net/sched/sch_plug.c

0001 // SPDX-License-Identifier: GPL-2.0-or-later
0002 /*
0003  * sch_plug.c Queue traffic until an explicit release command
0004  *
0005  * There are two ways to use this qdisc:
0006  * 1. A simple "instantaneous" plug/unplug operation, by issuing an alternating
0007  *    sequence of TCQ_PLUG_BUFFER & TCQ_PLUG_RELEASE_INDEFINITE commands.
0008  *
0009  * 2. For network output buffering (a.k.a output commit) functionality.
0010  *    Output commit property is commonly used by applications using checkpoint
0011  *    based fault-tolerance to ensure that the checkpoint from which a system
0012  *    is being restored is consistent w.r.t outside world.
0013  *
0014  *    Consider for e.g. Remus - a Virtual Machine checkpointing system,
0015  *    wherein a VM is checkpointed, say every 50ms. The checkpoint is replicated
0016  *    asynchronously to the backup host, while the VM continues executing the
0017  *    next epoch speculatively.
0018  *
0019  *    The following is a typical sequence of output buffer operations:
0020  *       1.At epoch i, start_buffer(i)
0021  *       2. At end of epoch i (i.e. after 50ms):
0022  *          2.1 Stop VM and take checkpoint(i).
0023  *          2.2 start_buffer(i+1) and Resume VM
0024  *       3. While speculatively executing epoch(i+1), asynchronously replicate
0025  *          checkpoint(i) to backup host.
0026  *       4. When checkpoint_ack(i) is received from backup, release_buffer(i)
0027  *    Thus, this Qdisc would receive the following sequence of commands:
0028  *       TCQ_PLUG_BUFFER (epoch i)
0029  *       .. TCQ_PLUG_BUFFER (epoch i+1)
0030  *       ....TCQ_PLUG_RELEASE_ONE (epoch i)
0031  *       ......TCQ_PLUG_BUFFER (epoch i+2)
0032  *       ........
0033  */
0034
0035 #include <linux/module.h>
0036 #include <linux/types.h>
0037 #include <linux/kernel.h>
0038 #include <linux/errno.h>
0039 #include <linux/netdevice.h>
0040 #include <linux/skbuff.h>
0041 #include <net/pkt_sched.h>
0042
0043 /*
0044  * State of the queue, when used for network output buffering:
0045  *
0046  *                 plug(i+1)            plug(i)          head
0047  * ------------------+--------------------+---------------->
0048  *                   |                    |
0049  *                   |                    |
0050  * pkts_current_epoch| pkts_last_epoch    |pkts_to_release
0051  * ----------------->|<--------+--------->|+--------------->
0052  *                   v                    v
0053  *
0054  */
0055
0056 struct plug_sched_data {
0057     /* If true, the dequeue function releases all packets
0058      * from head to end of the queue. The queue turns into
0059      * a pass-through queue for newly arriving packets.
0060      */
0061     bool unplug_indefinite;
0062
0063     bool throttled;
0064
0065     /* Queue Limit in bytes */
0066     u32 limit;
0067
0068     /* Number of packets (output) from the current speculatively
0069      * executing epoch.
0070      */
0071     u32 pkts_current_epoch;
0072
0073     /* Number of packets corresponding to the recently finished
0074      * epoch. These will be released when we receive a
0075      * TCQ_PLUG_RELEASE_ONE command. This command is typically
0076      * issued after committing a checkpoint at the target.
0077      */
0078     u32 pkts_last_epoch;
0079
0080     /*
0081      * Number of packets from the head of the queue, that can
0082      * be released (committed checkpoint).
0083      */
0084     u32 pkts_to_release;
0085 };
0086
0087 static int plug_enqueue(struct sk_buff *skb, struct Qdisc *sch,
0088             struct sk_buff **to_free)
0089 {
0090     struct plug_sched_data *q = qdisc_priv(sch);
0091
0092     if (likely(sch->qstats.backlog + skb->len <= q->limit)) {
0093         if (!q->unplug_indefinite)
0094             q->pkts_current_epoch++;
0095         return qdisc_enqueue_tail(skb, sch);
0096     }
0097
0098     return qdisc_drop(skb, sch, to_free);
0099 }
0100
0101 static struct sk_buff *plug_dequeue(struct Qdisc *sch)
0102 {
0103     struct plug_sched_data *q = qdisc_priv(sch);
0104
0105     if (q->throttled)
0106         return NULL;
0107
0108     if (!q->unplug_indefinite) {
0109         if (!q->pkts_to_release) {
0110             /* No more packets to dequeue. Block the queue
0111              * and wait for the next release command.
0112              */
0113             q->throttled = true;
0114             return NULL;
0115         }
0116         q->pkts_to_release--;
0117     }
0118
0119     return qdisc_dequeue_head(sch);
0120 }
0121
0122 static int plug_init(struct Qdisc *sch, struct nlattr *opt,
0123              struct netlink_ext_ack *extack)
0124 {
0125     struct plug_sched_data *q = qdisc_priv(sch);
0126
0127     q->pkts_current_epoch = 0;
0128     q->pkts_last_epoch = 0;
0129     q->pkts_to_release = 0;
0130     q->unplug_indefinite = false;
0131
0132     if (opt == NULL) {
0133         q->limit = qdisc_dev(sch)->tx_queue_len
0134                    * psched_mtu(qdisc_dev(sch));
0135     } else {
0136         struct tc_plug_qopt *ctl = nla_data(opt);
0137
0138         if (nla_len(opt) < sizeof(*ctl))
0139             return -EINVAL;
0140
0141         q->limit = ctl->limit;
0142     }
0143
0144     q->throttled = true;
0145     return 0;
0146 }
0147
0148 /* Receives 4 types of messages:
0149  * TCQ_PLUG_BUFFER: Inset a plug into the queue and
0150  *  buffer any incoming packets
0151  * TCQ_PLUG_RELEASE_ONE: Dequeue packets from queue head
0152  *   to beginning of the next plug.
0153  * TCQ_PLUG_RELEASE_INDEFINITE: Dequeue all packets from queue.
0154  *   Stop buffering packets until the next TCQ_PLUG_BUFFER
0155  *   command is received (just act as a pass-thru queue).
0156  * TCQ_PLUG_LIMIT: Increase/decrease queue size
0157  */
0158 static int plug_change(struct Qdisc *sch, struct nlattr *opt,
0159                struct netlink_ext_ack *extack)
0160 {
0161     struct plug_sched_data *q = qdisc_priv(sch);
0162     struct tc_plug_qopt *msg;
0163
0164     if (opt == NULL)
0165         return -EINVAL;
0166
0167     msg = nla_data(opt);
0168     if (nla_len(opt) < sizeof(*msg))
0169         return -EINVAL;
0170
0171     switch (msg->action) {
0172     case TCQ_PLUG_BUFFER:
0173         /* Save size of the current buffer */
0174         q->pkts_last_epoch = q->pkts_current_epoch;
0175         q->pkts_current_epoch = 0;
0176         if (q->unplug_indefinite)
0177             q->throttled = true;
0178         q->unplug_indefinite = false;
0179         break;
0180     case TCQ_PLUG_RELEASE_ONE:
0181         /* Add packets from the last complete buffer to the
0182          * packets to be released set.
0183          */
0184         q->pkts_to_release += q->pkts_last_epoch;
0185         q->pkts_last_epoch = 0;
0186         q->throttled = false;
0187         netif_schedule_queue(sch->dev_queue);
0188         break;
0189     case TCQ_PLUG_RELEASE_INDEFINITE:
0190         q->unplug_indefinite = true;
0191         q->pkts_to_release = 0;
0192         q->pkts_last_epoch = 0;
0193         q->pkts_current_epoch = 0;
0194         q->throttled = false;
0195         netif_schedule_queue(sch->dev_queue);
0196         break;
0197     case TCQ_PLUG_LIMIT:
0198         /* Limit is supplied in bytes */
0199         q->limit = msg->limit;
0200         break;
0201     default:
0202         return -EINVAL;
0203     }
0204
0205     return 0;
0206 }
0207
0208 static struct Qdisc_ops plug_qdisc_ops __read_mostly = {
0209     .id          =       "plug",
0210     .priv_size   =       sizeof(struct plug_sched_data),
0211     .enqueue     =       plug_enqueue,
0212     .dequeue     =       plug_dequeue,
0213     .peek        =       qdisc_peek_head,
0214     .init        =       plug_init,
0215     .change      =       plug_change,
0216     .reset       =       qdisc_reset_queue,
0217     .owner       =       THIS_MODULE,
0218 };
0219
0220 static int __init plug_module_init(void)
0221 {
0222     return register_qdisc(&plug_qdisc_ops);
0223 }
0224
0225 static void __exit plug_module_exit(void)
0226 {
0227     unregister_qdisc(&plug_qdisc_ops);
0228 }
0229 module_init(plug_module_init)
0230 module_exit(plug_module_exit)
0231 MODULE_LICENSE("GPL");