0) { /* a real error */ + CERROR("Can't initialize FMR pool for CPT %d: %d\n", + cpt, rc); + goto failed; + } + } + + if (i > 0) { + LASSERT(i == ncpts); + goto create_tx_pool; + } + + cfs_percpt_free(net->ibn_fmr_ps); + net->ibn_fmr_ps = NULL; + + CWARN("Device does not support FMR, failing back to PMR\n"); + + if (*kiblnd_tunables.kib_pmr_pool_size < + *kiblnd_tunables.kib_ntx / 4) { + CERROR("Can't set pmr pool size (%d) < ntx / 4(%d)\n", + *kiblnd_tunables.kib_pmr_pool_size, + *kiblnd_tunables.kib_ntx / 4); + rc = -EINVAL; + goto failed; + } + + net->ibn_pmr_ps = cfs_percpt_alloc(lnet_cpt_table(), + sizeof(kib_pmr_poolset_t)); + if (net->ibn_pmr_ps == NULL) { + CERROR("Failed to allocate PMR pool array\n"); + rc = -ENOMEM; + goto failed; + } + + for (i = 0; i < ncpts; i++) { + cpt = (cpts == NULL) ? i : cpts[i]; + rc = kiblnd_init_poolset(&net->ibn_pmr_ps[cpt]->pps_poolset, + cpt, net, "PMR", + kiblnd_pmr_pool_size(ncpts), + kiblnd_create_pmr_pool, + kiblnd_destroy_pmr_pool, NULL, NULL); + if (rc != 0) { + CERROR("Can't initialize PMR pool for CPT %d: %d\n", + cpt, rc); + goto failed; + } + } + + create_tx_pool: + net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(), + sizeof(kib_tx_poolset_t)); + if (net->ibn_tx_ps == NULL) { + CERROR("Failed to allocate tx pool array\n"); + rc = -ENOMEM; + goto failed; + } + + for (i = 0; i < ncpts; i++) { + cpt = (cpts == NULL) ? i : cpts[i]; + rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset, + cpt, net, "TX", + kiblnd_tx_pool_size(ncpts), + kiblnd_create_tx_pool, + kiblnd_destroy_tx_pool, + kiblnd_tx_init, NULL); + if (rc != 0) { + CERROR("Can't initialize TX pool for CPT %d: %d\n", + cpt, rc); + goto failed; + } + } + + return 0; + failed: + kiblnd_net_fini_pools(net); + LASSERT(rc != 0); + return rc; } static int @@ -2207,15 +2395,6 @@ kiblnd_hdev_get_attr(kib_hca_dev_t *hdev) return rc; } -#ifdef HAVE_OFED_TRANSPORT_IWARP - /* XXX We can't trust this value returned by Chelsio driver, it's wrong - * and we have reported the bug, remove these in the future when Chelsio - * bug got fixed. */ - if (rdma_node_get_transport(hdev->ibh_ibdev->node_type) == - RDMA_TRANSPORT_IWARP) - hdev->ibh_mr_size = (1ULL << 32) - 1; -#endif - if (hdev->ibh_mr_size == ~0ULL) { hdev->ibh_mr_shift = 64; return 0; @@ -2399,14 +2578,14 @@ kiblnd_dev_need_failover(kib_dev_t *dev) dstaddr.sin_family = AF_INET; rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr, (struct sockaddr *)&dstaddr, 1); - if (rc != 0) { - CERROR("Failed to bind %s to device: %d\n", - dev->ibd_ifname, rc); + if (rc != 0 || cmid->device == NULL) { + CERROR("Failed to bind %s:%u.%u.%u.%u to device(%p): %d\n", + dev->ibd_ifname, HIPQUAD(dev->ibd_ifip), + cmid->device, rc); rdma_destroy_id(cmid); return rc; } - LASSERT (cmid->device != NULL); if (dev->ibd_hdev->ibh_ibdev == cmid->device) { /* don't need device failover */ rdma_destroy_id(cmid); @@ -2430,6 +2609,7 @@ kiblnd_dev_failover(kib_dev_t *dev) struct sockaddr_in addr; unsigned long flags; int rc = 0; + int i; LASSERT (*kiblnd_tunables.kib_dev_failover > 1 || dev->ibd_can_failover || @@ -2445,13 +2625,13 @@ kiblnd_dev_failover(kib_dev_t *dev) * because we can fail to create new listener. * But we have to close it now, otherwise rdma_bind_addr * will return EADDRINUSE... How crap! */ - cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags); + write_lock_irqsave(&kiblnd_data.kib_global_lock, flags); - cmid = dev->ibd_hdev->ibh_cmid; - /* make next schedule of kiblnd_dev_need_failover - * will return 1 for me */ - dev->ibd_hdev->ibh_cmid = NULL; - cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); + cmid = dev->ibd_hdev->ibh_cmid; + /* make next schedule of kiblnd_dev_need_failover() + * return 1 for me */ + dev->ibd_hdev->ibh_cmid = NULL; + write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); rdma_destroy_id(cmid); } @@ -2471,21 +2651,22 @@ kiblnd_dev_failover(kib_dev_t *dev) /* Bind to failover device or port */ rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr); - if (rc != 0) { - CERROR("Failed to bind %s to device: %d\n", - dev->ibd_ifname, rc); + if (rc != 0 || cmid->device == NULL) { + CERROR("Failed to bind %s:%u.%u.%u.%u to device(%p): %d\n", + dev->ibd_ifname, HIPQUAD(dev->ibd_ifip), + cmid->device, rc); rdma_destroy_id(cmid); goto out; } - LIBCFS_ALLOC(hdev, sizeof(*hdev)); + LIBCFS_ALLOC(hdev, sizeof(*hdev)); if (hdev == NULL) { CERROR("Failed to allocate kib_hca_dev\n"); rdma_destroy_id(cmid); rc = -ENOMEM; goto out; } - memset(hdev, 0, sizeof(*hdev)); + atomic_set(&hdev->ibh_ref, 1); hdev->ibh_dev = dev; hdev->ibh_cmid = cmid; @@ -2512,21 +2693,29 @@ kiblnd_dev_failover(kib_dev_t *dev) goto out; } - cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags); + write_lock_irqsave(&kiblnd_data.kib_global_lock, flags); old = dev->ibd_hdev; dev->ibd_hdev = hdev; /* take over the refcount */ hdev = old; cfs_list_for_each_entry(net, &dev->ibd_nets, ibn_list) { - kiblnd_fail_pool_set(&net->ibn_tx_ps.tps_poolset, &zombie_tpo); - if (net->ibn_with_pmr) - kiblnd_fail_pool_set(&net->ibn_pmr_ps.pps_poolset, &zombie_ppo); - if (net->ibn_with_fmr) - kiblnd_fail_fmr_pool_set(&net->ibn_fmr_ps, &zombie_fpo); - } - - cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); + cfs_cpt_for_each(i, lnet_cpt_table()) { + kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset, + &zombie_tpo); + + if (net->ibn_fmr_ps != NULL) { + kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i], + &zombie_fpo); + + } else if (net->ibn_pmr_ps != NULL) { + kiblnd_fail_poolset(&net->ibn_pmr_ps[i]-> + pps_poolset, &zombie_ppo); + } + } + } + + write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); out: if (!cfs_list_empty(&zombie_tpo)) kiblnd_destroy_pool_list(&zombie_tpo); @@ -2619,9 +2808,10 @@ kiblnd_create_dev(char *ifname) } void -kiblnd_base_shutdown (void) +kiblnd_base_shutdown(void) { - int i; + struct kib_sched_info *sched; + int i; LASSERT (cfs_list_empty(&kiblnd_data.kib_devs)); @@ -2643,7 +2833,13 @@ kiblnd_base_shutdown (void) /* flag threads to terminate; wake and wait for them to die */ kiblnd_data.kib_shutdown = 1; - cfs_waitq_broadcast(&kiblnd_data.kib_sched_waitq); + + /* NB: we really want to stop scheduler threads net by net + * instead of the whole module, this should be improved + * with dynamic configuration LNet */ + cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) + cfs_waitq_broadcast(&sched->ibs_waitq); + cfs_waitq_broadcast(&kiblnd_data.kib_connd_waitq); cfs_waitq_broadcast(&kiblnd_data.kib_failover_waitq); @@ -2662,10 +2858,14 @@ kiblnd_base_shutdown (void) break; } - if (kiblnd_data.kib_peers != NULL) - LIBCFS_FREE(kiblnd_data.kib_peers, - sizeof(cfs_list_t) * - kiblnd_data.kib_peer_hash_size); + if (kiblnd_data.kib_peers != NULL) { + LIBCFS_FREE(kiblnd_data.kib_peers, + sizeof(cfs_list_t) * + kiblnd_data.kib_peer_hash_size); + } + + if (kiblnd_data.kib_scheds != NULL) + cfs_percpt_free(kiblnd_data.kib_scheds); CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n", cfs_atomic_read(&libcfs_kmemory)); @@ -2678,7 +2878,7 @@ void kiblnd_shutdown (lnet_ni_t *ni) { kib_net_t *net = ni->ni_data; - cfs_rwlock_t *g_lock = &kiblnd_data.kib_global_lock; + rwlock_t *g_lock = &kiblnd_data.kib_global_lock; int i; unsigned long flags; @@ -2690,9 +2890,9 @@ kiblnd_shutdown (lnet_ni_t *ni) CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n", cfs_atomic_read(&libcfs_kmemory)); - cfs_write_lock_irqsave(g_lock, flags); - net->ibn_shutdown = 1; - cfs_write_unlock_irqrestore(g_lock, flags); + write_lock_irqsave(g_lock, flags); + net->ibn_shutdown = 1; + write_unlock_irqrestore(g_lock, flags); switch (net->ibn_init) { default: @@ -2713,13 +2913,13 @@ kiblnd_shutdown (lnet_ni_t *ni) cfs_pause(cfs_time_seconds(1)); } - kiblnd_ni_fini_pools(net); + kiblnd_net_fini_pools(net); - cfs_write_lock_irqsave(g_lock, flags); - LASSERT (net->ibn_dev->ibd_nnets > 0); - net->ibn_dev->ibd_nnets--; - cfs_list_del(&net->ibn_list); - cfs_write_unlock_irqrestore(g_lock, flags); + write_lock_irqsave(g_lock, flags); + LASSERT(net->ibn_dev->ibd_nnets > 0); + net->ibn_dev->ibd_nnets--; + cfs_list_del(&net->ibn_list); + write_unlock_irqrestore(g_lock, flags); /* fall through */ @@ -2748,17 +2948,18 @@ out: } int -kiblnd_base_startup (void) +kiblnd_base_startup(void) { - int i; - int rc; + struct kib_sched_info *sched; + int rc; + int i; LASSERT (kiblnd_data.kib_init == IBLND_INIT_NOTHING); PORTAL_MODULE_USE; memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */ - cfs_rwlock_init(&kiblnd_data.kib_global_lock); + rwlock_init(&kiblnd_data.kib_global_lock); CFS_INIT_LIST_HEAD(&kiblnd_data.kib_devs); CFS_INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs); @@ -2773,15 +2974,36 @@ kiblnd_base_startup (void) for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) CFS_INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]); - cfs_spin_lock_init(&kiblnd_data.kib_connd_lock); + spin_lock_init(&kiblnd_data.kib_connd_lock); CFS_INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns); CFS_INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies); cfs_waitq_init(&kiblnd_data.kib_connd_waitq); + cfs_waitq_init(&kiblnd_data.kib_failover_waitq); + + kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(), + sizeof(*sched)); + if (kiblnd_data.kib_scheds == NULL) + goto failed; + + cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) { + int nthrs; - cfs_spin_lock_init(&kiblnd_data.kib_sched_lock); - CFS_INIT_LIST_HEAD(&kiblnd_data.kib_sched_conns); - cfs_waitq_init(&kiblnd_data.kib_sched_waitq); - cfs_waitq_init(&kiblnd_data.kib_failover_waitq); + spin_lock_init(&sched->ibs_lock); + CFS_INIT_LIST_HEAD(&sched->ibs_conns); + cfs_waitq_init(&sched->ibs_waitq); + + nthrs = cfs_cpt_weight(lnet_cpt_table(), i); + if (*kiblnd_tunables.kib_nscheds > 0) { + nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds); + } else { + /* max to half of CPUs, another half is reserved for + * upper layer modules */ + nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs); + } + + sched->ibs_nthreads_max = nthrs; + sched->ibs_cpt = i; + } kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR; @@ -2789,23 +3011,15 @@ kiblnd_base_startup (void) kiblnd_data.kib_init = IBLND_INIT_DATA; /*****************************************************/ - for (i = 0; i < IBLND_N_SCHED; i++) { - rc = kiblnd_thread_start(kiblnd_scheduler, (void *)((long)i)); - if (rc != 0) { - CERROR("Can't spawn o2iblnd scheduler[%d]: %d\n", - i, rc); - goto failed; - } - } - - rc = kiblnd_thread_start(kiblnd_connd, NULL); + rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd"); if (rc != 0) { CERROR("Can't spawn o2iblnd connd: %d\n", rc); goto failed; } - if (*kiblnd_tunables.kib_dev_failover != 0) - rc = kiblnd_thread_start(kiblnd_failover_thread, NULL); + if (*kiblnd_tunables.kib_dev_failover != 0) + rc = kiblnd_thread_start(kiblnd_failover_thread, NULL, + "kiblnd_failover"); if (rc != 0) { CERROR("Can't spawn o2iblnd failover thread: %d\n", rc); @@ -2824,15 +3038,115 @@ kiblnd_base_startup (void) } int +kiblnd_start_schedulers(struct kib_sched_info *sched) +{ + int rc = 0; + int nthrs; + int i; + + if (sched->ibs_nthreads == 0) { + if (*kiblnd_tunables.kib_nscheds > 0) { + nthrs = sched->ibs_nthreads_max; + } else { + nthrs = cfs_cpt_weight(lnet_cpt_table(), + sched->ibs_cpt); + nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs); + nthrs = min(IBLND_N_SCHED_HIGH, nthrs); + } + } else { + LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max); + /* increase one thread if there is new interface */ + nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max); + } + + for (i = 0; i < nthrs; i++) { + long id; + char name[20]; + id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i); + snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld", + KIB_THREAD_CPT(id), KIB_THREAD_TID(id)); + rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name); + if (rc == 0) + continue; + + CERROR("Can't spawn thread %d for scheduler[%d]: %d\n", + sched->ibs_cpt, sched->ibs_nthreads + i, rc); + break; + } + + sched->ibs_nthreads += i; + return rc; +} + +int +kiblnd_dev_start_threads(kib_dev_t *dev, int newdev, __u32 *cpts, int ncpts) +{ + int cpt; + int rc; + int i; + + for (i = 0; i < ncpts; i++) { + struct kib_sched_info *sched; + + cpt = (cpts == NULL) ? i : cpts[i]; + sched = kiblnd_data.kib_scheds[cpt]; + + if (!newdev && sched->ibs_nthreads > 0) + continue; + + rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]); + if (rc != 0) { + CERROR("Failed to start scheduler threads for %s\n", + dev->ibd_ifname); + return rc; + } + } + return 0; +} + +kib_dev_t * +kiblnd_dev_search(char *ifname) +{ + kib_dev_t *alias = NULL; + kib_dev_t *dev; + char *colon; + char *colon2; + + colon = strchr(ifname, ':'); + cfs_list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) { + if (strcmp(&dev->ibd_ifname[0], ifname) == 0) + return dev; + + if (alias != NULL) + continue; + + colon2 = strchr(dev->ibd_ifname, ':'); + if (colon != NULL) + *colon = 0; + if (colon2 != NULL) + *colon2 = 0; + + if (strcmp(&dev->ibd_ifname[0], ifname) == 0) + alias = dev; + + if (colon != NULL) + *colon = ':'; + if (colon2 != NULL) + *colon2 = ':'; + } + return alias; +} + +int kiblnd_startup (lnet_ni_t *ni) { char *ifname; kib_dev_t *ibdev = NULL; kib_net_t *net; - cfs_list_t *tmp; struct timeval tv; unsigned long flags; int rc; + int newdev; LASSERT (ni->ni_lnd == &the_o2iblnd); @@ -2876,16 +3190,11 @@ kiblnd_startup (lnet_ni_t *ni) goto failed; } - cfs_list_for_each (tmp, &kiblnd_data.kib_devs) { - ibdev = cfs_list_entry(tmp, kib_dev_t, ibd_list); + ibdev = kiblnd_dev_search(ifname); - if (!strcmp(&ibdev->ibd_ifname[0], ifname)) - break; - - ibdev = NULL; - } - - if (ibdev == NULL) + newdev = ibdev == NULL; + /* hmm...create kib_dev even for alias */ + if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0) ibdev = kiblnd_create_dev(ifname); if (ibdev == NULL) @@ -2894,16 +3203,21 @@ kiblnd_startup (lnet_ni_t *ni) net->ibn_dev = ibdev; ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip); - rc = kiblnd_net_init_pools(net); + rc = kiblnd_dev_start_threads(ibdev, newdev, + ni->ni_cpts, ni->ni_ncpts); + if (rc != 0) + goto failed; + + rc = kiblnd_net_init_pools(net, ni->ni_cpts, ni->ni_ncpts); if (rc != 0) { CERROR("Failed to initialize NI pools: %d\n", rc); goto failed; } - cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags); - ibdev->ibd_nnets++; - cfs_list_add_tail(&net->ibn_list, &ibdev->ibd_nets); - cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); + write_lock_irqsave(&kiblnd_data.kib_global_lock, flags); + ibdev->ibd_nnets++; + cfs_list_add_tail(&net->ibn_list, &ibdev->ibd_nets); + write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); net->ibn_init = IBLND_INIT_ALL;