* GPL HEADER END
*/
/*
- * Copyright 2008 Sun Microsystems, Inc. All rights reserved
+ * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
* Use is subject to license terms.
*/
/*
peer->ibp_nid = nid;
peer->ibp_error = 0;
peer->ibp_last_alive = 0;
- atomic_set(&peer->ibp_refcount, 1); /* 1 ref for caller */
+ cfs_atomic_set(&peer->ibp_refcount, 1); /* 1 ref for caller */
- INIT_LIST_HEAD(&peer->ibp_list); /* not in the peer table yet */
- INIT_LIST_HEAD(&peer->ibp_conns);
- INIT_LIST_HEAD(&peer->ibp_tx_queue);
+ CFS_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);
- write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
+ cfs_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);
/* npeers only grows with the global lock held */
- atomic_inc(&net->ibn_npeers);
+ cfs_atomic_inc(&net->ibn_npeers);
- write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
+ cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
*peerp = peer;
return 0;
kib_net_t *net = peer->ibp_ni->ni_data;
LASSERT (net != NULL);
- LASSERT (atomic_read(&peer->ibp_refcount) == 0);
+ LASSERT (cfs_atomic_read(&peer->ibp_refcount) == 0);
LASSERT (!kiblnd_peer_active(peer));
LASSERT (peer->ibp_connecting == 0);
LASSERT (peer->ibp_accepting == 0);
- LASSERT (list_empty(&peer->ibp_conns));
- LASSERT (list_empty(&peer->ibp_tx_queue));
+ LASSERT (cfs_list_empty(&peer->ibp_conns));
+ LASSERT (cfs_list_empty(&peer->ibp_tx_queue));
LIBCFS_FREE(peer, sizeof(*peer));
* they are destroyed, so we can be assured that _all_ state to do
* with this peer has been cleaned up when its refcount drops to
* zero. */
- atomic_dec(&net->ibn_npeers);
+ cfs_atomic_dec(&net->ibn_npeers);
}
kib_peer_t *
{
/* the caller is responsible for accounting the additional reference
* that this creates */
- struct list_head *peer_list = kiblnd_nid2peerlist(nid);
- struct list_head *tmp;
+ cfs_list_t *peer_list = kiblnd_nid2peerlist(nid);
+ cfs_list_t *tmp;
kib_peer_t *peer;
- list_for_each (tmp, peer_list) {
+ cfs_list_for_each (tmp, peer_list) {
- peer = list_entry(tmp, kib_peer_t, ibp_list);
+ peer = cfs_list_entry(tmp, kib_peer_t, ibp_list);
LASSERT (peer->ibp_connecting > 0 || /* creating conns */
peer->ibp_accepting > 0 ||
- !list_empty(&peer->ibp_conns)); /* active conn */
+ !cfs_list_empty(&peer->ibp_conns)); /* active conn */
if (peer->ibp_nid != nid)
continue;
CDEBUG(D_NET, "got peer [%p] -> %s (%d) version: %x\n",
peer, libcfs_nid2str(nid),
- atomic_read(&peer->ibp_refcount),
+ cfs_atomic_read(&peer->ibp_refcount),
peer->ibp_version);
return peer;
}
void
kiblnd_unlink_peer_locked (kib_peer_t *peer)
{
- LASSERT (list_empty(&peer->ibp_conns));
+ LASSERT (cfs_list_empty(&peer->ibp_conns));
LASSERT (kiblnd_peer_active(peer));
- list_del_init(&peer->ibp_list);
+ cfs_list_del_init(&peer->ibp_list);
/* lose peerlist's ref */
kiblnd_peer_decref(peer);
}
kiblnd_get_peer_info (lnet_ni_t *ni, int index,
lnet_nid_t *nidp, int *count)
{
- kib_peer_t *peer;
- struct list_head *ptmp;
- int i;
- unsigned long flags;
+ kib_peer_t *peer;
+ cfs_list_t *ptmp;
+ int i;
+ unsigned long flags;
- read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
+ cfs_read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
- list_for_each (ptmp, &kiblnd_data.kib_peers[i]) {
+ cfs_list_for_each (ptmp, &kiblnd_data.kib_peers[i]) {
- peer = list_entry(ptmp, kib_peer_t, ibp_list);
+ peer = cfs_list_entry(ptmp, kib_peer_t, ibp_list);
LASSERT (peer->ibp_connecting > 0 ||
peer->ibp_accepting > 0 ||
- !list_empty(&peer->ibp_conns));
+ !cfs_list_empty(&peer->ibp_conns));
if (peer->ibp_ni != ni)
continue;
continue;
*nidp = peer->ibp_nid;
- *count = atomic_read(&peer->ibp_refcount);
+ *count = cfs_atomic_read(&peer->ibp_refcount);
- read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
- flags);
+ cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
+ flags);
return 0;
}
}
- read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
+ cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
return -ENOENT;
}
void
kiblnd_del_peer_locked (kib_peer_t *peer)
{
- struct list_head *ctmp;
- struct list_head *cnxt;
- kib_conn_t *conn;
+ cfs_list_t *ctmp;
+ cfs_list_t *cnxt;
+ kib_conn_t *conn;
- if (list_empty(&peer->ibp_conns)) {
+ if (cfs_list_empty(&peer->ibp_conns)) {
kiblnd_unlink_peer_locked(peer);
} else {
- list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
- conn = list_entry(ctmp, kib_conn_t, ibc_list);
+ cfs_list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
+ conn = cfs_list_entry(ctmp, kib_conn_t, ibc_list);
kiblnd_close_conn_locked(conn, 0);
}
int
kiblnd_del_peer (lnet_ni_t *ni, lnet_nid_t nid)
{
- CFS_LIST_HEAD (zombies);
- struct list_head *ptmp;
- struct list_head *pnxt;
- kib_peer_t *peer;
- int lo;
- int hi;
- int i;
- unsigned long flags;
- int rc = -ENOENT;
-
- write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
+ 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;
+
+ cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
if (nid != LNET_NID_ANY) {
lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
}
for (i = lo; i <= hi; i++) {
- list_for_each_safe (ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
- peer = list_entry(ptmp, kib_peer_t, ibp_list);
+ 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 ||
- !list_empty(&peer->ibp_conns));
+ !cfs_list_empty(&peer->ibp_conns));
if (peer->ibp_ni != ni)
continue;
if (!(nid == LNET_NID_ANY || peer->ibp_nid == nid))
continue;
- if (!list_empty(&peer->ibp_tx_queue)) {
- LASSERT (list_empty(&peer->ibp_conns));
+ if (!cfs_list_empty(&peer->ibp_tx_queue)) {
+ LASSERT (cfs_list_empty(&peer->ibp_conns));
- list_splice_init(&peer->ibp_tx_queue, &zombies);
+ cfs_list_splice_init(&peer->ibp_tx_queue,
+ &zombies);
}
kiblnd_del_peer_locked(peer);
}
}
- write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
+ cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
kiblnd_txlist_done(ni, &zombies, -EIO);
kib_conn_t *
kiblnd_get_conn_by_idx (lnet_ni_t *ni, int index)
{
- kib_peer_t *peer;
- struct list_head *ptmp;
- kib_conn_t *conn;
- struct list_head *ctmp;
- int i;
- unsigned long flags;
+ kib_peer_t *peer;
+ cfs_list_t *ptmp;
+ kib_conn_t *conn;
+ cfs_list_t *ctmp;
+ int i;
+ unsigned long flags;
- read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
+ cfs_read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
- list_for_each (ptmp, &kiblnd_data.kib_peers[i]) {
+ cfs_list_for_each (ptmp, &kiblnd_data.kib_peers[i]) {
- peer = list_entry(ptmp, kib_peer_t, ibp_list);
+ peer = cfs_list_entry(ptmp, kib_peer_t, ibp_list);
LASSERT (peer->ibp_connecting > 0 ||
peer->ibp_accepting > 0 ||
- !list_empty(&peer->ibp_conns));
+ !cfs_list_empty(&peer->ibp_conns));
if (peer->ibp_ni != ni)
continue;
- list_for_each (ctmp, &peer->ibp_conns) {
+ cfs_list_for_each (ctmp, &peer->ibp_conns) {
if (index-- > 0)
continue;
- conn = list_entry(ctmp, kib_conn_t, ibc_list);
+ conn = cfs_list_entry(ctmp, kib_conn_t,
+ ibc_list);
kiblnd_conn_addref(conn);
- read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
- flags);
+ cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
+ flags);
return conn;
}
}
}
- read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
+ cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
return NULL;
}
void
kiblnd_debug_conn (kib_conn_t *conn)
{
- struct list_head *tmp;
- int i;
+ cfs_list_t *tmp;
+ int i;
- spin_lock(&conn->ibc_lock);
+ cfs_spin_lock(&conn->ibc_lock);
CDEBUG(D_CONSOLE, "conn[%d] %p [version %x] -> %s: \n",
- atomic_read(&conn->ibc_refcount), conn,
+ 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,
CDEBUG(D_CONSOLE, " comms_err %d\n", conn->ibc_comms_error);
CDEBUG(D_CONSOLE, " early_rxs:\n");
- list_for_each(tmp, &conn->ibc_early_rxs)
- kiblnd_debug_rx(list_entry(tmp, kib_rx_t, rx_list));
+ cfs_list_for_each(tmp, &conn->ibc_early_rxs)
+ kiblnd_debug_rx(cfs_list_entry(tmp, kib_rx_t, rx_list));
+
+ CDEBUG(D_CONSOLE, " tx_noops:\n");
+ 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_queue_nocred:\n");
- list_for_each(tmp, &conn->ibc_tx_queue_nocred)
- kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
+ 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_rsrvd:\n");
- list_for_each(tmp, &conn->ibc_tx_queue_rsrvd)
- kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
+ 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:\n");
- list_for_each(tmp, &conn->ibc_tx_queue)
- kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
+ cfs_list_for_each(tmp, &conn->ibc_tx_queue)
+ kiblnd_debug_tx(cfs_list_entry(tmp, kib_tx_t, tx_list));
CDEBUG(D_CONSOLE, " active_txs:\n");
- list_for_each(tmp, &conn->ibc_active_txs)
- kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
+ cfs_list_for_each(tmp, &conn->ibc_active_txs)
+ kiblnd_debug_tx(cfs_list_entry(tmp, kib_tx_t, tx_list));
CDEBUG(D_CONSOLE, " rxs:\n");
for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++)
kiblnd_debug_rx(&conn->ibc_rxs[i]);
- spin_unlock(&conn->ibc_lock);
+ cfs_spin_unlock(&conn->ibc_lock);
}
int
}
static void
-kiblnd_setup_mtu(struct rdma_cm_id *cmid)
+kiblnd_setup_mtu_locked(struct rdma_cm_id *cmid)
{
- unsigned long flags;
int mtu;
/* XXX There is no path record for iWARP, set by netdev->change_mtu? */
if (cmid->route.path_rec == NULL)
return;
- write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
-
mtu = kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu);
LASSERT (mtu >= 0);
if (mtu != 0)
cmid->route.path_rec->mtu = mtu;
-
- write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
}
kib_conn_t *
* she must dispose of 'cmid'. (Actually I'd block forever if I tried
* to destroy 'cmid' here since I'm called from the CM which still has
* its ref on 'cmid'). */
+ cfs_rwlock_t *glock = &kiblnd_data.kib_global_lock;
kib_net_t *net = peer->ibp_ni->ni_data;
+ kib_dev_t *dev = net->ibn_dev;
struct ib_qp_init_attr *init_qp_attr;
kib_conn_t *conn;
struct ib_cq *cq;
int i;
LASSERT (net != NULL);
- LASSERT (!in_interrupt());
+ LASSERT (!cfs_in_interrupt());
LIBCFS_ALLOC(init_qp_attr, sizeof(*init_qp_attr));
if (init_qp_attr == NULL) {
cmid->context = conn; /* for future CM callbacks */
conn->ibc_cmid = cmid;
- INIT_LIST_HEAD(&conn->ibc_early_rxs);
- INIT_LIST_HEAD(&conn->ibc_tx_queue);
- INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
- INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
- INIT_LIST_HEAD(&conn->ibc_active_txs);
- spin_lock_init(&conn->ibc_lock);
+ 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);
+ cfs_spin_lock_init(&conn->ibc_lock);
LIBCFS_ALLOC(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
if (conn->ibc_connvars == NULL) {
}
memset(conn->ibc_connvars, 0, sizeof(*conn->ibc_connvars));
+ cfs_write_lock_irqsave(glock, flags);
+ if (dev->ibd_failover) {
+ cfs_write_unlock_irqrestore(glock, flags);
+ CERROR("%s: failover in progress\n", dev->ibd_ifname);
+ goto failed_2;
+ }
+
+ 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);
+ }
+
+ cfs_write_unlock_irqrestore(glock, flags);
+ CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
+ cmid->device->name, dev->ibd_ifname);
+ goto failed_2;
+ }
+
+ kiblnd_hdev_addref_locked(dev->ibd_hdev);
+ conn->ibc_hdev = dev->ibd_hdev;
+
+ kiblnd_setup_mtu_locked(cmid);
+
+ cfs_write_unlock_irqrestore(glock, flags);
+
LIBCFS_ALLOC(conn->ibc_rxs, IBLND_RX_MSGS(version) * sizeof(kib_rx_t));
if (conn->ibc_rxs == NULL) {
CERROR("Cannot allocate RX buffers\n");
goto failed_2;
}
- kiblnd_setup_mtu(cmid);
-
memset(init_qp_attr, 0, sizeof(*init_qp_attr));
init_qp_attr->event_handler = kiblnd_qp_event;
init_qp_attr->qp_context = conn;
init_qp_attr->send_cq = cq;
init_qp_attr->recv_cq = cq;
- rc = rdma_create_qp(cmid, net->ibn_dev->ibd_pd, init_qp_attr);
+ 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,
LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
/* 1 ref for caller and each rxmsg */
- atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(version));
+ cfs_atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(version));
conn->ibc_nrx = IBLND_RX_MSGS(version);
/* post receives */
/* Make posted receives complete */
kiblnd_abort_receives(conn);
- /* correct # of posted buffers
+ /* correct # of posted buffers
* NB locking needed now I'm racing with completion */
- spin_lock_irqsave(&kiblnd_data.kib_sched_lock, flags);
+ cfs_spin_lock_irqsave(&kiblnd_data.kib_sched_lock,
+ flags);
conn->ibc_nrx -= IBLND_RX_MSGS(version) - i;
- spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock,
- flags);
+ cfs_spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock,
+ flags);
/* cmid will be destroyed by CM(ofed) after cm_callback
* returned, so we can't refer it anymore
conn->ibc_state = state;
/* 1 more conn */
- atomic_inc(&net->ibn_nconns);
+ cfs_atomic_inc(&net->ibn_nconns);
return conn;
failed_2:
kib_peer_t *peer = 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_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 (!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);
if (conn->ibc_connvars != NULL)
LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
+ if (conn->ibc_hdev != NULL)
+ kiblnd_hdev_decref(conn->ibc_hdev);
+
/* See CAVEAT EMPTOR above in kiblnd_create_conn */
if (conn->ibc_state != IBLND_CONN_INIT) {
kib_net_t *net = peer->ibp_ni->ni_data;
kiblnd_peer_decref(peer);
rdma_destroy_id(cmid);
- atomic_dec(&net->ibn_nconns);
+ cfs_atomic_dec(&net->ibn_nconns);
}
LIBCFS_FREE(conn, sizeof(*conn));
int
kiblnd_close_peer_conns_locked (kib_peer_t *peer, int why)
{
- kib_conn_t *conn;
- struct list_head *ctmp;
- struct list_head *cnxt;
- int count = 0;
+ kib_conn_t *conn;
+ cfs_list_t *ctmp;
+ cfs_list_t *cnxt;
+ int count = 0;
- list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
- conn = list_entry(ctmp, kib_conn_t, ibc_list);
+ cfs_list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
+ conn = cfs_list_entry(ctmp, kib_conn_t, ibc_list);
CDEBUG(D_NET, "Closing conn -> %s, "
"version: %x, reason: %d\n",
kiblnd_close_stale_conns_locked (kib_peer_t *peer,
int version, __u64 incarnation)
{
- kib_conn_t *conn;
- struct list_head *ctmp;
- struct list_head *cnxt;
- int count = 0;
+ kib_conn_t *conn;
+ cfs_list_t *ctmp;
+ cfs_list_t *cnxt;
+ int count = 0;
- list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
- conn = list_entry(ctmp, kib_conn_t, ibc_list);
+ cfs_list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
+ conn = cfs_list_entry(ctmp, kib_conn_t, ibc_list);
if (conn->ibc_version == version &&
conn->ibc_incarnation == incarnation)
int
kiblnd_close_matching_conns (lnet_ni_t *ni, lnet_nid_t nid)
{
- kib_peer_t *peer;
- struct list_head *ptmp;
- struct list_head *pnxt;
- int lo;
- int hi;
- int i;
- unsigned long flags;
- int count = 0;
+ kib_peer_t *peer;
+ cfs_list_t *ptmp;
+ cfs_list_t *pnxt;
+ int lo;
+ int hi;
+ int i;
+ unsigned long flags;
+ int count = 0;
- write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
+ cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
if (nid != LNET_NID_ANY)
lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
}
for (i = lo; i <= hi; i++) {
- list_for_each_safe (ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
+ cfs_list_for_each_safe (ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
- peer = list_entry(ptmp, kib_peer_t, ibp_list);
+ peer = cfs_list_entry(ptmp, kib_peer_t, ibp_list);
LASSERT (peer->ibp_connecting > 0 ||
peer->ibp_accepting > 0 ||
- !list_empty(&peer->ibp_conns));
+ !cfs_list_empty(&peer->ibp_conns));
if (peer->ibp_ni != ni)
continue;
}
}
- write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
+ cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
/* wildcards always succeed */
if (nid == LNET_NID_ANY)
}
void
-kiblnd_query (lnet_ni_t *ni, lnet_nid_t nid, time_t *when)
+kiblnd_query (lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when)
{
- cfs_time_t last_alive = 0;
- rwlock_t *glock = &kiblnd_data.kib_global_lock;
- kib_peer_t *peer;
- unsigned long flags;
+ cfs_time_t last_alive = 0;
+ cfs_time_t now = cfs_time_current();
+ cfs_rwlock_t *glock = &kiblnd_data.kib_global_lock;
+ kib_peer_t *peer;
+ unsigned long flags;
- read_lock_irqsave(glock, flags);
+ cfs_read_lock_irqsave(glock, flags);
peer = kiblnd_find_peer_locked(nid);
if (peer != NULL) {
LASSERT (peer->ibp_connecting > 0 || /* creating conns */
peer->ibp_accepting > 0 ||
- !list_empty(&peer->ibp_conns)); /* active conn */
+ !cfs_list_empty(&peer->ibp_conns)); /* active conn */
last_alive = peer->ibp_last_alive;
}
- read_unlock_irqrestore(glock, flags);
+ cfs_read_unlock_irqrestore(glock, flags);
if (last_alive != 0)
- *when = cfs_time_current_sec() -
- cfs_duration_sec(cfs_time_current() - last_alive);
+ *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);
+
+ 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;
}
int npages = p->ibp_npages;
int i;
- LASSERT (p->ibp_device == NULL);
-
for (i = 0; i < npages; i++)
if (p->ibp_pages[i] != NULL)
__free_page(p->ibp_pages[i]);
int i;
LASSERT (conn->ibc_rxs != NULL);
- LASSERT (conn->ibc_rx_pages->ibp_device != NULL);
+ LASSERT (conn->ibc_hdev != NULL);
for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++) {
rx = &conn->ibc_rxs[i];
LASSERT (rx->rx_nob >= 0); /* not posted */
- kiblnd_dma_unmap_single(conn->ibc_rx_pages->ibp_device,
+ kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
rx->rx_msgaddr),
IBLND_MSG_SIZE, DMA_FROM_DEVICE);
}
- conn->ibc_rx_pages->ibp_device = NULL;
-
kiblnd_free_pages(conn->ibc_rx_pages);
conn->ibc_rx_pages = NULL;
rx->rx_conn = conn;
rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
- rx->rx_msgaddr = kiblnd_dma_map_single(conn->ibc_cmid->device,
+ rx->rx_msgaddr = kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
rx->rx_msg, IBLND_MSG_SIZE,
DMA_FROM_DEVICE);
- LASSERT (!kiblnd_dma_mapping_error(conn->ibc_cmid->device,
+ LASSERT (!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
rx->rx_msgaddr));
KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
LASSERT (ipg <= IBLND_RX_MSG_PAGES(conn->ibc_version));
}
}
-
- conn->ibc_rx_pages->ibp_device = conn->ibc_cmid->device;
}
-void
-kiblnd_unmap_tx_descs(lnet_ni_t *ni)
+static void
+kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo)
{
- int i;
+ kib_hca_dev_t *hdev = tpo->tpo_hdev;
kib_tx_t *tx;
- kib_net_t *net = ni->ni_data;
+ int i;
- LASSERT (net->ibn_tx_pages != NULL);
- LASSERT (net->ibn_tx_pages->ibp_device != NULL);
+ LASSERT (tpo->tpo_pool.po_allocated == 0);
- for (i = 0; i < IBLND_TX_MSGS(); i++) {
- tx = &net->ibn_tx_descs[i];
+ if (hdev == NULL)
+ return;
- kiblnd_dma_unmap_single(net->ibn_tx_pages->ibp_device,
+ for (i = 0; i < tpo->tpo_pool.po_size; i++) {
+ tx = &tpo->tpo_tx_descs[i];
+ kiblnd_dma_unmap_single(hdev->ibh_ibdev,
KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
tx->tx_msgaddr),
IBLND_MSG_SIZE, DMA_TO_DEVICE);
}
- net->ibn_tx_pages->ibp_device = NULL;
+ kiblnd_hdev_decref(hdev);
+ tpo->tpo_hdev = NULL;
}
-void
-kiblnd_map_tx_descs (lnet_ni_t *ni)
+static kib_hca_dev_t *
+kiblnd_current_hdev(kib_dev_t *dev)
{
- kib_net_t *net = ni->ni_data;
+ kib_hca_dev_t *hdev;
+ unsigned long flags;
+ int i = 0;
+
+ cfs_read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
+ while (dev->ibd_failover) {
+ cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
+ if (i++ % 50 == 0)
+ CDEBUG(D_NET, "Wait for dev(%s) failover\n", dev->ibd_ifname);
+ cfs_schedule_timeout(cfs_time_seconds(1) / 100);
+
+ cfs_read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
+ }
+
+ kiblnd_hdev_addref_locked(dev->ibd_hdev);
+ hdev = dev->ibd_hdev;
+
+ cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
+
+ return hdev;
+}
+
+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 = net->ibn_dev;
struct page *page;
kib_tx_t *tx;
int page_offset;
/* No fancy arithmetic when we do the buffer calculations */
CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
+ tpo->tpo_hdev = kiblnd_current_hdev(dev);
- for (ipage = page_offset = i = 0; i < IBLND_TX_MSGS(); i++) {
- page = net->ibn_tx_pages->ibp_pages[ipage];
- tx = &net->ibn_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_msgaddr = kiblnd_dma_map_single(
- net->ibn_dev->ibd_cmid->device,
- tx->tx_msg, IBLND_MSG_SIZE, DMA_TO_DEVICE);
- LASSERT (!kiblnd_dma_mapping_error(net->ibn_dev->ibd_cmid->device,
+ 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);
- list_add(&tx->tx_list, &net->ibn_idle_txs);
+ cfs_list_add(&tx->tx_list, &pool->po_free_list);
page_offset += IBLND_MSG_SIZE;
LASSERT (page_offset <= PAGE_SIZE);
if (page_offset == PAGE_SIZE) {
page_offset = 0;
ipage++;
- LASSERT (ipage <= IBLND_TX_MSG_PAGES());
- }
- }
-
- net->ibn_tx_pages->ibp_device = net->ibn_dev->ibd_cmid->device;
-}
-
-void
-kiblnd_free_tx_descs (lnet_ni_t *ni)
-{
- int i;
- kib_net_t *net = ni->ni_data;
-
- LASSERT (net != NULL);
-
- if (net->ibn_tx_pages != NULL)
- kiblnd_free_pages(net->ibn_tx_pages);
-
- if (net->ibn_tx_descs == NULL)
- return;
-
- for (i = 0; i < IBLND_TX_MSGS(); i++) {
- kib_tx_t *tx = &net->ibn_tx_descs[i];
-
- if (tx->tx_pages != NULL)
- LIBCFS_FREE(tx->tx_pages,
- LNET_MAX_IOV *
- sizeof(*tx->tx_pages));
-
- if (tx->tx_ipb != NULL)
- LIBCFS_FREE(tx->tx_ipb,
- IBLND_MAX_RDMA_FRAGS *
- sizeof(*tx->tx_ipb));
-
- if (tx->tx_frags != NULL)
- LIBCFS_FREE(tx->tx_frags,
- 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_rd != NULL)
- LIBCFS_FREE(tx->tx_rd,
- offsetof(kib_rdma_desc_t,
- rd_frags[IBLND_MAX_RDMA_FRAGS]));
- }
-
- LIBCFS_FREE(net->ibn_tx_descs,
- IBLND_TX_MSGS() * sizeof(kib_tx_t));
-}
-
-int
-kiblnd_alloc_tx_descs (lnet_ni_t *ni)
-{
- int i;
- int rc;
- kib_net_t *net = ni->ni_data;
-
- LASSERT (net != NULL);
-
- rc = kiblnd_alloc_pages(&net->ibn_tx_pages, IBLND_TX_MSG_PAGES());
-
- if (rc != 0) {
- CERROR("Can't allocate tx pages\n");
- return rc;
- }
-
- LIBCFS_ALLOC (net->ibn_tx_descs,
- IBLND_TX_MSGS() * sizeof(kib_tx_t));
- if (net->ibn_tx_descs == NULL) {
- CERROR("Can't allocate %d tx descriptors\n", IBLND_TX_MSGS());
- return -ENOMEM;
- }
-
- memset(net->ibn_tx_descs, 0,
- IBLND_TX_MSGS() * sizeof(kib_tx_t));
-
- for (i = 0; i < IBLND_TX_MSGS(); i++) {
- kib_tx_t *tx = &net->ibn_tx_descs[i];
-
- if (net->ibn_fmrpool != NULL) {
- LIBCFS_ALLOC(tx->tx_pages, LNET_MAX_IOV *
- sizeof(*tx->tx_pages));
- if (tx->tx_pages == NULL)
- return -ENOMEM;
- }
-
- if (net->ibn_pmrpool != NULL) {
- LIBCFS_ALLOC(tx->tx_ipb,
- IBLND_MAX_RDMA_FRAGS *
- sizeof(*tx->tx_ipb));
- if (tx->tx_ipb == NULL)
- return -ENOMEM;
+ LASSERT (ipage <= txpgs->ibp_npages);
}
-
- LIBCFS_ALLOC(tx->tx_frags,
- IBLND_MAX_RDMA_FRAGS *
- sizeof(*tx->tx_frags));
- if (tx->tx_frags == NULL)
- return -ENOMEM;
-
- LIBCFS_ALLOC(tx->tx_wrq,
- (1 + IBLND_MAX_RDMA_FRAGS) *
- sizeof(*tx->tx_wrq));
- if (tx->tx_wrq == NULL)
- return -ENOMEM;
-
- LIBCFS_ALLOC(tx->tx_sge,
- (1 + IBLND_MAX_RDMA_FRAGS) *
- sizeof(*tx->tx_sge));
- if (tx->tx_sge == NULL)
- return -ENOMEM;
-
- LIBCFS_ALLOC(tx->tx_rd,
- offsetof(kib_rdma_desc_t,
- rd_frags[IBLND_MAX_RDMA_FRAGS]));
- if (tx->tx_rd == NULL)
- return -ENOMEM;
}
-
- return 0;
}
struct ib_mr *
-kiblnd_find_dma_mr(kib_net_t *net, __u64 addr, __u64 size)
+kiblnd_find_dma_mr(kib_hca_dev_t *hdev, __u64 addr, __u64 size)
{
__u64 index;
- LASSERT (net->ibn_dev->ibd_mrs[0] != NULL);
+ LASSERT (hdev->ibh_mrs[0] != NULL);
- if (net->ibn_dev->ibd_nmrs == 1)
- return net->ibn_dev->ibd_mrs[0];
+ if (hdev->ibh_nmrs == 1)
+ return hdev->ibh_mrs[0];
- index = addr >> net->ibn_dev->ibd_mr_shift;
+ index = addr >> hdev->ibh_mr_shift;
- if (index < net->ibn_dev->ibd_nmrs &&
- index == ((addr + size - 1) >> net->ibn_dev->ibd_mr_shift))
- return net->ibn_dev->ibd_mrs[index];
+ if (index < hdev->ibh_nmrs &&
+ index == ((addr + size - 1) >> hdev->ibh_mr_shift))
+ return hdev->ibh_mrs[index];
return NULL;
}
struct ib_mr *
-kiblnd_find_rd_dma_mr(kib_net_t *net, kib_rdma_desc_t *rd)
+kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev, kib_rdma_desc_t *rd)
{
struct ib_mr *prev_mr;
struct ib_mr *mr;
int i;
- LASSERT (net->ibn_dev->ibd_mrs[0] != NULL);
+ LASSERT (hdev->ibh_mrs[0] != NULL);
if (*kiblnd_tunables.kib_map_on_demand > 0 &&
- *kiblnd_tunables.kib_map_on_demand < rd->rd_nfrags)
+ *kiblnd_tunables.kib_map_on_demand <= rd->rd_nfrags)
return NULL;
- if (net->ibn_dev->ibd_nmrs == 1)
- return net->ibn_dev->ibd_mrs[0];
+ if (hdev->ibh_nmrs == 1)
+ return hdev->ibh_mrs[0];
for (i = 0, mr = prev_mr = NULL;
i < rd->rd_nfrags; i++) {
- mr = kiblnd_find_dma_mr(net,
+ mr = kiblnd_find_dma_mr(hdev,
rd->rd_frags[i].rf_addr,
rd->rd_frags[i].rf_nob);
if (prev_mr == NULL)
}
void
-kiblnd_dev_cleanup(kib_dev_t *ibdev)
+kiblnd_destroy_fmr_pool(kib_fmr_pool_t *pool)
{
- int i;
+ LASSERT (pool->fpo_map_count == 0);
- if (ibdev->ibd_mrs == NULL)
- return;
+ if (pool->fpo_fmr_pool != NULL)
+ ib_destroy_fmr_pool(pool->fpo_fmr_pool);
- for (i = 0; i < ibdev->ibd_nmrs; i++) {
- if (ibdev->ibd_mrs[i] == NULL)
- break;
+ if (pool->fpo_hdev != NULL)
+ kiblnd_hdev_decref(pool->fpo_hdev);
- ib_dereg_mr(ibdev->ibd_mrs[i]);
- }
+ LIBCFS_FREE(pool, sizeof(kib_fmr_pool_t));
+}
- LIBCFS_FREE(ibdev->ibd_mrs, sizeof(*ibdev->ibd_mrs) * ibdev->ibd_nmrs);
- ibdev->ibd_mrs = NULL;
+void
+kiblnd_destroy_fmr_pool_list(cfs_list_t *head)
+{
+ kib_fmr_pool_t *pool;
+
+ while (!cfs_list_empty(head)) {
+ pool = cfs_list_entry(head->next, kib_fmr_pool_t, fpo_list);
+ cfs_list_del(&pool->fpo_list);
+ kiblnd_destroy_fmr_pool(pool);
+ }
}
int
-kiblnd_ib_create_fmr_pool(kib_dev_t *ibdev, struct ib_fmr_pool **fmrpp)
+kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo)
{
/* FMR pool for RDMA */
- struct ib_fmr_pool *fmrpool;
+ 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,
.dirty_watermark = *kiblnd_tunables.kib_fmr_flush_trigger,
.flush_function = NULL,
.flush_arg = NULL,
- .cache = *kiblnd_tunables.kib_fmr_cache};
+ .cache = !!*kiblnd_tunables.kib_fmr_cache};
+ int rc;
- if (*kiblnd_tunables.kib_fmr_pool_size <
- *kiblnd_tunables.kib_ntx) {
- CERROR("Can't set fmr pool size (%d) < ntx(%d)\n",
- *kiblnd_tunables.kib_fmr_pool_size,
- *kiblnd_tunables.kib_ntx);
- return -EINVAL;
- }
+ LIBCFS_ALLOC(fpo, sizeof(kib_fmr_pool_t));
+ if (fpo == NULL)
+ return -ENOMEM;
- fmrpool = ib_create_fmr_pool(ibdev->ibd_pd, ¶m);
- if (IS_ERR(fmrpool))
- return PTR_ERR(fmrpool);
+ memset(fpo, 0, sizeof(kib_fmr_pool_t));
+ fpo->fpo_hdev = kiblnd_current_hdev(dev);
+ fpo->fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd, ¶m);
+ if (IS_ERR(fpo->fpo_fmr_pool)) {
+ CERROR("Failed to create FMR pool: %ld\n",
+ PTR_ERR(fpo->fpo_fmr_pool));
+ rc = PTR_ERR(fpo->fpo_fmr_pool);
+ kiblnd_hdev_decref(fpo->fpo_hdev);
+ LIBCFS_FREE(fpo, sizeof(kib_fmr_pool_t));
+ return rc;
+ }
- *fmrpp = fmrpool;
+ fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
+ fpo->fpo_owner = fps;
+ *pp_fpo = fpo;
return 0;
}
-void
-kiblnd_phys_mr_unmap(kib_net_t *net, kib_phys_mr_t *pmr)
+static void
+kiblnd_fail_fmr_pool_set(kib_fmr_poolset_t *fps, cfs_list_t *zombies)
{
- kib_phys_mr_pool_t *pool = net->ibn_pmrpool;
- struct ib_mr *mr;
-
- spin_lock(&pool->ibmp_lock);
-
- mr = pmr->ibpm_mr;
-
- list_add(&pmr->ibpm_link, &pool->ibmp_free_list);
- pool->ibmp_allocated --;
+ cfs_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);
+ }
- spin_unlock(&pool->ibmp_lock);
+ cfs_spin_unlock(&fps->fps_lock);
+}
- if (mr != NULL)
- ib_dereg_mr(mr);
+static void
+kiblnd_fini_fmr_pool_set(kib_fmr_poolset_t *fps)
+{
+ kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
+ kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
}
-kib_phys_mr_t *
-kiblnd_phys_mr_map(kib_net_t *net, kib_rdma_desc_t *rd,
- struct ib_phys_buf *ipb, __u64 *iova)
+static int
+kiblnd_init_fmr_pool_set(kib_fmr_poolset_t *fps, kib_net_t *net)
{
- kib_phys_mr_pool_t *pool = net->ibn_pmrpool;
- kib_phys_mr_t *pmr;
- int i;
+ kib_fmr_pool_t *fpo;
+ int rc;
- LASSERT (ipb != NULL);
+ memset(fps, 0, sizeof(kib_fmr_poolset_t));
- spin_lock(&pool->ibmp_lock);
- if (list_empty(&pool->ibmp_free_list)) {
- spin_unlock(&pool->ibmp_lock);
- CERROR("pre-allocated MRs is not enough\n");
+ fps->fps_net = net;
+ cfs_spin_lock_init(&fps->fps_lock);
+ CFS_INIT_LIST_HEAD(&fps->fps_pool_list);
+ CFS_INIT_LIST_HEAD(&fps->fps_failed_pool_list);
+ rc = kiblnd_create_fmr_pool(fps, &fpo);
+ if (rc == 0)
+ cfs_list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
- return NULL;
- }
+ return rc;
+}
- pmr = list_entry(pool->ibmp_free_list.next,
- kib_phys_mr_t, ibpm_link);
- list_del_init(&pmr->ibpm_link);
- pool->ibmp_allocated ++;
+static int
+kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, cfs_time_t now)
+{
+ if (fpo->fpo_map_count != 0) /* still in use */
+ return 0;
+ if (fpo->fpo_failed)
+ return 1;
+ return cfs_time_aftereq(now, fpo->fpo_deadline);
+}
- spin_unlock(&pool->ibmp_lock);
+void
+kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
+{
+ CFS_LIST_HEAD (zombies);
+ kib_fmr_pool_t *fpo = fmr->fmr_pool;
+ kib_fmr_poolset_t *fps = fpo->fpo_owner;
+ cfs_time_t now = cfs_time_current();
+ kib_fmr_pool_t *tmp;
+ int rc;
- for (i = 0; i < rd->rd_nfrags; i ++) {
- ipb[i].addr = rd->rd_frags[i].rf_addr;
- ipb[i].size = rd->rd_frags[i].rf_nob;
- }
+ rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
+ LASSERT (rc == 0);
- pmr->ibpm_mr = ib_reg_phys_mr(net->ibn_dev->ibd_pd, ipb,
- rd->rd_nfrags,
- IB_ACCESS_LOCAL_WRITE |
- IB_ACCESS_REMOTE_WRITE,
- iova);
- if (!IS_ERR(pmr->ibpm_mr)) {
- pmr->ibpm_iova = *iova;
- return pmr;
+ if (status != 0) {
+ rc = ib_flush_fmr_pool(fpo->fpo_fmr_pool);
+ LASSERT (rc == 0);
}
- CERROR("Failed ib_reg_phys_mr: %ld\n", PTR_ERR(pmr->ibpm_mr));
- pmr->ibpm_mr = NULL;
+ fmr->fmr_pool = NULL;
+ fmr->fmr_pfmr = NULL;
- spin_lock(&pool->ibmp_lock);
+ cfs_spin_lock(&fps->fps_lock);
+ fpo->fpo_map_count --; /* decref the pool */
- list_add(&pmr->ibpm_link, &pool->ibmp_free_list);
- pool->ibmp_allocated --;
+ 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;
- spin_unlock(&pool->ibmp_lock);
+ if (kiblnd_fmr_pool_is_idle(fpo, now)) {
+ cfs_list_move(&fpo->fpo_list, &zombies);
+ fps->fps_version ++;
+ }
+ }
+ cfs_spin_unlock(&fps->fps_lock);
- return NULL;
+ if (!cfs_list_empty(&zombies))
+ kiblnd_destroy_fmr_pool_list(&zombies);
}
-void
-kiblnd_destroy_pmr_pool(kib_phys_mr_pool_t *pool)
+int
+kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages,
+ __u64 iov, kib_fmr_t *fmr)
{
- kib_phys_mr_t *pmr;
-
- LASSERT (pool->ibmp_allocated == 0);
+ struct ib_pool_fmr *pfmr;
+ kib_fmr_pool_t *fpo;
+ __u64 version;
+ int rc;
+
+ LASSERT (fps->fps_net->ibn_with_fmr);
+ again:
+ cfs_spin_lock(&fps->fps_lock);
+ version = fps->fps_version;
+ cfs_list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
+ fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
+ fpo->fpo_map_count ++;
+ cfs_spin_unlock(&fps->fps_lock);
+
+ 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;
+ }
- while (!list_empty(&pool->ibmp_free_list)) {
- pmr = list_entry(pool->ibmp_free_list.next,
- kib_phys_mr_t, ibpm_link);
+ cfs_spin_lock(&fps->fps_lock);
+ fpo->fpo_map_count --;
+ if (PTR_ERR(pfmr) != -EAGAIN) {
+ cfs_spin_unlock(&fps->fps_lock);
+ return PTR_ERR(pfmr);
+ }
- LASSERT (pmr->ibpm_mr == NULL);
+ /* EAGAIN and ... */
+ if (version != fps->fps_version) {
+ cfs_spin_unlock(&fps->fps_lock);
+ goto again;
+ }
+ }
- list_del(&pmr->ibpm_link);
+ if (fps->fps_increasing) {
+ cfs_spin_unlock(&fps->fps_lock);
+ CDEBUG(D_NET, "Another thread is allocating new "
+ "FMR pool, waiting for her to complete\n");
+ cfs_schedule();
+ goto again;
- LIBCFS_FREE(pmr, sizeof(kib_phys_mr_t));
}
- LIBCFS_FREE(pool, sizeof(kib_phys_mr_pool_t));
-}
+ if (cfs_time_before(cfs_time_current(), fps->fps_next_retry)) {
+ /* someone failed recently */
+ cfs_spin_unlock(&fps->fps_lock);
+ return -EAGAIN;
+ }
-int
-kiblnd_create_pmr_pool(kib_dev_t *ibdev, kib_phys_mr_pool_t **poolpp)
-{
- kib_phys_mr_pool_t *pool;
- kib_phys_mr_t *pmr;
- int i;
+ fps->fps_increasing = 1;
+ cfs_spin_unlock(&fps->fps_lock);
- if (*kiblnd_tunables.kib_pmr_pool_size <
- *kiblnd_tunables.kib_ntx) {
- CERROR("Can't set pmr pool size (%d) < ntx(%d)\n",
- *kiblnd_tunables.kib_pmr_pool_size,
- *kiblnd_tunables.kib_ntx);
- return -EINVAL;
+ CDEBUG(D_NET, "Allocate new FMR pool\n");
+ rc = kiblnd_create_fmr_pool(fps, &fpo);
+ cfs_spin_lock(&fps->fps_lock);
+ fps->fps_increasing = 0;
+ if (rc == 0) {
+ fps->fps_version ++;
+ cfs_list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
+ } else {
+ fps->fps_next_retry = cfs_time_shift(10);
}
+ cfs_spin_unlock(&fps->fps_lock);
- LIBCFS_ALLOC(pool, sizeof(kib_phys_mr_pool_t));
- if (pool == NULL)
- return -ENOMEM;
+ goto again;
+}
- spin_lock_init(&pool->ibmp_lock);
+static void
+kiblnd_fini_pool(kib_pool_t *pool)
+{
+ LASSERT (cfs_list_empty(&pool->po_free_list));
+ LASSERT (pool->po_allocated == 0);
- pool->ibmp_allocated = 0;
- CFS_INIT_LIST_HEAD(&pool->ibmp_free_list);
+ CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
+}
- for (i = 0; i < *kiblnd_tunables.kib_pmr_pool_size; i++) {
- LIBCFS_ALLOC(pmr, sizeof(kib_phys_mr_t));
+static void
+kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
+{
+ CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
- if (pmr == NULL) {
- kiblnd_destroy_pmr_pool(pool);
- return -ENOMEM;
- }
+ 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(pmr, 0, sizeof(kib_phys_mr_t));
+void
+kiblnd_destroy_pool_list(cfs_list_t *head)
+{
+ kib_pool_t *pool;
+
+ while (!cfs_list_empty(head)) {
+ pool = cfs_list_entry(head->next, kib_pool_t, po_list);
+ cfs_list_del(&pool->po_list);
- list_add(&pmr->ibpm_link, &pool->ibmp_free_list);
+ LASSERT (pool->po_owner != NULL);
+ pool->po_owner->ps_pool_destroy(pool);
}
+}
- *poolpp = pool;
+static void
+kiblnd_fail_pool_set(kib_poolset_t *ps, cfs_list_t *zombies)
+{
+ cfs_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);
+ }
+ cfs_spin_unlock(&ps->ps_lock);
+}
- return 0;
+static void
+kiblnd_fini_pool_set(kib_poolset_t *ps)
+{
+ kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
+ kiblnd_destroy_pool_list(&ps->ps_pool_list);
}
static int
-kiblnd_dev_get_attr(kib_dev_t *ibdev)
+kiblnd_init_pool_set(kib_poolset_t *ps, kib_net_t *net,
+ char *name, int size,
+ kib_ps_pool_create_t po_create,
+ kib_ps_pool_destroy_t po_destroy,
+ kib_ps_node_init_t nd_init,
+ kib_ps_node_fini_t nd_fini)
{
- struct ib_device_attr *attr;
- int rc;
+ kib_pool_t *pool;
+ int rc;
+
+ memset(ps, 0, sizeof(kib_poolset_t));
+
+ ps->ps_net = net;
+ ps->ps_pool_create = po_create;
+ ps->ps_pool_destroy = po_destroy;
+ ps->ps_node_init = nd_init;
+ ps->ps_node_fini = nd_fini;
+ ps->ps_pool_size = size;
+ strncpy(ps->ps_name, name, IBLND_POOL_NAME_LEN);
+ cfs_spin_lock_init(&ps->ps_lock);
+ CFS_INIT_LIST_HEAD(&ps->ps_pool_list);
+ CFS_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);
- /* XXX here should be HCA's page shift/size/mask in the future? */
- ibdev->ibd_page_shift = PAGE_SHIFT;
- ibdev->ibd_page_size = 1 << PAGE_SHIFT;
- ibdev->ibd_page_mask = ~((__u64)ibdev->ibd_page_size - 1);
+ return rc;
+}
- LIBCFS_ALLOC(attr, sizeof(*attr));
- if (attr == NULL) {
- CERROR("Out of memory\n");
- return -ENOMEM;
+static int
+kiblnd_pool_is_idle(kib_pool_t *pool, cfs_time_t now)
+{
+ if (pool->po_allocated != 0) /* still in use */
+ return 0;
+ if (pool->po_failed)
+ return 1;
+ return cfs_time_aftereq(now, pool->po_deadline);
+}
+
+void
+kiblnd_pool_free_node(kib_pool_t *pool, cfs_list_t *node)
+{
+ CFS_LIST_HEAD (zombies);
+ kib_poolset_t *ps = pool->po_owner;
+ kib_pool_t *tmp;
+ cfs_time_t now = cfs_time_current();
+
+ cfs_spin_lock(&ps->ps_lock);
+
+ 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 --;
+
+ 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;
+
+ if (kiblnd_pool_is_idle(pool, now))
+ cfs_list_move(&pool->po_list, &zombies);
+ }
+ cfs_spin_unlock(&ps->ps_lock);
+
+ if (!cfs_list_empty(&zombies))
+ kiblnd_destroy_pool_list(&zombies);
+}
+
+cfs_list_t *
+kiblnd_pool_alloc_node(kib_poolset_t *ps)
+{
+ cfs_list_t *node;
+ kib_pool_t *pool;
+ int rc;
+
+ again:
+ cfs_spin_lock(&ps->ps_lock);
+ cfs_list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
+ if (cfs_list_empty(&pool->po_free_list))
+ continue;
+
+ pool->po_allocated ++;
+ pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
+ node = pool->po_free_list.next;
+ cfs_list_del(node);
+
+ if (ps->ps_node_init != NULL) {
+ /* still hold the lock */
+ ps->ps_node_init(pool, node);
+ }
+ cfs_spin_unlock(&ps->ps_lock);
+ return node;
+ }
+
+ /* no available tx pool and ... */
+ if (ps->ps_increasing) {
+ /* another thread is allocating a new pool */
+ cfs_spin_unlock(&ps->ps_lock);
+ CDEBUG(D_NET, "Another thread is allocating new "
+ "%s pool, waiting for her to complete\n",
+ ps->ps_name);
+ cfs_schedule();
+ goto again;
+ }
+
+ if (cfs_time_before(cfs_time_current(), ps->ps_next_retry)) {
+ /* someone failed recently */
+ cfs_spin_unlock(&ps->ps_lock);
+ return NULL;
}
- rc = ib_query_device(ibdev->ibd_cmid->device, attr);
+ ps->ps_increasing = 1;
+ cfs_spin_unlock(&ps->ps_lock);
+
+ CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
+
+ rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
+
+ cfs_spin_lock(&ps->ps_lock);
+ ps->ps_increasing = 0;
+ if (rc == 0) {
+ cfs_list_add_tail(&pool->po_list, &ps->ps_pool_list);
+ } else {
+ /* retry 10 seconds later */
+ ps->ps_next_retry = cfs_time_shift(10);
+ CERROR("Can't allocate new %s pool because out of memory\n",
+ ps->ps_name);
+ }
+ cfs_spin_unlock(&ps->ps_lock);
+
+ 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)
+{
+ 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 int
+kiblnd_create_pmr_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
+{
+ kib_pmr_pool_t *ppo;
+ kib_pool_t *pool;
+ kib_phys_mr_t *pmr;
+ int i;
+
+ LIBCFS_ALLOC(ppo, 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_ALLOC(pmr, sizeof(kib_phys_mr_t));
+ if (pmr == NULL)
+ break;
+
+ memset(pmr, 0, sizeof(kib_phys_mr_t));
+ pmr->pmr_pool = ppo;
+ LIBCFS_ALLOC(pmr->pmr_ipb,
+ IBLND_MAX_RDMA_FRAGS *
+ sizeof(struct ib_phys_buf));
+ 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;
+
+ LASSERT (pool->po_allocated == 0);
+
+ if (tpo->tpo_tx_pages != NULL) {
+ kiblnd_unmap_tx_pool(tpo);
+ kiblnd_free_pages(tpo->tpo_tx_pages);
+ }
+
+ if (tpo->tpo_tx_descs == NULL)
+ goto out;
+
+ for (i = 0; i < pool->po_size; i++) {
+ kib_tx_t *tx = &tpo->tpo_tx_descs[i];
+
+ cfs_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));
+ 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_rd != NULL)
+ LIBCFS_FREE(tx->tx_rd,
+ offsetof(kib_rdma_desc_t,
+ rd_frags[IBLND_MAX_RDMA_FRAGS]));
+ }
+
+ LIBCFS_FREE(tpo->tpo_tx_descs,
+ pool->po_size * sizeof(kib_tx_t));
+out:
+ kiblnd_fini_pool(pool);
+ LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
+}
+
+static int
+kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
+{
+ int i;
+ int npg;
+ kib_pool_t *pool;
+ kib_tx_pool_t *tpo;
+
+ LIBCFS_ALLOC(tpo, sizeof(kib_tx_pool_t));
+ if (tpo == NULL) {
+ CERROR("Failed to allocate TX pool\n");
+ return -ENOMEM;
+ }
+
+ pool = &tpo->tpo_pool;
+ kiblnd_init_pool(ps, pool, size);
+ tpo->tpo_tx_descs = NULL;
+ tpo->tpo_tx_pages = NULL;
+
+ npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
+ if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, npg) != 0) {
+ CERROR("Can't allocate tx pages: %d\n", npg);
+ LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
+ return -ENOMEM;
+ }
+
+ LIBCFS_ALLOC (tpo->tpo_tx_descs, size * sizeof(kib_tx_t));
+ if (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));
+
+ for (i = 0; i < size; i++) {
+ kib_tx_t *tx = &tpo->tpo_tx_descs[i];
+
+ tx->tx_pool = tpo;
+ if (ps->ps_net->ibn_with_fmr){
+ LIBCFS_ALLOC(tx->tx_pages, LNET_MAX_IOV *
+ sizeof(*tx->tx_pages));
+ if (tx->tx_pages == NULL)
+ break;
+ }
+
+ LIBCFS_ALLOC(tx->tx_frags,
+ IBLND_MAX_RDMA_FRAGS *
+ sizeof(*tx->tx_frags));
+ if (tx->tx_frags == NULL)
+ break;
+
+ LIBCFS_ALLOC(tx->tx_wrq,
+ (1 + IBLND_MAX_RDMA_FRAGS) *
+ sizeof(*tx->tx_wrq));
+ if (tx->tx_wrq == NULL)
+ break;
+
+ LIBCFS_ALLOC(tx->tx_sge,
+ (1 + IBLND_MAX_RDMA_FRAGS) *
+ sizeof(*tx->tx_sge));
+ if (tx->tx_sge == NULL)
+ break;
+
+ LIBCFS_ALLOC(tx->tx_rd,
+ offsetof(kib_rdma_desc_t,
+ rd_frags[IBLND_MAX_RDMA_FRAGS]));
+ if (tx->tx_rd == NULL)
+ break;
+ }
+
+ if (i == size) {
+ kiblnd_map_tx_pool(tpo);
+ *pp_po = pool;
+ return 0;
+ }
+
+ ps->ps_pool_destroy(pool);
+ return -ENOMEM;
+}
+
+static void
+kiblnd_tx_init(kib_pool_t *pool, cfs_list_t *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);
+
+ tx->tx_cookie = tps->tps_next_tx_cookie ++;
+}
+
+void
+kiblnd_ni_fini_pools(kib_net_t *net)
+{
+ kiblnd_fini_pool_set(&net->ibn_tx_ps.tps_poolset);
+ if (net->ibn_with_fmr)
+ kiblnd_fini_fmr_pool_set(&net->ibn_fmr_ps);
+ else if (net->ibn_with_pmr)
+ kiblnd_fini_pool_set(&net->ibn_pmr_ps.pps_poolset);
+}
+
+int
+kiblnd_net_init_pools(kib_net_t *net)
+{
+ kib_fmr_poolset_t *fps = &net->ibn_fmr_ps;
+ kib_pmr_poolset_t *pps = &net->ibn_pmr_ps;
+ kib_tx_poolset_t *tps = &net->ibn_tx_ps;
+ unsigned long flags;
+ int rc;
+
+ if (*kiblnd_tunables.kib_fmr_pool_size <
+ *kiblnd_tunables.kib_ntx / 4) {
+ CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
+ *kiblnd_tunables.kib_fmr_pool_size,
+ *kiblnd_tunables.kib_ntx / 4);
+ return -EINVAL;
+ }
+
+ if (*kiblnd_tunables.kib_pmr_pool_size <
+ *kiblnd_tunables.kib_ntx / 4) {
+ CERROR("Can't set pmr pool size (%d) < ntx / 4(%d)\n",
+ *kiblnd_tunables.kib_pmr_pool_size,
+ *kiblnd_tunables.kib_ntx / 4);
+ return -EINVAL;
+ }
+
+ cfs_read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
+ if (*kiblnd_tunables.kib_map_on_demand > 0 ||
+ net->ibn_dev->ibd_hdev->ibh_nmrs > 1) {
+ /* premapping can fail if ibd_nmr > 1, so we always create
+ * FMR/PMR pool and map-on-demand if premapping failed */
+ cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
+ rc = kiblnd_init_fmr_pool_set(fps, net);
+ if (rc == 0) {
+ net->ibn_with_fmr = 1;
+ } else if (rc == -ENOSYS) {
+ rc = kiblnd_init_pool_set(&pps->pps_poolset, net, "PMR",
+ *kiblnd_tunables.kib_pmr_pool_size,
+ kiblnd_create_pmr_pool,
+ kiblnd_destroy_pmr_pool,
+ NULL, NULL);
+ if (rc == 0)
+ net->ibn_with_pmr = 1;
+ }
+ if (rc != 0)
+ return rc;
+ } else {
+ cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
+ }
+
+ rc = kiblnd_init_pool_set(&tps->tps_poolset, net, "TX", IBLND_TX_MSGS(),
+ kiblnd_create_tx_pool, kiblnd_destroy_tx_pool,
+ kiblnd_tx_init, NULL);
if (rc == 0)
- ibdev->ibd_mr_size = attr->max_mr_size;
+ return 0;
+
+ if (net->ibn_with_fmr)
+ kiblnd_fini_fmr_pool_set(fps);
+ else if (net->ibn_with_pmr)
+ kiblnd_fini_pool_set(&pps->pps_poolset);
+
+ return rc;
+}
+
+static int
+kiblnd_hdev_get_attr(kib_hca_dev_t *hdev)
+{
+ struct ib_device_attr *attr;
+ int rc;
+
+ /* It's safe to assume a HCA can handle a page size
+ * matching that of the native system */
+ 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;
+ }
+
+ rc = ib_query_device(hdev->ibh_ibdev, attr);
+ if (rc == 0)
+ hdev->ibh_mr_size = attr->max_mr_size;
LIBCFS_FREE(attr, sizeof(*attr));
return rc;
}
-#if 1
+#ifdef HAVE_OFED_TRANSPORT_IWARP
/* XXX We can't trust this value returned by Chelsio driver, it's wrong
* and we have reported the bug, remove these in the future when Chelsio
* bug got fixed. */
- if (rdma_node_get_transport(ibdev->ibd_cmid->device->node_type) ==
+ if (rdma_node_get_transport(hdev->ibh_ibdev->node_type) ==
RDMA_TRANSPORT_IWARP)
- ibdev->ibd_mr_size = (1ULL << 32) - 1;
+ hdev->ibh_mr_size = (1ULL << 32) - 1;
#endif
- if (ibdev->ibd_mr_size == ~0ULL) {
- ibdev->ibd_mr_shift = 64;
+ if (hdev->ibh_mr_size == ~0ULL) {
+ hdev->ibh_mr_shift = 64;
return 0;
}
- for (ibdev->ibd_mr_shift = 0;
- ibdev->ibd_mr_shift < 64; ibdev->ibd_mr_shift ++) {
- if (ibdev->ibd_mr_size == (1ULL << ibdev->ibd_mr_shift) ||
- ibdev->ibd_mr_size == (1ULL << ibdev->ibd_mr_shift) - 1)
+ for (hdev->ibh_mr_shift = 0;
+ hdev->ibh_mr_shift < 64; hdev->ibh_mr_shift ++) {
+ if (hdev->ibh_mr_size == (1ULL << hdev->ibh_mr_shift) ||
+ hdev->ibh_mr_size == (1ULL << hdev->ibh_mr_shift) - 1)
return 0;
}
- CERROR("Invalid mr size: "LPX64"\n", ibdev->ibd_mr_size);
+ CERROR("Invalid mr size: "LPX64"\n", hdev->ibh_mr_size);
return -EINVAL;
}
+void
+kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev)
+{
+ int i;
+
+ if (hdev->ibh_nmrs == 0 || hdev->ibh_mrs == NULL)
+ return;
+
+ for (i = 0; i < hdev->ibh_nmrs; i++) {
+ if (hdev->ibh_mrs[i] == NULL)
+ break;
+
+ ib_dereg_mr(hdev->ibh_mrs[i]);
+ }
+
+ LIBCFS_FREE(hdev->ibh_mrs, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
+ hdev->ibh_mrs = NULL;
+ hdev->ibh_nmrs = 0;
+}
+
+void
+kiblnd_hdev_destroy(kib_hca_dev_t *hdev)
+{
+ kiblnd_hdev_cleanup_mrs(hdev);
+
+ if (hdev->ibh_pd != NULL)
+ ib_dealloc_pd(hdev->ibh_pd);
+
+ if (hdev->ibh_cmid != NULL)
+ rdma_destroy_id(hdev->ibh_cmid);
+
+ LIBCFS_FREE(hdev, sizeof(*hdev));
+}
+
int
-kiblnd_dev_setup(kib_dev_t *ibdev)
+kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev)
{
struct ib_mr *mr;
int i;
int acflags = IB_ACCESS_LOCAL_WRITE |
IB_ACCESS_REMOTE_WRITE;
- rc = kiblnd_dev_get_attr(ibdev);
+ rc = kiblnd_hdev_get_attr(hdev);
if (rc != 0)
return rc;
- if (ibdev->ibd_mr_shift == 64) {
- LIBCFS_ALLOC(ibdev->ibd_mrs, 1 * sizeof(*ibdev->ibd_mrs));
- if (ibdev->ibd_mrs == NULL) {
+ if (hdev->ibh_mr_shift == 64) {
+ LIBCFS_ALLOC(hdev->ibh_mrs, 1 * sizeof(*hdev->ibh_mrs));
+ if (hdev->ibh_mrs == NULL) {
CERROR("Failed to allocate MRs table\n");
return -ENOMEM;
}
- ibdev->ibd_mrs[0] = NULL;
- ibdev->ibd_nmrs = 1;
+ hdev->ibh_mrs[0] = NULL;
+ hdev->ibh_nmrs = 1;
- mr = ib_get_dma_mr(ibdev->ibd_pd, acflags);
+ mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
if (IS_ERR(mr)) {
CERROR("Failed ib_get_dma_mr : %ld\n", PTR_ERR(mr));
- kiblnd_dev_cleanup(ibdev);
+ kiblnd_hdev_cleanup_mrs(hdev);
return PTR_ERR(mr);
}
- ibdev->ibd_mrs[0] = mr;
+ hdev->ibh_mrs[0] = mr;
goto out;
}
- mr_size = (1ULL << ibdev->ibd_mr_shift);
+ mr_size = (1ULL << hdev->ibh_mr_shift);
mm_size = (unsigned long)high_memory - PAGE_OFFSET;
- ibdev->ibd_nmrs = (int)((mm_size + mr_size - 1) >> ibdev->ibd_mr_shift);
+ hdev->ibh_nmrs = (int)((mm_size + mr_size - 1) >> hdev->ibh_mr_shift);
- if (ibdev->ibd_mr_shift < 32 || ibdev->ibd_nmrs > 1024) {
+ if (hdev->ibh_mr_shift < 32 || hdev->ibh_nmrs > 1024) {
/* it's 4T..., assume we will re-code at that time */
CERROR("Can't support memory size: x"LPX64
" with MR size: x"LPX64"\n", mm_size, mr_size);
}
/* create an array of MRs to cover all memory */
- LIBCFS_ALLOC(ibdev->ibd_mrs, sizeof(*ibdev->ibd_mrs) * ibdev->ibd_nmrs);
- if (ibdev->ibd_mrs == NULL) {
+ LIBCFS_ALLOC(hdev->ibh_mrs, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
+ if (hdev->ibh_mrs == NULL) {
CERROR("Failed to allocate MRs' table\n");
return -ENOMEM;
}
- memset(ibdev->ibd_mrs, 0, sizeof(*ibdev->ibd_mrs) * ibdev->ibd_nmrs);
+ memset(hdev->ibh_mrs, 0, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
- for (i = 0; i < ibdev->ibd_nmrs; i++) {
+ for (i = 0; i < hdev->ibh_nmrs; i++) {
struct ib_phys_buf ipb;
__u64 iova;
- ipb.size = ibdev->ibd_mr_size;
+ ipb.size = hdev->ibh_mr_size;
ipb.addr = i * mr_size;
iova = ipb.addr;
- mr = ib_reg_phys_mr(ibdev->ibd_pd, &ipb, 1, acflags, &iova);
+ mr = ib_reg_phys_mr(hdev->ibh_pd, &ipb, 1, acflags, &iova);
if (IS_ERR(mr)) {
CERROR("Failed ib_reg_phys_mr addr "LPX64
" size "LPX64" : %ld\n",
ipb.addr, ipb.size, PTR_ERR(mr));
- kiblnd_dev_cleanup(ibdev);
+ kiblnd_hdev_cleanup_mrs(hdev);
return PTR_ERR(mr);
}
LASSERT (iova == ipb.addr);
- ibdev->ibd_mrs[i] = mr;
+ hdev->ibh_mrs[i] = mr;
}
out:
CDEBUG(D_CONSOLE, "Register global MR array, MR size: "
LPX64", array size: %d\n",
- ibdev->ibd_mr_size, ibdev->ibd_nmrs);
+ hdev->ibh_mr_size, hdev->ibh_nmrs);
+ return 0;
+}
- list_add_tail(&ibdev->ibd_list,
- &kiblnd_data.kib_devs);
+static int
+kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
+{ /* DUMMY */
return 0;
}
+static int
+kiblnd_dev_need_failover(kib_dev_t *dev)
+{
+ struct rdma_cm_id *cmid;
+ struct sockaddr_in srcaddr;
+ struct sockaddr_in dstaddr;
+ int rc;
+
+ if (dev->ibd_hdev == NULL || /* initializing */
+ dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
+ *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
+ return 1;
+
+ /* XXX: it's UGLY, but I don't have better way to find
+ * ib-bonding HCA failover because:
+ *
+ * a. no reliable CM event for HCA failover...
+ * b. no OFED API to get ib_device for current net_device...
+ *
+ * We have only two choices at this point:
+ *
+ * a. rdma_bind_addr(), it will conflict with listener cmid
+ * b. rdma_resolve_addr() to zero addr */
+ cmid = rdma_create_id(kiblnd_dummy_callback,
+ dev, RDMA_PS_TCP);
+ if (IS_ERR(cmid)) {
+ rc = PTR_ERR(cmid);
+ CERROR("Failed to create cmid for failover: %d\n", rc);
+ return rc;
+ }
+
+ memset(&srcaddr, 0, sizeof(srcaddr));
+ srcaddr.sin_family = AF_INET;
+ srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
+
+ memset(&dstaddr, 0, sizeof(dstaddr));
+ dstaddr.sin_family = AF_INET;
+ rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
+ (struct sockaddr *)&dstaddr, 1);
+ if (rc != 0) {
+ CERROR("Failed to bind %s to device: %d\n",
+ dev->ibd_ifname, rc);
+ rdma_destroy_id(cmid);
+ return rc;
+ }
+
+ LASSERT (cmid->device != NULL);
+ if (dev->ibd_hdev->ibh_ibdev == cmid->device) {
+ /* don't need device failover */
+ rdma_destroy_id(cmid);
+ return 0;
+ }
+
+ return 1;
+}
+
+int
+kiblnd_dev_failover(kib_dev_t *dev)
+{
+ CFS_LIST_HEAD (zombie_tpo);
+ CFS_LIST_HEAD (zombie_ppo);
+ CFS_LIST_HEAD (zombie_fpo);
+ struct rdma_cm_id *cmid = NULL;
+ kib_hca_dev_t *hdev = NULL;
+ kib_hca_dev_t *old;
+ struct ib_pd *pd;
+ kib_net_t *net;
+ struct sockaddr_in addr;
+ unsigned long flags;
+ int rc = 0;
+
+ LASSERT (*kiblnd_tunables.kib_dev_failover > 1 ||
+ dev->ibd_can_failover ||
+ dev->ibd_hdev == NULL);
+
+ rc = kiblnd_dev_need_failover(dev);
+ if (rc <= 0)
+ goto out;
+
+ if (dev->ibd_hdev != NULL &&
+ dev->ibd_hdev->ibh_cmid != NULL) {
+ /* XXX it's not good to close old listener at here,
+ * because we can fail to create new listener.
+ * But we have to close it now, otherwise rdma_bind_addr
+ * will return EADDRINUSE... How crap! */
+ cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
+
+ cmid = dev->ibd_hdev->ibh_cmid;
+ /* make next schedule of kiblnd_dev_need_failover
+ * will return 1 for me */
+ dev->ibd_hdev->ibh_cmid = NULL;
+ cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
+
+ rdma_destroy_id(cmid);
+ }
+
+ cmid = rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP);
+ if (IS_ERR(cmid)) {
+ rc = PTR_ERR(cmid);
+ CERROR("Failed to create cmid for failover: %d\n", rc);
+ goto out;
+ }
+
+ memset(&addr, 0, sizeof(addr));
+ addr.sin_family = AF_INET;
+ addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
+ addr.sin_port = htons(*kiblnd_tunables.kib_service);
+
+ /* Bind to failover device or port */
+ rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
+ if (rc != 0) {
+ CERROR("Failed to bind %s to device: %d\n",
+ dev->ibd_ifname, rc);
+ rdma_destroy_id(cmid);
+ goto out;
+ }
+
+ LIBCFS_ALLOC(hdev, sizeof(*hdev));
+ if (hdev == NULL) {
+ CERROR("Failed to allocate kib_hca_dev\n");
+ rdma_destroy_id(cmid);
+ rc = -ENOMEM;
+ goto out;
+ }
+ memset(hdev, 0, sizeof(*hdev));
+ atomic_set(&hdev->ibh_ref, 1);
+ hdev->ibh_dev = dev;
+ hdev->ibh_cmid = cmid;
+ hdev->ibh_ibdev = cmid->device;
+
+ pd = ib_alloc_pd(cmid->device);
+ if (IS_ERR(pd)) {
+ rc = PTR_ERR(pd);
+ CERROR("Can't allocate PD: %d\n", rc);
+ goto out;
+ }
+
+ hdev->ibh_pd = pd;
+
+ rc = rdma_listen(cmid, 0);
+ if (rc != 0) {
+ CERROR("Can't start new listener: %d\n", rc);
+ goto out;
+ }
+
+ rc = kiblnd_hdev_setup_mrs(hdev);
+ if (rc != 0) {
+ CERROR("Can't setup device: %d\n", rc);
+ goto out;
+ }
+
+ cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
+
+ old = dev->ibd_hdev;
+ dev->ibd_hdev = hdev; /* take over the refcount */
+ hdev = old;
+
+ cfs_list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
+ kiblnd_fail_pool_set(&net->ibn_tx_ps.tps_poolset, &zombie_tpo);
+ if (net->ibn_with_pmr)
+ kiblnd_fail_pool_set(&net->ibn_pmr_ps.pps_poolset, &zombie_ppo);
+ if (net->ibn_with_fmr)
+ kiblnd_fail_fmr_pool_set(&net->ibn_fmr_ps, &zombie_fpo);
+ }
+
+ cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
+ 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;
+
+ return rc;
+}
+
void
kiblnd_destroy_dev (kib_dev_t *dev)
{
LASSERT (dev->ibd_nnets == 0);
+ LASSERT (cfs_list_empty(&dev->ibd_nets));
- if (!list_empty(&dev->ibd_list)) /* on kib_devs? */
- list_del_init(&dev->ibd_list);
+ cfs_list_del(&dev->ibd_fail_list);
+ cfs_list_del(&dev->ibd_list);
- kiblnd_dev_cleanup(dev);
+ if (dev->ibd_hdev != NULL)
+ kiblnd_hdev_decref(dev->ibd_hdev);
- if (dev->ibd_pd != NULL)
- ib_dealloc_pd(dev->ibd_pd);
+ LIBCFS_FREE(dev, sizeof(*dev));
+}
- if (dev->ibd_cmid != NULL)
- rdma_destroy_id(dev->ibd_cmid);
+kib_dev_t *
+kiblnd_create_dev(char *ifname)
+{
+ struct net_device *netdev;
+ kib_dev_t *dev;
+ __u32 netmask;
+ __u32 ip;
+ int up;
+ int rc;
- LIBCFS_FREE(dev, sizeof(*dev));
+ rc = libcfs_ipif_query(ifname, &up, &ip, &netmask);
+ if (rc != 0) {
+ CERROR("Can't query IPoIB interface %s: %d\n",
+ ifname, rc);
+ return NULL;
+ }
+
+ if (!up) {
+ CERROR("Can't query IPoIB interface %s: it's down\n", ifname);
+ return NULL;
+ }
+
+ LIBCFS_ALLOC(dev, sizeof(*dev));
+ if (dev == NULL)
+ return NULL;
+
+ memset(dev, 0, sizeof(*dev));
+#ifdef HAVE_DEV_GET_BY_NAME_2ARG
+ if ((netdev = dev_get_by_name(&init_net, ifname)) == NULL) {
+#else
+ if ((netdev = dev_get_by_name(ifname)) == NULL) {
+#endif
+ CERROR("Can't find netdev of IF: %s\n", ifname);
+ LIBCFS_FREE(dev, sizeof(*dev));
+ return NULL;
+ }
+
+ 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
{
int i;
- LASSERT (list_empty(&kiblnd_data.kib_devs));
+ LASSERT (cfs_list_empty(&kiblnd_data.kib_devs));
CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
- atomic_read(&libcfs_kmemory));
+ cfs_atomic_read(&libcfs_kmemory));
switch (kiblnd_data.kib_init) {
default:
case IBLND_INIT_DATA:
LASSERT (kiblnd_data.kib_peers != NULL);
for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
- LASSERT (list_empty(&kiblnd_data.kib_peers[i]));
+ LASSERT (cfs_list_empty(&kiblnd_data.kib_peers[i]));
}
- LASSERT (list_empty(&kiblnd_data.kib_connd_zombies));
- LASSERT (list_empty(&kiblnd_data.kib_connd_conns));
+ LASSERT (cfs_list_empty(&kiblnd_data.kib_connd_zombies));
+ LASSERT (cfs_list_empty(&kiblnd_data.kib_connd_conns));
/* flag threads to terminate; wake and wait for them to die */
kiblnd_data.kib_shutdown = 1;
- wake_up_all(&kiblnd_data.kib_sched_waitq);
- wake_up_all(&kiblnd_data.kib_connd_waitq);
+ cfs_waitq_broadcast(&kiblnd_data.kib_sched_waitq);
+ cfs_waitq_broadcast(&kiblnd_data.kib_connd_waitq);
+ cfs_waitq_broadcast(&kiblnd_data.kib_failover_waitq);
i = 2;
- while (atomic_read(&kiblnd_data.kib_nthreads) != 0) {
+ 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",
- atomic_read(&kiblnd_data.kib_nthreads));
+ cfs_atomic_read(&kiblnd_data.kib_nthreads));
cfs_pause(cfs_time_seconds(1));
}
if (kiblnd_data.kib_peers != NULL)
LIBCFS_FREE(kiblnd_data.kib_peers,
- sizeof(struct list_head) *
+ sizeof(cfs_list_t) *
kiblnd_data.kib_peer_hash_size);
CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
- atomic_read(&libcfs_kmemory));
+ cfs_atomic_read(&libcfs_kmemory));
kiblnd_data.kib_init = IBLND_INIT_NOTHING;
PORTAL_MODULE_UNUSE;
kiblnd_shutdown (lnet_ni_t *ni)
{
kib_net_t *net = ni->ni_data;
- rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
+ cfs_rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
int i;
unsigned long flags;
goto out;
CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
- atomic_read(&libcfs_kmemory));
+ cfs_atomic_read(&libcfs_kmemory));
- write_lock_irqsave(g_lock, flags);
+ cfs_write_lock_irqsave(g_lock, flags);
net->ibn_shutdown = 1;
- write_unlock_irqrestore(g_lock, flags);
+ cfs_write_unlock_irqrestore(g_lock, flags);
switch (net->ibn_init) {
default:
/* Wait for all peer state to clean up */
i = 2;
- while (atomic_read(&net->ibn_npeers) != 0) {
+ 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),
- atomic_read(&net->ibn_npeers));
+ cfs_atomic_read(&net->ibn_npeers));
cfs_pause(cfs_time_seconds(1));
}
- kiblnd_unmap_tx_descs(ni);
+ kiblnd_ni_fini_pools(net);
+ cfs_write_lock_irqsave(g_lock, flags);
LASSERT (net->ibn_dev->ibd_nnets > 0);
net->ibn_dev->ibd_nnets--;
+ cfs_list_del(&net->ibn_list);
+ cfs_write_unlock_irqrestore(g_lock, flags);
/* fall through */
case IBLND_INIT_NOTHING:
- LASSERT (atomic_read(&net->ibn_nconns) == 0);
-
- if (net->ibn_fmrpool != NULL)
- ib_destroy_fmr_pool(net->ibn_fmrpool);
- if (net->ibn_pmrpool != NULL)
- kiblnd_destroy_pmr_pool(net->ibn_pmrpool);
+ LASSERT (cfs_atomic_read(&net->ibn_nconns) == 0);
if (net->ibn_dev != NULL &&
net->ibn_dev->ibd_nnets == 0)
break;
}
- kiblnd_free_tx_descs(ni);
-
CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
- atomic_read(&libcfs_kmemory));
+ cfs_atomic_read(&libcfs_kmemory));
net->ibn_init = IBLND_INIT_NOTHING;
ni->ni_data = NULL;
-
+
LIBCFS_FREE(net, sizeof(*net));
out:
- if (list_empty(&kiblnd_data.kib_devs))
+ if (cfs_list_empty(&kiblnd_data.kib_devs))
kiblnd_base_shutdown();
return;
}
PORTAL_MODULE_USE;
memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
- rwlock_init(&kiblnd_data.kib_global_lock);
+ cfs_rwlock_init(&kiblnd_data.kib_global_lock);
- INIT_LIST_HEAD(&kiblnd_data.kib_devs);
+ CFS_INIT_LIST_HEAD(&kiblnd_data.kib_devs);
+ CFS_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(struct list_head) * kiblnd_data.kib_peer_hash_size);
+ 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++)
- INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
+ CFS_INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
- spin_lock_init(&kiblnd_data.kib_connd_lock);
- INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
- INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
- init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
+ cfs_spin_lock_init(&kiblnd_data.kib_connd_lock);
+ CFS_INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
+ CFS_INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
+ cfs_waitq_init(&kiblnd_data.kib_connd_waitq);
- spin_lock_init(&kiblnd_data.kib_sched_lock);
- INIT_LIST_HEAD(&kiblnd_data.kib_sched_conns);
- init_waitqueue_head(&kiblnd_data.kib_sched_waitq);
+ cfs_spin_lock_init(&kiblnd_data.kib_sched_lock);
+ CFS_INIT_LIST_HEAD(&kiblnd_data.kib_sched_conns);
+ cfs_waitq_init(&kiblnd_data.kib_sched_waitq);
+ cfs_waitq_init(&kiblnd_data.kib_failover_waitq);
kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
goto failed;
}
+ if (*kiblnd_tunables.kib_dev_failover != 0)
+ rc = kiblnd_thread_start(kiblnd_failover_thread, NULL);
+
+ if (rc != 0) {
+ CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
+ goto failed;
+ }
+
/* flag everything initialised */
kiblnd_data.kib_init = IBLND_INIT_ALL;
/*****************************************************/
char *ifname;
kib_dev_t *ibdev = NULL;
kib_net_t *net;
- struct list_head *tmp;
+ cfs_list_t *tmp;
struct timeval tv;
+ unsigned long flags;
int rc;
LASSERT (ni->ni_lnd == &the_o2iblnd);
memset(net, 0, sizeof(*net));
- do_gettimeofday(&tv);
+ cfs_gettimeofday(&tv);
net->ibn_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
ni->ni_peertimeout = *kiblnd_tunables.kib_peertimeout;
ni->ni_peertxcredits = *kiblnd_tunables.kib_peertxcredits;
ni->ni_peerrtrcredits = *kiblnd_tunables.kib_peerrtrcredits;
- spin_lock_init(&net->ibn_tx_lock);
- INIT_LIST_HEAD(&net->ibn_idle_txs);
-
if (ni->ni_interfaces[0] != NULL) {
/* Use the IPoIB interface specified in 'networks=' */
goto failed;
}
- list_for_each (tmp, &kiblnd_data.kib_devs) {
- ibdev = list_entry(tmp, kib_dev_t, ibd_list);
+ cfs_list_for_each (tmp, &kiblnd_data.kib_devs) {
+ ibdev = cfs_list_entry(tmp, kib_dev_t, ibd_list);
if (!strcmp(&ibdev->ibd_ifname[0], ifname))
break;
ibdev = NULL;
}
- if (ibdev == NULL) {
- __u32 ip;
- __u32 netmask;
- int up;
- struct rdma_cm_id *id;
- struct ib_pd *pd;
- struct sockaddr_in addr;
-
- rc = libcfs_ipif_query(ifname, &up, &ip, &netmask);
- if (rc != 0) {
- CERROR("Can't query IPoIB interface %s: %d\n",
- ifname, rc);
- goto failed;
- }
-
- if (!up) {
- CERROR("Can't query IPoIB interface %s: it's down\n",
- ifname);
- goto failed;
- }
-
- LIBCFS_ALLOC(ibdev, sizeof(*ibdev));
- if (ibdev == NULL)
- goto failed;
-
- memset(ibdev, 0, sizeof(*ibdev));
-
- CFS_INIT_LIST_HEAD(&ibdev->ibd_list); /* not yet in kib_devs */
- ibdev->ibd_ifip = ip;
- strcpy(&ibdev->ibd_ifname[0], ifname);
-
- id = rdma_create_id(kiblnd_cm_callback, ibdev, RDMA_PS_TCP);
- if (IS_ERR(id)) {
- CERROR("Can't create listen ID: %ld\n", PTR_ERR(id));
- goto failed;
- }
-
- ibdev->ibd_cmid = id;
-
- memset(&addr, 0, sizeof(addr));
- addr.sin_family = AF_INET;
- addr.sin_port = htons(*kiblnd_tunables.kib_service);
- addr.sin_addr.s_addr = htonl(ip);
+ if (ibdev == NULL)
+ ibdev = kiblnd_create_dev(ifname);
- rc = rdma_bind_addr(id, (struct sockaddr *)&addr);
- if (rc != 0) {
- CERROR("Can't bind to %s: %d\n", ifname, rc);
- goto failed;
- }
-
- /* Binding should have assigned me an IB device */
- LASSERT (id->device != NULL);
-
- pd = ib_alloc_pd(id->device);
- if (IS_ERR(pd)) {
- CERROR("Can't allocate PD: %ld\n", PTR_ERR(pd));
- goto failed;
- }
-
- ibdev->ibd_pd = pd;
-
- rc = rdma_listen(id, 256);
- if (rc != 0) {
- CERROR("Can't start listener: %d\n", rc);
- goto failed;
- }
-
- rc = kiblnd_dev_setup(ibdev);
- if (rc != 0) {
- CERROR("Can't setup device: %d\n", rc);
- goto failed;
- }
- }
+ if (ibdev == NULL)
+ goto failed;
- ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
net->ibn_dev = ibdev;
+ ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
- if (*kiblnd_tunables.kib_map_on_demand > 0 ||
- ibdev->ibd_nmrs > 1) { /* premapping can fail if ibd_nmr > 1,
- * so we always create FMR/PMR pool and
- * map-on-demand if premapping failed */
- /* Map on demand */
- rc = kiblnd_ib_create_fmr_pool(ibdev, &net->ibn_fmrpool);
- if (rc == -ENOSYS) {
- CDEBUG(D_CONSOLE, "No FMR, creating physical mapping\n");
-
- rc = kiblnd_create_pmr_pool(ibdev, &net->ibn_pmrpool);
- }
-
- if (rc != 0) {
- CERROR("Can't create FMR or physical mapping pool: %d, "
- "please disable map_on_demand and retry\n", rc);
- goto failed;
- }
-
- }
-
- rc = kiblnd_alloc_tx_descs(ni);
+ rc = kiblnd_net_init_pools(net);
if (rc != 0) {
- CERROR("Can't allocate tx descs\n");
+ CERROR("Failed to initialize NI pools: %d\n", rc);
goto failed;
}
- kiblnd_map_tx_descs(ni);
-
+ cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
ibdev->ibd_nnets++;
+ cfs_list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
+ cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
+
net->ibn_init = IBLND_INIT_ALL;
return 0;