X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lnet%2Fklnds%2Fo2iblnd%2Fo2iblnd.c;h=17f0596fd5f6913746e068736c65e62ed98b9968;hb=d8792a7dab933def57f6069296234ad48ea0da09;hp=6986f2d010d8834a6274b6d81b62dd4caf852b52;hpb=dc2e27a6c3ae7d67ffbccfd07c975a08583f1399;p=fs%2Flustre-release.git
diff --git a/lnet/klnds/o2iblnd/o2iblnd.c b/lnet/klnds/o2iblnd/o2iblnd.c
index 6986f2d..17f0596 100644
--- a/lnet/klnds/o2iblnd/o2iblnd.c
+++ b/lnet/klnds/o2iblnd/o2iblnd.c
@@ -15,11 +15,7 @@
*
* You should have received a copy of the GNU General Public License
* version 2 along with this program; If not, see
- * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
- *
- * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
- * CA 95054 USA or visit www.sun.com if you need additional information or
- * have any questions.
+ * http://www.gnu.org/licenses/gpl-2.0.html
*
* GPL HEADER END
*/
@@ -27,7 +23,7 @@
* Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
* Use is subject to license terms.
*
- * Copyright (c) 2011, 2012, Intel Corporation.
+ * Copyright (c) 2011, 2016, Intel Corporation.
*/
/*
* This file is part of Lustre, http://www.lustre.org/
@@ -259,7 +255,7 @@ kiblnd_unpack_msg(kib_msg_t *msg, int nob)
msg->ibm_cksum = msg_cksum;
if (flip) {
- /* leave magic unflipped as a clue to peer endianness */
+ /* leave magic unflipped as a clue to peer_ni endianness */
msg->ibm_version = version;
CLASSERT (sizeof(msg->ibm_type) == 1);
CLASSERT (sizeof(msg->ibm_credits) == 1);
@@ -317,31 +313,33 @@ kiblnd_unpack_msg(kib_msg_t *msg, int nob)
}
int
-kiblnd_create_peer(lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid)
+kiblnd_create_peer(lnet_ni_t *ni, kib_peer_ni_t **peerp, lnet_nid_t nid)
{
- kib_peer_t *peer;
+ kib_peer_ni_t *peer_ni;
kib_net_t *net = ni->ni_data;
- int cpt = lnet_cpt_of_nid(nid);
+ int cpt = lnet_cpt_of_nid(nid, ni);
unsigned long flags;
LASSERT(net != NULL);
LASSERT(nid != LNET_NID_ANY);
- LIBCFS_CPT_ALLOC(peer, lnet_cpt_table(), cpt, sizeof(*peer));
- if (peer == NULL) {
- CERROR("Cannot allocate peer\n");
+ LIBCFS_CPT_ALLOC(peer_ni, lnet_cpt_table(), cpt, sizeof(*peer_ni));
+ if (peer_ni == NULL) {
+ CERROR("Cannot allocate peer_ni\n");
return -ENOMEM;
}
- peer->ibp_ni = ni;
- peer->ibp_nid = nid;
- peer->ibp_error = 0;
- peer->ibp_last_alive = 0;
- atomic_set(&peer->ibp_refcount, 1); /* 1 ref for caller */
+ peer_ni->ibp_ni = ni;
+ peer_ni->ibp_nid = nid;
+ peer_ni->ibp_error = 0;
+ peer_ni->ibp_last_alive = 0;
+ peer_ni->ibp_max_frags = kiblnd_cfg_rdma_frags(peer_ni->ibp_ni);
+ peer_ni->ibp_queue_depth = ni->ni_net->net_tunables.lct_peer_tx_credits;
+ atomic_set(&peer_ni->ibp_refcount, 1); /* 1 ref for caller */
- 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);
+ INIT_LIST_HEAD(&peer_ni->ibp_list); /* not in the peer_ni table yet */
+ INIT_LIST_HEAD(&peer_ni->ibp_conns);
+ INIT_LIST_HEAD(&peer_ni->ibp_tx_queue);
write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
@@ -353,77 +351,79 @@ kiblnd_create_peer(lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid)
write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
- *peerp = peer;
+ *peerp = peer_ni;
return 0;
}
void
-kiblnd_destroy_peer (kib_peer_t *peer)
+kiblnd_destroy_peer (kib_peer_ni_t *peer_ni)
{
- kib_net_t *net = peer->ibp_ni->ni_data;
+ kib_net_t *net = peer_ni->ibp_ni->ni_data;
LASSERT(net != NULL);
- LASSERT (atomic_read(&peer->ibp_refcount) == 0);
- LASSERT(!kiblnd_peer_active(peer));
- LASSERT(peer->ibp_connecting == 0);
- LASSERT(peer->ibp_accepting == 0);
- LASSERT(list_empty(&peer->ibp_conns));
- LASSERT(list_empty(&peer->ibp_tx_queue));
+ LASSERT (atomic_read(&peer_ni->ibp_refcount) == 0);
+ LASSERT(!kiblnd_peer_active(peer_ni));
+ LASSERT(kiblnd_peer_idle(peer_ni));
+ LASSERT(list_empty(&peer_ni->ibp_tx_queue));
- LIBCFS_FREE(peer, sizeof(*peer));
+ LIBCFS_FREE(peer_ni, sizeof(*peer_ni));
- /* NB a peer's connections keep a reference on their peer until
+ /* NB a peer_ni's connections keep a reference on their peer_ni until
* they are destroyed, so we can be assured that _all_ state to do
- * with this peer has been cleaned up when its refcount drops to
+ * with this peer_ni has been cleaned up when its refcount drops to
* zero. */
atomic_dec(&net->ibn_npeers);
}
-kib_peer_t *
-kiblnd_find_peer_locked (lnet_nid_t nid)
+kib_peer_ni_t *
+kiblnd_find_peer_locked(struct lnet_ni *ni, lnet_nid_t nid)
{
/* the caller is responsible for accounting the additional reference
* that this creates */
struct list_head *peer_list = kiblnd_nid2peerlist(nid);
struct list_head *tmp;
- kib_peer_t *peer;
+ kib_peer_ni_t *peer_ni;
list_for_each(tmp, peer_list) {
- peer = 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 */
-
- if (peer->ibp_nid != nid)
+ peer_ni = list_entry(tmp, kib_peer_ni_t, ibp_list);
+ LASSERT(!kiblnd_peer_idle(peer_ni));
+
+ /*
+ * Match a peer if its NID and the NID of the local NI it
+ * communicates over are the same. Otherwise don't match
+ * the peer, which will result in a new lnd peer being
+ * created.
+ */
+ if (peer_ni->ibp_nid != nid ||
+ peer_ni->ibp_ni->ni_nid != ni->ni_nid)
continue;
- CDEBUG(D_NET, "got peer [%p] -> %s (%d) version: %x\n",
- peer, libcfs_nid2str(nid),
- atomic_read(&peer->ibp_refcount),
- peer->ibp_version);
- return peer;
+ CDEBUG(D_NET, "got peer_ni [%p] -> %s (%d) version: %x\n",
+ peer_ni, libcfs_nid2str(nid),
+ atomic_read(&peer_ni->ibp_refcount),
+ peer_ni->ibp_version);
+ return peer_ni;
}
return NULL;
}
void
-kiblnd_unlink_peer_locked (kib_peer_t *peer)
+kiblnd_unlink_peer_locked (kib_peer_ni_t *peer_ni)
{
- LASSERT(list_empty(&peer->ibp_conns));
+ LASSERT(list_empty(&peer_ni->ibp_conns));
- LASSERT (kiblnd_peer_active(peer));
- list_del_init(&peer->ibp_list);
+ LASSERT (kiblnd_peer_active(peer_ni));
+ list_del_init(&peer_ni->ibp_list);
/* lose peerlist's ref */
- kiblnd_peer_decref(peer);
+ kiblnd_peer_decref(peer_ni);
}
static int
kiblnd_get_peer_info(lnet_ni_t *ni, int index,
lnet_nid_t *nidp, int *count)
{
- kib_peer_t *peer;
+ kib_peer_ni_t *peer_ni;
struct list_head *ptmp;
int i;
unsigned long flags;
@@ -434,19 +434,17 @@ kiblnd_get_peer_info(lnet_ni_t *ni, int index,
list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
- peer = list_entry(ptmp, kib_peer_t, ibp_list);
- LASSERT(peer->ibp_connecting > 0 ||
- peer->ibp_accepting > 0 ||
- !list_empty(&peer->ibp_conns));
+ peer_ni = list_entry(ptmp, kib_peer_ni_t, ibp_list);
+ LASSERT(!kiblnd_peer_idle(peer_ni));
- if (peer->ibp_ni != ni)
+ if (peer_ni->ibp_ni != ni)
continue;
if (index-- > 0)
continue;
- *nidp = peer->ibp_nid;
- *count = atomic_read(&peer->ibp_refcount);
+ *nidp = peer_ni->ibp_nid;
+ *count = atomic_read(&peer_ni->ibp_refcount);
read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
flags);
@@ -459,23 +457,23 @@ kiblnd_get_peer_info(lnet_ni_t *ni, int index,
}
static void
-kiblnd_del_peer_locked (kib_peer_t *peer)
+kiblnd_del_peer_locked (kib_peer_ni_t *peer_ni)
{
struct list_head *ctmp;
struct list_head *cnxt;
kib_conn_t *conn;
- if (list_empty(&peer->ibp_conns)) {
- kiblnd_unlink_peer_locked(peer);
+ if (list_empty(&peer_ni->ibp_conns)) {
+ kiblnd_unlink_peer_locked(peer_ni);
} else {
- list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
+ list_for_each_safe(ctmp, cnxt, &peer_ni->ibp_conns) {
conn = list_entry(ctmp, kib_conn_t, ibc_list);
kiblnd_close_conn_locked(conn, 0);
}
- /* NB closing peer's last conn unlinked it. */
+ /* NB closing peer_ni's last conn unlinked it. */
}
- /* NB peer now unlinked; might even be freed if the peer table had the
+ /* NB peer_ni now unlinked; might even be freed if the peer_ni table had the
* last ref on it. */
}
@@ -485,7 +483,7 @@ kiblnd_del_peer (lnet_ni_t *ni, lnet_nid_t nid)
struct list_head zombies = LIST_HEAD_INIT(zombies);
struct list_head *ptmp;
struct list_head *pnxt;
- kib_peer_t *peer;
+ kib_peer_ni_t *peer_ni;
int lo;
int hi;
int i;
@@ -503,25 +501,23 @@ kiblnd_del_peer (lnet_ni_t *ni, lnet_nid_t nid)
for (i = lo; i <= hi; i++) {
list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
- peer = list_entry(ptmp, kib_peer_t, ibp_list);
- LASSERT(peer->ibp_connecting > 0 ||
- peer->ibp_accepting > 0 ||
- !list_empty(&peer->ibp_conns));
+ peer_ni = list_entry(ptmp, kib_peer_ni_t, ibp_list);
+ LASSERT(!kiblnd_peer_idle(peer_ni));
- if (peer->ibp_ni != ni)
+ if (peer_ni->ibp_ni != ni)
continue;
- if (!(nid == LNET_NID_ANY || peer->ibp_nid == nid))
+ if (!(nid == LNET_NID_ANY || peer_ni->ibp_nid == nid))
continue;
- if (!list_empty(&peer->ibp_tx_queue)) {
- LASSERT(list_empty(&peer->ibp_conns));
+ if (!list_empty(&peer_ni->ibp_tx_queue)) {
+ LASSERT(list_empty(&peer_ni->ibp_conns));
- list_splice_init(&peer->ibp_tx_queue,
+ list_splice_init(&peer_ni->ibp_tx_queue,
&zombies);
}
- kiblnd_del_peer_locked(peer);
+ kiblnd_del_peer_locked(peer_ni);
rc = 0; /* matched something */
}
}
@@ -536,7 +532,7 @@ kiblnd_del_peer (lnet_ni_t *ni, lnet_nid_t nid)
static kib_conn_t *
kiblnd_get_conn_by_idx(lnet_ni_t *ni, int index)
{
- kib_peer_t *peer;
+ kib_peer_ni_t *peer_ni;
struct list_head *ptmp;
kib_conn_t *conn;
struct list_head *ctmp;
@@ -548,15 +544,13 @@ kiblnd_get_conn_by_idx(lnet_ni_t *ni, int index)
for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
- peer = list_entry(ptmp, kib_peer_t, ibp_list);
- LASSERT(peer->ibp_connecting > 0 ||
- peer->ibp_accepting > 0 ||
- !list_empty(&peer->ibp_conns));
+ peer_ni = list_entry(ptmp, kib_peer_ni_t, ibp_list);
+ LASSERT(!kiblnd_peer_idle(peer_ni));
- if (peer->ibp_ni != ni)
+ if (peer_ni->ibp_ni != ni)
continue;
- list_for_each(ctmp, &peer->ibp_conns) {
+ list_for_each(ctmp, &peer_ni->ibp_conns) {
if (index-- > 0)
continue;
@@ -585,7 +579,7 @@ static void
kiblnd_debug_tx (kib_tx_t *tx)
{
CDEBUG(D_CONSOLE, " %p snd %d q %d w %d rc %d dl %lx "
- "cookie "LPX64" msg %s%s type %x cred %d\n",
+ "cookie %#llx msg %s%s type %x cred %d\n",
tx, tx->tx_sending, tx->tx_queued, tx->tx_waiting,
tx->tx_status, tx->tx_deadline, tx->tx_cookie,
tx->tx_lntmsg[0] == NULL ? "-" : "!",
@@ -635,7 +629,7 @@ kiblnd_debug_conn (kib_conn_t *conn)
kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
CDEBUG(D_CONSOLE, " rxs:\n");
- for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++)
+ for (i = 0; i < IBLND_RX_MSGS(conn); i++)
kiblnd_debug_rx(&conn->ibc_rxs[i]);
spin_unlock(&conn->ibc_lock);
@@ -684,6 +678,7 @@ kiblnd_get_completion_vector(kib_conn_t *conn, int cpt)
int vectors;
int off;
int i;
+ lnet_nid_t ibp_nid;
vectors = conn->ibc_cmid->device->num_comp_vectors;
if (vectors <= 1)
@@ -692,7 +687,8 @@ kiblnd_get_completion_vector(kib_conn_t *conn, int cpt)
mask = cfs_cpt_cpumask(lnet_cpt_table(), cpt);
/* hash NID to CPU id in this partition... */
- off = conn->ibc_peer->ibp_nid % cpumask_weight(mask);
+ ibp_nid = conn->ibc_peer->ibp_nid;
+ off = do_div(ibp_nid, cpumask_weight(mask));
for_each_cpu(i, mask) {
if (off-- == 0)
return i % vectors;
@@ -703,21 +699,24 @@ kiblnd_get_completion_vector(kib_conn_t *conn, int cpt)
}
kib_conn_t *
-kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
- int state, int version)
+kiblnd_create_conn(kib_peer_ni_t *peer_ni, struct rdma_cm_id *cmid,
+ int state, int version)
{
/* CAVEAT EMPTOR:
* If the new conn is created successfully it takes over the caller's
- * ref on 'peer'. It also "owns" 'cmid' and destroys it when it itself
- * is destroyed. On failure, the caller's ref on 'peer' remains and
+ * ref on 'peer_ni'. It also "owns" 'cmid' and destroys it when it itself
+ * is destroyed. On failure, the caller's ref on 'peer_ni' remains and
* she must dispose of 'cmid'. (Actually I'd block forever if I tried
* to destroy 'cmid' here since I'm called from the CM which still has
* its ref on 'cmid'). */
rwlock_t *glock = &kiblnd_data.kib_global_lock;
- kib_net_t *net = peer->ibp_ni->ni_data;
+ kib_net_t *net = peer_ni->ibp_ni->ni_data;
kib_dev_t *dev;
struct ib_qp_init_attr *init_qp_attr;
struct kib_sched_info *sched;
+#ifdef HAVE_IB_CQ_INIT_ATTR
+ struct ib_cq_init_attr cq_attr = {};
+#endif
kib_conn_t *conn;
struct ib_cq *cq;
unsigned long flags;
@@ -730,7 +729,7 @@ kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
dev = net->ibn_dev;
- cpt = lnet_cpt_of_nid(peer->ibp_nid);
+ cpt = lnet_cpt_of_nid(peer_ni->ibp_nid, peer_ni->ibp_ni);
sched = kiblnd_data.kib_scheds[cpt];
LASSERT(sched->ibs_nthreads > 0);
@@ -739,22 +738,24 @@ kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
sizeof(*init_qp_attr));
if (init_qp_attr == NULL) {
CERROR("Can't allocate qp_attr for %s\n",
- libcfs_nid2str(peer->ibp_nid));
+ libcfs_nid2str(peer_ni->ibp_nid));
goto failed_0;
}
LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
if (conn == NULL) {
CERROR("Can't allocate connection for %s\n",
- libcfs_nid2str(peer->ibp_nid));
+ libcfs_nid2str(peer_ni->ibp_nid));
goto failed_1;
}
conn->ibc_state = IBLND_CONN_INIT;
conn->ibc_version = version;
- conn->ibc_peer = peer; /* I take the caller's ref */
+ conn->ibc_peer = peer_ni; /* I take the caller's ref */
cmid->context = conn; /* for future CM callbacks */
conn->ibc_cmid = cmid;
+ conn->ibc_max_frags = peer_ni->ibp_max_frags;
+ conn->ibc_queue_depth = peer_ni->ibp_queue_depth;
INIT_LIST_HEAD(&conn->ibc_early_rxs);
INIT_LIST_HEAD(&conn->ibc_tx_noops);
@@ -800,28 +801,36 @@ kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
write_unlock_irqrestore(glock, flags);
LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
- IBLND_RX_MSGS(version) * sizeof(kib_rx_t));
+ IBLND_RX_MSGS(conn) * sizeof(kib_rx_t));
if (conn->ibc_rxs == NULL) {
CERROR("Cannot allocate RX buffers\n");
goto failed_2;
}
rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
- IBLND_RX_MSG_PAGES(version));
+ IBLND_RX_MSG_PAGES(conn));
if (rc != 0)
goto failed_2;
kiblnd_map_rx_descs(conn);
+#ifdef HAVE_IB_CQ_INIT_ATTR
+ cq_attr.cqe = IBLND_CQ_ENTRIES(conn);
+ cq_attr.comp_vector = kiblnd_get_completion_vector(conn, cpt);
+ cq = ib_create_cq(cmid->device,
+ kiblnd_cq_completion, kiblnd_cq_event, conn,
+ &cq_attr);
+#else
cq = ib_create_cq(cmid->device,
kiblnd_cq_completion, kiblnd_cq_event, conn,
- IBLND_CQ_ENTRIES(version),
+ IBLND_CQ_ENTRIES(conn),
kiblnd_get_completion_vector(conn, cpt));
- if (IS_ERR(cq)) {
- CERROR("Can't create CQ: %ld, cqe: %d\n",
- PTR_ERR(cq), IBLND_CQ_ENTRIES(version));
- goto failed_2;
- }
+#endif
+ if (IS_ERR(cq)) {
+ CERROR("Failed to create CQ with %d CQEs: %ld\n",
+ IBLND_CQ_ENTRIES(conn), PTR_ERR(cq));
+ goto failed_2;
+ }
conn->ibc_cq = cq;
@@ -833,8 +842,8 @@ kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
init_qp_attr->event_handler = kiblnd_qp_event;
init_qp_attr->qp_context = conn;
- init_qp_attr->cap.max_send_wr = IBLND_SEND_WRS(version);
- init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(version);
+ init_qp_attr->cap.max_send_wr = IBLND_SEND_WRS(conn);
+ init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(conn);
init_qp_attr->cap.max_send_sge = 1;
init_qp_attr->cap.max_recv_sge = 1;
init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
@@ -844,34 +853,44 @@ kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
conn->ibc_sched = sched;
- rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr);
- if (rc != 0) {
- CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d\n",
- rc, init_qp_attr->cap.max_send_wr,
- init_qp_attr->cap.max_recv_wr);
- goto failed_2;
- }
+ do {
+ rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr);
+ if (!rc || init_qp_attr->cap.max_send_wr < 16)
+ break;
- LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
+ init_qp_attr->cap.max_send_wr -= init_qp_attr->cap.max_send_wr / 4;
+ } while (rc);
- /* 1 ref for caller and each rxmsg */
- atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(version));
- conn->ibc_nrx = IBLND_RX_MSGS(version);
+ if (rc) {
+ CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d\n",
+ rc, init_qp_attr->cap.max_send_wr,
+ init_qp_attr->cap.max_recv_wr);
+ goto failed_2;
+ }
- /* post receives */
- for (i = 0; i < IBLND_RX_MSGS(version); i++) {
- rc = kiblnd_post_rx(&conn->ibc_rxs[i],
- IBLND_POSTRX_NO_CREDIT);
- if (rc != 0) {
- CERROR("Can't post rxmsg: %d\n", rc);
+ if (init_qp_attr->cap.max_send_wr != IBLND_SEND_WRS(conn))
+ CDEBUG(D_NET, "original send wr %d, created with %d\n",
+ IBLND_SEND_WRS(conn), 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(conn));
+ conn->ibc_nrx = IBLND_RX_MSGS(conn);
+
+ /* post receives */
+ for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
+ rc = kiblnd_post_rx(&conn->ibc_rxs[i], IBLND_POSTRX_NO_CREDIT);
+ if (rc != 0) {
+ CERROR("Can't post rxmsg: %d\n", rc);
- /* Make posted receives complete */
- kiblnd_abort_receives(conn);
+ /* Make posted receives complete */
+ kiblnd_abort_receives(conn);
- /* correct # of posted buffers
- * NB locking needed now I'm racing with completion */
+ /* correct # of posted buffers
+ * NB locking needed now I'm racing with completion */
spin_lock_irqsave(&sched->ibs_lock, flags);
- conn->ibc_nrx -= IBLND_RX_MSGS(version) - i;
+ conn->ibc_nrx -= IBLND_RX_MSGS(conn) - i;
spin_unlock_irqrestore(&sched->ibs_lock, flags);
/* cmid will be destroyed by CM(ofed) after cm_callback
@@ -880,9 +899,9 @@ kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
rdma_destroy_qp(conn->ibc_cmid);
conn->ibc_cmid = NULL;
- /* Drop my own and unused rxbuffer refcounts */
- while (i++ <= IBLND_RX_MSGS(version))
- kiblnd_conn_decref(conn);
+ /* Drop my own and unused rxbuffer refcounts */
+ while (i++ <= IBLND_RX_MSGS(conn))
+ kiblnd_conn_decref(conn);
return NULL;
}
@@ -898,7 +917,7 @@ kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
return conn;
failed_2:
- kiblnd_destroy_conn(conn);
+ kiblnd_destroy_conn(conn, true);
failed_1:
LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
failed_0:
@@ -906,10 +925,10 @@ kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
}
void
-kiblnd_destroy_conn (kib_conn_t *conn)
+kiblnd_destroy_conn(kib_conn_t *conn, bool free_conn)
{
struct rdma_cm_id *cmid = conn->ibc_cmid;
- kib_peer_t *peer = conn->ibc_peer;
+ kib_peer_ni_t *peer_ni = conn->ibc_peer;
int rc;
LASSERT (!in_interrupt());
@@ -952,7 +971,7 @@ kiblnd_destroy_conn (kib_conn_t *conn)
if (conn->ibc_rxs != NULL) {
LIBCFS_FREE(conn->ibc_rxs,
- IBLND_RX_MSGS(conn->ibc_version) * sizeof(kib_rx_t));
+ IBLND_RX_MSGS(conn) * sizeof(kib_rx_t));
}
if (conn->ibc_connvars != NULL)
@@ -963,30 +982,31 @@ kiblnd_destroy_conn (kib_conn_t *conn)
/* See CAVEAT EMPTOR above in kiblnd_create_conn */
if (conn->ibc_state != IBLND_CONN_INIT) {
- kib_net_t *net = peer->ibp_ni->ni_data;
+ kib_net_t *net = peer_ni->ibp_ni->ni_data;
- kiblnd_peer_decref(peer);
+ kiblnd_peer_decref(peer_ni);
rdma_destroy_id(cmid);
atomic_dec(&net->ibn_nconns);
}
- LIBCFS_FREE(conn, sizeof(*conn));
+ if (free_conn)
+ LIBCFS_FREE(conn, sizeof(*conn));
}
int
-kiblnd_close_peer_conns_locked(kib_peer_t *peer, int why)
+kiblnd_close_peer_conns_locked(kib_peer_ni_t *peer_ni, int why)
{
kib_conn_t *conn;
struct list_head *ctmp;
struct list_head *cnxt;
int count = 0;
- list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
+ list_for_each_safe(ctmp, cnxt, &peer_ni->ibp_conns) {
conn = list_entry(ctmp, kib_conn_t, ibc_list);
CDEBUG(D_NET, "Closing conn -> %s, "
"version: %x, reason: %d\n",
- libcfs_nid2str(peer->ibp_nid),
+ libcfs_nid2str(peer_ni->ibp_nid),
conn->ibc_version, why);
kiblnd_close_conn_locked(conn, why);
@@ -997,7 +1017,7 @@ kiblnd_close_peer_conns_locked(kib_peer_t *peer, int why)
}
int
-kiblnd_close_stale_conns_locked(kib_peer_t *peer,
+kiblnd_close_stale_conns_locked(kib_peer_ni_t *peer_ni,
int version, __u64 incarnation)
{
kib_conn_t *conn;
@@ -1005,7 +1025,7 @@ kiblnd_close_stale_conns_locked(kib_peer_t *peer,
struct list_head *cnxt;
int count = 0;
- list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
+ list_for_each_safe(ctmp, cnxt, &peer_ni->ibp_conns) {
conn = list_entry(ctmp, kib_conn_t, ibc_list);
if (conn->ibc_version == version &&
@@ -1013,8 +1033,8 @@ kiblnd_close_stale_conns_locked(kib_peer_t *peer,
continue;
CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
- "incarnation:"LPX64"(%x, "LPX64")\n",
- libcfs_nid2str(peer->ibp_nid),
+ "incarnation:%#llx(%x, %#llx)\n",
+ libcfs_nid2str(peer_ni->ibp_nid),
conn->ibc_version, conn->ibc_incarnation,
version, incarnation);
@@ -1028,7 +1048,7 @@ kiblnd_close_stale_conns_locked(kib_peer_t *peer,
static int
kiblnd_close_matching_conns(lnet_ni_t *ni, lnet_nid_t nid)
{
- kib_peer_t *peer;
+ kib_peer_ni_t *peer_ni;
struct list_head *ptmp;
struct list_head *pnxt;
int lo;
@@ -1049,18 +1069,16 @@ kiblnd_close_matching_conns(lnet_ni_t *ni, lnet_nid_t nid)
for (i = lo; i <= hi; i++) {
list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
- peer = list_entry(ptmp, kib_peer_t, ibp_list);
- LASSERT(peer->ibp_connecting > 0 ||
- peer->ibp_accepting > 0 ||
- !list_empty(&peer->ibp_conns));
+ peer_ni = list_entry(ptmp, kib_peer_ni_t, ibp_list);
+ LASSERT(!kiblnd_peer_idle(peer_ni));
- if (peer->ibp_ni != ni)
+ if (peer_ni->ibp_ni != ni)
continue;
- if (!(nid == LNET_NID_ANY || nid == peer->ibp_nid))
+ if (!(nid == LNET_NID_ANY || nid == peer_ni->ibp_nid))
continue;
- count += kiblnd_close_peer_conns_locked(peer, 0);
+ count += kiblnd_close_peer_conns_locked(peer_ni, 0);
}
}
@@ -1133,31 +1151,27 @@ kiblnd_query(lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when)
cfs_time_t last_alive = 0;
cfs_time_t now = cfs_time_current();
rwlock_t *glock = &kiblnd_data.kib_global_lock;
- kib_peer_t *peer;
+ kib_peer_ni_t *peer_ni;
unsigned long flags;
read_lock_irqsave(glock, flags);
- peer = kiblnd_find_peer_locked(nid);
- if (peer != NULL) {
- LASSERT(peer->ibp_connecting > 0 || /* creating conns */
- peer->ibp_accepting > 0 ||
- !list_empty(&peer->ibp_conns)); /* active conn */
- last_alive = peer->ibp_last_alive;
- }
+ peer_ni = kiblnd_find_peer_locked(ni, nid);
+ if (peer_ni != NULL)
+ last_alive = peer_ni->ibp_last_alive;
read_unlock_irqrestore(glock, flags);
if (last_alive != 0)
*when = last_alive;
- /* peer is not persistent in hash, trigger peer creation
+ /* peer_ni is not persistent in hash, trigger peer_ni creation
* and connection establishment with a NULL tx */
- if (peer == NULL)
+ if (peer_ni == NULL)
kiblnd_launch_tx(ni, NULL, nid);
- CDEBUG(D_NET, "Peer %s %p, alive %ld secs ago\n",
- libcfs_nid2str(nid), peer,
+ CDEBUG(D_NET, "peer_ni %s %p, alive %ld secs ago\n",
+ libcfs_nid2str(nid), peer_ni,
last_alive ? cfs_duration_sec(now - last_alive) : -1);
return;
}
@@ -1215,16 +1229,16 @@ kiblnd_unmap_rx_descs(kib_conn_t *conn)
LASSERT (conn->ibc_rxs != NULL);
LASSERT (conn->ibc_hdev != NULL);
- for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++) {
- rx = &conn->ibc_rxs[i];
+ for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
+ rx = &conn->ibc_rxs[i];
- LASSERT (rx->rx_nob >= 0); /* not posted */
+ LASSERT(rx->rx_nob >= 0); /* not posted */
- kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
- KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
- rx->rx_msgaddr),
- IBLND_MSG_SIZE, DMA_FROM_DEVICE);
- }
+ kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
+ KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
+ rx->rx_msgaddr),
+ IBLND_MSG_SIZE, DMA_FROM_DEVICE);
+ }
kiblnd_free_pages(conn->ibc_rx_pages);
@@ -1240,34 +1254,34 @@ kiblnd_map_rx_descs(kib_conn_t *conn)
int ipg;
int i;
- for (pg_off = ipg = i = 0;
- i < IBLND_RX_MSGS(conn->ibc_version); i++) {
- pg = conn->ibc_rx_pages->ibp_pages[ipg];
- rx = &conn->ibc_rxs[i];
+ for (pg_off = ipg = i = 0; i < IBLND_RX_MSGS(conn); i++) {
+ pg = conn->ibc_rx_pages->ibp_pages[ipg];
+ rx = &conn->ibc_rxs[i];
- rx->rx_conn = conn;
- rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
+ rx->rx_conn = conn;
+ rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
- rx->rx_msgaddr = kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
- rx->rx_msg, IBLND_MSG_SIZE,
- DMA_FROM_DEVICE);
- LASSERT (!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
- rx->rx_msgaddr));
- KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
+ rx->rx_msgaddr =
+ kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
+ rx->rx_msg, IBLND_MSG_SIZE,
+ DMA_FROM_DEVICE);
+ LASSERT(!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
+ rx->rx_msgaddr));
+ KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
- CDEBUG(D_NET, "rx %d: %p "LPX64"("LPX64")\n",
+ CDEBUG(D_NET, "rx %d: %p %#llx(%#llx)\n",
i, rx->rx_msg, rx->rx_msgaddr,
(__u64)(page_to_phys(pg) + pg_off));
- pg_off += IBLND_MSG_SIZE;
- LASSERT (pg_off <= PAGE_SIZE);
+ pg_off += IBLND_MSG_SIZE;
+ LASSERT(pg_off <= PAGE_SIZE);
- if (pg_off == PAGE_SIZE) {
- pg_off = 0;
- ipg++;
- LASSERT (ipg <= IBLND_RX_MSG_PAGES(conn->ibc_version));
- }
- }
+ if (pg_off == PAGE_SIZE) {
+ pg_off = 0;
+ ipg++;
+ LASSERT(ipg <= IBLND_RX_MSG_PAGES(conn));
+ }
+ }
}
static void
@@ -1307,6 +1321,7 @@ kiblnd_current_hdev(kib_dev_t *dev)
if (i++ % 50 == 0)
CDEBUG(D_NET, "%s: Wait for failover\n",
dev->ibd_ifname);
+ set_current_state(TASK_INTERRUPTIBLE);
schedule_timeout(cfs_time_seconds(1) / 100);
read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
@@ -1374,137 +1389,269 @@ kiblnd_map_tx_pool(kib_tx_pool_t *tpo)
}
struct ib_mr *
-kiblnd_find_dma_mr(kib_hca_dev_t *hdev, __u64 addr, __u64 size)
-{
- __u64 index;
-
- LASSERT (hdev->ibh_mrs[0] != NULL);
-
- if (hdev->ibh_nmrs == 1)
- return hdev->ibh_mrs[0];
-
- index = addr >> hdev->ibh_mr_shift;
-
- 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_hca_dev_t *hdev, kib_rdma_desc_t *rd)
+kiblnd_find_rd_dma_mr(struct lnet_ni *ni, kib_rdma_desc_t *rd,
+ int negotiated_nfrags)
{
- struct ib_mr *prev_mr;
- struct ib_mr *mr;
- int i;
+ kib_net_t *net = ni->ni_data;
+ kib_hca_dev_t *hdev = net->ibn_dev->ibd_hdev;
+ struct lnet_ioctl_config_o2iblnd_tunables *tunables;
+ int mod;
+ __u16 nfrags;
- LASSERT (hdev->ibh_mrs[0] != NULL);
+ tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
+ mod = tunables->lnd_map_on_demand;
+ nfrags = (negotiated_nfrags != -1) ? negotiated_nfrags : mod;
- if (*kiblnd_tunables.kib_map_on_demand > 0 &&
- *kiblnd_tunables.kib_map_on_demand <= rd->rd_nfrags)
- return NULL;
+ LASSERT(hdev->ibh_mrs != NULL);
- 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(hdev,
- rd->rd_frags[i].rf_addr,
- rd->rd_frags[i].rf_nob);
- if (prev_mr == NULL)
- prev_mr = mr;
-
- if (mr == NULL || prev_mr != mr) {
- /* Can't covered by one single MR */
- mr = NULL;
- break;
- }
- }
+ if (mod > 0 && nfrags <= rd->rd_nfrags)
+ return NULL;
- return mr;
+ return hdev->ibh_mrs;
}
static void
-kiblnd_destroy_fmr_pool(kib_fmr_pool_t *pool)
+kiblnd_destroy_fmr_pool(kib_fmr_pool_t *fpo)
{
- LASSERT (pool->fpo_map_count == 0);
+ LASSERT(fpo->fpo_map_count == 0);
- if (pool->fpo_fmr_pool != NULL)
- ib_destroy_fmr_pool(pool->fpo_fmr_pool);
+ if (fpo->fpo_is_fmr) {
+ if (fpo->fmr.fpo_fmr_pool)
+ ib_destroy_fmr_pool(fpo->fmr.fpo_fmr_pool);
+ } else {
+ struct kib_fast_reg_descriptor *frd, *tmp;
+ int i = 0;
+
+ list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
+ frd_list) {
+ list_del(&frd->frd_list);
+#ifndef HAVE_IB_MAP_MR_SG
+ ib_free_fast_reg_page_list(frd->frd_frpl);
+#endif
+ ib_dereg_mr(frd->frd_mr);
+ LIBCFS_FREE(frd, sizeof(*frd));
+ i++;
+ }
+ if (i < fpo->fast_reg.fpo_pool_size)
+ CERROR("FastReg pool still has %d regions registered\n",
+ fpo->fast_reg.fpo_pool_size - i);
+ }
- if (pool->fpo_hdev != NULL)
- kiblnd_hdev_decref(pool->fpo_hdev);
+ if (fpo->fpo_hdev)
+ kiblnd_hdev_decref(fpo->fpo_hdev);
- LIBCFS_FREE(pool, sizeof(kib_fmr_pool_t));
+ LIBCFS_FREE(fpo, sizeof(*fpo));
}
static void
kiblnd_destroy_fmr_pool_list(struct list_head *head)
{
- kib_fmr_pool_t *pool;
+ kib_fmr_pool_t *fpo, *tmp;
- while (!list_empty(head)) {
- pool = list_entry(head->next, kib_fmr_pool_t, fpo_list);
- list_del(&pool->fpo_list);
- kiblnd_destroy_fmr_pool(pool);
+ list_for_each_entry_safe(fpo, tmp, head, fpo_list) {
+ list_del(&fpo->fpo_list);
+ kiblnd_destroy_fmr_pool(fpo);
}
}
-static int kiblnd_fmr_pool_size(int ncpts)
+static int
+kiblnd_fmr_pool_size(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
+ int ncpts)
{
- int size = *kiblnd_tunables.kib_fmr_pool_size / ncpts;
+ int size = tunables->lnd_fmr_pool_size / ncpts;
return max(IBLND_FMR_POOL, size);
}
-static int kiblnd_fmr_flush_trigger(int ncpts)
+static int
+kiblnd_fmr_flush_trigger(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
+ int ncpts)
{
- int size = *kiblnd_tunables.kib_fmr_flush_trigger / ncpts;
+ int size = tunables->lnd_fmr_flush_trigger / ncpts;
return max(IBLND_FMR_POOL_FLUSH, size);
}
-static int
-kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo)
+static int kiblnd_alloc_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t *fpo)
{
- /* FMR pool for RDMA */
- kib_dev_t *dev = fps->fps_net->ibn_dev;
- kib_fmr_pool_t *fpo;
- struct ib_fmr_pool_param param = {
- .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
- .page_shift = PAGE_SHIFT,
- .access = (IB_ACCESS_LOCAL_WRITE |
- IB_ACCESS_REMOTE_WRITE),
+ struct ib_fmr_pool_param param = {
+ .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
+ .page_shift = PAGE_SHIFT,
+ .access = (IB_ACCESS_LOCAL_WRITE |
+ IB_ACCESS_REMOTE_WRITE),
.pool_size = fps->fps_pool_size,
.dirty_watermark = fps->fps_flush_trigger,
.flush_function = NULL,
.flush_arg = NULL,
- .cache = !!*kiblnd_tunables.kib_fmr_cache};
+ .cache = !!fps->fps_cache };
+ int rc = 0;
+
+ fpo->fmr.fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd,
+ ¶m);
+ if (IS_ERR(fpo->fmr.fpo_fmr_pool)) {
+ rc = PTR_ERR(fpo->fmr.fpo_fmr_pool);
+ if (rc != -ENOSYS)
+ CERROR("Failed to create FMR pool: %d\n", rc);
+ else
+ CERROR("FMRs are not supported\n");
+ }
+
+ return rc;
+}
+
+static int kiblnd_alloc_freg_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t *fpo)
+{
+ struct kib_fast_reg_descriptor *frd, *tmp;
+ int i, rc;
+
+ INIT_LIST_HEAD(&fpo->fast_reg.fpo_pool_list);
+ fpo->fast_reg.fpo_pool_size = 0;
+ for (i = 0; i < fps->fps_pool_size; i++) {
+ LIBCFS_CPT_ALLOC(frd, lnet_cpt_table(), fps->fps_cpt,
+ sizeof(*frd));
+ if (!frd) {
+ CERROR("Failed to allocate a new fast_reg descriptor\n");
+ rc = -ENOMEM;
+ goto out;
+ }
+ frd->frd_mr = NULL;
+
+#ifndef HAVE_IB_MAP_MR_SG
+ frd->frd_frpl = ib_alloc_fast_reg_page_list(fpo->fpo_hdev->ibh_ibdev,
+ LNET_MAX_PAYLOAD/PAGE_SIZE);
+ if (IS_ERR(frd->frd_frpl)) {
+ rc = PTR_ERR(frd->frd_frpl);
+ CERROR("Failed to allocate ib_fast_reg_page_list: %d\n",
+ rc);
+ frd->frd_frpl = NULL;
+ goto out_middle;
+ }
+#endif
+
+#ifdef HAVE_IB_ALLOC_FAST_REG_MR
+ frd->frd_mr = ib_alloc_fast_reg_mr(fpo->fpo_hdev->ibh_pd,
+ LNET_MAX_PAYLOAD/PAGE_SIZE);
+#else
+ frd->frd_mr = ib_alloc_mr(fpo->fpo_hdev->ibh_pd,
+ IB_MR_TYPE_MEM_REG,
+ LNET_MAX_PAYLOAD/PAGE_SIZE);
+#endif
+ if (IS_ERR(frd->frd_mr)) {
+ rc = PTR_ERR(frd->frd_mr);
+ CERROR("Failed to allocate ib_fast_reg_mr: %d\n", rc);
+ frd->frd_mr = NULL;
+ goto out_middle;
+ }
+
+ /* There appears to be a bug in MLX5 code where you must
+ * invalidate the rkey of a new FastReg pool before first
+ * using it. Thus, I am marking the FRD invalid here. */
+ frd->frd_valid = false;
+
+ list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
+ fpo->fast_reg.fpo_pool_size++;
+ }
+
+ return 0;
+
+out_middle:
+ if (frd->frd_mr)
+ ib_dereg_mr(frd->frd_mr);
+#ifndef HAVE_IB_MAP_MR_SG
+ if (frd->frd_frpl)
+ ib_free_fast_reg_page_list(frd->frd_frpl);
+#endif
+ LIBCFS_FREE(frd, sizeof(*frd));
+
+out:
+ list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
+ frd_list) {
+ list_del(&frd->frd_list);
+#ifndef HAVE_IB_MAP_MR_SG
+ ib_free_fast_reg_page_list(frd->frd_frpl);
+#endif
+ ib_dereg_mr(frd->frd_mr);
+ LIBCFS_FREE(frd, sizeof(*frd));
+ }
+
+ return rc;
+}
+
+static int
+kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo)
+{
+ struct ib_device_attr *dev_attr;
+ kib_dev_t *dev = fps->fps_net->ibn_dev;
+ kib_fmr_pool_t *fpo;
int rc;
- LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
- if (fpo == NULL)
+#ifndef HAVE_IB_DEVICE_ATTRS
+ dev_attr = kmalloc(sizeof(*dev_attr), GFP_KERNEL);
+ if (!dev_attr)
return -ENOMEM;
+#endif
+
+ LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
+ if (!fpo) {
+ rc = -ENOMEM;
+ goto out_dev_attr;
+ }
fpo->fpo_hdev = kiblnd_current_hdev(dev);
- fpo->fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd, ¶m);
- if (IS_ERR(fpo->fpo_fmr_pool)) {
- rc = PTR_ERR(fpo->fpo_fmr_pool);
- CERROR("Failed to create FMR pool: %d\n", rc);
+#ifdef HAVE_IB_DEVICE_ATTRS
+ dev_attr = &fpo->fpo_hdev->ibh_ibdev->attrs;
+#else
+ rc = ib_query_device(fpo->fpo_hdev->ibh_ibdev, dev_attr);
+ if (rc) {
+ CERROR("Query device failed for %s: %d\n",
+ fpo->fpo_hdev->ibh_ibdev->name, rc);
+ goto out_dev_attr;
+ }
+#endif
+
+ /* Check for FMR or FastReg support */
+ fpo->fpo_is_fmr = 0;
+ if (fpo->fpo_hdev->ibh_ibdev->alloc_fmr &&
+ fpo->fpo_hdev->ibh_ibdev->dealloc_fmr &&
+ fpo->fpo_hdev->ibh_ibdev->map_phys_fmr &&
+ fpo->fpo_hdev->ibh_ibdev->unmap_fmr) {
+ LCONSOLE_INFO("Using FMR for registration\n");
+ fpo->fpo_is_fmr = 1;
+ } else if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
+ LCONSOLE_INFO("Using FastReg for registration\n");
+ } else {
+ rc = -ENOSYS;
+ LCONSOLE_ERROR_MSG(rc, "IB device does not support FMRs nor FastRegs, can't register memory\n");
+ goto out_dev_attr;
+ }
- kiblnd_hdev_decref(fpo->fpo_hdev);
- LIBCFS_FREE(fpo, sizeof(kib_fmr_pool_t));
- return rc;
- }
+ if (fpo->fpo_is_fmr)
+ rc = kiblnd_alloc_fmr_pool(fps, fpo);
+ else
+ rc = kiblnd_alloc_freg_pool(fps, fpo);
+ if (rc)
+ goto out_fpo;
- fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
- fpo->fpo_owner = fps;
- *pp_fpo = fpo;
+#ifndef HAVE_IB_DEVICE_ATTRS
+ kfree(dev_attr);
+#endif
+ fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
+ fpo->fpo_owner = fps;
+ *pp_fpo = fpo;
- return 0;
+ return 0;
+
+out_fpo:
+ kiblnd_hdev_decref(fpo->fpo_hdev);
+ LIBCFS_FREE(fpo, sizeof(*fpo));
+
+out_dev_attr:
+#ifndef HAVE_IB_DEVICE_ATTRS
+ kfree(dev_attr);
+#endif
+
+ return rc;
}
static void
@@ -1539,8 +1686,9 @@ kiblnd_fini_fmr_poolset(kib_fmr_poolset_t *fps)
}
static int
-kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt, kib_net_t *net,
- int pool_size, int flush_trigger)
+kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt, int ncpts,
+ kib_net_t *net,
+ struct lnet_ioctl_config_o2iblnd_tunables *tunables)
{
kib_fmr_pool_t *fpo;
int rc;
@@ -1549,8 +1697,11 @@ kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt, kib_net_t *net,
fps->fps_net = net;
fps->fps_cpt = cpt;
- fps->fps_pool_size = pool_size;
- fps->fps_flush_trigger = flush_trigger;
+
+ fps->fps_pool_size = kiblnd_fmr_pool_size(tunables, ncpts);
+ fps->fps_flush_trigger = kiblnd_fmr_flush_trigger(tunables, ncpts);
+ fps->fps_cache = tunables->lnd_fmr_cache;
+
spin_lock_init(&fps->fps_lock);
INIT_LIST_HEAD(&fps->fps_pool_list);
INIT_LIST_HEAD(&fps->fps_failed_pool_list);
@@ -1572,26 +1723,65 @@ kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, cfs_time_t now)
return cfs_time_aftereq(now, fpo->fpo_deadline);
}
+static int
+kiblnd_map_tx_pages(kib_tx_t *tx, kib_rdma_desc_t *rd)
+{
+ kib_hca_dev_t *hdev;
+ __u64 *pages = tx->tx_pages;
+ int npages;
+ int size;
+ int i;
+
+ hdev = tx->tx_pool->tpo_hdev;
+
+ for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
+ for (size = 0; size < rd->rd_frags[i].rf_nob;
+ size += hdev->ibh_page_size) {
+ pages[npages++] = (rd->rd_frags[i].rf_addr &
+ hdev->ibh_page_mask) + size;
+ }
+ }
+
+ return npages;
+}
+
void
kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
{
struct list_head zombies = LIST_HEAD_INIT(zombies);
kib_fmr_pool_t *fpo = fmr->fmr_pool;
- kib_fmr_poolset_t *fps = fpo->fpo_owner;
+ kib_fmr_poolset_t *fps;
cfs_time_t now = cfs_time_current();
kib_fmr_pool_t *tmp;
int rc;
- rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
- LASSERT(rc == 0);
+ if (!fpo)
+ return;
- if (status != 0) {
- rc = ib_flush_fmr_pool(fpo->fpo_fmr_pool);
- LASSERT(rc == 0);
- }
+ fps = fpo->fpo_owner;
+ if (fpo->fpo_is_fmr) {
+ if (fmr->fmr_pfmr) {
+ rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
+ LASSERT(!rc);
+ fmr->fmr_pfmr = NULL;
+ }
+ if (status) {
+ rc = ib_flush_fmr_pool(fpo->fmr.fpo_fmr_pool);
+ LASSERT(!rc);
+ }
+ } else {
+ struct kib_fast_reg_descriptor *frd = fmr->fmr_frd;
+
+ if (frd) {
+ frd->frd_valid = false;
+ spin_lock(&fps->fps_lock);
+ list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
+ spin_unlock(&fps->fps_lock);
+ fmr->fmr_frd = NULL;
+ }
+ }
fmr->fmr_pool = NULL;
- fmr->fmr_pfmr = NULL;
spin_lock(&fps->fps_lock);
fpo->fpo_map_count--; /* decref the pool */
@@ -1613,13 +1803,16 @@ kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
}
int
-kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages,
- __u64 iov, kib_fmr_t *fmr)
+kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, kib_tx_t *tx, kib_rdma_desc_t *rd,
+ __u32 nob, __u64 iov, kib_fmr_t *fmr)
{
- struct ib_pool_fmr *pfmr;
- kib_fmr_pool_t *fpo;
- __u64 version;
- int rc;
+ kib_fmr_pool_t *fpo;
+ __u64 *pages = tx->tx_pages;
+ __u64 version;
+ bool is_rx = (rd != tx->tx_rd);
+ bool tx_pages_mapped = 0;
+ int npages = 0;
+ int rc;
again:
spin_lock(&fps->fps_lock);
@@ -1627,21 +1820,138 @@ again:
list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
fpo->fpo_map_count++;
- spin_unlock(&fps->fps_lock);
- pfmr = ib_fmr_pool_map_phys(fpo->fpo_fmr_pool,
- pages, npages, iov);
- if (likely(!IS_ERR(pfmr))) {
- fmr->fmr_pool = fpo;
- fmr->fmr_pfmr = pfmr;
- return 0;
- }
+ if (fpo->fpo_is_fmr) {
+ struct ib_pool_fmr *pfmr;
+
+ spin_unlock(&fps->fps_lock);
+
+ if (!tx_pages_mapped) {
+ npages = kiblnd_map_tx_pages(tx, rd);
+ tx_pages_mapped = 1;
+ }
+
+ pfmr = ib_fmr_pool_map_phys(fpo->fmr.fpo_fmr_pool,
+ pages, npages, iov);
+ if (likely(!IS_ERR(pfmr))) {
+ fmr->fmr_key = is_rx ? pfmr->fmr->rkey
+ : pfmr->fmr->lkey;
+ fmr->fmr_frd = NULL;
+ fmr->fmr_pfmr = pfmr;
+ fmr->fmr_pool = fpo;
+ return 0;
+ }
+ rc = PTR_ERR(pfmr);
+ } else {
+ if (!list_empty(&fpo->fast_reg.fpo_pool_list)) {
+ struct kib_fast_reg_descriptor *frd;
+#ifdef HAVE_IB_MAP_MR_SG
+ struct ib_reg_wr *wr;
+ int n;
+#else
+ struct ib_rdma_wr *wr;
+ struct ib_fast_reg_page_list *frpl;
+#endif
+ struct ib_mr *mr;
+
+ frd = list_first_entry(&fpo->fast_reg.fpo_pool_list,
+ struct kib_fast_reg_descriptor,
+ frd_list);
+ list_del(&frd->frd_list);
+ spin_unlock(&fps->fps_lock);
+
+#ifndef HAVE_IB_MAP_MR_SG
+ frpl = frd->frd_frpl;
+#endif
+ mr = frd->frd_mr;
+
+ if (!frd->frd_valid) {
+ struct ib_rdma_wr *inv_wr;
+ __u32 key = is_rx ? mr->rkey : mr->lkey;
+
+ inv_wr = &frd->frd_inv_wr;
+ memset(inv_wr, 0, sizeof(*inv_wr));
+
+ inv_wr->wr.opcode = IB_WR_LOCAL_INV;
+ inv_wr->wr.wr_id = IBLND_WID_MR;
+ inv_wr->wr.ex.invalidate_rkey = key;
+
+ /* Bump the key */
+ key = ib_inc_rkey(key);
+ ib_update_fast_reg_key(mr, key);
+ }
+
+#ifdef HAVE_IB_MAP_MR_SG
+#ifdef HAVE_IB_MAP_MR_SG_5ARGS
+ n = ib_map_mr_sg(mr, tx->tx_frags,
+ tx->tx_nfrags, NULL, PAGE_SIZE);
+#else
+ n = ib_map_mr_sg(mr, tx->tx_frags,
+ tx->tx_nfrags, PAGE_SIZE);
+#endif
+ if (unlikely(n != tx->tx_nfrags)) {
+ CERROR("Failed to map mr %d/%d "
+ "elements\n", n, tx->tx_nfrags);
+ return n < 0 ? n : -EINVAL;
+ }
+
+ mr->iova = iov;
+
+ wr = &frd->frd_fastreg_wr;
+ memset(wr, 0, sizeof(*wr));
+
+ wr->wr.opcode = IB_WR_REG_MR;
+ wr->wr.wr_id = IBLND_WID_MR;
+ wr->wr.num_sge = 0;
+ wr->wr.send_flags = 0;
+ wr->mr = mr;
+ wr->key = is_rx ? mr->rkey : mr->lkey;
+ wr->access = (IB_ACCESS_LOCAL_WRITE |
+ IB_ACCESS_REMOTE_WRITE);
+#else
+ if (!tx_pages_mapped) {
+ npages = kiblnd_map_tx_pages(tx, rd);
+ tx_pages_mapped = 1;
+ }
+
+ LASSERT(npages <= frpl->max_page_list_len);
+ memcpy(frpl->page_list, pages,
+ sizeof(*pages) * npages);
+
+ /* Prepare FastReg WR */
+ wr = &frd->frd_fastreg_wr;
+ memset(wr, 0, sizeof(*wr));
+
+ wr->wr.opcode = IB_WR_FAST_REG_MR;
+ wr->wr.wr_id = IBLND_WID_MR;
+
+ wr->wr.wr.fast_reg.iova_start = iov;
+ wr->wr.wr.fast_reg.page_list = frpl;
+ wr->wr.wr.fast_reg.page_list_len = npages;
+ wr->wr.wr.fast_reg.page_shift = PAGE_SHIFT;
+ wr->wr.wr.fast_reg.length = nob;
+ wr->wr.wr.fast_reg.rkey =
+ is_rx ? mr->rkey : mr->lkey;
+ wr->wr.wr.fast_reg.access_flags =
+ (IB_ACCESS_LOCAL_WRITE |
+ IB_ACCESS_REMOTE_WRITE);
+#endif
+
+ fmr->fmr_key = is_rx ? mr->rkey : mr->lkey;
+ fmr->fmr_frd = frd;
+ fmr->fmr_pfmr = NULL;
+ fmr->fmr_pool = fpo;
+ return 0;
+ }
+ spin_unlock(&fps->fps_lock);
+ rc = -EBUSY;
+ }
spin_lock(&fps->fps_lock);
fpo->fpo_map_count--;
- if (PTR_ERR(pfmr) != -EAGAIN) {
+ if (rc != -EAGAIN) {
spin_unlock(&fps->fps_lock);
- return PTR_ERR(pfmr);
+ return rc;
}
/* EAGAIN and ... */
@@ -1831,6 +2141,9 @@ kiblnd_pool_alloc_node(kib_poolset_t *ps)
struct list_head *node;
kib_pool_t *pool;
int rc;
+ unsigned int interval = 1;
+ cfs_time_t time_before;
+ unsigned int trips = 0;
again:
spin_lock(&ps->ps_lock);
@@ -1855,10 +2168,17 @@ again:
if (ps->ps_increasing) {
/* another thread is allocating a new pool */
spin_unlock(&ps->ps_lock);
+ trips++;
CDEBUG(D_NET, "Another thread is allocating new "
- "%s pool, waiting for her to complete\n",
- ps->ps_name);
- schedule();
+ "%s pool, waiting %d HZs for her to complete."
+ "trips = %d\n",
+ ps->ps_name, interval, trips);
+
+ set_current_state(TASK_INTERRUPTIBLE);
+ schedule_timeout(interval);
+ if (interval < cfs_time_seconds(1))
+ interval *= 2;
+
goto again;
}
@@ -1872,8 +2192,10 @@ again:
spin_unlock(&ps->ps_lock);
CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
-
+ time_before = cfs_time_current();
rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
+ CDEBUG(D_NET, "ps_pool_create took %lu HZ to complete",
+ cfs_time_current() - time_before);
spin_lock(&ps->ps_lock);
ps->ps_increasing = 0;
@@ -1889,145 +2211,6 @@ again:
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;
- struct list_head *node;
- int rc;
- int i;
-
- node = kiblnd_pool_alloc_node(&pps->pps_poolset);
- if (node == NULL) {
- CERROR("Failed to allocate PMR descriptor\n");
- return -ENOMEM;
- }
-
- pmr = container_of(node, kib_phys_mr_t, pmr_list);
- if (pmr->pmr_pool->ppo_hdev != hdev) {
- kiblnd_pool_free_node(&pmr->pmr_pool->ppo_pool, node);
- return -EAGAIN;
- }
-
- 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 (!list_empty(&pool->po_free_list)) {
- pmr = list_entry(pool->po_free_list.next,
- kib_phys_mr_t, pmr_list);
-
- LASSERT (pmr->pmr_mr == NULL);
- list_del(&pmr->pmr_list);
-
- if (pmr->pmr_ipb != NULL) {
- LIBCFS_FREE(pmr->pmr_ipb,
- IBLND_MAX_RDMA_FRAGS *
- sizeof(struct ib_phys_buf));
- }
-
- LIBCFS_FREE(pmr, sizeof(kib_phys_mr_t));
- }
-
- kiblnd_fini_pool(pool);
- if (ppo->ppo_hdev != NULL)
- kiblnd_hdev_decref(ppo->ppo_hdev);
-
- LIBCFS_FREE(ppo, sizeof(kib_pmr_pool_t));
-}
-
-static inline int kiblnd_pmr_pool_size(int ncpts)
-{
- int size = *kiblnd_tunables.kib_pmr_pool_size / ncpts;
-
- return max(IBLND_PMR_POOL, size);
-}
-
-static int
-kiblnd_create_pmr_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
-{
- struct kib_pmr_pool *ppo;
- struct kib_pool *pool;
- kib_phys_mr_t *pmr;
- int i;
-
- LIBCFS_CPT_ALLOC(ppo, lnet_cpt_table(),
- ps->ps_cpt, sizeof(kib_pmr_pool_t));
- if (ppo == NULL) {
- CERROR("Failed to allocate PMR pool\n");
- return -ENOMEM;
- }
-
- pool = &ppo->ppo_pool;
- kiblnd_init_pool(ps, pool, size);
-
- for (i = 0; i < size; i++) {
- LIBCFS_CPT_ALLOC(pmr, lnet_cpt_table(),
- ps->ps_cpt, sizeof(kib_phys_mr_t));
- if (pmr == NULL)
- break;
-
- pmr->pmr_pool = ppo;
- LIBCFS_CPT_ALLOC(pmr->pmr_ipb, lnet_cpt_table(), ps->ps_cpt,
- IBLND_MAX_RDMA_FRAGS * sizeof(*pmr->pmr_ipb));
- if (pmr->pmr_ipb == NULL)
- break;
-
- 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)
{
@@ -2054,8 +2237,8 @@ kiblnd_destroy_tx_pool(kib_pool_t *pool)
sizeof(*tx->tx_pages));
if (tx->tx_frags != NULL)
LIBCFS_FREE(tx->tx_frags,
- IBLND_MAX_RDMA_FRAGS *
- sizeof(*tx->tx_frags));
+ (1 + IBLND_MAX_RDMA_FRAGS) *
+ sizeof(*tx->tx_frags));
if (tx->tx_wrq != NULL)
LIBCFS_FREE(tx->tx_wrq,
(1 + IBLND_MAX_RDMA_FRAGS) *
@@ -2133,11 +2316,12 @@ kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
}
LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
- IBLND_MAX_RDMA_FRAGS * sizeof(*tx->tx_frags));
+ (1 + IBLND_MAX_RDMA_FRAGS) *
+ sizeof(*tx->tx_frags));
if (tx->tx_frags == NULL)
break;
- sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS);
+ sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS + 1);
LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
(1 + IBLND_MAX_RDMA_FRAGS) *
@@ -2186,7 +2370,6 @@ kiblnd_net_fini_pools(kib_net_t *net)
cfs_cpt_for_each(i, lnet_cpt_table()) {
kib_tx_poolset_t *tps;
kib_fmr_poolset_t *fps;
- kib_pmr_poolset_t *pps;
if (net->ibn_tx_ps != NULL) {
tps = net->ibn_tx_ps[i];
@@ -2197,11 +2380,6 @@ kiblnd_net_fini_pools(kib_net_t *net)
fps = net->ibn_fmr_ps[i];
kiblnd_fini_fmr_poolset(fps);
}
-
- if (net->ibn_pmr_ps != NULL) {
- pps = net->ibn_pmr_ps[i];
- kiblnd_fini_poolset(&pps->pps_poolset);
- }
}
if (net->ibn_tx_ps != NULL) {
@@ -2213,24 +2391,21 @@ kiblnd_net_fini_pools(kib_net_t *net)
cfs_percpt_free(net->ibn_fmr_ps);
net->ibn_fmr_ps = NULL;
}
-
- if (net->ibn_pmr_ps != NULL) {
- cfs_percpt_free(net->ibn_pmr_ps);
- net->ibn_pmr_ps = NULL;
- }
}
static int
-kiblnd_net_init_pools(kib_net_t *net, __u32 *cpts, int ncpts)
+kiblnd_net_init_pools(kib_net_t *net, lnet_ni_t *ni, __u32 *cpts, int ncpts)
{
+ struct lnet_ioctl_config_o2iblnd_tunables *tunables;
unsigned long flags;
int cpt;
int rc;
int i;
+ tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
+
read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
- if (*kiblnd_tunables.kib_map_on_demand == 0 &&
- net->ibn_dev->ibd_hdev->ibh_nmrs == 1) {
+ if (tunables->lnd_map_on_demand == 0) {
read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
flags);
goto create_tx_pool;
@@ -2238,21 +2413,20 @@ kiblnd_net_init_pools(kib_net_t *net, __u32 *cpts, int ncpts)
read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
- if (*kiblnd_tunables.kib_fmr_pool_size <
- *kiblnd_tunables.kib_ntx / 4) {
+ if (tunables->lnd_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,
+ tunables->lnd_fmr_pool_size,
*kiblnd_tunables.kib_ntx / 4);
rc = -EINVAL;
goto failed;
}
- /* TX pool must be created later than FMR/PMR, see LU-2268
+ /* TX pool must be created later than FMR, see LU-2268
* for details */
LASSERT(net->ibn_tx_ps == NULL);
/* premapping can fail if ibd_nmr > 1, so we always create
- * FMR/PMR pool and map-on-demand if premapping failed */
+ * FMR pool and map-on-demand if premapping failed */
net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
sizeof(kib_fmr_poolset_t));
@@ -2264,59 +2438,17 @@ kiblnd_net_init_pools(kib_net_t *net, __u32 *cpts, int ncpts)
for (i = 0; i < ncpts; i++) {
cpt = (cpts == NULL) ? i : cpts[i];
- rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, net,
- kiblnd_fmr_pool_size(ncpts),
- kiblnd_fmr_flush_trigger(ncpts));
- if (rc == -ENOSYS && i == 0) /* no FMR */
- break; /* create PMR pool */
-
- if (rc != 0) { /* a real error */
+ rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, ncpts,
+ net, tunables);
+ if (rc != 0) {
CERROR("Can't initialize FMR pool for CPT %d: %d\n",
cpt, rc);
goto failed;
}
}
- if (i > 0) {
+ if (i > 0)
LASSERT(i == ncpts);
- goto create_tx_pool;
- }
-
- cfs_percpt_free(net->ibn_fmr_ps);
- net->ibn_fmr_ps = NULL;
-
- CWARN("Device does not support FMR, failing back to PMR\n");
-
- if (*kiblnd_tunables.kib_pmr_pool_size <
- *kiblnd_tunables.kib_ntx / 4) {
- CERROR("Can't set pmr pool size (%d) < ntx / 4(%d)\n",
- *kiblnd_tunables.kib_pmr_pool_size,
- *kiblnd_tunables.kib_ntx / 4);
- rc = -EINVAL;
- goto failed;
- }
-
- net->ibn_pmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
- sizeof(kib_pmr_poolset_t));
- if (net->ibn_pmr_ps == NULL) {
- CERROR("Failed to allocate PMR pool array\n");
- rc = -ENOMEM;
- goto failed;
- }
-
- for (i = 0; i < ncpts; i++) {
- cpt = (cpts == NULL) ? i : cpts[i];
- rc = kiblnd_init_poolset(&net->ibn_pmr_ps[cpt]->pps_poolset,
- cpt, net, "PMR",
- kiblnd_pmr_pool_size(ncpts),
- kiblnd_create_pmr_pool,
- kiblnd_destroy_pmr_pool, NULL, NULL);
- if (rc != 0) {
- CERROR("Can't initialize PMR pool for CPT %d: %d\n",
- cpt, rc);
- goto failed;
- }
- }
create_tx_pool:
net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
@@ -2352,8 +2484,10 @@ kiblnd_net_init_pools(kib_net_t *net, __u32 *cpts, int ncpts)
static int
kiblnd_hdev_get_attr(kib_hca_dev_t *hdev)
{
- struct ib_device_attr *attr;
- int rc;
+#ifndef HAVE_IB_DEVICE_ATTRS
+ struct ib_device_attr *attr;
+ int rc;
+#endif
/* It's safe to assume a HCA can handle a page size
* matching that of the native system */
@@ -2361,6 +2495,9 @@ kiblnd_hdev_get_attr(kib_hca_dev_t *hdev)
hdev->ibh_page_size = 1 << PAGE_SHIFT;
hdev->ibh_page_mask = ~((__u64)hdev->ibh_page_size - 1);
+#ifdef HAVE_IB_DEVICE_ATTRS
+ hdev->ibh_mr_size = hdev->ibh_ibdev->attrs.max_mr_size;
+#else
LIBCFS_ALLOC(attr, sizeof(*attr));
if (attr == NULL) {
CERROR("Out of memory\n");
@@ -2377,41 +2514,26 @@ kiblnd_hdev_get_attr(kib_hca_dev_t *hdev)
CERROR("Failed to query IB device: %d\n", rc);
return rc;
}
+#endif
if (hdev->ibh_mr_size == ~0ULL) {
hdev->ibh_mr_shift = 64;
return 0;
}
- 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", hdev->ibh_mr_size);
+ CERROR("Invalid mr size: %#llx\n", hdev->ibh_mr_size);
return -EINVAL;
}
static void
kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev)
{
- int i;
-
- if (hdev->ibh_nmrs == 0 || hdev->ibh_mrs == NULL)
- return;
-
- for (i = 0; i < hdev->ibh_nmrs; i++) {
- if (hdev->ibh_mrs[i] == NULL)
- break;
+ if (hdev->ibh_mrs == NULL)
+ return;
- ib_dereg_mr(hdev->ibh_mrs[i]);
- }
+ ib_dereg_mr(hdev->ibh_mrs);
- LIBCFS_FREE(hdev->ibh_mrs, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
- hdev->ibh_mrs = NULL;
- hdev->ibh_nmrs = 0;
+ hdev->ibh_mrs = NULL;
}
void
@@ -2431,89 +2553,25 @@ kiblnd_hdev_destroy(kib_hca_dev_t *hdev)
static int
kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev)
{
- struct ib_mr *mr;
- int i;
- int rc;
- __u64 mm_size;
- __u64 mr_size;
- int acflags = IB_ACCESS_LOCAL_WRITE |
- IB_ACCESS_REMOTE_WRITE;
-
- rc = kiblnd_hdev_get_attr(hdev);
- if (rc != 0)
- return rc;
-
- if (hdev->ibh_mr_shift == 64) {
- LIBCFS_ALLOC(hdev->ibh_mrs, 1 * sizeof(*hdev->ibh_mrs));
- if (hdev->ibh_mrs == NULL) {
- CERROR("Failed to allocate MRs table\n");
- return -ENOMEM;
- }
-
- hdev->ibh_mrs[0] = NULL;
- hdev->ibh_nmrs = 1;
+ struct ib_mr *mr;
+ int rc;
+ int acflags = IB_ACCESS_LOCAL_WRITE |
+ IB_ACCESS_REMOTE_WRITE;
- mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
- if (IS_ERR(mr)) {
- CERROR("Failed ib_get_dma_mr : %ld\n", PTR_ERR(mr));
- kiblnd_hdev_cleanup_mrs(hdev);
- return PTR_ERR(mr);
- }
-
- hdev->ibh_mrs[0] = mr;
-
- goto out;
- }
-
- mr_size = (1ULL << hdev->ibh_mr_shift);
- mm_size = (unsigned long)high_memory - PAGE_OFFSET;
-
- hdev->ibh_nmrs = (int)((mm_size + mr_size - 1) >> hdev->ibh_mr_shift);
-
- if (hdev->ibh_mr_shift < 32 || hdev->ibh_nmrs > 1024) {
- /* it's 4T..., assume we will re-code at that time */
- CERROR("Can't support memory size: x"LPX64
- " with MR size: x"LPX64"\n", mm_size, mr_size);
- return -EINVAL;
- }
-
- /* create an array of MRs to cover all memory */
- LIBCFS_ALLOC(hdev->ibh_mrs, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
- if (hdev->ibh_mrs == NULL) {
- CERROR("Failed to allocate MRs' table\n");
- return -ENOMEM;
- }
-
- memset(hdev->ibh_mrs, 0, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
-
- for (i = 0; i < hdev->ibh_nmrs; i++) {
- struct ib_phys_buf ipb;
- __u64 iova;
-
- ipb.size = hdev->ibh_mr_size;
- ipb.addr = i * mr_size;
- iova = ipb.addr;
-
- mr = ib_reg_phys_mr(hdev->ibh_pd, &ipb, 1, acflags, &iova);
- if (IS_ERR(mr)) {
- CERROR("Failed ib_reg_phys_mr addr "LPX64
- " size "LPX64" : %ld\n",
- ipb.addr, ipb.size, PTR_ERR(mr));
- kiblnd_hdev_cleanup_mrs(hdev);
- return PTR_ERR(mr);
- }
+ rc = kiblnd_hdev_get_attr(hdev);
+ if (rc != 0)
+ return rc;
- LASSERT (iova == ipb.addr);
+ mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
+ if (IS_ERR(mr)) {
+ CERROR("Failed ib_get_dma_mr: %ld\n", PTR_ERR(mr));
+ kiblnd_hdev_cleanup_mrs(hdev);
+ return PTR_ERR(mr);
+ }
- hdev->ibh_mrs[i] = mr;
- }
+ hdev->ibh_mrs = mr;
-out:
- if (hdev->ibh_mr_size != ~0ULL || hdev->ibh_nmrs != 1)
- LCONSOLE_INFO("Register global MR array, MR size: "
- LPX64", array size: %d\n",
- hdev->ibh_mr_size, hdev->ibh_nmrs);
- return 0;
+ return 0;
}
static int
@@ -2683,14 +2741,9 @@ kiblnd_dev_failover(kib_dev_t *dev)
kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
&zombie_tpo);
- if (net->ibn_fmr_ps != NULL) {
+ if (net->ibn_fmr_ps != NULL)
kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
&zombie_fpo);
-
- } else if (net->ibn_pmr_ps != NULL) {
- kiblnd_fail_poolset(&net->ibn_pmr_ps[i]->
- pps_poolset, &zombie_ppo);
- }
}
}
@@ -2754,7 +2807,6 @@ kiblnd_create_dev(char *ifname)
if (dev == NULL)
return NULL;
- memset(dev, 0, sizeof(*dev));
netdev = dev_get_by_name(&init_net, ifname);
if (netdev == NULL) {
dev->ibd_can_failover = 0;
@@ -2805,6 +2857,8 @@ kiblnd_base_shutdown(void)
}
LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
+ LASSERT(list_empty(&kiblnd_data.kib_reconn_list));
+ LASSERT(list_empty(&kiblnd_data.kib_reconn_wait));
/* flag threads to terminate; wake and wait for them to die */
kiblnd_data.kib_shutdown = 1;
@@ -2879,7 +2933,7 @@ kiblnd_shutdown (lnet_ni_t *ni)
/* nuke all existing peers within this net */
kiblnd_del_peer(ni, LNET_NID_ANY);
- /* Wait for all peer state to clean up */
+ /* Wait for all peer_ni state to clean up */
i = 2;
while (atomic_read(&net->ibn_npeers) != 0) {
i++;
@@ -2956,6 +3010,9 @@ kiblnd_base_startup(void)
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_LIST_HEAD(&kiblnd_data.kib_reconn_list);
+ INIT_LIST_HEAD(&kiblnd_data.kib_reconn_wait);
+
init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
@@ -3126,8 +3183,9 @@ kiblnd_startup (lnet_ni_t *ni)
unsigned long flags;
int rc;
int newdev;
+ int node_id;
- LASSERT (ni->ni_lnd == &the_o2iblnd);
+ LASSERT (ni->ni_net->net_lnd == &the_o2iblnd);
if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
rc = kiblnd_base_startup();
@@ -3140,15 +3198,10 @@ kiblnd_startup (lnet_ni_t *ni)
if (net == NULL)
goto failed;
- memset(net, 0, sizeof(*net));
-
do_gettimeofday(&tv);
net->ibn_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
- ni->ni_peertimeout = *kiblnd_tunables.kib_peertimeout;
- ni->ni_maxtxcredits = *kiblnd_tunables.kib_credits;
- ni->ni_peertxcredits = *kiblnd_tunables.kib_peertxcredits;
- ni->ni_peerrtrcredits = *kiblnd_tunables.kib_peerrtrcredits;
+ kiblnd_tunables_setup(ni);
if (ni->ni_interfaces[0] != NULL) {
/* Use the IPoIB interface specified in 'networks=' */
@@ -3174,20 +3227,23 @@ kiblnd_startup (lnet_ni_t *ni)
newdev = ibdev == NULL;
/* hmm...create kib_dev even for alias */
if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0)
- ibdev = kiblnd_create_dev(ifname);
+ ibdev = kiblnd_create_dev(ifname);
- if (ibdev == NULL)
- goto failed;
+ if (ibdev == NULL)
+ goto failed;
- net->ibn_dev = ibdev;
- ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
+ node_id = dev_to_node(ibdev->ibd_hdev->ibh_ibdev->dma_device);
+ ni->dev_cpt = cfs_cpt_of_node(lnet_cpt_table(), node_id);
+
+ net->ibn_dev = ibdev;
+ ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
rc = kiblnd_dev_start_threads(ibdev, newdev,
ni->ni_cpts, ni->ni_ncpts);
if (rc != 0)
goto failed;
- rc = kiblnd_net_init_pools(net, ni->ni_cpts, ni->ni_ncpts);
+ rc = kiblnd_net_init_pools(net, ni, ni->ni_cpts, ni->ni_ncpts);
if (rc != 0) {
CERROR("Failed to initialize NI pools: %d\n", rc);
goto failed;
@@ -3222,36 +3278,36 @@ static lnd_t the_o2iblnd = {
.lnd_recv = kiblnd_recv,
};
-static void __exit
-kiblnd_module_fini (void)
+static void __exit ko2iblnd_exit(void)
{
- lnet_unregister_lnd(&the_o2iblnd);
- kiblnd_tunables_fini();
+ lnet_unregister_lnd(&the_o2iblnd);
}
-static int __init
-kiblnd_module_init (void)
+static int __init ko2iblnd_init(void)
{
- int rc;
+ int rc;
- CLASSERT (sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
- CLASSERT (offsetof(kib_msg_t, ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
- <= IBLND_MSG_SIZE);
- CLASSERT (offsetof(kib_msg_t, ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
- <= IBLND_MSG_SIZE);
+ CLASSERT(sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
+ CLASSERT(offsetof(kib_msg_t,
+ ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) <=
+ IBLND_MSG_SIZE);
+ CLASSERT(offsetof(kib_msg_t,
+ ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
+ <= IBLND_MSG_SIZE);
- rc = kiblnd_tunables_init();
- if (rc != 0)
- return rc;
+ rc = kiblnd_tunables_init();
+ if (rc != 0)
+ return rc;
- lnet_register_lnd(&the_o2iblnd);
+ lnet_register_lnd(&the_o2iblnd);
- return 0;
+ return 0;
}
-MODULE_AUTHOR("Sun Microsystems, Inc. ");
-MODULE_DESCRIPTION("Kernel OpenIB gen2 LND v2.00");
+MODULE_AUTHOR("OpenSFS, Inc. ");
+MODULE_DESCRIPTION("OpenIB gen2 LNet Network Driver");
+MODULE_VERSION("2.8.0");
MODULE_LICENSE("GPL");
-module_init(kiblnd_module_init);
-module_exit(kiblnd_module_fini);
+module_init(ko2iblnd_init);
+module_exit(ko2iblnd_exit);