Thomas Gleixner | 2874c5f | 2019-05-27 08:55:01 +0200 | [diff] [blame] | 1 | // SPDX-License-Identifier: GPL-2.0-or-later |
Shriram Rajagopalan | c3059be | 2012-02-05 13:51:32 +0000 | [diff] [blame] | 2 | /* |
| 3 | * sch_plug.c Queue traffic until an explicit release command |
| 4 | * |
Shriram Rajagopalan | c3059be | 2012-02-05 13:51:32 +0000 | [diff] [blame] | 5 | * There are two ways to use this qdisc: |
| 6 | * 1. A simple "instantaneous" plug/unplug operation, by issuing an alternating |
| 7 | * sequence of TCQ_PLUG_BUFFER & TCQ_PLUG_RELEASE_INDEFINITE commands. |
| 8 | * |
| 9 | * 2. For network output buffering (a.k.a output commit) functionality. |
| 10 | * Output commit property is commonly used by applications using checkpoint |
| 11 | * based fault-tolerance to ensure that the checkpoint from which a system |
| 12 | * is being restored is consistent w.r.t outside world. |
| 13 | * |
| 14 | * Consider for e.g. Remus - a Virtual Machine checkpointing system, |
| 15 | * wherein a VM is checkpointed, say every 50ms. The checkpoint is replicated |
| 16 | * asynchronously to the backup host, while the VM continues executing the |
| 17 | * next epoch speculatively. |
| 18 | * |
| 19 | * The following is a typical sequence of output buffer operations: |
| 20 | * 1.At epoch i, start_buffer(i) |
| 21 | * 2. At end of epoch i (i.e. after 50ms): |
| 22 | * 2.1 Stop VM and take checkpoint(i). |
| 23 | * 2.2 start_buffer(i+1) and Resume VM |
| 24 | * 3. While speculatively executing epoch(i+1), asynchronously replicate |
| 25 | * checkpoint(i) to backup host. |
| 26 | * 4. When checkpoint_ack(i) is received from backup, release_buffer(i) |
| 27 | * Thus, this Qdisc would receive the following sequence of commands: |
| 28 | * TCQ_PLUG_BUFFER (epoch i) |
| 29 | * .. TCQ_PLUG_BUFFER (epoch i+1) |
| 30 | * ....TCQ_PLUG_RELEASE_ONE (epoch i) |
| 31 | * ......TCQ_PLUG_BUFFER (epoch i+2) |
| 32 | * ........ |
| 33 | */ |
| 34 | |
| 35 | #include <linux/module.h> |
| 36 | #include <linux/types.h> |
| 37 | #include <linux/kernel.h> |
| 38 | #include <linux/errno.h> |
| 39 | #include <linux/netdevice.h> |
| 40 | #include <linux/skbuff.h> |
| 41 | #include <net/pkt_sched.h> |
| 42 | |
| 43 | /* |
| 44 | * State of the queue, when used for network output buffering: |
| 45 | * |
| 46 | * plug(i+1) plug(i) head |
| 47 | * ------------------+--------------------+----------------> |
| 48 | * | | |
| 49 | * | | |
| 50 | * pkts_current_epoch| pkts_last_epoch |pkts_to_release |
| 51 | * ----------------->|<--------+--------->|+---------------> |
| 52 | * v v |
| 53 | * |
| 54 | */ |
| 55 | |
| 56 | struct plug_sched_data { |
| 57 | /* If true, the dequeue function releases all packets |
| 58 | * from head to end of the queue. The queue turns into |
| 59 | * a pass-through queue for newly arriving packets. |
| 60 | */ |
| 61 | bool unplug_indefinite; |
| 62 | |
Eric Dumazet | 8fe6a79 | 2016-06-10 16:41:36 -0700 | [diff] [blame] | 63 | bool throttled; |
| 64 | |
Shriram Rajagopalan | c3059be | 2012-02-05 13:51:32 +0000 | [diff] [blame] | 65 | /* Queue Limit in bytes */ |
| 66 | u32 limit; |
| 67 | |
| 68 | /* Number of packets (output) from the current speculatively |
| 69 | * executing epoch. |
| 70 | */ |
| 71 | u32 pkts_current_epoch; |
| 72 | |
| 73 | /* Number of packets corresponding to the recently finished |
| 74 | * epoch. These will be released when we receive a |
| 75 | * TCQ_PLUG_RELEASE_ONE command. This command is typically |
| 76 | * issued after committing a checkpoint at the target. |
| 77 | */ |
| 78 | u32 pkts_last_epoch; |
| 79 | |
| 80 | /* |
| 81 | * Number of packets from the head of the queue, that can |
| 82 | * be released (committed checkpoint). |
| 83 | */ |
| 84 | u32 pkts_to_release; |
| 85 | }; |
| 86 | |
Eric Dumazet | 520ac30 | 2016-06-21 23:16:49 -0700 | [diff] [blame] | 87 | static int plug_enqueue(struct sk_buff *skb, struct Qdisc *sch, |
| 88 | struct sk_buff **to_free) |
Shriram Rajagopalan | c3059be | 2012-02-05 13:51:32 +0000 | [diff] [blame] | 89 | { |
| 90 | struct plug_sched_data *q = qdisc_priv(sch); |
| 91 | |
| 92 | if (likely(sch->qstats.backlog + skb->len <= q->limit)) { |
| 93 | if (!q->unplug_indefinite) |
| 94 | q->pkts_current_epoch++; |
| 95 | return qdisc_enqueue_tail(skb, sch); |
| 96 | } |
| 97 | |
Eric Dumazet | 520ac30 | 2016-06-21 23:16:49 -0700 | [diff] [blame] | 98 | return qdisc_drop(skb, sch, to_free); |
Shriram Rajagopalan | c3059be | 2012-02-05 13:51:32 +0000 | [diff] [blame] | 99 | } |
| 100 | |
| 101 | static struct sk_buff *plug_dequeue(struct Qdisc *sch) |
| 102 | { |
| 103 | struct plug_sched_data *q = qdisc_priv(sch); |
| 104 | |
Eric Dumazet | 8fe6a79 | 2016-06-10 16:41:36 -0700 | [diff] [blame] | 105 | if (q->throttled) |
Shriram Rajagopalan | c3059be | 2012-02-05 13:51:32 +0000 | [diff] [blame] | 106 | return NULL; |
| 107 | |
| 108 | if (!q->unplug_indefinite) { |
| 109 | if (!q->pkts_to_release) { |
| 110 | /* No more packets to dequeue. Block the queue |
| 111 | * and wait for the next release command. |
| 112 | */ |
Eric Dumazet | 8fe6a79 | 2016-06-10 16:41:36 -0700 | [diff] [blame] | 113 | q->throttled = true; |
Shriram Rajagopalan | c3059be | 2012-02-05 13:51:32 +0000 | [diff] [blame] | 114 | return NULL; |
| 115 | } |
| 116 | q->pkts_to_release--; |
| 117 | } |
| 118 | |
| 119 | return qdisc_dequeue_head(sch); |
| 120 | } |
| 121 | |
Alexander Aring | e63d7dfd | 2017-12-20 12:35:13 -0500 | [diff] [blame] | 122 | static int plug_init(struct Qdisc *sch, struct nlattr *opt, |
| 123 | struct netlink_ext_ack *extack) |
Shriram Rajagopalan | c3059be | 2012-02-05 13:51:32 +0000 | [diff] [blame] | 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) { |
Phil Sutter | 348e343 | 2015-08-18 10:30:49 +0200 | [diff] [blame] | 133 | q->limit = qdisc_dev(sch)->tx_queue_len |
| 134 | * psched_mtu(qdisc_dev(sch)); |
Shriram Rajagopalan | c3059be | 2012-02-05 13:51:32 +0000 | [diff] [blame] | 135 | } else { |
| 136 | struct tc_plug_qopt *ctl = nla_data(opt); |
| 137 | |
| 138 | if (nla_len(opt) < sizeof(*ctl)) |
| 139 | return -EINVAL; |
| 140 | |
| 141 | q->limit = ctl->limit; |
| 142 | } |
| 143 | |
Eric Dumazet | 8fe6a79 | 2016-06-10 16:41:36 -0700 | [diff] [blame] | 144 | q->throttled = true; |
Shriram Rajagopalan | c3059be | 2012-02-05 13:51:32 +0000 | [diff] [blame] | 145 | return 0; |
| 146 | } |
| 147 | |
| 148 | /* Receives 4 types of messages: |
| 149 | * TCQ_PLUG_BUFFER: Inset a plug into the queue and |
| 150 | * buffer any incoming packets |
| 151 | * TCQ_PLUG_RELEASE_ONE: Dequeue packets from queue head |
| 152 | * to beginning of the next plug. |
| 153 | * TCQ_PLUG_RELEASE_INDEFINITE: Dequeue all packets from queue. |
| 154 | * Stop buffering packets until the next TCQ_PLUG_BUFFER |
| 155 | * command is received (just act as a pass-thru queue). |
| 156 | * TCQ_PLUG_LIMIT: Increase/decrease queue size |
| 157 | */ |
Alexander Aring | 2030721 | 2017-12-20 12:35:14 -0500 | [diff] [blame] | 158 | static int plug_change(struct Qdisc *sch, struct nlattr *opt, |
| 159 | struct netlink_ext_ack *extack) |
Shriram Rajagopalan | c3059be | 2012-02-05 13:51:32 +0000 | [diff] [blame] | 160 | { |
| 161 | struct plug_sched_data *q = qdisc_priv(sch); |
| 162 | struct tc_plug_qopt *msg; |
| 163 | |
Shriram Rajagopalan | c3059be | 2012-02-05 13:51:32 +0000 | [diff] [blame] | 164 | msg = nla_data(opt); |
| 165 | if (nla_len(opt) < sizeof(*msg)) |
| 166 | return -EINVAL; |
| 167 | |
| 168 | switch (msg->action) { |
| 169 | case TCQ_PLUG_BUFFER: |
| 170 | /* Save size of the current buffer */ |
| 171 | q->pkts_last_epoch = q->pkts_current_epoch; |
| 172 | q->pkts_current_epoch = 0; |
| 173 | if (q->unplug_indefinite) |
Eric Dumazet | 8fe6a79 | 2016-06-10 16:41:36 -0700 | [diff] [blame] | 174 | q->throttled = true; |
Shriram Rajagopalan | c3059be | 2012-02-05 13:51:32 +0000 | [diff] [blame] | 175 | q->unplug_indefinite = false; |
| 176 | break; |
| 177 | case TCQ_PLUG_RELEASE_ONE: |
| 178 | /* Add packets from the last complete buffer to the |
| 179 | * packets to be released set. |
| 180 | */ |
| 181 | q->pkts_to_release += q->pkts_last_epoch; |
| 182 | q->pkts_last_epoch = 0; |
Eric Dumazet | 8fe6a79 | 2016-06-10 16:41:36 -0700 | [diff] [blame] | 183 | q->throttled = false; |
Shriram Rajagopalan | c3059be | 2012-02-05 13:51:32 +0000 | [diff] [blame] | 184 | netif_schedule_queue(sch->dev_queue); |
| 185 | break; |
| 186 | case TCQ_PLUG_RELEASE_INDEFINITE: |
| 187 | q->unplug_indefinite = true; |
| 188 | q->pkts_to_release = 0; |
| 189 | q->pkts_last_epoch = 0; |
| 190 | q->pkts_current_epoch = 0; |
Eric Dumazet | 8fe6a79 | 2016-06-10 16:41:36 -0700 | [diff] [blame] | 191 | q->throttled = false; |
Shriram Rajagopalan | c3059be | 2012-02-05 13:51:32 +0000 | [diff] [blame] | 192 | netif_schedule_queue(sch->dev_queue); |
| 193 | break; |
| 194 | case TCQ_PLUG_LIMIT: |
| 195 | /* Limit is supplied in bytes */ |
| 196 | q->limit = msg->limit; |
| 197 | break; |
| 198 | default: |
| 199 | return -EINVAL; |
| 200 | } |
| 201 | |
| 202 | return 0; |
| 203 | } |
| 204 | |
Eric Dumazet | 2132cf6 | 2012-02-13 05:40:45 +0000 | [diff] [blame] | 205 | static struct Qdisc_ops plug_qdisc_ops __read_mostly = { |
Shriram Rajagopalan | c3059be | 2012-02-05 13:51:32 +0000 | [diff] [blame] | 206 | .id = "plug", |
| 207 | .priv_size = sizeof(struct plug_sched_data), |
| 208 | .enqueue = plug_enqueue, |
| 209 | .dequeue = plug_dequeue, |
valis | 8fc134f | 2023-09-01 12:22:37 -0400 | [diff] [blame] | 210 | .peek = qdisc_peek_dequeued, |
Shriram Rajagopalan | c3059be | 2012-02-05 13:51:32 +0000 | [diff] [blame] | 211 | .init = plug_init, |
| 212 | .change = plug_change, |
WANG Cong | fe6bea7 | 2015-07-21 16:31:53 -0700 | [diff] [blame] | 213 | .reset = qdisc_reset_queue, |
Shriram Rajagopalan | c3059be | 2012-02-05 13:51:32 +0000 | [diff] [blame] | 214 | .owner = THIS_MODULE, |
| 215 | }; |
Michal Koutný | 241a94a | 2024-02-01 14:09:41 +0100 | [diff] [blame] | 216 | MODULE_ALIAS_NET_SCH("plug"); |
Shriram Rajagopalan | c3059be | 2012-02-05 13:51:32 +0000 | [diff] [blame] | 217 | |
| 218 | static int __init plug_module_init(void) |
| 219 | { |
| 220 | return register_qdisc(&plug_qdisc_ops); |
| 221 | } |
| 222 | |
| 223 | static void __exit plug_module_exit(void) |
| 224 | { |
| 225 | unregister_qdisc(&plug_qdisc_ops); |
| 226 | } |
| 227 | module_init(plug_module_init) |
| 228 | module_exit(plug_module_exit) |
| 229 | MODULE_LICENSE("GPL"); |
Victor Nogueira | f96118c | 2023-10-27 08:50:45 -0700 | [diff] [blame] | 230 | MODULE_DESCRIPTION("Qdisc to plug and unplug traffic via netlink control"); |