X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lnet%2Fklnds%2Fo2iblnd%2Fo2iblnd.c;h=87341202ce074e5a7a7c2f1a94cfb058327bf979;hb=1436c2e3309a50800e8578fd9157926279db6234;hp=2251f23ff404dc16804e5fd940778fd8a7b74525;hpb=6e3ec5812ebd1b5ecf7cae584f429b013ffe7431;p=fs%2Flustre-release.git diff --git a/lnet/klnds/o2iblnd/o2iblnd.c b/lnet/klnds/o2iblnd/o2iblnd.c index 2251f23..8734120 100644 --- a/lnet/klnds/o2iblnd/o2iblnd.c +++ b/lnet/klnds/o2iblnd/o2iblnd.c @@ -1,6 +1,4 @@ -/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*- - * vim:expandtab:shiftwidth=8:tabstop=8: - * +/* * GPL HEADER START * * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. @@ -26,8 +24,10 @@ * GPL HEADER END */ /* - * Copyright 2008 Sun Microsystems, Inc. All rights reserved + * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. + * + * Copyright (c) 2011, 2012, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -40,19 +40,11 @@ #include "o2iblnd.h" -lnd_t the_o2iblnd = { - .lnd_type = O2IBLND, - .lnd_startup = kiblnd_startup, - .lnd_shutdown = kiblnd_shutdown, - .lnd_ctl = kiblnd_ctl, - .lnd_query = kiblnd_query, - .lnd_send = kiblnd_send, - .lnd_recv = kiblnd_recv, -}; +static lnd_t the_o2iblnd; kib_data_t kiblnd_data; -__u32 +static __u32 kiblnd_cksum (void *ptr, int nob) { char *c = ptr; @@ -324,183 +316,182 @@ 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; } - memset(peer, 0, sizeof(*peer)); /* zero flags etc */ - - peer->ibp_ni = ni; - peer->ibp_nid = nid; - peer->ibp_error = 0; - peer->ibp_last_alive = 0; - cfs_atomic_set(&peer->ibp_refcount, 1); /* 1 ref for caller */ + peer->ibp_ni = ni; + peer->ibp_nid = nid; + peer->ibp_error = 0; + peer->ibp_last_alive = 0; + atomic_set(&peer->ibp_refcount, 1); /* 1 ref for caller */ - CFS_INIT_LIST_HEAD(&peer->ibp_list); /* not in the peer table yet */ - CFS_INIT_LIST_HEAD(&peer->ibp_conns); - CFS_INIT_LIST_HEAD(&peer->ibp_tx_queue); + INIT_LIST_HEAD(&peer->ibp_list); /* not in the peer table yet */ + INIT_LIST_HEAD(&peer->ibp_conns); + 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); + /* always called with a ref on ni, which prevents ni being shutdown */ + LASSERT(net->ibn_shutdown == 0); - /* npeers only grows with the global lock held */ - cfs_atomic_inc(&net->ibn_npeers); + /* npeers only grows with the global lock held */ + 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; + *peerp = peer; + return 0; } void kiblnd_destroy_peer (kib_peer_t *peer) { - kib_net_t *net = peer->ibp_ni->ni_data; + kib_net_t *net = peer->ibp_ni->ni_data; - LASSERT (net != NULL); - LASSERT (cfs_atomic_read(&peer->ibp_refcount) == 0); - LASSERT (!kiblnd_peer_active(peer)); - LASSERT (peer->ibp_connecting == 0); - LASSERT (peer->ibp_accepting == 0); - LASSERT (cfs_list_empty(&peer->ibp_conns)); - LASSERT (cfs_list_empty(&peer->ibp_tx_queue)); + 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(list_empty(&peer->ibp_tx_queue)); - LIBCFS_FREE(peer, sizeof(*peer)); + LIBCFS_FREE(peer, sizeof(*peer)); - /* NB a peer's connections keep a reference on their peer until - * they are destroyed, so we can be assured that _all_ state to do - * with this peer has been cleaned up when its refcount drops to - * zero. */ - cfs_atomic_dec(&net->ibn_npeers); + /* NB a peer's connections keep a reference on their peer until + * they are destroyed, so we can be assured that _all_ state to do + * with this peer has been cleaned up when its refcount drops to + * zero. */ + atomic_dec(&net->ibn_npeers); } kib_peer_t * kiblnd_find_peer_locked (lnet_nid_t nid) { - /* the caller is responsible for accounting the additional reference - * that this creates */ - cfs_list_t *peer_list = kiblnd_nid2peerlist(nid); - cfs_list_t *tmp; - kib_peer_t *peer; + /* the caller is responsible for accounting the additional reference + * that this creates */ + struct list_head *peer_list = kiblnd_nid2peerlist(nid); + struct list_head *tmp; + kib_peer_t *peer; - cfs_list_for_each (tmp, peer_list) { + list_for_each(tmp, peer_list) { - peer = cfs_list_entry(tmp, kib_peer_t, ibp_list); + peer = list_entry(tmp, kib_peer_t, ibp_list); - LASSERT (peer->ibp_connecting > 0 || /* creating conns */ - peer->ibp_accepting > 0 || - !cfs_list_empty(&peer->ibp_conns)); /* active conn */ + LASSERT(peer->ibp_connecting > 0 || /* creating conns */ + peer->ibp_accepting > 0 || + !list_empty(&peer->ibp_conns)); /* active conn */ - if (peer->ibp_nid != nid) - continue; + if (peer->ibp_nid != nid) + continue; - CDEBUG(D_NET, "got peer [%p] -> %s (%d) version: %x\n", - peer, libcfs_nid2str(nid), - cfs_atomic_read(&peer->ibp_refcount), - peer->ibp_version); - return peer; - } - return NULL; + CDEBUG(D_NET, "got peer [%p] -> %s (%d) version: %x\n", + peer, libcfs_nid2str(nid), + atomic_read(&peer->ibp_refcount), + peer->ibp_version); + return peer; + } + return NULL; } void kiblnd_unlink_peer_locked (kib_peer_t *peer) { - LASSERT (cfs_list_empty(&peer->ibp_conns)); + LASSERT(list_empty(&peer->ibp_conns)); LASSERT (kiblnd_peer_active(peer)); - cfs_list_del_init(&peer->ibp_list); + list_del_init(&peer->ibp_list); /* lose peerlist's ref */ kiblnd_peer_decref(peer); } -int -kiblnd_get_peer_info (lnet_ni_t *ni, int index, - lnet_nid_t *nidp, int *count) +static int +kiblnd_get_peer_info(lnet_ni_t *ni, int index, + lnet_nid_t *nidp, int *count) { - kib_peer_t *peer; - cfs_list_t *ptmp; - int i; - unsigned long flags; + kib_peer_t *peer; + struct list_head *ptmp; + 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]) { + list_for_each(ptmp, &kiblnd_data.kib_peers[i]) { - peer = cfs_list_entry(ptmp, kib_peer_t, ibp_list); - LASSERT (peer->ibp_connecting > 0 || - peer->ibp_accepting > 0 || - !cfs_list_empty(&peer->ibp_conns)); + peer = list_entry(ptmp, kib_peer_t, ibp_list); + LASSERT(peer->ibp_connecting > 0 || + peer->ibp_accepting > 0 || + !list_empty(&peer->ibp_conns)); - if (peer->ibp_ni != ni) - continue; + if (peer->ibp_ni != ni) + continue; - if (index-- > 0) - continue; + if (index-- > 0) + continue; - *nidp = peer->ibp_nid; - *count = cfs_atomic_read(&peer->ibp_refcount); + *nidp = peer->ibp_nid; + *count = 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); - return -ENOENT; + read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); + return -ENOENT; } -void +static void kiblnd_del_peer_locked (kib_peer_t *peer) { - cfs_list_t *ctmp; - cfs_list_t *cnxt; - kib_conn_t *conn; + struct list_head *ctmp; + struct list_head *cnxt; + kib_conn_t *conn; - if (cfs_list_empty(&peer->ibp_conns)) { - kiblnd_unlink_peer_locked(peer); - } else { - cfs_list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) { - conn = cfs_list_entry(ctmp, kib_conn_t, ibc_list); + if (list_empty(&peer->ibp_conns)) { + kiblnd_unlink_peer_locked(peer); + } else { + list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) { + conn = list_entry(ctmp, kib_conn_t, ibc_list); - kiblnd_close_conn_locked(conn, 0); - } - /* NB closing peer's last conn unlinked it. */ - } - /* NB peer now unlinked; might even be freed if the peer table had the - * last ref on it. */ + kiblnd_close_conn_locked(conn, 0); + } + /* NB closing peer's last conn unlinked it. */ + } + /* NB peer now unlinked; might even be freed if the peer table had the + * last ref on it. */ } -int +static int kiblnd_del_peer (lnet_ni_t *ni, lnet_nid_t nid) { - CFS_LIST_HEAD (zombies); - cfs_list_t *ptmp; - cfs_list_t *pnxt; - kib_peer_t *peer; - int lo; - int hi; - int i; - unsigned long flags; - int rc = -ENOENT; + struct list_head zombies = LIST_HEAD_INIT(zombies); + struct list_head *ptmp; + struct list_head *pnxt; + kib_peer_t *peer; + int lo; + int hi; + int i; + 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; @@ -509,80 +500,79 @@ kiblnd_del_peer (lnet_ni_t *ni, lnet_nid_t nid) hi = kiblnd_data.kib_peer_hash_size - 1; } - for (i = lo; i <= hi; i++) { - cfs_list_for_each_safe (ptmp, pnxt, &kiblnd_data.kib_peers[i]) { - peer = cfs_list_entry(ptmp, kib_peer_t, ibp_list); - LASSERT (peer->ibp_connecting > 0 || - peer->ibp_accepting > 0 || - !cfs_list_empty(&peer->ibp_conns)); + 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)); - if (peer->ibp_ni != ni) - continue; + if (peer->ibp_ni != ni) + continue; - if (!(nid == LNET_NID_ANY || peer->ibp_nid == nid)) - continue; + if (!(nid == LNET_NID_ANY || peer->ibp_nid == nid)) + continue; - if (!cfs_list_empty(&peer->ibp_tx_queue)) { - LASSERT (cfs_list_empty(&peer->ibp_conns)); + if (!list_empty(&peer->ibp_tx_queue)) { + LASSERT(list_empty(&peer->ibp_conns)); - cfs_list_splice_init(&peer->ibp_tx_queue, - &zombies); - } + list_splice_init(&peer->ibp_tx_queue, + &zombies); + } - kiblnd_del_peer_locked(peer); - rc = 0; /* matched something */ - } - } + kiblnd_del_peer_locked(peer); + rc = 0; /* matched something */ + } + } - 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); + kiblnd_txlist_done(ni, &zombies, -EIO); - return rc; + return rc; } -kib_conn_t * -kiblnd_get_conn_by_idx (lnet_ni_t *ni, int index) +static kib_conn_t * +kiblnd_get_conn_by_idx(lnet_ni_t *ni, int index) { - kib_peer_t *peer; - cfs_list_t *ptmp; - kib_conn_t *conn; - cfs_list_t *ctmp; - int i; - unsigned long flags; + kib_peer_t *peer; + struct list_head *ptmp; + kib_conn_t *conn; + struct list_head *ctmp; + 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]) { - - peer = cfs_list_entry(ptmp, kib_peer_t, ibp_list); - LASSERT (peer->ibp_connecting > 0 || - peer->ibp_accepting > 0 || - !cfs_list_empty(&peer->ibp_conns)); - - if (peer->ibp_ni != ni) - continue; - - cfs_list_for_each (ctmp, &peer->ibp_conns) { - if (index-- > 0) - continue; - - 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; - } - } - } + for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) { + 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)); - cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); + if (peer->ibp_ni != ni) + continue; + + list_for_each(ctmp, &peer->ibp_conns) { + if (index-- > 0) + continue; + + conn = list_entry(ctmp, kib_conn_t, ibc_list); + kiblnd_conn_addref(conn); + read_unlock_irqrestore(&kiblnd_data.kib_global_lock, + flags); + return conn; + } + } + } + + read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); return NULL; } -void +static void kiblnd_debug_rx (kib_rx_t *rx) { CDEBUG(D_CONSOLE, " %p status %d msg_type %x cred %d\n", @@ -590,7 +580,7 @@ kiblnd_debug_rx (kib_rx_t *rx) rx->rx_msg->ibm_credits); } -void +static void kiblnd_debug_tx (kib_tx_t *tx) { CDEBUG(D_CONSOLE, " %p snd %d q %d w %d rc %d dl %lx " @@ -605,45 +595,49 @@ kiblnd_debug_tx (kib_tx_t *tx) void kiblnd_debug_conn (kib_conn_t *conn) { - cfs_list_t *tmp; - int i; + struct list_head *tmp; + int i; + + spin_lock(&conn->ibc_lock); - cfs_spin_lock(&conn->ibc_lock); + CDEBUG(D_CONSOLE, "conn[%d] %p [version %x] -> %s:\n", + atomic_read(&conn->ibc_refcount), conn, + conn->ibc_version, libcfs_nid2str(conn->ibc_peer->ibp_nid)); + CDEBUG(D_CONSOLE, " state %d nposted %d/%d cred %d o_cred %d " + " r_cred %d\n", conn->ibc_state, conn->ibc_noops_posted, + conn->ibc_nsends_posted, conn->ibc_credits, + conn->ibc_outstanding_credits, conn->ibc_reserved_credits); + CDEBUG(D_CONSOLE, " comms_err %d\n", conn->ibc_comms_error); - CDEBUG(D_CONSOLE, "conn[%d] %p [version %x] -> %s: \n", - cfs_atomic_read(&conn->ibc_refcount), conn, - conn->ibc_version, libcfs_nid2str(conn->ibc_peer->ibp_nid)); - CDEBUG(D_CONSOLE, " state %d nposted %d/%d cred %d o_cred %d r_cred %d\n", - conn->ibc_state, conn->ibc_noops_posted, - conn->ibc_nsends_posted, conn->ibc_credits, - conn->ibc_outstanding_credits, conn->ibc_reserved_credits); - CDEBUG(D_CONSOLE, " comms_err %d\n", conn->ibc_comms_error); + CDEBUG(D_CONSOLE, " early_rxs:\n"); + list_for_each(tmp, &conn->ibc_early_rxs) + kiblnd_debug_rx(list_entry(tmp, kib_rx_t, rx_list)); - CDEBUG(D_CONSOLE, " early_rxs:\n"); - cfs_list_for_each(tmp, &conn->ibc_early_rxs) - kiblnd_debug_rx(cfs_list_entry(tmp, kib_rx_t, rx_list)); + CDEBUG(D_CONSOLE, " tx_noops:\n"); + list_for_each(tmp, &conn->ibc_tx_noops) + kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list)); - CDEBUG(D_CONSOLE, " tx_queue_nocred:\n"); - cfs_list_for_each(tmp, &conn->ibc_tx_queue_nocred) - kiblnd_debug_tx(cfs_list_entry(tmp, kib_tx_t, tx_list)); + CDEBUG(D_CONSOLE, " tx_queue_nocred:\n"); + list_for_each(tmp, &conn->ibc_tx_queue_nocred) + kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list)); - CDEBUG(D_CONSOLE, " tx_queue_rsrvd:\n"); - cfs_list_for_each(tmp, &conn->ibc_tx_queue_rsrvd) - kiblnd_debug_tx(cfs_list_entry(tmp, kib_tx_t, tx_list)); + CDEBUG(D_CONSOLE, " tx_queue_rsrvd:\n"); + list_for_each(tmp, &conn->ibc_tx_queue_rsrvd) + kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list)); - CDEBUG(D_CONSOLE, " tx_queue:\n"); - cfs_list_for_each(tmp, &conn->ibc_tx_queue) - kiblnd_debug_tx(cfs_list_entry(tmp, kib_tx_t, tx_list)); + CDEBUG(D_CONSOLE, " tx_queue:\n"); + list_for_each(tmp, &conn->ibc_tx_queue) + kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list)); - CDEBUG(D_CONSOLE, " active_txs:\n"); - cfs_list_for_each(tmp, &conn->ibc_active_txs) - kiblnd_debug_tx(cfs_list_entry(tmp, kib_tx_t, tx_list)); + CDEBUG(D_CONSOLE, " active_txs:\n"); + list_for_each(tmp, &conn->ibc_active_txs) + kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list)); - CDEBUG(D_CONSOLE, " rxs:\n"); - for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++) - kiblnd_debug_rx(&conn->ibc_rxs[i]); + CDEBUG(D_CONSOLE, " rxs:\n"); + 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 @@ -668,101 +662,163 @@ kiblnd_translate_mtu(int value) } static void -kiblnd_setup_mtu(struct rdma_cm_id *cmid) +kiblnd_setup_mtu_locked(struct rdma_cm_id *cmid) { - unsigned long flags; int mtu; /* XXX There is no path record for iWARP, set by netdev->change_mtu? */ if (cmid->route.path_rec == NULL) return; - cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags); - mtu = kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu); LASSERT (mtu >= 0); if (mtu != 0) cmid->route.path_rec->mtu = mtu; - - cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); } -kib_conn_t * -kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid, - int state, int version) +#ifdef HAVE_OFED_IB_COMP_VECTOR +static int +kiblnd_get_completion_vector(kib_conn_t *conn, int cpt) { - /* CAVEAT EMPTOR: - * If the new conn is created successfully it takes over the caller's - * ref on 'peer'. It also "owns" 'cmid' and destroys it when it itself - * is destroyed. On failure, the caller's ref on 'peer' remains and - * 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'). */ - kib_net_t *net = peer->ibp_ni->ni_data; - struct ib_qp_init_attr *init_qp_attr; - kib_conn_t *conn; - struct ib_cq *cq; - unsigned long flags; - int rc; - int i; + cpumask_t *mask; + int vectors; + int off; + int i; - 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; - } - - LIBCFS_ALLOC(conn, 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 */ - cmid->context = conn; /* for future CM callbacks */ - conn->ibc_cmid = cmid; + vectors = conn->ibc_cmid->device->num_comp_vectors; + if (vectors <= 1) + return 0; - CFS_INIT_LIST_HEAD(&conn->ibc_early_rxs); - CFS_INIT_LIST_HEAD(&conn->ibc_tx_queue); - 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); + mask = cfs_cpt_cpumask(lnet_cpt_table(), cpt); - 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)); + /* 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; + } - 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)); - - rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, - IBLND_RX_MSG_PAGES(version)); - if (rc != 0) - goto failed_2; + LBUG(); + return 1; +} +#endif - kiblnd_map_rx_descs(conn); +kib_conn_t * +kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid, + int state, int version) +{ + /* CAVEAT EMPTOR: + * If the new conn is created successfully it takes over the caller's + * ref on 'peer'. It also "owns" 'cmid' and destroys it when it itself + * is destroyed. On failure, the caller's ref on 'peer' remains and + * 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'). */ + rwlock_t *glock = &kiblnd_data.kib_global_lock; + kib_net_t *net = peer->ibp_ni->ni_data; + kib_dev_t *dev; + struct ib_qp_init_attr *init_qp_attr; + 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(!in_interrupt()); + + dev = net->ibn_dev; + + cpt = lnet_cpt_of_nid(peer->ibp_nid); + sched = kiblnd_data.kib_scheds[cpt]; + + 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; + } + + conn->ibc_state = IBLND_CONN_INIT; + conn->ibc_version = version; + conn->ibc_peer = peer; /* I take the caller's ref */ + cmid->context = conn; /* for future CM callbacks */ + conn->ibc_cmid = cmid; + + INIT_LIST_HEAD(&conn->ibc_early_rxs); + INIT_LIST_HEAD(&conn->ibc_tx_noops); + INIT_LIST_HEAD(&conn->ibc_tx_queue); + INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd); + INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred); + INIT_LIST_HEAD(&conn->ibc_active_txs); + 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 */ + if (kiblnd_dev_can_failover(dev)) { + list_add_tail(&dev->ibd_fail_list, + &kiblnd_data.kib_failed_devs); + wake_up(&kiblnd_data.kib_failover_waitq); + } + + write_unlock_irqrestore(glock, flags); + CERROR("cmid HCA(%s), kib_dev(%s) need failover\n", + cmid->device->name, dev->ibd_ifname); + goto failed_2; + } + + kiblnd_hdev_addref_locked(dev->ibd_hdev); + conn->ibc_hdev = dev->ibd_hdev; + + kiblnd_setup_mtu_locked(cmid); + + write_unlock_irqrestore(glock, flags); + + 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, cpt, + IBLND_RX_MSG_PAGES(version)); + if (rc != 0) + goto failed_2; + + 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, @@ -782,9 +838,6 @@ kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid, goto failed_2; } - kiblnd_setup_mtu(cmid); - - 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); @@ -796,7 +849,9 @@ kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid, init_qp_attr->send_cq = cq; init_qp_attr->recv_cq = cq; - rc = rdma_create_qp(cmid, net->ibn_dev->ibd_pd, init_qp_attr); + 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, @@ -807,7 +862,7 @@ kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid, LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr)); /* 1 ref for caller and each rxmsg */ - cfs_atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(version)); + atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(version)); conn->ibc_nrx = IBLND_RX_MSGS(version); /* post receives */ @@ -822,11 +877,9 @@ 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 @@ -848,7 +901,7 @@ kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid, conn->ibc_state = state; /* 1 more conn */ - cfs_atomic_inc(&net->ibn_nconns); + atomic_inc(&net->ibn_nconns); return conn; failed_2: @@ -862,168 +915,172 @@ kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid, void kiblnd_destroy_conn (kib_conn_t *conn) { - struct rdma_cm_id *cmid = conn->ibc_cmid; - kib_peer_t *peer = conn->ibc_peer; - int rc; + struct rdma_cm_id *cmid = conn->ibc_cmid; + kib_peer_t *peer = conn->ibc_peer; + int rc; - LASSERT (!cfs_in_interrupt()); - LASSERT (cfs_atomic_read(&conn->ibc_refcount) == 0); - LASSERT (cfs_list_empty(&conn->ibc_early_rxs)); - LASSERT (cfs_list_empty(&conn->ibc_tx_queue)); - LASSERT (cfs_list_empty(&conn->ibc_tx_queue_rsrvd)); - LASSERT (cfs_list_empty(&conn->ibc_tx_queue_nocred)); - LASSERT (cfs_list_empty(&conn->ibc_active_txs)); - LASSERT (conn->ibc_noops_posted == 0); - LASSERT (conn->ibc_nsends_posted == 0); - - switch (conn->ibc_state) { - default: - /* conn must be completely disengaged from the network */ - LBUG(); + LASSERT (!in_interrupt()); + LASSERT (atomic_read(&conn->ibc_refcount) == 0); + LASSERT(list_empty(&conn->ibc_early_rxs)); + LASSERT(list_empty(&conn->ibc_tx_noops)); + LASSERT(list_empty(&conn->ibc_tx_queue)); + LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd)); + LASSERT(list_empty(&conn->ibc_tx_queue_nocred)); + LASSERT(list_empty(&conn->ibc_active_txs)); + LASSERT (conn->ibc_noops_posted == 0); + LASSERT (conn->ibc_nsends_posted == 0); - case IBLND_CONN_DISCONNECTED: - /* connvars should have been freed already */ - LASSERT (conn->ibc_connvars == NULL); - break; + switch (conn->ibc_state) { + default: + /* conn must be completely disengaged from the network */ + LBUG(); - case IBLND_CONN_INIT: - break; - } + case IBLND_CONN_DISCONNECTED: + /* connvars should have been freed already */ + LASSERT (conn->ibc_connvars == NULL); + break; - /* conn->ibc_cmid might be destroyed by CM already */ - if (cmid != NULL && cmid->qp != NULL) - rdma_destroy_qp(cmid); + case IBLND_CONN_INIT: + break; + } - if (conn->ibc_cq != NULL) { - rc = ib_destroy_cq(conn->ibc_cq); - if (rc != 0) - CWARN("Error destroying CQ: %d\n", rc); - } + /* conn->ibc_cmid might be destroyed by CM already */ + if (cmid != NULL && cmid->qp != NULL) + rdma_destroy_qp(cmid); - if (conn->ibc_rx_pages != NULL) - kiblnd_unmap_rx_descs(conn); + if (conn->ibc_cq != NULL) { + rc = ib_destroy_cq(conn->ibc_cq); + if (rc != 0) + CWARN("Error destroying CQ: %d\n", rc); + } - if (conn->ibc_rxs != NULL) { - LIBCFS_FREE(conn->ibc_rxs, - IBLND_RX_MSGS(conn->ibc_version) * sizeof(kib_rx_t)); - } + if (conn->ibc_rx_pages != NULL) + kiblnd_unmap_rx_descs(conn); - if (conn->ibc_connvars != NULL) - LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars)); + if (conn->ibc_rxs != NULL) { + LIBCFS_FREE(conn->ibc_rxs, + IBLND_RX_MSGS(conn->ibc_version) * sizeof(kib_rx_t)); + } - /* See CAVEAT EMPTOR above in kiblnd_create_conn */ - if (conn->ibc_state != IBLND_CONN_INIT) { - kib_net_t *net = peer->ibp_ni->ni_data; + if (conn->ibc_connvars != NULL) + LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars)); - kiblnd_peer_decref(peer); - rdma_destroy_id(cmid); - cfs_atomic_dec(&net->ibn_nconns); - } + if (conn->ibc_hdev != NULL) + kiblnd_hdev_decref(conn->ibc_hdev); + + /* See CAVEAT EMPTOR above in kiblnd_create_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); + } - LIBCFS_FREE(conn, sizeof(*conn)); + LIBCFS_FREE(conn, sizeof(*conn)); } int -kiblnd_close_peer_conns_locked (kib_peer_t *peer, int why) +kiblnd_close_peer_conns_locked(kib_peer_t *peer, int why) { - kib_conn_t *conn; - cfs_list_t *ctmp; - cfs_list_t *cnxt; - int count = 0; + kib_conn_t *conn; + struct list_head *ctmp; + struct list_head *cnxt; + int count = 0; - cfs_list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) { - conn = cfs_list_entry(ctmp, kib_conn_t, ibc_list); + list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) { + conn = list_entry(ctmp, kib_conn_t, ibc_list); - CDEBUG(D_NET, "Closing conn -> %s, " - "version: %x, reason: %d\n", - libcfs_nid2str(peer->ibp_nid), - conn->ibc_version, why); + CDEBUG(D_NET, "Closing conn -> %s, " + "version: %x, reason: %d\n", + libcfs_nid2str(peer->ibp_nid), + conn->ibc_version, why); - kiblnd_close_conn_locked(conn, why); - count++; - } + kiblnd_close_conn_locked(conn, why); + count++; + } - return count; + return count; } int -kiblnd_close_stale_conns_locked (kib_peer_t *peer, - int version, __u64 incarnation) +kiblnd_close_stale_conns_locked(kib_peer_t *peer, + int version, __u64 incarnation) { - kib_conn_t *conn; - cfs_list_t *ctmp; - cfs_list_t *cnxt; - int count = 0; + kib_conn_t *conn; + struct list_head *ctmp; + struct list_head *cnxt; + int count = 0; - cfs_list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) { - conn = cfs_list_entry(ctmp, kib_conn_t, ibc_list); + list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) { + conn = list_entry(ctmp, kib_conn_t, ibc_list); - if (conn->ibc_version == version && - conn->ibc_incarnation == incarnation) - continue; + if (conn->ibc_version == version && + conn->ibc_incarnation == incarnation) + continue; - CDEBUG(D_NET, "Closing stale conn -> %s version: %x, " - "incarnation:"LPX64"(%x, "LPX64")\n", - libcfs_nid2str(peer->ibp_nid), - conn->ibc_version, conn->ibc_incarnation, - version, incarnation); + CDEBUG(D_NET, "Closing stale conn -> %s version: %x, " + "incarnation:"LPX64"(%x, "LPX64")\n", + libcfs_nid2str(peer->ibp_nid), + conn->ibc_version, conn->ibc_incarnation, + version, incarnation); - kiblnd_close_conn_locked(conn, -ESTALE); - count++; - } + kiblnd_close_conn_locked(conn, -ESTALE); + count++; + } - return count; + return count; } -int -kiblnd_close_matching_conns (lnet_ni_t *ni, lnet_nid_t nid) +static int +kiblnd_close_matching_conns(lnet_ni_t *ni, lnet_nid_t nid) { - kib_peer_t *peer; - cfs_list_t *ptmp; - cfs_list_t *pnxt; - int lo; - int hi; - int i; - unsigned long flags; - int count = 0; + kib_peer_t *peer; + struct list_head *ptmp; + struct list_head *pnxt; + int lo; + int hi; + int i; + 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; - else { - lo = 0; - hi = kiblnd_data.kib_peer_hash_size - 1; - } + if (nid != LNET_NID_ANY) + lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers; + else { + lo = 0; + hi = kiblnd_data.kib_peer_hash_size - 1; + } - for (i = lo; i <= hi; i++) { - cfs_list_for_each_safe (ptmp, pnxt, &kiblnd_data.kib_peers[i]) { + for (i = lo; i <= hi; i++) { + list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) { - peer = cfs_list_entry(ptmp, kib_peer_t, ibp_list); - LASSERT (peer->ibp_connecting > 0 || - peer->ibp_accepting > 0 || - !cfs_list_empty(&peer->ibp_conns)); + peer = list_entry(ptmp, kib_peer_t, ibp_list); + LASSERT(peer->ibp_connecting > 0 || + peer->ibp_accepting > 0 || + !list_empty(&peer->ibp_conns)); - if (peer->ibp_ni != ni) - continue; + if (peer->ibp_ni != ni) + continue; - if (!(nid == LNET_NID_ANY || nid == peer->ibp_nid)) - continue; + if (!(nid == LNET_NID_ANY || nid == peer->ibp_nid)) + continue; - count += kiblnd_close_peer_conns_locked(peer, 0); - } - } + count += kiblnd_close_peer_conns_locked(peer, 0); + } + } - 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) - return 0; + /* wildcards always succeed */ + if (nid == LNET_NID_ANY) + return 0; - return (count == 0) ? -ENOENT : 0; + return (count == 0) ? -ENOENT : 0; } -int +static int kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg) { struct libcfs_ioctl_data *data = arg; @@ -1077,58 +1134,63 @@ kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg) return rc; } -void -kiblnd_query (lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when) +static void +kiblnd_query(lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when) { - cfs_time_t last_alive = 0; - 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) { - LASSERT (peer->ibp_connecting > 0 || /* creating conns */ - peer->ibp_accepting > 0 || - !cfs_list_empty(&peer->ibp_conns)); /* active conn */ - last_alive = peer->ibp_last_alive; - } + 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 */ + last_alive = peer->ibp_last_alive; + } - cfs_read_unlock_irqrestore(glock, flags); + read_unlock_irqrestore(glock, flags); - if (last_alive != 0) - *when = last_alive; + if (last_alive != 0) + *when = last_alive; - /* peer is not persistent in hash, trigger peer creation - * and connection establishment with a NULL tx */ - if (peer == NULL) - kiblnd_launch_tx(ni, NULL, nid); - return; + /* peer is not persistent in hash, trigger peer creation + * and connection establishment with a NULL tx */ + if (peer == NULL) + kiblnd_launch_tx(ni, NULL, nid); + + CDEBUG(D_NET, "Peer %s %p, alive %ld secs ago\n", + libcfs_nid2str(nid), peer, + last_alive ? cfs_duration_sec(now - last_alive) : -1); + return; } -void -kiblnd_free_pages (kib_pages_t *p) +static void +kiblnd_free_pages(kib_pages_t *p) { - int npages = p->ibp_npages; - int i; - - LASSERT (p->ibp_device == NULL); + 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) + __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; @@ -1138,7 +1200,8 @@ 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, + GFP_NOFS); if (p->ibp_pages[i] == NULL) { CERROR("Can't allocate page %d of %d\n", i, npages); kiblnd_free_pages(p); @@ -1157,21 +1220,19 @@ kiblnd_unmap_rx_descs(kib_conn_t *conn) int i; LASSERT (conn->ibc_rxs != NULL); - LASSERT (conn->ibc_rx_pages->ibp_device != NULL); + LASSERT (conn->ibc_hdev != NULL); for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++) { rx = &conn->ibc_rxs[i]; LASSERT (rx->rx_nob >= 0); /* not posted */ - kiblnd_dma_unmap_single(conn->ibc_rx_pages->ibp_device, + kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev, KIBLND_UNMAP_ADDR(rx, rx_msgunmap, rx->rx_msgaddr), IBLND_MSG_SIZE, DMA_FROM_DEVICE); } - conn->ibc_rx_pages->ibp_device = NULL; - kiblnd_free_pages(conn->ibc_rx_pages); conn->ibc_rx_pages = NULL; @@ -1194,10 +1255,10 @@ kiblnd_map_rx_descs(kib_conn_t *conn) rx->rx_conn = conn; rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off); - rx->rx_msgaddr = kiblnd_dma_map_single(conn->ibc_cmid->device, + rx->rx_msgaddr = kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev, rx->rx_msg, IBLND_MSG_SIZE, DMA_FROM_DEVICE); - LASSERT (!kiblnd_dma_mapping_error(conn->ibc_cmid->device, + LASSERT (!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev, rx->rx_msgaddr)); KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr); @@ -1214,27 +1275,56 @@ kiblnd_map_rx_descs(kib_conn_t *conn) LASSERT (ipg <= IBLND_RX_MSG_PAGES(conn->ibc_version)); } } - - conn->ibc_rx_pages->ibp_device = conn->ibc_cmid->device; } static void kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo) { + kib_hca_dev_t *hdev = tpo->tpo_hdev; kib_tx_t *tx; int i; LASSERT (tpo->tpo_pool.po_allocated == 0); - LASSERT (tpo->tpo_tx_pages->ibp_device != NULL); + + if (hdev == NULL) + return; for (i = 0; i < tpo->tpo_pool.po_size; i++) { tx = &tpo->tpo_tx_descs[i]; - kiblnd_dma_unmap_single(tpo->tpo_tx_pages->ibp_device, + kiblnd_dma_unmap_single(hdev->ibh_ibdev, KIBLND_UNMAP_ADDR(tx, tx_msgunmap, tx->tx_msgaddr), IBLND_MSG_SIZE, DMA_TO_DEVICE); } - tpo->tpo_tx_pages->ibp_device = NULL; + + kiblnd_hdev_decref(hdev); + tpo->tpo_hdev = NULL; +} + +static kib_hca_dev_t * +kiblnd_current_hdev(kib_dev_t *dev) +{ + kib_hca_dev_t *hdev; + unsigned long flags; + int i = 0; + + 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); + schedule_timeout(cfs_time_seconds(1) / 100); + + read_lock_irqsave(&kiblnd_data.kib_global_lock, flags); + } + + kiblnd_hdev_addref_locked(dev->ibd_hdev); + hdev = dev->ibd_hdev; + + read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); + + return hdev; } static void @@ -1243,6 +1333,7 @@ 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; struct page *page; kib_tx_t *tx; int page_offset; @@ -1250,8 +1341,8 @@ kiblnd_map_tx_pool(kib_tx_pool_t *tpo) int i; LASSERT (net != NULL); - LASSERT (net->ibn_dev->ibd_cmid != NULL && - net->ibn_dev->ibd_cmid->device != NULL); + + dev = net->ibn_dev; /* pre-mapped messages are not bigger than 1 page */ CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE); @@ -1259,73 +1350,74 @@ kiblnd_map_tx_pool(kib_tx_pool_t *tpo) /* No fancy arithmetic when we do the buffer calculations */ CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0); - txpgs->ibp_device = net->ibn_dev->ibd_cmid->device; + tpo->tpo_hdev = kiblnd_current_hdev(dev); - for (ipage = page_offset = i = 0; i < pool->po_size; i++) { - page = txpgs->ibp_pages[ipage]; - tx = &tpo->tpo_tx_descs[i]; + for (ipage = page_offset = i = 0; i < pool->po_size; i++) { + page = txpgs->ibp_pages[ipage]; + tx = &tpo->tpo_tx_descs[i]; - tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) + - page_offset); + tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) + + page_offset); - tx->tx_msgaddr = kiblnd_dma_map_single( - txpgs->ibp_device, tx->tx_msg, - IBLND_MSG_SIZE, DMA_TO_DEVICE); - LASSERT (!kiblnd_dma_mapping_error(txpgs->ibp_device, - tx->tx_msgaddr)); - KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr); + tx->tx_msgaddr = kiblnd_dma_map_single(tpo->tpo_hdev->ibh_ibdev, + tx->tx_msg, + IBLND_MSG_SIZE, + DMA_TO_DEVICE); + LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev, + tx->tx_msgaddr)); + KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr); - cfs_list_add(&tx->tx_list, &pool->po_free_list); + list_add(&tx->tx_list, &pool->po_free_list); - page_offset += IBLND_MSG_SIZE; - LASSERT (page_offset <= PAGE_SIZE); + page_offset += IBLND_MSG_SIZE; + LASSERT(page_offset <= PAGE_SIZE); - if (page_offset == PAGE_SIZE) { - page_offset = 0; - ipage++; - LASSERT (ipage <= txpgs->ibp_npages); - } - } + if (page_offset == PAGE_SIZE) { + page_offset = 0; + ipage++; + LASSERT(ipage <= txpgs->ibp_npages); + } + } } struct ib_mr * -kiblnd_find_dma_mr(kib_net_t *net, __u64 addr, __u64 size) +kiblnd_find_dma_mr(kib_hca_dev_t *hdev, __u64 addr, __u64 size) { __u64 index; - LASSERT (net->ibn_dev->ibd_mrs[0] != NULL); + LASSERT (hdev->ibh_mrs[0] != NULL); - if (net->ibn_dev->ibd_nmrs == 1) - return net->ibn_dev->ibd_mrs[0]; + if (hdev->ibh_nmrs == 1) + return hdev->ibh_mrs[0]; - index = addr >> net->ibn_dev->ibd_mr_shift; + index = addr >> hdev->ibh_mr_shift; - if (index < net->ibn_dev->ibd_nmrs && - index == ((addr + size - 1) >> net->ibn_dev->ibd_mr_shift)) - return net->ibn_dev->ibd_mrs[index]; + if (index < hdev->ibh_nmrs && + index == ((addr + size - 1) >> hdev->ibh_mr_shift)) + return hdev->ibh_mrs[index]; return NULL; } struct ib_mr * -kiblnd_find_rd_dma_mr(kib_net_t *net, kib_rdma_desc_t *rd) +kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev, kib_rdma_desc_t *rd) { struct ib_mr *prev_mr; struct ib_mr *mr; int i; - LASSERT (net->ibn_dev->ibd_mrs[0] != NULL); + LASSERT (hdev->ibh_mrs[0] != NULL); if (*kiblnd_tunables.kib_map_on_demand > 0 && *kiblnd_tunables.kib_map_on_demand <= rd->rd_nfrags) return NULL; - if (net->ibn_dev->ibd_nmrs == 1) - return net->ibn_dev->ibd_mrs[0]; + if (hdev->ibh_nmrs == 1) + return hdev->ibh_mrs[0]; for (i = 0, mr = prev_mr = NULL; i < rd->rd_nfrags; i++) { - mr = kiblnd_find_dma_mr(net, + mr = kiblnd_find_dma_mr(hdev, rd->rd_frags[i].rf_addr, rd->rd_frags[i].rf_nob); if (prev_mr == NULL) @@ -1341,7 +1433,7 @@ kiblnd_find_rd_dma_mr(kib_net_t *net, kib_rdma_desc_t *rd) return mr; } -void +static void kiblnd_destroy_fmr_pool(kib_fmr_pool_t *pool) { LASSERT (pool->fpo_map_count == 0); @@ -1349,51 +1441,68 @@ kiblnd_destroy_fmr_pool(kib_fmr_pool_t *pool) if (pool->fpo_fmr_pool != NULL) ib_destroy_fmr_pool(pool->fpo_fmr_pool); + if (pool->fpo_hdev != NULL) + kiblnd_hdev_decref(pool->fpo_hdev); + LIBCFS_FREE(pool, sizeof(kib_fmr_pool_t)); } -void -kiblnd_destroy_fmr_pool_list(cfs_list_t *head) +static void +kiblnd_destroy_fmr_pool_list(struct list_head *head) { - kib_fmr_pool_t *pool; + kib_fmr_pool_t *pool; - while (!cfs_list_empty(head)) { - pool = cfs_list_entry(head->next, kib_fmr_pool_t, fpo_list); - cfs_list_del(&pool->fpo_list); - kiblnd_destroy_fmr_pool(pool); - } + 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); + } } -int +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); +} + +static int kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_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), - .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; - - LASSERT (fps->fps_net->ibn_dev != NULL && - fps->fps_net->ibn_dev->ibd_pd != NULL); - - 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_fmr_pool = ib_create_fmr_pool(fps->fps_net->ibn_dev->ibd_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; } @@ -1406,67 +1515,108 @@ kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo) } static void -kiblnd_fini_fmr_pool_set(kib_fmr_poolset_t *fps) +kiblnd_fail_fmr_poolset(kib_fmr_poolset_t *fps, struct list_head *zombies) +{ + if (fps->fps_net == NULL) /* intialized? */ + return; + + spin_lock(&fps->fps_lock); + + while (!list_empty(&fps->fps_pool_list)) { + kib_fmr_pool_t *fpo = list_entry(fps->fps_pool_list.next, + kib_fmr_pool_t, fpo_list); + fpo->fpo_failed = 1; + list_del(&fpo->fpo_list); + if (fpo->fpo_map_count == 0) + list_add(&fpo->fpo_list, zombies); + else + list_add(&fpo->fpo_list, &fps->fps_failed_pool_list); + } + + spin_unlock(&fps->fps_lock); +} + +static void +kiblnd_fini_fmr_poolset(kib_fmr_poolset_t *fps) { - 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; + kib_fmr_pool_t *fpo; + int rc; - memset(fps, 0, sizeof(kib_fmr_poolset_t)); + 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); - rc = kiblnd_create_fmr_pool(fps, &fpo); - if (rc == 0) - cfs_list_add_tail(&fpo->fpo_list, &fps->fps_pool_list); + fps->fps_net = net; + fps->fps_cpt = cpt; + fps->fps_pool_size = pool_size; + fps->fps_flush_trigger = flush_trigger; + spin_lock_init(&fps->fps_lock); + INIT_LIST_HEAD(&fps->fps_pool_list); + INIT_LIST_HEAD(&fps->fps_failed_pool_list); - return rc; + rc = kiblnd_create_fmr_pool(fps, &fpo); + if (rc == 0) + list_add_tail(&fpo->fpo_list, &fps->fps_pool_list); + + return rc; +} + +static int +kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, cfs_time_t now) +{ + if (fpo->fpo_map_count != 0) /* still in use */ + return 0; + if (fpo->fpo_failed) + return 1; + return cfs_time_aftereq(now, fpo->fpo_deadline); } void kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status) { - CFS_LIST_HEAD (zombies); - kib_fmr_pool_t *fpo = fmr->fmr_pool; - kib_fmr_poolset_t *fps = fpo->fpo_owner; - kib_fmr_pool_t *tmp; - int rc; + struct list_head zombies = LIST_HEAD_INIT(zombies); + kib_fmr_pool_t *fpo = fmr->fmr_pool; + kib_fmr_poolset_t *fps = fpo->fpo_owner; + 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); + rc = ib_fmr_pool_unmap(fmr->fmr_pfmr); + LASSERT(rc == 0); - if (status != 0) { - rc = ib_flush_fmr_pool(fpo->fpo_fmr_pool); - LASSERT (rc == 0); - } + if (status != 0) { + rc = ib_flush_fmr_pool(fpo->fpo_fmr_pool); + LASSERT(rc == 0); + } - fmr->fmr_pool = NULL; - fmr->fmr_pfmr = NULL; + fmr->fmr_pool = NULL; + fmr->fmr_pfmr = NULL; - cfs_spin_lock(&fps->fps_lock); - fpo->fpo_map_count --; /* decref the pool */ + 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) { - /* the first pool is persistent */ - if (fps->fps_pool_list.next == &fpo->fpo_list) - continue; + list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) { + /* the first pool is persistent */ + if (fps->fps_pool_list.next == &fpo->fpo_list) + continue; - if (fpo->fpo_map_count == 0 && /* no more reference */ - cfs_time_aftereq(cfs_time_current(), fpo->fpo_deadline)) { - cfs_list_move(&fpo->fpo_list, &zombies); - fps->fps_version ++; - } - } - cfs_spin_unlock(&fps->fps_lock); + if (kiblnd_fmr_pool_is_idle(fpo, now)) { + list_move(&fpo->fpo_list, &zombies); + fps->fps_version++; + } + } + spin_unlock(&fps->fps_lock); - if (!cfs_list_empty(&zombies)) - kiblnd_destroy_fmr_pool_list(&zombies); + if (!list_empty(&zombies)) + kiblnd_destroy_fmr_pool_list(&zombies); } int @@ -1478,14 +1628,13 @@ 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); +again: + spin_lock(&fps->fps_lock); + version = fps->fps_version; + 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); @@ -1495,219 +1644,256 @@ 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(10); - } - 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"); + 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++; + 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 kiblnd_fini_pool(kib_pool_t *pool) { - LASSERT (cfs_list_empty(&pool->po_free_list)); - LASSERT (pool->po_allocated == 0); + LASSERT(list_empty(&pool->po_free_list)); + LASSERT(pool->po_allocated == 0); - CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name); + CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name); } static void kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size) { - CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name); + CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name); - memset(pool, 0, sizeof(kib_pool_t)); - CFS_INIT_LIST_HEAD(&pool->po_free_list); - pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE); - pool->po_owner = ps; - pool->po_size = size; + memset(pool, 0, sizeof(kib_pool_t)); + INIT_LIST_HEAD(&pool->po_free_list); + pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE); + pool->po_owner = ps; + pool->po_size = size; } -void -kiblnd_destroy_pool_list(kib_poolset_t *ps, cfs_list_t *head) +static void +kiblnd_destroy_pool_list(struct list_head *head) { - kib_pool_t *pool; + kib_pool_t *pool; - while (!cfs_list_empty(head)) { - pool = cfs_list_entry(head->next, kib_pool_t, po_list); - cfs_list_del(&pool->po_list); - ps->ps_pool_destroy(pool); - } + while (!list_empty(head)) { + pool = list_entry(head->next, kib_pool_t, po_list); + list_del(&pool->po_list); + + LASSERT(pool->po_owner != NULL); + pool->po_owner->ps_pool_destroy(pool); + } } static void -kiblnd_fini_pool_set(kib_poolset_t *ps) +kiblnd_fail_poolset(kib_poolset_t *ps, struct list_head *zombies) { - kiblnd_destroy_pool_list(ps, &ps->ps_pool_list); + if (ps->ps_net == NULL) /* intialized? */ + return; + + spin_lock(&ps->ps_lock); + while (!list_empty(&ps->ps_pool_list)) { + kib_pool_t *po = list_entry(ps->ps_pool_list.next, + kib_pool_t, po_list); + po->po_failed = 1; + list_del(&po->po_list); + if (po->po_allocated == 0) + list_add(&po->po_list, zombies); + else + list_add(&po->po_list, &ps->ps_failed_pool_list); + } + spin_unlock(&ps->ps_lock); +} + +static void +kiblnd_fini_poolset(kib_poolset_t *ps) +{ + 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); - CFS_INIT_LIST_HEAD(&ps->ps_pool_list); + if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name)) + >= sizeof(ps->ps_name)) + return -E2BIG; + spin_lock_init(&ps->ps_lock); + INIT_LIST_HEAD(&ps->ps_pool_list); + INIT_LIST_HEAD(&ps->ps_failed_pool_list); - rc = ps->ps_pool_create(ps, size, &pool); - if (rc == 0) - cfs_list_add(&pool->po_list, &ps->ps_pool_list); - else - CERROR("Failed to create the first pool for %s\n", ps->ps_name); + rc = ps->ps_pool_create(ps, size, &pool); + if (rc == 0) + list_add(&pool->po_list, &ps->ps_pool_list); + else + CERROR("Failed to create the first pool for %s\n", ps->ps_name); - return rc; + return rc; +} + +static int +kiblnd_pool_is_idle(kib_pool_t *pool, cfs_time_t now) +{ + if (pool->po_allocated != 0) /* still in use */ + return 0; + if (pool->po_failed) + return 1; + return cfs_time_aftereq(now, pool->po_deadline); } void -kiblnd_pool_free_node(kib_pool_t *pool, cfs_list_t *node) +kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node) { - CFS_LIST_HEAD (zombies); - kib_poolset_t *ps = pool->po_owner; - kib_pool_t *tmp; - cfs_time_t now = cfs_time_current(); + struct list_head zombies = LIST_HEAD_INIT(zombies); + kib_poolset_t *ps = pool->po_owner; + 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); + if (ps->ps_node_fini != NULL) + ps->ps_node_fini(pool, node); - LASSERT (pool->po_allocated > 0); - cfs_list_add(node, &pool->po_free_list); - pool->po_allocated --; + LASSERT(pool->po_allocated > 0); + list_add(node, &pool->po_free_list); + pool->po_allocated--; - cfs_list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) { - /* the first pool is persistent */ - if (ps->ps_pool_list.next == &pool->po_list) - continue; + list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) { + /* the first pool is persistent */ + if (ps->ps_pool_list.next == &pool->po_list) + continue; - if (pool->po_allocated == 0 && - cfs_time_aftereq(now, pool->po_deadline)) - cfs_list_move(&pool->po_list, &zombies); - } - cfs_spin_unlock(&ps->ps_lock); + if (kiblnd_pool_is_idle(pool, now)) + list_move(&pool->po_list, &zombies); + } + spin_unlock(&ps->ps_lock); - if (!cfs_list_empty(&zombies)) - kiblnd_destroy_pool_list(ps, &zombies); + if (!list_empty(&zombies)) + kiblnd_destroy_pool_list(&zombies); } -cfs_list_t * +struct list_head * kiblnd_pool_alloc_node(kib_poolset_t *ps) { - cfs_list_t *node; - kib_pool_t *pool; - int rc; - - again: - cfs_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; - - pool->po_allocated ++; - pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE); - node = pool->po_free_list.next; - cfs_list_del(node); - - if (ps->ps_node_init != NULL) { - /* 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); + struct list_head *node; + kib_pool_t *pool; + int rc; + +again: + spin_lock(&ps->ps_lock); + list_for_each_entry(pool, &ps->ps_pool_list, po_list) { + if (list_empty(&pool->po_free_list)) + continue; + + pool->po_allocated++; + pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE); + node = pool->po_free_list.next; + list_del(node); + + if (ps->ps_node_init != NULL) { + /* still hold the lock */ + ps->ps_node_init(pool, node); + } + 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); - cfs_schedule(); + schedule(); 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); - ps->ps_increasing = 0; - if (rc == 0) { - cfs_list_add_tail(&pool->po_list, &ps->ps_pool_list); - } else { - /* retry 10 seconds later */ - ps->ps_next_retry = cfs_time_shift(10); - CERROR("Can't allocate new %s pool because out of memory\n", - ps->ps_name); - } - cfs_spin_unlock(&ps->ps_lock); + spin_lock(&ps->ps_lock); + ps->ps_increasing = 0; + if (rc == 0) { + list_add_tail(&pool->po_list, &ps->ps_pool_list); + } else { + ps->ps_next_retry = cfs_time_shift(IBLND_POOL_RETRY); + CERROR("Can't allocate new %s pool because out of memory\n", + ps->ps_name); + } + spin_unlock(&ps->ps_lock); - goto again; + goto again; } void @@ -1723,27 +1909,32 @@ kiblnd_pmr_pool_unmap(kib_phys_mr_t *pmr) } int -kiblnd_pmr_pool_map(kib_pmr_poolset_t *pps, kib_rdma_desc_t *rd, - __u64 *iova, kib_phys_mr_t **pp_pmr) -{ - kib_phys_mr_t *pmr; - cfs_list_t *node; - int rc; - int i; - - node = kiblnd_pool_alloc_node(&pps->pps_poolset); - if (node == NULL) { - CERROR("Failed to allocate PMR descriptor\n"); - return -ENOMEM; - } +kiblnd_pmr_pool_map(kib_pmr_poolset_t *pps, kib_hca_dev_t *hdev, + kib_rdma_desc_t *rd, __u64 *iova, kib_phys_mr_t **pp_pmr) +{ + kib_phys_mr_t *pmr; + struct list_head *node; + int rc; + int i; + + node = kiblnd_pool_alloc_node(&pps->pps_poolset); + if (node == NULL) { + CERROR("Failed to allocate PMR descriptor\n"); + return -ENOMEM; + } + + pmr = container_of(node, kib_phys_mr_t, pmr_list); + if (pmr->pmr_pool->ppo_hdev != hdev) { + kiblnd_pool_free_node(&pmr->pmr_pool->ppo_pool, node); + return -EAGAIN; + } - pmr = container_of(node, kib_phys_mr_t, pmr_list); for (i = 0; i < rd->rd_nfrags; i ++) { pmr->pmr_ipb[i].addr = rd->rd_frags[i].rf_addr; pmr->pmr_ipb[i].size = rd->rd_frags[i].rf_nob; } - pmr->pmr_mr = ib_reg_phys_mr(pps->pps_poolset.ps_net->ibn_dev->ibd_pd, + pmr->pmr_mr = ib_reg_phys_mr(hdev->ibh_pd, pmr->pmr_ipb, rd->rd_nfrags, IB_ACCESS_LOCAL_WRITE | IB_ACCESS_REMOTE_WRITE, @@ -1754,13 +1945,13 @@ kiblnd_pmr_pool_map(kib_pmr_poolset_t *pps, kib_rdma_desc_t *rd, return 0; } - rc = PTR_ERR(pmr->pmr_mr); - CERROR("Failed ib_reg_phys_mr: %d\n", rc); + rc = PTR_ERR(pmr->pmr_mr); + CERROR("Failed ib_reg_phys_mr: %d\n", rc); - pmr->pmr_mr = NULL; - kiblnd_pool_free_node(&pmr->pmr_pool->ppo_pool, node); + pmr->pmr_mr = NULL; + kiblnd_pool_free_node(&pmr->pmr_pool->ppo_pool, node); - return rc; + return rc; } static void @@ -1771,12 +1962,12 @@ kiblnd_destroy_pmr_pool(kib_pool_t *pool) LASSERT (pool->po_allocated == 0); - while (!cfs_list_empty(&pool->po_free_list)) { - pmr = cfs_list_entry(pool->po_free_list.next, + while (!list_empty(&pool->po_free_list)) { + pmr = list_entry(pool->po_free_list.next, kib_phys_mr_t, pmr_list); LASSERT (pmr->pmr_mr == NULL); - cfs_list_del(&pmr->pmr_list); + list_del(&pmr->pmr_list); if (pmr->pmr_ipb != NULL) { LIBCFS_FREE(pmr->pmr_ipb, @@ -1788,18 +1979,29 @@ kiblnd_destroy_pmr_pool(kib_pool_t *pool) } kiblnd_fini_pool(pool); + if (ppo->ppo_hdev != NULL) + kiblnd_hdev_decref(ppo->ppo_hdev); + 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; @@ -1809,19 +2011,18 @@ 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; - cfs_list_add(&pmr->pmr_list, &pool->po_free_list); + list_add(&pmr->pmr_list, &pool->po_free_list); } if (i < size) { @@ -1829,6 +2030,7 @@ kiblnd_create_pmr_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po) return -ENOMEM; } + ppo->ppo_hdev = kiblnd_current_hdev(ps->ps_net->ibn_dev); *pp_po = pool; return 0; } @@ -1842,8 +2044,7 @@ kiblnd_destroy_tx_pool(kib_pool_t *pool) LASSERT (pool->po_allocated == 0); if (tpo->tpo_tx_pages != NULL) { - if (tpo->tpo_tx_pages->ibp_device != NULL) - kiblnd_unmap_tx_pool(tpo); + kiblnd_unmap_tx_pool(tpo); kiblnd_free_pages(tpo->tpo_tx_pages); } @@ -1853,7 +2054,7 @@ kiblnd_destroy_tx_pool(kib_pool_t *pool) for (i = 0; i < pool->po_size; i++) { kib_tx_t *tx = &tpo->tpo_tx_descs[i]; - cfs_list_del(&tx->tx_list); + list_del(&tx->tx_list); if (tx->tx_pages != NULL) LIBCFS_FREE(tx->tx_pages, LNET_MAX_IOV * @@ -1883,6 +2084,13 @@ 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) { @@ -1891,7 +2099,7 @@ 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; @@ -1903,13 +2111,14 @@ 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); @@ -1922,36 +2131,38 @@ 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) { @@ -1965,113 +2176,197 @@ kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po) } static void -kiblnd_tx_init(kib_pool_t *pool, cfs_list_t *node) +kiblnd_tx_init(kib_pool_t *pool, struct list_head *node) { - kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t, - tps_poolset); - kib_tx_t *tx = cfs_list_entry(node, kib_tx_t, tx_list); + kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t, + tps_poolset); + kib_tx_t *tx = list_entry(node, kib_tx_t, tx_list); - tx->tx_cookie = tps->tps_next_tx_cookie ++; + tx->tx_cookie = tps->tps_next_tx_cookie++; } -void -kiblnd_ni_fini_pools(kib_net_t *net) +static void +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; - 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); + } - if (*kiblnd_tunables.kib_map_on_demand > 0 || - net->ibn_dev->ibd_nmrs > 1) { /* premapping can fail if ibd_nmr > 1, - * so we always create FMR/PMR pool and - * map-on-demand if premapping failed */ - 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; - } + 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; + } } -void -kiblnd_dev_cleanup(kib_dev_t *ibdev) -{ - int i; - - if (ibdev->ibd_mrs == NULL) - return; - - for (i = 0; i < ibdev->ibd_nmrs; i++) { - if (ibdev->ibd_mrs[i] == NULL) - break; - - ib_dereg_mr(ibdev->ibd_mrs[i]); - } - - LIBCFS_FREE(ibdev->ibd_mrs, sizeof(*ibdev->ibd_mrs) * ibdev->ibd_nmrs); - ibdev->ibd_mrs = NULL; +static 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 -kiblnd_dev_get_attr(kib_dev_t *ibdev) +kiblnd_hdev_get_attr(kib_hca_dev_t *hdev) { struct ib_device_attr *attr; int rc; /* It's safe to assume a HCA can handle a page size * matching that of the native system */ - ibdev->ibd_page_shift = PAGE_SHIFT; - ibdev->ibd_page_size = 1 << PAGE_SHIFT; - ibdev->ibd_page_mask = ~((__u64)ibdev->ibd_page_size - 1); + hdev->ibh_page_shift = PAGE_SHIFT; + hdev->ibh_page_size = 1 << PAGE_SHIFT; + hdev->ibh_page_mask = ~((__u64)hdev->ibh_page_size - 1); LIBCFS_ALLOC(attr, sizeof(*attr)); if (attr == NULL) { @@ -2079,9 +2374,9 @@ kiblnd_dev_get_attr(kib_dev_t *ibdev) return -ENOMEM; } - rc = ib_query_device(ibdev->ibd_cmid->device, attr); + rc = ib_query_device(hdev->ibh_ibdev, attr); if (rc == 0) - ibdev->ibd_mr_size = attr->max_mr_size; + hdev->ibh_mr_size = attr->max_mr_size; LIBCFS_FREE(attr, sizeof(*attr)); @@ -2090,33 +2385,58 @@ kiblnd_dev_get_attr(kib_dev_t *ibdev) 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(ibdev->ibd_cmid->device->node_type) == - RDMA_TRANSPORT_IWARP) - ibdev->ibd_mr_size = (1ULL << 32) - 1; -#endif - - if (ibdev->ibd_mr_size == ~0ULL) { - ibdev->ibd_mr_shift = 64; + if (hdev->ibh_mr_size == ~0ULL) { + hdev->ibh_mr_shift = 64; return 0; } - for (ibdev->ibd_mr_shift = 0; - ibdev->ibd_mr_shift < 64; ibdev->ibd_mr_shift ++) { - if (ibdev->ibd_mr_size == (1ULL << ibdev->ibd_mr_shift) || - ibdev->ibd_mr_size == (1ULL << ibdev->ibd_mr_shift) - 1) + for (hdev->ibh_mr_shift = 0; + hdev->ibh_mr_shift < 64; hdev->ibh_mr_shift ++) { + if (hdev->ibh_mr_size == (1ULL << hdev->ibh_mr_shift) || + hdev->ibh_mr_size == (1ULL << hdev->ibh_mr_shift) - 1) return 0; } - CERROR("Invalid mr size: "LPX64"\n", ibdev->ibd_mr_size); + CERROR("Invalid mr size: "LPX64"\n", hdev->ibh_mr_size); return -EINVAL; } -int -kiblnd_dev_setup(kib_dev_t *ibdev) +static void +kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev) +{ + int i; + + if (hdev->ibh_nmrs == 0 || hdev->ibh_mrs == NULL) + return; + + for (i = 0; i < hdev->ibh_nmrs; i++) { + if (hdev->ibh_mrs[i] == NULL) + break; + + ib_dereg_mr(hdev->ibh_mrs[i]); + } + + LIBCFS_FREE(hdev->ibh_mrs, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs); + hdev->ibh_mrs = NULL; + hdev->ibh_nmrs = 0; +} + +void +kiblnd_hdev_destroy(kib_hca_dev_t *hdev) +{ + kiblnd_hdev_cleanup_mrs(hdev); + + if (hdev->ibh_pd != NULL) + ib_dealloc_pd(hdev->ibh_pd); + + if (hdev->ibh_cmid != NULL) + rdma_destroy_id(hdev->ibh_cmid); + + LIBCFS_FREE(hdev, sizeof(*hdev)); +} + +static int +kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev) { struct ib_mr *mr; int i; @@ -2126,38 +2446,38 @@ kiblnd_dev_setup(kib_dev_t *ibdev) int acflags = IB_ACCESS_LOCAL_WRITE | IB_ACCESS_REMOTE_WRITE; - rc = kiblnd_dev_get_attr(ibdev); + rc = kiblnd_hdev_get_attr(hdev); if (rc != 0) return rc; - if (ibdev->ibd_mr_shift == 64) { - LIBCFS_ALLOC(ibdev->ibd_mrs, 1 * sizeof(*ibdev->ibd_mrs)); - if (ibdev->ibd_mrs == NULL) { + if (hdev->ibh_mr_shift == 64) { + LIBCFS_ALLOC(hdev->ibh_mrs, 1 * sizeof(*hdev->ibh_mrs)); + if (hdev->ibh_mrs == NULL) { CERROR("Failed to allocate MRs table\n"); return -ENOMEM; } - ibdev->ibd_mrs[0] = NULL; - ibdev->ibd_nmrs = 1; + hdev->ibh_mrs[0] = NULL; + hdev->ibh_nmrs = 1; - mr = ib_get_dma_mr(ibdev->ibd_pd, acflags); + mr = ib_get_dma_mr(hdev->ibh_pd, acflags); if (IS_ERR(mr)) { CERROR("Failed ib_get_dma_mr : %ld\n", PTR_ERR(mr)); - kiblnd_dev_cleanup(ibdev); + kiblnd_hdev_cleanup_mrs(hdev); return PTR_ERR(mr); } - ibdev->ibd_mrs[0] = mr; + hdev->ibh_mrs[0] = mr; goto out; } - mr_size = (1ULL << ibdev->ibd_mr_shift); + mr_size = (1ULL << hdev->ibh_mr_shift); mm_size = (unsigned long)high_memory - PAGE_OFFSET; - ibdev->ibd_nmrs = (int)((mm_size + mr_size - 1) >> ibdev->ibd_mr_shift); + hdev->ibh_nmrs = (int)((mm_size + mr_size - 1) >> hdev->ibh_mr_shift); - if (ibdev->ibd_mr_shift < 32 || ibdev->ibd_nmrs > 1024) { + if (hdev->ibh_mr_shift < 32 || hdev->ibh_nmrs > 1024) { /* it's 4T..., assume we will re-code at that time */ CERROR("Can't support memory size: x"LPX64 " with MR size: x"LPX64"\n", mm_size, mr_size); @@ -2165,74 +2485,328 @@ kiblnd_dev_setup(kib_dev_t *ibdev) } /* create an array of MRs to cover all memory */ - LIBCFS_ALLOC(ibdev->ibd_mrs, sizeof(*ibdev->ibd_mrs) * ibdev->ibd_nmrs); - if (ibdev->ibd_mrs == NULL) { + LIBCFS_ALLOC(hdev->ibh_mrs, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs); + if (hdev->ibh_mrs == NULL) { CERROR("Failed to allocate MRs' table\n"); return -ENOMEM; } - memset(ibdev->ibd_mrs, 0, sizeof(*ibdev->ibd_mrs) * ibdev->ibd_nmrs); + memset(hdev->ibh_mrs, 0, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs); - for (i = 0; i < ibdev->ibd_nmrs; i++) { + for (i = 0; i < hdev->ibh_nmrs; i++) { struct ib_phys_buf ipb; __u64 iova; - ipb.size = ibdev->ibd_mr_size; + ipb.size = hdev->ibh_mr_size; ipb.addr = i * mr_size; iova = ipb.addr; - mr = ib_reg_phys_mr(ibdev->ibd_pd, &ipb, 1, acflags, &iova); + mr = ib_reg_phys_mr(hdev->ibh_pd, &ipb, 1, acflags, &iova); if (IS_ERR(mr)) { CERROR("Failed ib_reg_phys_mr addr "LPX64 " size "LPX64" : %ld\n", ipb.addr, ipb.size, PTR_ERR(mr)); - kiblnd_dev_cleanup(ibdev); + kiblnd_hdev_cleanup_mrs(hdev); return PTR_ERR(mr); } LASSERT (iova == ipb.addr); - ibdev->ibd_mrs[i] = mr; + hdev->ibh_mrs[i] = mr; } out: - CDEBUG(D_CONSOLE, "Register global MR array, MR size: " - LPX64", array size: %d\n", - ibdev->ibd_mr_size, ibdev->ibd_nmrs); + if (hdev->ibh_mr_size != ~0ULL || hdev->ibh_nmrs != 1) + LCONSOLE_INFO("Register global MR array, MR size: " + LPX64", array size: %d\n", + hdev->ibh_mr_size, hdev->ibh_nmrs); + return 0; +} - cfs_list_add_tail(&ibdev->ibd_list, - &kiblnd_data.kib_devs); +static int +kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event) +{ /* DUMMY */ return 0; } +static int +kiblnd_dev_need_failover(kib_dev_t *dev) +{ + struct rdma_cm_id *cmid; + struct sockaddr_in srcaddr; + struct sockaddr_in dstaddr; + int rc; + + if (dev->ibd_hdev == NULL || /* initializing */ + dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */ + *kiblnd_tunables.kib_dev_failover > 1) /* debugging */ + return 1; + + /* XXX: it's UGLY, but I don't have better way to find + * ib-bonding HCA failover because: + * + * a. no reliable CM event for HCA failover... + * b. no OFED API to get ib_device for current net_device... + * + * We have only two choices at this point: + * + * a. rdma_bind_addr(), it will conflict with listener cmid + * b. rdma_resolve_addr() to zero addr */ + cmid = kiblnd_rdma_create_id(kiblnd_dummy_callback, dev, RDMA_PS_TCP, + IB_QPT_RC); + if (IS_ERR(cmid)) { + rc = PTR_ERR(cmid); + CERROR("Failed to create cmid for failover: %d\n", rc); + return rc; + } + + memset(&srcaddr, 0, sizeof(srcaddr)); + srcaddr.sin_family = AF_INET; + srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip); + + memset(&dstaddr, 0, sizeof(dstaddr)); + dstaddr.sin_family = AF_INET; + rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr, + (struct sockaddr *)&dstaddr, 1); + 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; + } + + if (dev->ibd_hdev->ibh_ibdev == cmid->device) { + /* don't need device failover */ + rdma_destroy_id(cmid); + return 0; + } + + return 1; +} + +int +kiblnd_dev_failover(kib_dev_t *dev) +{ + struct list_head zombie_tpo = LIST_HEAD_INIT(zombie_tpo); + struct list_head zombie_ppo = LIST_HEAD_INIT(zombie_ppo); + struct list_head zombie_fpo = LIST_HEAD_INIT(zombie_fpo); + struct rdma_cm_id *cmid = NULL; + kib_hca_dev_t *hdev = NULL; + kib_hca_dev_t *old; + struct ib_pd *pd; + kib_net_t *net; + struct sockaddr_in addr; + unsigned long flags; + int rc = 0; + int i; + + LASSERT (*kiblnd_tunables.kib_dev_failover > 1 || + dev->ibd_can_failover || + dev->ibd_hdev == NULL); + + rc = kiblnd_dev_need_failover(dev); + if (rc <= 0) + goto out; + + if (dev->ibd_hdev != NULL && + dev->ibd_hdev->ibh_cmid != NULL) { + /* XXX it's not good to close old listener at here, + * because we can fail to create new listener. + * But we have to close it now, otherwise rdma_bind_addr + * will return EADDRINUSE... How crap! */ + write_lock_irqsave(&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); + } + + cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP, + IB_QPT_RC); + if (IS_ERR(cmid)) { + rc = PTR_ERR(cmid); + CERROR("Failed to create cmid for failover: %d\n", rc); + goto out; + } + + memset(&addr, 0, sizeof(addr)); + addr.sin_family = AF_INET; + addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip); + addr.sin_port = htons(*kiblnd_tunables.kib_service); + + /* Bind to failover device or port */ + rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr); + 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)); + if (hdev == NULL) { + CERROR("Failed to allocate kib_hca_dev\n"); + rdma_destroy_id(cmid); + rc = -ENOMEM; + goto out; + } + + atomic_set(&hdev->ibh_ref, 1); + hdev->ibh_dev = dev; + hdev->ibh_cmid = cmid; + hdev->ibh_ibdev = cmid->device; + + pd = ib_alloc_pd(cmid->device); + if (IS_ERR(pd)) { + rc = PTR_ERR(pd); + CERROR("Can't allocate PD: %d\n", rc); + goto out; + } + + hdev->ibh_pd = pd; + + rc = rdma_listen(cmid, 0); + if (rc != 0) { + CERROR("Can't start new listener: %d\n", rc); + goto out; + } + + rc = kiblnd_hdev_setup_mrs(hdev); + if (rc != 0) { + CERROR("Can't setup device: %d\n", rc); + goto out; + } + + write_lock_irqsave(&kiblnd_data.kib_global_lock, flags); + + old = dev->ibd_hdev; + dev->ibd_hdev = hdev; /* take over the refcount */ + hdev = old; + + list_for_each_entry(net, &dev->ibd_nets, ibn_list) { + 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 (!list_empty(&zombie_tpo)) + kiblnd_destroy_pool_list(&zombie_tpo); + if (!list_empty(&zombie_ppo)) + kiblnd_destroy_pool_list(&zombie_ppo); + if (!list_empty(&zombie_fpo)) + kiblnd_destroy_fmr_pool_list(&zombie_fpo); + if (hdev != NULL) + kiblnd_hdev_decref(hdev); + + if (rc != 0) + dev->ibd_failed_failover++; + else + dev->ibd_failed_failover = 0; + + return rc; +} + void kiblnd_destroy_dev (kib_dev_t *dev) { LASSERT (dev->ibd_nnets == 0); + LASSERT(list_empty(&dev->ibd_nets)); - if (!cfs_list_empty(&dev->ibd_list)) /* on kib_devs? */ - cfs_list_del_init(&dev->ibd_list); + list_del(&dev->ibd_fail_list); + list_del(&dev->ibd_list); - kiblnd_dev_cleanup(dev); + if (dev->ibd_hdev != NULL) + kiblnd_hdev_decref(dev->ibd_hdev); - if (dev->ibd_pd != NULL) - ib_dealloc_pd(dev->ibd_pd); + LIBCFS_FREE(dev, sizeof(*dev)); +} - if (dev->ibd_cmid != NULL) - rdma_destroy_id(dev->ibd_cmid); +static kib_dev_t * +kiblnd_create_dev(char *ifname) +{ + struct net_device *netdev; + kib_dev_t *dev; + __u32 netmask; + __u32 ip; + int up; + int rc; - LIBCFS_FREE(dev, sizeof(*dev)); + rc = libcfs_ipif_query(ifname, &up, &ip, &netmask); + if (rc != 0) { + CERROR("Can't query IPoIB interface %s: %d\n", + ifname, rc); + return NULL; + } + + if (!up) { + CERROR("Can't query IPoIB interface %s: it's down\n", ifname); + return NULL; + } + + LIBCFS_ALLOC(dev, sizeof(*dev)); + if (dev == NULL) + return NULL; + + memset(dev, 0, sizeof(*dev)); +#ifdef HAVE_DEV_GET_BY_NAME_2ARG + netdev = dev_get_by_name(&init_net, ifname); +#else + netdev = dev_get_by_name(ifname); +#endif + if (netdev == NULL) { + dev->ibd_can_failover = 0; + } else { + dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER); + dev_put(netdev); + } + + INIT_LIST_HEAD(&dev->ibd_nets); + INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */ + INIT_LIST_HEAD(&dev->ibd_fail_list); + dev->ibd_ifip = ip; + strcpy(&dev->ibd_ifname[0], ifname); + + /* initialize the device */ + rc = kiblnd_dev_failover(dev); + if (rc != 0) { + CERROR("Can't initialize device: %d\n", rc); + LIBCFS_FREE(dev, sizeof(*dev)); + return NULL; + } + + list_add_tail(&dev->ibd_list, + &kiblnd_data.kib_devs); + return dev; } -void -kiblnd_base_shutdown (void) +static void +kiblnd_base_shutdown(void) { - int i; + struct kib_sched_info *sched; + int i; - LASSERT (cfs_list_empty(&kiblnd_data.kib_devs)); + LASSERT(list_empty(&kiblnd_data.kib_devs)); CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n", - cfs_atomic_read(&libcfs_kmemory)); + atomic_read(&libcfs_kmemory)); switch (kiblnd_data.kib_init) { default: @@ -2242,22 +2816,29 @@ kiblnd_base_shutdown (void) case IBLND_INIT_DATA: LASSERT (kiblnd_data.kib_peers != NULL); for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) { - LASSERT (cfs_list_empty(&kiblnd_data.kib_peers[i])); + LASSERT(list_empty(&kiblnd_data.kib_peers[i])); } - LASSERT (cfs_list_empty(&kiblnd_data.kib_connd_zombies)); - LASSERT (cfs_list_empty(&kiblnd_data.kib_connd_conns)); + LASSERT(list_empty(&kiblnd_data.kib_connd_zombies)); + LASSERT(list_empty(&kiblnd_data.kib_connd_conns)); - /* flag threads to terminate; wake and wait for them to die */ - kiblnd_data.kib_shutdown = 1; - cfs_waitq_broadcast(&kiblnd_data.kib_sched_waitq); - cfs_waitq_broadcast(&kiblnd_data.kib_connd_waitq); + /* flag threads to terminate; wake and wait for them to die */ + kiblnd_data.kib_shutdown = 1; - i = 2; - while (cfs_atomic_read(&kiblnd_data.kib_nthreads) != 0) { + /* 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) + wake_up_all(&sched->ibs_waitq); + + wake_up_all(&kiblnd_data.kib_connd_waitq); + wake_up_all(&kiblnd_data.kib_failover_waitq); + + i = 2; + while (atomic_read(&kiblnd_data.kib_nthreads) != 0) { i++; CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */ "Waiting for %d threads to terminate\n", - cfs_atomic_read(&kiblnd_data.kib_nthreads)); + atomic_read(&kiblnd_data.kib_nthreads)); cfs_pause(cfs_time_seconds(1)); } @@ -2267,23 +2848,27 @@ 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(struct list_head) * + 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)); + atomic_read(&libcfs_kmemory)); - kiblnd_data.kib_init = IBLND_INIT_NOTHING; - PORTAL_MODULE_UNUSE; + kiblnd_data.kib_init = IBLND_INIT_NOTHING; + module_put(THIS_MODULE); } -void +static 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; @@ -2293,11 +2878,11 @@ kiblnd_shutdown (lnet_ni_t *ni) goto out; CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n", - cfs_atomic_read(&libcfs_kmemory)); + 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: @@ -2309,24 +2894,27 @@ kiblnd_shutdown (lnet_ni_t *ni) /* Wait for all peer state to clean up */ i = 2; - while (cfs_atomic_read(&net->ibn_npeers) != 0) { + while (atomic_read(&net->ibn_npeers) != 0) { i++; CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* 2**n? */ "%s: waiting for %d peers to disconnect\n", libcfs_nid2str(ni->ni_nid), - cfs_atomic_read(&net->ibn_npeers)); + atomic_read(&net->ibn_npeers)); cfs_pause(cfs_time_seconds(1)); } - kiblnd_ni_fini_pools(net); + kiblnd_net_fini_pools(net); - LASSERT (net->ibn_dev->ibd_nnets > 0); - net->ibn_dev->ibd_nnets--; + write_lock_irqsave(g_lock, flags); + LASSERT(net->ibn_dev->ibd_nnets > 0); + net->ibn_dev->ibd_nnets--; + list_del(&net->ibn_list); + write_unlock_irqrestore(g_lock, flags); /* fall through */ case IBLND_INIT_NOTHING: - LASSERT (cfs_atomic_read(&net->ibn_nconns) == 0); + LASSERT (atomic_read(&net->ibn_nconns) == 0); if (net->ibn_dev != NULL && net->ibn_dev->ibd_nnets == 0) @@ -2336,7 +2924,7 @@ kiblnd_shutdown (lnet_ni_t *ni) } CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n", - cfs_atomic_read(&libcfs_kmemory)); + atomic_read(&libcfs_kmemory)); net->ibn_init = IBLND_INIT_NOTHING; ni->ni_data = NULL; @@ -2344,44 +2932,68 @@ kiblnd_shutdown (lnet_ni_t *ni) LIBCFS_FREE(net, sizeof(*net)); out: - if (cfs_list_empty(&kiblnd_data.kib_devs)) + if (list_empty(&kiblnd_data.kib_devs)) kiblnd_base_shutdown(); return; } -int -kiblnd_base_startup (void) +static int +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); + LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING); - PORTAL_MODULE_USE; - memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */ + try_module_get(THIS_MODULE); + 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); + INIT_LIST_HEAD(&kiblnd_data.kib_devs); + INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs); - kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE; - LIBCFS_ALLOC(kiblnd_data.kib_peers, - sizeof(cfs_list_t) * - kiblnd_data.kib_peer_hash_size); - if (kiblnd_data.kib_peers == NULL) { - goto failed; - } - for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) - CFS_INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]); + kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE; + LIBCFS_ALLOC(kiblnd_data.kib_peers, + sizeof(struct list_head) * + kiblnd_data.kib_peer_hash_size); + if (kiblnd_data.kib_peers == NULL) + goto failed; + + for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) + INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]); + + 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_waitqueue_head(&kiblnd_data.kib_connd_waitq); + init_waitqueue_head(&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_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_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); + spin_lock_init(&sched->ibs_lock); + INIT_LIST_HEAD(&sched->ibs_conns); + init_waitqueue_head(&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; @@ -2389,18 +3001,18 @@ 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, "kiblnd_connd"); + if (rc != 0) { + CERROR("Can't spawn o2iblnd connd: %d\n", rc); + goto failed; } - rc = kiblnd_thread_start(kiblnd_connd, 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 connd: %d\n", rc); + CERROR("Can't spawn o2iblnd failover thread: %d\n", rc); goto failed; } @@ -2415,15 +3027,116 @@ kiblnd_base_startup (void) return -ENETDOWN; } -int +static 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; +} + +static 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; +} + +static kib_dev_t * +kiblnd_dev_search(char *ifname) +{ + kib_dev_t *alias = NULL; + kib_dev_t *dev; + char *colon; + char *colon2; + + colon = strchr(ifname, ':'); + 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; +} + +static 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); @@ -2438,10 +3151,10 @@ kiblnd_startup (lnet_ni_t *ni) if (net == NULL) goto failed; - memset(net, 0, sizeof(*net)); + memset(net, 0, sizeof(*net)); - cfs_gettimeofday(&tv); - net->ibn_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec; + 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; @@ -2467,107 +3180,41 @@ 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); - - if (!strcmp(&ibdev->ibd_ifname[0], ifname)) - break; - - ibdev = NULL; - } - - if (ibdev == NULL) { - __u32 ip; - __u32 netmask; - int up; - struct rdma_cm_id *id; - struct ib_pd *pd; - struct sockaddr_in addr; + ibdev = kiblnd_dev_search(ifname); - rc = libcfs_ipif_query(ifname, &up, &ip, &netmask); - if (rc != 0) { - CERROR("Can't query IPoIB interface %s: %d\n", - ifname, rc); - goto failed; - } + 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 (!up) { - CERROR("Can't query IPoIB interface %s: it's down\n", - ifname); - goto failed; - } - - LIBCFS_ALLOC(ibdev, sizeof(*ibdev)); - if (ibdev == NULL) - goto failed; - - memset(ibdev, 0, sizeof(*ibdev)); - - CFS_INIT_LIST_HEAD(&ibdev->ibd_list); /* not yet in kib_devs */ - ibdev->ibd_ifip = ip; - strcpy(&ibdev->ibd_ifname[0], ifname); - - id = rdma_create_id(kiblnd_cm_callback, ibdev, RDMA_PS_TCP); - if (IS_ERR(id)) { - CERROR("Can't create listen ID: %ld\n", PTR_ERR(id)); - goto failed; - } - - ibdev->ibd_cmid = id; - - memset(&addr, 0, sizeof(addr)); - addr.sin_family = AF_INET; - addr.sin_port = htons(*kiblnd_tunables.kib_service); - addr.sin_addr.s_addr = htonl(ip); - - rc = rdma_bind_addr(id, (struct sockaddr *)&addr); - if (rc != 0) { - CERROR("Can't bind to %s: %d\n", ifname, rc); - goto failed; - } - - /* Binding should have assigned me an IB device */ - LASSERT (id->device != NULL); - CDEBUG(D_CONSOLE, "Listener bound to %s:%u.%u.%u.%u:%d:%s\n", - ifname, HIPQUAD(ip), *kiblnd_tunables.kib_service, - id->device->name); - - pd = ib_alloc_pd(id->device); - if (IS_ERR(pd)) { - CERROR("Can't allocate PD: %ld\n", PTR_ERR(pd)); - goto failed; - } - - ibdev->ibd_pd = pd; - - rc = rdma_listen(id, 256); - if (rc != 0) { - CERROR("Can't start listener: %d\n", rc); - goto failed; - } - - rc = kiblnd_dev_setup(ibdev); - if (rc != 0) { - CERROR("Can't setup device: %d\n", rc); - goto failed; - } - } + if (ibdev == NULL) + goto failed; - ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip); 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; } - ibdev->ibd_nnets++; + + write_lock_irqsave(&kiblnd_data.kib_global_lock, flags); + ibdev->ibd_nnets++; + list_add_tail(&net->ibn_list, &ibdev->ibd_nets); + write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); + net->ibn_init = IBLND_INIT_ALL; return 0; failed: - if (net->ibn_dev == NULL && ibdev != NULL) + if (net != NULL && net->ibn_dev == NULL && ibdev != NULL) kiblnd_destroy_dev(ibdev); kiblnd_shutdown(ni); @@ -2576,14 +3223,24 @@ failed: return -ENETDOWN; } -void __exit +static lnd_t the_o2iblnd = { + .lnd_type = O2IBLND, + .lnd_startup = kiblnd_startup, + .lnd_shutdown = kiblnd_shutdown, + .lnd_ctl = kiblnd_ctl, + .lnd_query = kiblnd_query, + .lnd_send = kiblnd_send, + .lnd_recv = kiblnd_recv, +}; + +static void __exit kiblnd_module_fini (void) { lnet_unregister_lnd(&the_o2iblnd); kiblnd_tunables_fini(); } -int __init +static int __init kiblnd_module_init (void) { int rc;