Add the rt linux 4.1.3-rt3 as base
[kvmfornfv.git] / kernel / net / sched / sch_plug.c
diff --git a/kernel/net/sched/sch_plug.c b/kernel/net/sched/sch_plug.c
new file mode 100644 (file)
index 0000000..89f8fcf
--- /dev/null
@@ -0,0 +1,233 @@
+/*
+ * sch_plug.c Queue traffic until an explicit release command
+ *
+ *             This program is free software; you can redistribute it and/or
+ *             modify it under the terms of the GNU General Public License
+ *             as published by the Free Software Foundation; either version
+ *             2 of the License, or (at your option) any later version.
+ *
+ * There are two ways to use this qdisc:
+ * 1. A simple "instantaneous" plug/unplug operation, by issuing an alternating
+ *    sequence of TCQ_PLUG_BUFFER & TCQ_PLUG_RELEASE_INDEFINITE commands.
+ *
+ * 2. For network output buffering (a.k.a output commit) functionality.
+ *    Output commit property is commonly used by applications using checkpoint
+ *    based fault-tolerance to ensure that the checkpoint from which a system
+ *    is being restored is consistent w.r.t outside world.
+ *
+ *    Consider for e.g. Remus - a Virtual Machine checkpointing system,
+ *    wherein a VM is checkpointed, say every 50ms. The checkpoint is replicated
+ *    asynchronously to the backup host, while the VM continues executing the
+ *    next epoch speculatively.
+ *
+ *    The following is a typical sequence of output buffer operations:
+ *       1.At epoch i, start_buffer(i)
+ *       2. At end of epoch i (i.e. after 50ms):
+ *          2.1 Stop VM and take checkpoint(i).
+ *          2.2 start_buffer(i+1) and Resume VM
+ *       3. While speculatively executing epoch(i+1), asynchronously replicate
+ *          checkpoint(i) to backup host.
+ *       4. When checkpoint_ack(i) is received from backup, release_buffer(i)
+ *    Thus, this Qdisc would receive the following sequence of commands:
+ *       TCQ_PLUG_BUFFER (epoch i)
+ *       .. TCQ_PLUG_BUFFER (epoch i+1)
+ *       ....TCQ_PLUG_RELEASE_ONE (epoch i)
+ *       ......TCQ_PLUG_BUFFER (epoch i+2)
+ *       ........
+ */
+
+#include <linux/module.h>
+#include <linux/types.h>
+#include <linux/kernel.h>
+#include <linux/errno.h>
+#include <linux/netdevice.h>
+#include <linux/skbuff.h>
+#include <net/pkt_sched.h>
+
+/*
+ * State of the queue, when used for network output buffering:
+ *
+ *                 plug(i+1)            plug(i)          head
+ * ------------------+--------------------+---------------->
+ *                   |                    |
+ *                   |                    |
+ * pkts_current_epoch| pkts_last_epoch    |pkts_to_release
+ * ----------------->|<--------+--------->|+--------------->
+ *                   v                    v
+ *
+ */
+
+struct plug_sched_data {
+       /* If true, the dequeue function releases all packets
+        * from head to end of the queue. The queue turns into
+        * a pass-through queue for newly arriving packets.
+        */
+       bool unplug_indefinite;
+
+       /* Queue Limit in bytes */
+       u32 limit;
+
+       /* Number of packets (output) from the current speculatively
+        * executing epoch.
+        */
+       u32 pkts_current_epoch;
+
+       /* Number of packets corresponding to the recently finished
+        * epoch. These will be released when we receive a
+        * TCQ_PLUG_RELEASE_ONE command. This command is typically
+        * issued after committing a checkpoint at the target.
+        */
+       u32 pkts_last_epoch;
+
+       /*
+        * Number of packets from the head of the queue, that can
+        * be released (committed checkpoint).
+        */
+       u32 pkts_to_release;
+};
+
+static int plug_enqueue(struct sk_buff *skb, struct Qdisc *sch)
+{
+       struct plug_sched_data *q = qdisc_priv(sch);
+
+       if (likely(sch->qstats.backlog + skb->len <= q->limit)) {
+               if (!q->unplug_indefinite)
+                       q->pkts_current_epoch++;
+               return qdisc_enqueue_tail(skb, sch);
+       }
+
+       return qdisc_reshape_fail(skb, sch);
+}
+
+static struct sk_buff *plug_dequeue(struct Qdisc *sch)
+{
+       struct plug_sched_data *q = qdisc_priv(sch);
+
+       if (qdisc_is_throttled(sch))
+               return NULL;
+
+       if (!q->unplug_indefinite) {
+               if (!q->pkts_to_release) {
+                       /* No more packets to dequeue. Block the queue
+                        * and wait for the next release command.
+                        */
+                       qdisc_throttled(sch);
+                       return NULL;
+               }
+               q->pkts_to_release--;
+       }
+
+       return qdisc_dequeue_head(sch);
+}
+
+static int plug_init(struct Qdisc *sch, struct nlattr *opt)
+{
+       struct plug_sched_data *q = qdisc_priv(sch);
+
+       q->pkts_current_epoch = 0;
+       q->pkts_last_epoch = 0;
+       q->pkts_to_release = 0;
+       q->unplug_indefinite = false;
+
+       if (opt == NULL) {
+               /* We will set a default limit of 100 pkts (~150kB)
+                * in case tx_queue_len is not available. The
+                * default value is completely arbitrary.
+                */
+               u32 pkt_limit = qdisc_dev(sch)->tx_queue_len ? : 100;
+               q->limit = pkt_limit * psched_mtu(qdisc_dev(sch));
+       } else {
+               struct tc_plug_qopt *ctl = nla_data(opt);
+
+               if (nla_len(opt) < sizeof(*ctl))
+                       return -EINVAL;
+
+               q->limit = ctl->limit;
+       }
+
+       qdisc_throttled(sch);
+       return 0;
+}
+
+/* Receives 4 types of messages:
+ * TCQ_PLUG_BUFFER: Inset a plug into the queue and
+ *  buffer any incoming packets
+ * TCQ_PLUG_RELEASE_ONE: Dequeue packets from queue head
+ *   to beginning of the next plug.
+ * TCQ_PLUG_RELEASE_INDEFINITE: Dequeue all packets from queue.
+ *   Stop buffering packets until the next TCQ_PLUG_BUFFER
+ *   command is received (just act as a pass-thru queue).
+ * TCQ_PLUG_LIMIT: Increase/decrease queue size
+ */
+static int plug_change(struct Qdisc *sch, struct nlattr *opt)
+{
+       struct plug_sched_data *q = qdisc_priv(sch);
+       struct tc_plug_qopt *msg;
+
+       if (opt == NULL)
+               return -EINVAL;
+
+       msg = nla_data(opt);
+       if (nla_len(opt) < sizeof(*msg))
+               return -EINVAL;
+
+       switch (msg->action) {
+       case TCQ_PLUG_BUFFER:
+               /* Save size of the current buffer */
+               q->pkts_last_epoch = q->pkts_current_epoch;
+               q->pkts_current_epoch = 0;
+               if (q->unplug_indefinite)
+                       qdisc_throttled(sch);
+               q->unplug_indefinite = false;
+               break;
+       case TCQ_PLUG_RELEASE_ONE:
+               /* Add packets from the last complete buffer to the
+                * packets to be released set.
+                */
+               q->pkts_to_release += q->pkts_last_epoch;
+               q->pkts_last_epoch = 0;
+               qdisc_unthrottled(sch);
+               netif_schedule_queue(sch->dev_queue);
+               break;
+       case TCQ_PLUG_RELEASE_INDEFINITE:
+               q->unplug_indefinite = true;
+               q->pkts_to_release = 0;
+               q->pkts_last_epoch = 0;
+               q->pkts_current_epoch = 0;
+               qdisc_unthrottled(sch);
+               netif_schedule_queue(sch->dev_queue);
+               break;
+       case TCQ_PLUG_LIMIT:
+               /* Limit is supplied in bytes */
+               q->limit = msg->limit;
+               break;
+       default:
+               return -EINVAL;
+       }
+
+       return 0;
+}
+
+static struct Qdisc_ops plug_qdisc_ops __read_mostly = {
+       .id          =       "plug",
+       .priv_size   =       sizeof(struct plug_sched_data),
+       .enqueue     =       plug_enqueue,
+       .dequeue     =       plug_dequeue,
+       .peek        =       qdisc_peek_head,
+       .init        =       plug_init,
+       .change      =       plug_change,
+       .owner       =       THIS_MODULE,
+};
+
+static int __init plug_module_init(void)
+{
+       return register_qdisc(&plug_qdisc_ops);
+}
+
+static void __exit plug_module_exit(void)
+{
+       unregister_qdisc(&plug_qdisc_ops);
+}
+module_init(plug_module_init)
+module_exit(plug_module_exit)
+MODULE_LICENSE("GPL");