X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lnet%2Fklnds%2Fo2iblnd%2Fo2iblnd.c;h=eaf59fce3a2aa88c3464dd7d3a3b8eeb8060e1bc;hb=52d1671d98c29bd54430718b6e41df136e596c74;hp=7d3c025dda098dee122a6ddbc27e18a6a1f32875;hpb=cc5ef6ae5412c3e94061d949ef684036eb003f27;p=fs%2Flustre-release.git diff --git a/lnet/klnds/o2iblnd/o2iblnd.c b/lnet/klnds/o2iblnd/o2iblnd.c index 7d3c025..eaf59fc 100644 --- a/lnet/klnds/o2iblnd/o2iblnd.c +++ b/lnet/klnds/o2iblnd/o2iblnd.c @@ -15,11 +15,7 @@ * * You should have received a copy of the GNU General Public License * version 2 along with this program; If not, see - * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf - * - * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, - * CA 95054 USA or visit www.sun.com if you need additional information or - * have any questions. + * http://www.gnu.org/licenses/gpl-2.0.html * * GPL HEADER END */ @@ -27,7 +23,7 @@ * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. * - * Copyright (c) 2011, 2012, Intel Corporation. + * Copyright (c) 2011, 2017, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -38,21 +34,16 @@ * Author: Eric Barton */ +#include +#include + #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 struct lnet_lnd the_o2iblnd; -kib_data_t kiblnd_data; +struct kib_data kiblnd_data; -__u32 +static __u32 kiblnd_cksum (void *ptr, int nob) { char *c = ptr; @@ -107,41 +98,40 @@ kiblnd_msgtype2str(int type) static int kiblnd_msgtype2size(int type) { - const int hdr_size = offsetof(kib_msg_t, ibm_u); + const int hdr_size = offsetof(struct kib_msg, ibm_u); switch (type) { case IBLND_MSG_CONNREQ: case IBLND_MSG_CONNACK: - return hdr_size + sizeof(kib_connparams_t); + return hdr_size + sizeof(struct kib_connparams); case IBLND_MSG_NOOP: return hdr_size; case IBLND_MSG_IMMEDIATE: - return offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[0]); + return offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[0]); case IBLND_MSG_PUT_REQ: - return hdr_size + sizeof(kib_putreq_msg_t); + return hdr_size + sizeof(struct kib_putreq_msg); case IBLND_MSG_PUT_ACK: - return hdr_size + sizeof(kib_putack_msg_t); + return hdr_size + sizeof(struct kib_putack_msg); case IBLND_MSG_GET_REQ: - return hdr_size + sizeof(kib_get_msg_t); + return hdr_size + sizeof(struct kib_get_msg); case IBLND_MSG_PUT_NAK: case IBLND_MSG_PUT_DONE: case IBLND_MSG_GET_DONE: - return hdr_size + sizeof(kib_completion_msg_t); + return hdr_size + sizeof(struct kib_completion_msg); default: return -1; } } -static int -kiblnd_unpack_rd(kib_msg_t *msg, int flip) +static int kiblnd_unpack_rd(struct kib_msg *msg, int flip) { - kib_rdma_desc_t *rd; + struct kib_rdma_desc *rd; int nob; int n; int i; @@ -166,7 +156,7 @@ kiblnd_unpack_rd(kib_msg_t *msg, int flip) return 1; } - nob = offsetof (kib_msg_t, ibm_u) + + nob = offsetof(struct kib_msg, ibm_u) + kiblnd_rd_msg_size(rd, msg->ibm_type, n); if (msg->ibm_nob < nob) { @@ -186,11 +176,10 @@ kiblnd_unpack_rd(kib_msg_t *msg, int flip) return 0; } -void -kiblnd_pack_msg (lnet_ni_t *ni, kib_msg_t *msg, int version, - int credits, lnet_nid_t dstnid, __u64 dststamp) +void kiblnd_pack_msg(struct lnet_ni *ni, struct kib_msg *msg, int version, + int credits, lnet_nid_t dstnid, __u64 dststamp) { - kib_net_t *net = ni->ni_data; + struct kib_net *net = ni->ni_data; /* CAVEAT EMPTOR! all message fields not set here should have been * initialised previously. */ @@ -211,10 +200,9 @@ kiblnd_pack_msg (lnet_ni_t *ni, kib_msg_t *msg, int version, } } -int -kiblnd_unpack_msg(kib_msg_t *msg, int nob) +int kiblnd_unpack_msg(struct kib_msg *msg, int nob) { - const int hdr_size = offsetof(kib_msg_t, ibm_u); + const int hdr_size = offsetof(struct kib_msg, ibm_u); __u32 msg_cksum; __u16 version; int msg_nob; @@ -266,7 +254,7 @@ kiblnd_unpack_msg(kib_msg_t *msg, int nob) msg->ibm_cksum = msg_cksum; if (flip) { - /* leave magic unflipped as a clue to peer endianness */ + /* leave magic unflipped as a clue to peer_ni endianness */ msg->ibm_version = version; CLASSERT (sizeof(msg->ibm_type) == 1); CLASSERT (sizeof(msg->ibm_credits) == 1); @@ -324,138 +312,139 @@ 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(struct lnet_ni *ni, struct kib_peer_ni **peerp, + lnet_nid_t nid) { - kib_peer_t *peer; - kib_net_t *net = ni->ni_data; - int cpt = lnet_cpt_of_nid(nid); - unsigned long flags; + struct kib_peer_ni *peer_ni; + struct kib_net *net = ni->ni_data; + int cpt = lnet_cpt_of_nid(nid, ni); + unsigned long flags; LASSERT(net != NULL); LASSERT(nid != LNET_NID_ANY); - LIBCFS_CPT_ALLOC(peer, lnet_cpt_table(), cpt, sizeof(*peer)); - if (peer == NULL) { - CERROR("Cannot allocate peer\n"); + LIBCFS_CPT_ALLOC(peer_ni, lnet_cpt_table(), cpt, sizeof(*peer_ni)); + if (peer_ni == NULL) { + CERROR("Cannot allocate peer_ni\n"); return -ENOMEM; } - memset(peer, 0, sizeof(*peer)); /* zero flags etc */ + peer_ni->ibp_ni = ni; + peer_ni->ibp_nid = nid; + peer_ni->ibp_error = 0; + peer_ni->ibp_last_alive = 0; + peer_ni->ibp_max_frags = IBLND_MAX_RDMA_FRAGS; + peer_ni->ibp_queue_depth = ni->ni_net->net_tunables.lct_peer_tx_credits; + atomic_set(&peer_ni->ibp_refcount, 1); /* 1 ref for caller */ - 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 */ - - 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_ni->ibp_list); /* not in the peer_ni table yet */ + INIT_LIST_HEAD(&peer_ni->ibp_conns); + INIT_LIST_HEAD(&peer_ni->ibp_tx_queue); 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); write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); - *peerp = peer; - return 0; + *peerp = peer_ni; + return 0; } void -kiblnd_destroy_peer (kib_peer_t *peer) -{ - 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)); - - 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); -} - -kib_peer_t * -kiblnd_find_peer_locked (lnet_nid_t nid) +kiblnd_destroy_peer(struct kib_peer_ni *peer_ni) { - /* 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; + struct kib_net *net = peer_ni->ibp_ni->ni_data; - cfs_list_for_each (tmp, peer_list) { - - peer = cfs_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 */ - - if (peer->ibp_nid != nid) - continue; + LASSERT(net != NULL); + LASSERT (atomic_read(&peer_ni->ibp_refcount) == 0); + LASSERT(!kiblnd_peer_active(peer_ni)); + LASSERT(kiblnd_peer_idle(peer_ni)); + LASSERT(list_empty(&peer_ni->ibp_tx_queue)); + + LIBCFS_FREE(peer_ni, sizeof(*peer_ni)); + + /* NB a peer_ni's connections keep a reference on their peer_ni until + * they are destroyed, so we can be assured that _all_ state to do + * with this peer_ni has been cleaned up when its refcount drops to + * zero. */ + atomic_dec(&net->ibn_npeers); +} + +struct kib_peer_ni * +kiblnd_find_peer_locked(struct lnet_ni *ni, lnet_nid_t nid) +{ + /* the caller is responsible for accounting the additional reference + * that this creates */ + struct list_head *peer_list = kiblnd_nid2peerlist(nid); + struct list_head *tmp; + struct kib_peer_ni *peer_ni; + + list_for_each(tmp, peer_list) { + + peer_ni = list_entry(tmp, struct kib_peer_ni, ibp_list); + LASSERT(!kiblnd_peer_idle(peer_ni)); + + /* + * Match a peer if its NID and the NID of the local NI it + * communicates over are the same. Otherwise don't match + * the peer, which will result in a new lnd peer being + * created. + */ + if (peer_ni->ibp_nid != nid || + peer_ni->ibp_ni->ni_nid != ni->ni_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_ni [%p] -> %s (%d) version: %x\n", + peer_ni, libcfs_nid2str(nid), + atomic_read(&peer_ni->ibp_refcount), + peer_ni->ibp_version); + return peer_ni; + } + return NULL; } void -kiblnd_unlink_peer_locked (kib_peer_t *peer) +kiblnd_unlink_peer_locked(struct kib_peer_ni *peer_ni) { - LASSERT (cfs_list_empty(&peer->ibp_conns)); + LASSERT(list_empty(&peer_ni->ibp_conns)); - LASSERT (kiblnd_peer_active(peer)); - cfs_list_del_init(&peer->ibp_list); + LASSERT (kiblnd_peer_active(peer_ni)); + list_del_init(&peer_ni->ibp_list); /* lose peerlist's ref */ - kiblnd_peer_decref(peer); + kiblnd_peer_decref(peer_ni); } -int -kiblnd_get_peer_info (lnet_ni_t *ni, int index, - lnet_nid_t *nidp, int *count) +static int +kiblnd_get_peer_info(struct lnet_ni *ni, int index, + lnet_nid_t *nidp, int *count) { - kib_peer_t *peer; - cfs_list_t *ptmp; - int i; - unsigned long flags; + struct kib_peer_ni *peer_ni; + struct list_head *ptmp; + int i; + unsigned long 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_ni = list_entry(ptmp, struct kib_peer_ni, ibp_list); + LASSERT(!kiblnd_peer_idle(peer_ni)); - if (peer->ibp_ni != ni) - continue; + if (peer_ni->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_ni->ibp_nid; + *count = atomic_read(&peer_ni->ibp_refcount); read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); @@ -464,42 +453,42 @@ kiblnd_get_peer_info (lnet_ni_t *ni, int index, } read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); - return -ENOENT; + return -ENOENT; } -void -kiblnd_del_peer_locked (kib_peer_t *peer) +static void +kiblnd_del_peer_locked(struct kib_peer_ni *peer_ni) { - cfs_list_t *ctmp; - cfs_list_t *cnxt; - kib_conn_t *conn; + struct list_head *ctmp; + struct list_head *cnxt; + struct kib_conn *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_ni->ibp_conns)) { + kiblnd_unlink_peer_locked(peer_ni); + } else { + list_for_each_safe(ctmp, cnxt, &peer_ni->ibp_conns) { + conn = list_entry(ctmp, struct kib_conn, 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_ni's last conn unlinked it. */ + } + /* NB peer_ni now unlinked; might even be freed if the peer_ni table had the + * last ref on it. */ } -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; +static int +kiblnd_del_peer(struct lnet_ni *ni, lnet_nid_t nid) +{ + struct list_head zombies = LIST_HEAD_INIT(zombies); + struct list_head *ptmp; + struct list_head *pnxt; + struct kib_peer_ni *peer_ni; + int lo; + int hi; + int i; + unsigned long flags; + int rc = -ENOENT; write_lock_irqsave(&kiblnd_data.kib_global_lock, flags); @@ -510,68 +499,63 @@ 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_ni = list_entry(ptmp, struct kib_peer_ni, ibp_list); + LASSERT(!kiblnd_peer_idle(peer_ni)); - if (peer->ibp_ni != ni) - continue; + if (peer_ni->ibp_ni != ni) + continue; - if (!(nid == LNET_NID_ANY || peer->ibp_nid == nid)) - continue; + if (!(nid == LNET_NID_ANY || peer_ni->ibp_nid == nid)) + continue; - if (!cfs_list_empty(&peer->ibp_tx_queue)) { - LASSERT (cfs_list_empty(&peer->ibp_conns)); + if (!list_empty(&peer_ni->ibp_tx_queue)) { + LASSERT(list_empty(&peer_ni->ibp_conns)); - cfs_list_splice_init(&peer->ibp_tx_queue, - &zombies); - } + list_splice_init(&peer_ni->ibp_tx_queue, + &zombies); + } - kiblnd_del_peer_locked(peer); - rc = 0; /* matched something */ - } - } + kiblnd_del_peer_locked(peer_ni); + rc = 0; /* matched something */ + } + } write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); - kiblnd_txlist_done(ni, &zombies, -EIO); + kiblnd_txlist_done(&zombies, -EIO, LNET_MSG_STATUS_LOCAL_ERROR); - return rc; + return rc; } -kib_conn_t * -kiblnd_get_conn_by_idx (lnet_ni_t *ni, int index) +static struct kib_conn * +kiblnd_get_conn_by_idx(struct lnet_ni *ni, int index) { - kib_peer_t *peer; - cfs_list_t *ptmp; - kib_conn_t *conn; - cfs_list_t *ctmp; - int i; - unsigned long flags; + struct kib_peer_ni *peer_ni; + struct list_head *ptmp; + struct kib_conn *conn; + struct list_head *ctmp; + int i; + unsigned long 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]) { + for (i = 0; i < kiblnd_data.kib_peer_hash_size; 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_ni = list_entry(ptmp, struct kib_peer_ni, ibp_list); + LASSERT(!kiblnd_peer_idle(peer_ni)); - if (peer->ibp_ni != ni) - continue; + if (peer_ni->ibp_ni != ni) + continue; - cfs_list_for_each (ctmp, &peer->ibp_conns) { - if (index-- > 0) - continue; + list_for_each(ctmp, &peer_ni->ibp_conns) { + if (index-- > 0) + continue; - conn = cfs_list_entry(ctmp, kib_conn_t, - ibc_list); - kiblnd_conn_addref(conn); + conn = list_entry(ctmp, struct kib_conn, ibc_list); + kiblnd_conn_addref(conn); read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); return conn; @@ -583,70 +567,70 @@ kiblnd_get_conn_by_idx (lnet_ni_t *ni, int index) return NULL; } -void -kiblnd_debug_rx (kib_rx_t *rx) +static void +kiblnd_debug_rx(struct kib_rx *rx) { CDEBUG(D_CONSOLE, " %p status %d msg_type %x cred %d\n", rx, rx->rx_status, rx->rx_msg->ibm_type, rx->rx_msg->ibm_credits); } -void -kiblnd_debug_tx (kib_tx_t *tx) +static void +kiblnd_debug_tx(struct kib_tx *tx) { - CDEBUG(D_CONSOLE, " %p snd %d q %d w %d rc %d dl %lx " - "cookie "LPX64" msg %s%s type %x cred %d\n", + CDEBUG(D_CONSOLE, " %p snd %d q %d w %d rc %d dl %lld " + "cookie %#llx msg %s%s type %x cred %d\n", tx, tx->tx_sending, tx->tx_queued, tx->tx_waiting, - tx->tx_status, tx->tx_deadline, tx->tx_cookie, + tx->tx_status, ktime_to_ns(tx->tx_deadline), tx->tx_cookie, tx->tx_lntmsg[0] == NULL ? "-" : "!", tx->tx_lntmsg[1] == NULL ? "-" : "!", tx->tx_msg->ibm_type, tx->tx_msg->ibm_credits); } void -kiblnd_debug_conn (kib_conn_t *conn) +kiblnd_debug_conn(struct kib_conn *conn) { - cfs_list_t *tmp; - int i; + struct list_head *tmp; + int i; spin_lock(&conn->ibc_lock); - 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, "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, " 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, " early_rxs:\n"); + list_for_each(tmp, &conn->ibc_early_rxs) + kiblnd_debug_rx(list_entry(tmp, struct kib_rx, rx_list)); - CDEBUG(D_CONSOLE, " tx_noops:\n"); - cfs_list_for_each(tmp, &conn->ibc_tx_noops) - kiblnd_debug_tx(cfs_list_entry(tmp, kib_tx_t, tx_list)); + CDEBUG(D_CONSOLE, " tx_noops:\n"); + list_for_each(tmp, &conn->ibc_tx_noops) + kiblnd_debug_tx(list_entry(tmp, struct kib_tx, 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, struct kib_tx, 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, struct kib_tx, 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, struct kib_tx, 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, struct kib_tx, 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); i++) + kiblnd_debug_rx(&conn->ibc_rxs[i]); spin_unlock(&conn->ibc_lock); } @@ -687,14 +671,14 @@ kiblnd_setup_mtu_locked(struct rdma_cm_id *cmid) cmid->route.path_rec->mtu = mtu; } -#ifdef HAVE_OFED_IB_COMP_VECTOR static int -kiblnd_get_completion_vector(kib_conn_t *conn, int cpt) +kiblnd_get_completion_vector(struct kib_conn *conn, int cpt) { cpumask_t *mask; int vectors; int off; int i; + lnet_nid_t ibp_nid; vectors = conn->ibc_cmid->device->num_comp_vectors; if (vectors <= 1) @@ -703,8 +687,9 @@ kiblnd_get_completion_vector(kib_conn_t *conn, int cpt) mask = cfs_cpt_cpumask(lnet_cpt_table(), cpt); /* hash NID to CPU id in this partition... */ - off = conn->ibc_peer->ibp_nid % cpus_weight(*mask); - for_each_cpu_mask(i, *mask) { + ibp_nid = conn->ibc_peer->ibp_nid; + off = do_div(ibp_nid, cpumask_weight(mask)); + for_each_cpu(i, mask) { if (off-- == 0) return i % vectors; } @@ -712,25 +697,75 @@ kiblnd_get_completion_vector(kib_conn_t *conn, int cpt) LBUG(); return 1; } -#endif -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; +/* + * Get the scheduler bound to this CPT. If the scheduler has no + * threads, which means that the CPT has no CPUs, then grab the + * next scheduler that we can use. + * + * This case would be triggered if a NUMA node is configured with + * no associated CPUs. + */ +static struct kib_sched_info * +kiblnd_get_scheduler(int cpt) +{ + struct kib_sched_info *sched; + int i; + + sched = kiblnd_data.kib_scheds[cpt]; + + if (sched->ibs_nthreads > 0) + return sched; + + cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) { + if (sched->ibs_nthreads > 0) { + CDEBUG(D_NET, "scheduler[%d] has no threads. selected scheduler[%d]\n", + cpt, sched->ibs_cpt); + return sched; + } + } + + return NULL; +} + +static unsigned int kiblnd_send_wrs(struct kib_conn *conn) +{ + /* + * One WR for the LNet message + * And ibc_max_frags for the transfer WRs + */ + unsigned int ret = 1 + conn->ibc_max_frags; + enum kib_dev_caps dev_caps = conn->ibc_hdev->ibh_dev->ibd_dev_caps; + + /* FastReg needs two extra WRs for map and invalidate */ + if (dev_caps & IBLND_DEV_CAPS_FASTREG_ENABLED) + ret += 2; + + /* account for a maximum of ibc_queue_depth in-flight transfers */ + ret *= conn->ibc_queue_depth; + return ret; +} + +struct kib_conn * +kiblnd_create_conn(struct kib_peer_ni *peer_ni, 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_ni'. It also "owns" 'cmid' and destroys it when it itself + * is destroyed. On failure, the caller's ref on 'peer_ni' 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; + struct kib_net *net = peer_ni->ibp_ni->ni_data; + struct kib_dev *dev; + struct ib_qp_init_attr *init_qp_attr; struct kib_sched_info *sched; - kib_conn_t *conn; +#ifdef HAVE_IB_CQ_INIT_ATTR + struct ib_cq_init_attr cq_attr = {}; +#endif + struct kib_conn *conn; struct ib_cq *cq; unsigned long flags; int cpt; @@ -738,42 +773,56 @@ kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid, int i; LASSERT(net != NULL); - LASSERT(!cfs_in_interrupt()); + LASSERT(!in_interrupt()); dev = net->ibn_dev; - cpt = lnet_cpt_of_nid(peer->ibp_nid); - sched = kiblnd_data.kib_scheds[cpt]; + cpt = lnet_cpt_of_nid(peer_ni->ibp_nid, peer_ni->ibp_ni); + sched = kiblnd_get_scheduler(cpt); + + if (sched == NULL) { + CERROR("no schedulers available. node is unhealthy\n"); + goto failed_0; + } - LASSERT(sched->ibs_nthreads > 0); + /* + * The cpt might have changed if we ended up selecting a non cpt + * native scheduler. So use the scheduler's cpt instead. + */ + cpt = sched->ibs_cpt; 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)); + libcfs_nid2str(peer_ni->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; - } + if (conn == NULL) { + CERROR("Can't allocate connection for %s\n", + libcfs_nid2str(peer_ni->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; - - CFS_INIT_LIST_HEAD(&conn->ibc_early_rxs); - CFS_INIT_LIST_HEAD(&conn->ibc_tx_noops); - 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); + conn->ibc_state = IBLND_CONN_INIT; + conn->ibc_version = version; + conn->ibc_peer = peer_ni; /* I take the caller's ref */ + cmid->context = conn; /* for future CM callbacks */ + conn->ibc_cmid = cmid; + conn->ibc_max_frags = peer_ni->ibp_max_frags; + conn->ibc_queue_depth = peer_ni->ibp_queue_depth; + conn->ibc_rxs = NULL; + conn->ibc_rx_pages = NULL; + + 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); + INIT_LIST_HEAD(&conn->ibc_zombie_txs); spin_lock_init(&conn->ibc_lock); LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt, @@ -790,19 +839,19 @@ kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid, goto failed_2; } - if (dev->ibd_hdev->ibh_ibdev != cmid->device) { - /* wakeup failover thread and teardown connection */ - if (kiblnd_dev_can_failover(dev)) { - cfs_list_add_tail(&dev->ibd_fail_list, - &kiblnd_data.kib_failed_devs); - cfs_waitq_signal(&kiblnd_data.kib_failover_waitq); - } + 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; - } + 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; @@ -811,85 +860,110 @@ kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *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 +#ifdef HAVE_IB_CQ_INIT_ATTR + cq_attr.cqe = IBLND_CQ_ENTRIES(conn); + cq_attr.comp_vector = kiblnd_get_completion_vector(conn, cpt); cq = ib_create_cq(cmid->device, kiblnd_cq_completion, kiblnd_cq_event, conn, - IBLND_CQ_ENTRIES(version), - kiblnd_get_completion_vector(conn, cpt)); + &cq_attr); #else - cq = ib_create_cq(cmid->device, - kiblnd_cq_completion, kiblnd_cq_event, conn, - IBLND_CQ_ENTRIES(version)); + cq = ib_create_cq(cmid->device, + kiblnd_cq_completion, kiblnd_cq_event, conn, + IBLND_CQ_ENTRIES(conn), + kiblnd_get_completion_vector(conn, cpt)); #endif - if (IS_ERR(cq)) { - CERROR("Can't create CQ: %ld, cqe: %d\n", - PTR_ERR(cq), IBLND_CQ_ENTRIES(version)); - goto failed_2; - } + if (IS_ERR(cq)) { + /* + * on MLX-5 (possibly MLX-4 as well) this error could be + * hit if the concurrent_sends and/or peer_tx_credits is set + * too high. Or due to an MLX-5 bug which tries to + * allocate 256kb via kmalloc for WR cookie array + */ + CERROR("Failed to create CQ with %d CQEs: %ld\n", + IBLND_CQ_ENTRIES(conn), PTR_ERR(cq)); + goto failed_2; + } conn->ibc_cq = cq; - rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP); - if (rc != 0) { - CERROR("Can't request completion notificiation: %d\n", rc); - goto failed_2; - } + rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP); + if (rc != 0) { + CERROR("Can't request completion notification: %d\n", rc); + goto failed_2; + } - 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); - init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(version); - init_qp_attr->cap.max_send_sge = 1; - init_qp_attr->cap.max_recv_sge = 1; - init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR; - init_qp_attr->qp_type = IB_QPT_RC; - init_qp_attr->send_cq = cq; - init_qp_attr->recv_cq = cq; + init_qp_attr->event_handler = kiblnd_qp_event; + init_qp_attr->qp_context = conn; + init_qp_attr->cap.max_send_sge = *kiblnd_tunables.kib_wrq_sge; + init_qp_attr->cap.max_recv_sge = 1; + init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR; + init_qp_attr->qp_type = IB_QPT_RC; + init_qp_attr->send_cq = cq; + init_qp_attr->recv_cq = cq; conn->ibc_sched = sched; - rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr); - if (rc != 0) { - CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d\n", - rc, init_qp_attr->cap.max_send_wr, - init_qp_attr->cap.max_recv_wr); - goto failed_2; - } + do { + init_qp_attr->cap.max_send_wr = kiblnd_send_wrs(conn); + init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(conn); - LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr)); + rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr); + if (!rc || conn->ibc_queue_depth < 2) + break; + + conn->ibc_queue_depth--; + } while (rc); + + if (rc) { + CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d, " + "send_sge: %d, recv_sge: %d\n", + rc, init_qp_attr->cap.max_send_wr, + init_qp_attr->cap.max_recv_wr, + init_qp_attr->cap.max_send_sge, + init_qp_attr->cap.max_recv_sge); + goto failed_2; + } + + if (conn->ibc_queue_depth != peer_ni->ibp_queue_depth) + CWARN("peer %s - queue depth reduced from %u to %u" + " to allow for qp creation\n", + libcfs_nid2str(peer_ni->ibp_nid), + peer_ni->ibp_queue_depth, + conn->ibc_queue_depth); + + LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt, + IBLND_RX_MSGS(conn) * sizeof(struct kib_rx)); + 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(conn)); + if (rc != 0) + goto failed_2; + + kiblnd_map_rx_descs(conn); + + 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)); - conn->ibc_nrx = IBLND_RX_MSGS(version); + /* 1 ref for caller and each rxmsg */ + atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(conn)); + conn->ibc_nrx = IBLND_RX_MSGS(conn); - /* post receives */ - for (i = 0; i < IBLND_RX_MSGS(version); i++) { - rc = kiblnd_post_rx(&conn->ibc_rxs[i], - IBLND_POSTRX_NO_CREDIT); - if (rc != 0) { - CERROR("Can't post rxmsg: %d\n", rc); + /* post receives */ + for (i = 0; i < IBLND_RX_MSGS(conn); i++) { + rc = kiblnd_post_rx(&conn->ibc_rxs[i], IBLND_POSTRX_NO_CREDIT); + if (rc != 0) { + CERROR("Can't post rxmsg: %d\n", rc); - /* Make posted receives complete */ - kiblnd_abort_receives(conn); + /* Make posted receives complete */ + kiblnd_abort_receives(conn); - /* correct # of posted buffers - * NB locking needed now I'm racing with completion */ + /* correct # of posted buffers + * NB locking needed now I'm racing with completion */ spin_lock_irqsave(&sched->ibs_lock, flags); - conn->ibc_nrx -= IBLND_RX_MSGS(version) - i; + conn->ibc_nrx -= IBLND_RX_MSGS(conn) - i; spin_unlock_irqrestore(&sched->ibs_lock, flags); /* cmid will be destroyed by CM(ofed) after cm_callback @@ -898,9 +972,9 @@ kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid, rdma_destroy_qp(conn->ibc_cmid); conn->ibc_cmid = NULL; - /* Drop my own and unused rxbuffer refcounts */ - while (i++ <= IBLND_RX_MSGS(version)) - kiblnd_conn_decref(conn); + /* Drop my own and unused rxbuffer refcounts */ + while (i++ <= IBLND_RX_MSGS(conn)) + kiblnd_conn_decref(conn); return NULL; } @@ -912,11 +986,12 @@ 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: - kiblnd_destroy_conn(conn); + kiblnd_destroy_conn(conn); + LIBCFS_FREE(conn, sizeof(*conn)); failed_1: LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr)); failed_0: @@ -924,175 +999,174 @@ 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; - - 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_noops)); - 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(); - - case IBLND_CONN_DISCONNECTED: - /* connvars should have been freed already */ - LASSERT (conn->ibc_connvars == NULL); - break; +kiblnd_destroy_conn(struct kib_conn *conn) +{ + struct rdma_cm_id *cmid = conn->ibc_cmid; + struct kib_peer_ni *peer_ni = conn->ibc_peer; + int rc; + + 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); + + switch (conn->ibc_state) { + default: + /* conn must be completely disengaged from the network */ + LBUG(); + + case IBLND_CONN_DISCONNECTED: + /* connvars should have been freed already */ + LASSERT (conn->ibc_connvars == NULL); + break; - case IBLND_CONN_INIT: - break; - } + case IBLND_CONN_INIT: + break; + } - /* conn->ibc_cmid might be destroyed by CM already */ - if (cmid != NULL && cmid->qp != NULL) - rdma_destroy_qp(cmid); + /* conn->ibc_cmid might be destroyed by CM already */ + if (cmid != NULL && cmid->qp != NULL) + rdma_destroy_qp(cmid); - 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_cq != NULL) { + rc = ib_destroy_cq(conn->ibc_cq); + if (rc != 0) + CWARN("Error destroying CQ: %d\n", rc); + } - if (conn->ibc_rx_pages != NULL) - kiblnd_unmap_rx_descs(conn); + kiblnd_txlist_done(&conn->ibc_zombie_txs, -ECONNABORTED, + LNET_MSG_STATUS_OK); - 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) * sizeof(struct kib_rx)); + } - if (conn->ibc_hdev != NULL) - kiblnd_hdev_decref(conn->ibc_hdev); + if (conn->ibc_connvars != NULL) + LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars)); - /* 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_hdev != NULL) + kiblnd_hdev_decref(conn->ibc_hdev); - kiblnd_peer_decref(peer); - rdma_destroy_id(cmid); - cfs_atomic_dec(&net->ibn_nconns); - } + /* See CAVEAT EMPTOR above in kiblnd_create_conn */ + if (conn->ibc_state != IBLND_CONN_INIT) { + struct kib_net *net = peer_ni->ibp_ni->ni_data; - LIBCFS_FREE(conn, sizeof(*conn)); + kiblnd_peer_decref(peer_ni); + rdma_destroy_id(cmid); + atomic_dec(&net->ibn_nconns); + } } int -kiblnd_close_peer_conns_locked (kib_peer_t *peer, int why) +kiblnd_close_peer_conns_locked(struct kib_peer_ni *peer_ni, int why) { - kib_conn_t *conn; - cfs_list_t *ctmp; - cfs_list_t *cnxt; - int count = 0; + struct kib_conn *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_ni->ibp_conns) { + conn = list_entry(ctmp, struct kib_conn, 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_ni->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(struct kib_peer_ni *peer_ni, + int version, __u64 incarnation) { - kib_conn_t *conn; - cfs_list_t *ctmp; - cfs_list_t *cnxt; - int count = 0; + struct kib_conn *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_ni->ibp_conns) { + conn = list_entry(ctmp, struct kib_conn, 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:%#llx(%x, %#llx)\n", + libcfs_nid2str(peer_ni->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(struct lnet_ni *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; + struct kib_peer_ni *peer_ni; + struct list_head *ptmp; + struct list_head *pnxt; + int lo; + int hi; + int i; + unsigned long flags; + int count = 0; 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_ni = list_entry(ptmp, struct kib_peer_ni, ibp_list); + LASSERT(!kiblnd_peer_idle(peer_ni)); - if (peer->ibp_ni != ni) - continue; + if (peer_ni->ibp_ni != ni) + continue; - if (!(nid == LNET_NID_ANY || nid == peer->ibp_nid)) - continue; + if (!(nid == LNET_NID_ANY || nid == peer_ni->ibp_nid)) + continue; - count += kiblnd_close_peer_conns_locked(peer, 0); - } - } + count += kiblnd_close_peer_conns_locked(peer_ni, 0); + } + } 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 -kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg) +static int +kiblnd_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg) { struct libcfs_ioctl_data *data = arg; int rc = -EINVAL; @@ -1114,7 +1188,7 @@ kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg) break; } case IOC_LIBCFS_GET_CONN: { - kib_conn_t *conn; + struct kib_conn *conn; rc = 0; conn = kiblnd_get_conn_by_idx(ni, data->ioc_count); @@ -1123,15 +1197,15 @@ kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg) break; } - LASSERT (conn->ibc_cmid != NULL); - data->ioc_nid = conn->ibc_peer->ibp_nid; - if (conn->ibc_cmid->route.path_rec == NULL) - data->ioc_u32[0] = 0; /* iWarp has no path MTU */ - else - data->ioc_u32[0] = - ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu); - kiblnd_conn_decref(conn); - break; + LASSERT(conn->ibc_cmid != NULL); + data->ioc_nid = conn->ibc_peer->ibp_nid; + if (conn->ibc_cmid->route.path_rec == NULL) + data->ioc_u32[0] = 0; /* iWarp has no path MTU */ + else + data->ioc_u32[0] = + ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu); + kiblnd_conn_decref(conn); + break; } case IOC_LIBCFS_CLOSE_CONNECTION: { rc = kiblnd_close_matching_conns(ni, data->ioc_nid); @@ -1145,74 +1219,70 @@ 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(struct lnet_ni *ni, lnet_nid_t nid, time64_t *when) { - 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; + time64_t last_alive = 0; + time64_t now = ktime_get_seconds(); + rwlock_t *glock = &kiblnd_data.kib_global_lock; + struct kib_peer_ni *peer_ni; + unsigned long 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_ni = kiblnd_find_peer_locked(ni, nid); + if (peer_ni != NULL) + last_alive = peer_ni->ibp_last_alive; 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); + /* peer_ni is not persistent in hash, trigger peer_ni creation + * and connection establishment with a NULL tx */ + if (peer_ni == 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; + CDEBUG(D_NET, "peer_ni %s %p, alive %lld secs ago\n", + libcfs_nid2str(nid), peer_ni, + last_alive ? now - last_alive : -1); + return; } -void -kiblnd_free_pages(kib_pages_t *p) +static void +kiblnd_free_pages(struct kib_pages *p) { int npages = p->ibp_npages; int i; for (i = 0; i < npages; i++) { if (p->ibp_pages[i] != NULL) - cfs_free_page(p->ibp_pages[i]); + __free_page(p->ibp_pages[i]); } - LIBCFS_FREE(p, offsetof(kib_pages_t, ibp_pages[npages])); + LIBCFS_FREE(p, offsetof(struct kib_pages, ibp_pages[npages])); } int -kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages) +kiblnd_alloc_pages(struct kib_pages **pp, int cpt, int npages) { - kib_pages_t *p; - int i; + struct kib_pages *p; + int i; LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt, - offsetof(kib_pages_t, ibp_pages[npages])); + offsetof(struct kib_pages, ibp_pages[npages])); if (p == NULL) { CERROR("Can't allocate descriptor for %d pages\n", npages); return -ENOMEM; } - memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages])); + memset(p, 0, offsetof(struct kib_pages, ibp_pages[npages])); p->ibp_npages = npages; for (i = 0; i < npages; i++) { p->ibp_pages[i] = cfs_page_cpt_alloc(lnet_cpt_table(), cpt, - CFS_ALLOC_IO); + GFP_NOFS); if (p->ibp_pages[i] == NULL) { CERROR("Can't allocate page %d of %d\n", i, npages); kiblnd_free_pages(p); @@ -1225,24 +1295,24 @@ kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages) } void -kiblnd_unmap_rx_descs(kib_conn_t *conn) +kiblnd_unmap_rx_descs(struct kib_conn *conn) { - kib_rx_t *rx; + struct kib_rx *rx; int i; LASSERT (conn->ibc_rxs != NULL); LASSERT (conn->ibc_hdev != NULL); - for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++) { - rx = &conn->ibc_rxs[i]; + for (i = 0; i < IBLND_RX_MSGS(conn); i++) { + rx = &conn->ibc_rxs[i]; - LASSERT (rx->rx_nob >= 0); /* not posted */ + LASSERT(rx->rx_nob >= 0); /* not posted */ - kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev, - KIBLND_UNMAP_ADDR(rx, rx_msgunmap, - rx->rx_msgaddr), - IBLND_MSG_SIZE, DMA_FROM_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); + } kiblnd_free_pages(conn->ibc_rx_pages); @@ -1250,50 +1320,50 @@ kiblnd_unmap_rx_descs(kib_conn_t *conn) } void -kiblnd_map_rx_descs(kib_conn_t *conn) +kiblnd_map_rx_descs(struct kib_conn *conn) { - kib_rx_t *rx; + struct kib_rx *rx; struct page *pg; int pg_off; int ipg; int i; - for (pg_off = ipg = i = 0; - i < IBLND_RX_MSGS(conn->ibc_version); i++) { - pg = conn->ibc_rx_pages->ibp_pages[ipg]; - rx = &conn->ibc_rxs[i]; + for (pg_off = ipg = i = 0; i < IBLND_RX_MSGS(conn); i++) { + pg = conn->ibc_rx_pages->ibp_pages[ipg]; + rx = &conn->ibc_rxs[i]; - rx->rx_conn = conn; - rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off); + rx->rx_conn = conn; + rx->rx_msg = (struct kib_msg *)(((char *)page_address(pg)) + pg_off); - 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_hdev->ibh_ibdev, - rx->rx_msgaddr)); - KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr); + 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_hdev->ibh_ibdev, + rx->rx_msgaddr)); + KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr); - CDEBUG(D_NET,"rx %d: %p "LPX64"("LPX64")\n", - i, rx->rx_msg, rx->rx_msgaddr, - lnet_page2phys(pg) + pg_off); + CDEBUG(D_NET, "rx %d: %p %#llx(%#llx)\n", + i, rx->rx_msg, rx->rx_msgaddr, + (__u64)(page_to_phys(pg) + pg_off)); - pg_off += IBLND_MSG_SIZE; - LASSERT (pg_off <= PAGE_SIZE); + pg_off += IBLND_MSG_SIZE; + LASSERT(pg_off <= PAGE_SIZE); - if (pg_off == PAGE_SIZE) { - pg_off = 0; - ipg++; - LASSERT (ipg <= IBLND_RX_MSG_PAGES(conn->ibc_version)); - } - } + if (pg_off == PAGE_SIZE) { + pg_off = 0; + ipg++; + LASSERT(ipg <= IBLND_RX_MSG_PAGES(conn)); + } + } } static void -kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo) +kiblnd_unmap_tx_pool(struct kib_tx_pool *tpo) { - kib_hca_dev_t *hdev = tpo->tpo_hdev; - kib_tx_t *tx; - int i; + struct kib_hca_dev *hdev = tpo->tpo_hdev; + struct kib_tx *tx; + int i; LASSERT (tpo->tpo_pool.po_allocated == 0); @@ -1312,10 +1382,10 @@ kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo) tpo->tpo_hdev = NULL; } -static kib_hca_dev_t * -kiblnd_current_hdev(kib_dev_t *dev) +static struct kib_hca_dev * +kiblnd_current_hdev(struct kib_dev *dev) { - kib_hca_dev_t *hdev; + struct kib_hca_dev *hdev; unsigned long flags; int i = 0; @@ -1325,7 +1395,8 @@ kiblnd_current_hdev(kib_dev_t *dev) if (i++ % 50 == 0) CDEBUG(D_NET, "%s: Wait for failover\n", dev->ibd_ifname); - cfs_schedule_timeout(cfs_time_seconds(1) / 100); + set_current_state(TASK_INTERRUPTIBLE); + schedule_timeout(cfs_time_seconds(1) / 100); read_lock_irqsave(&kiblnd_data.kib_global_lock, flags); } @@ -1339,14 +1410,14 @@ kiblnd_current_hdev(kib_dev_t *dev) } static void -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; +kiblnd_map_tx_pool(struct kib_tx_pool *tpo) +{ + struct kib_pages *txpgs = tpo->tpo_tx_pages; + struct kib_pool *pool = &tpo->tpo_pool; + struct kib_net *net = pool->po_owner->ps_net; + struct kib_dev *dev; + struct page *page; + struct kib_tx *tx; int page_offset; int ipage; int i; @@ -1363,191 +1434,283 @@ kiblnd_map_tx_pool(kib_tx_pool_t *tpo) 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 = (struct kib_msg *)(((char *)page_address(page)) + + page_offset); - 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); + 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_hca_dev_t *hdev, __u64 addr, __u64 size) +static void +kiblnd_destroy_fmr_pool(struct kib_fmr_pool *fpo) { - __u64 index; + LASSERT(fpo->fpo_map_count == 0); - LASSERT (hdev->ibh_mrs[0] != NULL); + if (fpo->fpo_is_fmr && fpo->fmr.fpo_fmr_pool) { + ib_destroy_fmr_pool(fpo->fmr.fpo_fmr_pool); + } else { + struct kib_fast_reg_descriptor *frd, *tmp; + int i = 0; + + list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list, + frd_list) { + list_del(&frd->frd_list); +#ifndef HAVE_IB_MAP_MR_SG + ib_free_fast_reg_page_list(frd->frd_frpl); +#endif + ib_dereg_mr(frd->frd_mr); + LIBCFS_FREE(frd, sizeof(*frd)); + i++; + } + if (i < fpo->fast_reg.fpo_pool_size) + CERROR("FastReg pool still has %d regions registered\n", + fpo->fast_reg.fpo_pool_size - i); + } - if (hdev->ibh_nmrs == 1) - return hdev->ibh_mrs[0]; + if (fpo->fpo_hdev) + kiblnd_hdev_decref(fpo->fpo_hdev); - index = addr >> hdev->ibh_mr_shift; + LIBCFS_FREE(fpo, sizeof(*fpo)); +} - if (index < hdev->ibh_nmrs && - index == ((addr + size - 1) >> hdev->ibh_mr_shift)) - return hdev->ibh_mrs[index]; +static void +kiblnd_destroy_fmr_pool_list(struct list_head *head) +{ + struct kib_fmr_pool *fpo, *tmp; - return NULL; + list_for_each_entry_safe(fpo, tmp, head, fpo_list) { + list_del(&fpo->fpo_list); + kiblnd_destroy_fmr_pool(fpo); + } } -struct ib_mr * -kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev, kib_rdma_desc_t *rd) +static int +kiblnd_fmr_pool_size(struct lnet_ioctl_config_o2iblnd_tunables *tunables, + int ncpts) { - struct ib_mr *prev_mr; - struct ib_mr *mr; - int i; - - LASSERT (hdev->ibh_mrs[0] != NULL); + int size = tunables->lnd_fmr_pool_size / ncpts; - if (*kiblnd_tunables.kib_map_on_demand > 0 && - *kiblnd_tunables.kib_map_on_demand <= rd->rd_nfrags) - return NULL; + return max(IBLND_FMR_POOL, size); +} - if (hdev->ibh_nmrs == 1) - return hdev->ibh_mrs[0]; +static int +kiblnd_fmr_flush_trigger(struct lnet_ioctl_config_o2iblnd_tunables *tunables, + int ncpts) +{ + int size = tunables->lnd_fmr_flush_trigger / ncpts; - for (i = 0, mr = prev_mr = NULL; - i < rd->rd_nfrags; i++) { - mr = kiblnd_find_dma_mr(hdev, - rd->rd_frags[i].rf_addr, - rd->rd_frags[i].rf_nob); - if (prev_mr == NULL) - prev_mr = mr; + return max(IBLND_FMR_POOL_FLUSH, size); +} - if (mr == NULL || prev_mr != mr) { - /* Can't covered by one single MR */ - mr = NULL; - break; - } - } +static int kiblnd_alloc_fmr_pool(struct kib_fmr_poolset *fps, + struct kib_fmr_pool *fpo) +{ + struct ib_fmr_pool_param param = { + .max_pages_per_fmr = LNET_MAX_IOV, + .page_shift = PAGE_SHIFT, + .access = (IB_ACCESS_LOCAL_WRITE | + IB_ACCESS_REMOTE_WRITE), + .pool_size = fps->fps_pool_size, + .dirty_watermark = fps->fps_flush_trigger, + .flush_function = NULL, + .flush_arg = NULL, + .cache = !!fps->fps_cache }; + int rc = 0; + + fpo->fmr.fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd, + ¶m); + if (IS_ERR(fpo->fmr.fpo_fmr_pool)) { + rc = PTR_ERR(fpo->fmr.fpo_fmr_pool); + if (rc != -ENOSYS) + CERROR("Failed to create FMR pool: %d\n", rc); + else + CERROR("FMRs are not supported\n"); + } + fpo->fpo_is_fmr = true; - return mr; + return rc; } -void -kiblnd_destroy_fmr_pool(kib_fmr_pool_t *pool) +static int kiblnd_alloc_freg_pool(struct kib_fmr_poolset *fps, + struct kib_fmr_pool *fpo, + enum kib_dev_caps dev_caps) { - LASSERT (pool->fpo_map_count == 0); + struct kib_fast_reg_descriptor *frd, *tmp; + int i, rc; - if (pool->fpo_fmr_pool != NULL) - ib_destroy_fmr_pool(pool->fpo_fmr_pool); + fpo->fpo_is_fmr = false; - if (pool->fpo_hdev != NULL) - kiblnd_hdev_decref(pool->fpo_hdev); + INIT_LIST_HEAD(&fpo->fast_reg.fpo_pool_list); + fpo->fast_reg.fpo_pool_size = 0; + for (i = 0; i < fps->fps_pool_size; i++) { + LIBCFS_CPT_ALLOC(frd, lnet_cpt_table(), fps->fps_cpt, + sizeof(*frd)); + if (!frd) { + CERROR("Failed to allocate a new fast_reg descriptor\n"); + rc = -ENOMEM; + goto out; + } + frd->frd_mr = NULL; + +#ifndef HAVE_IB_MAP_MR_SG + frd->frd_frpl = ib_alloc_fast_reg_page_list(fpo->fpo_hdev->ibh_ibdev, + LNET_MAX_IOV); + if (IS_ERR(frd->frd_frpl)) { + rc = PTR_ERR(frd->frd_frpl); + CERROR("Failed to allocate ib_fast_reg_page_list: %d\n", + rc); + frd->frd_frpl = NULL; + goto out_middle; + } +#endif - LIBCFS_FREE(pool, sizeof(kib_fmr_pool_t)); -} +#ifdef HAVE_IB_ALLOC_FAST_REG_MR + frd->frd_mr = ib_alloc_fast_reg_mr(fpo->fpo_hdev->ibh_pd, + LNET_MAX_IOV); +#else + /* + * it is expected to get here if this is an MLX-5 card. + * MLX-4 cards will always use FMR and MLX-5 cards will + * always use fast_reg. It turns out that some MLX-5 cards + * (possibly due to older FW versions) do not natively support + * gaps. So we will need to track them here. + */ + frd->frd_mr = ib_alloc_mr(fpo->fpo_hdev->ibh_pd, +#ifdef IB_MR_TYPE_SG_GAPS + ((*kiblnd_tunables.kib_use_fastreg_gaps == 1) && + (dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT)) ? + IB_MR_TYPE_SG_GAPS : + IB_MR_TYPE_MEM_REG, +#else + IB_MR_TYPE_MEM_REG, +#endif + LNET_MAX_IOV); + if ((*kiblnd_tunables.kib_use_fastreg_gaps == 1) && + (dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT)) + CWARN("using IB_MR_TYPE_SG_GAPS, expect a performance drop\n"); +#endif + if (IS_ERR(frd->frd_mr)) { + rc = PTR_ERR(frd->frd_mr); + CERROR("Failed to allocate ib_fast_reg_mr: %d\n", rc); + frd->frd_mr = NULL; + goto out_middle; + } -void -kiblnd_destroy_fmr_pool_list(cfs_list_t *head) -{ - kib_fmr_pool_t *pool; + /* There appears to be a bug in MLX5 code where you must + * invalidate the rkey of a new FastReg pool before first + * using it. Thus, I am marking the FRD invalid here. */ + frd->frd_valid = false; - 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); - } -} + list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list); + fpo->fast_reg.fpo_pool_size++; + } -static int kiblnd_fmr_pool_size(int ncpts) -{ - int size = *kiblnd_tunables.kib_fmr_pool_size / ncpts; + return 0; - return max(IBLND_FMR_POOL, size); -} +out_middle: + if (frd->frd_mr) + ib_dereg_mr(frd->frd_mr); +#ifndef HAVE_IB_MAP_MR_SG + if (frd->frd_frpl) + ib_free_fast_reg_page_list(frd->frd_frpl); +#endif + LIBCFS_FREE(frd, sizeof(*frd)); -static int kiblnd_fmr_flush_trigger(int ncpts) -{ - int size = *kiblnd_tunables.kib_fmr_flush_trigger / ncpts; +out: + list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list, + frd_list) { + list_del(&frd->frd_list); +#ifndef HAVE_IB_MAP_MR_SG + ib_free_fast_reg_page_list(frd->frd_frpl); +#endif + ib_dereg_mr(frd->frd_mr); + LIBCFS_FREE(frd, sizeof(*frd)); + } - return max(IBLND_FMR_POOL_FLUSH, size); + return rc; } -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 = fps->fps_pool_size, - .dirty_watermark = fps->fps_flush_trigger, - .flush_function = NULL, - .flush_arg = NULL, - .cache = !!*kiblnd_tunables.kib_fmr_cache}; +static int kiblnd_create_fmr_pool(struct kib_fmr_poolset *fps, + struct kib_fmr_pool **pp_fpo) +{ + struct kib_dev *dev = fps->fps_net->ibn_dev; + struct kib_fmr_pool *fpo; int rc; LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo)); - if (fpo == NULL) + if (!fpo) { return -ENOMEM; + } + memset(fpo, 0, sizeof(*fpo)); 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); + if (dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED) + rc = kiblnd_alloc_fmr_pool(fps, fpo); + else + rc = kiblnd_alloc_freg_pool(fps, fpo, dev->ibd_dev_caps); + if (rc) + goto out_fpo; - kiblnd_hdev_decref(fpo->fpo_hdev); - LIBCFS_FREE(fpo, sizeof(kib_fmr_pool_t)); - return rc; - } + fpo->fpo_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE; + fpo->fpo_owner = fps; + *pp_fpo = fpo; - fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE); - fpo->fpo_owner = fps; - *pp_fpo = fpo; + return 0; - return 0; +out_fpo: + kiblnd_hdev_decref(fpo->fpo_hdev); + LIBCFS_FREE(fpo, sizeof(*fpo)); + return rc; } static void -kiblnd_fail_fmr_poolset(kib_fmr_poolset_t *fps, cfs_list_t *zombies) +kiblnd_fail_fmr_poolset(struct kib_fmr_poolset *fps, struct list_head *zombies) { if (fps->fps_net == NULL) /* intialized? */ return; spin_lock(&fps->fps_lock); - while (!cfs_list_empty(&fps->fps_pool_list)) { - kib_fmr_pool_t *fpo = cfs_list_entry(fps->fps_pool_list.next, - kib_fmr_pool_t, fpo_list); - fpo->fpo_failed = 1; - cfs_list_del(&fpo->fpo_list); - if (fpo->fpo_map_count == 0) - cfs_list_add(&fpo->fpo_list, zombies); - else - cfs_list_add(&fpo->fpo_list, &fps->fps_failed_pool_list); - } + while (!list_empty(&fps->fps_pool_list)) { + struct kib_fmr_pool *fpo = list_entry(fps->fps_pool_list.next, + struct kib_fmr_pool, + 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_fini_fmr_poolset(struct kib_fmr_poolset *fps) { if (fps->fps_net != NULL) { /* initialized? */ kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list); @@ -1556,109 +1719,269 @@ kiblnd_fini_fmr_poolset(kib_fmr_poolset_t *fps) } static int -kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt, kib_net_t *net, - int pool_size, int flush_trigger) +kiblnd_init_fmr_poolset(struct kib_fmr_poolset *fps, int cpt, int ncpts, + struct kib_net *net, + struct lnet_ioctl_config_o2iblnd_tunables *tunables) { - kib_fmr_pool_t *fpo; - int rc; + struct kib_fmr_pool *fpo; + int rc; - memset(fps, 0, sizeof(kib_fmr_poolset_t)); + memset(fps, 0, sizeof(struct kib_fmr_poolset)); - fps->fps_net = net; + fps->fps_net = net; fps->fps_cpt = cpt; - fps->fps_pool_size = pool_size; - fps->fps_flush_trigger = flush_trigger; + + fps->fps_pool_size = kiblnd_fmr_pool_size(tunables, ncpts); + fps->fps_flush_trigger = kiblnd_fmr_flush_trigger(tunables, ncpts); + fps->fps_cache = tunables->lnd_fmr_cache; + spin_lock_init(&fps->fps_lock); - CFS_INIT_LIST_HEAD(&fps->fps_pool_list); - CFS_INIT_LIST_HEAD(&fps->fps_failed_pool_list); + INIT_LIST_HEAD(&fps->fps_pool_list); + INIT_LIST_HEAD(&fps->fps_failed_pool_list); - rc = kiblnd_create_fmr_pool(fps, &fpo); - if (rc == 0) - cfs_list_add_tail(&fpo->fpo_list, &fps->fps_pool_list); + rc = kiblnd_create_fmr_pool(fps, &fpo); + if (rc == 0) + list_add_tail(&fpo->fpo_list, &fps->fps_pool_list); - return rc; + return rc; } static int -kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, cfs_time_t now) +kiblnd_fmr_pool_is_idle(struct kib_fmr_pool *fpo, time64_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); + return now >= fpo->fpo_deadline; +} + +static int +kiblnd_map_tx_pages(struct kib_tx *tx, struct kib_rdma_desc *rd) +{ + struct kib_hca_dev *hdev; + __u64 *pages = tx->tx_pages; + int npages; + int size; + int i; + + hdev = tx->tx_pool->tpo_hdev; + + for (i = 0, npages = 0; i < rd->rd_nfrags; i++) { + for (size = 0; size < rd->rd_frags[i].rf_nob; + size += hdev->ibh_page_size) { + pages[npages++] = (rd->rd_frags[i].rf_addr & + hdev->ibh_page_mask) + size; + } + } + + return npages; } void -kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status) +kiblnd_fmr_pool_unmap(struct kib_fmr *fmr, int status) { - CFS_LIST_HEAD (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; + struct list_head zombies = LIST_HEAD_INIT(zombies); + struct kib_fmr_pool *fpo = fmr->fmr_pool; + struct kib_fmr_poolset *fps; + time64_t now = ktime_get_seconds(); + struct kib_fmr_pool *tmp; + int rc; - rc = ib_fmr_pool_unmap(fmr->fmr_pfmr); - LASSERT (rc == 0); + if (!fpo) + return; - if (status != 0) { - rc = ib_flush_fmr_pool(fpo->fpo_fmr_pool); - LASSERT (rc == 0); - } + fps = fpo->fpo_owner; + if (fpo->fpo_is_fmr) { + if (fmr->fmr_pfmr) { + ib_fmr_pool_unmap(fmr->fmr_pfmr); + fmr->fmr_pfmr = NULL; + } + + if (status) { + rc = ib_flush_fmr_pool(fpo->fmr.fpo_fmr_pool); + LASSERT(!rc); + } + } else { + struct kib_fast_reg_descriptor *frd = fmr->fmr_frd; - fmr->fmr_pool = NULL; - fmr->fmr_pfmr = NULL; + if (frd) { + frd->frd_valid = false; + spin_lock(&fps->fps_lock); + list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list); + spin_unlock(&fps->fps_lock); + fmr->fmr_frd = NULL; + } + } + fmr->fmr_pool = NULL; spin_lock(&fps->fps_lock); - fpo->fpo_map_count --; /* decref the pool */ + 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 (kiblnd_fmr_pool_is_idle(fpo, now)) { - cfs_list_move(&fpo->fpo_list, &zombies); - fps->fps_version ++; - } - } + 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 -kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages, - __u64 iov, kib_fmr_t *fmr) +int kiblnd_fmr_pool_map(struct kib_fmr_poolset *fps, struct kib_tx *tx, + struct kib_rdma_desc *rd, u32 nob, u64 iov, + struct kib_fmr *fmr) { - struct ib_pool_fmr *pfmr; - kib_fmr_pool_t *fpo; - __u64 version; - int rc; + struct kib_fmr_pool *fpo; + __u64 *pages = tx->tx_pages; + __u64 version; + bool is_rx = (rd != tx->tx_rd); + bool tx_pages_mapped = 0; + int npages = 0; + int rc; - again: +again: 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); + list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) { + fpo->fpo_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE; fpo->fpo_map_count++; - spin_unlock(&fps->fps_lock); - pfmr = ib_fmr_pool_map_phys(fpo->fpo_fmr_pool, - pages, npages, iov); - if (likely(!IS_ERR(pfmr))) { - fmr->fmr_pool = fpo; - fmr->fmr_pfmr = pfmr; - return 0; - } + if (fpo->fpo_is_fmr) { + struct ib_pool_fmr *pfmr; + + spin_unlock(&fps->fps_lock); + + if (!tx_pages_mapped) { + npages = kiblnd_map_tx_pages(tx, rd); + tx_pages_mapped = 1; + } + + pfmr = kib_fmr_pool_map(fpo->fmr.fpo_fmr_pool, + pages, npages, iov); + if (likely(!IS_ERR(pfmr))) { + fmr->fmr_key = is_rx ? pfmr->fmr->rkey + : pfmr->fmr->lkey; + fmr->fmr_frd = NULL; + fmr->fmr_pfmr = pfmr; + fmr->fmr_pool = fpo; + return 0; + } + rc = PTR_ERR(pfmr); + } else { + if (!list_empty(&fpo->fast_reg.fpo_pool_list)) { + struct kib_fast_reg_descriptor *frd; +#ifdef HAVE_IB_MAP_MR_SG + struct ib_reg_wr *wr; + int n; +#else + struct ib_rdma_wr *wr; + struct ib_fast_reg_page_list *frpl; +#endif + struct ib_mr *mr; + + frd = list_first_entry(&fpo->fast_reg.fpo_pool_list, + struct kib_fast_reg_descriptor, + frd_list); + list_del(&frd->frd_list); + spin_unlock(&fps->fps_lock); + +#ifndef HAVE_IB_MAP_MR_SG + frpl = frd->frd_frpl; +#endif + mr = frd->frd_mr; + + if (!frd->frd_valid) { + struct ib_rdma_wr *inv_wr; + __u32 key = is_rx ? mr->rkey : mr->lkey; + + inv_wr = &frd->frd_inv_wr; + memset(inv_wr, 0, sizeof(*inv_wr)); + + inv_wr->wr.opcode = IB_WR_LOCAL_INV; + inv_wr->wr.wr_id = IBLND_WID_MR; + inv_wr->wr.ex.invalidate_rkey = key; + + /* Bump the key */ + key = ib_inc_rkey(key); + ib_update_fast_reg_key(mr, key); + } + +#ifdef HAVE_IB_MAP_MR_SG +#ifdef HAVE_IB_MAP_MR_SG_5ARGS + n = ib_map_mr_sg(mr, tx->tx_frags, + tx->tx_nfrags, NULL, PAGE_SIZE); +#else + n = ib_map_mr_sg(mr, tx->tx_frags, + tx->tx_nfrags, PAGE_SIZE); +#endif + if (unlikely(n != tx->tx_nfrags)) { + CERROR("Failed to map mr %d/%d " + "elements\n", n, tx->tx_nfrags); + return n < 0 ? n : -EINVAL; + } + + wr = &frd->frd_fastreg_wr; + memset(wr, 0, sizeof(*wr)); + + wr->wr.opcode = IB_WR_REG_MR; + wr->wr.wr_id = IBLND_WID_MR; + wr->wr.num_sge = 0; + wr->wr.send_flags = 0; + wr->mr = mr; + wr->key = is_rx ? mr->rkey : mr->lkey; + wr->access = (IB_ACCESS_LOCAL_WRITE | + IB_ACCESS_REMOTE_WRITE); +#else + if (!tx_pages_mapped) { + npages = kiblnd_map_tx_pages(tx, rd); + tx_pages_mapped = 1; + } + + LASSERT(npages <= frpl->max_page_list_len); + memcpy(frpl->page_list, pages, + sizeof(*pages) * npages); + + /* Prepare FastReg WR */ + wr = &frd->frd_fastreg_wr; + memset(wr, 0, sizeof(*wr)); + + wr->wr.opcode = IB_WR_FAST_REG_MR; + wr->wr.wr_id = IBLND_WID_MR; + + wr->wr.wr.fast_reg.iova_start = iov; + wr->wr.wr.fast_reg.page_list = frpl; + wr->wr.wr.fast_reg.page_list_len = npages; + wr->wr.wr.fast_reg.page_shift = PAGE_SHIFT; + wr->wr.wr.fast_reg.length = nob; + wr->wr.wr.fast_reg.rkey = + is_rx ? mr->rkey : mr->lkey; + wr->wr.wr.fast_reg.access_flags = + (IB_ACCESS_LOCAL_WRITE | + IB_ACCESS_REMOTE_WRITE); +#endif + + fmr->fmr_key = is_rx ? mr->rkey : mr->lkey; + fmr->fmr_frd = frd; + fmr->fmr_pfmr = NULL; + fmr->fmr_pool = fpo; + return 0; + } + spin_unlock(&fps->fps_lock); + rc = -EAGAIN; + } spin_lock(&fps->fps_lock); fpo->fpo_map_count--; - if (PTR_ERR(pfmr) != -EAGAIN) { + if (rc != -EAGAIN) { spin_unlock(&fps->fps_lock); - return PTR_ERR(pfmr); + return rc; } /* EAGAIN and ... */ @@ -1672,12 +1995,12 @@ kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages, spin_unlock(&fps->fps_lock); CDEBUG(D_NET, "Another thread is allocating new " "FMR pool, waiting for her to complete\n"); - cfs_schedule(); + schedule(); goto again; } - if (cfs_time_before(cfs_time_current(), fps->fps_next_retry)) { + if (ktime_get_seconds() < fps->fps_next_retry) { /* someone failed recently */ spin_unlock(&fps->fps_lock); return -EAGAIN; @@ -1692,9 +2015,9 @@ kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages, fps->fps_increasing = 0; if (rc == 0) { fps->fps_version++; - cfs_list_add_tail(&fpo->fpo_list, &fps->fps_pool_list); + list_add_tail(&fpo->fpo_list, &fps->fps_pool_list); } else { - fps->fps_next_retry = cfs_time_shift(IBLND_POOL_RETRY); + fps->fps_next_retry = ktime_get_seconds() + IBLND_POOL_RETRY; } spin_unlock(&fps->fps_lock); @@ -1702,62 +2025,63 @@ kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages, } static void -kiblnd_fini_pool(kib_pool_t *pool) +kiblnd_fini_pool(struct kib_pool *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) +kiblnd_init_pool(struct kib_poolset *ps, struct kib_pool *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(struct kib_pool)); + INIT_LIST_HEAD(&pool->po_free_list); + pool->po_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE; + pool->po_owner = ps; + pool->po_size = size; } -void -kiblnd_destroy_pool_list(cfs_list_t *head) +static void +kiblnd_destroy_pool_list(struct list_head *head) { - kib_pool_t *pool; + struct kib_pool *pool; - while (!cfs_list_empty(head)) { - pool = cfs_list_entry(head->next, kib_pool_t, po_list); - cfs_list_del(&pool->po_list); + while (!list_empty(head)) { + pool = list_entry(head->next, struct kib_pool, po_list); + list_del(&pool->po_list); - LASSERT (pool->po_owner != NULL); - pool->po_owner->ps_pool_destroy(pool); - } + LASSERT(pool->po_owner != NULL); + pool->po_owner->ps_pool_destroy(pool); + } } static void -kiblnd_fail_poolset(kib_poolset_t *ps, cfs_list_t *zombies) +kiblnd_fail_poolset(struct kib_poolset *ps, struct list_head *zombies) { if (ps->ps_net == NULL) /* intialized? */ return; spin_lock(&ps->ps_lock); - while (!cfs_list_empty(&ps->ps_pool_list)) { - kib_pool_t *po = cfs_list_entry(ps->ps_pool_list.next, - kib_pool_t, po_list); - po->po_failed = 1; - cfs_list_del(&po->po_list); - if (po->po_allocated == 0) - cfs_list_add(&po->po_list, zombies); - else - cfs_list_add(&po->po_list, &ps->ps_failed_pool_list); - } + while (!list_empty(&ps->ps_pool_list)) { + struct kib_pool *po = list_entry(ps->ps_pool_list.next, + struct kib_pool, 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) +kiblnd_fini_poolset(struct kib_poolset *ps) { if (ps->ps_net != NULL) { /* initialized? */ kiblnd_destroy_pool_list(&ps->ps_failed_pool_list); @@ -1766,17 +2090,17 @@ kiblnd_fini_poolset(kib_poolset_t *ps) } static int -kiblnd_init_poolset(kib_poolset_t *ps, int cpt, - kib_net_t *net, char *name, int size, +kiblnd_init_poolset(struct kib_poolset *ps, int cpt, + struct kib_net *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; + struct kib_pool *pool; + int rc; - memset(ps, 0, sizeof(kib_poolset_t)); + memset(ps, 0, sizeof(struct kib_poolset)); ps->ps_cpt = cpt; ps->ps_net = net; @@ -1789,81 +2113,85 @@ kiblnd_init_poolset(kib_poolset_t *ps, int cpt, >= sizeof(ps->ps_name)) return -E2BIG; spin_lock_init(&ps->ps_lock); - CFS_INIT_LIST_HEAD(&ps->ps_pool_list); - CFS_INIT_LIST_HEAD(&ps->ps_failed_pool_list); + 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) +kiblnd_pool_is_idle(struct kib_pool *pool, time64_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); + return now >= pool->po_deadline; } void -kiblnd_pool_free_node(kib_pool_t *pool, cfs_list_t *node) +kiblnd_pool_free_node(struct kib_pool *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); + struct kib_poolset *ps = pool->po_owner; + struct kib_pool *tmp; + time64_t now = ktime_get_seconds(); 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 (kiblnd_pool_is_idle(pool, now)) - cfs_list_move(&pool->po_list, &zombies); - } + if (kiblnd_pool_is_idle(pool, now)) + list_move(&pool->po_list, &zombies); + } spin_unlock(&ps->ps_lock); - if (!cfs_list_empty(&zombies)) + if (!list_empty(&zombies)) kiblnd_destroy_pool_list(&zombies); } -cfs_list_t * -kiblnd_pool_alloc_node(kib_poolset_t *ps) +struct list_head * +kiblnd_pool_alloc_node(struct kib_poolset *ps) { - cfs_list_t *node; - kib_pool_t *pool; - int rc; + struct list_head *node; + struct kib_pool *pool; + int rc; + unsigned int interval = 1; + ktime_t time_before; + unsigned int trips = 0; - again: +again: 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); - } + 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 = ktime_get_seconds() + + 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; } @@ -1872,14 +2200,21 @@ kiblnd_pool_alloc_node(kib_poolset_t *ps) if (ps->ps_increasing) { /* another thread is allocating a new pool */ spin_unlock(&ps->ps_lock); + trips++; CDEBUG(D_NET, "Another thread is allocating new " - "%s pool, waiting for her to complete\n", - ps->ps_name); - cfs_schedule(); + "%s pool, waiting %d HZs for her to complete." + "trips = %d\n", + ps->ps_name, interval, trips); + + set_current_state(TASK_INTERRUPTIBLE); + schedule_timeout(interval); + if (interval < cfs_time_seconds(1)) + interval *= 2; + goto again; } - if (cfs_time_before(cfs_time_current(), ps->ps_next_retry)) { + if (ktime_get_seconds() < ps->ps_next_retry) { /* someone failed recently */ spin_unlock(&ps->ps_lock); return NULL; @@ -1889,167 +2224,31 @@ kiblnd_pool_alloc_node(kib_poolset_t *ps) spin_unlock(&ps->ps_lock); CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name); - + time_before = ktime_get(); rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool); + CDEBUG(D_NET, "ps_pool_create took %lld ms to complete", + ktime_ms_delta(ktime_get(), time_before)); spin_lock(&ps->ps_lock); - ps->ps_increasing = 0; - if (rc == 0) { - cfs_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); - } + ps->ps_increasing = 0; + if (rc == 0) { + list_add_tail(&pool->po_list, &ps->ps_pool_list); + } else { + ps->ps_next_retry = ktime_get_seconds() + 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; } -void -kiblnd_pmr_pool_unmap(kib_phys_mr_t *pmr) -{ - kib_pmr_pool_t *ppo = pmr->pmr_pool; - struct ib_mr *mr = pmr->pmr_mr; - - pmr->pmr_mr = NULL; - kiblnd_pool_free_node(&ppo->ppo_pool, &pmr->pmr_list); - if (mr != NULL) - ib_dereg_mr(mr); -} - -int -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; - 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; - } - - 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; - } - - 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(hdev->ibh_pd, - pmr->pmr_ipb, rd->rd_nfrags, - IB_ACCESS_LOCAL_WRITE | - IB_ACCESS_REMOTE_WRITE, - iova); - if (!IS_ERR(pmr->pmr_mr)) { - pmr->pmr_iova = *iova; - *pp_pmr = pmr; - return 0; - } - - 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); - - return rc; -} - static void -kiblnd_destroy_pmr_pool(kib_pool_t *pool) +kiblnd_destroy_tx_pool(struct kib_pool *pool) { - kib_pmr_pool_t *ppo = container_of(pool, kib_pmr_pool_t, ppo_pool); - kib_phys_mr_t *pmr; - - LASSERT (pool->po_allocated == 0); - - while (!cfs_list_empty(&pool->po_free_list)) { - pmr = cfs_list_entry(pool->po_free_list.next, - kib_phys_mr_t, pmr_list); - - LASSERT (pmr->pmr_mr == NULL); - cfs_list_del(&pmr->pmr_list); - - if (pmr->pmr_ipb != NULL) { - LIBCFS_FREE(pmr->pmr_ipb, - IBLND_MAX_RDMA_FRAGS * - sizeof(struct ib_phys_buf)); - } - - LIBCFS_FREE(pmr, sizeof(kib_phys_mr_t)); - } - - 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) -{ - struct kib_pmr_pool *ppo; - struct kib_pool *pool; - kib_phys_mr_t *pmr; - int i; - - 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; - } - - pool = &ppo->ppo_pool; - kiblnd_init_pool(ps, pool, size); - - for (i = 0; i < size; i++) { - 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); - } - - if (i < size) { - ps->ps_pool_destroy(pool); - return -ENOMEM; - } - - ppo->ppo_hdev = kiblnd_current_hdev(ps->ps_net->ibn_dev); - *pp_po = pool; - return 0; -} - -static void -kiblnd_destroy_tx_pool(kib_pool_t *pool) -{ - kib_tx_pool_t *tpo = container_of(pool, kib_tx_pool_t, tpo_pool); - int i; + struct kib_tx_pool *tpo = container_of(pool, struct kib_tx_pool, + tpo_pool); + int i; LASSERT (pool->po_allocated == 0); @@ -2062,52 +2261,57 @@ kiblnd_destroy_tx_pool(kib_pool_t *pool) goto out; for (i = 0; i < pool->po_size; i++) { - kib_tx_t *tx = &tpo->tpo_tx_descs[i]; + struct kib_tx *tx = &tpo->tpo_tx_descs[i]; + int wrq_sge = *kiblnd_tunables.kib_wrq_sge; - cfs_list_del(&tx->tx_list); + list_del(&tx->tx_list); if (tx->tx_pages != NULL) LIBCFS_FREE(tx->tx_pages, LNET_MAX_IOV * sizeof(*tx->tx_pages)); if (tx->tx_frags != NULL) LIBCFS_FREE(tx->tx_frags, - IBLND_MAX_RDMA_FRAGS * - sizeof(*tx->tx_frags)); + (1 + IBLND_MAX_RDMA_FRAGS) * + sizeof(*tx->tx_frags)); if (tx->tx_wrq != NULL) LIBCFS_FREE(tx->tx_wrq, (1 + IBLND_MAX_RDMA_FRAGS) * sizeof(*tx->tx_wrq)); - if (tx->tx_sge != NULL) - LIBCFS_FREE(tx->tx_sge, - (1 + IBLND_MAX_RDMA_FRAGS) * - sizeof(*tx->tx_sge)); + if (tx->tx_sge != NULL) + LIBCFS_FREE(tx->tx_sge, + (1 + IBLND_MAX_RDMA_FRAGS) * wrq_sge * + sizeof(*tx->tx_sge)); if (tx->tx_rd != NULL) LIBCFS_FREE(tx->tx_rd, - offsetof(kib_rdma_desc_t, + offsetof(struct kib_rdma_desc, rd_frags[IBLND_MAX_RDMA_FRAGS])); } LIBCFS_FREE(tpo->tpo_tx_descs, - pool->po_size * sizeof(kib_tx_t)); + pool->po_size * sizeof(struct kib_tx)); out: kiblnd_fini_pool(pool); - LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t)); + LIBCFS_FREE(tpo, sizeof(struct kib_tx_pool)); } -static int kiblnd_tx_pool_size(int ncpts) +static int kiblnd_tx_pool_size(struct lnet_ni *ni, int ncpts) { - int ntx = *kiblnd_tunables.kib_ntx / ncpts; + struct lnet_ioctl_config_o2iblnd_tunables *tunables; + int ntx; + + tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib; + ntx = tunables->lnd_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) +kiblnd_create_tx_pool(struct kib_poolset *ps, int size, struct kib_pool **pp_po) { int i; int npg; - kib_pool_t *pool; - kib_tx_pool_t *tpo; + struct kib_pool *pool; + struct kib_tx_pool *tpo; LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo)); if (tpo == NULL) { @@ -2123,22 +2327,23 @@ kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po) npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE; 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)); + LIBCFS_FREE(tpo, sizeof(struct kib_tx_pool)); return -ENOMEM; } LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt, - size * sizeof(kib_tx_t)); + size * sizeof(struct kib_tx)); if (tpo->tpo_tx_descs == NULL) { CERROR("Can't allocate %d tx descriptors\n", size); ps->ps_pool_destroy(pool); return -ENOMEM; } - memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t)); + memset(tpo->tpo_tx_descs, 0, size * sizeof(struct kib_tx)); for (i = 0; i < size; i++) { - kib_tx_t *tx = &tpo->tpo_tx_descs[i]; + struct kib_tx *tx = &tpo->tpo_tx_descs[i]; + int wrq_sge = *kiblnd_tunables.kib_wrq_sge; tx->tx_pool = tpo; if (ps->ps_net->ibn_fmr_ps != NULL) { @@ -2150,11 +2355,12 @@ kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po) } LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt, - IBLND_MAX_RDMA_FRAGS * sizeof(*tx->tx_frags)); + (1 + IBLND_MAX_RDMA_FRAGS) * + sizeof(*tx->tx_frags)); if (tx->tx_frags == NULL) break; - sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS); + sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS + 1); LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt, (1 + IBLND_MAX_RDMA_FRAGS) * @@ -2163,13 +2369,13 @@ kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po) break; LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt, - (1 + IBLND_MAX_RDMA_FRAGS) * + (1 + IBLND_MAX_RDMA_FRAGS) * wrq_sge * 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, + offsetof(struct kib_rdma_desc, rd_frags[IBLND_MAX_RDMA_FRAGS])); if (tx->tx_rd == NULL) break; @@ -2186,24 +2392,24 @@ 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(struct kib_pool *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); + struct kib_tx_poolset *tps = container_of(pool->po_owner, + struct kib_tx_poolset, + tps_poolset); + struct kib_tx *tx = list_entry(node, struct kib_tx, tx_list); - tx->tx_cookie = tps->tps_next_tx_cookie ++; + tx->tx_cookie = tps->tps_next_tx_cookie++; } -void -kiblnd_net_fini_pools(kib_net_t *net) +static void +kiblnd_net_fini_pools(struct kib_net *net) { int i; cfs_cpt_for_each(i, lnet_cpt_table()) { - kib_tx_poolset_t *tps; - kib_fmr_poolset_t *fps; - kib_pmr_poolset_t *pps; + struct kib_tx_poolset *tps; + struct kib_fmr_poolset *fps; if (net->ibn_tx_ps != NULL) { tps = net->ibn_tx_ps[i]; @@ -2214,11 +2420,6 @@ kiblnd_net_fini_pools(kib_net_t *net) fps = net->ibn_fmr_ps[i]; kiblnd_fini_fmr_poolset(fps); } - - if (net->ibn_pmr_ps != NULL) { - pps = net->ibn_pmr_ps[i]; - kiblnd_fini_poolset(&pps->pps_poolset); - } } if (net->ibn_tx_ps != NULL) { @@ -2230,49 +2431,55 @@ kiblnd_net_fini_pools(kib_net_t *net) cfs_percpt_free(net->ibn_fmr_ps); net->ibn_fmr_ps = NULL; } - - if (net->ibn_pmr_ps != NULL) { - cfs_percpt_free(net->ibn_pmr_ps); - net->ibn_pmr_ps = NULL; - } } -int -kiblnd_net_init_pools(kib_net_t *net, __u32 *cpts, int ncpts) +static int +kiblnd_net_init_pools(struct kib_net *net, struct lnet_ni *ni, __u32 *cpts, + int ncpts) { + struct lnet_ioctl_config_o2iblnd_tunables *tunables; +#ifdef HAVE_IB_GET_DMA_MR unsigned long flags; +#endif int cpt; int rc; int i; + tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib; + +#ifdef HAVE_IB_GET_DMA_MR 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) { + /* + * if lnd_map_on_demand is zero then we have effectively disabled + * FMR or FastReg and we're using global memory regions + * exclusively. + */ + if (!tunables->lnd_map_on_demand) { read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); goto create_tx_pool; } read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); +#endif - if (*kiblnd_tunables.kib_fmr_pool_size < - *kiblnd_tunables.kib_ntx / 4) { + if (tunables->lnd_fmr_pool_size < tunables->lnd_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); + tunables->lnd_fmr_pool_size, + tunables->lnd_ntx / 4); rc = -EINVAL; goto failed; } - /* TX pool must be created later than FMR/PMR, see LU-2268 + /* TX pool must be created later than FMR, 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 */ + * FMR pool and map-on-demand if premapping failed */ net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(), - sizeof(kib_fmr_poolset_t)); + sizeof(struct kib_fmr_poolset)); if (net->ibn_fmr_ps == NULL) { CERROR("Failed to allocate FMR pool array\n"); rc = -ENOMEM; @@ -2281,63 +2488,23 @@ kiblnd_net_init_pools(kib_net_t *net, __u32 *cpts, int ncpts) for (i = 0; i < ncpts; i++) { cpt = (cpts == NULL) ? i : cpts[i]; - rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, net, - kiblnd_fmr_pool_size(ncpts), - kiblnd_fmr_flush_trigger(ncpts)); - if (rc == -ENOSYS && i == 0) /* no FMR */ - break; /* create PMR pool */ - - if (rc != 0) { /* a real error */ + rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, ncpts, + net, tunables); + if (rc != 0) { CERROR("Can't initialize FMR pool for CPT %d: %d\n", cpt, rc); goto failed; } } - if (i > 0) { + 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; - } - } +#ifdef HAVE_IB_GET_DMA_MR create_tx_pool: +#endif net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(), - sizeof(kib_tx_poolset_t)); + sizeof(struct kib_tx_poolset)); if (net->ibn_tx_ps == NULL) { CERROR("Failed to allocate tx pool array\n"); rc = -ENOMEM; @@ -2348,7 +2515,7 @@ kiblnd_net_init_pools(kib_net_t *net, __u32 *cpts, int ncpts) 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_tx_pool_size(ni, ncpts), kiblnd_create_tx_pool, kiblnd_destroy_tx_pool, kiblnd_tx_init, NULL); @@ -2367,74 +2534,99 @@ kiblnd_net_init_pools(kib_net_t *net, __u32 *cpts, int ncpts) } static int -kiblnd_hdev_get_attr(kib_hca_dev_t *hdev) +kiblnd_hdev_get_attr(struct kib_hca_dev *hdev) { - struct ib_device_attr *attr; - int rc; + struct ib_device_attr *dev_attr; + int rc = 0; - /* It's safe to assume a HCA can handle a page size - * matching that of the native system */ - hdev->ibh_page_shift = PAGE_SHIFT; - hdev->ibh_page_size = 1 << PAGE_SHIFT; - hdev->ibh_page_mask = ~((__u64)hdev->ibh_page_size - 1); + /* It's safe to assume a HCA can handle a page size + * matching that of the native system */ + 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) { - CERROR("Out of memory\n"); - return -ENOMEM; - } +#ifndef HAVE_IB_DEVICE_ATTRS + LIBCFS_ALLOC(dev_attr, sizeof(*dev_attr)); + if (dev_attr == NULL) { + CERROR("Out of memory\n"); + return -ENOMEM; + } - rc = ib_query_device(hdev->ibh_ibdev, attr); - if (rc == 0) - hdev->ibh_mr_size = attr->max_mr_size; + rc = ib_query_device(hdev->ibh_ibdev, dev_attr); + if (rc != 0) { + CERROR("Failed to query IB device: %d\n", rc); + goto out_clean_attr; + } +#else + dev_attr = &hdev->ibh_ibdev->attrs; +#endif - LIBCFS_FREE(attr, sizeof(*attr)); + hdev->ibh_mr_size = dev_attr->max_mr_size; - if (rc != 0) { - CERROR("Failed to query IB device: %d\n", rc); - return rc; - } + /* Setup device Memory Registration capabilities */ +#ifdef HAVE_IB_DEVICE_OPS + if (hdev->ibh_ibdev->ops.alloc_fmr && + hdev->ibh_ibdev->ops.dealloc_fmr && + hdev->ibh_ibdev->ops.map_phys_fmr && + hdev->ibh_ibdev->ops.unmap_fmr) { +#else + if (hdev->ibh_ibdev->alloc_fmr && + hdev->ibh_ibdev->dealloc_fmr && + hdev->ibh_ibdev->map_phys_fmr && + hdev->ibh_ibdev->unmap_fmr) { +#endif + LCONSOLE_INFO("Using FMR for registration\n"); + hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FMR_ENABLED; + } else if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) { + LCONSOLE_INFO("Using FastReg for registration\n"); + hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_ENABLED; +#ifndef HAVE_IB_ALLOC_FAST_REG_MR +#ifdef IB_DEVICE_SG_GAPS_REG + if (dev_attr->device_cap_flags & IB_DEVICE_SG_GAPS_REG) + hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT; +#endif +#endif + } else { + rc = -ENOSYS; + } - if (hdev->ibh_mr_size == ~0ULL) { - hdev->ibh_mr_shift = 64; - return 0; - } + if (rc == 0 && hdev->ibh_mr_size == ~0ULL) + hdev->ibh_mr_shift = 64; + else if (rc != 0) + rc = -EINVAL; - 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; - } +#ifndef HAVE_IB_DEVICE_ATTRS +out_clean_attr: + LIBCFS_FREE(dev_attr, sizeof(*dev_attr)); +#endif - CERROR("Invalid mr size: "LPX64"\n", hdev->ibh_mr_size); - return -EINVAL; + if (rc == -ENOSYS) + CERROR("IB device does not support FMRs nor FastRegs, can't " + "register memory: %d\n", rc); + else if (rc == -EINVAL) + CERROR("Invalid mr size: %#llx\n", hdev->ibh_mr_size); + return rc; } -void -kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev) +#ifdef HAVE_IB_GET_DMA_MR +static void +kiblnd_hdev_cleanup_mrs(struct kib_hca_dev *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; + if (hdev->ibh_mrs == NULL) + return; - ib_dereg_mr(hdev->ibh_mrs[i]); - } + ib_dereg_mr(hdev->ibh_mrs); - LIBCFS_FREE(hdev->ibh_mrs, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs); - hdev->ibh_mrs = NULL; - hdev->ibh_nmrs = 0; + hdev->ibh_mrs = NULL; } +#endif void -kiblnd_hdev_destroy(kib_hca_dev_t *hdev) +kiblnd_hdev_destroy(struct kib_hca_dev *hdev) { +#ifdef HAVE_IB_GET_DMA_MR kiblnd_hdev_cleanup_mrs(hdev); +#endif if (hdev->ibh_pd != NULL) ib_dealloc_pd(hdev->ibh_pd); @@ -2445,93 +2637,26 @@ kiblnd_hdev_destroy(kib_hca_dev_t *hdev) LIBCFS_FREE(hdev, sizeof(*hdev)); } -int -kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev) -{ - struct ib_mr *mr; - int i; - int rc; - __u64 mm_size; - __u64 mr_size; - int acflags = IB_ACCESS_LOCAL_WRITE | - IB_ACCESS_REMOTE_WRITE; - - rc = kiblnd_hdev_get_attr(hdev); - if (rc != 0) - return rc; - - 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; - } - - hdev->ibh_mrs[0] = NULL; - hdev->ibh_nmrs = 1; - - 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_hdev_cleanup_mrs(hdev); - return PTR_ERR(mr); - } - - hdev->ibh_mrs[0] = mr; - - goto out; - } - - mr_size = (1ULL << hdev->ibh_mr_shift); - mm_size = (unsigned long)high_memory - PAGE_OFFSET; - - hdev->ibh_nmrs = (int)((mm_size + mr_size - 1) >> hdev->ibh_mr_shift); - - 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); - return -EINVAL; - } - - /* create an array of MRs to cover all memory */ - 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(hdev->ibh_mrs, 0, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs); - - for (i = 0; i < hdev->ibh_nmrs; i++) { - struct ib_phys_buf ipb; - __u64 iova; - - ipb.size = hdev->ibh_mr_size; - ipb.addr = i * mr_size; - iova = ipb.addr; - - 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_hdev_cleanup_mrs(hdev); - return PTR_ERR(mr); - } +#ifdef HAVE_IB_GET_DMA_MR +static int +kiblnd_hdev_setup_mrs(struct kib_hca_dev *hdev) +{ + struct ib_mr *mr; + int acflags = IB_ACCESS_LOCAL_WRITE | + IB_ACCESS_REMOTE_WRITE; - LASSERT (iova == ipb.addr); + 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_hdev_cleanup_mrs(hdev); + return PTR_ERR(mr); + } - hdev->ibh_mrs[i] = mr; - } + hdev->ibh_mrs = mr; -out: - 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; + return 0; } +#endif static int kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event) @@ -2540,7 +2665,7 @@ kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event) } static int -kiblnd_dev_need_failover(kib_dev_t *dev) +kiblnd_dev_need_failover(struct kib_dev *dev) { struct rdma_cm_id *cmid; struct sockaddr_in srcaddr; @@ -2579,33 +2704,29 @@ kiblnd_dev_need_failover(kib_dev_t *dev) 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), + CERROR("Failed to bind %s:%pI4h to device(%p): %d\n", + dev->ibd_ifname, &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; + rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */ + rdma_destroy_id(cmid); + return rc; } int -kiblnd_dev_failover(kib_dev_t *dev) +kiblnd_dev_failover(struct kib_dev *dev) { - CFS_LIST_HEAD (zombie_tpo); - CFS_LIST_HEAD (zombie_ppo); - CFS_LIST_HEAD (zombie_fpo); + 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 kib_hca_dev *hdev = NULL; + struct kib_hca_dev *old; struct ib_pd *pd; - kib_net_t *net; + struct kib_net *net; struct sockaddr_in addr; unsigned long flags; int rc = 0; @@ -2652,8 +2773,8 @@ kiblnd_dev_failover(kib_dev_t *dev) /* 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), + CERROR("Failed to bind %s:%pI4h to device(%p): %d\n", + dev->ibd_ifname, &dev->ibd_ifip, cmid->device, rc); rdma_destroy_id(cmid); goto out; @@ -2672,12 +2793,16 @@ kiblnd_dev_failover(kib_dev_t *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; - } +#ifdef HAVE_IB_ALLOC_PD_2ARGS + pd = ib_alloc_pd(cmid->device, 0); +#else + pd = ib_alloc_pd(cmid->device); +#endif + if (IS_ERR(pd)) { + rc = PTR_ERR(pd); + CERROR("Can't allocate PD: %d\n", rc); + goto out; + } hdev->ibh_pd = pd; @@ -2687,61 +2812,64 @@ kiblnd_dev_failover(kib_dev_t *dev) goto out; } - rc = kiblnd_hdev_setup_mrs(hdev); - if (rc != 0) { - CERROR("Can't setup device: %d\n", rc); - goto out; - } + rc = kiblnd_hdev_get_attr(hdev); + if (rc != 0) { + CERROR("Can't get device attributes: %d\n", rc); + goto out; + } + +#ifdef HAVE_IB_GET_DMA_MR + rc = kiblnd_hdev_setup_mrs(hdev); + if (rc != 0) { + CERROR("Can't setup device: %d\n", rc); + goto out; + } +#endif write_lock_irqsave(&kiblnd_data.kib_global_lock, flags); - old = dev->ibd_hdev; - dev->ibd_hdev = hdev; /* take over the refcount */ - hdev = old; + old = dev->ibd_hdev; + dev->ibd_hdev = hdev; /* take over the refcount */ + hdev = old; - cfs_list_for_each_entry(net, &dev->ibd_nets, ibn_list) { + 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) { + if (net->ibn_fmr_ps != NULL) kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i], &zombie_fpo); - - } else if (net->ibn_pmr_ps != NULL) { - kiblnd_fail_poolset(&net->ibn_pmr_ps[i]-> - pps_poolset, &zombie_ppo); - } } } write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); out: - if (!cfs_list_empty(&zombie_tpo)) - kiblnd_destroy_pool_list(&zombie_tpo); - if (!cfs_list_empty(&zombie_ppo)) - kiblnd_destroy_pool_list(&zombie_ppo); - if (!cfs_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; + 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); - return rc; + if (rc != 0) + dev->ibd_failed_failover++; + else + dev->ibd_failed_failover = 0; + + return rc; } void -kiblnd_destroy_dev (kib_dev_t *dev) +kiblnd_destroy_dev(struct kib_dev *dev) { - LASSERT (dev->ibd_nnets == 0); - LASSERT (cfs_list_empty(&dev->ibd_nets)); + LASSERT(dev->ibd_nnets == 0); + LASSERT(list_empty(&dev->ibd_nets)); - cfs_list_del(&dev->ibd_fail_list); - cfs_list_del(&dev->ibd_list); + list_del(&dev->ibd_fail_list); + list_del(&dev->ibd_list); if (dev->ibd_hdev != NULL) kiblnd_hdev_decref(dev->ibd_hdev); @@ -2749,74 +2877,16 @@ kiblnd_destroy_dev (kib_dev_t *dev) LIBCFS_FREE(dev, sizeof(*dev)); } -kib_dev_t * -kiblnd_create_dev(char *ifname) -{ - struct net_device *netdev; - kib_dev_t *dev; - __u32 netmask; - __u32 ip; - int up; - int rc; - - 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); - } - - CFS_INIT_LIST_HEAD(&dev->ibd_nets); - CFS_INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */ - CFS_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; - } - - cfs_list_add_tail(&dev->ibd_list, - &kiblnd_data.kib_devs); - return dev; -} - -void +static void kiblnd_base_shutdown(void) { 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: @@ -2826,31 +2896,35 @@ 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)); + LASSERT(list_empty(&kiblnd_data.kib_reconn_list)); + LASSERT(list_empty(&kiblnd_data.kib_reconn_wait)); - /* flag threads to terminate; wake and wait for them to die */ - kiblnd_data.kib_shutdown = 1; + /* flag threads to terminate; wake and wait for them to die */ + kiblnd_data.kib_shutdown = 1; /* NB: we really want to stop scheduler threads net by net * instead of the whole module, this should be improved * with dynamic configuration LNet */ cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) - cfs_waitq_broadcast(&sched->ibs_waitq); - - cfs_waitq_broadcast(&kiblnd_data.kib_connd_waitq); - cfs_waitq_broadcast(&kiblnd_data.kib_failover_waitq); - - i = 2; - while (cfs_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)); - cfs_pause(cfs_time_seconds(1)); - } + 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++; + /* power of 2? */ + CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, + "Waiting for %d threads to terminate\n", + atomic_read(&kiblnd_data.kib_nthreads)); + set_current_state(TASK_UNINTERRUPTIBLE); + schedule_timeout(cfs_time_seconds(1)); + } /* fall through */ @@ -2860,7 +2934,7 @@ kiblnd_base_shutdown(void) if (kiblnd_data.kib_peers != NULL) { LIBCFS_FREE(kiblnd_data.kib_peers, - sizeof(cfs_list_t) * + sizeof(struct list_head) * kiblnd_data.kib_peer_hash_size); } @@ -2868,16 +2942,16 @@ kiblnd_base_shutdown(void) 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 -kiblnd_shutdown (lnet_ni_t *ni) +static void +kiblnd_shutdown(struct lnet_ni *ni) { - kib_net_t *net = ni->ni_data; + struct kib_net *net = ni->ni_data; rwlock_t *g_lock = &kiblnd_data.kib_global_lock; int i; unsigned long flags; @@ -2888,7 +2962,7 @@ 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)); write_lock_irqsave(g_lock, flags); net->ibn_shutdown = 1; @@ -2902,29 +2976,31 @@ kiblnd_shutdown (lnet_ni_t *ni) /* nuke all existing peers within this net */ kiblnd_del_peer(ni, LNET_NID_ANY); - /* Wait for all peer state to clean up */ - i = 2; - while (cfs_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)); - cfs_pause(cfs_time_seconds(1)); - } + /* Wait for all peer_ni state to clean up */ + i = 2; + while (atomic_read(&net->ibn_npeers) != 0) { + i++; + /* power of 2? */ + CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, + "%s: waiting for %d peers to disconnect\n", + libcfs_nid2str(ni->ni_nid), + atomic_read(&net->ibn_npeers)); + set_current_state(TASK_UNINTERRUPTIBLE); + schedule_timeout(cfs_time_seconds(1)); + } kiblnd_net_fini_pools(net); write_lock_irqsave(g_lock, flags); LASSERT(net->ibn_dev->ibd_nnets > 0); net->ibn_dev->ibd_nnets--; - cfs_list_del(&net->ibn_list); + 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) @@ -2934,7 +3010,7 @@ 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; @@ -2942,43 +3018,46 @@ 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 +static int kiblnd_base_startup(void) { 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 */ rwlock_init(&kiblnd_data.kib_global_lock); - CFS_INIT_LIST_HEAD(&kiblnd_data.kib_devs); - CFS_INIT_LIST_HEAD(&kiblnd_data.kib_failed_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); - CFS_INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns); - CFS_INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies); - cfs_waitq_init(&kiblnd_data.kib_connd_waitq); - cfs_waitq_init(&kiblnd_data.kib_failover_waitq); + INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns); + INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies); + INIT_LIST_HEAD(&kiblnd_data.kib_reconn_list); + INIT_LIST_HEAD(&kiblnd_data.kib_reconn_wait); + + init_waitqueue_head(&kiblnd_data.kib_connd_waitq); + init_waitqueue_head(&kiblnd_data.kib_failover_waitq); kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(), sizeof(*sched)); @@ -2989,8 +3068,8 @@ kiblnd_base_startup(void) int nthrs; spin_lock_init(&sched->ibs_lock); - CFS_INIT_LIST_HEAD(&sched->ibs_conns); - cfs_waitq_init(&sched->ibs_waitq); + 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) { @@ -3011,14 +3090,15 @@ kiblnd_base_startup(void) kiblnd_data.kib_init = IBLND_INIT_DATA; /*****************************************************/ - rc = kiblnd_thread_start(kiblnd_connd, NULL); + rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd"); if (rc != 0) { CERROR("Can't spawn o2iblnd connd: %d\n", rc); goto failed; } - if (*kiblnd_tunables.kib_dev_failover != 0) - rc = kiblnd_thread_start(kiblnd_failover_thread, NULL); + if (*kiblnd_tunables.kib_dev_failover != 0) + rc = kiblnd_thread_start(kiblnd_failover_thread, NULL, + "kiblnd_failover"); if (rc != 0) { CERROR("Can't spawn o2iblnd failover thread: %d\n", rc); @@ -3036,7 +3116,7 @@ kiblnd_base_startup(void) return -ENETDOWN; } -int +static int kiblnd_start_schedulers(struct kib_sched_info *sched) { int rc = 0; @@ -3060,9 +3140,11 @@ kiblnd_start_schedulers(struct kib_sched_info *sched) for (i = 0; i < nthrs; i++) { long id; - + char name[20]; id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i); - rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id); + 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; @@ -3075,8 +3157,7 @@ kiblnd_start_schedulers(struct kib_sched_info *sched) return rc; } -int -kiblnd_dev_start_threads(kib_dev_t *dev, int newdev, __u32 *cpts, int ncpts) +static int kiblnd_dev_start_threads(struct kib_dev *dev, u32 *cpts, int ncpts) { int cpt; int rc; @@ -3088,7 +3169,7 @@ kiblnd_dev_start_threads(kib_dev_t *dev, int newdev, __u32 *cpts, int ncpts) cpt = (cpts == NULL) ? i : cpts[i]; sched = kiblnd_data.kib_scheds[cpt]; - if (!newdev && sched->ibs_nthreads > 0) + if (sched->ibs_nthreads > 0) continue; rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]); @@ -3101,51 +3182,18 @@ kiblnd_dev_start_threads(kib_dev_t *dev, int newdev, __u32 *cpts, int ncpts) return 0; } -kib_dev_t * -kiblnd_dev_search(char *ifname) -{ - kib_dev_t *alias = NULL; - kib_dev_t *dev; - char *colon; - char *colon2; - - colon = strchr(ifname, ':'); - cfs_list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) { - if (strcmp(&dev->ibd_ifname[0], ifname) == 0) - return dev; - - if (alias != NULL) - continue; - - colon2 = strchr(dev->ibd_ifname, ':'); - if (colon != NULL) - *colon = 0; - if (colon2 != NULL) - *colon2 = 0; - - if (strcmp(&dev->ibd_ifname[0], ifname) == 0) - alias = dev; - - if (colon != NULL) - *colon = ':'; - if (colon2 != NULL) - *colon2 = ':'; - } - return alias; -} - -int -kiblnd_startup (lnet_ni_t *ni) +static int +kiblnd_startup(struct lnet_ni *ni) { char *ifname; - kib_dev_t *ibdev = NULL; - kib_net_t *net; - struct timeval tv; + struct lnet_inetdev *ifaces = NULL; + struct kib_dev *ibdev = NULL; + struct kib_net *net; unsigned long flags; int rc; - int newdev; + int i; - LASSERT (ni->ni_lnd == &the_o2iblnd); + LASSERT (ni->ni_net->net_lnd == &the_o2iblnd); if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) { rc = kiblnd_base_startup(); @@ -3158,54 +3206,81 @@ kiblnd_startup (lnet_ni_t *ni) if (net == NULL) goto failed; - memset(net, 0, sizeof(*net)); - - cfs_gettimeofday(&tv); - net->ibn_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec; - - ni->ni_peertimeout = *kiblnd_tunables.kib_peertimeout; - ni->ni_maxtxcredits = *kiblnd_tunables.kib_credits; - ni->ni_peertxcredits = *kiblnd_tunables.kib_peertxcredits; - ni->ni_peerrtrcredits = *kiblnd_tunables.kib_peerrtrcredits; + net->ibn_incarnation = ktime_get_real_ns() / NSEC_PER_USEC; - if (ni->ni_interfaces[0] != NULL) { - /* Use the IPoIB interface specified in 'networks=' */ + kiblnd_tunables_setup(ni); - CLASSERT (LNET_MAX_INTERFACES > 1); - if (ni->ni_interfaces[1] != NULL) { - CERROR("Multiple interfaces not supported\n"); - goto failed; - } + /* + * ni_interfaces is only to support legacy pre Multi-Rail + * tcp bonding for ksocklnd. Multi-Rail wants each secondary + * IP to be treated as an unique 'struct ni' interfaces instead. + */ + if (ni->ni_interfaces[0] != NULL) { + /* Use the IPoIB interface specified in 'networks=' */ + if (ni->ni_interfaces[1] != NULL) { + CERROR("ko2iblnd: Multiple interfaces not supported\n"); + goto failed; + } - ifname = ni->ni_interfaces[0]; - } else { - ifname = *kiblnd_tunables.kib_default_ipif; - } + ifname = ni->ni_interfaces[0]; + } else { + ifname = *kiblnd_tunables.kib_default_ipif; + } if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) { CERROR("IPoIB interface name too long: %s\n", ifname); goto failed; } - ibdev = kiblnd_dev_search(ifname); + rc = lnet_inet_enumerate(&ifaces); + if (rc < 0) + goto failed; + + for (i = 0; i < rc; i++) { + if (strcmp(ifname, ifaces[i].li_name) == 0) + break; + } - 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 (i == rc) { + CERROR("ko2iblnd: No matching interfaces\n"); + rc = -ENOENT; + goto failed; + } - if (ibdev == NULL) - goto failed; + LIBCFS_ALLOC(ibdev, sizeof(*ibdev)); + if (!ibdev) { + rc = -ENOMEM; + goto failed; + } - net->ibn_dev = ibdev; - ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip); + ibdev->ibd_ifip = ifaces[i].li_ipaddr; + strlcpy(ibdev->ibd_ifname, ifaces[i].li_name, + sizeof(ibdev->ibd_ifname)); + ibdev->ibd_can_failover = !!(ifaces[i].li_flags & IFF_MASTER); - rc = kiblnd_dev_start_threads(ibdev, newdev, - ni->ni_cpts, ni->ni_ncpts); + INIT_LIST_HEAD(&ibdev->ibd_nets); + INIT_LIST_HEAD(&ibdev->ibd_list); /* not yet in kib_devs */ + INIT_LIST_HEAD(&ibdev->ibd_fail_list); + + /* initialize the device */ + rc = kiblnd_dev_failover(ibdev); + if (rc) { + CERROR("ko2iblnd: Can't initialize device: rc = %d\n", rc); + goto failed; + } + + list_add_tail(&ibdev->ibd_list, &kiblnd_data.kib_devs); + + net->ibn_dev = ibdev; + ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip); + + ni->ni_dev_cpt = ifaces[i].li_cpt; + + rc = kiblnd_dev_start_threads(ibdev, ni->ni_cpts, ni->ni_ncpts); if (rc != 0) goto failed; - rc = kiblnd_net_init_pools(net, ni->ni_cpts, ni->ni_ncpts); + rc = kiblnd_net_init_pools(net, ni, ni->ni_cpts, ni->ni_ncpts); if (rc != 0) { CERROR("Failed to initialize NI pools: %d\n", rc); goto failed; @@ -3213,7 +3288,7 @@ kiblnd_startup (lnet_ni_t *ni) write_lock_irqsave(&kiblnd_data.kib_global_lock, flags); ibdev->ibd_nnets++; - cfs_list_add_tail(&net->ibn_list, &ibdev->ibd_nets); + list_add_tail(&net->ibn_list, &ibdev->ibd_nets); write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); net->ibn_init = IBLND_INIT_ALL; @@ -3221,45 +3296,56 @@ kiblnd_startup (lnet_ni_t *ni) return 0; failed: - if (net->ibn_dev == NULL && ibdev != NULL) + if (net != NULL && net->ibn_dev == NULL && ibdev != NULL) kiblnd_destroy_dev(ibdev); + kfree(ifaces); kiblnd_shutdown(ni); CDEBUG(D_NET, "kiblnd_startup failed\n"); return -ENETDOWN; } -void __exit -kiblnd_module_fini (void) +static struct lnet_lnd 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 ko2iblnd_exit(void) { - lnet_unregister_lnd(&the_o2iblnd); - kiblnd_tunables_fini(); + lnet_unregister_lnd(&the_o2iblnd); } -int __init -kiblnd_module_init (void) +static int __init ko2iblnd_init(void) { - int rc; + int rc; - CLASSERT (sizeof(kib_msg_t) <= IBLND_MSG_SIZE); - CLASSERT (offsetof(kib_msg_t, ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) - <= IBLND_MSG_SIZE); - CLASSERT (offsetof(kib_msg_t, ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) - <= IBLND_MSG_SIZE); + CLASSERT(sizeof(struct kib_msg) <= IBLND_MSG_SIZE); + CLASSERT(offsetof(struct kib_msg, + ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) <= + IBLND_MSG_SIZE); + CLASSERT(offsetof(struct kib_msg, + ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) + <= IBLND_MSG_SIZE); - rc = kiblnd_tunables_init(); - if (rc != 0) - return rc; + rc = kiblnd_tunables_init(); + if (rc != 0) + return rc; - lnet_register_lnd(&the_o2iblnd); + lnet_register_lnd(&the_o2iblnd); - return 0; + return 0; } -MODULE_AUTHOR("Sun Microsystems, Inc. "); -MODULE_DESCRIPTION("Kernel OpenIB gen2 LND v2.00"); +MODULE_AUTHOR("OpenSFS, Inc. "); +MODULE_DESCRIPTION("OpenIB gen2 LNet Network Driver"); +MODULE_VERSION("2.8.0"); MODULE_LICENSE("GPL"); -module_init(kiblnd_module_init); -module_exit(kiblnd_module_fini); +module_init(ko2iblnd_init); +module_exit(ko2iblnd_exit);