xref: /freebsd/sys/netinet/ip_dummynet.h (revision 6af83ee0d2941d18880b6aaa2b4facd1d30c6106)
1 /*-
2  * Copyright (c) 1998-2002 Luigi Rizzo, Universita` di Pisa
3  * Portions Copyright (c) 2000 Akamba Corp.
4  * All rights reserved
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25  * SUCH DAMAGE.
26  *
27  * $FreeBSD$
28  */
29 
30 #ifndef _IP_DUMMYNET_H
31 #define _IP_DUMMYNET_H
32 
33 /*
34  * Definition of dummynet data structures. In the structures, I decided
35  * not to use the macros in <sys/queue.h> in the hope of making the code
36  * easier to port to other architectures. The type of lists and queue we
37  * use here is pretty simple anyways.
38  */
39 
40 /*
41  * We start with a heap, which is used in the scheduler to decide when
42  * to transmit packets etc.
43  *
44  * The key for the heap is used for two different values:
45  *
46  * 1. timer ticks- max 10K/second, so 32 bits are enough;
47  *
48  * 2. virtual times. These increase in steps of len/x, where len is the
49  *    packet length, and x is either the weight of the flow, or the
50  *    sum of all weights.
51  *    If we limit to max 1000 flows and a max weight of 100, then
52  *    x needs 17 bits. The packet size is 16 bits, so we can easily
53  *    overflow if we do not allow errors.
54  * So we use a key "dn_key" which is 64 bits. Some macros are used to
55  * compare key values and handle wraparounds.
56  * MAX64 returns the largest of two key values.
57  * MY_M is used as a shift count when doing fixed point arithmetic
58  * (a better name would be useful...).
59  */
60 typedef u_int64_t dn_key ;      /* sorting key */
61 #define DN_KEY_LT(a,b)     ((int64_t)((a)-(b)) < 0)
62 #define DN_KEY_LEQ(a,b)    ((int64_t)((a)-(b)) <= 0)
63 #define DN_KEY_GT(a,b)     ((int64_t)((a)-(b)) > 0)
64 #define DN_KEY_GEQ(a,b)    ((int64_t)((a)-(b)) >= 0)
65 #define MAX64(x,y)  (( (int64_t) ( (y)-(x) )) > 0 ) ? (y) : (x)
66 #define MY_M	16 /* number of left shift to obtain a larger precision */
67 
68 /*
69  * XXX With this scaling, max 1000 flows, max weight 100, 1Gbit/s, the
70  * virtual time wraps every 15 days.
71  */
72 
73 /*
74  * The OFFSET_OF macro is used to return the offset of a field within
75  * a structure. It is used by the heap management routines.
76  */
77 #define OFFSET_OF(type, field) ((int)&( ((type *)0)->field) )
78 
79 /*
80  * The maximum hash table size for queues.  This value must be a power
81  * of 2.
82  */
83 #define DN_MAX_HASH_SIZE 65536
84 
85 /*
86  * A heap entry is made of a key and a pointer to the actual
87  * object stored in the heap.
88  * The heap is an array of dn_heap_entry entries, dynamically allocated.
89  * Current size is "size", with "elements" actually in use.
90  * The heap normally supports only ordered insert and extract from the top.
91  * If we want to extract an object from the middle of the heap, we
92  * have to know where the object itself is located in the heap (or we
93  * need to scan the whole array). To this purpose, an object has a
94  * field (int) which contains the index of the object itself into the
95  * heap. When the object is moved, the field must also be updated.
96  * The offset of the index in the object is stored in the 'offset'
97  * field in the heap descriptor. The assumption is that this offset
98  * is non-zero if we want to support extract from the middle.
99  */
100 struct dn_heap_entry {
101     dn_key key ;	/* sorting key. Topmost element is smallest one */
102     void *object ;	/* object pointer */
103 } ;
104 
105 struct dn_heap {
106     int size ;
107     int elements ;
108     int offset ; /* XXX if > 0 this is the offset of direct ptr to obj */
109     struct dn_heap_entry *p ;	/* really an array of "size" entries */
110 } ;
111 
112 #ifdef _KERNEL
113 /*
114  * Packets processed by dummynet have an mbuf tag associated with
115  * them that carries their dummynet state.  This is used within
116  * the dummynet code as well as outside when checking for special
117  * processing requirements.
118  */
119 struct dn_pkt_tag {
120     struct ip_fw *rule;		/* matching rule */
121     int dn_dir;			/* action when packet comes out. */
122 #define DN_TO_IP_OUT	1
123 #define DN_TO_IP_IN	2
124 #define DN_TO_BDG_FWD	3
125 #define DN_TO_ETH_DEMUX	4
126 #define DN_TO_ETH_OUT	5
127 
128     dn_key output_time;		/* when the pkt is due for delivery	*/
129     struct ifnet *ifp;		/* interface, for ip_output		*/
130     int flags ;			/* flags, for ip_output (IPv6 ?)	*/
131 };
132 #endif /* _KERNEL */
133 
134 /*
135  * Overall structure of dummynet (with WF2Q+):
136 
137 In dummynet, packets are selected with the firewall rules, and passed
138 to two different objects: PIPE or QUEUE.
139 
140 A QUEUE is just a queue with configurable size and queue management
141 policy. It is also associated with a mask (to discriminate among
142 different flows), a weight (used to give different shares of the
143 bandwidth to different flows) and a "pipe", which essentially
144 supplies the transmit clock for all queues associated with that
145 pipe.
146 
147 A PIPE emulates a fixed-bandwidth link, whose bandwidth is
148 configurable.  The "clock" for a pipe can come from either an
149 internal timer, or from the transmit interrupt of an interface.
150 A pipe is also associated with one (or more, if masks are used)
151 queue, where all packets for that pipe are stored.
152 
153 The bandwidth available on the pipe is shared by the queues
154 associated with that pipe (only one in case the packet is sent
155 to a PIPE) according to the WF2Q+ scheduling algorithm and the
156 configured weights.
157 
158 In general, incoming packets are stored in the appropriate queue,
159 which is then placed into one of a few heaps managed by a scheduler
160 to decide when the packet should be extracted.
161 The scheduler (a function called dummynet()) is run at every timer
162 tick, and grabs queues from the head of the heaps when they are
163 ready for processing.
164 
165 There are three data structures definining a pipe and associated queues:
166 
167  + dn_pipe, which contains the main configuration parameters related
168    to delay and bandwidth;
169  + dn_flow_set, which contains WF2Q+ configuration, flow
170    masks, plr and RED configuration;
171  + dn_flow_queue, which is the per-flow queue (containing the packets)
172 
173 Multiple dn_flow_set can be linked to the same pipe, and multiple
174 dn_flow_queue can be linked to the same dn_flow_set.
175 All data structures are linked in a linear list which is used for
176 housekeeping purposes.
177 
178 During configuration, we create and initialize the dn_flow_set
179 and dn_pipe structures (a dn_pipe also contains a dn_flow_set).
180 
181 At runtime: packets are sent to the appropriate dn_flow_set (either
182 WFQ ones, or the one embedded in the dn_pipe for fixed-rate flows),
183 which in turn dispatches them to the appropriate dn_flow_queue
184 (created dynamically according to the masks).
185 
186 The transmit clock for fixed rate flows (ready_event()) selects the
187 dn_flow_queue to be used to transmit the next packet. For WF2Q,
188 wfq_ready_event() extract a pipe which in turn selects the right
189 flow using a number of heaps defined into the pipe itself.
190 
191  *
192  */
193 
194 /*
195  * per flow queue. This contains the flow identifier, the queue
196  * of packets, counters, and parameters used to support both RED and
197  * WF2Q+.
198  *
199  * A dn_flow_queue is created and initialized whenever a packet for
200  * a new flow arrives.
201  */
202 struct dn_flow_queue {
203     struct dn_flow_queue *next ;
204     struct ipfw_flow_id id ;
205 
206     struct mbuf *head, *tail ;	/* queue of packets */
207     u_int len ;
208     u_int len_bytes ;
209     u_long numbytes ;		/* credit for transmission (dynamic queues) */
210 
211     u_int64_t tot_pkts ;	/* statistics counters	*/
212     u_int64_t tot_bytes ;
213     u_int32_t drops ;
214 
215     int hash_slot ;		/* debugging/diagnostic */
216 
217     /* RED parameters */
218     int avg ;                   /* average queue length est. (scaled) */
219     int count ;                 /* arrivals since last RED drop */
220     int random ;                /* random value (scaled) */
221     u_int32_t q_time ;          /* start of queue idle time */
222 
223     /* WF2Q+ support */
224     struct dn_flow_set *fs ;	/* parent flow set */
225     int heap_pos ;		/* position (index) of struct in heap */
226     dn_key sched_time ;		/* current time when queue enters ready_heap */
227 
228     dn_key S,F ;		/* start time, finish time */
229     /*
230      * Setting F < S means the timestamp is invalid. We only need
231      * to test this when the queue is empty.
232      */
233 } ;
234 
235 /*
236  * flow_set descriptor. Contains the "template" parameters for the
237  * queue configuration, and pointers to the hash table of dn_flow_queue's.
238  *
239  * The hash table is an array of lists -- we identify the slot by
240  * hashing the flow-id, then scan the list looking for a match.
241  * The size of the hash table (buckets) is configurable on a per-queue
242  * basis.
243  *
244  * A dn_flow_set is created whenever a new queue or pipe is created (in the
245  * latter case, the structure is located inside the struct dn_pipe).
246  */
247 struct dn_flow_set {
248     struct dn_flow_set *next; /* next flow set in all_flow_sets list */
249 
250     u_short fs_nr ;             /* flow_set number       */
251     u_short flags_fs;
252 #define DN_HAVE_FLOW_MASK	0x0001
253 #define DN_IS_RED		0x0002
254 #define DN_IS_GENTLE_RED	0x0004
255 #define DN_QSIZE_IS_BYTES	0x0008	/* queue size is measured in bytes */
256 #define DN_NOERROR		0x0010	/* do not report ENOBUFS on drops  */
257 #define DN_IS_PIPE		0x4000
258 #define DN_IS_QUEUE		0x8000
259 
260     struct dn_pipe *pipe ;	/* pointer to parent pipe */
261     u_short parent_nr ;		/* parent pipe#, 0 if local to a pipe */
262 
263     int weight ;		/* WFQ queue weight */
264     int qsize ;			/* queue size in slots or bytes */
265     int plr ;			/* pkt loss rate (2^31-1 means 100%) */
266 
267     struct ipfw_flow_id flow_mask ;
268 
269     /* hash table of queues onto this flow_set */
270     int rq_size ;		/* number of slots */
271     int rq_elements ;		/* active elements */
272     struct dn_flow_queue **rq;	/* array of rq_size entries */
273 
274     u_int32_t last_expired ;	/* do not expire too frequently */
275     int backlogged ;		/* #active queues for this flowset */
276 
277         /* RED parameters */
278 #define SCALE_RED               16
279 #define SCALE(x)                ( (x) << SCALE_RED )
280 #define SCALE_VAL(x)            ( (x) >> SCALE_RED )
281 #define SCALE_MUL(x,y)          ( ( (x) * (y) ) >> SCALE_RED )
282     int w_q ;			/* queue weight (scaled) */
283     int max_th ;		/* maximum threshold for queue (scaled) */
284     int min_th ;		/* minimum threshold for queue (scaled) */
285     int max_p ;			/* maximum value for p_b (scaled) */
286     u_int c_1 ;			/* max_p/(max_th-min_th) (scaled) */
287     u_int c_2 ;			/* max_p*min_th/(max_th-min_th) (scaled) */
288     u_int c_3 ;			/* for GRED, (1-max_p)/max_th (scaled) */
289     u_int c_4 ;			/* for GRED, 1 - 2*max_p (scaled) */
290     u_int * w_q_lookup ;	/* lookup table for computing (1-w_q)^t */
291     u_int lookup_depth ;	/* depth of lookup table */
292     int lookup_step ;		/* granularity inside the lookup table */
293     int lookup_weight ;		/* equal to (1-w_q)^t / (1-w_q)^(t+1) */
294     int avg_pkt_size ;		/* medium packet size */
295     int max_pkt_size ;		/* max packet size */
296 } ;
297 
298 /*
299  * Pipe descriptor. Contains global parameters, delay-line queue,
300  * and the flow_set used for fixed-rate queues.
301  *
302  * For WF2Q+ support it also has 3 heaps holding dn_flow_queue:
303  *   not_eligible_heap, for queues whose start time is higher
304  *	than the virtual time. Sorted by start time.
305  *   scheduler_heap, for queues eligible for scheduling. Sorted by
306  *	finish time.
307  *   idle_heap, all flows that are idle and can be removed. We
308  *	do that on each tick so we do not slow down too much
309  *	operations during forwarding.
310  *
311  */
312 struct dn_pipe {		/* a pipe */
313     struct dn_pipe *next ;
314 
315     int	pipe_nr ;		/* number	*/
316     int bandwidth;		/* really, bytes/tick.	*/
317     int	delay ;			/* really, ticks	*/
318 
319     struct	mbuf *head, *tail ;	/* packets in delay line */
320 
321     /* WF2Q+ */
322     struct dn_heap scheduler_heap ; /* top extract - key Finish time*/
323     struct dn_heap not_eligible_heap; /* top extract- key Start time */
324     struct dn_heap idle_heap ; /* random extract - key Start=Finish time */
325 
326     dn_key V ;			/* virtual time */
327     int sum;			/* sum of weights of all active sessions */
328     int numbytes;		/* bits I can transmit (more or less). */
329 
330     dn_key sched_time ;		/* time pipe was scheduled in ready_heap */
331 
332     /*
333      * When the tx clock come from an interface (if_name[0] != '\0'), its name
334      * is stored below, whereas the ifp is filled when the rule is configured.
335      */
336     char if_name[IFNAMSIZ];
337     struct ifnet *ifp ;
338     int ready ; /* set if ifp != NULL and we got a signal from it */
339 
340     struct dn_flow_set fs ; /* used with fixed-rate flows */
341 };
342 
343 #ifdef _KERNEL
344 typedef	int ip_dn_ctl_t(struct sockopt *); /* raw_ip.c */
345 typedef	void ip_dn_ruledel_t(void *); /* ip_fw.c */
346 typedef	int ip_dn_io_t(struct mbuf *m, int dir, struct ip_fw_args *fwa);
347 extern	ip_dn_ctl_t *ip_dn_ctl_ptr;
348 extern	ip_dn_ruledel_t *ip_dn_ruledel_ptr;
349 extern	ip_dn_io_t *ip_dn_io_ptr;
350 #define	DUMMYNET_LOADED	(ip_dn_io_ptr != NULL)
351 
352 /*
353  * Return the IPFW rule associated with the dummynet tag; if any.
354  * Make sure that the dummynet tag is not reused by lower layers.
355  */
356 static __inline struct ip_fw *
357 ip_dn_claim_rule(struct mbuf *m)
358 {
359 	struct m_tag *mtag = m_tag_find(m, PACKET_TAG_DUMMYNET, NULL);
360 	if (mtag != NULL) {
361 		mtag->m_tag_id = PACKET_TAG_NONE;
362 		return (((struct dn_pkt_tag *)(mtag+1))->rule);
363 	} else
364 		return (NULL);
365 }
366 #endif
367 #endif /* _IP_DUMMYNET_H */
368