/*
* Copyright (C) 2012 Cray, Inc.
*
+ * Copyright (c) 2013, 2014, Intel Corporation.
+ *
* Author: Nic Henke <nic@cray.com>
* Author: James Shimek <jshimek@cray.com>
*
/* 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,
kgn_data_t kgnilnd_data;
+int
+kgnilnd_thread_start(int(*fn)(void *arg), void *arg, char *name, int id)
+{
+ struct task_struct *thrd;
+
+ thrd = kthread_run(fn, arg, "%s_%02d", name, id);
+ if (IS_ERR(thrd))
+ return PTR_ERR(thrd);
+
+ atomic_inc(&kgnilnd_data.kgn_nthreads);
+ return 0;
+}
+
+/* bind scheduler threads to cpus */
+int
+kgnilnd_start_sd_threads(void)
+{
+ int cpu;
+ int i = 0;
+ struct task_struct *task;
+
+ for_each_online_cpu(cpu) {
+ /* don't bind to cpu 0 - all interrupts are processed here */
+ if (cpu == 0)
+ continue;
+
+ task = kthread_create(kgnilnd_scheduler, (void *)((long)i),
+ "%s_%02d", "kgnilnd_sd", i);
+ if (!IS_ERR(task)) {
+ kthread_bind(task, cpu);
+ wake_up_process(task);
+ } else {
+ CERROR("Can't spawn gnilnd scheduler[%d] %ld\n", i,
+ PTR_ERR(task));
+ return PTR_ERR(task);
+ }
+ atomic_inc(&kgnilnd_data.kgn_nthreads);
+
+ if (++i >= *kgnilnd_tunables.kgn_sched_threads) {
+ break;
+ }
+ }
+
+ return 0;
+}
+
/* needs write_lock on kgn_peer_conn_lock */
int
kgnilnd_close_stale_conns_locked(kgn_peer_t *peer, kgn_conn_t *newconn)
GOTO(failed, rc = -ENOMEM);
}
+ mutex_init(&conn->gnc_smsg_mutex);
+ mutex_init(&conn->gnc_rdma_mutex);
atomic_set(&conn->gnc_refcount, 1);
atomic_set(&conn->gnc_reaper_noop, 0);
atomic_set(&conn->gnc_sched_noop, 0);
}
void
-kgnilnd_peer_notify(kgn_peer_t *peer, int error)
+kgnilnd_peer_notify(kgn_peer_t *peer, int error, int alive)
{
int tell_lnet = 0;
int nnets = 0;
peer, libcfs_nid2str(peer->gnp_nid), peer->gnp_connecting, conn,
kgnilnd_data.kgn_in_reset, error);
- if ((peer->gnp_connecting == GNILND_PEER_IDLE) &&
+ if (((peer->gnp_connecting == GNILND_PEER_IDLE) &&
(conn == NULL) &&
(!kgnilnd_data.kgn_in_reset) &&
- (!kgnilnd_conn_clean_errno(error))) {
+ (!kgnilnd_conn_clean_errno(error))) || alive) {
tell_lnet = 1;
}
peer, libcfs_nid2str(peer_nid), peer->gnp_last_alive,
cfs_duration_sec(jiffies - peer->gnp_last_alive));
- lnet_notify(net->gnn_ni, peer_nid, 0, peer->gnp_last_alive);
-
+ lnet_notify(net->gnn_ni, peer_nid, alive,
+ peer->gnp_last_alive);
kgnilnd_net_decref(net);
}
/* I'm telling Mommy! - use peer_error if they initiated close */
kgnilnd_peer_notify(conn->gnc_peer,
- conn->gnc_error == -ECONNRESET ? conn->gnc_peer_error
- : conn->gnc_error);
+ conn->gnc_error == -ECONNRESET ?
+ conn->gnc_peer_error : conn->gnc_error, 0);
EXIT;
}
* 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];
* make sure we tell LNet - if this is from other context,
* the checks in the function will prevent an errant
* notification */
- kgnilnd_peer_notify(conn->gnc_peer, conn->gnc_error);
+ kgnilnd_peer_notify(conn->gnc_peer, conn->gnc_error, 0);
list_for_each_entry_safe(gmp, gmpN, &conn->gnc_mdd_list,
gmp_list) {
{
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);
atomic_read(&kgnilnd_data.kgn_npending_detach) ||
atomic_read(&kgnilnd_data.kgn_npending_unlink)) {
- cfs_pause(cfs_time_seconds(1));
+ set_current_state(TASK_UNINTERRUPTIBLE);
+ schedule_timeout(cfs_time_seconds(1));
i++;
CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, "Waiting on %d peers %d closes %d detaches\n",
*device_id = conn->gnc_device->gnd_host_id;
*peerstamp = conn->gnc_peerstamp;
- *tx_seq = conn->gnc_tx_seq;
- *rx_seq = conn->gnc_rx_seq;
+ *tx_seq = atomic_read(&conn->gnc_tx_seq);
+ *rx_seq = atomic_read(&conn->gnc_rx_seq);
*fmaq_len = kgnilnd_count_list(&conn->gnc_fmaq);
*nfma = atomic_read(&conn->gnc_nlive_fma);
*nrdma = atomic_read(&conn->gnc_nlive_rdma);
* kgnilnd_tx_done
*/
kgnilnd_txlist_done(&zombies, -ENETRESET);
-
- if (*kgnilnd_tunables.kgn_peer_health) {
- kgnilnd_peer_notify(peer, -ECONNRESET);
- }
+ kgnilnd_peer_notify(peer, -ECONNRESET, 0);
+ LCONSOLE_INFO("Recieved down event for nid %lld\n", nid);
}
- CDEBUG(D_INFO, "marking nid %lld %s\n", nid, down ? "down" : "up");
return 0;
}
kgn_tx_t *tx;
kgn_peer_t *peer = NULL;
kgn_conn_t *conn = NULL;
- lnet_process_id_t id = {.nid = nid, .pid = LUSTRE_SRV_LNET_PID};
+ lnet_process_id_t id = {
+ .nid = nid,
+ .pid = LNET_PID_LUSTRE,
+ };
ENTRY;
/* I expect to find him, so only take a read lock */
cq_size = *kgnilnd_tunables.kgn_credits * 2 * 3;
rrc = kgnilnd_cdm_create(dev->gnd_id, *kgnilnd_tunables.kgn_ptag,
- GNILND_COOKIE, 0,
+ *kgnilnd_tunables.kgn_pkey, 0,
&dev->gnd_domain);
if (rrc != GNI_RC_SUCCESS) {
CERROR("Can't create CDM %d (%d)\n", dev->gnd_id, rrc);
}
}
+ rrc = sock_create_kern(PF_INET, SOCK_DGRAM, IPPROTO_IP, &kgnilnd_data.kgn_sock);
+ if (rrc < 0) {
+ CERROR("sock_create returned %d\n", rrc);
+ GOTO(failed, rrc);
+ }
+
rc = kgnilnd_nicaddr_to_nid(dev->gnd_host_id, &dev->gnd_nid);
if (rc < 0) {
/* log messages during startup */
}
CDEBUG(D_NET, "NIC %x -> NID %d\n", dev->gnd_host_id, dev->gnd_nid);
- rrc = kgnilnd_cq_create(dev->gnd_handle, cq_size,
+ rrc = kgnilnd_cq_create(dev->gnd_handle, *kgnilnd_tunables.kgn_credits,
0, kgnilnd_device_callback,
dev->gnd_id, &dev->gnd_snd_rdma_cqh);
if (rrc != GNI_RC_SUCCESS) {
dev->gnd_domain = NULL;
}
+ if (kgnilnd_data.kgn_sock)
+ sock_release(kgnilnd_data.kgn_sock);
+
EXIT;
}
-
int kgnilnd_base_startup(void)
{
struct timeval tv;
int i;
kgn_device_t *dev;
struct task_struct *thrd;
+
+#if defined(CONFIG_CRAY_XT) && !defined(CONFIG_CRAY_COMPUTE)
+ /* limit how much memory can be allocated for fma blocks in
+ * instances where many nodes need to reconnects at the same time */
+ struct sysinfo si;
+ si_meminfo(&si);
+ kgnilnd_data.free_pages_limit = si.totalram/4;
+#endif
+
ENTRY;
LASSERTF(kgnilnd_data.kgn_init == GNILND_INIT_NOTHING,
/* zero pointers, flags etc */
memset(&kgnilnd_data, 0, sizeof(kgnilnd_data));
+ kgnilnd_check_kgni_version();
/* CAVEAT EMPTOR: Every 'Fma' message includes the sender's NID and
* a unique (for all time) connstamp so we can uniquely identify
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);
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);
}
/* threads will load balance across devs as they are available */
- for (i = 0; i < *kgnilnd_tunables.kgn_sched_threads; i++) {
- rc = kgnilnd_thread_start(kgnilnd_scheduler, (void *)((long)i),
- "kgnilnd_sd", i);
- if (rc != 0) {
- CERROR("Can't spawn gnilnd scheduler[%d]: %d\n",
- i, rc);
+ if (*kgnilnd_tunables.kgn_thread_affinity) {
+ rc = kgnilnd_start_sd_threads();
+ if (rc != 0)
GOTO(failed, rc);
+ } else {
+ for (i = 0; i < *kgnilnd_tunables.kgn_sched_threads; i++) {
+ rc = kgnilnd_thread_start(kgnilnd_scheduler,
+ (void *)((long)i),
+ "kgnilnd_sd", i);
+ if (rc != 0) {
+ CERROR("Can't spawn gnilnd scheduler[%d]: %d\n",
+ i, rc);
+ GOTO(failed, rc);
+ }
}
}
}
}
-
-
/* flag everything initialised */
kgnilnd_data.kgn_init = GNILND_INIT_ALL;
/*****************************************************/
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));
+ set_current_state(TASK_UNINTERRUPTIBLE);
+ schedule_timeout(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
i++;
CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
"Waiting for ruhroh thread to terminate\n");
- cfs_pause(cfs_time_seconds(1));
+ set_current_state(TASK_UNINTERRUPTIBLE);
+ schedule_timeout(cfs_time_seconds(1));
}
/* Flag threads to terminate */
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);
wake_up_all(&kgnilnd_data.kgn_reaper_waitq);
spin_unlock(&kgnilnd_data.kgn_reaper_lock);
- kgnilnd_wakeup_rca_thread();
+ if (atomic_read(&kgnilnd_data.kgn_nthreads))
+ kgnilnd_wakeup_rca_thread();
/* Wait for threads to exit */
i = 2;
CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
"Waiting for %d threads to terminate\n",
atomic_read(&kgnilnd_data.kgn_nthreads));
- cfs_pause(cfs_time_seconds(1));
+ set_current_state(TASK_UNINTERRUPTIBLE);
+ schedule_timeout(cfs_time_seconds(1));
}
LASSERTF(atomic_read(&kgnilnd_data.kgn_npeers) == 0,
}
/* Serialize with shutdown. */
- down(&kgnilnd_data.kgn_quiesce_sem);
+ mutex_lock(&kgnilnd_data.kgn_quiesce_mutex);
LIBCFS_ALLOC(net, sizeof(*net));
if (net == NULL) {
/* 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));
"Waiting for %d references to clear on net %d\n",
atomic_read(&net->gnn_refcount),
net->gnn_netnum);
- cfs_pause(cfs_time_seconds(1));
+ set_current_state(TASK_UNINTERRUPTIBLE);
+ schedule_timeout(cfs_time_seconds(1));
}
/* release ref from kgnilnd_startup */
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;
}
git://git.whamcloud.com / fs / lustre-release.git / blobdiff