xref: /linux/net/sched/sch_plug.c (revision e9f0878c4b2004ac19581274c1ae4c61ae3ca70e)
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 
60 struct 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 	bool throttled;
68 
69 	/* Queue Limit in bytes */
70 	u32 limit;
71 
72 	/* Number of packets (output) from the current speculatively
73 	 * executing epoch.
74 	 */
75 	u32 pkts_current_epoch;
76 
77 	/* Number of packets corresponding to the recently finished
78 	 * epoch. These will be released when we receive a
79 	 * TCQ_PLUG_RELEASE_ONE command. This command is typically
80 	 * issued after committing a checkpoint at the target.
81 	 */
82 	u32 pkts_last_epoch;
83 
84 	/*
85 	 * Number of packets from the head of the queue, that can
86 	 * be released (committed checkpoint).
87 	 */
88 	u32 pkts_to_release;
89 };
90 
91 static int plug_enqueue(struct sk_buff *skb, struct Qdisc *sch,
92 			struct sk_buff **to_free)
93 {
94 	struct plug_sched_data *q = qdisc_priv(sch);
95 
96 	if (likely(sch->qstats.backlog + skb->len <= q->limit)) {
97 		if (!q->unplug_indefinite)
98 			q->pkts_current_epoch++;
99 		return qdisc_enqueue_tail(skb, sch);
100 	}
101 
102 	return qdisc_drop(skb, sch, to_free);
103 }
104 
105 static struct sk_buff *plug_dequeue(struct Qdisc *sch)
106 {
107 	struct plug_sched_data *q = qdisc_priv(sch);
108 
109 	if (q->throttled)
110 		return NULL;
111 
112 	if (!q->unplug_indefinite) {
113 		if (!q->pkts_to_release) {
114 			/* No more packets to dequeue. Block the queue
115 			 * and wait for the next release command.
116 			 */
117 			q->throttled = true;
118 			return NULL;
119 		}
120 		q->pkts_to_release--;
121 	}
122 
123 	return qdisc_dequeue_head(sch);
124 }
125 
126 static int plug_init(struct Qdisc *sch, struct nlattr *opt,
127 		     struct netlink_ext_ack *extack)
128 {
129 	struct plug_sched_data *q = qdisc_priv(sch);
130 
131 	q->pkts_current_epoch = 0;
132 	q->pkts_last_epoch = 0;
133 	q->pkts_to_release = 0;
134 	q->unplug_indefinite = false;
135 
136 	if (opt == NULL) {
137 		q->limit = qdisc_dev(sch)->tx_queue_len
138 		           * 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 	q->throttled = true;
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  */
162 static int plug_change(struct Qdisc *sch, struct nlattr *opt,
163 		       struct netlink_ext_ack *extack)
164 {
165 	struct plug_sched_data *q = qdisc_priv(sch);
166 	struct tc_plug_qopt *msg;
167 
168 	if (opt == NULL)
169 		return -EINVAL;
170 
171 	msg = nla_data(opt);
172 	if (nla_len(opt) < sizeof(*msg))
173 		return -EINVAL;
174 
175 	switch (msg->action) {
176 	case TCQ_PLUG_BUFFER:
177 		/* Save size of the current buffer */
178 		q->pkts_last_epoch = q->pkts_current_epoch;
179 		q->pkts_current_epoch = 0;
180 		if (q->unplug_indefinite)
181 			q->throttled = true;
182 		q->unplug_indefinite = false;
183 		break;
184 	case TCQ_PLUG_RELEASE_ONE:
185 		/* Add packets from the last complete buffer to the
186 		 * packets to be released set.
187 		 */
188 		q->pkts_to_release += q->pkts_last_epoch;
189 		q->pkts_last_epoch = 0;
190 		q->throttled = false;
191 		netif_schedule_queue(sch->dev_queue);
192 		break;
193 	case TCQ_PLUG_RELEASE_INDEFINITE:
194 		q->unplug_indefinite = true;
195 		q->pkts_to_release = 0;
196 		q->pkts_last_epoch = 0;
197 		q->pkts_current_epoch = 0;
198 		q->throttled = false;
199 		netif_schedule_queue(sch->dev_queue);
200 		break;
201 	case TCQ_PLUG_LIMIT:
202 		/* Limit is supplied in bytes */
203 		q->limit = msg->limit;
204 		break;
205 	default:
206 		return -EINVAL;
207 	}
208 
209 	return 0;
210 }
211 
212 static struct Qdisc_ops plug_qdisc_ops __read_mostly = {
213 	.id          =       "plug",
214 	.priv_size   =       sizeof(struct plug_sched_data),
215 	.enqueue     =       plug_enqueue,
216 	.dequeue     =       plug_dequeue,
217 	.peek        =       qdisc_peek_head,
218 	.init        =       plug_init,
219 	.change      =       plug_change,
220 	.reset       =	     qdisc_reset_queue,
221 	.owner       =       THIS_MODULE,
222 };
223 
224 static int __init plug_module_init(void)
225 {
226 	return register_qdisc(&plug_qdisc_ops);
227 }
228 
229 static void __exit plug_module_exit(void)
230 {
231 	unregister_qdisc(&plug_qdisc_ops);
232 }
233 module_init(plug_module_init)
234 module_exit(plug_module_exit)
235 MODULE_LICENSE("GPL");
236