/*
* Copyright (C) 2012 Cray, Inc.
*
- * Copyright (c) 2013, 2014, Intel Corporation.
+ * Copyright (c) 2013, 2017, Intel Corporation.
*
* Author: Nic Henke <nic@cray.com>
* Author: James Shimek <jshimek@cray.com>
#include "gnilnd.h"
/* Primary entry points from LNET. There are no guarantees against reentrance. */
-lnd_t the_kgnilnd = {
-#ifdef CONFIG_CRAY_XT
+const struct lnet_lnd the_kgnilnd = {
.lnd_type = GNILND,
-#else
- .lnd_type = GNIIPLND,
-#endif
.lnd_startup = kgnilnd_startup,
.lnd_shutdown = kgnilnd_shutdown,
.lnd_ctl = kgnilnd_ctl,
.lnd_send = kgnilnd_send,
.lnd_recv = kgnilnd_recv,
.lnd_eager_recv = kgnilnd_eager_recv,
- .lnd_query = kgnilnd_query,
};
kgn_data_t kgnilnd_data;
int
kgnilnd_close_stale_conns_locked(kgn_peer_t *peer, kgn_conn_t *newconn)
{
- kgn_conn_t *conn;
- struct list_head *ctmp, *cnxt;
+ kgn_conn_t *conn, *cnxt;
int loopback;
int count = 0;
- loopback = peer->gnp_nid == peer->gnp_net->gnn_ni->ni_nid;
-
- list_for_each_safe(ctmp, cnxt, &peer->gnp_conns) {
- conn = list_entry(ctmp, kgn_conn_t, gnc_list);
+ loopback = (peer->gnp_nid ==
+ lnet_nid_to_nid4(&peer->gnp_net->gnn_ni->ni_nid));
+ list_for_each_entry_safe(conn, cnxt, &peer->gnp_conns, gnc_list) {
if (conn->gnc_state != GNILND_CONN_ESTABLISHED)
continue;
newconn->gnc_peer_connstamp == conn->gnc_my_connstamp) {
CDEBUG(D_NET, "skipping prune of %p, "
"loopback and matching stamps"
- " connstamp "LPU64"("LPU64")"
- " peerstamp "LPU64"("LPU64")\n",
+ " connstamp %llu(%llu)"
+ " peerstamp %llu(%llu)\n",
conn, newconn->gnc_my_connstamp,
conn->gnc_peer_connstamp,
newconn->gnc_peer_connstamp,
if (conn->gnc_peerstamp != newconn->gnc_peerstamp) {
LASSERTF(conn->gnc_peerstamp < newconn->gnc_peerstamp,
- "conn 0x%p peerstamp "LPU64" >= "
- "newconn 0x%p peerstamp "LPU64"\n",
+ "conn 0x%p peerstamp %llu >= "
+ "newconn 0x%p peerstamp %llu\n",
conn, conn->gnc_peerstamp,
newconn, newconn->gnc_peerstamp);
CDEBUG(D_NET, "Closing stale conn nid: %s "
- " peerstamp:"LPX64"("LPX64")\n",
+ " peerstamp:%#llx(%#llx)\n",
libcfs_nid2str(peer->gnp_nid),
conn->gnc_peerstamp, newconn->gnc_peerstamp);
} else {
LASSERTF(conn->gnc_peer_connstamp < newconn->gnc_peer_connstamp,
- "conn 0x%p peer_connstamp "LPU64" >= "
- "newconn 0x%p peer_connstamp "LPU64"\n",
+ "conn 0x%p peer_connstamp %llu >= "
+ "newconn 0x%p peer_connstamp %llu\n",
conn, conn->gnc_peer_connstamp,
newconn, newconn->gnc_peer_connstamp);
CDEBUG(D_NET, "Closing stale conn nid: %s"
- " connstamp:"LPU64"("LPU64")\n",
+ " connstamp:%llu(%llu)\n",
libcfs_nid2str(peer->gnp_nid),
conn->gnc_peer_connstamp, newconn->gnc_peer_connstamp);
}
kgnilnd_conn_isdup_locked(kgn_peer_t *peer, kgn_conn_t *newconn)
{
kgn_conn_t *conn;
- struct list_head *tmp;
int loopback;
ENTRY;
- loopback = peer->gnp_nid == peer->gnp_net->gnn_ni->ni_nid;
+ loopback = (peer->gnp_nid ==
+ lnet_nid_to_nid4(&peer->gnp_net->gnn_ni->ni_nid));
- list_for_each(tmp, &peer->gnp_conns) {
- conn = list_entry(tmp, kgn_conn_t, gnc_list);
+ list_for_each_entry(conn, &peer->gnp_conns, gnc_list) {
CDEBUG(D_NET, "checking conn 0x%p for peer %s"
- " lo %d new "LPU64" existing "LPU64
- " new peer "LPU64" existing peer "LPU64
+ " lo %d new %llu existing %llu"
+ " new peer %llu existing peer %llu"
" new dev %p existing dev %p\n",
conn, libcfs_nid2str(peer->gnp_nid),
loopback,
return -ENOMEM;
}
- LIBCFS_ALLOC(conn->gnc_tx_ref_table, GNILND_MAX_MSG_ID * sizeof(void *));
+ conn->gnc_tx_ref_table =
+ kgnilnd_vzalloc(GNILND_MAX_MSG_ID * sizeof(void *));
if (conn->gnc_tx_ref_table == NULL) {
CERROR("Can't allocate conn tx_ref_table\n");
GOTO(failed, rc = -ENOMEM);
INIT_LIST_HEAD(&conn->gnc_schedlist);
INIT_LIST_HEAD(&conn->gnc_fmaq);
INIT_LIST_HEAD(&conn->gnc_mdd_list);
+ INIT_LIST_HEAD(&conn->gnc_delaylist);
spin_lock_init(&conn->gnc_list_lock);
spin_lock_init(&conn->gnc_tx_lock);
conn->gnc_magic = GNILND_CONN_MAGIC;
conn->gnc_next_tx = (int) GNILND_MAX_MSG_ID - 10;
/* if this fails, we have conflicts and MAX_TX is too large */
- CLASSERT(GNILND_MAX_MSG_ID < GNILND_MSGID_CLOSE);
+ BUILD_BUG_ON(GNILND_MAX_MSG_ID >= GNILND_MSGID_CLOSE);
/* get a new unique CQ id for this conn */
write_lock(&kgnilnd_data.kgn_peer_conn_lock);
* check context */
conn->gnc_device = dev;
- conn->gnc_timeout = MAX(*kgnilnd_tunables.kgn_timeout,
- GNILND_MIN_TIMEOUT);
+ conn->gnc_timeout = max(*kgnilnd_tunables.kgn_timeout,
+ GNILND_MIN_TIMEOUT);
kgnilnd_update_reaper_timeout(conn->gnc_timeout);
/* this is the ep_handle for doing SMSG & BTE */
failed:
atomic_dec(&kgnilnd_data.kgn_nconns);
- LIBCFS_FREE(conn->gnc_tx_ref_table, GNILND_MAX_MSG_ID * sizeof(void *));
+ kgnilnd_vfree(conn->gnc_tx_ref_table,
+ GNILND_MAX_MSG_ID * sizeof(void *));
LIBCFS_FREE(conn, sizeof(*conn));
return rc;
}
list_empty(&conn->gnc_hashlist) &&
list_empty(&conn->gnc_schedlist) &&
list_empty(&conn->gnc_mdd_list) &&
+ list_empty(&conn->gnc_delaylist) &&
conn->gnc_magic == GNILND_CONN_MAGIC,
- "conn 0x%p->%s IRQ %d sched %d purg %d ep 0x%p Mg %d lists %d/%d/%d/%d\n",
+ "conn 0x%p->%s IRQ %d sched %d purg %d ep 0x%p Mg %d lists %d/%d/%d/%d/%d\n",
conn, conn->gnc_peer ? libcfs_nid2str(conn->gnc_peer->gnp_nid)
: "<?>",
!!in_interrupt(), conn->gnc_scheduled,
list_empty(&conn->gnc_list),
list_empty(&conn->gnc_hashlist),
list_empty(&conn->gnc_schedlist),
- list_empty(&conn->gnc_mdd_list));
+ list_empty(&conn->gnc_mdd_list),
+ list_empty(&conn->gnc_delaylist));
/* Tripping these is especially bad, as it means we have items on the
* lists that didn't keep their refcount on the connection - or
kgnilnd_peer_decref(conn->gnc_peer);
if (conn->gnc_tx_ref_table != NULL) {
- LIBCFS_FREE(conn->gnc_tx_ref_table,
- GNILND_MAX_MSG_ID * sizeof(void *));
+ kgnilnd_vfree(conn->gnc_tx_ref_table,
+ GNILND_MAX_MSG_ID * sizeof(void *));
}
LIBCFS_FREE(conn, sizeof(*conn));
void
kgnilnd_peer_alive(kgn_peer_t *peer)
{
- set_mb(peer->gnp_last_alive, jiffies);
+ time64_t now = ktime_get_seconds();
+
+ set_mb(peer->gnp_last_alive, now);
}
void
*/
for (i = 0; i < *kgnilnd_tunables.kgn_net_hash_size; i++) {
- list_for_each_entry(net , &kgnilnd_data.kgn_nets[i], gnn_list) {
+ list_for_each_entry(net, &kgnilnd_data.kgn_nets[i], gnn_list) {
/* if gnn_shutdown set for any net shutdown is in progress just return */
if (net->gnn_shutdown) {
up_read(&kgnilnd_data.kgn_net_rw_sem);
return;
}
- LIBCFS_ALLOC(nets, nnets * sizeof(*nets));
+ CFS_ALLOC_PTR_ARRAY(nets, nnets);
if (nets == NULL) {
up_read(&kgnilnd_data.kgn_net_rw_sem);
up_read(&kgnilnd_data.kgn_net_rw_sem);
for (i = 0; i < nnets; i++) {
- lnet_nid_t peer_nid;
+ struct lnet_nid peer_nid;
net = nets[i];
- peer_nid = kgnilnd_lnd2lnetnid(net->gnn_ni->ni_nid,
- peer->gnp_nid);
+ lnet_nid4_to_nid(kgnilnd_lnd2lnetnid(
+ lnet_nid_to_nid4(&net->gnn_ni->ni_nid),
+ peer->gnp_nid),
+ &peer_nid);
- CDEBUG(D_NET, "peer 0x%p->%s last_alive %lu (%lus ago)\n",
+ CDEBUG(D_NET, "peer 0x%p->%s last_alive %lld (%llds ago)\n",
peer, libcfs_nid2str(peer_nid), peer->gnp_last_alive,
- cfs_duration_sec(jiffies - peer->gnp_last_alive));
+ ktime_get_seconds() - peer->gnp_last_alive);
- lnet_notify(net->gnn_ni, peer_nid, alive,
+ lnet_notify(net->gnn_ni, &peer_nid, alive, true,
peer->gnp_last_alive);
kgnilnd_net_decref(net);
}
- LIBCFS_FREE(nets, nnets * sizeof(*nets));
+ CFS_FREE_PTR_ARRAY(nets, nnets);
}
}
/* if we NETERROR, make sure it is rate limited */
if (!kgnilnd_conn_clean_errno(error) &&
- peer->gnp_down == GNILND_RCA_NODE_UP) {
+ peer->gnp_state != GNILND_PEER_DOWN) {
CNETERR("closing conn to %s: error %d\n",
libcfs_nid2str(peer->gnp_nid), error);
} else {
conn->gnc_state = GNILND_CONN_CLOSING;
}
+ if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RDMA_CQ_ERROR)) {
+ msleep_interruptible(MSEC_PER_SEC);
+ }
+
/* leave on peer->gnp_conns to make sure we don't let the reaper
* or others try to unlink this peer until the conn is fully
* processed for closing */
kgnilnd_conn_state2str(conn));
LASSERT(list_empty(&conn->gnc_hashlist));
+ /* We shouldnt be on the delay list, the conn can
+ * get added to this list during a retransmit, and retransmits
+ * only occur within scheduler threads.
+ */
+ LASSERT(list_empty(&conn->gnc_delaylist));
/* we've sent the close, start nuking */
if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SCHEDULE_COMPLETE))
logmsg = (nlive + nrdma + nq_rdma);
if (logmsg) {
- if (conn->gnc_peer->gnp_down == GNILND_RCA_NODE_UP) {
- CNETERR("Closed conn 0x%p->%s (errno %d, peer errno %d): "
- "canceled %d TX, %d/%d RDMA\n",
- conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
- conn->gnc_error, conn->gnc_peer_error,
- nlive, nq_rdma, nrdma);
- } else {
- CDEBUG(D_NET, "Closed conn 0x%p->%s (errno %d,"
- " peer errno %d): canceled %d TX, %d/%d RDMA\n",
- conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
- conn->gnc_error, conn->gnc_peer_error,
- nlive, nq_rdma, nrdma);
- }
+ int level = conn->gnc_peer->gnp_state == GNILND_PEER_UP ?
+ D_NETERROR : D_NET;
+ CDEBUG(level, "Closed conn 0x%p->%s (errno %d,"
+ " peer errno %d): canceled %d TX, %d/%d RDMA\n",
+ conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
+ conn->gnc_error, conn->gnc_peer_error,
+ nlive, nq_rdma, nrdma);
}
kgnilnd_destroy_conn_ep(conn);
/* set timeout vals in conn early so we can use them for the NAK */
/* use max of the requested and our timeout, peer will do the same */
- conn->gnc_timeout = MAX(conn->gnc_timeout, connreq->gncr_timeout);
+ conn->gnc_timeout = max(conn->gnc_timeout, connreq->gncr_timeout);
/* only ep_bind really mucks around with the CQ */
/* only ep bind if we are not connecting to ourself and the dstnid is not a wildcard. this check
/* log this for help in debuggin SMSG buffer re-use */
CDEBUG(D_NET, "conn %p src %s dst %s smsg %p acquired"
- " local cqid %u SMSG %p->%u hndl "LPX64"."LPX64
- " remote cqid %u SMSG %p->%u hndl "LPX64"."LPX64"\n",
+ " local cqid %u SMSG %p->%u hndl %#llx.%#llx"
+ " remote cqid %u SMSG %p->%u hndl %#llx.%#llx\n",
conn, libcfs_nid2str(connreq->gncr_srcnid),
libcfs_nid2str(connreq->gncr_dstnid),
&conn->gnpr_smsg_attr,
return -ENOMEM;
}
peer->gnp_nid = nid;
- peer->gnp_down = node_state;
+ peer->gnp_state = node_state;
/* translate from nid to nic addr & store */
rc = kgnilnd_nid_to_nicaddrs(LNET_NIDADDR(nid), 1, &peer->gnp_host_id);
list_for_each_entry_safe(gmp, gmpN, &conn->gnc_mdd_list,
gmp_list) {
CDEBUG(D_NET,
- "dev %p releasing held mdd "LPX64"."LPX64"\n",
+ "dev %p releasing held mdd %#llx.%#llx\n",
conn->gnc_device, gmp->gmp_map_key.qword1,
gmp->gmp_map_key.qword2);
current_to += *kgnilnd_tunables.kgn_min_reconnect_interval / 2;
}
- current_to = MIN(current_to,
- *kgnilnd_tunables.kgn_max_reconnect_interval);
+ current_to = min(current_to,
+ *kgnilnd_tunables.kgn_max_reconnect_interval);
peer->gnp_reconnect_interval = current_to;
CDEBUG(D_NET, "peer %s can reconnect at %lu interval %lu\n",
lnet_nid_t *id, __u32 *nic_addr,
int *refcount, int *connecting)
{
- struct list_head *ptmp;
kgn_peer_t *peer;
int i;
int rc = -ENOENT;
read_lock(&kgnilnd_data.kgn_peer_conn_lock);
for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++) {
-
- list_for_each(ptmp, &kgnilnd_data.kgn_peers[i]) {
- peer = list_entry(ptmp, kgn_peer_t, gnp_list);
-
+ list_for_each_entry(peer, &kgnilnd_data.kgn_peers[i], gnp_list) {
if (index-- > 0)
continue;
{
LIST_HEAD (souls);
LIST_HEAD (zombies);
- struct list_head *ptmp, *pnxt;
- kgn_peer_t *peer;
+ kgn_peer_t *peer, *pnxt;
int lo;
int hi;
int i;
}
for (i = lo; i <= hi; i++) {
- list_for_each_safe(ptmp, pnxt, &kgnilnd_data.kgn_peers[i]) {
- peer = list_entry(ptmp, kgn_peer_t, gnp_list);
-
+ list_for_each_entry_safe(peer, pnxt, &kgnilnd_data.kgn_peers[i],
+ gnp_list) {
LASSERTF(peer->gnp_net != NULL,
"peer %p (%s) with NULL net\n",
peer, libcfs_nid2str(peer->gnp_nid));
return rc;
}
- i = 4;
- while (atomic_read(&kgnilnd_data.kgn_npending_conns) ||
- atomic_read(&kgnilnd_data.kgn_npending_detach) ||
- atomic_read(&kgnilnd_data.kgn_npending_unlink)) {
-
- set_current_state(TASK_UNINTERRUPTIBLE);
- schedule_timeout(cfs_time_seconds(1));
- i++;
-
- CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, "Waiting on %d peers %d closes %d detaches\n",
+ wait_var_event_warning(&kgnilnd_data,
+ !atomic_read(&kgnilnd_data.kgn_npending_conns) &&
+ !atomic_read(&kgnilnd_data.kgn_npending_detach) &&
+ !atomic_read(&kgnilnd_data.kgn_npending_unlink),
+ "Waiting on %d peers %d closes %d detaches\n",
atomic_read(&kgnilnd_data.kgn_npending_unlink),
atomic_read(&kgnilnd_data.kgn_npending_conns),
atomic_read(&kgnilnd_data.kgn_npending_detach));
- }
return rc;
}
kgnilnd_get_conn_by_idx(int index)
{
kgn_peer_t *peer;
- struct list_head *ptmp;
kgn_conn_t *conn;
- struct list_head *ctmp;
int i;
for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++) {
read_lock(&kgnilnd_data.kgn_peer_conn_lock);
- list_for_each(ptmp, &kgnilnd_data.kgn_peers[i]) {
-
- peer = list_entry(ptmp, kgn_peer_t, gnp_list);
-
- list_for_each(ctmp, &peer->gnp_conns) {
- conn = list_entry(ctmp, kgn_conn_t, gnc_list);
-
+ list_for_each_entry(peer, &kgnilnd_data.kgn_peers[i], gnp_list) {
+ list_for_each_entry(conn, &peer->gnp_conns, gnc_list) {
if (conn->gnc_state != GNILND_CONN_ESTABLISHED)
continue;
write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
/* Don't add a peer for node up events */
- if (down == GNILND_RCA_NODE_UP) {
+ if (down == GNILND_PEER_UP)
return 0;
- }
/* find any valid net - we don't care which one... */
down_read(&kgnilnd_data.kgn_net_rw_sem);
/* The nid passed in does not yet contain the net portion.
* Let's build it up now
*/
- nid = LNET_MKNID(LNET_NIDNET(net->gnn_ni->ni_nid), nid);
+ nid = LNET_MKNID(LNET_NID_NET(&net->gnn_ni->ni_nid), nid);
rc = kgnilnd_add_peer(net, nid, &new_peer);
if (rc) {
}
}
- peer->gnp_down = down;
+ peer->gnp_state = down;
- if (down == GNILND_RCA_NODE_DOWN) {
+ if (down == GNILND_PEER_DOWN) {
kgn_conn_t *conn;
peer->gnp_down_event_time = jiffies;
write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
- if (down == GNILND_RCA_NODE_DOWN) {
+ if (down == GNILND_PEER_DOWN) {
/* using ENETRESET so we don't get messages from
* kgnilnd_tx_done
*/
kgnilnd_txlist_done(&zombies, -ENETRESET);
kgnilnd_peer_notify(peer, -ECONNRESET, 0);
- LCONSOLE_INFO("Recieved down event for nid %lld\n", nid);
+ LCONSOLE_INFO("Received down event for nid %d\n",
+ LNET_NIDADDR(nid));
}
return 0;
}
int
-kgnilnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
+kgnilnd_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg)
{
struct libcfs_ioctl_data *data = arg;
kgn_net_t *net = ni->ni_data;
* LNET assumes a conn and peer per net, the LNET_MKNID/LNET_NIDADDR allows us to let Lnet see what it
* wants to see instead of the underlying network that is being used to send the data
*/
- data->ioc_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), LNET_NIDADDR(nid));
+ data->ioc_nid = LNET_MKNID(LNET_NID_NET(&ni->ni_nid),
+ LNET_NIDADDR(nid));
data->ioc_flags = peer_connecting;
data->ioc_count = peer_refcount;
/* LNET_MKNID is used to build the correct address based on what LNET wants to see instead of
* the generic connection that is used to send the data
*/
- data->ioc_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), LNET_NIDADDR(conn->gnc_peer->gnp_nid));
+ data->ioc_nid = LNET_MKNID(LNET_NID_NET(&ni->ni_nid),
+ LNET_NIDADDR(conn->gnc_peer->gnp_nid));
data->ioc_u32[0] = conn->gnc_device->gnd_id;
kgnilnd_conn_decref(conn);
}
}
case IOC_LIBCFS_REGISTER_MYNID: {
/* Ignore if this is a noop */
- if (data->ioc_nid == ni->ni_nid) {
+ if (data->ioc_nid == lnet_nid_to_nid4(&ni->ni_nid)) {
rc = 0;
} else {
CERROR("obsolete IOC_LIBCFS_REGISTER_MYNID: %s(%s)\n",
libcfs_nid2str(data->ioc_nid),
- libcfs_nid2str(ni->ni_nid));
+ libcfs_nidstr(&ni->ni_nid));
rc = -EINVAL;
}
break;
return rc;
}
-void
-kgnilnd_query(lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when)
-{
- kgn_net_t *net = ni->ni_data;
- kgn_tx_t *tx;
- kgn_peer_t *peer = NULL;
- kgn_conn_t *conn = NULL;
- lnet_process_id_t id = {
- .nid = nid,
- .pid = LNET_PID_LUSTRE,
- };
- ENTRY;
-
- /* I expect to find him, so only take a read lock */
- read_lock(&kgnilnd_data.kgn_peer_conn_lock);
- peer = kgnilnd_find_peer_locked(nid);
- if (peer != NULL) {
- /* LIE if in a quiesce - we will update the timeouts after,
- * but we don't want sends failing during it */
- if (kgnilnd_data.kgn_quiesce_trigger) {
- *when = jiffies;
- read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
- GOTO(out, 0);
- }
-
- /* Update to best guess, might refine on later checks */
- *when = peer->gnp_last_alive;
-
- /* we have a peer, how about a conn? */
- conn = kgnilnd_find_conn_locked(peer);
-
- if (conn == NULL) {
- /* if there is no conn, check peer last errno to see if clean disconnect
- * - if it was, we lie to LNet because we believe a TX would complete
- * on reconnect */
- if (kgnilnd_conn_clean_errno(peer->gnp_last_errno)) {
- *when = jiffies;
- }
- /* we still want to fire a TX and new conn in this case */
- } else {
- /* gnp_last_alive is valid, run for the hills */
- read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
- GOTO(out, 0);
- }
- }
- /* if we get here, either we have no peer or no conn for him, so fire off
- * new TX to trigger conn setup */
- read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
-
- /* if we couldn't find him, we'll fire up a TX and get connected -
- * if we don't do this, after ni_peer_timeout, LNet will declare him dead.
- * So really we treat kgnilnd_query as a bit of a 'connect now' type
- * event because it'll only do this when it wants to send
- *
- * Use a real TX for this to get the proper gnp_tx_queue behavior, etc
- * normally we'd use kgnilnd_send_ctlmsg for this, but we don't really
- * care that this goes out quickly since we already know we need a new conn
- * formed */
- if (CFS_FAIL_CHECK(CFS_FAIL_GNI_NOOP_SEND))
- return;
-
- tx = kgnilnd_new_tx_msg(GNILND_MSG_NOOP, ni->ni_nid);
- if (tx != NULL) {
- kgnilnd_launch_tx(tx, net, &id);
- }
-out:
- CDEBUG(D_NETTRACE, "peer 0x%p->%s when %lu\n", peer,
- libcfs_nid2str(nid), *when);
- EXIT;
-}
-
int
kgnilnd_dev_init(kgn_device_t *dev)
{
cq_size = *kgnilnd_tunables.kgn_credits * 2 * 3;
rrc = kgnilnd_cdm_create(dev->gnd_id, *kgnilnd_tunables.kgn_ptag,
- GNI_JOB_CREATE_COOKIE(GNI_PKEY_LND, 0), 0,
+ *kgnilnd_tunables.kgn_pkey, 0,
&dev->gnd_domain);
if (rrc != GNI_RC_SUCCESS) {
CERROR("Can't create CDM %d (%d)\n", dev->gnd_id, rrc);
}
}
- rrc = sock_create_kern(PF_INET, SOCK_DGRAM, IPPROTO_IP, &kgnilnd_data.kgn_sock);
- if (rrc < 0) {
- CERROR("sock_create returned %d\n", rrc);
- GOTO(failed, rrc);
- }
-
rc = kgnilnd_nicaddr_to_nid(dev->gnd_host_id, &dev->gnd_nid);
if (rc < 0) {
/* log messages during startup */
GOTO(failed, rc = -EINVAL);
}
+ rrc = kgnilnd_register_smdd_buf(dev);
+ if (rrc != GNI_RC_SUCCESS) {
+ GOTO(failed, rc = -EINVAL);
+ }
+
RETURN(0);
failed:
/* At quiesce or rest time, need to loop through and clear gnd_ready_conns ?*/
LASSERTF(list_empty(&dev->gnd_ready_conns) &&
list_empty(&dev->gnd_map_tx) &&
- list_empty(&dev->gnd_rdmaq),
- "dev 0x%p ready_conns %d@0x%p map_tx %d@0x%p rdmaq %d@0x%p\n",
+ list_empty(&dev->gnd_rdmaq) &&
+ list_empty(&dev->gnd_delay_conns),
+ "dev 0x%p ready_conns %d@0x%p delay_conns %d@0x%p"
+ "map_tx %d@0x%p rdmaq %d@0x%p\n",
dev, kgnilnd_count_list(&dev->gnd_ready_conns), &dev->gnd_ready_conns,
+ kgnilnd_count_list(&dev->gnd_delay_conns), &dev->gnd_delay_conns,
kgnilnd_count_list(&dev->gnd_map_tx), &dev->gnd_map_tx,
kgnilnd_count_list(&dev->gnd_rdmaq), &dev->gnd_rdmaq);
"%d physical mappings of %d pages still mapped\n",
dev->gnd_map_nphys, dev->gnd_map_physnop);
- LASSERTF(dev->gnd_map_nvirt == 0 && dev->gnd_map_virtnob == 0,
- "%d virtual mappings of "LPU64" bytes still mapped\n",
- dev->gnd_map_nvirt, dev->gnd_map_virtnob);
-
LASSERTF(atomic_read(&dev->gnd_n_mdd) == 0 &&
atomic_read(&dev->gnd_n_mdd_held) == 0 &&
atomic64_read(&dev->gnd_nbytes_map) == 0,
- "%d SMSG mappings of %ld bytes still mapped or held %d\n",
+ "%d SMSG mappings of %lld bytes still mapped or held %d\n",
atomic_read(&dev->gnd_n_mdd),
- atomic64_read(&dev->gnd_nbytes_map), atomic_read(&dev->gnd_n_mdd_held));
+ (u64)atomic64_read(&dev->gnd_nbytes_map),
+ atomic_read(&dev->gnd_n_mdd_held));
LASSERT(list_empty(&dev->gnd_map_list));
atomic_read(&kgnilnd_data.kgn_nthreads) == 0,
"tried to shutdown with threads active\n");
+ if (dev->gnd_smdd_hold_buf) {
+ rrc = kgnilnd_deregister_smdd_buf(dev);
+ LASSERTF(rrc == GNI_RC_SUCCESS,
+ "bad rc from deregistion of sMDD buffer: %d\n", rrc);
+ dev->gnd_smdd_hold_buf = NULL;
+ }
+
if (dev->gnd_rcv_fma_cqh) {
rrc = kgnilnd_cq_destroy(dev->gnd_rcv_fma_cqh);
LASSERTF(rrc == GNI_RC_SUCCESS,
dev->gnd_domain = NULL;
}
- if (kgnilnd_data.kgn_sock)
- sock_release(kgnilnd_data.kgn_sock);
-
EXIT;
}
int kgnilnd_base_startup(void)
{
- struct timeval tv;
- int pkmem = atomic_read(&libcfs_kmemory);
+ long long pkmem = libcfs_kmem_read();
int rc;
int i;
kgn_device_t *dev;
* initialised with seconds + microseconds at startup time. So we
* rely on NOT creating connections more frequently on average than
* 1MHz to ensure we don't use old connstamps when we reboot. */
- do_gettimeofday(&tv);
kgnilnd_data.kgn_connstamp =
kgnilnd_data.kgn_peerstamp =
- (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
+ ktime_get_seconds();
init_rwsem(&kgnilnd_data.kgn_net_rw_sem);
dev->gnd_id = i;
INIT_LIST_HEAD(&dev->gnd_ready_conns);
+ INIT_LIST_HEAD(&dev->gnd_delay_conns);
INIT_LIST_HEAD(&dev->gnd_map_tx);
INIT_LIST_HEAD(&dev->gnd_fma_buffs);
mutex_init(&dev->gnd_cq_mutex);
init_rwsem(&dev->gnd_conn_sem);
/* alloc & setup nid based dgram table */
- LIBCFS_ALLOC(dev->gnd_dgrams,
- sizeof(struct list_head) * *kgnilnd_tunables.kgn_peer_hash_size);
+ CFS_ALLOC_PTR_ARRAY(dev->gnd_dgrams,
+ *kgnilnd_tunables.kgn_peer_hash_size);
if (dev->gnd_dgrams == NULL)
GOTO(failed, rc = -ENOMEM);
atomic_set(&dev->gnd_ndgrams, 0);
atomic_set(&dev->gnd_nwcdgrams, 0);
/* setup timer for RDMAQ processing */
- setup_timer(&dev->gnd_rdmaq_timer, kgnilnd_schedule_device_timer,
- (unsigned long)dev);
+ cfs_timer_setup(&dev->gnd_rdmaq_timer,
+ kgnilnd_schedule_device_timer,
+ (unsigned long)dev, 0);
/* setup timer for mapping processing */
- setup_timer(&dev->gnd_map_timer, kgnilnd_schedule_device_timer,
- (unsigned long)dev);
+ cfs_timer_setup(&dev->gnd_map_timer,
+ kgnilnd_schedule_device_timer,
+ (unsigned long)dev, 0);
}
/* OK to call kgnilnd_api_shutdown() to cleanup now */
kgnilnd_data.kgn_init = GNILND_INIT_DATA;
- try_module_get(THIS_MODULE);
+ if (!try_module_get(THIS_MODULE))
+ GOTO(failed, rc = -ENOENT);
rwlock_init(&kgnilnd_data.kgn_peer_conn_lock);
- LIBCFS_ALLOC(kgnilnd_data.kgn_peers,
- sizeof(struct list_head) * *kgnilnd_tunables.kgn_peer_hash_size);
+ CFS_ALLOC_PTR_ARRAY(kgnilnd_data.kgn_peers,
+ *kgnilnd_tunables.kgn_peer_hash_size);
if (kgnilnd_data.kgn_peers == NULL)
GOTO(failed, rc = -ENOMEM);
INIT_LIST_HEAD(&kgnilnd_data.kgn_peers[i]);
}
- LIBCFS_ALLOC(kgnilnd_data.kgn_conns,
- sizeof(struct list_head) * *kgnilnd_tunables.kgn_peer_hash_size);
+ CFS_ALLOC_PTR_ARRAY(kgnilnd_data.kgn_conns,
+ *kgnilnd_tunables.kgn_peer_hash_size);
if (kgnilnd_data.kgn_conns == NULL)
GOTO(failed, rc = -ENOMEM);
INIT_LIST_HEAD(&kgnilnd_data.kgn_conns[i]);
}
- LIBCFS_ALLOC(kgnilnd_data.kgn_nets,
- sizeof(struct list_head) * *kgnilnd_tunables.kgn_net_hash_size);
+ CFS_ALLOC_PTR_ARRAY(kgnilnd_data.kgn_nets,
+ *kgnilnd_tunables.kgn_net_hash_size);
if (kgnilnd_data.kgn_nets == NULL)
GOTO(failed, rc = -ENOMEM);
}
kgnilnd_data.kgn_mbox_cache =
- kmem_cache_create("kgn_mbox_block", KMALLOC_MAX_SIZE, 0,
+ kmem_cache_create("kgn_mbox_block", GNILND_MBOX_SIZE, 0,
SLAB_HWCACHE_ALIGN, NULL);
if (kgnilnd_data.kgn_mbox_cache == NULL) {
CERROR("Can't create slab for physical mbox blocks\n");
kgnilnd_data.kgn_tx_phys_cache =
kmem_cache_create("kgn_tx_phys",
- LNET_MAX_IOV * sizeof(gni_mem_segment_t),
+ GNILND_MAX_IOV * sizeof(gni_mem_segment_t),
0, 0, NULL);
if (kgnilnd_data.kgn_tx_phys_cache == NULL) {
CERROR("Can't create slab for kgn_tx_phys\n");
kgnilnd_data.kgn_cksum_npages * sizeof (struct page *));
for (i = 0; i < kgnilnd_data.kgn_cksum_npages; i++) {
- kgnilnd_data.kgn_cksum_map_pages[i] = kmalloc(LNET_MAX_IOV * sizeof (struct page *),
+ kgnilnd_data.kgn_cksum_map_pages[i] = kmalloc(GNILND_MAX_IOV * sizeof (struct page *),
GFP_KERNEL);
if (kgnilnd_data.kgn_cksum_map_pages[i] == NULL) {
CERROR("Can't allocate vmap cksum pages for cpu %d\n", i);
kgnilnd_data.kgn_init = GNILND_INIT_ALL;
/*****************************************************/
- CDEBUG(D_MALLOC, "initial kmem %d\n", pkmem);
+ CDEBUG(D_MALLOC, "initial kmem %lld\n", pkmem);
RETURN(0);
failed:
CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
"Waiting for conns to be cleaned up %d\n",atomic_read(&kgnilnd_data.kgn_nconns));
- set_current_state(TASK_UNINTERRUPTIBLE);
- schedule_timeout(cfs_time_seconds(1));
+ schedule_timeout_uninterruptible(cfs_time_seconds(1));
}
/* Peer state all cleaned up BEFORE setting shutdown, so threads don't
* have to worry about shutdown races. NB connections may be created
i++;
CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
"Waiting for ruhroh thread to terminate\n");
- set_current_state(TASK_UNINTERRUPTIBLE);
- schedule_timeout(cfs_time_seconds(1));
+ schedule_timeout_uninterruptible(cfs_time_seconds(1));
}
/* Flag threads to terminate */
kgnilnd_unmap_fma_blocks(dev);
kgnilnd_schedule_device(dev);
- wake_up_all(&dev->gnd_dgram_waitq);
- wake_up_all(&dev->gnd_dgping_waitq);
+ wake_up(&dev->gnd_dgram_waitq);
+ wake_up(&dev->gnd_dgping_waitq);
LASSERT(list_empty(&dev->gnd_connd_peers));
}
spin_lock(&kgnilnd_data.kgn_reaper_lock);
- wake_up_all(&kgnilnd_data.kgn_reaper_waitq);
+ wake_up(&kgnilnd_data.kgn_reaper_waitq);
spin_unlock(&kgnilnd_data.kgn_reaper_lock);
if (atomic_read(&kgnilnd_data.kgn_nthreads))
CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
"Waiting for %d threads to terminate\n",
atomic_read(&kgnilnd_data.kgn_nthreads));
- set_current_state(TASK_UNINTERRUPTIBLE);
- schedule_timeout(cfs_time_seconds(1));
+ schedule_timeout_uninterruptible(cfs_time_seconds(1));
}
LASSERTF(atomic_read(&kgnilnd_data.kgn_npeers) == 0,
for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++)
LASSERT(list_empty(&kgnilnd_data.kgn_peers[i]));
- LIBCFS_FREE(kgnilnd_data.kgn_peers,
- sizeof (struct list_head) *
- *kgnilnd_tunables.kgn_peer_hash_size);
+ CFS_FREE_PTR_ARRAY(kgnilnd_data.kgn_peers,
+ *kgnilnd_tunables.kgn_peer_hash_size);
}
down_write(&kgnilnd_data.kgn_net_rw_sem);
for (i = 0; i < *kgnilnd_tunables.kgn_net_hash_size; i++)
LASSERT(list_empty(&kgnilnd_data.kgn_nets[i]));
- LIBCFS_FREE(kgnilnd_data.kgn_nets,
- sizeof (struct list_head) *
- *kgnilnd_tunables.kgn_net_hash_size);
+ CFS_FREE_PTR_ARRAY(kgnilnd_data.kgn_nets,
+ *kgnilnd_tunables.kgn_net_hash_size);
}
up_write(&kgnilnd_data.kgn_net_rw_sem);
for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++)
LASSERT(list_empty(&kgnilnd_data.kgn_conns[i]));
- LIBCFS_FREE(kgnilnd_data.kgn_conns,
- sizeof (struct list_head) *
- *kgnilnd_tunables.kgn_peer_hash_size);
+ CFS_FREE_PTR_ARRAY(kgnilnd_data.kgn_conns,
+ *kgnilnd_tunables.kgn_peer_hash_size);
}
for (i = 0; i < kgnilnd_data.kgn_ndevs; i++) {
"dgrams left %d\n", atomic_read(&dev->gnd_ndgrams));
if (dev->gnd_dgrams != NULL) {
- for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++)
+ for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size;
+ i++)
LASSERT(list_empty(&dev->gnd_dgrams[i]));
- LIBCFS_FREE(dev->gnd_dgrams,
- sizeof (struct list_head) *
- *kgnilnd_tunables.kgn_peer_hash_size);
+ CFS_FREE_PTR_ARRAY(dev->gnd_dgrams,
+ *kgnilnd_tunables.kgn_peer_hash_size);
}
kgnilnd_free_phys_fmablk(dev);
kfree(kgnilnd_data.kgn_cksum_map_pages);
}
- CDEBUG(D_MALLOC, "after NAL cleanup: kmem %d\n",
- atomic_read(&libcfs_kmemory));
+ CDEBUG(D_MALLOC, "after NAL cleanup: kmem %lld\n",
+ libcfs_kmem_read());
kgnilnd_data.kgn_init = GNILND_INIT_NOTHING;
module_put(THIS_MODULE);
}
int
-kgnilnd_startup(lnet_ni_t *ni)
+kgnilnd_startup(struct lnet_ni *ni)
{
int rc, devno;
kgn_net_t *net;
ENTRY;
- LASSERTF(ni->ni_lnd == &the_kgnilnd,
+ LASSERTF(ni->ni_net->net_lnd == &the_kgnilnd,
"bad LND 0x%p != the_kgnilnd @ 0x%p\n",
- ni->ni_lnd, &the_kgnilnd);
+ ni->ni_net->net_lnd, &the_kgnilnd);
if (kgnilnd_data.kgn_init == GNILND_INIT_NOTHING) {
rc = kgnilnd_base_startup();
INIT_LIST_HEAD(&net->gnn_list);
ni->ni_data = net;
net->gnn_ni = ni;
- ni->ni_maxtxcredits = *kgnilnd_tunables.kgn_credits;
- ni->ni_peertxcredits = *kgnilnd_tunables.kgn_peer_credits;
+
+ kgnilnd_tunables_setup(ni);
+
+ if (!ni->ni_interface) {
+ rc = lnet_ni_add_interface(ni, "ipogif0");
+ if (rc < 0)
+ CWARN("gnilnd failed to allocate ni_interface\n");
+ }
if (*kgnilnd_tunables.kgn_peer_health) {
int fudge;
fudge = (GNILND_TO2KA(*kgnilnd_tunables.kgn_timeout) / GNILND_REAPER_NCHECKS);
timeout = *kgnilnd_tunables.kgn_timeout + fudge;
- if (*kgnilnd_tunables.kgn_peer_timeout >= timeout)
- ni->ni_peertimeout = *kgnilnd_tunables.kgn_peer_timeout;
- else if (*kgnilnd_tunables.kgn_peer_timeout > -1) {
+ if (*kgnilnd_tunables.kgn_peer_timeout >= timeout) {
+ ni->ni_net->net_tunables.lct_peer_timeout =
+ *kgnilnd_tunables.kgn_peer_timeout;
+ } else if (*kgnilnd_tunables.kgn_peer_timeout > -1) {
LCONSOLE_ERROR("Peer_timeout is set to %d but needs to be >= %d\n",
*kgnilnd_tunables.kgn_peer_timeout,
timeout);
LIBCFS_FREE(net, sizeof(*net));
GOTO(failed, rc = -EINVAL);
} else
- ni->ni_peertimeout = timeout;
+ ni->ni_net->net_tunables.lct_peer_timeout = timeout;
LCONSOLE_INFO("Enabling LNet peer health for gnilnd, timeout %ds\n",
- ni->ni_peertimeout);
+ ni->ni_net->net_tunables.lct_peer_timeout);
}
atomic_set(&net->gnn_refcount, 1);
/* if we have multiple devices, spread the nets around */
- net->gnn_netnum = LNET_NETNUM(LNET_NIDNET(ni->ni_nid));
+ net->gnn_netnum = LNET_NETNUM(LNET_NID_NET(&ni->ni_nid));
- devno = LNET_NIDNET(ni->ni_nid) % GNILND_MAXDEVS;
+ devno = LNET_NID_NET(&ni->ni_nid) % GNILND_MAXDEVS;
net->gnn_dev = &kgnilnd_data.kgn_devices[devno];
/* allocate a 'dummy' cdm for datagram use. We can only have a single
/* the instance id for the cdm is the NETNUM offset by MAXDEVS -
* ensuring we'll have a unique id */
-
- ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), net->gnn_dev->gnd_nid);
+ ni->ni_nid.nid_addr[0] =
+ cpu_to_be32(LNET_NIDADDR(net->gnn_dev->gnd_nid));
CDEBUG(D_NET, "adding net %p nid=%s on dev %d \n",
- net, libcfs_nid2str(ni->ni_nid), net->gnn_dev->gnd_id);
+ net, libcfs_nidstr(&ni->ni_nid), net->gnn_dev->gnd_id);
/* until the gnn_list is set, we need to cleanup ourselves as
* kgnilnd_shutdown is just gonna get confused */
}
void
-kgnilnd_shutdown(lnet_ni_t *ni)
+kgnilnd_shutdown(struct lnet_ni *ni)
{
kgn_net_t *net = ni->ni_data;
int i;
/* Serialize with startup. */
mutex_lock(&kgnilnd_data.kgn_quiesce_mutex);
- CDEBUG(D_MALLOC, "before NAL cleanup: kmem %d\n",
- atomic_read(&libcfs_kmemory));
+ CDEBUG(D_MALLOC, "before NAL cleanup: kmem %lld\n",
+ libcfs_kmem_read());
if (net == NULL) {
CERROR("got NULL net for ni %p\n", ni);
"Waiting for %d references to clear on net %d\n",
atomic_read(&net->gnn_refcount),
net->gnn_netnum);
- set_current_state(TASK_UNINTERRUPTIBLE);
- schedule_timeout(cfs_time_seconds(1));
+ schedule_timeout_uninterruptible(cfs_time_seconds(1));
}
/* release ref from kgnilnd_startup */
kgnilnd_base_shutdown();
}
}
- CDEBUG(D_MALLOC, "after NAL cleanup: kmem %d\n",
- atomic_read(&libcfs_kmemory));
+ CDEBUG(D_MALLOC, "after NAL cleanup: kmem %lld\n",
+ libcfs_kmem_read());
mutex_unlock(&kgnilnd_data.kgn_quiesce_mutex);
EXIT;
}
-void __exit
-kgnilnd_module_fini(void)
+static void __exit kgnilnd_exit(void)
{
lnet_unregister_lnd(&the_kgnilnd);
kgnilnd_proc_fini();
kgnilnd_remove_sysctl();
- kgnilnd_tunables_fini();
}
-int __init
-kgnilnd_module_init(void)
+static int __init kgnilnd_init(void)
{
int rc;
if (rc != 0)
return rc;
- printk(KERN_INFO "Lustre: kgnilnd build version: "KGNILND_BUILD_REV"\n");
+ LCONSOLE_INFO("Lustre: kgnilnd build version: "LUSTRE_VERSION_STRING"\n");
kgnilnd_insert_sysctl();
kgnilnd_proc_init();
}
MODULE_AUTHOR("Cray, Inc. <nic@cray.com>");
-MODULE_DESCRIPTION("Kernel Gemini LND v"KGNILND_BUILD_REV);
+MODULE_DESCRIPTION("Gemini LNet Network Driver");
+MODULE_VERSION(LUSTRE_VERSION_STRING);
MODULE_LICENSE("GPL");
-module_init(kgnilnd_module_init);
-module_exit(kgnilnd_module_fini);
+module_init(kgnilnd_init);
+module_exit(kgnilnd_exit);
git://git.whamcloud.com / fs / lustre-release.git / blobdiff