/* Primary entry points from LNET. There are no guarantees against reentrance. */
lnd_t the_kgnilnd = {
+#ifdef CONFIG_CRAY_XT
.lnd_type = GNILND,
+#else
+ .lnd_type = GNIIPLND,
+#endif
.lnd_startup = kgnilnd_startup,
.lnd_shutdown = kgnilnd_shutdown,
.lnd_ctl = kgnilnd_ctl,
LIBCFS_ALLOC(conn->gnc_tx_ref_table, GNILND_MAX_MSG_ID * sizeof(void *));
if (conn->gnc_tx_ref_table == NULL) {
CERROR("Can't allocate conn tx_ref_table\n");
- rc = -ENOMEM;
- GOTO(failed, rc);
+ GOTO(failed, rc = -ENOMEM);
}
atomic_set(&conn->gnc_refcount, 1);
if (conn->gnc_cqid == 0) {
CERROR("Could not allocate unique CQ ID for conn 0x%p\n", conn);
- rc = -E2BIG;
- GOTO(failed, rc);
+ GOTO(failed, rc = -E2BIG);
}
CDEBUG(D_NET, "alloc cqid %u for conn 0x%p\n",
rrc = kgnilnd_ep_create(dev->gnd_handle, dev->gnd_snd_fma_cqh,
&conn->gnc_ephandle);
mutex_unlock(&dev->gnd_cq_mutex);
- if (rrc != GNI_RC_SUCCESS) {
- rc = -ENETDOWN;
- GOTO(failed, rc);
- }
+ if (rrc != GNI_RC_SUCCESS)
+ GOTO(failed, rc = -ENETDOWN);
CDEBUG(D_NET, "created conn 0x%p ep_hndl 0x%p\n",
conn, conn->gnc_ephandle);
* kgn_peer_conn_lock is held, we guarantee that nobody calls
* kgnilnd_add_peer_locked without checking gnn_shutdown */
int
-kgnilnd_create_peer_safe(kgn_peer_t **peerp, lnet_nid_t nid, kgn_net_t *net)
+kgnilnd_create_peer_safe(kgn_peer_t **peerp,
+ lnet_nid_t nid,
+ kgn_net_t *net,
+ int node_state)
{
kgn_peer_t *peer;
int rc;
return -ENOMEM;
}
peer->gnp_nid = nid;
- peer->gnp_down = GNILND_RCA_NODE_UP;
+ peer->gnp_down = node_state;
/* translate from nid to nic addr & store */
rc = kgnilnd_nid_to_nicaddrs(LNET_NIDADDR(nid), 1, &peer->gnp_host_id);
CDEBUG(D_NET, "conn %p peer %p dev %p\n", conn, peer,
conn->gnc_device);
+ LASSERTF(conn->gnc_in_purgatory == 0,
+ "Conn already in purgatory\n");
conn->gnc_in_purgatory = 1;
mbox = &conn->gnc_fma_blk->gnm_mbox_info[conn->gnc_mbox_id];
{
kgn_peer_t *peer;
int rc;
+ int node_state;
ENTRY;
if (nid == LNET_NID_ANY)
return -EINVAL;
+ node_state = kgnilnd_get_node_state(LNET_NIDADDR(nid));
+
/* NB - this will not block during normal operations -
* the only writer of this is in the startup/shutdown path. */
rc = down_read_trylock(&kgnilnd_data.kgn_net_rw_sem);
rc = -ESHUTDOWN;
RETURN(rc);
}
- rc = kgnilnd_create_peer_safe(&peer, nid, net);
+ rc = kgnilnd_create_peer_safe(&peer, nid, net, node_state);
if (rc != 0) {
up_read(&kgnilnd_data.kgn_net_rw_sem);
RETURN(rc);
write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
- /* release all of the souls found held in purgatory */
- kgnilnd_release_purgatory_list(&souls);
-
/* nuke peer TX */
kgnilnd_txlist_done(&zombies, error);
{
int rc;
kgn_peer_t *peer, *new_peer;
- CFS_LIST_HEAD(zombies);
+ LIST_HEAD(zombies);
write_lock(&kgnilnd_data.kgn_peer_conn_lock);
peer = kgnilnd_find_peer_locked(nid);
&dev->gnd_domain);
if (rrc != GNI_RC_SUCCESS) {
CERROR("Can't create CDM %d (%d)\n", dev->gnd_id, rrc);
- rc = -ENODEV;
- GOTO(failed, rc);
+ GOTO(failed, rc = -ENODEV);
}
rrc = kgnilnd_cdm_attach(dev->gnd_domain, dev->gnd_id,
if (rrc != GNI_RC_SUCCESS) {
CERROR("Can't attach CDM to device %d (%d)\n",
dev->gnd_id, rrc);
- rc = -ENODEV;
- GOTO(failed, rc);
+ GOTO(failed, rc = -ENODEV);
}
/* a bit gross, but not much we can do - Aries Sim doesn't have
* hardcoded NIC/NID that we can use */
rc = kgnilnd_setup_nic_translation(dev->gnd_host_id);
- if (rc != 0) {
- rc = -ENODEV;
- GOTO(failed, rc);
- }
+ if (rc != 0)
+ GOTO(failed, rc = -ENODEV);
/* only dev 0 gets the errors - no need to reset the stack twice
* - this works because we have a single PTAG, if we had more
if (rrc != GNI_RC_SUCCESS) {
CERROR("Can't subscribe for errors on device %d: rc %d\n",
dev->gnd_id, rrc);
- rc = -ENODEV;
- GOTO(failed, rc);
+ GOTO(failed, rc = -ENODEV);
}
rc = kgnilnd_set_quiesce_callback(dev->gnd_handle,
if (rc != GNI_RC_SUCCESS) {
CERROR("Can't subscribe for quiesce callback on device %d: rc %d\n",
dev->gnd_id, rrc);
- rc = -ENODEV;
- GOTO(failed, rc);
+ GOTO(failed, rc = -ENODEV);
}
}
+ rrc = sock_create_kern(PF_INET, SOCK_DGRAM, IPPROTO_IP, &kgnilnd_data.kgn_sock);
+ if (rrc < 0) {
+ CERROR("sock_create returned %d\n", rrc);
+ GOTO(failed, rrc);
+ }
+
rc = kgnilnd_nicaddr_to_nid(dev->gnd_host_id, &dev->gnd_nid);
if (rc < 0) {
/* log messages during startup */
CERROR("couldn't translate host_id 0x%x to nid. rc %d\n",
dev->gnd_host_id, rc);
}
- rc = -ESRCH;
- GOTO(failed, rc);
+ GOTO(failed, rc = -ESRCH);
}
CDEBUG(D_NET, "NIC %x -> NID %d\n", dev->gnd_host_id, dev->gnd_nid);
if (rrc != GNI_RC_SUCCESS) {
CERROR("Can't create rdma send cq size %u for device "
"%d (%d)\n", cq_size, dev->gnd_id, rrc);
- rc = -EINVAL;
- GOTO(failed, rc);
+ GOTO(failed, rc = -EINVAL);
}
rrc = kgnilnd_cq_create(dev->gnd_handle, cq_size,
if (rrc != GNI_RC_SUCCESS) {
CERROR("Can't create fma send cq size %u for device %d (%d)\n",
cq_size, dev->gnd_id, rrc);
- rc = -EINVAL;
- GOTO(failed, rc);
+ GOTO(failed, rc = -EINVAL);
}
/* This one we size differently - overflows are possible and it needs to be
if (rrc != GNI_RC_SUCCESS) {
CERROR("Can't create fma cq size %d for device %d (%d)\n",
*kgnilnd_tunables.kgn_fma_cq_size, dev->gnd_id, rrc);
- rc = -EINVAL;
- GOTO(failed, rc);
+ GOTO(failed, rc = -EINVAL);
}
RETURN(0);
dev->gnd_domain = NULL;
}
+ sock_release(kgnilnd_data.kgn_sock);
+
EXIT;
}
INIT_LIST_HEAD(&dev->gnd_map_tx);
INIT_LIST_HEAD(&dev->gnd_fma_buffs);
mutex_init(&dev->gnd_cq_mutex);
- sema_init(&dev->gnd_fmablk_sem, 1);
+ mutex_init(&dev->gnd_fmablk_mutex);
spin_lock_init(&dev->gnd_fmablk_lock);
init_waitqueue_head(&dev->gnd_waitq);
init_waitqueue_head(&dev->gnd_dgram_waitq);
LIBCFS_ALLOC(dev->gnd_dgrams,
sizeof(struct list_head) * *kgnilnd_tunables.kgn_peer_hash_size);
- if (dev->gnd_dgrams == NULL) {
- rc = -ENOMEM;
- GOTO(failed, rc);
- }
+ if (dev->gnd_dgrams == NULL)
+ GOTO(failed, rc = -ENOMEM);
for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++) {
INIT_LIST_HEAD(&dev->gnd_dgrams[i]);
init_waitqueue_head(&kgnilnd_data.kgn_ruhroh_waitq);
spin_lock_init(&kgnilnd_data.kgn_reaper_lock);
- sema_init(&kgnilnd_data.kgn_quiesce_sem, 1);
+ mutex_init(&kgnilnd_data.kgn_quiesce_mutex);
atomic_set(&kgnilnd_data.kgn_nquiesce, 0);
atomic_set(&kgnilnd_data.kgn_npending_conns, 0);
atomic_set(&kgnilnd_data.kgn_npending_unlink, 0);
/* OK to call kgnilnd_api_shutdown() to cleanup now */
kgnilnd_data.kgn_init = GNILND_INIT_DATA;
- PORTAL_MODULE_USE;
+ try_module_get(THIS_MODULE);
rwlock_init(&kgnilnd_data.kgn_peer_conn_lock);
LIBCFS_ALLOC(kgnilnd_data.kgn_peers,
sizeof(struct list_head) * *kgnilnd_tunables.kgn_peer_hash_size);
- if (kgnilnd_data.kgn_peers == NULL) {
- rc = -ENOMEM;
- GOTO(failed, rc);
- }
+ if (kgnilnd_data.kgn_peers == NULL)
+ GOTO(failed, rc = -ENOMEM);
for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++) {
INIT_LIST_HEAD(&kgnilnd_data.kgn_peers[i]);
LIBCFS_ALLOC(kgnilnd_data.kgn_conns,
sizeof(struct list_head) * *kgnilnd_tunables.kgn_peer_hash_size);
- if (kgnilnd_data.kgn_conns == NULL) {
- rc = -ENOMEM;
- GOTO(failed, rc);
- }
+ if (kgnilnd_data.kgn_conns == NULL)
+ GOTO(failed, rc = -ENOMEM);
for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++) {
INIT_LIST_HEAD(&kgnilnd_data.kgn_conns[i]);
LIBCFS_ALLOC(kgnilnd_data.kgn_nets,
sizeof(struct list_head) * *kgnilnd_tunables.kgn_net_hash_size);
- if (kgnilnd_data.kgn_nets == NULL) {
- rc = -ENOMEM;
- GOTO(failed, rc);
- }
+ if (kgnilnd_data.kgn_nets == NULL)
+ GOTO(failed, rc = -ENOMEM);
for (i = 0; i < *kgnilnd_tunables.kgn_net_hash_size; i++) {
INIT_LIST_HEAD(&kgnilnd_data.kgn_nets[i]);
}
kgnilnd_data.kgn_mbox_cache =
- cfs_mem_cache_create("kgn_mbox_block",
- KMALLOC_MAX_SIZE,
- 0, /* offset */
- SLAB_HWCACHE_ALIGN); /* flags */
+ kmem_cache_create("kgn_mbox_block", KMALLOC_MAX_SIZE, 0,
+ SLAB_HWCACHE_ALIGN, NULL);
if (kgnilnd_data.kgn_mbox_cache == NULL) {
CERROR("Can't create slab for physical mbox blocks\n");
- rc = -ENOMEM;
- GOTO(failed, rc);
+ GOTO(failed, rc = -ENOMEM);
}
kgnilnd_data.kgn_rx_cache =
- cfs_mem_cache_create("kgn_rx_t",
- sizeof(kgn_rx_t),
- 0, /* offset */
- 0); /* flags */
+ kmem_cache_create("kgn_rx_t", sizeof(kgn_rx_t), 0, 0, NULL);
if (kgnilnd_data.kgn_rx_cache == NULL) {
CERROR("Can't create slab for kgn_rx_t descriptors\n");
- rc = -ENOMEM;
- GOTO(failed, rc);
+ GOTO(failed, rc = -ENOMEM);
}
kgnilnd_data.kgn_tx_cache =
- cfs_mem_cache_create("kgn_tx_t",
- sizeof(kgn_tx_t),
- 0, /* offset */
- 0); /* flags */
+ kmem_cache_create("kgn_tx_t", sizeof(kgn_tx_t), 0, 0, NULL);
if (kgnilnd_data.kgn_tx_cache == NULL) {
CERROR("Can't create slab for kgn_tx_t\n");
- rc = -ENOMEM;
- GOTO(failed, rc);
+ GOTO(failed, rc = -ENOMEM);
}
kgnilnd_data.kgn_tx_phys_cache =
- cfs_mem_cache_create("kgn_tx_phys",
- LNET_MAX_IOV * sizeof(gni_mem_segment_t),
- 0, /* offset */
- 0); /* flags */
+ kmem_cache_create("kgn_tx_phys",
+ LNET_MAX_IOV * sizeof(gni_mem_segment_t),
+ 0, 0, NULL);
if (kgnilnd_data.kgn_tx_phys_cache == NULL) {
CERROR("Can't create slab for kgn_tx_phys\n");
- rc = -ENOMEM;
- GOTO(failed, rc);
+ GOTO(failed, rc = -ENOMEM);
}
kgnilnd_data.kgn_dgram_cache =
- cfs_mem_cache_create("kgn_dgram_t",
- sizeof(kgn_dgram_t),
- 0, /* offset */
- 0); /* flags */
+ kmem_cache_create("kgn_dgram_t", sizeof(kgn_dgram_t), 0, 0, NULL);
if (kgnilnd_data.kgn_dgram_cache == NULL) {
CERROR("Can't create slab for outgoing datagrams\n");
- rc = -ENOMEM;
- GOTO(failed, rc);
+ GOTO(failed, rc = -ENOMEM);
}
/* allocate a MAX_IOV array of page pointers for each cpu */
GFP_KERNEL);
if (kgnilnd_data.kgn_cksum_map_pages == NULL) {
CERROR("Can't allocate vmap cksum pages\n");
- rc = -ENOMEM;
- GOTO(failed, rc);
+ GOTO(failed, rc = -ENOMEM);
}
kgnilnd_data.kgn_cksum_npages = num_possible_cpus();
memset(kgnilnd_data.kgn_cksum_map_pages, 0,
GFP_KERNEL);
if (kgnilnd_data.kgn_cksum_map_pages[i] == NULL) {
CERROR("Can't allocate vmap cksum pages for cpu %d\n", i);
- rc = -ENOMEM;
- GOTO(failed, rc);
+ GOTO(failed, rc = -ENOMEM);
}
}
kgnilnd_data.kgn_ndevs++;
rc = kgnilnd_allocate_phys_fmablk(dev);
- if (rc) {
+ if (rc)
GOTO(failed, rc);
- }
}
}
if (kgnilnd_data.kgn_ndevs == 0) {
CERROR("Can't initialise any GNI devices\n");
- rc = -ENODEV;
- GOTO(failed, rc);
+ GOTO(failed, rc = -ENODEV);
}
rc = kgnilnd_thread_start(kgnilnd_reaper, NULL, "kgnilnd_rpr", 0);
void
kgnilnd_base_shutdown(void)
{
- int i;
+ int i, j;
ENTRY;
while (CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_PAUSE_SHUTDOWN, 1)) {};
for (i = 0; i < kgnilnd_data.kgn_ndevs; i++) {
kgn_device_t *dev = &kgnilnd_data.kgn_devices[i];
kgnilnd_cancel_wc_dgrams(dev);
+ kgnilnd_cancel_dgrams(dev);
kgnilnd_del_conn_or_peer(NULL, LNET_NID_ANY, GNILND_DEL_PEER, -ESHUTDOWN);
kgnilnd_wait_for_canceled_dgrams(dev);
}
+ /* We need to verify there are no conns left before we let the threads
+ * shut down otherwise we could clean up the peers but still have
+ * some outstanding conns due to orphaned datagram conns that are
+ * being cleaned up.
+ */
+ i = 2;
+ while (atomic_read(&kgnilnd_data.kgn_nconns) != 0) {
+ i++;
+
+ for(j = 0; j < kgnilnd_data.kgn_ndevs; ++j) {
+ kgn_device_t *dev = &kgnilnd_data.kgn_devices[j];
+ kgnilnd_schedule_device(dev);
+ }
+
+ CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
+ "Waiting for conns to be cleaned up %d\n",atomic_read(&kgnilnd_data.kgn_nconns));
+ cfs_pause(cfs_time_seconds(1));
+ }
/* Peer state all cleaned up BEFORE setting shutdown, so threads don't
* have to worry about shutdown races. NB connections may be created
* while there are still active connds, but these will be temporary
kgn_device_t *dev = &kgnilnd_data.kgn_devices[i];
/* should clear all the MDDs */
- kgnilnd_unmap_phys_fmablk(dev);
+ kgnilnd_unmap_fma_blocks(dev);
kgnilnd_schedule_device(dev);
wake_up_all(&dev->gnd_dgram_waitq);
kgnilnd_free_phys_fmablk(dev);
}
- if (kgnilnd_data.kgn_mbox_cache != NULL) {
- i = cfs_mem_cache_destroy(kgnilnd_data.kgn_mbox_cache);
- LASSERTF(i == 0, "rc %d destroying kgn_mbox_cache\n", i);
- }
+ if (kgnilnd_data.kgn_mbox_cache != NULL)
+ kmem_cache_destroy(kgnilnd_data.kgn_mbox_cache);
- if (kgnilnd_data.kgn_rx_cache != NULL) {
- i = cfs_mem_cache_destroy(kgnilnd_data.kgn_rx_cache);
- LASSERTF(i == 0, "rc %d destroying kgn_rx_cache\n", i);
- }
+ if (kgnilnd_data.kgn_rx_cache != NULL)
+ kmem_cache_destroy(kgnilnd_data.kgn_rx_cache);
- if (kgnilnd_data.kgn_tx_cache != NULL) {
- i = cfs_mem_cache_destroy(kgnilnd_data.kgn_tx_cache);
- LASSERTF(i == 0, "rc %d destroying kgn_tx_cache\n", i);
- }
+ if (kgnilnd_data.kgn_tx_cache != NULL)
+ kmem_cache_destroy(kgnilnd_data.kgn_tx_cache);
- if (kgnilnd_data.kgn_tx_phys_cache != NULL) {
- i = cfs_mem_cache_destroy(kgnilnd_data.kgn_tx_phys_cache);
- LASSERTF(i == 0, "rc %d destroying kgn_tx_phys_cache\n", i);
- }
+ if (kgnilnd_data.kgn_tx_phys_cache != NULL)
+ kmem_cache_destroy(kgnilnd_data.kgn_tx_phys_cache);
- if (kgnilnd_data.kgn_dgram_cache != NULL) {
- i = cfs_mem_cache_destroy(kgnilnd_data.kgn_dgram_cache);
- LASSERTF(i == 0, "rc %d destroying kgn_dgram_cache\n", i);
- }
+ if (kgnilnd_data.kgn_dgram_cache != NULL)
+ kmem_cache_destroy(kgnilnd_data.kgn_dgram_cache);
if (kgnilnd_data.kgn_cksum_map_pages != NULL) {
for (i = 0; i < kgnilnd_data.kgn_cksum_npages; i++) {
atomic_read(&libcfs_kmemory));
kgnilnd_data.kgn_init = GNILND_INIT_NOTHING;
- PORTAL_MODULE_UNUSE;
+ module_put(THIS_MODULE);
EXIT;
}
}
/* Serialize with shutdown. */
- down(&kgnilnd_data.kgn_quiesce_sem);
+ mutex_lock(&kgnilnd_data.kgn_quiesce_mutex);
LIBCFS_ALLOC(net, sizeof(*net));
if (net == NULL) {
CERROR("could not allocate net for new interface instance\n");
- rc = -ENOMEM;
/* no need to cleanup the CDM... */
- GOTO(failed, rc);
+ GOTO(failed, rc = -ENOMEM);
}
INIT_LIST_HEAD(&net->gnn_list);
ni->ni_data = net;
timeout);
ni->ni_data = NULL;
LIBCFS_FREE(net, sizeof(*net));
- rc = -EINVAL;
- GOTO(failed, rc);
+ GOTO(failed, rc = -EINVAL);
} else
ni->ni_peertimeout = timeout;
/* we need a separate thread to call probe_wait_by_id until
* we get a function callback notifier from kgni */
- up(&kgnilnd_data.kgn_quiesce_sem);
+ mutex_unlock(&kgnilnd_data.kgn_quiesce_mutex);
RETURN(0);
failed:
- up(&kgnilnd_data.kgn_quiesce_sem);
+ mutex_unlock(&kgnilnd_data.kgn_quiesce_mutex);
kgnilnd_shutdown(ni);
RETURN(rc);
}
"init %d\n", kgnilnd_data.kgn_init);
/* Serialize with startup. */
- down(&kgnilnd_data.kgn_quiesce_sem);
+ mutex_lock(&kgnilnd_data.kgn_quiesce_mutex);
CDEBUG(D_MALLOC, "before NAL cleanup: kmem %d\n",
atomic_read(&libcfs_kmemory));
if (net == NULL) {
CERROR("got NULL net for ni %p\n", ni);
- rc = -EINVAL;
- GOTO(out, rc);
+ GOTO(out, rc = -EINVAL);
}
LASSERTF(ni == net->gnn_ni,
CDEBUG(D_MALLOC, "after NAL cleanup: kmem %d\n",
atomic_read(&libcfs_kmemory));
- up(&kgnilnd_data.kgn_quiesce_sem);
+ mutex_unlock(&kgnilnd_data.kgn_quiesce_mutex);
EXIT;
- return;
}
void __exit
git://git.whamcloud.com / fs / lustre-release.git / blobdiff