linux/net/sched/sch_plug.c
<<
>>
Prefs
   1/*
   2 * sch_plug.c Queue traffic until an explicit release command
   3 *
   4 *             This program is free software; you can redistribute it and/or
   5 *             modify it under the terms of the GNU General Public License
   6 *             as published by the Free Software Foundation; either version
   7 *             2 of the License, or (at your option) any later version.
   8 *
   9 * There are two ways to use this qdisc:
  10 * 1. A simple "instantaneous" plug/unplug operation, by issuing an alternating
  11 *    sequence of TCQ_PLUG_BUFFER & TCQ_PLUG_RELEASE_INDEFINITE commands.
  12 *
  13 * 2. For network output buffering (a.k.a output commit) functionality.
  14 *    Output commit property is commonly used by applications using checkpoint
  15 *    based fault-tolerance to ensure that the checkpoint from which a system
  16 *    is being restored is consistent w.r.t outside world.
  17 *
  18 *    Consider for e.g. Remus - a Virtual Machine checkpointing system,
  19 *    wherein a VM is checkpointed, say every 50ms. The checkpoint is replicated
  20 *    asynchronously to the backup host, while the VM continues executing the
  21 *    next epoch speculatively.
  22 *
  23 *    The following is a typical sequence of output buffer operations:
  24 *       1.At epoch i, start_buffer(i)
  25 *       2. At end of epoch i (i.e. after 50ms):
  26 *          2.1 Stop VM and take checkpoint(i).
  27 *          2.2 start_buffer(i+1) and Resume VM
  28 *       3. While speculatively executing epoch(i+1), asynchronously replicate
  29 *          checkpoint(i) to backup host.
  30 *       4. When checkpoint_ack(i) is received from backup, release_buffer(i)
  31 *    Thus, this Qdisc would receive the following sequence of commands:
  32 *       TCQ_PLUG_BUFFER (epoch i)
  33 *       .. TCQ_PLUG_BUFFER (epoch i+1)
  34 *       ....TCQ_PLUG_RELEASE_ONE (epoch i)
  35 *       ......TCQ_PLUG_BUFFER (epoch i+2)
  36 *       ........
  37 */
  38
  39#include <linux/module.h>
  40#include <linux/types.h>
  41#include <linux/kernel.h>
  42#include <linux/errno.h>
  43#include <linux/netdevice.h>
  44#include <linux/skbuff.h>
  45#include <net/pkt_sched.h>
  46
  47/*
  48 * State of the queue, when used for network output buffering:
  49 *
  50 *                 plug(i+1)            plug(i)          head
  51 * ------------------+--------------------+---------------->
  52 *                   |                    |
  53 *                   |                    |
  54 * pkts_current_epoch| pkts_last_epoch    |pkts_to_release
  55 * ----------------->|<--------+--------->|+--------------->
  56 *                   v                    v
  57 *
  58 */
  59
  60struct plug_sched_data {
  61        /* If true, the dequeue function releases all packets
  62         * from head to end of the queue. The queue turns into
  63         * a pass-through queue for newly arriving packets.
  64         */
  65        bool unplug_indefinite;
  66
  67        /* Queue Limit in bytes */
  68        u32 limit;
  69
  70        /* Number of packets (output) from the current speculatively
  71         * executing epoch.
  72         */
  73        u32 pkts_current_epoch;
  74
  75        /* Number of packets corresponding to the recently finished
  76         * epoch. These will be released when we receive a
  77         * TCQ_PLUG_RELEASE_ONE command. This command is typically
  78         * issued after committing a checkpoint at the target.
  79         */
  80        u32 pkts_last_epoch;
  81
  82        /*
  83         * Number of packets from the head of the queue, that can
  84         * be released (committed checkpoint).
  85         */
  86        u32 pkts_to_release;
  87};
  88
  89static int plug_enqueue(struct sk_buff *skb, struct Qdisc *sch)
  90{
  91        struct plug_sched_data *q = qdisc_priv(sch);
  92
  93        if (likely(sch->qstats.backlog + skb->len <= q->limit)) {
  94                if (!q->unplug_indefinite)
  95                        q->pkts_current_epoch++;
  96                return qdisc_enqueue_tail(skb, sch);
  97        }
  98
  99        return qdisc_reshape_fail(skb, sch);
 100}
 101
 102static struct sk_buff *plug_dequeue(struct Qdisc *sch)
 103{
 104        struct plug_sched_data *q = qdisc_priv(sch);
 105
 106        if (qdisc_is_throttled(sch))
 107                return NULL;
 108
 109        if (!q->unplug_indefinite) {
 110                if (!q->pkts_to_release) {
 111                        /* No more packets to dequeue. Block the queue
 112                         * and wait for the next release command.
 113                         */
 114                        qdisc_throttled(sch);
 115                        return NULL;
 116                }
 117                q->pkts_to_release--;
 118        }
 119
 120        return qdisc_dequeue_head(sch);
 121}
 122
 123static int plug_init(struct Qdisc *sch, struct nlattr *opt)
 124{
 125        struct plug_sched_data *q = qdisc_priv(sch);
 126
 127        q->pkts_current_epoch = 0;
 128        q->pkts_last_epoch = 0;
 129        q->pkts_to_release = 0;
 130        q->unplug_indefinite = false;
 131
 132        if (opt == NULL) {
 133                /* We will set a default limit of 100 pkts (~150kB)
 134                 * in case tx_queue_len is not available. The
 135                 * default value is completely arbitrary.
 136                 */
 137                u32 pkt_limit = qdisc_dev(sch)->tx_queue_len ? : 100;
 138                q->limit = pkt_limit * psched_mtu(qdisc_dev(sch));
 139        } else {
 140                struct tc_plug_qopt *ctl = nla_data(opt);
 141
 142                if (nla_len(opt) < sizeof(*ctl))
 143                        return -EINVAL;
 144
 145                q->limit = ctl->limit;
 146        }
 147
 148        qdisc_throttled(sch);
 149        return 0;
 150}
 151
 152/* Receives 4 types of messages:
 153 * TCQ_PLUG_BUFFER: Inset a plug into the queue and
 154 *  buffer any incoming packets
 155 * TCQ_PLUG_RELEASE_ONE: Dequeue packets from queue head
 156 *   to beginning of the next plug.
 157 * TCQ_PLUG_RELEASE_INDEFINITE: Dequeue all packets from queue.
 158 *   Stop buffering packets until the next TCQ_PLUG_BUFFER
 159 *   command is received (just act as a pass-thru queue).
 160 * TCQ_PLUG_LIMIT: Increase/decrease queue size
 161 */
 162static int plug_change(struct Qdisc *sch, struct nlattr *opt)
 163{
 164        struct plug_sched_data *q = qdisc_priv(sch);
 165        struct tc_plug_qopt *msg;
 166
 167        if (opt == NULL)
 168                return -EINVAL;
 169
 170        msg = nla_data(opt);
 171        if (nla_len(opt) < sizeof(*msg))
 172                return -EINVAL;
 173
 174        switch (msg->action) {
 175        case TCQ_PLUG_BUFFER:
 176                /* Save size of the current buffer */
 177                q->pkts_last_epoch = q->pkts_current_epoch;
 178                q->pkts_current_epoch = 0;
 179                if (q->unplug_indefinite)
 180                        qdisc_throttled(sch);
 181                q->unplug_indefinite = false;
 182                break;
 183        case TCQ_PLUG_RELEASE_ONE:
 184                /* Add packets from the last complete buffer to the
 185                 * packets to be released set.
 186                 */
 187                q->pkts_to_release += q->pkts_last_epoch;
 188                q->pkts_last_epoch = 0;
 189                qdisc_unthrottled(sch);
 190                netif_schedule_queue(sch->dev_queue);
 191                break;
 192        case TCQ_PLUG_RELEASE_INDEFINITE:
 193                q->unplug_indefinite = true;
 194                q->pkts_to_release = 0;
 195                q->pkts_last_epoch = 0;
 196                q->pkts_current_epoch = 0;
 197                qdisc_unthrottled(sch);
 198                netif_schedule_queue(sch->dev_queue);
 199                break;
 200        case TCQ_PLUG_LIMIT:
 201                /* Limit is supplied in bytes */
 202                q->limit = msg->limit;
 203                break;
 204        default:
 205                return -EINVAL;
 206        }
 207
 208        return 0;
 209}
 210
 211static struct Qdisc_ops plug_qdisc_ops __read_mostly = {
 212        .id          =       "plug",
 213        .priv_size   =       sizeof(struct plug_sched_data),
 214        .enqueue     =       plug_enqueue,
 215        .dequeue     =       plug_dequeue,
 216        .peek        =       qdisc_peek_head,
 217        .init        =       plug_init,
 218        .change      =       plug_change,
 219        .owner       =       THIS_MODULE,
 220};
 221
 222static int __init plug_module_init(void)
 223{
 224        return register_qdisc(&plug_qdisc_ops);
 225}
 226
 227static void __exit plug_module_exit(void)
 228{
 229        unregister_qdisc(&plug_qdisc_ops);
 230}
 231module_init(plug_module_init)
 232module_exit(plug_module_exit)
 233MODULE_LICENSE("GPL");
 234