summaryrefslogtreecommitdiff
path: root/net/rxrpc/conn_client.c
diff options
context:
space:
mode:
Diffstat (limited to 'net/rxrpc/conn_client.c')
-rw-r--r--net/rxrpc/conn_client.c910
1 files changed, 795 insertions, 115 deletions
diff --git a/net/rxrpc/conn_client.c b/net/rxrpc/conn_client.c
index 6e1099ed1dbd..349402b08e5a 100644
--- a/net/rxrpc/conn_client.c
+++ b/net/rxrpc/conn_client.c
@@ -7,6 +7,68 @@
* modify it under the terms of the GNU General Public Licence
* as published by the Free Software Foundation; either version
* 2 of the Licence, or (at your option) any later version.
+ *
+ *
+ * Client connections need to be cached for a little while after they've made a
+ * call so as to handle retransmitted DATA packets in case the server didn't
+ * receive the final ACK or terminating ABORT we sent it.
+ *
+ * Client connections can be in one of a number of cache states:
+ *
+ * (1) INACTIVE - The connection is not held in any list and may not have been
+ * exposed to the world. If it has been previously exposed, it was
+ * discarded from the idle list after expiring.
+ *
+ * (2) WAITING - The connection is waiting for the number of client conns to
+ * drop below the maximum capacity. Calls may be in progress upon it from
+ * when it was active and got culled.
+ *
+ * The connection is on the rxrpc_waiting_client_conns list which is kept
+ * in to-be-granted order. Culled conns with waiters go to the back of
+ * the queue just like new conns.
+ *
+ * (3) ACTIVE - The connection has at least one call in progress upon it, it
+ * may freely grant available channels to new calls and calls may be
+ * waiting on it for channels to become available.
+ *
+ * The connection is on the rxrpc_active_client_conns list which is kept
+ * in activation order for culling purposes.
+ *
+ * rxrpc_nr_active_client_conns is held incremented also.
+ *
+ * (4) CULLED - The connection got summarily culled to try and free up
+ * capacity. Calls currently in progress on the connection are allowed to
+ * continue, but new calls will have to wait. There can be no waiters in
+ * this state - the conn would have to go to the WAITING state instead.
+ *
+ * (5) IDLE - The connection has no calls in progress upon it and must have
+ * been exposed to the world (ie. the EXPOSED flag must be set). When it
+ * expires, the EXPOSED flag is cleared and the connection transitions to
+ * the INACTIVE state.
+ *
+ * The connection is on the rxrpc_idle_client_conns list which is kept in
+ * order of how soon they'll expire.
+ *
+ * There are flags of relevance to the cache:
+ *
+ * (1) EXPOSED - The connection ID got exposed to the world. If this flag is
+ * set, an extra ref is added to the connection preventing it from being
+ * reaped when it has no calls outstanding. This flag is cleared and the
+ * ref dropped when a conn is discarded from the idle list.
+ *
+ * This allows us to move terminal call state retransmission to the
+ * connection and to discard the call immediately we think it is done
+ * with. It also give us a chance to reuse the connection.
+ *
+ * (2) DONT_REUSE - The connection should be discarded as soon as possible and
+ * should not be reused. This is set when an exclusive connection is used
+ * or a call ID counter overflows.
+ *
+ * The caching state may only be changed if the cache lock is held.
+ *
+ * There are two idle client connection expiry durations. If the total number
+ * of connections is below the reap threshold, we use the normal duration; if
+ * it's above, we use the fast duration.
*/
#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
@@ -16,22 +78,37 @@
#include <linux/timer.h>
#include "ar-internal.h"
+__read_mostly unsigned int rxrpc_max_client_connections = 1000;
+__read_mostly unsigned int rxrpc_reap_client_connections = 900;
+__read_mostly unsigned int rxrpc_conn_idle_client_expiry = 2 * 60 * HZ;
+__read_mostly unsigned int rxrpc_conn_idle_client_fast_expiry = 2 * HZ;
+
+static unsigned int rxrpc_nr_client_conns;
+static unsigned int rxrpc_nr_active_client_conns;
+static __read_mostly bool rxrpc_kill_all_client_conns;
+
+static DEFINE_SPINLOCK(rxrpc_client_conn_cache_lock);
+static DEFINE_SPINLOCK(rxrpc_client_conn_discard_mutex);
+static LIST_HEAD(rxrpc_waiting_client_conns);
+static LIST_HEAD(rxrpc_active_client_conns);
+static LIST_HEAD(rxrpc_idle_client_conns);
+
/*
* We use machine-unique IDs for our client connections.
*/
DEFINE_IDR(rxrpc_client_conn_ids);
static DEFINE_SPINLOCK(rxrpc_conn_id_lock);
+static void rxrpc_cull_active_client_conns(void);
+static void rxrpc_discard_expired_client_conns(struct work_struct *);
+
+static DECLARE_DELAYED_WORK(rxrpc_client_conn_reap,
+ rxrpc_discard_expired_client_conns);
+
/*
* Get a connection ID and epoch for a client connection from the global pool.
* The connection struct pointer is then recorded in the idr radix tree. The
* epoch is changed if this wraps.
- *
- * TODO: The IDR tree gets very expensive on memory if the connection IDs are
- * widely scattered throughout the number space, so we shall need to retire
- * connections that have, say, an ID more than four times the maximum number of
- * client conns away from the current allocation point to try and keep the IDs
- * concentrated. We will also need to retire connections from an old epoch.
*/
static int rxrpc_get_client_connection_id(struct rxrpc_connection *conn,
gfp_t gfp)
@@ -114,8 +191,7 @@ void rxrpc_destroy_client_conn_ids(void)
}
/*
- * Allocate a client connection. The caller must take care to clear any
- * padding bytes in *cp.
+ * Allocate a client connection.
*/
static struct rxrpc_connection *
rxrpc_alloc_client_connection(struct rxrpc_conn_parameters *cp, gfp_t gfp)
@@ -131,6 +207,10 @@ rxrpc_alloc_client_connection(struct rxrpc_conn_parameters *cp, gfp_t gfp)
return ERR_PTR(-ENOMEM);
}
+ atomic_set(&conn->usage, 1);
+ if (conn->params.exclusive)
+ __set_bit(RXRPC_CONN_DONT_REUSE, &conn->flags);
+
conn->params = *cp;
conn->out_clientflag = RXRPC_CLIENT_INITIATED;
conn->state = RXRPC_CONN_CLIENT;
@@ -148,7 +228,6 @@ rxrpc_alloc_client_connection(struct rxrpc_conn_parameters *cp, gfp_t gfp)
goto error_2;
write_lock(&rxrpc_connection_lock);
- list_add_tail(&conn->link, &rxrpc_connections);
list_add_tail(&conn->proc_link, &rxrpc_connection_proc_list);
write_unlock(&rxrpc_connection_lock);
@@ -171,32 +250,68 @@ error_0:
}
/*
- * find a connection for a call
- * - called in process context with IRQs enabled
+ * Determine if a connection may be reused.
*/
-int rxrpc_connect_call(struct rxrpc_call *call,
- struct rxrpc_conn_parameters *cp,
- struct sockaddr_rxrpc *srx,
- gfp_t gfp)
+static bool rxrpc_may_reuse_conn(struct rxrpc_connection *conn)
+{
+ int id_cursor, id, distance, limit;
+
+ if (test_bit(RXRPC_CONN_DONT_REUSE, &conn->flags))
+ goto dont_reuse;
+
+ if (conn->proto.epoch != rxrpc_epoch)
+ goto mark_dont_reuse;
+
+ /* The IDR tree gets very expensive on memory if the connection IDs are
+ * widely scattered throughout the number space, so we shall want to
+ * kill off connections that, say, have an ID more than about four
+ * times the maximum number of client conns away from the current
+ * allocation point to try and keep the IDs concentrated.
+ */
+ id_cursor = READ_ONCE(rxrpc_client_conn_ids.cur);
+ id = conn->proto.cid >> RXRPC_CIDSHIFT;
+ distance = id - id_cursor;
+ if (distance < 0)
+ distance = -distance;
+ limit = round_up(rxrpc_max_client_connections, IDR_SIZE) * 4;
+ if (distance > limit)
+ goto mark_dont_reuse;
+
+ return true;
+
+mark_dont_reuse:
+ set_bit(RXRPC_CONN_DONT_REUSE, &conn->flags);
+dont_reuse:
+ return false;
+}
+
+/*
+ * Create or find a client connection to use for a call.
+ *
+ * If we return with a connection, the call will be on its waiting list. It's
+ * left to the caller to assign a channel and wake up the call.
+ */
+static int rxrpc_get_client_conn(struct rxrpc_call *call,
+ struct rxrpc_conn_parameters *cp,
+ struct sockaddr_rxrpc *srx,
+ gfp_t gfp)
{
struct rxrpc_connection *conn, *candidate = NULL;
struct rxrpc_local *local = cp->local;
struct rb_node *p, **pp, *parent;
long diff;
- int chan;
-
- DECLARE_WAITQUEUE(myself, current);
+ int ret = -ENOMEM;
_enter("{%d,%lx},", call->debug_id, call->user_call_ID);
cp->peer = rxrpc_lookup_peer(cp->local, srx, gfp);
if (!cp->peer)
- return -ENOMEM;
+ goto error;
+ /* If the connection is not meant to be exclusive, search the available
+ * connections to see if the connection we want to use already exists.
+ */
if (!cp->exclusive) {
- /* Search for a existing client connection unless this is going
- * to be a connection that's used exclusively for a single call.
- */
_debug("search 1");
spin_lock(&local->client_conns_lock);
p = local->client_conns.rb_node;
@@ -207,39 +322,55 @@ int rxrpc_connect_call(struct rxrpc_call *call,
diff = (cmp(peer) ?:
cmp(key) ?:
cmp(security_level));
- if (diff < 0)
+#undef cmp
+ if (diff < 0) {
p = p->rb_left;
- else if (diff > 0)
+ } else if (diff > 0) {
p = p->rb_right;
- else
- goto found_extant_conn;
+ } else {
+ if (rxrpc_may_reuse_conn(conn) &&
+ rxrpc_get_connection_maybe(conn))
+ goto found_extant_conn;
+ /* The connection needs replacing. It's better
+ * to effect that when we have something to
+ * replace it with so that we don't have to
+ * rebalance the tree twice.
+ */
+ break;
+ }
}
spin_unlock(&local->client_conns_lock);
}
- /* We didn't find a connection or we want an exclusive one. */
- _debug("get new conn");
+ /* There wasn't a connection yet or we need an exclusive connection.
+ * We need to create a candidate and then potentially redo the search
+ * in case we're racing with another thread also trying to connect on a
+ * shareable connection.
+ */
+ _debug("new conn");
candidate = rxrpc_alloc_client_connection(cp, gfp);
- if (!candidate) {
- _leave(" = -ENOMEM");
- return -ENOMEM;
+ if (IS_ERR(candidate)) {
+ ret = PTR_ERR(candidate);
+ goto error_peer;
}
+ /* Add the call to the new connection's waiting list in case we're
+ * going to have to wait for the connection to come live. It's our
+ * connection, so we want first dibs on the channel slots. We would
+ * normally have to take channel_lock but we do this before anyone else
+ * can see the connection.
+ */
+ list_add_tail(&call->chan_wait_link, &candidate->waiting_calls);
+
if (cp->exclusive) {
- /* Assign the call on an exclusive connection to channel 0 and
- * don't add the connection to the endpoint's shareable conn
- * lookup tree.
- */
- _debug("exclusive chan 0");
- conn = candidate;
- atomic_set(&conn->avail_chans, RXRPC_MAXCALLS - 1);
- spin_lock(&conn->channel_lock);
- chan = 0;
- goto found_channel;
+ call->conn = candidate;
+ _leave(" = 0 [exclusive %d]", candidate->debug_id);
+ return 0;
}
- /* We need to redo the search before attempting to add a new connection
- * lest we race with someone else adding a conflicting instance.
+ /* Publish the new connection for userspace to find. We need to redo
+ * the search before doing this lest we race with someone else adding a
+ * conflicting instance.
*/
_debug("search 2");
spin_lock(&local->client_conns_lock);
@@ -250,123 +381,672 @@ int rxrpc_connect_call(struct rxrpc_call *call,
parent = *pp;
conn = rb_entry(parent, struct rxrpc_connection, client_node);
+#define cmp(X) ((long)conn->params.X - (long)candidate->params.X)
diff = (cmp(peer) ?:
cmp(key) ?:
cmp(security_level));
- if (diff < 0)
+#undef cmp
+ if (diff < 0) {
pp = &(*pp)->rb_left;
- else if (diff > 0)
+ } else if (diff > 0) {
pp = &(*pp)->rb_right;
- else
- goto found_extant_conn;
+ } else {
+ if (rxrpc_may_reuse_conn(conn) &&
+ rxrpc_get_connection_maybe(conn))
+ goto found_extant_conn;
+ /* The old connection is from an outdated epoch. */
+ _debug("replace conn");
+ clear_bit(RXRPC_CONN_IN_CLIENT_CONNS, &conn->flags);
+ rb_replace_node(&conn->client_node,
+ &candidate->client_node,
+ &local->client_conns);
+ goto candidate_published;
+ }
}
- /* The second search also failed; simply add the new connection with
- * the new call in channel 0. Note that we need to take the channel
- * lock before dropping the client conn lock.
- */
_debug("new conn");
- set_bit(RXRPC_CONN_IN_CLIENT_CONNS, &candidate->flags);
rb_link_node(&candidate->client_node, parent, pp);
rb_insert_color(&candidate->client_node, &local->client_conns);
-attached:
- conn = candidate;
- candidate = NULL;
- atomic_set(&conn->avail_chans, RXRPC_MAXCALLS - 1);
- spin_lock(&conn->channel_lock);
+candidate_published:
+ set_bit(RXRPC_CONN_IN_CLIENT_CONNS, &candidate->flags);
+ call->conn = candidate;
spin_unlock(&local->client_conns_lock);
- chan = 0;
+ _leave(" = 0 [new %d]", candidate->debug_id);
+ return 0;
-found_channel:
- _debug("found chan");
- call->conn = conn;
- call->peer = rxrpc_get_peer(conn->params.peer);
- call->cid = conn->proto.cid | chan;
- call->call_id = ++conn->channels[chan].call_counter;
- conn->channels[chan].call_id = call->call_id;
- rcu_assign_pointer(conn->channels[chan].call, call);
+ /* We come here if we found a suitable connection already in existence.
+ * Discard any candidate we may have allocated, and try to get a
+ * channel on this one.
+ */
+found_extant_conn:
+ _debug("found conn");
+ spin_unlock(&local->client_conns_lock);
- _net("CONNECT call %d on conn %d", call->debug_id, conn->debug_id);
+ rxrpc_put_connection(candidate);
+ candidate = NULL;
+ spin_lock(&conn->channel_lock);
+ call->conn = conn;
+ list_add(&call->chan_wait_link, &conn->waiting_calls);
spin_unlock(&conn->channel_lock);
+ _leave(" = 0 [extant %d]", conn->debug_id);
+ return 0;
+
+error_peer:
rxrpc_put_peer(cp->peer);
cp->peer = NULL;
- _leave(" = %p {u=%d}", conn, atomic_read(&conn->usage));
- return 0;
+error:
+ _leave(" = %d", ret);
+ return ret;
+}
- /* We found a potentially suitable connection already in existence. If
- * we can reuse it (ie. its usage count hasn't been reduced to 0 by the
- * reaper), discard any candidate we may have allocated, and try to get
- * a channel on this one, otherwise we have to replace it.
- */
-found_extant_conn:
- _debug("found conn");
- if (!rxrpc_get_connection_maybe(conn)) {
- set_bit(RXRPC_CONN_IN_CLIENT_CONNS, &candidate->flags);
- rb_replace_node(&conn->client_node,
- &candidate->client_node,
- &local->client_conns);
- clear_bit(RXRPC_CONN_IN_CLIENT_CONNS, &conn->flags);
- goto attached;
+/*
+ * Activate a connection.
+ */
+static void rxrpc_activate_conn(struct rxrpc_connection *conn)
+{
+ conn->cache_state = RXRPC_CONN_CLIENT_ACTIVE;
+ rxrpc_nr_active_client_conns++;
+ list_move_tail(&conn->cache_link, &rxrpc_active_client_conns);
+}
+
+/*
+ * Attempt to animate a connection for a new call.
+ *
+ * If it's not exclusive, the connection is in the endpoint tree, and we're in
+ * the conn's list of those waiting to grab a channel. There is, however, a
+ * limit on the number of live connections allowed at any one time, so we may
+ * have to wait for capacity to become available.
+ *
+ * Note that a connection on the waiting queue might *also* have active
+ * channels if it has been culled to make space and then re-requested by a new
+ * call.
+ */
+static void rxrpc_animate_client_conn(struct rxrpc_connection *conn)
+{
+ unsigned int nr_conns;
+
+ _enter("%d,%d", conn->debug_id, conn->cache_state);
+
+ if (conn->cache_state == RXRPC_CONN_CLIENT_ACTIVE)
+ goto out;
+
+ spin_lock(&rxrpc_client_conn_cache_lock);
+
+ nr_conns = rxrpc_nr_client_conns;
+ if (!test_and_set_bit(RXRPC_CONN_COUNTED, &conn->flags))
+ rxrpc_nr_client_conns = nr_conns + 1;
+
+ switch (conn->cache_state) {
+ case RXRPC_CONN_CLIENT_ACTIVE:
+ case RXRPC_CONN_CLIENT_WAITING:
+ break;
+
+ case RXRPC_CONN_CLIENT_INACTIVE:
+ case RXRPC_CONN_CLIENT_CULLED:
+ case RXRPC_CONN_CLIENT_IDLE:
+ if (nr_conns >= rxrpc_max_client_connections)
+ goto wait_for_capacity;
+ goto activate_conn;
+
+ default:
+ BUG();
}
- spin_unlock(&local->client_conns_lock);
+out_unlock:
+ spin_unlock(&rxrpc_client_conn_cache_lock);
+out:
+ _leave(" [%d]", conn->cache_state);
+ return;
- rxrpc_put_connection(candidate);
+activate_conn:
+ _debug("activate");
+ rxrpc_activate_conn(conn);
+ goto out_unlock;
+
+wait_for_capacity:
+ _debug("wait");
+ conn->cache_state = RXRPC_CONN_CLIENT_WAITING;
+ list_move_tail(&conn->cache_link, &rxrpc_waiting_client_conns);
+ goto out_unlock;
+}
+
+/*
+ * Deactivate a channel.
+ */
+static void rxrpc_deactivate_one_channel(struct rxrpc_connection *conn,
+ unsigned int channel)
+{
+ struct rxrpc_channel *chan = &conn->channels[channel];
+
+ rcu_assign_pointer(chan->call, NULL);
+ conn->active_chans &= ~(1 << channel);
+}
+
+/*
+ * Assign a channel to the call at the front of the queue and wake the call up.
+ * We don't increment the callNumber counter until this number has been exposed
+ * to the world.
+ */
+static void rxrpc_activate_one_channel(struct rxrpc_connection *conn,
+ unsigned int channel)
+{
+ struct rxrpc_channel *chan = &conn->channels[channel];
+ struct rxrpc_call *call = list_entry(conn->waiting_calls.next,
+ struct rxrpc_call, chan_wait_link);
+ u32 call_id = chan->call_counter + 1;
+
+ list_del_init(&call->chan_wait_link);
+ conn->active_chans |= 1 << channel;
+ call->peer = rxrpc_get_peer(conn->params.peer);
+ call->cid = conn->proto.cid | channel;
+ call->call_id = call_id;
+
+ _net("CONNECT call %08x:%08x as call %d on conn %d",
+ call->cid, call->call_id, call->debug_id, conn->debug_id);
+
+ /* Paired with the read barrier in rxrpc_wait_for_channel(). This
+ * orders cid and epoch in the connection wrt to call_id without the
+ * need to take the channel_lock.
+ *
+ * We provisionally assign a callNumber at this point, but we don't
+ * confirm it until the call is about to be exposed.
+ *
+ * TODO: Pair with a barrier in the data_ready handler when that looks
+ * at the call ID through a connection channel.
+ */
+ smp_wmb();
+ chan->call_id = call_id;
+ rcu_assign_pointer(chan->call, call);
+ wake_up(&call->waitq);
+}
+
+/*
+ * Assign channels and callNumbers to waiting calls.
+ */
+static void rxrpc_activate_channels(struct rxrpc_connection *conn)
+{
+ unsigned char mask;
+
+ _enter("%d", conn->debug_id);
+
+ if (conn->cache_state != RXRPC_CONN_CLIENT_ACTIVE ||
+ conn->active_chans == RXRPC_ACTIVE_CHANS_MASK)
+ return;
+
+ spin_lock(&conn->channel_lock);
+
+ while (!list_empty(&conn->waiting_calls) &&
+ (mask = ~conn->active_chans,
+ mask &= RXRPC_ACTIVE_CHANS_MASK,
+ mask != 0))
+ rxrpc_activate_one_channel(conn, __ffs(mask));
+
+ spin_unlock(&conn->channel_lock);
+ _leave("");
+}
+
+/*
+ * Wait for a callNumber and a channel to be granted to a call.
+ */
+static int rxrpc_wait_for_channel(struct rxrpc_call *call, gfp_t gfp)
+{
+ int ret = 0;
+
+ _enter("%d", call->debug_id);
+
+ if (!call->call_id) {
+ DECLARE_WAITQUEUE(myself, current);
- if (!atomic_add_unless(&conn->avail_chans, -1, 0)) {
if (!gfpflags_allow_blocking(gfp)) {
- rxrpc_put_connection(conn);
- _leave(" = -EAGAIN");
- return -EAGAIN;
+ ret = -EAGAIN;
+ goto out;
}
- add_wait_queue(&conn->channel_wq, &myself);
+ add_wait_queue_exclusive(&call->waitq, &myself);
for (;;) {
set_current_state(TASK_INTERRUPTIBLE);
- if (atomic_add_unless(&conn->avail_chans, -1, 0))
+ if (call->call_id)
+ break;
+ if (signal_pending(current)) {
+ ret = -ERESTARTSYS;
break;
- if (signal_pending(current))
- goto interrupted;
+ }
schedule();
}
- remove_wait_queue(&conn->channel_wq, &myself);
+ remove_wait_queue(&call->waitq, &myself);
__set_current_state(TASK_RUNNING);
}
- /* The connection allegedly now has a free channel and we can now
- * attach the call to it.
- */
+ /* Paired with the write barrier in rxrpc_activate_one_channel(). */
+ smp_rmb();
+
+out:
+ _leave(" = %d", ret);
+ return ret;
+}
+
+/*
+ * find a connection for a call
+ * - called in process context with IRQs enabled
+ */
+int rxrpc_connect_call(struct rxrpc_call *call,
+ struct rxrpc_conn_parameters *cp,
+ struct sockaddr_rxrpc *srx,
+ gfp_t gfp)
+{
+ int ret;
+
+ _enter("{%d,%lx},", call->debug_id, call->user_call_ID);
+
+ rxrpc_discard_expired_client_conns(NULL);
+ rxrpc_cull_active_client_conns();
+
+ ret = rxrpc_get_client_conn(call, cp, srx, gfp);
+ if (ret < 0)
+ return ret;
+
+ rxrpc_animate_client_conn(call->conn);
+ rxrpc_activate_channels(call->conn);
+
+ ret = rxrpc_wait_for_channel(call, gfp);
+ if (ret < 0)
+ rxrpc_disconnect_client_call(call);
+
+ _leave(" = %d", ret);
+ return ret;
+}
+
+/*
+ * Note that a connection is about to be exposed to the world. Once it is
+ * exposed, we maintain an extra ref on it that stops it from being summarily
+ * discarded before it's (a) had a chance to deal with retransmission and (b)
+ * had a chance at re-use (the per-connection security negotiation is
+ * expensive).
+ */
+static void rxrpc_expose_client_conn(struct rxrpc_connection *conn)
+{
+ if (!test_and_set_bit(RXRPC_CONN_EXPOSED, &conn->flags))
+ rxrpc_get_connection(conn);
+}
+
+/*
+ * Note that a call, and thus a connection, is about to be exposed to the
+ * world.
+ */
+void rxrpc_expose_client_call(struct rxrpc_call *call)
+{
+ struct rxrpc_connection *conn = call->conn;
+ struct rxrpc_channel *chan =
+ &conn->channels[call->cid & RXRPC_CHANNELMASK];
+
+ if (!test_and_set_bit(RXRPC_CALL_EXPOSED, &call->flags)) {
+ /* Mark the call ID as being used. If the callNumber counter
+ * exceeds ~2 billion, we kill the connection after its
+ * outstanding calls have finished so that the counter doesn't
+ * wrap.
+ */
+ chan->call_counter++;
+ if (chan->call_counter >= INT_MAX)
+ set_bit(RXRPC_CONN_DONT_REUSE, &conn->flags);
+ rxrpc_expose_client_conn(conn);
+ }
+}
+
+/*
+ * Disconnect a client call.
+ */
+void rxrpc_disconnect_client_call(struct rxrpc_call *call)
+{
+ unsigned int channel = call->cid & RXRPC_CHANNELMASK;
+ struct rxrpc_connection *conn = call->conn;
+ struct rxrpc_channel *chan = &conn->channels[channel];
+
+ call->conn = NULL;
+
spin_lock(&conn->channel_lock);
- for (chan = 0; chan < RXRPC_MAXCALLS; chan++)
- if (!conn->channels[chan].call)
- goto found_channel;
- BUG();
+ /* Calls that have never actually been assigned a channel can simply be
+ * discarded. If the conn didn't get used either, it will follow
+ * immediately unless someone else grabs it in the meantime.
+ */
+ if (!list_empty(&call->chan_wait_link)) {
+ _debug("call is waiting");
+ ASSERTCMP(call->call_id, ==, 0);
+ ASSERT(!test_bit(RXRPC_CALL_EXPOSED, &call->flags));
+ list_del_init(&call->chan_wait_link);
+
+ /* We must deactivate or idle the connection if it's now
+ * waiting for nothing.
+ */
+ spin_lock(&rxrpc_client_conn_cache_lock);
+ if (conn->cache_state == RXRPC_CONN_CLIENT_WAITING &&
+ list_empty(&conn->waiting_calls) &&
+ !conn->active_chans)
+ goto idle_connection;
+ goto out;
+ }
+
+ ASSERTCMP(rcu_access_pointer(chan->call), ==, call);
+ ASSERTCMP(atomic_read(&conn->usage), >=, 2);
+
+ /* If a client call was exposed to the world, we save the result for
+ * retransmission.
+ *
+ * We use a barrier here so that the call number and abort code can be
+ * read without needing to take a lock.
+ *
+ * TODO: Make the incoming packet handler check this and handle
+ * terminal retransmission without requiring access to the call.
+ */
+ if (test_bit(RXRPC_CALL_EXPOSED, &call->flags)) {
+ _debug("exposed %u,%u", call->call_id, call->local_abort);
+ __rxrpc_disconnect_call(conn, call);
+ }
+
+ /* See if we can pass the channel directly to another call. */
+ if (conn->cache_state == RXRPC_CONN_CLIENT_ACTIVE &&
+ !list_empty(&conn->waiting_calls)) {
+ _debug("pass chan");
+ rxrpc_activate_one_channel(conn, channel);
+ goto out_2;
+ }
+
+ /* Things are more complex and we need the cache lock. We might be
+ * able to simply idle the conn or it might now be lurking on the wait
+ * list. It might even get moved back to the active list whilst we're
+ * waiting for the lock.
+ */
+ spin_lock(&rxrpc_client_conn_cache_lock);
+
+ switch (conn->cache_state) {
+ case RXRPC_CONN_CLIENT_ACTIVE:
+ if (list_empty(&conn->waiting_calls)) {
+ rxrpc_deactivate_one_channel(conn, channel);
+ if (!conn->active_chans) {
+ rxrpc_nr_active_client_conns--;
+ goto idle_connection;
+ }
+ goto out;
+ }
+
+ _debug("pass chan 2");
+ rxrpc_activate_one_channel(conn, channel);
+ goto out;
+
+ case RXRPC_CONN_CLIENT_CULLED:
+ rxrpc_deactivate_one_channel(conn, channel);
+ ASSERT(list_empty(&conn->waiting_calls));
+ if (!conn->active_chans)
+ goto idle_connection;
+ goto out;
+
+ case RXRPC_CONN_CLIENT_WAITING:
+ rxrpc_deactivate_one_channel(conn, channel);
+ goto out;
+
+ default:
+ BUG();
+ }
-interrupted:
- remove_wait_queue(&conn->channel_wq, &myself);
- __set_current_state(TASK_RUNNING);
+out:
+ spin_unlock(&rxrpc_client_conn_cache_lock);
+out_2:
+ spin_unlock(&conn->channel_lock);
rxrpc_put_connection(conn);
- rxrpc_put_peer(cp->peer);
- cp->peer = NULL;
- _leave(" = -ERESTARTSYS");
- return -ERESTARTSYS;
+ _leave("");
+ return;
+
+idle_connection:
+ /* As no channels remain active, the connection gets deactivated
+ * immediately or moved to the idle list for a short while.
+ */
+ if (test_bit(RXRPC_CONN_EXPOSED, &conn->flags)) {
+ _debug("make idle");
+ conn->idle_timestamp = jiffies;
+ conn->cache_state = RXRPC_CONN_CLIENT_IDLE;
+ list_move_tail(&conn->cache_link, &rxrpc_idle_client_conns);
+ if (rxrpc_idle_client_conns.next == &conn->cache_link &&
+ !rxrpc_kill_all_client_conns)
+ queue_delayed_work(rxrpc_workqueue,
+ &rxrpc_client_conn_reap,
+ rxrpc_conn_idle_client_expiry);
+ } else {
+ _debug("make inactive");
+ conn->cache_state = RXRPC_CONN_CLIENT_INACTIVE;
+ list_del_init(&conn->cache_link);
+ }
+ goto out;
}
/*
- * Remove a client connection from the local endpoint's tree, thereby removing
- * it as a target for reuse for new client calls.
+ * Clean up a dead client connection.
*/
-void rxrpc_unpublish_client_conn(struct rxrpc_connection *conn)
+static struct rxrpc_connection *
+rxrpc_put_one_client_conn(struct rxrpc_connection *conn)
{
+ struct rxrpc_connection *next;
struct rxrpc_local *local = conn->params.local;
+ unsigned int nr_conns;
- spin_lock(&local->client_conns_lock);
- if (test_and_clear_bit(RXRPC_CONN_IN_CLIENT_CONNS, &conn->flags))
- rb_erase(&conn->client_node, &local->client_conns);
- spin_unlock(&local->client_conns_lock);
+ if (test_bit(RXRPC_CONN_IN_CLIENT_CONNS, &conn->flags)) {
+ spin_lock(&local->client_conns_lock);
+ if (test_and_clear_bit(RXRPC_CONN_IN_CLIENT_CONNS,
+ &conn->flags))
+ rb_erase(&conn->client_node, &local->client_conns);
+ spin_unlock(&local->client_conns_lock);
+ }
rxrpc_put_client_connection_id(conn);
+
+ ASSERTCMP(conn->cache_state, ==, RXRPC_CONN_CLIENT_INACTIVE);
+
+ if (!test_bit(RXRPC_CONN_COUNTED, &conn->flags))
+ return NULL;
+
+ spin_lock(&rxrpc_client_conn_cache_lock);
+ nr_conns = --rxrpc_nr_client_conns;
+
+ next = NULL;
+ if (nr_conns < rxrpc_max_client_connections &&
+ !list_empty(&rxrpc_waiting_client_conns)) {
+ next = list_entry(rxrpc_waiting_client_conns.next,
+ struct rxrpc_connection, cache_link);
+ rxrpc_get_connection(next);
+ rxrpc_activate_conn(next);
+ }
+
+ spin_unlock(&rxrpc_client_conn_cache_lock);
+ rxrpc_kill_connection(conn);
+
+ if (next)
+ rxrpc_activate_channels(next);
+
+ /* We need to get rid of the temporary ref we took upon next, but we
+ * can't call rxrpc_put_connection() recursively.
+ */
+ return next;
+}
+
+/*
+ * Clean up a dead client connections.
+ */
+void rxrpc_put_client_conn(struct rxrpc_connection *conn)
+{
+ struct rxrpc_connection *next;
+
+ do {
+ _enter("%p{u=%d,d=%d}",
+ conn, atomic_read(&conn->usage), conn->debug_id);
+
+ next = rxrpc_put_one_client_conn(conn);
+
+ if (!next)
+ break;
+ conn = next;
+ } while (atomic_dec_and_test(&conn->usage));
+
+ _leave("");
+}
+
+/*
+ * Kill the longest-active client connections to make room for new ones.
+ */
+static void rxrpc_cull_active_client_conns(void)
+{
+ struct rxrpc_connection *conn;
+ unsigned int nr_conns = rxrpc_nr_client_conns;
+ unsigned int nr_active, limit;
+
+ _enter("");
+
+ ASSERTCMP(nr_conns, >=, 0);
+ if (nr_conns < rxrpc_max_client_connections) {
+ _leave(" [ok]");
+ return;
+ }
+ limit = rxrpc_reap_client_connections;
+
+ spin_lock(&rxrpc_client_conn_cache_lock);
+ nr_active = rxrpc_nr_active_client_conns;
+
+ while (nr_active > limit) {
+ ASSERT(!list_empty(&rxrpc_active_client_conns));
+ conn = list_entry(rxrpc_active_client_conns.next,
+ struct rxrpc_connection, cache_link);
+ ASSERTCMP(conn->cache_state, ==, RXRPC_CONN_CLIENT_ACTIVE);
+
+ if (list_empty(&conn->waiting_calls)) {
+ conn->cache_state = RXRPC_CONN_CLIENT_CULLED;
+ list_del_init(&conn->cache_link);
+ } else {
+ conn->cache_state = RXRPC_CONN_CLIENT_WAITING;
+ list_move_tail(&conn->cache_link,
+ &rxrpc_waiting_client_conns);
+ }
+
+ nr_active--;
+ }
+
+ rxrpc_nr_active_client_conns = nr_active;
+ spin_unlock(&rxrpc_client_conn_cache_lock);
+ ASSERTCMP(nr_active, >=, 0);
+ _leave(" [culled]");
+}
+
+/*
+ * Discard expired client connections from the idle list. Each conn in the
+ * idle list has been exposed and holds an extra ref because of that.
+ *
+ * This may be called from conn setup or from a work item so cannot be
+ * considered non-reentrant.
+ */
+static void rxrpc_discard_expired_client_conns(struct work_struct *work)
+{
+ struct rxrpc_connection *conn;
+ unsigned long expiry, conn_expires_at, now;
+ unsigned int nr_conns;
+ bool did_discard = false;
+
+ _enter("%c", work ? 'w' : 'n');
+
+ if (list_empty(&rxrpc_idle_client_conns)) {
+ _leave(" [empty]");
+ return;
+ }
+
+ /* Don't double up on the discarding */
+ if (!spin_trylock(&rxrpc_client_conn_discard_mutex)) {
+ _leave(" [already]");
+ return;
+ }
+
+ /* We keep an estimate of what the number of conns ought to be after
+ * we've discarded some so that we don't overdo the discarding.
+ */
+ nr_conns = rxrpc_nr_client_conns;
+
+next:
+ spin_lock(&rxrpc_client_conn_cache_lock);
+
+ if (list_empty(&rxrpc_idle_client_conns))
+ goto out;
+
+ conn = list_entry(rxrpc_idle_client_conns.next,
+ struct rxrpc_connection, cache_link);
+ ASSERT(test_bit(RXRPC_CONN_EXPOSED, &conn->flags));
+
+ if (!rxrpc_kill_all_client_conns) {
+ /* If the number of connections is over the reap limit, we
+ * expedite discard by reducing the expiry timeout. We must,
+ * however, have at least a short grace period to be able to do
+ * final-ACK or ABORT retransmission.
+ */
+ expiry = rxrpc_conn_idle_client_expiry;
+ if (nr_conns > rxrpc_reap_client_connections)
+ expiry = rxrpc_conn_idle_client_fast_expiry;
+
+ conn_expires_at = conn->idle_timestamp + expiry;
+
+ now = READ_ONCE(jiffies);
+ if (time_after(conn_expires_at, now))
+ goto not_yet_expired;
+ }
+
+ _debug("discard conn %d", conn->debug_id);
+ if (!test_and_clear_bit(RXRPC_CONN_EXPOSED, &conn->flags))
+ BUG();
+ conn->cache_state = RXRPC_CONN_CLIENT_INACTIVE;
+ list_del_init(&conn->cache_link);
+
+ spin_unlock(&rxrpc_client_conn_cache_lock);
+
+ /* When we cleared the EXPOSED flag, we took on responsibility for the
+ * reference that that had on the usage count. We deal with that here.
+ * If someone re-sets the flag and re-gets the ref, that's fine.
+ */
+ rxrpc_put_connection(conn);
+ did_discard = true;
+ nr_conns--;
+ goto next;
+
+not_yet_expired:
+ /* The connection at the front of the queue hasn't yet expired, so
+ * schedule the work item for that point if we discarded something.
+ *
+ * We don't worry if the work item is already scheduled - it can look
+ * after rescheduling itself at a later time. We could cancel it, but
+ * then things get messier.
+ */
+ _debug("not yet");
+ if (!rxrpc_kill_all_client_conns)
+ queue_delayed_work(rxrpc_workqueue,
+ &rxrpc_client_conn_reap,
+ conn_expires_at - now);
+
+out:
+ spin_unlock(&rxrpc_client_conn_cache_lock);
+ spin_unlock(&rxrpc_client_conn_discard_mutex);
+ _leave("");
+}
+
+/*
+ * Preemptively destroy all the client connection records rather than waiting
+ * for them to time out
+ */
+void __exit rxrpc_destroy_all_client_connections(void)
+{
+ _enter("");
+
+ spin_lock(&rxrpc_client_conn_cache_lock);
+ rxrpc_kill_all_client_conns = true;
+ spin_unlock(&rxrpc_client_conn_cache_lock);
+
+ cancel_delayed_work(&rxrpc_client_conn_reap);
+
+ if (!queue_delayed_work(rxrpc_workqueue, &rxrpc_client_conn_reap, 0))
+ _debug("destroy: queue failed");
+
+ _leave("");
}